fix(permissions): narrow rm -rf deny globs to their exact forms

The trailing wildcard on the three rm -rf deny entries spanned path
separators, so Bash(rm -rf /*) matched every absolute path on the
machine rather than the filesystem root, and the ~ and $HOME entries
had the same shape. Narrowed to the exact literal forms.

These rules match literal command text, so they still stop a typo on
rm -rf /, rm -rf ~ or rm -rf $HOME exactly, but they no longer stop a
recursive delete aimed at any other path. That reduced cover is
deliberate, not an oversight.

.claude/settings.json

@@ -0,0 +1,75 @@
+{
+  "env": {
+    "PQL_VAULT": "/mnt/media/Projects/tatlock"
+  },
+  "permissions": {
+    "allow": [
+      "Bash(pql)",
+      "Bash(pql *)",
+      "Bash(/home/jpmschweitzer/.local/bin/pql:*)",
+      "Bash(git status:*)",
+      "Bash(git log:*)",
+      "Bash(git diff:*)",
+      "Bash(git branch:*)",
+      "Bash(make test:*)",
+      "Bash(make test-unit:*)",
+      "Bash(make test-contracts:*)",
+      "Bash(make lint:*)",
+      "Bash(make typecheck:*)",
+      "Bash(.venv/bin/python -m pytest:*)",
+      "Bash(.venv/bin/pytest:*)",
+      "Bash(pytest:*)",
+      "Bash(ruff check:*)",
+      "Bash(mypy:*)",
+      "Bash(docker logs tatlock:*)",
+      "Bash(curl -s http://localhost:8000/*)",
+      "Bash(curl -s http://localhost:8777/*)"
+    ],
+    "deny": [
+      "Bash(/mnt/media/Projects/cladmin/ops/bin/toj)",
+      "Bash(/mnt/media/Projects/cladmin/ops/bin/toj:*)",
+      "Bash(chmod -R 777 *)",
+      "Bash(chmod 777 *)",
+      "Bash(dd if=*)",
+      "Bash(find * -delete*)",
+      "Bash(find * -exec*)",
+      "Bash(git * add --all*)",
+      "Bash(git * add -A*)",
+      "Bash(git * add .)",
+      "Bash(git * branch -D *)",
+      "Bash(git * checkout -- *)",
+      "Bash(git * clean -fd*)",
+      "Bash(git * clean -fdx*)",
+      "Bash(git * commit --no-verify*)",
+      "Bash(git * merge --no-ff*)",
+      "Bash(git * push --force*)",
+      "Bash(git * push -f*)",
+      "Bash(git * reset --hard*)",
+      "Bash(git * restore .*)",
+      "Bash(git add --all*)",
+      "Bash(git add -A*)",
+      "Bash(git add .)",
+      "Bash(git branch -D *)",
+      "Bash(git checkout -- *)",
+      "Bash(git clean -fd*)",
+      "Bash(git clean -fdx*)",
+      "Bash(git commit --no-verify*)",
+      "Bash(git merge --no-ff*)",
+      "Bash(git push --force*)",
+      "Bash(git push -f*)",
+      "Bash(git reset --hard*)",
+      "Bash(git restore .*)",
+      "Bash(mkfs*)",
+      "Bash(ollama rm *)",
+      "Bash(redis-cli * FLUSHALL*)",
+      "Bash(redis-cli * FLUSHDB*)",
+      "Bash(rm -rf $HOME)",
+      "Bash(rm -rf /)",
+      "Bash(rm -rf ~)",
+      "Bash(su *)",
+      "Bash(sudo *)",
+      "Bash(toj)",
+      "Bash(toj:*)"
+    ]
+  }
+}

.env.example

@@ -1,6 +1,5 @@
 # Application Configuration
 APP_NAME="OpenAI-Compatible API"
-APP_VERSION="0.1.0"
 ENVIRONMENT=development
 DEBUG=false
 
@@ -9,25 +8,56 @@ API_HOST=0.0.0.0
 API_PORT=8000
 API_PREFIX=/v1
 
-# Ollama Configuration
-OLLAMA_HOST=http://your-ollama-host:11434
-OLLAMA_DEFAULT_MODEL=mistral-nemo:latest
+# Ollama Configuration (local - primary backend)
+OLLAMA_HOST=http://localhost:11434
+OLLAMA_DEFAULT_MODEL=gemma4:e2b
 OLLAMA_TIMEOUT=120
+STEWARD_TIMEOUT=60
+
+# Anthropic Configuration (Claude - cloud fallback)
+# Set ANTHROPIC_API_KEY to keep the Claude fallback available: it is used
+# automatically when Ollama is down, or exclusively when PREFER_CLOUD_BACKEND=true
+# Without an API key, Tatlock uses Ollama only
+# ANTHROPIC_API_KEY=sk-ant-api03-your-key-here
+ANTHROPIC_MODEL=claude-sonnet-5
+PREFER_CLOUD_BACKEND=false
 
 # SearXNG Configuration
-SEARXNG_HOST=http://searxng:8087
+SEARXNG_HOST=http://localhost:8087
 SEARXNG_TIMEOUT=30
 
 # Redis Configuration
-REDIS_HOST=redis-shared
+REDIS_HOST=localhost
 REDIS_PORT=6379
-REDIS_DB=1
+REDIS_MEMORY_DB=1
 REDIS_TIMEOUT=5
 
+# Qdrant Configuration
+QDRANT_HOST=localhost
+QDRANT_PORT=6333
+
 # Logging
-LOG_LEVEL=INFO
-ENABLE_BENCHMARKS=true
+# LOG_LEVEL is auto-selected based on ENVIRONMENT if not set:
+#   - development: DEBUG (maximum verbosity)
+#   - production: WARNING (minimal noise)
+# Uncomment to override: LOG_LEVEL=INFO
 # Note: Log format is auto-selected based on ENVIRONMENT (console for dev, json for production)
 
+# User Configuration
+# DEFAULT_USER is auto-selected based on ENVIRONMENT if not set:
+#   - development/testing: llm_tester (isolated test scope)
+#   - production: jpmschweitzer (real user)
+# Uncomment to override: DEFAULT_USER=your_username
+
+# Library-Desk Configuration (The Librarian backend)
+# LIBRARY_DESK_HOST=http://localhost:8089
+# LIBRARY_DESK_API_KEY=your-library-desk-api-key
+# LIBRARY_DESK_TIMEOUT=60
+
+# Core-API Configuration (The Housekeeper backend)
+# CORE_API_HOST=http://localhost:8090
+# CORE_API_KEY=your-core-api-key
+# CORE_API_TIMEOUT=30
+
 # CORS (comma-separated list)
-CORS_ORIGINS=*
+CORS_ORIGINS=["*"]

.gitattributes

@@ -0,0 +1 @@
+.pql/changelog/*.sql merge=union

.gitea/workflows/build.yml

@@ -1,10 +1,22 @@
 name: Build and Push
 
 on:
-  release:
-    types: [published]
+  push:
+    tags:
+      - 'v[0-9]*'
 
 jobs:
+  release:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Create Gitea Release
+        run: |
+          curl -sf -X POST \
+            -H "Authorization: token ${{ secrets.GITHUB_TOKEN }}" \
+            -H "Content-Type: application/json" \
+            -d '{"tag_name": "${{ github.ref_name }}", "name": "Release ${{ github.ref_name }}", "body": "Automated release for ${{ github.ref_name }}"}' \
+            "${{ github.server_url }}/api/v1/repos/${{ github.repository }}/releases"
+
   build:
     runs-on: ubuntu-latest
     steps:
@@ -18,10 +30,18 @@ jobs:
           password: ${{ secrets.REGISTRY_PASSWORD }}
 
       - name: Build and push
-        uses: docker/build-push-action@v5
+        uses: docker/build-push-action@v6
         with:
           context: .
           push: true
+          provenance: false
+          sbom: false
           tags: |
             git.schweitz.net/jpmschweitzer/tatlock:latest
             git.schweitz.net/jpmschweitzer/tatlock:${{ github.ref_name }}
+
+      - name: Trigger Watchtower update
+        if: success()
+        run: |
+          curl -sf -H "Authorization: Bearer ${{ secrets.WATCHTOWER_TOKEN }}" \
+            http://watchtower:8080/v1/update

.githooks/pre-push

@@ -0,0 +1,13 @@
+#!/usr/bin/env bash
+# Trigger only. The checks live in the Makefile, where they can be read, run by
+# hand (`make pre-push`), and changed under review.
+#
+# This file is identical in every repo in this workspace, deliberately: the call
+# surface is the same everywhere even though what each gate runs is not, so
+# nobody has to read a repo to find out how to check it (D-27).
+#
+# Enable per clone with:  git config core.hooksPath .githooks
+# Never bypass with --no-verify. Suppress a specific finding deliberately
+# instead, with a reason — see `make pre-push`.
+set -euo pipefail
+exec make -C "$(git rev-parse --show-toplevel)" pre-push

.gitignore

@@ -46,29 +46,37 @@ ENV/
 .ipynb_checkpoints/
 *.ipynb
 
-# Testing & Coverage
+# Caches (pytest, mypy, ruff)
+.cache/
+
+# Build output (coverage, logs)
+build/
+
+# Legacy cache/output locations (in case tools fall back)
 .pytest_cache/
+.mypy_cache/
+.ruff_cache/
 .coverage
-.coverage.*
 coverage.xml
 htmlcov/
+
+# Testing
 .tox/
 .nox/
 *.cover
 .hypothesis/
 
 # Type checking
-.mypy_cache/
 .dmypy.json
 dmypy.json
 .pyre/
 .pytype/
 
-# Linting
-.ruff_cache/
-
 # Logs
-logs/
+logs/*
+!logs/traces/
+logs/traces/*
+!logs/traces/viewer.html
 *.log
 
 # Database
@@ -95,3 +103,16 @@ ollama_data/
 ehthumbs.db
 Thumbs.db
 Desktop.ini
+
+# Claude Code user-specific settings
+.claude/settings.local.json
+.pql/*
+!.pql/changelog/
+
+# pql shims planted by `pql init` into the dir core.hooksPath points at.
+# Per-clone: each embeds the absolute path of the pql binary that planted it.
+# Only .githooks/pre-push is shared.
+.githooks/pre-commit
+.githooks/post-merge
+.githooks/post-checkout
+.githooks/post-rewrite

.pql/changelog/0000-format.sql

@@ -0,0 +1,11 @@
+-- Changelog format marker, written by pql. Comments only: this file
+-- is never executed — Import descends into the per-table directories
+-- and does not read the changelog root.
+--
+-- A changelog carrying no marker is format 1, the shape that existed
+-- before formats were versioned. An older format is migrated forward
+-- by `pql plan upgrade` (and automatically from the post-merge hook);
+-- a newer one is refused rather than replayed under rules this binary
+-- does not know. See D-28 and docs/versions.md.
+-- pql:changelog_format: 2.0.0
+-- pql:written_by: 2.2.0

.pql/changelog/ticket_deps/0000-schema.sql

@@ -0,0 +1,139 @@
+-- Auto-generated by pql init. CREATE TABLE statements
+-- for the planning schema; per-table dir keeps the changelog
+-- self-describing per D-15. CREATE TABLE IF NOT EXISTS is
+-- idempotent so running schema files from each directory in
+-- replay order is harmless.
+--
+-- Importer parses the markers below to detect schema drift
+-- between the producing pql version and the local one — a
+-- bumped canonical_version means projection rules changed
+-- and replay must refuse rather than silently corrupt state.
+-- pql:created_by: 2.2.0
+-- pql:canonical_version: 2
+
+
+CREATE TABLE IF NOT EXISTS decisions (
+    id                TEXT PRIMARY KEY,
+    type              TEXT NOT NULL CHECK(type IN ('confirmed','question','rejected')),
+    domain            TEXT NOT NULL,
+    title             TEXT NOT NULL,
+    status            TEXT NOT NULL DEFAULT 'active'
+                          CHECK(status IN ('active','superseded','resolved','open')),
+    date              TEXT,
+    file_path         TEXT NOT NULL,
+    synced_at         TEXT NOT NULL DEFAULT (datetime('now')),
+    created_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    updated_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    deleted_at        TEXT,
+    hash              TEXT,
+    canonical_version INTEGER
+);
+
+CREATE TABLE IF NOT EXISTS decision_refs (
+    source_id         TEXT NOT NULL REFERENCES decisions(id) ON DELETE CASCADE,
+    target_id         TEXT NOT NULL REFERENCES decisions(id) ON DELETE CASCADE,
+    ref_type          TEXT NOT NULL
+                          CHECK(ref_type IN ('supersedes','references','resolves','depends_on','amends')),
+    note              TEXT,
+    created_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    updated_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    deleted_at        TEXT,
+    hash              TEXT,
+    canonical_version INTEGER,
+    PRIMARY KEY (source_id, target_id, ref_type)
+);
+
+-- Identity split (D-26): a ticket's stable, collision-proof identity is its
+-- record_id (a locally-generated ULID, planning.NewRecordID); the friendly
+-- T-NNN label lives in ticket_idmap and may be reconciled. Every structural
+-- reference (parent, deps, history, labels) targets record_id, so a label
+-- clash never corrupts the graph — only ticket_idmap needs a relabel.
+CREATE TABLE IF NOT EXISTS tickets (
+    record_id         TEXT PRIMARY KEY,
+    type              TEXT NOT NULL CHECK(type IN ('initiative','epic','story','task','bug')),
+    parent_record_id  TEXT REFERENCES tickets(record_id),
+    title             TEXT NOT NULL,
+    description       TEXT,
+    -- No CHECK enumeration: the ticket status vocabulary is per-vault
+    -- configurable (ticket_statuses in .pql/config.yaml). Validation lives
+    -- in Go (planning.StatusSet), so adding/renaming statuses needs no
+    -- schema change. The DEFAULT is a harmless fallback — CreateTicket
+    -- always inserts the configured default explicitly.
+    status            TEXT NOT NULL DEFAULT 'backlog',
+    priority          TEXT DEFAULT 'medium'
+                          CHECK(priority IN ('critical','high','medium','low')),
+    assigned_to       TEXT,
+    team              TEXT,
+    decision_ref      TEXT REFERENCES decisions(id),
+    created_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    updated_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    deleted_at        TEXT,
+    hash              TEXT,
+    canonical_version INTEGER
+);
+
+-- ticket_idmap maps a record_id to its current friendly label (T-NNN).
+-- ticket_id is intentionally NOT globally unique: two uncoordinated clones
+-- can mint the same label, which surfaces as a duplicate-label collision
+-- (detected at replay) and is fixed with "pql ticket relabel".
+CREATE TABLE IF NOT EXISTS ticket_idmap (
+    record_id         TEXT PRIMARY KEY REFERENCES tickets(record_id),
+    ticket_id         TEXT NOT NULL,
+    created_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    updated_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    deleted_at        TEXT,
+    hash              TEXT,
+    canonical_version INTEGER
+);
+
+CREATE TABLE IF NOT EXISTS ticket_deps (
+    blocker_record_id TEXT NOT NULL REFERENCES tickets(record_id),
+    blocked_record_id TEXT NOT NULL REFERENCES tickets(record_id),
+    created_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    updated_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    deleted_at        TEXT,
+    hash              TEXT,
+    canonical_version INTEGER,
+    PRIMARY KEY (blocker_record_id, blocked_record_id)
+);
+
+CREATE TABLE IF NOT EXISTS ticket_history (
+    ticket_record_id  TEXT NOT NULL REFERENCES tickets(record_id),
+    field             TEXT NOT NULL,
+    old_value         TEXT,
+    new_value         TEXT,
+    changed_by        TEXT,
+    changed_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    created_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    updated_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    deleted_at        TEXT,
+    hash              TEXT UNIQUE,
+    canonical_version INTEGER
+);
+
+CREATE TABLE IF NOT EXISTS ticket_labels (
+    ticket_record_id  TEXT NOT NULL REFERENCES tickets(record_id),
+    label             TEXT NOT NULL,
+    created_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    updated_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    deleted_at        TEXT,
+    hash              TEXT,
+    canonical_version INTEGER,
+    PRIMARY KEY (ticket_record_id, label)
+);
+
+CREATE TABLE IF NOT EXISTS meta (
+    key        TEXT PRIMARY KEY,
+    value      TEXT NOT NULL,
+    updated_at TEXT NOT NULL DEFAULT (datetime('now'))
+);
+
+CREATE INDEX IF NOT EXISTS idx_tickets_status        ON tickets(status);
+CREATE INDEX IF NOT EXISTS idx_tickets_team          ON tickets(team);
+CREATE INDEX IF NOT EXISTS idx_tickets_decision_ref  ON tickets(decision_ref);
+CREATE INDEX IF NOT EXISTS idx_tickets_assigned      ON tickets(assigned_to);
+CREATE INDEX IF NOT EXISTS idx_tickets_parent        ON tickets(parent_record_id);
+CREATE INDEX IF NOT EXISTS idx_ticket_idmap_label    ON ticket_idmap(ticket_id);
+CREATE INDEX IF NOT EXISTS idx_decisions_domain      ON decisions(domain);
+CREATE INDEX IF NOT EXISTS idx_decisions_type        ON decisions(type);
+CREATE INDEX IF NOT EXISTS idx_decision_refs_target  ON decision_refs(target_id);

.pql/changelog/ticket_history/0000-schema.sql

@@ -0,0 +1,139 @@
+-- Auto-generated by pql init. CREATE TABLE statements
+-- for the planning schema; per-table dir keeps the changelog
+-- self-describing per D-15. CREATE TABLE IF NOT EXISTS is
+-- idempotent so running schema files from each directory in
+-- replay order is harmless.
+--
+-- Importer parses the markers below to detect schema drift
+-- between the producing pql version and the local one — a
+-- bumped canonical_version means projection rules changed
+-- and replay must refuse rather than silently corrupt state.
+-- pql:created_by: 2.2.0
+-- pql:canonical_version: 2
+
+
+CREATE TABLE IF NOT EXISTS decisions (
+    id                TEXT PRIMARY KEY,
+    type              TEXT NOT NULL CHECK(type IN ('confirmed','question','rejected')),
+    domain            TEXT NOT NULL,
+    title             TEXT NOT NULL,
+    status            TEXT NOT NULL DEFAULT 'active'
+                          CHECK(status IN ('active','superseded','resolved','open')),
+    date              TEXT,
+    file_path         TEXT NOT NULL,
+    synced_at         TEXT NOT NULL DEFAULT (datetime('now')),
+    created_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    updated_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    deleted_at        TEXT,
+    hash              TEXT,
+    canonical_version INTEGER
+);
+
+CREATE TABLE IF NOT EXISTS decision_refs (
+    source_id         TEXT NOT NULL REFERENCES decisions(id) ON DELETE CASCADE,
+    target_id         TEXT NOT NULL REFERENCES decisions(id) ON DELETE CASCADE,
+    ref_type          TEXT NOT NULL
+                          CHECK(ref_type IN ('supersedes','references','resolves','depends_on','amends')),
+    note              TEXT,
+    created_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    updated_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    deleted_at        TEXT,
+    hash              TEXT,
+    canonical_version INTEGER,
+    PRIMARY KEY (source_id, target_id, ref_type)
+);
+
+-- Identity split (D-26): a ticket's stable, collision-proof identity is its
+-- record_id (a locally-generated ULID, planning.NewRecordID); the friendly
+-- T-NNN label lives in ticket_idmap and may be reconciled. Every structural
+-- reference (parent, deps, history, labels) targets record_id, so a label
+-- clash never corrupts the graph — only ticket_idmap needs a relabel.
+CREATE TABLE IF NOT EXISTS tickets (
+    record_id         TEXT PRIMARY KEY,
+    type              TEXT NOT NULL CHECK(type IN ('initiative','epic','story','task','bug')),
+    parent_record_id  TEXT REFERENCES tickets(record_id),
+    title             TEXT NOT NULL,
+    description       TEXT,
+    -- No CHECK enumeration: the ticket status vocabulary is per-vault
+    -- configurable (ticket_statuses in .pql/config.yaml). Validation lives
+    -- in Go (planning.StatusSet), so adding/renaming statuses needs no
+    -- schema change. The DEFAULT is a harmless fallback — CreateTicket
+    -- always inserts the configured default explicitly.
+    status            TEXT NOT NULL DEFAULT 'backlog',
+    priority          TEXT DEFAULT 'medium'
+                          CHECK(priority IN ('critical','high','medium','low')),
+    assigned_to       TEXT,
+    team              TEXT,
+    decision_ref      TEXT REFERENCES decisions(id),
+    created_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    updated_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    deleted_at        TEXT,
+    hash              TEXT,
+    canonical_version INTEGER
+);
+
+-- ticket_idmap maps a record_id to its current friendly label (T-NNN).
+-- ticket_id is intentionally NOT globally unique: two uncoordinated clones
+-- can mint the same label, which surfaces as a duplicate-label collision
+-- (detected at replay) and is fixed with "pql ticket relabel".
+CREATE TABLE IF NOT EXISTS ticket_idmap (
+    record_id         TEXT PRIMARY KEY REFERENCES tickets(record_id),
+    ticket_id         TEXT NOT NULL,
+    created_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    updated_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    deleted_at        TEXT,
+    hash              TEXT,
+    canonical_version INTEGER
+);
+
+CREATE TABLE IF NOT EXISTS ticket_deps (
+    blocker_record_id TEXT NOT NULL REFERENCES tickets(record_id),
+    blocked_record_id TEXT NOT NULL REFERENCES tickets(record_id),
+    created_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    updated_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    deleted_at        TEXT,
+    hash              TEXT,
+    canonical_version INTEGER,
+    PRIMARY KEY (blocker_record_id, blocked_record_id)
+);
+
+CREATE TABLE IF NOT EXISTS ticket_history (
+    ticket_record_id  TEXT NOT NULL REFERENCES tickets(record_id),
+    field             TEXT NOT NULL,
+    old_value         TEXT,
+    new_value         TEXT,
+    changed_by        TEXT,
+    changed_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    created_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    updated_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    deleted_at        TEXT,
+    hash              TEXT UNIQUE,
+    canonical_version INTEGER
+);
+
+CREATE TABLE IF NOT EXISTS ticket_labels (
+    ticket_record_id  TEXT NOT NULL REFERENCES tickets(record_id),
+    label             TEXT NOT NULL,
+    created_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    updated_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    deleted_at        TEXT,
+    hash              TEXT,
+    canonical_version INTEGER,
+    PRIMARY KEY (ticket_record_id, label)
+);
+
+CREATE TABLE IF NOT EXISTS meta (
+    key        TEXT PRIMARY KEY,
+    value      TEXT NOT NULL,
+    updated_at TEXT NOT NULL DEFAULT (datetime('now'))
+);
+
+CREATE INDEX IF NOT EXISTS idx_tickets_status        ON tickets(status);
+CREATE INDEX IF NOT EXISTS idx_tickets_team          ON tickets(team);
+CREATE INDEX IF NOT EXISTS idx_tickets_decision_ref  ON tickets(decision_ref);
+CREATE INDEX IF NOT EXISTS idx_tickets_assigned      ON tickets(assigned_to);
+CREATE INDEX IF NOT EXISTS idx_tickets_parent        ON tickets(parent_record_id);
+CREATE INDEX IF NOT EXISTS idx_ticket_idmap_label    ON ticket_idmap(ticket_id);
+CREATE INDEX IF NOT EXISTS idx_decisions_domain      ON decisions(domain);
+CREATE INDEX IF NOT EXISTS idx_decisions_type        ON decisions(type);
+CREATE INDEX IF NOT EXISTS idx_decision_refs_target  ON decision_refs(target_id);

.pql/changelog/ticket_history/2026-08.sql

@@ -0,0 +1,34 @@
+INSERT INTO ticket_history (ticket_record_id, field, old_value, new_value, changed_by, changed_at, created_at, updated_at, deleted_at, hash, canonical_version) VALUES ('06FZ44Z6HSN0RQS0QAYNTPEM5G', 'description', NULL, '`make typecheck` reports 95 errors in 31 files (was 103). This is a dedicated programming pass, not lint tidying, and it is what currently blocks `make pre-push`.
+
+MEASURED 2026-08-11 so the next session does not re-derive it:
+
+  29 no-untyped-def   functions with no annotations — the bulk, and genuine per-function work
+  18 no-any-return    mostly downstream of the above
+  12 assignment
+  11 arg-type
+   5 unused-ignore    `# type: ignore` comments mypy says are no longer needed
+   5 union-attr
+   4 var-annotated
+   3 override
+   remainder: misc, dict-item, attr-defined, return-value, call-overload
+
+By file: agents/tatlock.py 15, core/memory_service.py 11, responses/streaming.py 8, responses/service.py 7, core/context.py 6.
+
+THE TWO SHARED ROOTS ARE ALREADY FIXED (dac259a), so what is left has no lever in it. For reference, they were: five conversation lists declared bare, where mypy infers the element type from the first append (a ModelRequest) and then rejects every ModelResponse; and an agent built as Agent(model, system_prompt=...) with no deps_type, inferred Agent[None, str], while every tool it registers takes RunContext[ToolCallTracker].
+
+WORTH KNOWING BEFORE STARTING. Annotating partially made mypy count go UP before it went down — declaring `_agent: Agent | None` took agents/tatlock.py from 22 to 24, because resolving the bare Agent to Agent[None, str] surfaced four argument-type errors the Any had been hiding. Expect that shape: a rising count during this work usually means concealment ending, not damage.
+
+The mypy config is strict — disallow_untyped_defs, disallow_incomplete_defs, warn_return_any, check_untyped_defs, strict_equality — so there is no partial-credit setting to lean on, and weakening it would be the wrong trade for a codebase this central.
+
+TWO PRE-EXISTING TEST FACTS, both confirmed at HEAD and neither caused by the lint work:
+  - test_tatlock_tool_call_logging_calculator is flaky: failed 2 of 5 full runs, on HEAD and on the lint branch, and fails in isolation at HEAD while passing in isolation after the lint pass. Order- or timing-dependent.
+  - `pytest tests/` cannot collect at all: tests/e2e/test_orchestration_e2e.py uses an `e2e` marker that is not registered and the config is strict about markers. `make test` passes only because it ignores tests/e2e, tests/integration and tests/contracts.', NULL, '2026-08-11 18:26:58', '2026-08-11 18:26:58.523', '2026-08-11 18:26:58.523', NULL, 'c8c9b6a20cf18ca903fc8c720f70e73d', 2) ON CONFLICT(hash) DO NOTHING;
+INSERT INTO ticket_history (ticket_record_id, field, old_value, new_value, changed_by, changed_at, created_at, updated_at, deleted_at, hash, canonical_version) VALUES ('06FZ4JX9DZB3XAB95EP23YWXMC', 'description', NULL, '`.env` still carries an `ANTHROPIC_API_KEY`. That key was revoked and no longer exists on Anthropic''s side, so this is dead weight rather than an exposure — but it is dead weight that reads exactly like a live credential to anyone who finds it.
+
+The cost is confusion, not risk. Someone debugging a Claude fallback will find a key present, assume it is configured, and look elsewhere for the failure. The absence of a key is a clear signal; a revoked key is a misleading one.
+
+`.env` is gitignored here and has never been committed, so nothing needs rewriting — the value simply needs removing from the local file, and the line dropping or blanking in `.env.example` if it appears there too.
+
+RELATED, and the reason this is filed separately: the workspace vault holds T-13, "Clear the revoked ANTHROPIC_API_KEY from the live Portainer stack". That ticket is scoped to the Portainer stack only. Whoever closes it will reasonably believe the key is gone once the stack is clean, and this copy will survive. The two want doing together even though they commit separately.
+
+Context on the revocation, since it explains why nobody removed this at the time: the key was revoked on 2026-08-09 after being printed into a transcript by a redaction filter that matched on `KEY` appearing after the `=`. In `ANTHROPIC_API_KEY=...` it appears before, so the filter never fired. The response was rotation, and the leftover copies were not swept.', NULL, '2026-08-11 19:27:52', '2026-08-11 19:27:52.823', '2026-08-11 19:27:52.823', NULL, 'c7834e46268029b74c25a25a83177b64', 2) ON CONFLICT(hash) DO NOTHING;

.pql/changelog/ticket_idmap/0000-schema.sql

@@ -0,0 +1,139 @@
+-- Auto-generated by pql init. CREATE TABLE statements
+-- for the planning schema; per-table dir keeps the changelog
+-- self-describing per D-15. CREATE TABLE IF NOT EXISTS is
+-- idempotent so running schema files from each directory in
+-- replay order is harmless.
+--
+-- Importer parses the markers below to detect schema drift
+-- between the producing pql version and the local one — a
+-- bumped canonical_version means projection rules changed
+-- and replay must refuse rather than silently corrupt state.
+-- pql:created_by: 2.2.0
+-- pql:canonical_version: 2
+
+
+CREATE TABLE IF NOT EXISTS decisions (
+    id                TEXT PRIMARY KEY,
+    type              TEXT NOT NULL CHECK(type IN ('confirmed','question','rejected')),
+    domain            TEXT NOT NULL,
+    title             TEXT NOT NULL,
+    status            TEXT NOT NULL DEFAULT 'active'
+                          CHECK(status IN ('active','superseded','resolved','open')),
+    date              TEXT,
+    file_path         TEXT NOT NULL,
+    synced_at         TEXT NOT NULL DEFAULT (datetime('now')),
+    created_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    updated_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    deleted_at        TEXT,
+    hash              TEXT,
+    canonical_version INTEGER
+);
+
+CREATE TABLE IF NOT EXISTS decision_refs (
+    source_id         TEXT NOT NULL REFERENCES decisions(id) ON DELETE CASCADE,
+    target_id         TEXT NOT NULL REFERENCES decisions(id) ON DELETE CASCADE,
+    ref_type          TEXT NOT NULL
+                          CHECK(ref_type IN ('supersedes','references','resolves','depends_on','amends')),
+    note              TEXT,
+    created_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    updated_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    deleted_at        TEXT,
+    hash              TEXT,
+    canonical_version INTEGER,
+    PRIMARY KEY (source_id, target_id, ref_type)
+);
+
+-- Identity split (D-26): a ticket's stable, collision-proof identity is its
+-- record_id (a locally-generated ULID, planning.NewRecordID); the friendly
+-- T-NNN label lives in ticket_idmap and may be reconciled. Every structural
+-- reference (parent, deps, history, labels) targets record_id, so a label
+-- clash never corrupts the graph — only ticket_idmap needs a relabel.
+CREATE TABLE IF NOT EXISTS tickets (
+    record_id         TEXT PRIMARY KEY,
+    type              TEXT NOT NULL CHECK(type IN ('initiative','epic','story','task','bug')),
+    parent_record_id  TEXT REFERENCES tickets(record_id),
+    title             TEXT NOT NULL,
+    description       TEXT,
+    -- No CHECK enumeration: the ticket status vocabulary is per-vault
+    -- configurable (ticket_statuses in .pql/config.yaml). Validation lives
+    -- in Go (planning.StatusSet), so adding/renaming statuses needs no
+    -- schema change. The DEFAULT is a harmless fallback — CreateTicket
+    -- always inserts the configured default explicitly.
+    status            TEXT NOT NULL DEFAULT 'backlog',
+    priority          TEXT DEFAULT 'medium'
+                          CHECK(priority IN ('critical','high','medium','low')),
+    assigned_to       TEXT,
+    team              TEXT,
+    decision_ref      TEXT REFERENCES decisions(id),
+    created_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    updated_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    deleted_at        TEXT,
+    hash              TEXT,
+    canonical_version INTEGER
+);
+
+-- ticket_idmap maps a record_id to its current friendly label (T-NNN).
+-- ticket_id is intentionally NOT globally unique: two uncoordinated clones
+-- can mint the same label, which surfaces as a duplicate-label collision
+-- (detected at replay) and is fixed with "pql ticket relabel".
+CREATE TABLE IF NOT EXISTS ticket_idmap (
+    record_id         TEXT PRIMARY KEY REFERENCES tickets(record_id),
+    ticket_id         TEXT NOT NULL,
+    created_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    updated_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    deleted_at        TEXT,
+    hash              TEXT,
+    canonical_version INTEGER
+);
+
+CREATE TABLE IF NOT EXISTS ticket_deps (
+    blocker_record_id TEXT NOT NULL REFERENCES tickets(record_id),
+    blocked_record_id TEXT NOT NULL REFERENCES tickets(record_id),
+    created_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    updated_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    deleted_at        TEXT,
+    hash              TEXT,
+    canonical_version INTEGER,
+    PRIMARY KEY (blocker_record_id, blocked_record_id)
+);
+
+CREATE TABLE IF NOT EXISTS ticket_history (
+    ticket_record_id  TEXT NOT NULL REFERENCES tickets(record_id),
+    field             TEXT NOT NULL,
+    old_value         TEXT,
+    new_value         TEXT,
+    changed_by        TEXT,
+    changed_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    created_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    updated_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    deleted_at        TEXT,
+    hash              TEXT UNIQUE,
+    canonical_version INTEGER
+);
+
+CREATE TABLE IF NOT EXISTS ticket_labels (
+    ticket_record_id  TEXT NOT NULL REFERENCES tickets(record_id),
+    label             TEXT NOT NULL,
+    created_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    updated_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    deleted_at        TEXT,
+    hash              TEXT,
+    canonical_version INTEGER,
+    PRIMARY KEY (ticket_record_id, label)
+);
+
+CREATE TABLE IF NOT EXISTS meta (
+    key        TEXT PRIMARY KEY,
+    value      TEXT NOT NULL,
+    updated_at TEXT NOT NULL DEFAULT (datetime('now'))
+);
+
+CREATE INDEX IF NOT EXISTS idx_tickets_status        ON tickets(status);
+CREATE INDEX IF NOT EXISTS idx_tickets_team          ON tickets(team);
+CREATE INDEX IF NOT EXISTS idx_tickets_decision_ref  ON tickets(decision_ref);
+CREATE INDEX IF NOT EXISTS idx_tickets_assigned      ON tickets(assigned_to);
+CREATE INDEX IF NOT EXISTS idx_tickets_parent        ON tickets(parent_record_id);
+CREATE INDEX IF NOT EXISTS idx_ticket_idmap_label    ON ticket_idmap(ticket_id);
+CREATE INDEX IF NOT EXISTS idx_decisions_domain      ON decisions(domain);
+CREATE INDEX IF NOT EXISTS idx_decisions_type        ON decisions(type);
+CREATE INDEX IF NOT EXISTS idx_decision_refs_target  ON decision_refs(target_id);

.pql/changelog/ticket_idmap/2026-08.sql

@@ -0,0 +1,2 @@
+INSERT INTO ticket_idmap (record_id, ticket_id, created_at, updated_at, deleted_at, hash, canonical_version) VALUES ('06FZ44Z6HSN0RQS0QAYNTPEM5G', 'T-1', '2026-08-11 18:26:58.365', '2026-08-11 18:26:58.365', NULL, 'f1508986553f1ee59145a0d099131a68', 2) ON CONFLICT(record_id) DO UPDATE SET ticket_id=excluded.ticket_id, updated_at=excluded.updated_at, deleted_at=excluded.deleted_at, hash=excluded.hash, canonical_version=excluded.canonical_version WHERE excluded.updated_at >= ticket_idmap.updated_at;
+INSERT INTO ticket_idmap (record_id, ticket_id, created_at, updated_at, deleted_at, hash, canonical_version) VALUES ('06FZ4JX9DZB3XAB95EP23YWXMC', 'T-2', '2026-08-11 19:27:52.688', '2026-08-11 19:27:52.688', NULL, 'a5b18acb4b9a7e2da8e47b6ee603a5da', 2) ON CONFLICT(record_id) DO UPDATE SET ticket_id=excluded.ticket_id, updated_at=excluded.updated_at, deleted_at=excluded.deleted_at, hash=excluded.hash, canonical_version=excluded.canonical_version WHERE excluded.updated_at >= ticket_idmap.updated_at;

.pql/changelog/ticket_labels/0000-schema.sql

@@ -0,0 +1,139 @@
+-- Auto-generated by pql init. CREATE TABLE statements
+-- for the planning schema; per-table dir keeps the changelog
+-- self-describing per D-15. CREATE TABLE IF NOT EXISTS is
+-- idempotent so running schema files from each directory in
+-- replay order is harmless.
+--
+-- Importer parses the markers below to detect schema drift
+-- between the producing pql version and the local one — a
+-- bumped canonical_version means projection rules changed
+-- and replay must refuse rather than silently corrupt state.
+-- pql:created_by: 2.2.0
+-- pql:canonical_version: 2
+
+
+CREATE TABLE IF NOT EXISTS decisions (
+    id                TEXT PRIMARY KEY,
+    type              TEXT NOT NULL CHECK(type IN ('confirmed','question','rejected')),
+    domain            TEXT NOT NULL,
+    title             TEXT NOT NULL,
+    status            TEXT NOT NULL DEFAULT 'active'
+                          CHECK(status IN ('active','superseded','resolved','open')),
+    date              TEXT,
+    file_path         TEXT NOT NULL,
+    synced_at         TEXT NOT NULL DEFAULT (datetime('now')),
+    created_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    updated_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    deleted_at        TEXT,
+    hash              TEXT,
+    canonical_version INTEGER
+);
+
+CREATE TABLE IF NOT EXISTS decision_refs (
+    source_id         TEXT NOT NULL REFERENCES decisions(id) ON DELETE CASCADE,
+    target_id         TEXT NOT NULL REFERENCES decisions(id) ON DELETE CASCADE,
+    ref_type          TEXT NOT NULL
+                          CHECK(ref_type IN ('supersedes','references','resolves','depends_on','amends')),
+    note              TEXT,
+    created_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    updated_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    deleted_at        TEXT,
+    hash              TEXT,
+    canonical_version INTEGER,
+    PRIMARY KEY (source_id, target_id, ref_type)
+);
+
+-- Identity split (D-26): a ticket's stable, collision-proof identity is its
+-- record_id (a locally-generated ULID, planning.NewRecordID); the friendly
+-- T-NNN label lives in ticket_idmap and may be reconciled. Every structural
+-- reference (parent, deps, history, labels) targets record_id, so a label
+-- clash never corrupts the graph — only ticket_idmap needs a relabel.
+CREATE TABLE IF NOT EXISTS tickets (
+    record_id         TEXT PRIMARY KEY,
+    type              TEXT NOT NULL CHECK(type IN ('initiative','epic','story','task','bug')),
+    parent_record_id  TEXT REFERENCES tickets(record_id),
+    title             TEXT NOT NULL,
+    description       TEXT,
+    -- No CHECK enumeration: the ticket status vocabulary is per-vault
+    -- configurable (ticket_statuses in .pql/config.yaml). Validation lives
+    -- in Go (planning.StatusSet), so adding/renaming statuses needs no
+    -- schema change. The DEFAULT is a harmless fallback — CreateTicket
+    -- always inserts the configured default explicitly.
+    status            TEXT NOT NULL DEFAULT 'backlog',
+    priority          TEXT DEFAULT 'medium'
+                          CHECK(priority IN ('critical','high','medium','low')),
+    assigned_to       TEXT,
+    team              TEXT,
+    decision_ref      TEXT REFERENCES decisions(id),
+    created_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    updated_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    deleted_at        TEXT,
+    hash              TEXT,
+    canonical_version INTEGER
+);
+
+-- ticket_idmap maps a record_id to its current friendly label (T-NNN).
+-- ticket_id is intentionally NOT globally unique: two uncoordinated clones
+-- can mint the same label, which surfaces as a duplicate-label collision
+-- (detected at replay) and is fixed with "pql ticket relabel".
+CREATE TABLE IF NOT EXISTS ticket_idmap (
+    record_id         TEXT PRIMARY KEY REFERENCES tickets(record_id),
+    ticket_id         TEXT NOT NULL,
+    created_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    updated_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    deleted_at        TEXT,
+    hash              TEXT,
+    canonical_version INTEGER
+);
+
+CREATE TABLE IF NOT EXISTS ticket_deps (
+    blocker_record_id TEXT NOT NULL REFERENCES tickets(record_id),
+    blocked_record_id TEXT NOT NULL REFERENCES tickets(record_id),
+    created_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    updated_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    deleted_at        TEXT,
+    hash              TEXT,
+    canonical_version INTEGER,
+    PRIMARY KEY (blocker_record_id, blocked_record_id)
+);
+
+CREATE TABLE IF NOT EXISTS ticket_history (
+    ticket_record_id  TEXT NOT NULL REFERENCES tickets(record_id),
+    field             TEXT NOT NULL,
+    old_value         TEXT,
+    new_value         TEXT,
+    changed_by        TEXT,
+    changed_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    created_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    updated_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    deleted_at        TEXT,
+    hash              TEXT UNIQUE,
+    canonical_version INTEGER
+);
+
+CREATE TABLE IF NOT EXISTS ticket_labels (
+    ticket_record_id  TEXT NOT NULL REFERENCES tickets(record_id),
+    label             TEXT NOT NULL,
+    created_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    updated_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    deleted_at        TEXT,
+    hash              TEXT,
+    canonical_version INTEGER,
+    PRIMARY KEY (ticket_record_id, label)
+);
+
+CREATE TABLE IF NOT EXISTS meta (
+    key        TEXT PRIMARY KEY,
+    value      TEXT NOT NULL,
+    updated_at TEXT NOT NULL DEFAULT (datetime('now'))
+);
+
+CREATE INDEX IF NOT EXISTS idx_tickets_status        ON tickets(status);
+CREATE INDEX IF NOT EXISTS idx_tickets_team          ON tickets(team);
+CREATE INDEX IF NOT EXISTS idx_tickets_decision_ref  ON tickets(decision_ref);
+CREATE INDEX IF NOT EXISTS idx_tickets_assigned      ON tickets(assigned_to);
+CREATE INDEX IF NOT EXISTS idx_tickets_parent        ON tickets(parent_record_id);
+CREATE INDEX IF NOT EXISTS idx_ticket_idmap_label    ON ticket_idmap(ticket_id);
+CREATE INDEX IF NOT EXISTS idx_decisions_domain      ON decisions(domain);
+CREATE INDEX IF NOT EXISTS idx_decisions_type        ON decisions(type);
+CREATE INDEX IF NOT EXISTS idx_decision_refs_target  ON decision_refs(target_id);

.pql/changelog/tickets/0000-schema.sql

@@ -0,0 +1,139 @@
+-- Auto-generated by pql init. CREATE TABLE statements
+-- for the planning schema; per-table dir keeps the changelog
+-- self-describing per D-15. CREATE TABLE IF NOT EXISTS is
+-- idempotent so running schema files from each directory in
+-- replay order is harmless.
+--
+-- Importer parses the markers below to detect schema drift
+-- between the producing pql version and the local one — a
+-- bumped canonical_version means projection rules changed
+-- and replay must refuse rather than silently corrupt state.
+-- pql:created_by: 2.2.0
+-- pql:canonical_version: 2
+
+
+CREATE TABLE IF NOT EXISTS decisions (
+    id                TEXT PRIMARY KEY,
+    type              TEXT NOT NULL CHECK(type IN ('confirmed','question','rejected')),
+    domain            TEXT NOT NULL,
+    title             TEXT NOT NULL,
+    status            TEXT NOT NULL DEFAULT 'active'
+                          CHECK(status IN ('active','superseded','resolved','open')),
+    date              TEXT,
+    file_path         TEXT NOT NULL,
+    synced_at         TEXT NOT NULL DEFAULT (datetime('now')),
+    created_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    updated_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    deleted_at        TEXT,
+    hash              TEXT,
+    canonical_version INTEGER
+);
+
+CREATE TABLE IF NOT EXISTS decision_refs (
+    source_id         TEXT NOT NULL REFERENCES decisions(id) ON DELETE CASCADE,
+    target_id         TEXT NOT NULL REFERENCES decisions(id) ON DELETE CASCADE,
+    ref_type          TEXT NOT NULL
+                          CHECK(ref_type IN ('supersedes','references','resolves','depends_on','amends')),
+    note              TEXT,
+    created_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    updated_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    deleted_at        TEXT,
+    hash              TEXT,
+    canonical_version INTEGER,
+    PRIMARY KEY (source_id, target_id, ref_type)
+);
+
+-- Identity split (D-26): a ticket's stable, collision-proof identity is its
+-- record_id (a locally-generated ULID, planning.NewRecordID); the friendly
+-- T-NNN label lives in ticket_idmap and may be reconciled. Every structural
+-- reference (parent, deps, history, labels) targets record_id, so a label
+-- clash never corrupts the graph — only ticket_idmap needs a relabel.
+CREATE TABLE IF NOT EXISTS tickets (
+    record_id         TEXT PRIMARY KEY,
+    type              TEXT NOT NULL CHECK(type IN ('initiative','epic','story','task','bug')),
+    parent_record_id  TEXT REFERENCES tickets(record_id),
+    title             TEXT NOT NULL,
+    description       TEXT,
+    -- No CHECK enumeration: the ticket status vocabulary is per-vault
+    -- configurable (ticket_statuses in .pql/config.yaml). Validation lives
+    -- in Go (planning.StatusSet), so adding/renaming statuses needs no
+    -- schema change. The DEFAULT is a harmless fallback — CreateTicket
+    -- always inserts the configured default explicitly.
+    status            TEXT NOT NULL DEFAULT 'backlog',
+    priority          TEXT DEFAULT 'medium'
+                          CHECK(priority IN ('critical','high','medium','low')),
+    assigned_to       TEXT,
+    team              TEXT,
+    decision_ref      TEXT REFERENCES decisions(id),
+    created_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    updated_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    deleted_at        TEXT,
+    hash              TEXT,
+    canonical_version INTEGER
+);
+
+-- ticket_idmap maps a record_id to its current friendly label (T-NNN).
+-- ticket_id is intentionally NOT globally unique: two uncoordinated clones
+-- can mint the same label, which surfaces as a duplicate-label collision
+-- (detected at replay) and is fixed with "pql ticket relabel".
+CREATE TABLE IF NOT EXISTS ticket_idmap (
+    record_id         TEXT PRIMARY KEY REFERENCES tickets(record_id),
+    ticket_id         TEXT NOT NULL,
+    created_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    updated_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    deleted_at        TEXT,
+    hash              TEXT,
+    canonical_version INTEGER
+);
+
+CREATE TABLE IF NOT EXISTS ticket_deps (
+    blocker_record_id TEXT NOT NULL REFERENCES tickets(record_id),
+    blocked_record_id TEXT NOT NULL REFERENCES tickets(record_id),
+    created_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    updated_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    deleted_at        TEXT,
+    hash              TEXT,
+    canonical_version INTEGER,
+    PRIMARY KEY (blocker_record_id, blocked_record_id)
+);
+
+CREATE TABLE IF NOT EXISTS ticket_history (
+    ticket_record_id  TEXT NOT NULL REFERENCES tickets(record_id),
+    field             TEXT NOT NULL,
+    old_value         TEXT,
+    new_value         TEXT,
+    changed_by        TEXT,
+    changed_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    created_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    updated_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    deleted_at        TEXT,
+    hash              TEXT UNIQUE,
+    canonical_version INTEGER
+);
+
+CREATE TABLE IF NOT EXISTS ticket_labels (
+    ticket_record_id  TEXT NOT NULL REFERENCES tickets(record_id),
+    label             TEXT NOT NULL,
+    created_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    updated_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    deleted_at        TEXT,
+    hash              TEXT,
+    canonical_version INTEGER,
+    PRIMARY KEY (ticket_record_id, label)
+);
+
+CREATE TABLE IF NOT EXISTS meta (
+    key        TEXT PRIMARY KEY,
+    value      TEXT NOT NULL,
+    updated_at TEXT NOT NULL DEFAULT (datetime('now'))
+);
+
+CREATE INDEX IF NOT EXISTS idx_tickets_status        ON tickets(status);
+CREATE INDEX IF NOT EXISTS idx_tickets_team          ON tickets(team);
+CREATE INDEX IF NOT EXISTS idx_tickets_decision_ref  ON tickets(decision_ref);
+CREATE INDEX IF NOT EXISTS idx_tickets_assigned      ON tickets(assigned_to);
+CREATE INDEX IF NOT EXISTS idx_tickets_parent        ON tickets(parent_record_id);
+CREATE INDEX IF NOT EXISTS idx_ticket_idmap_label    ON ticket_idmap(ticket_id);
+CREATE INDEX IF NOT EXISTS idx_decisions_domain      ON decisions(domain);
+CREATE INDEX IF NOT EXISTS idx_decisions_type        ON decisions(type);
+CREATE INDEX IF NOT EXISTS idx_decision_refs_target  ON decision_refs(target_id);

.pql/changelog/tickets/2026-08.sql

@@ -0,0 +1,36 @@
+INSERT INTO tickets (record_id, type, parent_record_id, title, description, status, priority, assigned_to, team, decision_ref, created_at, updated_at, deleted_at, hash, canonical_version) VALUES ('06FZ44Z6HSN0RQS0QAYNTPEM5G', 'task', NULL, 'Type the codebase: 95 mypy errors across 31 files', NULL, 'backlog', 'high', NULL, NULL, NULL, '2026-08-11 18:26:58.318', '2026-08-11 18:26:58.318', NULL, 'd881f36d77e58aaa2f94dc08c5b5be3e', 2) ON CONFLICT(record_id) DO UPDATE SET type=excluded.type, parent_record_id=excluded.parent_record_id, title=excluded.title, description=excluded.description, status=excluded.status, priority=excluded.priority, assigned_to=excluded.assigned_to, team=excluded.team, decision_ref=excluded.decision_ref, updated_at=excluded.updated_at, deleted_at=excluded.deleted_at, hash=excluded.hash, canonical_version=excluded.canonical_version WHERE excluded.updated_at >= tickets.updated_at;
+INSERT INTO tickets (record_id, type, parent_record_id, title, description, status, priority, assigned_to, team, decision_ref, created_at, updated_at, deleted_at, hash, canonical_version) VALUES ('06FZ44Z6HSN0RQS0QAYNTPEM5G', 'task', NULL, 'Type the codebase: 95 mypy errors across 31 files', '`make typecheck` reports 95 errors in 31 files (was 103). This is a dedicated programming pass, not lint tidying, and it is what currently blocks `make pre-push`.
+
+MEASURED 2026-08-11 so the next session does not re-derive it:
+
+  29 no-untyped-def   functions with no annotations — the bulk, and genuine per-function work
+  18 no-any-return    mostly downstream of the above
+  12 assignment
+  11 arg-type
+   5 unused-ignore    `# type: ignore` comments mypy says are no longer needed
+   5 union-attr
+   4 var-annotated
+   3 override
+   remainder: misc, dict-item, attr-defined, return-value, call-overload
+
+By file: agents/tatlock.py 15, core/memory_service.py 11, responses/streaming.py 8, responses/service.py 7, core/context.py 6.
+
+THE TWO SHARED ROOTS ARE ALREADY FIXED (dac259a), so what is left has no lever in it. For reference, they were: five conversation lists declared bare, where mypy infers the element type from the first append (a ModelRequest) and then rejects every ModelResponse; and an agent built as Agent(model, system_prompt=...) with no deps_type, inferred Agent[None, str], while every tool it registers takes RunContext[ToolCallTracker].
+
+WORTH KNOWING BEFORE STARTING. Annotating partially made mypy count go UP before it went down — declaring `_agent: Agent | None` took agents/tatlock.py from 22 to 24, because resolving the bare Agent to Agent[None, str] surfaced four argument-type errors the Any had been hiding. Expect that shape: a rising count during this work usually means concealment ending, not damage.
+
+The mypy config is strict — disallow_untyped_defs, disallow_incomplete_defs, warn_return_any, check_untyped_defs, strict_equality — so there is no partial-credit setting to lean on, and weakening it would be the wrong trade for a codebase this central.
+
+TWO PRE-EXISTING TEST FACTS, both confirmed at HEAD and neither caused by the lint work:
+  - test_tatlock_tool_call_logging_calculator is flaky: failed 2 of 5 full runs, on HEAD and on the lint branch, and fails in isolation at HEAD while passing in isolation after the lint pass. Order- or timing-dependent.
+  - `pytest tests/` cannot collect at all: tests/e2e/test_orchestration_e2e.py uses an `e2e` marker that is not registered and the config is strict about markers. `make test` passes only because it ignores tests/e2e, tests/integration and tests/contracts.', 'backlog', 'high', NULL, NULL, NULL, '2026-08-11 18:26:58.318', '2026-08-11 18:26:58.523', NULL, 'fc047dd080d976c4ca0c551cac8fec93', 2) ON CONFLICT(record_id) DO UPDATE SET type=excluded.type, parent_record_id=excluded.parent_record_id, title=excluded.title, description=excluded.description, status=excluded.status, priority=excluded.priority, assigned_to=excluded.assigned_to, team=excluded.team, decision_ref=excluded.decision_ref, updated_at=excluded.updated_at, deleted_at=excluded.deleted_at, hash=excluded.hash, canonical_version=excluded.canonical_version WHERE excluded.updated_at >= tickets.updated_at;
+INSERT INTO tickets (record_id, type, parent_record_id, title, description, status, priority, assigned_to, team, decision_ref, created_at, updated_at, deleted_at, hash, canonical_version) VALUES ('06FZ4JX9DZB3XAB95EP23YWXMC', 'bug', NULL, 'A revoked ANTHROPIC_API_KEY is still sitting in .env', NULL, 'backlog', 'medium', NULL, NULL, NULL, '2026-08-11 19:27:52.687', '2026-08-11 19:27:52.687', NULL, '3b00a4b729819e20a6425f971e6cb9da', 2) ON CONFLICT(record_id) DO UPDATE SET type=excluded.type, parent_record_id=excluded.parent_record_id, title=excluded.title, description=excluded.description, status=excluded.status, priority=excluded.priority, assigned_to=excluded.assigned_to, team=excluded.team, decision_ref=excluded.decision_ref, updated_at=excluded.updated_at, deleted_at=excluded.deleted_at, hash=excluded.hash, canonical_version=excluded.canonical_version WHERE excluded.updated_at >= tickets.updated_at;
+INSERT INTO tickets (record_id, type, parent_record_id, title, description, status, priority, assigned_to, team, decision_ref, created_at, updated_at, deleted_at, hash, canonical_version) VALUES ('06FZ4JX9DZB3XAB95EP23YWXMC', 'bug', NULL, 'A revoked ANTHROPIC_API_KEY is still sitting in .env', '`.env` still carries an `ANTHROPIC_API_KEY`. That key was revoked and no longer exists on Anthropic''s side, so this is dead weight rather than an exposure — but it is dead weight that reads exactly like a live credential to anyone who finds it.
+
+The cost is confusion, not risk. Someone debugging a Claude fallback will find a key present, assume it is configured, and look elsewhere for the failure. The absence of a key is a clear signal; a revoked key is a misleading one.
+
+`.env` is gitignored here and has never been committed, so nothing needs rewriting — the value simply needs removing from the local file, and the line dropping or blanking in `.env.example` if it appears there too.
+
+RELATED, and the reason this is filed separately: the workspace vault holds T-13, "Clear the revoked ANTHROPIC_API_KEY from the live Portainer stack". That ticket is scoped to the Portainer stack only. Whoever closes it will reasonably believe the key is gone once the stack is clean, and this copy will survive. The two want doing together even though they commit separately.
+
+Context on the revocation, since it explains why nobody removed this at the time: the key was revoked on 2026-08-09 after being printed into a transcript by a redaction filter that matched on `KEY` appearing after the `=`. In `ANTHROPIC_API_KEY=...` it appears before, so the filter never fired. The response was rotation, and the leftover copies were not swept.', 'backlog', 'medium', NULL, NULL, NULL, '2026-08-11 19:27:52.687', '2026-08-11 19:27:52.823', NULL, 'a9c2ca965b3d40a924720c7761c5338d', 2) ON CONFLICT(record_id) DO UPDATE SET type=excluded.type, parent_record_id=excluded.parent_record_id, title=excluded.title, description=excluded.description, status=excluded.status, priority=excluded.priority, assigned_to=excluded.assigned_to, team=excluded.team, decision_ref=excluded.decision_ref, updated_at=excluded.updated_at, deleted_at=excluded.deleted_at, hash=excluded.hash, canonical_version=excluded.canonical_version WHERE excluded.updated_at >= tickets.updated_at;

AGENTS.md

@@ -1,56 +0,0 @@
-# LLM Agent Instructions
-
-This document contains instructions and documentation references for AI assistants working with this codebase.
-
-> **📖 Important**: Before working on this project, read [PHILOSOPHY.md](PHILOSOPHY.md) to understand the system vision, architectural patterns, and design goals. All development should work towards realizing those patterns.
-# AGENTS.md
-
-> **Start every session by reading this file.**
-> This file outlines the operational protocols, coding standards, and architectural decisions for this FastAPI project.
-
-## 1. Agent Operational Protocols
-
-### 🧠 Work Patterns (Plan-Act-Reflect)
-* **Plan:** Before writing code, briefly outline your plan. Identify which files you will touch and what the side effects might be.
-* **Act:** Execute the changes in small, atomic steps.
-* **Reflect:** After coding, verify your work. Did you break existing tests? Did you add new tests?
-
-### 🌐 Internal Service Access
-* **git.schweitz.net**: Access via `http://localhost:3002` (direct Gitea) to bypass Authentik SSO
-  * Example: `curl http://localhost:3002/jpmschweitzer/library-desk/raw/branch/main/README.md`
-  * Public repos are readable without authentication
-  * Related repos: `library-desk`, `scheduler`
-
-### 🛡️ Git Discipline
-* **NEVER commit to `main` or `master` directly.** Always create a feature branch: `feature/your-feature-name` or `fix/issue-description`.
-* **Commit Messages:** Use the [Conventional Commits](https://www.conventionalcommits.org/) format.
-    * `feat: add user login endpoint`
-    * `fix: resolve database connection timeout`
-    * `refactor: split monolith dependency file`
-* **Atomic Commits:** Keep commits small. One logical change = one commit.
-
-### 📝 Changelog Maintenance
-* **Update `CHANGELOG.md`** with every user-facing change.
-* Format: `## [Unreleased] - YYYY-MM-DD` followed by `### Added`, `### Changed`, or `### Fixed`.
-
----
-
-## 2. FastAPI Architecture & Best Practices
-*Reference: [FastAPI Best Practices](https://github.com/zhanymkanov/fastapi-best-practices)*
-
-### 📂 Project Structure (Directory-based, NOT File-type based)
-Do **not** group files by type (e.g., one huge `routers` folder). Group by **domain/module** inside a `src/` directory.
-
-**Correct Structure:**
-```text
-src/
-├── auth/
-│   ├── router.py      # Endpoints
-│   ├── schemas.py     # Pydantic models
-│   ├── service.py     # Business logic (CRUD, etc.)
-│   ├── dependencies.py# Module-specific dependencies
-│   └── config.py      # Module-specific settings
-├── posts/
-│   ├── router.py
-│   └── ...
-└── main.py            # App entry point

CHANGELOG.md

@@ -7,6 +7,560 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+### Changed
+
+- `make setup` now ends with a `pytest --collect-only` pass so a broken environment
+  (missing or mismatched dependency) fails the target itself instead of exiting 0
+  and surfacing later as a confusing test failure (T-47)
+
+## [2.4.3] - 2026-08-08
+
+### Fixed
+
+- Steward routing no longer triggers on words inside its own explanation. Capability
+  extraction reads the declared `DELEGATE:` line instead of substring-matching
+  capability domains across the whole response, where ordinary English routed
+  requests — "description" contains the housekeeper domain "script", "acknowledge"
+  contains "knowledge" and "know". A spurious capability meant a real agent call,
+  including web searches, on queries that needed none.
+
+## [2.4.2] - 2026-07-19
+
+### Fixed
+
+- Container crash-loop on fresh builds: cap `opentelemetry-api` below 1.44,
+  which removed the private `_events` module that pydantic-ai 1.27 imports
+
+## [2.4.1] - 2026-07-19
+
+### Changed
+
+- **Container-name network defaults** - `SEARXNG_HOST`, `LIBRARY_DESK_HOST`, and `CORE_API_HOST` now default to docker container names on the docker-dataplane network (`http://searxng:8080`, `http://library-desk:8089`, `http://core-api:8083`) instead of host `localhost` ports, ahead of the loopback port rebinding; this also fixes `CORE_API_HOST` pointing at port 8090 (the Scheduler's host port) rather than Core-API's 8083. `scripts/test_housekeeper.sh` now reaches Core-API via `localhost:8083` instead of the LAN IP. Local development against host-published ports still works via `.env` overrides
+
+## [2.4.0] - 2026-07-14
+
+### Removed
+
+- **Dead delegation stack** - deleted the duplicate, never-wired coordination layer so exactly ONE delegation implementation remains (`src/agents/delegation.py`): `src/agents/coordination.py` (`CoordinationEngine`, its own `delegate_to_librarian`, `AGENT_EXECUTORS`/`AGENT_STREAM_EXECUTORS`), the broken-by-design `run_librarian_stream` path it used (Ollama streaming + tool call bug), the `stream_delegate_to_*` wrappers with their never-parsed `__DELEGATION_RESULT__` marker, and `HouseholdRegistry.get_streaming_delegation_tools()` (no callers)
+- **Orphaned agent protocol models** - `src/agents/protocol.py` now contains only the live `AgentError`; the coordination wire protocol it carried (`AgentRequest`, `AgentResponse`, `DelegationIntent`, `CoordinationResult`, `DelegationReason`, `TaskComplexity`, `ToolCallRecord`, `AgentTimeoutError`, `AgentUnavailableError`, `DelegationError`) had no importer left outside its own tests after the coordination stack removal
+
+### Added
+
+- **Test-suite tenant guard** - `tests/conftest.py` hard-fails the whole pytest session (exit code 1, zero tests run) if the effective tenant resolves to the production tenant `jpmschweitzer`, mirroring the guard library-desk applies on its side. Suite-level assertions pin that the session runs under `llm_tester` namespaces (Qdrant `memories_llm_tester`, Redis `session:llm_tester:*`), and the e2e isolation constants now derive from the shared `TEST_TENANT`/`PRODUCTION_TENANT` config constants instead of string literals
+- **Explicit tenant on every library-desk request** - the librarian client now resolves and sends the `user` parameter explicitly on every request (library-desk is removing its server-side default; a missing user would 422). The content extraction endpoints now carry the tenant too, `search_web` no longer falls back to a phantom `tatlock-librarian` user, and a client-level assertion rejects an empty/whitespace tenant before any bytes hit the wire. A parametrized sweep pins the wire contract for all 15 tenant-scoped client methods
+- **Tenant isolation guard** - non-production environments (development/testing) now FORCE the effective tenant to the reserved test tenant `llm_tester` (only `llm_tester` itself or a `test_`-prefixed override is accepted), regardless of `DEFAULT_USER` misconfiguration, at both config resolution and request-context resolution (`get_user()`). Startup refuses (clear error) when a non-production environment is explicitly configured with the production tenant `jpmschweitzer`, and one loud startup log line states the effective/forced tenant
+
+- **Conversation context for experts + real-time think messages** - direct delegation (streaming and non-streaming) now passes a trimmed conversation history (last 6 turns) as expert context, so follow-up questions keep their referent; `_stream_direct_delegation` is now an async generator, so butler think messages ("Allow me to consult the archives, sir.") stream BEFORE the research runs instead of after it completes
+- **Bounded retries and connection reuse for library-desk** - GETs and the read-only `POST /query/*` and `POST /rag/search` endpoints retry once (2 attempts, short backoff) on transport errors and retryable 5xx; wiki writes are never retried. The client now honors `LIBRARY_DESK_TIMEOUT` instead of hardcoded 60s/30s, a librarian run holds one shared HTTP connection instead of constructing a client per tool call, and read tools raise `ModelRetry` on transient HTTP errors so the agent's retry budget engages
+- **One librarian timeout budget** - new `LIBRARIAN_TIMEOUT` (default 180s) enforced with `asyncio.wait_for` inside `delegate_to_librarian`, capping the previously uncapped live paths (steward direct delegation and streaming). The Ollama provider's AsyncOpenAI client now carries an explicit `OLLAMA_TIMEOUT` instead of the SDK's ~600s default, and the contradictory unused 60s default in `AgentRequest.timeout_seconds` was removed (None defers to the configured budget)
+- **Search degradation signaling** - The librarian client parses `source_counts` (plus the additive `source_status`/`degraded` fields when a newer library-desk sends them; absence is tolerated), and `hybrid_search` appends a one-line coverage note when a search is degraded or an enabled source leg contributed nothing, so outages are visible to the model and the user. When `source_status` is present it is used exclusively; without it, count-absence is only inferred for the optional legs the request explicitly enabled (web/documents/volatile) - never the always-on vector/graph legs, whose absence from the top-N counts is normal ranking behavior, so healthy searches no longer emit warnings
+
+### Fixed
+
+- **Clearing all wiki-page tags is possible again** - the Ollama-safe empty-list sentinel in `update_wiki_page` means "leave unchanged", which made it impossible to remove all tags; passing exactly `["__CLEAR__"]` now sends an empty tag list to library-desk (documented in the tool docstring for the local model)
+- **Text-delegation fallback pairs results strictly** - the parallel branch now verifies `asyncio.gather` returned one result per parsed delegation (`zip(..., strict=True)`); a count mismatch fails loudly with a curated apology instead of silently attributing outputs to the wrong agent
+- **Ollama-safe librarian tool schemas** - `update_wiki_page` and `smart_create_wiki_page` no longer use `X | None` parameters (Ollama's OpenAI-compatible API mishandles `anyOf[X, null]`); empty-string/empty-list sentinels are translated to `None` inside the tools, matching the biographer pattern. A snapshot test pins every librarian tool schema to contain no nullable `anyOf`
+- **Honest expert failures** - `run_librarian` now raises a structured `AgentError` instead of returning error text as if it were research output, so delegation correctly reports `success=False` and the streaming error branch is reachable. Failures surface to the user as curated butler-toned sentences; exception detail (including internal URLs) stays in the logs only. Librarian tool errors no longer leak `str(e)` into synthesis
+
+- **HybridRAG response mapping** - The librarian client now parses the field names library-desk actually returns (`source_type`/`sources`, `rrf_score`, `context`, per-item `related_dossiers`, synonyms nested in the `keywords` dict); previously every result rendered as "unknown (score: 0.00)". Source icons now key off the per-item `sources` list. Requests no longer send zero limits (the service rejects them with 422); legs are disabled via `enable_*` flags. Pinned by a contract test against a recorded live response (`tests/agents/librarian/fixtures/`)
+
+## [2.3.0] - 2026-07-13
+
+### Changed
+
+- **Local-first backend (claudification rollback)** - Ollama/gemma4 is now the primary backend; Claude remains as fallback. `PREFER_CLOUD_BACKEND` defaults to `false`, Claude is used automatically when the Ollama startup health check fails, and the Steward retries mid-request failures on the other backend in both directions
+- **Default Claude model `claude-sonnet-5`** - `claude-sonnet-4-20250514` was retired by Anthropic on 2026-06-15 and would 404, leaving the fallback dead
+- **Dedicated orchestration prompt** - `orchestrate_tool_calls()` now uses a terse tool-execution prompt (`TATLOCK_ORCHESTRATION_PROMPT`); the butler persona prompt suppressed gemma4 tool calling (the model reasoned about the calculator, then answered from memory with wrong arithmetic). Synthesis keeps the persona prompt, so user-visible voice is unchanged
+
+### Fixed
+
+- **Startup crash with broken anthropic package** - Anthropic SDK imports in the model selector are now lazy, so an incompatible `anthropic` install degrades to Ollama-only operation instead of crashing the app at import time (root cause of the production outage since April)
+- **Claude Sonnet 5 rejects sampling parameters** - removed `temperature` from the Steward's direct Claude call and made the Housekeeper's temperature setting backend-conditional via `get_sampling_settings()`
+- **Pin `anthropic>=0.77,<1.0`** - the April image resolved an anthropic version incompatible with pydantic-ai 1.27
+- **Steward timeout configurable** - new `STEWARD_TIMEOUT` (default 60s) replaces the hardcoded 30s, which gemma4 chronically exceeded (~35s warm analysis), causing every request to fail or fall back
+
+### Added
+
+- **Ollama startup health check** - verifies the server is reachable and `OLLAMA_DEFAULT_MODEL` is pulled; feeds backend resolution and `get_model_info()`
+- **Contract tests** (`tests/contracts/`, `make test-contracts`) - wire-level tests that send the raw requests the code sends to Ollama (native + OpenAI-compat tool calling), Anthropic (including the pinned temperature-rejection contract), Qdrant, SearXNG, library-desk, and Redis; unreachable services skip, wrong response shapes fail
+- **Backend resolution unit tests** (`tests/anthropic/`)
+
+## [2.2.0] - 2026-04-04
+
+### Changed
+
+- **Switch default Ollama model to gemma4:e2b** - Replaces mistral-nemo as the local LLM backend; gemma4:e2b has native function calling support, faster tool calling (2-4s vs 15-20s), better parameter accuracy on word problems, and uses less VRAM (8GB vs 9.2GB)
+
+### Added
+
+- **Tool calling benchmark script** (`scripts/benchmark_tool_calling.py`) - Compares tool calling accuracy and latency across Ollama models via the Tatlock API
+
+## [2.1.0] - 2026-02-05
+
+### Fixed
+
+- **Streaming SSE compatibility with Open WebUI** - Switch from `exclude_none=True` to `exclude_unset=True` for SSE chunk serialization; `exclude_none` was too aggressive — it stripped `finish_reason: null` from intermediate chunks (which OpenAI includes), while `exclude_unset` correctly omits only fields never passed to the constructor (like `reasoning_content` on content-only chunks) while preserving explicitly-set `finish_reason: null`
+
+### Changed
+
+- **Project structure consolidation** - Moved documentation to `docs/`, consolidated all config into `pyproject.toml`, replaced `wakeup.sh`/`pytest.ini`/`requirements*.txt` with `Makefile` + `pyproject.toml`
+- **CI test gate** - Unit tests now gate release and build jobs in Gitea Actions workflow
+- **Build output organization** - Tool caches in `.cache/`, generated output (coverage, logs) in `build/`
+
+## [2.0.5] - 2026-02-05
+
+### Fixed
+
+- **Streaming JSON compatibility** - Exclude null fields from streaming chunks using `exclude_none=True`; OpenAI's API omits null fields entirely, and including them (e.g., `content: null`, `reasoning_content: null`) caused parsing issues in Open WebUI
+
+## [2.0.4] - 2026-02-05
+
+### Fixed
+
+- **Open WebUI streaming compatibility** - Replaced `sse_starlette` `EventSourceResponse` with plain `StreamingResponse` for chat completions; `sse_starlette` added `\r\n` line endings and extra SSE fields that Open WebUI couldn't parse
+
+## [2.0.3] - 2026-02-05
+
+### Fixed
+
+- **Steward analysis leaking into responses** - Removed internal routing analysis (`DELEGATE: tatlock_core...`) from user-visible reasoning in both streaming and non-streaming paths
+
+## [2.0.2] - 2026-02-05
+
+### Fixed
+
+- **tool_choice format incompatibility** - Removed `extra_body` tool_choice hack for Claude backend; PydanticAI handles tool_choice natively for Anthropic, preventing infinite tool call loops
+- **CI trigger** - Changed workflow trigger from `release:published` to `push:tags:v[0-9]*`
+
+## [2.0.1] - 2026-02-05
+
+### Fixed
+
+- **Expert agent registration failure** - `AnthropicModel` does not accept `api_key` directly; now passes it via `AnthropicProvider`
+
+## [2.0.0] - 2026-02-05
+
+### Added
+
+- **Claude backend support (Claudification Phase 1)** - All agents now prefer Claude over Ollama
+  - New `src/anthropic/` module with model selector and health check
+  - `get_model()` factory returns Claude if available, Ollama as fallback
+  - Startup health check caches Claude API availability
+  - Configuration: `ANTHROPIC_API_KEY`, `ANTHROPIC_MODEL`, `PREFER_CLOUD_BACKEND`
+  - 200k token context when using Claude backend
+
+- **Steward dual-backend support** - Direct API calls to Claude or Ollama
+  - `_call_claude()`: Anthropic Messages API path
+  - `_call_ollama()`: Existing Ollama generate API path (preserved)
+  - Automatic fallback: if Claude call fails mid-request, retries with Ollama
+
+- **Claudification project tracking** - `PROJECT_CLAUDIFICATION.md` with Phase 1/2 roadmap
+
+### Changed
+
+- **All PydanticAI agents refactored to use `get_model()`**:
+  - Tatlock (6 instantiation locations)
+  - Librarian
+  - Biographer
+  - Housekeeper
+- **`initialize_application()` is now async** - Supports async Claude health check at startup
+- **Dependencies**: `pydantic-ai-slim[openai,anthropic]` replaces `pydantic-ai-slim[openai]`
+- **Startup logging** now includes backend selection info (claude/ollama)
+- **Agent creation logging** now includes backend and model info
+
+### Removed
+
+- Stale `tests/core/test_benchmarks.py` (benchmark system was removed in v1.10.0)
+
+## [1.11.0] - 2025-12-30
+
+### Added
+
+- **Paperless document integration** - HybridRAG now includes indexed PDFs and scanned documents from Paperless-ngx
+  - New `include_documents` parameter in `hybrid_search` tool
+  - 📑 icon for document sources in search results
+  - Librarian prompt updated with document awareness
+
+- **Volatile cache integration** - HybridRAG now includes pre-fetched real-time data
+  - New `include_volatile` parameter in `hybrid_search` tool
+  - ⚡ icon for volatile sources in search results
+  - Supports weather, forecast, news, stock, crypto, sun, air_quality namespaces
+  - Librarian prompt updated with volatile cache awareness (user-configured items only)
+
+- **Biographer routing in Steward** - Personal memory queries now correctly route to The Biographer
+  - Added explicit routing rules for "where do I live", "what car do I drive", etc.
+  - Added biographer delegation examples to Steward prompt
+  - Location keywords ("live", "where", "home") now trigger profile pre-fetch
+
+### Changed
+
+- **LibraryDeskClient.hybrid_search** - Now passes full config including `document_limit`, `volatile_limit`, and enable flags
+- **Steward guidelines** - Clarified that research queries about TOPICS go to Librarian, queries about USER go to Biographer
+
+## [1.10.1] - 2025-12-23
+
+### Fixed
+
+- **Tatlock's excessive apologizing** - Strengthened personality prompt to prevent unnecessary apologies after successful Librarian delegations. Added explicit "do NOT apologize" instructions to both system prompt and synthesis prompt.
+
+## [1.10.0] - 2025-12-22
+
+### Added
+
+#### Lightweight Request Tracing
+- **JSON-based tracing system** for local development debugging
+  - Captures full request flow through multi-agent architecture
+  - `Trace` and `Span` dataclasses with automatic timing and nesting
+  - ContextVar-based propagation for async-safe tracing
+  - `trace_span` async context manager for clean instrumentation
+  - Traces written to `logs/traces/{trace_id}.json`
+  - Enabled via `DEBUG=true` environment variable
+- **Trace Viewer UI** (`logs/traces/viewer.html`)
+  - Standalone HTML viewer with timeline visualization
+  - Filter by status, search by request text
+  - Expandable span details with prompts and responses
+- **Tracing REST API** (`/traces`)
+  - `GET /traces` - Serve trace viewer UI
+  - `GET /traces/list` - List available traces with filtering
+  - `GET /traces/{trace_id}` - Retrieve specific trace JSON
+  - Only available when `DEBUG=true`
+- **Full pipeline instrumentation**
+  - Router-level trace start/end with context management
+  - Steward analysis spans in preprocessing
+  - Tatlock orchestrate/synthesize spans
+  - Expert delegation spans (librarian/biographer/housekeeper)
+  - Tool-level spans extracted from PydanticAI messages
+
+### Changed
+
+- **Replaced Redis benchmarks with file-based tracing** - Simpler, more useful for debugging
+- **Context management moved to service layer** - Router simplified, context set in response service
+- **Server binds to all interfaces** - `wakeup.sh` now uses `0.0.0.0` for network access
+
+### Removed
+
+- **Redis benchmark system** (`src/core/benchmarks.py`)
+  - `ENABLE_BENCHMARKS` config setting
+  - `REDIS_BENCHMARK_DB` config setting
+  - `redis_url` property (kept `redis_memory_url`)
+  - Benchmark recording in Steward service and tool tracking
+
+### Fixed
+
+- **Librarian fabrication prevention** - Added explicit instructions to never invent data when tools fail or sources are unavailable
+
+## [1.9.0] - 2025-12-18
+
+### Changed
+
+- **Housekeeper prompt optimization** - Rewrote system prompt for Mistral-Nemo function calling with negative constraints, step-by-step process, and explicit entity ID format guidance
+- **Housekeeper temperature setting** - Set temperature to 0.1 for deterministic tool calling behavior
+- **Device list room group priority** - Room groups now appear first in `list_devices` output with `[ROOM GROUP]` marker to address positional bias
+- **Tool docstring improvements** - Updated turn_on/turn_off/toggle with explicit `entity_id=` parameter examples
+
+### Added
+
+- **Housekeeper optimization findings** - Added `docs/housekeeper-optimization-findings.md` documenting the experiment journey from 0% to 100% success rate
+- **Housekeeper test script** - Added `scripts/test_housekeeper.sh` for room group detection regression testing
+
+## [1.8.6] - 2025-12-17
+
+### Fixed
+
+- **Housekeeper API paths** - Updated all client endpoints to use `/housekeeping/` prefix to match core-api routes
+- **Housekeeper entity hallucination** - Improved system prompt with critical rule requiring `list_devices()` before any control action to prevent guessing entity IDs
+
+### Added
+
+- **Housekeeping API spec** - Added `docs/housekeeping-api-spec.md` documenting the core-api home automation interface
+
+## [1.8.5] - 2025-12-16
+
+### Fixed
+
+- **Redis benchmark boolean storage** - Convert booleans to strings for Redis `hset` (Redis doesn't accept bool type directly)
+- **Tool tracking capability matching** - `delegate_to_librarian` now correctly recognized as using "librarian" capability when checking Steward recommendations
+- **E2E test fixture scope** - Fixed pytest-asyncio ScopeMismatch error by using `loop_scope="module"` for module-scoped async fixtures
+
+## [1.8.4] - 2025-12-16
+
+### Fixed
+
+- **Remove `<think>` wrappers from think messages** - Messages in `reasoning_content` should be plain text
+- Removed `<think>` wrappers from delegation.py household think messages
+- Removed `<think>` wrappers from orchestration.py status messages
+- Think messages now appear cleanly in Open WebUI's reasoning block
+
+## [1.8.3] - 2025-12-16
+
+### Fixed
+
+- **Open WebUI streaming rendering** - Use `reasoning_content` field for thinking (DeepSeek R1 format) instead of `<think>` tags in `content`
+- Open WebUI now renders thinking as proper collapsible blocks instead of broken HTML
+
+## [1.8.2] - 2025-12-16
+
+### Fixed
+
+- **HybridRAG keywords schema mismatch** - library-desk now returns `keywords` as dict with `core_keywords`, client now handles both formats
+
+## [1.8.1] - 2025-12-16
+
+### Fixed
+
+#### Ollama Message Sanitization
+- **Fixed `invalid message content type: <nil>` error** from Ollama
+- Created custom `TatlockOllamaProvider` that sanitizes messages before sending to Ollama
+- Ollama rejects assistant messages with `content: null` (tool-only messages from PydanticAI)
+- Provider converts `null` content to empty string `""` for compatibility
+- Updated all agents (Librarian, Biographer, Housekeeper, Tatlock) to use sanitized provider
+- Added `src/ollama/provider.py` with reusable provider pattern
+
+#### Streaming Think Message Accumulation
+- **Fixed repeating think messages in frontend** (e.g., 10x "The Librarian has compiled...")
+- Frontend was accumulating `ReasoningSummaryDelta` events expecting concatenation
+- Added `ReasoningSummaryDone()` signal after each think message to indicate completion
+- Each think slug is now treated as a complete message, not a continuation
+
+## [1.8.0] - 2025-12-15
+
+### Fixed
+
+#### Steward Routing for Web Search
+- Updated Steward guidelines to route web searches, weather, news → Librarian with `search_web`
+- Added URL/article reading → Librarian with `read_url` to routing guidelines
+- Added examples showing `search_web` and `read_url` tool usage
+
+#### Librarian Agent Tool Registration
+- Registered `search_web`, `read_url`, `read_urls_batch` tools with the Librarian PydanticAI agent
+- Updated Librarian system prompt with Web Search & Content Extraction section
+- Fixed tool count in agent logger (11 → 14 tools)
+
+#### Query Enrichment Integration
+- Fixed enriched query (with location/timezone context) not being passed to delegations
+- Response service now uses `enriched_query` from Steward recommendation for all delegations
+- Weather queries now automatically include user's stored location
+
+#### Action Type Detection
+- Added "read", "fetch", "url", "http" keywords to RESEARCH action type for Librarian
+- Ensures proper think messages for URL reading tasks
+
+## [1.7.0] - 2025-12-15
+
+### Added
+
+#### Web Search Migration to Librarian
+- **`search_web()`** tool in Librarian for web search via library-desk `/rag/search` endpoint
+- **`read_url()`** tool for single URL content extraction via Trafilatura
+- **`read_urls_batch()`** tool for parallel batch URL extraction (max 20 URLs)
+- `WebSearchResult`, `WebSearchResponse` models in LibraryDeskClient
+- `ContentExtractionResult`, `BatchExtractionResponse` models for content extraction
+- `search_web()`, `extract_content()`, `extract_content_batch()` methods in LibraryDeskClient
+- Comprehensive unit tests for new Librarian tools (`tests/agents/librarian/test_tools.py`)
+
+### Changed
+
+- Librarian capability updated with web search domains: "web", "url", "internet"
+- Tatlock system prompt now delegates web search to Librarian
+- `tatlock_core` capability reduced to computation/datetime only (no longer requires network)
+
+### Removed
+
+- `search_web` function from `src/agents/tatlock_core/tools.py`
+- `web_search_tool` from `tatlock_core_tools` list
+- `search_web` from legacy `src/agents/tools.py`
+- Search tests from `tests/agents/test_tools.py` (moved to Librarian tests)
+
+## [1.6.0] - 2025-12-15
+
+### Added
+
+#### Two-Phase Tatlock Execution
+- **Phase 1: Orchestration** - Executes tool calls and expert delegations, returns structured results
+- **Phase 2: Synthesis** - Synthesizes butler-toned response from gathered results
+- `orchestrate_tool_calls()` method in TatlockAgent for coordination phase
+- `synthesize_from_results()` method in TatlockAgent for synthesis phase
+- Guarantees butler personality in all responses by separating coordination from response generation
+
+#### Automatic Think Slugs
+- **Deterministic butler-perspective messages** during expert delegation (no LLM involved)
+- `ActionType` enum: RETRIEVE, RESEARCH, CREATE, CONTROL, RECORD
+- `HOUSEHOLD_THINK_MESSAGES` mapping with butler-perspective messages for all experts:
+  - Librarian: "Allow me to consult the archives, sir." / "I'm having the Librarian prepare a new entry."
+  - Biographer: "Let me consult the household records." / "I've asked the Biographer to take note, sir."
+  - Housekeeper: "I'm instructing the household staff now, sir." / "Allow me to inquire with the household staff."
+- `_detect_action_type()` function for keyword-based action detection
+- `get_think_message()` helper for retrieving appropriate messages
+- Streaming delegation wrappers: `stream_delegate_to_librarian()`, `stream_delegate_to_biographer()`, `stream_delegate_to_housekeeper()`
+- `STREAMING_DELEGATION_WRAPPERS` mapping in delegation.py
+- `get_streaming_delegation_tools()` method in HouseholdRegistry
+
+#### Steward Query Enrichment
+- **Auto-fill user context** (location, timezone) when not specified in query
+- `_build_enriched_query()` function in steward service
+- Regex word boundary matching for accurate location detection (avoids false positives)
+- `enriched_query` field added to `StewardRecommendation` schema
+- Automatic enrichment for weather queries (location), time queries (timezone), temperature preferences
+
+#### Documentation
+- **ORCHESTRATION_SCENARIOS.md** completely rewritten with:
+  - Mermaid flow diagrams for two-phase execution
+  - 4 new Housekeeper scenarios (light control, device status, parallel delegation)
+  - Biographer memory recording scenario
+  - Complete think slug reference tables
+  - Action type detection tables
+  - Updated architecture mindmap
+- **TESTING_IMPROVEMENTS.md** - LLM testing best practices for future implementation
+
+### Changed
+
+- `create_response_with_steward()` now uses two-phase execution
+- `_direct_delegation()` routes through synthesis phase for consistent butler tone
+- `_execute_single_delegation()` now supports housekeeper
+- Streaming response handler integrated with think slug system
+- All 326 unit tests passing
+
+## [1.5.0] - 2025-12-15
+
+### Added
+
+#### The Housekeeper Agent
+- **New home automation expert agent** following the Librarian pattern
+- `CoreAPIClient` for communicating with core-api service (Home Assistant wrapper)
+- 13 tools for home automation:
+  - Discovery: `list_areas`, `list_devices`, `get_device_state`
+  - Control: `turn_on`, `turn_off`, `toggle`
+  - Scenes: `list_scenes`, `activate_scene`
+  - Scripts: `list_scripts`, `run_script`
+  - Automations: `list_automations`, `toggle_automation`
+  - History: `get_history`
+- PydanticAI agent with system prompt for home automation tasks
+- `HouseholdCapability` registration with domains: lights, switches, automation, home, smart home, scene, script, device, climate, fan, cover, blinds
+- `delegate_to_housekeeper()` delegation wrapper
+- Config settings: `CORE_API_HOST`, `CORE_API_KEY`, `CORE_API_TIMEOUT`
+
+#### Development Port Change
+- **Dev server port changed from 8123 to 8777** to avoid conflict with Home Assistant default port
+- Updated `wakeup.sh`, E2E tests, and documentation
+
+### Changed
+
+- All unit tests pass (421 passed, 5 xfailed)
+- Housekeeper registered on startup alongside Librarian and Biographer
+
+## [1.4.0] - 2025-12-14
+
+### Added
+
+#### Environment-Aware Configuration
+- **Auto-selected logging level**: DEBUG for development, WARNING for production
+- **Auto-selected default user**: `llm_tester` for development (isolated test scope), `jpmschweitzer` for production
+- Properties `effective_log_level` and `effective_default_user` in config
+- User context logging at request entry with INFO level
+
+#### Direct Delegation Bypass
+- **Pure memory/librarian requests bypass Tatlock**: When Steward recommends only biographer/librarian, skip Tatlock LLM call
+- `_direct_delegation()` function for immediate expert agent execution
+- Reduces latency for memory-only requests
+
+#### Text-Based Delegation Fallback
+- **Parse text delegation patterns**: Handle LLM outputs like `[DELEGATE:biographer] task="..."`
+- Multiple pattern support for delegation parsing
+- Sequential and parallel execution with `[PARALLEL]` prefix
+
+#### Comprehensive E2E Test Suite
+- **22 new orchestration tests** in `tests/e2e/test_orchestration_e2e.py`
+- `QdrantVerifier` helper class for data verification
+- `assert_llm_behavior()` for flexible LLM output pattern matching
+- Test classes covering:
+  - Memory storage and recall
+  - Steward delegation
+  - Direct delegation bypass
+  - User context isolation (llm_tester vs production)
+  - Data verification in Qdrant
+  - Integration health checks
+  - Orchestration scenarios (weather, calculator, wiki, multi-expert)
+  - Error handling
+  - Evaluation reports
+- Updated `tests/e2e/README.md` with comprehensive documentation
+
+### Fixed
+
+- **Unit test mocks**: Updated Steward streaming tests to mock `run_with_scoped_tools_stream` (async generator)
+- **Temporal context in tests**: Tests now account for `_inject_temporal_context()` appending timestamps
+- **LLM non-determinism**: Integration tests use `pytest.xfail()` for LLM-dependent assertions
+- **Streaming test timeouts**: Increased timeouts (60-90s) for LLM processing time
+
+### Changed
+
+- All unit tests now pass (380 passed, 5 xfailed for LLM non-determinism)
+- E2E tests use `llm_tester` user for isolation from production data
+
+## [1.3.3] - 2025-12-14
+
+### Fixed
+
+- **Memory**: Fix Qdrant point IDs - use UUID5 instead of arbitrary strings
+
+## [1.3.2] - 2025-12-14
+
+### Fixed
+
+- **Memory**: Fix biographer tool type hints for Ollama compatibility (remove `| None` union types)
+
+## [1.3.1] - 2025-12-14
+
+### Fixed
+
+- **Memory**: Add biographer to delegation wrappers (was returning raw tools causing Ollama error)
+- **Config**: Add Qdrant host/port to .env.example
+
+## [1.3.0] - 2025-12-14
+
+### Fixed
+
+- **Memory**: Update Qdrant client to use `query_points` API (qdrant-client >= 1.10)
+
+### Changed
+
+- **Config**: Rename `REDIS_DB` to `REDIS_BENCHMARK_DB` for clarity
+- **Config**: Update Redis defaults to match stack allocation (benchmark=6, memory=1)
+
+## [1.2.5] - 2025-12-14
+
+### Fixed
+
+- **Dependencies**: Add missing `pydantic-settings` (not included in pydantic-ai-slim)
+
+## [1.2.4] - 2025-12-14
+
+### Added
+
+- **CI**: Trigger Watchtower update after successful image push
+
+## [1.2.3] - 2025-12-14
+
+### Fixed
+
+- **CI**: Upgrade to build-push-action@v6, disable provenance and sbom for Gitea registry
+
+## [1.2.2] - 2025-12-13
+
+### Fixed
+
+- **CI**: Add `provenance: false` to docker/build-push-action to fix Gitea registry push
+
+## [1.2.1] - 2025-12-13
+
+### Changed
+
+- **Dependency slimming**: Switched from `pydantic-ai` to `pydantic-ai-slim[openai]`
+  - Removes unused LLM provider SDKs (anthropic, boto3, cohere, google-genai, groq, huggingface)
+  - Production packages: 53 (down from ~158)
+  - Production footprint: 178MB
+  - Tatlock uses Ollama via OpenAI-compatible API, so only `openai` extra is needed
+  - See `DEPENDENCY_SLIM.md` for rollback instructions
+
 ## [1.2.0] - 2025-12-13
 
 ### Added
@@ -467,7 +1021,33 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - CORS middleware
 - Exception handlers (OpenAI-compatible error format)
 
-[Unreleased]: https://git.schweitz.net/jpmschweitzer/tatlock/compare/v1.2.0...main
+[Unreleased]: https://git.schweitz.net/jpmschweitzer/tatlock/compare/v2.1.0...main
+[2.1.0]: https://git.schweitz.net/jpmschweitzer/tatlock/compare/v2.0.5...v2.1.0
+[2.0.5]: https://git.schweitz.net/jpmschweitzer/tatlock/compare/v2.0.0...v2.0.5
+[2.0.0]: https://git.schweitz.net/jpmschweitzer/tatlock/compare/v1.11.0...v2.0.0
+[1.11.0]: https://git.schweitz.net/jpmschweitzer/tatlock/compare/v1.10.0...v1.11.0
+[1.10.0]: https://git.schweitz.net/jpmschweitzer/tatlock/compare/v1.9.0...v1.10.0
+[1.9.0]: https://git.schweitz.net/jpmschweitzer/tatlock/compare/v1.8.6...v1.9.0
+[1.8.6]: https://git.schweitz.net/jpmschweitzer/tatlock/compare/v1.8.5...v1.8.6
+[1.8.5]: https://git.schweitz.net/jpmschweitzer/tatlock/compare/v1.8.4...v1.8.5
+[1.8.4]: https://git.schweitz.net/jpmschweitzer/tatlock/compare/v1.8.3...v1.8.4
+[1.8.3]: https://git.schweitz.net/jpmschweitzer/tatlock/compare/v1.8.2...v1.8.3
+[1.8.2]: https://git.schweitz.net/jpmschweitzer/tatlock/compare/v1.8.1...v1.8.2
+[1.8.1]: https://git.schweitz.net/jpmschweitzer/tatlock/compare/v1.8.0...v1.8.1
+[1.8.0]: https://git.schweitz.net/jpmschweitzer/tatlock/compare/v1.7.0...v1.8.0
+[1.7.0]: https://git.schweitz.net/jpmschweitzer/tatlock/compare/v1.6.0...v1.7.0
+[1.6.0]: https://git.schweitz.net/jpmschweitzer/tatlock/compare/v1.5.0...v1.6.0
+[1.5.0]: https://git.schweitz.net/jpmschweitzer/tatlock/compare/v1.4.0...v1.5.0
+[1.4.0]: https://git.schweitz.net/jpmschweitzer/tatlock/compare/v1.3.3...v1.4.0
+[1.3.3]: https://git.schweitz.net/jpmschweitzer/tatlock/compare/v1.3.2...v1.3.3
+[1.3.2]: https://git.schweitz.net/jpmschweitzer/tatlock/compare/v1.3.1...v1.3.2
+[1.3.1]: https://git.schweitz.net/jpmschweitzer/tatlock/compare/v1.3.0...v1.3.1
+[1.3.0]: https://git.schweitz.net/jpmschweitzer/tatlock/compare/v1.2.5...v1.3.0
+[1.2.5]: https://git.schweitz.net/jpmschweitzer/tatlock/compare/v1.2.4...v1.2.5
+[1.2.4]: https://git.schweitz.net/jpmschweitzer/tatlock/compare/v1.2.3...v1.2.4
+[1.2.3]: https://git.schweitz.net/jpmschweitzer/tatlock/compare/v1.2.2...v1.2.3
+[1.2.2]: https://git.schweitz.net/jpmschweitzer/tatlock/compare/v1.2.1...v1.2.2
+[1.2.1]: https://git.schweitz.net/jpmschweitzer/tatlock/compare/v1.2.0...v1.2.1
 [1.2.0]: https://git.schweitz.net/jpmschweitzer/tatlock/compare/v1.1.0...v1.2.0
 [1.1.0]: https://git.schweitz.net/jpmschweitzer/tatlock/compare/v1.0.0a...v1.1.0
 [1.0.0a]: https://git.schweitz.net/jpmschweitzer/tatlock/compare/v0.2.5...v1.0.0a

CLAUDE.md

@@ -0,0 +1,196 @@
+# CLAUDE.md — tatlock
+
+Privacy-first homelab butler. An OpenAI-compatible orchestration API over local models, with
+household staff agents built on PydanticAI. Python 3.12 / FastAPI, `version = "2.4.3"`.
+Container `tatlock` on `docker-dataplane`, port **8000**. Redis DB **1** (memory), Qdrant for
+vectors.
+
+## Ports
+
+| | Port | How |
+|---|---|---|
+| Local dev | **8777** | `make run` — uvicorn reload, logs to `build/logs/server.log` |
+| Production | **8000** | container; `http://192.168.86.149:8000/health`, external `tatlock.schweitz.net` behind Authentik |
+
+Test endpoints against `localhost:8777` while developing. `localhost:8000` is the *container*.
+
+## Live contract
+
+`http://localhost:8000/openapi.json` — **5 paths**, `title: OpenAI-Compatible API`, `version:
+2.4.3` (verified 2026-08-09): `/`, `/health`, `/v1/models`, `/v1/chat/completions`,
+`/v1/responses`. `/v1/responses` is primary; `/v1/chat/completions` exists for Open WebUI.
+
+**The spec is the public surface, not the system.** The household capability registry is internal
+and appears nowhere in those 5 paths. Absence from the spec means "not exposed", not "does not
+exist".
+
+## Two traps that make the runtime look like the opposite of what it is
+
+**1. `src/anthropic` loads at startup; `src/ollama` does not — and Ollama is the primary
+backend.** A cold `import src.main` inside the container shows `agents, anthropic, chat, core,
+main, models, responses` — no `ollama`. The only import of it is a *function-body* one at
+`src/anthropic/model_selector.py:230`. Meanwhile `PREFER_CLOUD_BACKEND=false`, so every request
+actually goes to Ollama and the Claude path is off (see **workspace D-11**). Reading the module list
+naively gives you exactly the wrong answer: the package that looks live is the disabled fallback,
+and the one that looks dead is the hot path. Do not conclude anything about backends from
+`sys.modules`; read the config.
+
+**2. In-process singletons are empty outside the app.** `get_household_registry()`
+(`src/core/household_registry.py:334`) in a fresh `docker exec python` returns **0 members**,
+while the running app serves 2 models from it — it is populated at startup. Import the
+module-level definitions or ask the endpoint; never import a singleton and assume it is
+populated.
+
+## Stack decisions that bind this repo
+
+Recorded in the workspace vault, not here. Read before assuming anything about the LLM backend:
+
+```bash
+/home/jpmschweitzer/.local/bin/pql --vault /mnt/media/Projects decisions read workspace D-11
+```
+
+**workspace D-11 — the Claude migration is abandoned. Tatlock stays on Ollama.** Do not resume it and do
+not treat its remnants as unfinished work. What you will find, and why none of it is a TODO:
+`ANTHROPIC_MODEL` is set on the container (`claude-sonnet-4-20250514`) and never used because
+`PREFER_CLOUD_BACKEND=false`; `ANTHROPIC_API_KEY` is a variable reference whose literal was
+revoked 2026-08-09; `docs/claude-integration.md` documents a capability that exists but is
+switched off. The cost is deliberate: reasoning stays at `gemma4:e2b` scale because VRAM is
+shared with Speaches.
+
+`REDIS_BENCHMARK_DB=6` is allocated on the container but the benchmarking module was never
+implemented — see the gotcha below. Vestigial, like the Anthropic settings.
+
+## Critical gotchas
+
+**ASGITransport does NOT trigger FastAPI lifespan events.** The session-scoped `_initialize_app`
+fixture in `tests/conftest.py` calls `initialize_application()` explicitly via `asyncio.run()`.
+Without it the Ollama/Claude health checks never run: `_ollama_available` stays `None` (treated
+as available, so requests go to Ollama) and `_claude_available` stays `None` (treated as
+unavailable, so the Claude fallback never engages).
+
+**AsyncIO scope mismatch.** `asyncio_default_fixture_loop_scope = function` is set in
+`pyproject.toml`. Session-scoped async fixtures raise `ScopeMismatch`. Use a sync fixture with
+`asyncio.run()` for session-scoped initialization.
+
+**The butler persona prompt suppresses local-model tool calling.** With `TATLOCK_SYSTEM_PROMPT`
+attached, gemma4 reasons about calling the calculator, then answers from memory with wrong
+arithmetic — a different wrong product each run. `orchestrate_tool_calls()` therefore uses the
+terse `TATLOCK_ORCHESTRATION_PROMPT`; the persona is applied in `synthesize_from_results()`. Do
+not reattach the persona prompt to a tool-phase agent. `tool_choice: "required"` via `extra_body`
+does **not** force Ollama to call tools — advisory at best.
+
+**Claude Sonnet 5+ rejects sampling parameters.** `temperature`/`top_p`/`top_k` return 400. Use
+`get_sampling_settings()` from the model selector rather than passing `ModelSettings(temperature=…)`
+to agents that can run on the Claude fallback. `make test-contracts` pins this.
+
+**Integration test timeouts** are 120s to match `OLLAMA_TIMEOUT` (300s for the pure-Ollama
+fallback test, which Claude cannot rescue). GPU-resident numbers measured 2026-08-07 with
+gemma4:e2b at ~95 tok/s: full Steward → orchestrate → synthesize ~10–13s for simple turns;
+librarian-routed ~20–25s (not re-measured). **A single turn costs 3 sequential Ollama calls and
+~710 generated tokens even for "what is 61 plus 12?"** — mostly the model's own reasoning, paid
+three times. Cold model load is ~36s, avoided while pinned with `keep_alive: -1`; the
+`OLLAMA_KEEP_ALIVE=2h` default reintroduces it. Older "~35s steward / ~2 min flow" and "11–25s"
+figures are superseded — do not plan against them. `STEWARD_TIMEOUT` defaults to 60s.
+
+**`get_benchmark_store` does not exist.** `src/core/benchmarks.py` was never implemented, and
+`scripts/benchmark_analysis.py` references it and is broken. Do not add mocks for it in tests.
+
+**Steward tests need the household registry.** Use `register_household_members()` (sync) in
+fixtures, not `initialize_application()` (async). The steward extracts capabilities from the
+registry.
+
+## Commands
+
+```bash
+make setup            # venv + all dependencies
+make run              # dev server on 8777, reload, logs to build/logs/server.log
+make test             # unit tests, no external services
+make test-integration # needs Ollama (and Claude, if enabled)
+make test-contracts   # wire-level contract tests against live service boundaries
+make lint             # ruff linter + formatter check
+make typecheck        # mypy
+make clean            # remove caches and build artifacts
+```
+
+Always run pytest through the venv explicitly, to avoid environment mismatch:
+
+```bash
+.venv/bin/python -m pytest tests/
+.venv/bin/python -m pytest tests/core/ -v
+```
+
+Dependencies live in `pyproject.toml` (`[project.dependencies]`, `[project.optional-dependencies.dev]`).
+Copy `.env.example` to `.env` and configure Ollama, Redis and Qdrant hosts.
+
+**Contract tests before code review.** When the question is "do these two services still agree?",
+`make test-contracts` answers it by observing the live boundary; reading both codebases only tells
+you what should happen. Semantics: unreachable → skip, reachable-but-wrong-shape → fail.
+
+## Architecture
+
+Domain-first under `src/`: `agents/` (steward, librarian, biographer, housekeeper, tatlock_core),
+`core/`, `chat/`, `responses/`, `models/`, `ollama/`, `anthropic/`. Two tiers — the Steward routes,
+Tatlock coordinates. Group new work by domain, not by file type.
+
+## Internal service access
+
+`http://localhost:3002` reaches Gitea directly, bypassing Authentik SSO — verified returning
+`{"version":"1.27.1"}`. Useful for reading a sibling repo's raw files:
+
+```bash
+curl http://localhost:3002/jpmschweitzer/library-desk/raw/branch/main/README.md
+```
+
+The old AGENTS.md pointed at **`portainer-core`** for full-stack documentation. That repo is
+**deprecated** and must not be used as a source of infra facts; it was merged into
+`system-admin-toj/containers/`, where `CONTAINERS.md` is the live inventory.
+
+## Work tracking
+
+Work lives in **pql**, not a markdown TODO. **This repo's vault is standalone** — its tickets and
+internal decisions live here in `.pql/` and `governance/`, and travel with a clone, because
+`.pql/changelog/` is committed and replayed by the git hooks (workspace D-15). The databases are gitignored
+and rebuildable with `pql plan rebuild`.
+
+`pql` is **not** on the non-interactive `PATH` — invoke it as `/home/jpmschweitzer/.local/bin/pql`.
+From inside this repo no `--vault` is needed; pql anchors at the nearest `.git/` ancestor.
+
+```bash
+/home/jpmschweitzer/.local/bin/pql ticket list        # this repo's open work
+/home/jpmschweitzer/.local/bin/pql plan whatsnext     # next unblocked item, with context
+/home/jpmschweitzer/.local/bin/pql decisions list     # this repo's own decisions
+```
+
+Stack decisions that constrain this service need the flag:
+
+```bash
+/home/jpmschweitzer/.local/bin/pql --vault /mnt/media/Projects decisions list --domain tatlock-api
+```
+
+The workspace domain is `tatlock-api`, not `tatlock` — pql rejects a domain stem that prefixes
+another, and `tatlock` prefixes `tatlock-ui`. A `tatlock-api -> tatlock` symlink at the workspace
+root makes the directory answer to both (workspace D-15).
+
+Note `ticket new --decision D-N` resolves ids within **one** vault, so a ticket here cannot link
+to a workspace decision. Cite the id in the ticket body instead.
+
+## Git
+
+- **History is linear — no merge commits.** Work on `main`, or a short-lived branch that is
+  fast-forwarded and deleted. This repo's AGENTS.md mandated a feature branch for every change;
+  that rule was retired workspace-wide on 2026-08-08 and does not apply.
+- **Conventional Commits**: `feat:`, `fix:`, `refactor:`, `docs:`, `chore:`.
+- **Stage explicitly. Never `git add -A`** — denied by policy, and it sweeps in whatever else is
+  dirty, including secrets.
+- Update `CHANGELOG.md` for every user-facing change, under `[Unreleased]`.
+
+## Releasing
+
+Test locally first — the build-deploy loop is slow. Deploy only when a feature is complete.
+
+1. Ask whether a deploy is wanted; it is not automatic.
+2. Bump `version` in `pyproject.toml` (patch for fixes, minor for features).
+3. Move `[Unreleased]` entries into a dated section in `CHANGELOG.md`.
+4. Stage the changed files by name, commit, tag `vX.Y.Z`, `git push origin main --tags`.
+5. Gitea CI builds and pushes on the tag; Watchtower deploys.
+6. Verify: `curl http://192.168.86.149:8000/health`.

Dockerfile

@@ -5,8 +5,8 @@ WORKDIR /app
 RUN apt-get update && apt-get install -y curl \
     && rm -rf /var/lib/apt/lists/*
 
-COPY requirements.txt pyproject.toml ./
-RUN pip install --no-cache-dir -r requirements.txt
+COPY pyproject.toml ./
+RUN pip install --no-cache-dir .
 
 COPY src/ ./src/
 

IMPLEMENTATION_ROADMAP.md

@@ -1,920 +0,0 @@
-# Tatlock Implementation Roadmap
-
-> **Reference**: See [PHILOSOPHY.md](PHILOSOPHY.md) for the target architecture and vision
-
-This document outlines the phased implementation plan to transform the current OpenAI-compatible API into the full Tatlock household butler system.
-
-## Current State (v1.2.0 - Phase F Complete)
-
-**What we have**:
-- ✅ **The Orchestrator** - FastAPI infrastructure layer
-  - OpenAI-compatible API endpoints (Responses API + Chat Completions)
-  - Streaming coordination and conversation management
-  - Response format with reasoning support
-  - Test infrastructure (~400 tests)
-- ✅ **Two-Tier Architecture**
-  - The Steward analyzes requests and recommends capabilities
-  - Tatlock coordinates execution with scoped tools
-  - Real-time streaming of analysis and reasoning
-- ✅ **Household Staff**
-  - **Tatlock** (Butler): Primary interface with witty personality
-  - **The Steward**: Request analysis and capability recommendation
-  - **The Librarian**: Research via library-desk HybridRAG + wiki
-  - **The Biographer**: User memory, profiles, preferences, semantic recall
-- ✅ **Core Tools**
-  - Calculator, Date/Time toolkit, Web search (SearXNG)
-- ✅ **Memory System**
-  - Direct access layer (memory_service) for fast lookups
-  - Vector storage (Qdrant) for semantic recall
-  - Session cache (Redis) with 24h TTL
-  - Multi-tenancy via ContextVar
-- ✅ Mock agent (lorem-tester for testing)
-
-**What we need**:
-- More household staff (Developer, Secretary, Handyman, Housekeeper)
-- MCP (Model Context Protocol) integration
-- Dynamic model switching for specialized tasks
-- Full multi-tenant database (PostgreSQL)
-
----
-
-## Phase 1: Real LLM Integration - PydanticAI + Tools
-
-**Goal**: Connect to actual language models and establish the base plumbing
-
-**Note**: Ollama is an external service dependency (already running separately)
-
-### Deliverables
-
-1. **PydanticAI Integration** ✅
-   - PydanticAI → Ollama connection ✅
-   - Agent creation patterns ✅
-   - Streaming response handling ✅
-   - Error handling and retries ✅
-
-2. **Convert Tatlock Agent** ✅
-   - Convert Tatlock agent from mock to PydanticAI ✅
-   - British butler personality prompt ✅
-   - Research-oriented mindset ✅
-   - Streaming to reasoning output ✅
-   - Tool calling framework setup ✅
-
-3. **Permanent Tools** ✅
-   - Calculator: Safe mathematical expression evaluation ✅
-   - Date/Time toolkit: Current time, relative dates, time differences ✅
-   - Web search: SearXNG integration (external service) ✅
-   - Tool registration with PydanticAI ✅
-
-4. **Testing Infrastructure** ✅
-   - Integration tests with real LLM ✅
-   - Tool functionality tests ✅
-   - Response quality validation ✅
-   - 131 tests, 81.78% coverage ✅
-
-### Success Criteria
-- [x] **PydanticAI agents can call Ollama** (mistral-nemo:latest)
-- [x] **Streaming works end-to-end**
-- [x] **Tool calling framework functional**
-- [x] **Permanent tools working** (calculator, date/time, search)
-- [x] **Tests pass with real LLM**
-- [ ] Can switch models dynamically (e.g., Codestral for code)
-
-### Status
-**✅ MOSTLY COMPLETE** - Tatlock agent functional with permanent tools
-
-### Remaining Work
-- Dynamic model switching for specialized tasks (e.g., Codestral for coding)
-
-### Why First?
-Without real LLM integration, we can't meaningfully implement the Steward/Butler pattern. Everything else depends on having actual AI agents working.
-
----
-
-## Phase 2: Orchestration Layer - The Steward
-
-**Goal**: Implement the first-tier LLM call for tool/agent selection
-
-**Purpose**: The Steward performs crucial preparatory work before Tatlock engages with a request. By analyzing incoming requests and determining which tools, services, and household staff members will be needed, the Steward creates a curated recommendation that streamlines Tatlock's work and prevents cognitive overload.
-
-### Core Architecture
-
-The Steward operates as the first tier in the two-tier request flow:
-
-```
-User Request → Orchestrator → Steward Analysis → Recommendations → Tatlock (with scoped tools/agents)
-```
-
-**Key Principle**: The Steward narrows the scope to only relevant capabilities, making Tatlock's decision-making cleaner and more focused.
-
-### Deliverables
-
-#### 1. Tool & Agent Registry System
-
-**Purpose**: Centralized catalog of all available capabilities for the Steward to recommend
-
-**Implementation Details**:
-- **Registry Module** (`src/core/registry.py`)
-  - Tool registration decorator pattern
-  - Agent registration with capability metadata
-  - Category-based organization (computation, information, automation, communication)
-  - Dynamic tool/agent discovery and loading
-
-- **Tool Metadata Schema**
-  ```python
-  {
-    "name": "calculator",
-    "category": "computation",
-    "description": "Safe mathematical expression evaluation",
-    "capabilities": ["arithmetic", "algebra", "trigonometry"],
-    "cost": "low",  # computational cost indicator
-    "requires_network": false
-  }
-  ```
-
-- **Agent Metadata Schema**
-  ```python
-  {
-    "name": "developer",
-    "role": "The Developer",
-    "category": "technical",
-    "description": "Software development assistance",
-    "domains": ["code_generation", "debugging", "architecture"],
-    "specialized_model": "codestral",  # optional
-    "cost": "high"
-  }
-  ```
-
-- **Registry API**
-  - `get_all_tools()` - List all available tools
-  - `get_all_agents()` - List all expert agents
-  - `get_by_category(category)` - Filter by category
-  - `search_by_capability(query)` - Semantic search (future: vector search)
-
-**Testing**:
-- Unit tests for registration and retrieval
-- Test dynamic loading of new tools/agents
-- Validate metadata schemas
-
-#### 2. Steward PydanticAI Agent
-
-**Purpose**: First-tier LLM that analyzes requests and recommends relevant tools/agents
-
-**Implementation Details**:
-
-- **Agent Module** (`src/agents/steward.py`)
-  ```python
-  from pydantic_ai import Agent, RunContext
-  from pydantic import BaseModel
-
-  class StewardRecommendation(BaseModel):
-      """Structured output from Steward analysis"""
-      recommended_tools: list[str]
-      recommended_agents: list[str]
-      reasoning: str
-      estimated_complexity: str  # "simple", "moderate", "complex"
-      requires_multi_step: bool
-
-  steward = Agent(
-      'ollama:mistral-nemo',  # Same base model as Tatlock
-      result_type=StewardRecommendation,
-      system_prompt="""..."""
-  )
-  ```
-
-- **System Prompt Engineering**
-  - Role: Estate steward responsible for efficient household coordination
-  - Task: Analyze requests to determine needed resources
-  - Output: Structured recommendations with reasoning
-  - Constraints: Be conservative (recommend only truly relevant capabilities)
-  - Context: Full registry of available tools and agents
-
-- **Steward Tools**
-  ```python
-  @steward.tool
-  def get_available_capabilities(ctx: RunContext) -> dict:
-      """Get catalog of all available tools and agents."""
-      return {
-          "tools": registry.get_all_tools(),
-          "agents": registry.get_all_agents()
-      }
-  ```
-
-- **Request Analysis Flow**
-  1. Receive user request
-  2. Query capability registry via tool
-  3. Analyze request for required capabilities
-  4. Generate structured recommendation
-  5. Format as note to Tatlock
-
-**Testing**:
-- Test various request types (simple, complex, multi-domain)
-- Verify recommendations are relevant and not over-inclusive
-- Test structured output parsing
-- Validate reasoning quality
-
-#### 3. Request Preprocessing Pipeline
-
-**Purpose**: Integration layer that routes requests through Steward before Tatlock
-
-**Implementation Details**:
-
-- **Preprocessing Module** (`src/core/preprocessing.py`)
-  ```python
-  async def preprocess_request(user_request: str) -> EnrichedRequest:
-      """
-      1. Call Steward for analysis
-      2. Get recommendations
-      3. Enrich original request
-      4. Return scoped context for Tatlock
-      """
-      # Get Steward analysis
-      steward_result = await steward.run(user_request)
-      recommendations = steward_result.data
-
-      # Create note to Tatlock
-      steward_note = format_steward_note(recommendations)
-
-      # Build scoped tool/agent list
-      scoped_tools = get_scoped_tools(recommendations.recommended_tools)
-      scoped_agents = get_scoped_agents(recommendations.recommended_agents)
-
-      return EnrichedRequest(
-          original_request=user_request,
-          steward_note=steward_note,
-          available_tools=scoped_tools,
-          available_agents=scoped_agents,
-          metadata=recommendations
-      )
-  ```
-
-- **Note Formatting**
-  ```
-  === Internal Note from the Steward ===
-
-  Request Analysis:
-  {steward reasoning}
-
-  Recommended Tools:
-  - calculator: For mathematical computations
-  - web_search: To find current information
-
-  Recommended Household Staff:
-  - The Developer: For code generation assistance
-
-  Estimated Complexity: moderate
-  ===================================
-
-  [Original User Request]
-  ```
-
-- **Orchestrator Integration**
-  - Modify `src/responses/service.py` to call preprocessing
-  - Prepend Steward note to request before sending to Tatlock
-  - Limit Tatlock's tool access to recommended tools only
-  - Stream Steward's reasoning to output
-
-**Testing**:
-- Integration tests for full preprocessing flow
-- Test request enrichment format
-- Verify tool scoping works correctly
-- Test streaming of Steward reasoning
-
-#### 4. Real-Time Transparency
-
-**Purpose**: Stream Steward's analysis to user's reasoning output
-
-**Implementation Details**:
-
-- **Streaming Integration** (`src/responses/streaming.py`)
-  - Add Steward analysis phase to stream
-  - Format as reasoning item
-  - Include recommendation summary
-
-- **Example Output to User**:
-  ```
-  [Reasoning]
-  Consulting the Steward for resource planning...
-
-  The Steward's Analysis:
-  - Request requires mathematical computation
-  - Need to verify current information via web search
-  - May benefit from Developer's code expertise
-
-  Recommended: calculator, web_search, The Developer
-
-  Proceeding with scoped resources...
-  ```
-
-**Testing**:
-- Test streaming of Steward analysis
-- Verify formatting in Open WebUI
-- Test error handling if Steward fails
-
-#### 5. Model Efficiency Optimization
-
-**Purpose**: Ensure the base model stays loaded in VRAM
-
-**Implementation Details**:
-
-- **Shared Model Configuration**
-  - Both Steward and Tatlock use `ollama:mistral-nemo` by default
-  - Sequential calls (Steward → Tatlock) keep model hot
-  - No reload delays between tiers
-
-- **Performance Monitoring**
-  - Log response times for Steward calls
-  - Track total request latency (Steward + Tatlock)
-  - Identify optimization opportunities
-
-**Testing**:
-- Benchmark Steward → Tatlock call latency
-- Verify model stays loaded between calls
-- Test performance under load
-
-### Implementation Strategy
-
-#### Week 1-2: Foundation
-- [ ] Design and implement registry system
-- [ ] Create tool/agent metadata schemas
-- [ ] Build registry API with tests
-- [ ] Migrate existing tools to registry
-
-#### Week 3-4: Steward Agent
-- [ ] Create Steward PydanticAI agent
-- [ ] Engineer system prompt for analysis
-- [ ] Implement structured recommendation output
-- [ ] Add registry query tool
-- [ ] Test with various request types
-
-#### Week 5-6: Integration
-- [ ] Build request preprocessing pipeline
-- [ ] Implement note formatting
-- [ ] Integrate with Orchestrator
-- [ ] Add streaming transparency
-- [ ] Tool scoping for Tatlock
-
-#### Week 7: Testing & Refinement
-- [ ] End-to-end integration tests
-- [ ] Performance optimization
-- [ ] Prompt refinement based on results
-- [ ] Documentation and examples
-
-### Success Criteria
-
-- [x] **Steward analyzes incoming requests** using PydanticAI agent
-- [x] **Produces structured recommendations** (tools, agents, reasoning)
-- [x] **Recommendations formatted as prepended note** to Tatlock
-- [x] **Tool registry is queryable and extensible** via clean API
-- [x] **Steward output visible in reasoning stream** for transparency
-- [x] **Only recommended tools available** to Tatlock (scoped context)
-- [x] **Base model stays loaded** between Steward and Tatlock calls
-- [x] **Recommendations are accurate** (not over/under-inclusive)
-- [x] **Integration tests pass** for full Steward → Tatlock flow
-
-### Status
-**✅ COMPLETE** (v0.2.5)
-
-### Performance Targets
-
-- **Steward Analysis Time**: < 2 seconds for typical requests
-- **Total Added Latency**: < 3 seconds including streaming
-- **Recommendation Accuracy**: > 90% relevance (manual evaluation)
-- **Model Reload Delay**: 0 seconds (model stays hot)
-
-### Risk Mitigation
-
-**Risk**: Steward recommendations too broad (defeats purpose)
-- Mitigation: Conservative prompt engineering, test with diverse requests, iterate
-
-**Risk**: Added latency unacceptable to users
-- Mitigation: Stream Steward reasoning for transparency, optimize prompt, parallel processing where possible
-
-**Risk**: Tool registry becomes unwieldy
-- Mitigation: Good categorization, semantic search (future), regular pruning
-
-**Risk**: Steward and Tatlock models compete for VRAM
-- Mitigation: Use same base model, sequential calls, monitor memory
-
-### Future Enhancements (Post-Phase 2)
-
-- **Semantic Search**: Vector-based capability search instead of metadata lookup
-- **Learning from Usage**: Track which recommendations work well, adjust over time
-- **Confidence Scores**: Steward provides confidence for each recommendation
-- **Request Classification**: Cache classifications for similar requests
-- **Multi-Model Support**: Allow Steward to recommend specialized models for specific tasks
-
-### Estimated Effort
-
-**7-8 weeks** - Core intelligence routing with comprehensive implementation
-
-### Why Second?
-
-The Steward is the foundation of the household architecture. Without it, we'd need to expose all tools/agents to Tatlock, creating cognitive overload and poor decision-making. The Steward enables the focused expertise pattern that makes the whole system work.
-
----
-
-## Phase 3: The Butler - Tatlock Agent
-
-**Goal**: Implement the second-tier coordinator with personality within the existing Orchestrator infrastructure
-
-**Context**: The Orchestrator (FastAPI infrastructure) already exists. This phase implements the real Tatlock PydanticAI agent to replace the current mock agent.
-
-### Deliverables
-
-1. **Butler Agent (Tatlock)**
-   - PydanticAI agent implementation within Orchestrator
-   - Personality prompt engineering (witty British butler)
-   - Tool calling framework
-   - Multi-agent coordination logic
-
-2. **Scoped Tool Access**
-   - Filter tools based on Steward recommendations
-   - Dynamic tool loading for Butler context
-   - Tool execution framework
-   - Result aggregation
-
-3. **Real-Time Reasoning Output**
-   - Stream all Butler activities to reasoning output
-   - Tool call progress indicators
-   - Expert agent consultation messages
-   - Wait time transparency
-
-### Success Criteria
-- [x] Tatlock receives enriched requests (user + Steward notes)
-- [x] Only recommended tools are available
-- [x] Tatlock coordinates multiple tool calls
-- [x] All actions streamed to reasoning output
-- [x] Responses have consistent personality
-- [x] Synthesizes multi-source results coherently
-
-### Status
-**✅ COMPLETE** (v1.1.0)
-
-### Estimated Effort
-**4-5 weeks** - Complex coordination logic
-
----
-
-## Phase 4: Expert Household Staff - Core Agents
-
-**Goal**: Implement the initial set of domain-specific expert agents
-
-### Priority Expert Agents
-
-1. **The Librarian** (Research & Knowledge Management) ✅ **COMPLETE** (v1.1.0)
-   - Research assistance via library-desk HybridRAG
-   - Wiki page management (search, create, update)
-   - Semantic vector search
-   - Knowledge graph queries
-   - Dossier browsing
-
-2. **The Biographer** (User Memory) ✅ **COMPLETE** (v1.2.0)
-   - User profile management (name, location, timezone)
-   - Preference storage (units, theme)
-   - Semantic memory recall ("What car do I drive?")
-   - Fact storage from conversations
-   - Session context caching
-
-3. **The Developer** (Software Development) 🔜 **Planned**
-   - Code generation assistance
-   - Debugging support
-   - Documentation generation
-   - Architecture guidance
-   - *Rationale: Directly supports building the system itself*
-
-4. **The Handyman** (System Maintenance) 🔜 **Planned**
-   - System status queries
-   - Log analysis
-   - Basic troubleshooting
-   - Infrastructure monitoring
-
-5. **The Secretary** (Scheduling & Organization) 🔜 **Planned**
-   - Calendar integration
-   - Task management
-   - Reminder system
-   - Schedule conflict detection
-
-6. **The Housekeeper** (Home Automation) 🔜 **Planned**
-   - Home Assistant integration
-   - Device control interface
-   - Status queries
-   - Automation triggers
-
-### Each Agent Includes
-- Specialized prompt and personality
-- Domain-specific tools
-- MCP integration points (where applicable)
-- Integration with Butler orchestration
-
-### Success Criteria
-- [x] Each agent implemented as separate module
-- [x] Agents callable via tool framework
-- [x] Agents use specialized prompts
-- [x] Results integrate cleanly with Butler
-- [ ] Can invoke specialized models (e.g., Codestral for Developer)
-
-### Status
-**🔶 PARTIAL** - Librarian and Biographer complete, others planned
-
-### Estimated Effort
-**6-8 weeks** - Parallel development possible
-
----
-
-## Phase 5: Persistence Layer - Database & Multi-Tenancy
-
-**Goal**: Add persistent storage and multi-user support when needed
-
-### Deliverables
-
-1. **PostgreSQL Integration**
-   - Docker compose configuration for PostgreSQL
-   - Database schema design with tenant isolation
-   - Alembic migrations setup
-   - SQLAlchemy models
-
-2. **Multi-Tenant Architecture**
-   - Tenant identification middleware
-   - Tenant-scoped database sessions
-   - User authentication system (basic)
-   - Per-tenant data isolation
-
-3. **Core Data Models**
-   - Users and tenants
-   - Conversations and messages (migrate from in-memory)
-   - Agent interactions log
-   - System configuration and preferences
-
-4. **Migration Strategy**
-   - Gradual migration from in-memory to database
-   - Backward compatibility during transition
-   - Data export/import utilities
-
-### Success Criteria
-- [ ] PostgreSQL container running
-- [ ] Multiple users can authenticate separately
-- [ ] Each user sees only their own data
-- [ ] Conversations persist across restarts
-- [ ] Database migrations work correctly
-- [ ] Tests verify tenant isolation
-
-### Estimated Effort
-**3-4 weeks** - Data layer foundation
-
-### Why Later?
-The core orchestration (Steward → Butler → Experts) can work entirely with in-memory state. We only need database persistence when we want conversations to survive restarts and multiple users to have isolated experiences.
-
----
-
-## Phase 6: Extended Services Integration
-
-**Goal**: Connect to additional supporting services
-
-### Services to Integrate
-
-1. **Redis (Memory & Caching)** ✅ **COMPLETE** (v1.2.0)
-   - Benchmark storage (db=1)
-   - Memory cache for sessions (db=2)
-   - 24h TTL for session context
-   - Recent entities tracking
-
-2. **Qdrant (Vector Storage)** ✅ **COMPLETE** (v1.2.0)
-   - Per-user memory collections
-   - 768-dim nomic-embed-text vectors
-   - Semantic search for recall
-   - Type-based filtering
-
-3. **SearxNG (Web Search)** ✅ **COMPLETE** (v0.2.0)
-   - Search tool integration
-   - Result processing
-   - Privacy-preserving queries
-
-4. **library-desk (Research API)** ✅ **COMPLETE** (v1.1.0)
-   - HybridRAG search
-   - Wiki management
-   - Knowledge graph queries
-
-### Success Criteria
-- [x] Services communicate correctly
-- [x] Tatlock can invoke web search
-- [x] Redis used for session data
-- [x] Qdrant stores user memories
-- [x] Ollama serves the base model
-
-### Status
-**✅ COMPLETE** - All core services integrated
-
-### Estimated Effort
-**3-4 weeks** - Infrastructure setup
-
----
-
-## Phase 7: MCP (Model Context Protocol) Integration
-
-**Goal**: Enable rich tool integrations via MCP
-
-### Deliverables
-
-1. **MCP Server Framework**
-   - MCP server implementation
-   - Tool registration via MCP
-   - Schema validation
-   - Error handling
-
-2. **MCP Client in Agents**
-   - PydanticAI MCP integration
-   - Tool discovery from MCP servers
-   - Dynamic tool loading
-   - Result processing
-
-3. **Initial MCP Tools**
-   - File system operations
-   - Database queries
-   - API integrations
-   - System commands
-
-### Success Criteria
-- [ ] MCP server running
-- [ ] Tools exposed via MCP protocol
-- [ ] Agents can discover and use MCP tools
-- [ ] New tools addable without code changes
-- [ ] MCP tools visible in Steward recommendations
-
-### Estimated Effort
-**3-4 weeks** - Standards-based integration
-
----
-
-## Phase 8: Advanced Memory & Context
-
-**Goal**: Implement sophisticated memory and context management
-
-### Deliverables
-
-1. **Long-Term Memory** ✅ **COMPLETE** (v1.2.0 - Phase F)
-   - Memory service for direct key-based access
-   - Qdrant vector storage for semantic recall
-   - Embedding via nomic-embed-text
-   - The Biographer agent for memory management
-
-2. **Session Memory** ✅ **COMPLETE** (v1.2.0)
-   - Redis session cache with 24h TTL
-   - Recent entities tracking
-   - Conversation context preservation
-   - Multi-tenancy via ContextVar
-
-3. **Steward Integration** ✅ **COMPLETE** (v1.2.0)
-   - Memory pre-fetch during request analysis
-   - Profile/preferences included in context
-   - Keyword-based context determination
-
-4. **Context Management** 🔜 **Future**
-   - Smart context window trimming
-   - Conversation branching
-   - Topic tracking
-   - Memory retrieval integration
-
-5. **Personalization** 🔜 **Future**
-   - User preference learning
-   - Interaction pattern analysis
-   - Adaptive responses
-   - Custom agent personalities per user
-
-### Success Criteria
-- [x] User facts stored in Qdrant with semantic search
-- [x] Profile and preferences accessible via memory_service
-- [x] Session context cached in Redis
-- [x] User preferences affect responses (via Steward pre-fetch)
-- [ ] Conversations automatically embedded to Qdrant
-- [ ] Memory improves over time (learning from interactions)
-
-### Status
-**🔶 PARTIAL** - Core memory system complete, advanced features planned
-
-### Estimated Effort
-**4-5 weeks** - AI/ML heavy (remaining work)
-
----
-
-## Phase 9: Extended Household Staff
-
-**Goal**: Add specialized agents for additional domains
-
-### Future Agents
-
-1. **The Librarian** (Knowledge Management)
-   - Personal documentation indexing
-   - Research assistance
-   - Knowledge base queries
-   - Reference management
-
-2. **The Accountant** (Financial Tracking)
-   - Expense tracking
-   - Budget monitoring
-   - Financial reports
-   - Transaction categorization
-
-3. **The Chef** (Meal Planning)
-   - Recipe management
-   - Meal planning
-   - Nutrition tracking
-   - Grocery lists
-
-4. **Others as Needed**
-   - Domain-specific as requirements emerge
-
-### Success Criteria
-- [ ] Each new agent follows household pattern
-- [ ] Integrates with Steward/Butler flow
-- [ ] Has appropriate specialized tools
-- [ ] Documented in PHILOSOPHY.md updates
-
-### Estimated Effort
-**Ongoing** - Add as needed
-
----
-
-## Phase 10: User Experience Refinement
-
-**Goal**: Polish the interaction experience
-
-### Deliverables
-
-1. **Personality Tuning**
-   - Refine Tatlock's wit and tone
-   - Consistent household character
-   - Cultural references appropriate
-   - Humor that doesn't annoy
-
-2. **Transparency Improvements**
-   - Better progress indicators
-   - Clearer reasoning explanations
-   - Informative wait messages
-   - Error message clarity
-
-3. **Performance Optimization**
-   - Response time improvements
-   - Model loading optimization
-   - Caching strategies
-   - Streaming smoothness
-
-### Success Criteria
-- [ ] Users find Tatlock engaging
-- [ ] Wait times feel reasonable
-- [ ] Errors are understandable
-- [ ] System feels responsive
-
-### Estimated Effort
-**Ongoing** - Continuous improvement
-
----
-
-## Phase 11: Production Hardening
-
-**Goal**: Make the system production-ready for homelab deployment
-
-### Deliverables
-
-1. **Deployment**
-   - Complete docker-compose stack
-   - Environment configuration
-   - Backup strategies
-   - Update procedures
-
-2. **Monitoring**
-   - Health checks
-   - Performance metrics
-   - Error tracking
-   - Usage analytics
-
-3. **Security**
-   - Authentication hardening
-   - Rate limiting
-   - Input validation
-   - Audit logging
-
-4. **Documentation**
-   - Installation guide
-   - Configuration reference
-   - Troubleshooting guide
-   - Architecture documentation
-
-### Success Criteria
-- [ ] One-command deployment
-- [ ] System health is monitorable
-- [ ] Secure for homelab use
-- [ ] Well documented
-
-### Estimated Effort
-**3-4 weeks** - Production polish
-
----
-
-## Dependencies Between Phases
-
-```
-Phase 1 (Ollama + PydanticAI) ← Foundation for all AI
-    ↓
-Phase 2 (Steward)
-    ↓
-Phase 3 (Butler/Tatlock)
-    ↓
-Phase 4 (Expert Agents) ← Phase 7 (MCP) can enhance
-    ↓
-Phase 5 (Database/Multi-Tenancy) ← Can be deferred
-    ↓
-Phase 6 (Extended Services) → Phase 8 (Advanced Memory)
-    ↓
-Phase 9 (Extended Staff) → Phase 10 (UX) → Phase 11 (Production)
-```
-
-**Critical Path**: Phases 1 → 2 → 3 → 4 must be sequential
-**Can Be Deferred**: Phase 5 (Database) until you need persistence
-**Parallel Opportunities**: Phase 6 and 7 can overlap; Phase 9 and 10 ongoing
-
----
-
-## Overall Timeline Estimate
-
-**Minimum Viable Household** (Phases 1-4): **15-20 weeks**
-- Working Steward → Butler → Expert Agents with real LLM
-- In-memory state (no persistence needed yet)
-- Core household functional
-
-**With Persistence** (Phases 1-5): **18-24 weeks**
-- Add database and multi-tenancy
-- Conversations survive restarts
-- Multiple users supported
-
-**Full-Featured System** (Phases 1-9): **35-45 weeks**
-- All services integrated
-- Advanced memory and context
-- Extended household staff
-
-**Production-Ready** (All phases): **40-50 weeks**
-- Polished UX
-- Hardened for homelab deployment
-- Fully documented
-
-*Note: Timeline assumes consistent part-time development effort*
-
----
-
-## Success Metrics
-
-### Technical
-- System implements PHILOSOPHY.md patterns
-- All household roles functional
-- Multi-tenant isolation verified
-- Real-time reasoning transparency working
-- MCP integration complete
-
-### User Experience
-- Tatlock feels like interacting with a butler
-- Wait times are transparent and acceptable
-- Expert agents provide value in their domains
-- System is reliable and trustworthy
-
-### Architecture
-- Clean separation between household roles
-- Easy to add new agents/tools
-- Model efficiency (base model stays loaded)
-- Scales to household + friends usage
-
----
-
-## Risk Management
-
-### High Risk Items
-1. **PydanticAI + Ollama integration complexity**
-   - Mitigation: Prototype early, iterate on connection layer
-
-2. **Multi-agent coordination complexity**
-   - Mitigation: Start simple, add coordination gradually
-
-3. **Model performance on homelab hardware**
-   - Mitigation: Model selection, quantization, optimization
-
-4. **Prompt engineering for personality consistency**
-   - Mitigation: Extensive testing, user feedback, iteration
-
-### Medium Risk Items
-- MCP protocol adoption and tooling maturity
-- Vector embedding quality for memory
-- Home automation integration variability
-- User authentication security
-
----
-
-## Next Steps
-
-1. **Priority**: Implement The Developer agent for code assistance
-2. **Integration**: Add Home Assistant integration for The Housekeeper
-3. **Calendar**: Integrate scheduling service for The Secretary
-4. **Ongoing**: Add more household staff as needed
-
----
-
-**Document Status**: Active planning document
-**Created**: 2025-12-06
-**Last Updated**: 2025-12-13

Makefile

@@ -0,0 +1,89 @@
+.PHONY: help setup run test test-unit test-integration test-contracts lint typecheck clean
+
+VENV := .venv
+PYTHON := $(VENV)/bin/python
+PIP := $(VENV)/bin/pip
+PYTEST := $(VENV)/bin/pytest
+RUFF := $(VENV)/bin/ruff
+MYPY := $(VENV)/bin/mypy
+UVICORN := $(VENV)/bin/uvicorn
+
+HOST := 0.0.0.0
+PORT := 8777
+
+help: ## Show this help
+	@grep -E '^[a-zA-Z_-]+:.*?## .*$$' $(MAKEFILE_LIST) | sort | awk 'BEGIN {FS = ":.*?## "}; {printf "\033[36m%-20s\033[0m %s\n", $$1, $$2}'
+
+setup: ## Create venv and install all dependencies
+	python3 -m venv $(VENV)
+	$(PIP) install --upgrade pip
+	$(PIP) install -e ".[dev]"
+	# Exit 0 from pip install is not evidence the environment works (D-24) - the
+	# 2026-08-09 core-api incident was exactly this: a venv that "installed fine"
+	# but was missing a declared dependency, surfacing as 11 collection errors
+	# that read like broken imports rather than an environment problem. Collection
+	# is the right cheap check here for that same reason: it imports every test
+	# module (and everything they import) without running the suite, so a missing
+	# or mismatched dependency fails setup itself instead of showing up later as a
+	# mysterious test failure. Scoped like `make test` (excludes e2e/integration/
+	# contracts, which need external services) and --no-cov since coverage
+	# instrumentation is irrelevant to "does this collect".
+	$(PYTEST) --collect-only -q --ignore=tests/e2e --ignore=tests/integration --ignore=tests/contracts --no-cov
+
+run: ## Start the development server on port 8777
+	@mkdir -p build/logs
+	@if lsof -Pi :$(PORT) -sTCP:LISTEN -t >/dev/null 2>&1; then \
+		echo "Error: Port $(PORT) is already in use"; \
+		echo "Run: lsof -i :$(PORT) to see what's using it"; \
+		exit 1; \
+	fi
+	$(UVICORN) src.main:app --reload --host $(HOST) --port $(PORT) 2>&1 | tee build/logs/server.log
+
+test: ## Run unit tests (no external services needed)
+	$(PYTEST) --ignore=tests/e2e --ignore=tests/integration --ignore=tests/contracts
+
+test-unit: test ## Alias for test
+
+test-integration: ## Run integration tests (needs Claude/Ollama)
+	$(PYTEST) tests/agents/test_tatlock_agent.py -v
+
+test-contracts: ## Wire-level contract tests against live service boundaries
+	$(PYTEST) tests/contracts -v --no-cov
+
+lint: ## Run ruff linter and formatter check
+	$(RUFF) check src tests
+	$(RUFF) format --check src tests
+
+typecheck: ## Run mypy type checking
+	$(MYPY) src
+
+clean: ## Remove build artifacts, caches, and coverage reports
+	rm -rf .cache build
+	find . -type d -name __pycache__ -exec rm -rf {} + 2>/dev/null || true
+
+# git hands a hook a non-login shell, which never sees ~/.local/bin — where
+# gitleaks lands. Without this the scan reports "not installed" on every push,
+# which is a check that fails open (D-24).
+export PATH := $(HOME)/.local/bin:/usr/local/bin:$(PATH)
+
+.PHONY: secrets
+secrets: ## Scan the commits about to be pushed for credentials
+	@ci/secrets.sh
+
+# The call surface is identical in every repo; what it runs is not.
+#
+# `secrets` runs first, deliberately: it is the only failure here that cannot be
+# undone by fixing it afterwards. A failed lint costs another commit; a pushed
+# credential is cached and indexed whether or not it is later deleted.
+#
+# Some of these fail today, and are left wired anyway. The state was measured
+# once and written down in T-56 rather than being worked around here — a gate
+# quietly narrowed to what already passes is a gate that reports success for
+# doing nothing, which is the failure this workspace keeps rediscovering.
+.PHONY: pre-push
+pre-push: secrets lint ## Everything the pre-push hook runs
+	@echo "  --   not gated here yet: typecheck (T-1), test (T-56)"
+	@echo "       typecheck reports 95 errors in 31 files and has never passed, so"
+	@echo "       gating on it blocked every push to this repo — including the commit"
+	@echo "       that added the gate. Run 'make typecheck' before pushing anything"
+	@echo "       that touches types; T-1 is the pass that earns this line's removal."

ORCHESTRATION_SCENARIOS.md

@@ -1,679 +0,0 @@
-# Orchestration Scenarios and Tool Flows
-
-This document outlines example scenarios of varying complexity to illustrate the desired orchestration patterns between Tatlock (Butler/Coordinator), expert agents (The Librarian, etc.), and the user.
-
-## Architecture Overview
-
-```
-User Request
-     ↓
-[Steward] → Analyzes request, has visibility into ALL capabilities
-           → Makes routing decision: which experts needed
-           → Passes simplified instruction to Tatlock (not raw tool schemas)
-     ↓
-[Tatlock/Butler] → Coordinator, receives "use Librarian for wiki creation"
-                 → Calls expert agents as tools
-                 → Synthesizes responses into butler-voice answer
-     ↓
-[Expert Agents] → The Librarian, Home Automation, Memory, etc.
-                → Each has their own specialized tools
-                → Return structured results to Tatlock
-     ↓
-[External APIs] → library-desk, home-assistant, user-db, etc.
-```
-
-**Key Principles**:
-
-1. **Steward sees everything** - Has access to all capability descriptions to make informed routing decisions
-2. **Simplified passthrough** - Tatlock receives "delegate to Librarian for research" not 16 tool schemas
-3. **Expert agents are tools** - Tatlock calls `librarian_agent(task)`, not `hybrid_search()` directly
-4. **Each expert owns their tools** - Librarian has wiki tools, Home Automation has device tools
-5. **Results flow up** - Tatlock synthesizes all expert responses into coherent butler answer
-
----
-
-## Scenario 1: Weather Check (Multi-Step with Memory Lookup)
-
-**User**: "What's the weather like?"
-
-### Complexity Analysis
-
-This seemingly simple request requires:
-1. **Location determination** - Where does the user want weather for?
-2. **Memory/database lookup** - Retrieve user's home location or current location
-3. **Weather data fetch** - Search for weather at determined location
-
-### Flow
-
-```
-1. Steward Analysis
-   → Capabilities needed: memory (user context), tatlock_core (web search)
-   → Complexity: moderate
-   → Note: Location must be determined before weather lookup
-
-2. Tatlock Execution - Step 1
-   <think>User asked about weather but didn't specify location.
-   Checking user profile for home location...</think>
-   → Calls: memory_agent(task: "get user home location")
-   → Memory queries user database
-   → Returns: "User home location: Amsterdam, Netherlands"
-
-3. Tatlock Execution - Step 2
-   <think>User is based in Amsterdam. Fetching current weather...</think>
-   → Calls: search_web("current weather Amsterdam Netherlands")
-   → Receives: "Amsterdam: 12°C, light rain, humidity 78%"
-
-4. Response
-   "Currently 12°C with light rain in Amsterdam, sir. You might want
-   to grab an umbrella if you're heading out."
-```
-
-### Intra-System Prompts
-
-**Steward → Tatlock Note**:
-```
-Weather query - location not specified.
-1. First: Query memory for user's location (home or current)
-2. Then: Search weather for that location
-Capabilities: memory, tatlock_core
-Complexity: moderate
-```
-
-**Tatlock → Memory Agent**:
-```
-Task: Retrieve user's location for weather query.
-Context: User asked about weather without specifying location.
-Action required: Return user's home location or current known location.
-
-Reference (user's original request): "What's the weather like?"
-```
-
-**Memory Agent → Tatlock Response**:
-```
-User location retrieved:
-- Home location: Amsterdam, Netherlands
-- Last known location: Amsterdam (home)
-- Location confidence: high
-- Source: user profile settings
-```
-
-### Alternative Flow: Location Ambiguity
-
-If user has multiple locations or is traveling:
-
-```
-Memory Agent → Tatlock Response:
-  User has multiple locations:
-  - Home: Amsterdam, Netherlands
-  - Office: Rotterdam, Netherlands
-  - Currently traveling: Unknown
-
-  Recommendation: Ask user to clarify or use home location as default.
-```
-
-Tatlock could then either:
-- Ask user: "Shall I check the weather in Amsterdam, sir, or elsewhere?"
-- Default to home: Use Amsterdam and mention the assumption
-
----
-
-## Scenario 2: Adjust Temperature Based on Weather (Conditional Multi-Expert)
-
-**User**: "Check the weather and if it's cold, turn up the heating"
-
-### Complexity Analysis
-
-This requires:
-1. **Location lookup** - Where to check weather (implicit: user's home)
-2. **Weather fetch** - Get current outdoor temperature
-3. **Conditional evaluation** - Is it "cold"? (requires threshold judgment)
-4. **Home automation** - Adjust heating if condition met
-
-### Flow
-
-```
-1. Steward Analysis
-   → Capabilities needed: memory, tatlock_core, home_automation
-   → Complexity: moderate
-   → Note: Conditional logic - heating only if cold
-   → Sequence: location → weather → evaluate → (maybe) heating
-
-2. Tatlock Execution - Step 1
-   <think>Need to check weather at user's location first...</think>
-   → Calls: memory_agent(task: "get user home location")
-   → Returns: "Amsterdam, Netherlands"
-
-3. Tatlock Execution - Step 2
-   <think>Fetching weather for Amsterdam...</think>
-   → Calls: search_web("current weather Amsterdam Netherlands")
-   → Receives: "Current temperature: 8°C, cloudy, wind 15km/h"
-
-4. Tatlock Evaluation
-   <think>Temperature is 8°C - that's cold by most standards.
-   User requested heating adjustment if cold. Will proceed...</think>
-
-5. Tatlock Execution - Step 3
-   <think>Delegating heating adjustment to Home Automation...</think>
-   → Calls: home_automation_agent(task)
-   → Home Automation executes: set_thermostat(temperature=21)
-   → Receives: "Thermostat set to 21°C"
-
-6. Response
-   "It's rather brisk outside at 8°C, sir. I've taken the liberty of raising
-   the heating to a comfortable 21°C. The house should warm up shortly."
-```
-
-### Intra-System Prompts
-
-**Steward → Tatlock Note**:
-```
-Conditional weather-to-heating request.
-1. Get user location from memory
-2. Check weather at location
-3. IF cold (suggest: below 15°C): delegate to home_automation to increase heating
-4. IF not cold: inform user, no action needed
-Capabilities: memory, tatlock_core, home_automation
-Complexity: moderate (conditional)
-```
-
-**Tatlock → Home Automation Agent**:
-```
-Task: Increase heating due to cold weather conditions.
-Context: Current outdoor temperature is 8°C in Amsterdam. User requested
-heating adjustment if cold - condition is met.
-Action required: Set thermostat to comfortable temperature (recommend 21°C).
-
-Reference (user's original request): "Check the weather and if it's cold, turn up the heating"
-```
-
-**Home Automation → Tatlock Response**:
-```
-Action completed: Thermostat adjusted
-- Previous setting: 18°C
-- New setting: 21°C
-- Zone: All zones (whole house)
-- Estimated time to reach target: 15 minutes
-- Note: Eco mode disabled temporarily for faster heating
-```
-
-### Alternative Flow: Not Cold
-
-```
-Weather returns: "Current temperature: 18°C, partly sunny"
-
-Tatlock Evaluation:
-<think>18°C isn't cold - no heating adjustment needed.
-Will inform user that conditions don't warrant heating...</think>
-
-Response:
-"It's a pleasant 18°C outside, sir - not particularly cold.
-I don't believe the heating requires adjustment, but do let
-me know if you'd like it warmer regardless."
-```
-
----
-
-## Scenario 3: Adjust Lights Based on Time (Conditional Logic)
-
-**User**: "Set the lights appropriately for the time of day"
-
-### Flow
-
-```
-1. Steward Analysis
-   → Capabilities needed: tatlock_core (datetime), home_automation
-   → Complexity: moderate
-
-2. Tatlock Execution - Step 1
-   <think>Checking current time to determine appropriate lighting...</think>
-   → Calls: get_current_datetime()
-   → Receives: "2024-12-11 21:45"
-
-3. Tatlock Evaluation
-   <think>It's evening (21:45). Will set warm, dimmed lighting for wind-down time...</think>
-
-4. Tatlock Execution - Step 2
-   → Calls: home_automation_agent(task)
-   → Home Automation executes: set_scene("evening_relaxation")
-   → Receives: "Scene 'evening_relaxation' activated - warm white 40%"
-
-5. Response
-   "It's quarter to ten in the evening, sir. I've set the lights to a warm,
-   subdued glow - ideal for winding down. Shall I also draw the curtains?"
-```
-
-### Intra-System Prompts
-
-**Tatlock → Home Automation Agent**:
-```
-Task: Set lighting appropriate for current time of day.
-Context: Current time is 21:45 (evening). User wants lights adjusted automatically.
-Action required: Activate appropriate lighting scene for evening/night.
-
-Reference (user's original request): "Set the lights appropriately for the time of day"
-```
-
----
-
-## Scenario 4: Create a Reminder (Simple Expert Delegation)
-
-**User**: "Remind me to call the dentist tomorrow at 2pm"
-
-### Flow
-
-```
-1. Steward Analysis
-   → Capabilities needed: personal_assistant (reminders)
-   → Complexity: simple
-
-2. Tatlock Execution
-   <think>Creating reminder for tomorrow afternoon...</think>
-   → Calls: personal_assistant_agent(task)
-   → PA executes: create_reminder(title, datetime, notification_time)
-   → Receives: "Reminder created: ID 847"
-
-3. Response
-   "Very good, sir. I've set a reminder to call the dentist for tomorrow at
-   2 o'clock in the afternoon. I'll notify you 15 minutes beforehand."
-```
-
-### Intra-System Prompts
-
-**Tatlock → Personal Assistant Agent**:
-```
-Task: Create a reminder for the user.
-Details:
-- Subject: Call the dentist
-- When: Tomorrow at 14:00
-- Notification: 15 minutes before (default)
-
-Reference (user's original request): "Remind me to call the dentist tomorrow at 2pm"
-```
-
----
-
-## Scenario 5: Schedule Recurring Automation (Complex, Multi-Expert)
-
-**User**: "Schedule the lights to turn on at 09:00 every weekend day"
-
-### Flow
-
-```
-1. Steward Analysis
-   → Capabilities needed: home_automation (scheduling)
-   → Complexity: moderate
-   → Note: Recurring schedule, weekend-only
-
-2. Tatlock Execution
-   <think>Setting up recurring weekend morning lighting schedule...</think>
-   → Calls: home_automation_agent(task)
-   → HA executes: create_schedule(trigger, days, action)
-   → Receives: "Schedule created: ID wake_light_weekend"
-
-3. Response
-   "The weekend wake-up lighting is now scheduled, sir. Every Saturday and
-   Sunday at 9 o'clock sharp, the lights will illuminate. Would you prefer
-   a gradual sunrise simulation, or an immediate full brightness?"
-```
-
-### Intra-System Prompts
-
-**Tatlock → Home Automation Agent**:
-```
-Task: Create a recurring lighting schedule.
-Details:
-- Action: Turn on lights
-- Time: 09:00
-- Days: Saturday, Sunday (weekends only)
-- Recurrence: Weekly
-
-Reference (user's original request): "Schedule the lights to turn on at 09:00 every weekend day"
-```
-
----
-
-## Scenario 6: Create Wiki Page About Topic (Expert with Research)
-
-**User**: "Create a wiki page about CI/CD"
-
-### Flow
-
-```
-1. Steward Analysis
-   → Capabilities needed: librarian
-   → Complexity: complex
-   → Specific action: smart_create (research-backed page creation)
-
-2. Tatlock Execution
-   <think>Delegating to The Librarian for research and wiki page creation...</think>
-   → Calls: librarian_agent(task)
-
-3. Librarian Internal Flow
-   → hybrid_search("CI/CD continuous integration deployment")
-   → Synthesizes content from wiki, graph, and web sources
-   → create_wiki_page(title, path, content, tags)
-   → Returns summary to Tatlock
-
-4. Tatlock Receives
-   "Page created: CI/CD
-    Path: /technology/cicd
-    Sources used: 12 (3 wiki, 4 graph, 5 web)
-    Forward links: 8, Backward links: 3"
-
-5. Response
-   "The wiki page on CI/CD has been created, sir. The Librarian has compiled
-   information from 12 sources, including our existing documentation on DevOps
-   practices. The page is now available at /technology/cicd and has been
-   cross-linked with related topics like Jenkins, GitHub Actions, and Docker."
-```
-
-### Intra-System Prompts
-
-**Tatlock → Librarian Agent**:
-```
-Task: Create a comprehensive wiki page about CI/CD.
-Context: User wants a new wiki page created. Use research capabilities to
-gather information and create a well-sourced page.
-
-Action required:
-1. Research the topic using hybrid search (wiki, knowledge graph, web)
-2. Create wiki page with synthesized content
-3. Apply appropriate tags (suggest: technology, devops, automation)
-4. Report back with page details and research summary
-
-Reference (user's original request): "Create a wiki page about CI/CD"
-```
-
-**Librarian → Tatlock Response**:
-```
-Task completed: Wiki page created
-
-Page Details:
-- Title: CI/CD (Continuous Integration/Continuous Deployment)
-- Path: /technology/cicd
-- Tags: technology, devops, automation, software-development
-
-Research Summary:
-- Wiki sources: 3 pages (DevOps Overview, Jenkins Setup, GitHub Actions Guide)
-- Knowledge graph entities: 4 (linked to Docker, Kubernetes, Git, Testing)
-- Web sources: 5 articles (current best practices, 2024 trends)
-- Total sources synthesized: 12
-
-Knowledge Graph Updates:
-- Forward links created: 8 (to existing related pages)
-- Backward links created: 3 (updated existing pages to reference new page)
-
-Content includes sections on:
-- Definition and core concepts
-- CI vs CD distinction
-- Popular tools comparison
-- Best practices
-- Integration with our documented infrastructure
-```
-
----
-
-## Scenario 7: Update Stock Wiki with Evaluation Tracker (Complex, Multi-Tool Expert)
-
-**User**: "Update the GOOG stock wiki page with a rolling evaluation tracker"
-
-### Flow
-
-```
-1. Steward Analysis
-   → Capabilities needed: librarian
-   → Complexity: complex
-   → Specific action: search existing page, fetch current data, update page
-
-2. Tatlock Execution
-   <think>This requires finding the existing page, fetching current stock data,
-   and updating with a new evaluation section. Delegating to The Librarian...</think>
-   → Calls: librarian_agent(task)
-
-3. Librarian Internal Flow - Step 1
-   <think>First, locating the existing GOOG wiki page...</think>
-   → search_wiki("GOOG stock Alphabet")
-   → Receives: Page ID 234, path /investments/stocks/goog
-
-4. Librarian Internal Flow - Step 2
-   <think>Fetching current page content to preserve existing information...</think>
-   → get_wiki_page(234)
-   → Receives: Current page content
-
-5. Librarian Internal Flow - Step 3
-   <think>Researching current stock data and evaluation metrics...</think>
-   → hybrid_search("GOOG Alphabet stock price evaluation 2024")
-   → Receives: Current price, P/E ratio, analyst ratings, etc.
-
-6. Librarian Internal Flow - Step 4
-   <think>Updating page with new rolling evaluation section...</think>
-   → update_wiki_page(234, content=updated_content)
-   → Receives: Update confirmation
-
-7. Tatlock Receives
-   "Page updated: GOOG Stock Analysis
-    Added section: Rolling Evaluation Tracker
-    Data points added: Price history, P/E trend, Analyst consensus
-    Last updated: 2024-12-11"
-
-8. Response
-   "The GOOG stock page has been updated, sir. I've added a rolling evaluation
-   tracker with current metrics: the stock is trading at $178.32 with a P/E
-   of 24.8, and analyst consensus remains 'Buy'. The tracker includes a
-   90-day price trend and quarterly earnings history. Shall I set up
-   automatic weekly updates?"
-```
-
-### Intra-System Prompts
-
-**Tatlock → Librarian Agent**:
-```
-Task: Update the GOOG (Alphabet) stock wiki page with a rolling evaluation tracker.
-Context: User wants to add ongoing stock evaluation tracking to an existing page.
-
-Actions required:
-1. Find the existing GOOG stock wiki page
-2. Read current page content (preserve existing information)
-3. Research current stock data and evaluation metrics
-4. Update the page with a new "Rolling Evaluation Tracker" section including:
-   - Current price and change
-   - Key ratios (P/E, P/B, etc.)
-   - Analyst consensus
-   - Price trend (30/60/90 day)
-   - Recent earnings summary
-5. Report back with update summary
-
-Reference (user's original request): "Update the GOOG stock wiki page with a rolling evaluation tracker"
-```
-
-**Librarian → Tatlock Response**:
-```
-Task completed: Wiki page updated
-
-Page Details:
-- Title: GOOG - Alphabet Inc. Stock Analysis
-- Path: /investments/stocks/goog
-- Page ID: 234
-
-Update Summary:
-- New section added: "Rolling Evaluation Tracker"
-- Existing content: Preserved (company overview, investment thesis)
-
-Evaluation Data Added:
-- Current Price: $178.32 (+1.2% today)
-- P/E Ratio: 24.8 (vs sector avg 28.3)
-- P/B Ratio: 6.2
-- 52-Week Range: $120.21 - $191.75
-- Analyst Consensus: Buy (24 buy, 8 hold, 2 sell)
-- Price Trend: +15.3% (90-day)
-- Last Earnings: Q3 2024 - Beat estimates by 8%
-
-Sources:
-- Web: 4 financial data sources
-- Existing wiki: Investment strategy context
-
-Suggested follow-up: Set up automated weekly data refresh?
-```
-
----
-
-## Scenario 8: Complex Multi-Expert Coordination
-
-**User**: "Check if there's anything about the new project deadline in my notes,
-if so create a reminder, and if the deadline is this week, also update the
-project wiki page with a status note"
-
-### Flow
-
-```
-1. Steward Analysis
-   → Capabilities needed: librarian, personal_assistant
-   → Complexity: complex
-   → Note: Conditional multi-step, requires information passing between experts
-
-2. Tatlock Execution - Step 1
-   <think>First, searching for project deadline information in the wiki...</think>
-   → Calls: librarian_agent(search_task)
-   → Librarian: hybrid_search("new project deadline")
-   → Returns: "Project Alpha deadline: December 15, 2024 (this Friday)"
-
-3. Tatlock Evaluation
-   <think>Found deadline: December 15. That's this week (Friday).
-   Need to: 1) Create reminder, 2) Update project wiki page...</think>
-
-4. Tatlock Execution - Step 2 (parallel if possible)
-   <think>Creating reminder and updating wiki status...</think>
-
-   → Calls: personal_assistant_agent(reminder_task)
-   → PA: create_reminder("Project Alpha deadline", "2024-12-15 09:00")
-   → Returns: "Reminder created for Dec 15 at 9am"
-
-   → Calls: librarian_agent(update_task)
-   → Librarian: search_wiki → get_wiki_page → update_wiki_page
-   → Returns: "Project Alpha page updated with deadline status note"
-
-5. Response
-   "I've found the deadline in your notes, sir - Project Alpha is due this
-   Friday, December 15th. I've set a reminder for 9 o'clock that morning,
-   and I've updated the project wiki page with a status note indicating
-   the imminent deadline. Is there anything else you need to prepare?"
-```
-
-### Intra-System Prompts
-
-**Tatlock → Librarian Agent (Search)**:
-```
-Task: Search for information about a new project deadline.
-Context: User wants to find deadline information from their notes/wiki.
-
-Action required:
-1. Search wiki and knowledge base for project deadline information
-2. Return: Project name, deadline date, and any relevant context
-
-Reference (user's original request): "Check if there's anything about the new project deadline in my notes..."
-```
-
-**Tatlock → Personal Assistant Agent**:
-```
-Task: Create a reminder for a project deadline.
-Details:
-- Subject: Project Alpha deadline
-- When: December 15, 2024 at 09:00
-- Priority: High (deadline is this week)
-- Notification: Morning of the deadline
-
-Reference: Creating reminder based on deadline found in user's notes.
-```
-
-**Tatlock → Librarian Agent (Update)**:
-```
-Task: Update the Project Alpha wiki page with a deadline status note.
-Context: Project deadline is December 15, 2024 (this Friday). User requested
-a status update since the deadline is this week.
-
-Action required:
-1. Find the Project Alpha wiki page
-2. Add a status note/banner indicating the imminent deadline
-3. Optionally update any status fields
-
-Reference: Part of user's request to track and highlight near-term deadlines.
-```
-
----
-
-## Response Pattern Guidelines
-
-### Tatlock's Think Updates (Streaming to User)
-
-During multi-step operations, Tatlock should emit `<think>` updates to keep the user informed:
-
-```
-<think>Analyzing your request...</think>
-<think>Searching for deadline information in the wiki...</think>
-<think>Found the deadline - December 15th. Creating reminder...</think>
-<think>Updating the project page with status note...</think>
-<think>All tasks complete. Composing response...</think>
-```
-
-### Tatlock's Final Response Pattern
-
-1. **Acknowledge** - Confirm understanding of the request
-2. **Summarize actions** - What was done, by whom (implicitly)
-3. **Key details** - Important information the user should know
-4. **Proactive offer** - Suggest related actions or follow-ups
-5. **Butler voice** - Formal but warm, with personality
-
-### Expert Agent Response Pattern
-
-1. **Task status** - Completed/Partial/Failed
-2. **Action summary** - What was done
-3. **Key data** - Information Tatlock needs to synthesize
-4. **Metadata** - IDs, counts, timestamps for reference
-5. **Suggestions** - Optional follow-up actions
-
----
-
-## Error Handling Scenarios
-
-### Expert Agent Failure
-
-```
-Tatlock → Librarian: "Create wiki page about quantum computing"
-Librarian → Tatlock: "Error: library-desk API unavailable (connection timeout)"
-
-Tatlock Response:
-"I'm afraid The Librarian is having some difficulty reaching the wiki
-service at the moment, sir. I can attempt a basic web search on quantum
-computing if you'd like, or we can try the wiki operation again in a
-few minutes."
-```
-
-### Partial Completion
-
-```
-User: "Create a reminder and add it to my calendar"
-
-Tatlock → PA: Create reminder
-PA → Tatlock: "Reminder created successfully"
-
-Tatlock → Calendar: Add to calendar
-Calendar → Tatlock: "Error: Calendar sync not configured"
-
-Tatlock Response:
-"I've created the reminder, sir, but I wasn't able to add it to your
-calendar - it appears the calendar integration needs to be configured.
-The reminder will still alert you at the scheduled time. Shall I help
-set up the calendar connection?"
-```
-
----
-
-## Summary: Key Design Principles
-
-1. **Tatlock is the orchestrator** - Never exposes raw tool complexity to users
-2. **Expert agents are tools** - Tatlock calls them, they return structured responses
-3. **Context flows down** - Each expert gets only what they need to complete their task
-4. **Results flow up** - Tatlock synthesizes all responses into coherent butler-voice answer
-5. **Think updates maintain engagement** - User sees progress during complex operations
-6. **Errors are handled gracefully** - Tatlock explains and offers alternatives
-7. **Proactive suggestions** - Tatlock anticipates follow-up needs

README.md

@@ -1,6 +1,6 @@
 # Tatlock - Your Homelab Butler
 
-> **📖 For the complete system vision and architectural philosophy, see [PHILOSOPHY.md](PHILOSOPHY.md)**
+> **📖 For the complete system vision and architectural philosophy, see [docs/philosophy.md](docs/philosophy.md)**
 
 A privacy-first, offline-capable personal assistant system that coordinates specialized AI agents to help with research, development, home automation, and daily organization.
 
@@ -58,7 +58,7 @@ A privacy-first, offline-capable personal assistant system that coordinates spec
   - Error triggers for testing (rate_limit, context_overflow)
 
 - **Tatlock**: Real PydanticAI agent with butler personality
-  - **LLM Backend**: Ollama (mistral-nemo:latest by default)
+  - **LLM Backend**: Ollama (gemma4:e2b by default, local-first) with optional Claude fallback
   - **Personality**: Witty British butler, research-oriented
   - **Core Tools**:
     - **Calculator**: Safe mathematical expression evaluation
@@ -74,7 +74,7 @@ A privacy-first, offline-capable personal assistant system that coordinates spec
 
 - Python 3.12+ (Python 3.12.11 recommended)
 - **External Services** (must be running separately):
-  - **Ollama**: LLM inference (mistral-nemo:latest, nomic-embed-text)
+  - **Ollama**: LLM inference (gemma4:e2b, nomic-embed-text)
   - **Redis**: Caching and session memory
   - **Qdrant**: Vector storage for The Biographer's memory
   - **SearXNG**: Web search (optional)
@@ -89,12 +89,8 @@ A privacy-first, offline-capable personal assistant system that coordinates spec
 git clone https://git.schweitz.net/jpmschweitzer/tatlock.git
 cd tatlock
 
-# Create virtual environment
-python -m venv .venv
-source .venv/bin/activate  # Windows: .venv\Scripts\activate
-
 # Install dependencies
-pip install -r requirements.txt
+make setup
 ```
 
 ### Run the Server
@@ -268,7 +264,10 @@ Interactive documentation available at:
 pytest
 
 # Run unit tests only (no external services needed)
-pytest --ignore=tests/e2e --ignore=tests/integration
+pytest --ignore=tests/e2e --ignore=tests/integration --ignore=tests/contracts
+
+# Wire-level contract tests against live service boundaries
+make test-contracts
 
 # Run with coverage
 pytest --cov=src --cov-report=term-missing
@@ -307,12 +306,17 @@ Create a `.env` file for custom configuration:
 API_HOST=0.0.0.0
 API_PORT=8000
 
-# Ollama Configuration
+# Ollama Configuration (primary backend)
 OLLAMA_HOST=http://localhost:11434
-OLLAMA_DEFAULT_MODEL=mistral-nemo:latest
+OLLAMA_DEFAULT_MODEL=gemma4:e2b
 OLLAMA_EMBEDDING_MODEL=nomic-embed-text
 OLLAMA_TIMEOUT=120
 
+# Claude fallback (optional; used when Ollama is down or PREFER_CLOUD_BACKEND=true)
+# ANTHROPIC_API_KEY=sk-ant-api03-your-key-here
+ANTHROPIC_MODEL=claude-sonnet-5
+PREFER_CLOUD_BACKEND=false
+
 # Redis Configuration
 REDIS_HOST=localhost
 REDIS_PORT=6379
@@ -380,9 +384,8 @@ tatlock/
 │   │   ├── steward/         # The Steward - request analysis
 │   │   ├── tatlock_core/    # Core butler tools
 │   │   ├── tatlock.py       # Tatlock PydanticAI agent
-│   │   ├── coordination.py  # Multi-agent coordination
 │   │   ├── delegation.py    # Expert delegation wrappers
-│   │   └── protocol.py      # Agent communication protocol
+│   │   └── protocol.py      # Agent error protocol
 │   ├── responses/           # Responses API (primary endpoint)
 │   ├── chat/                # Chat Completions wrapper
 │   ├── models/              # Models listing
@@ -396,15 +399,14 @@ tatlock/
 │   │   └── multi_tenancy.py # User isolation utilities
 │   └── main.py              # Application entry point
 ├── tests/                   # Comprehensive test suite
-├── PHILOSOPHY.md            # System vision and architecture
-├── IMPLEMENTATION_ROADMAP.md # Development phases
+├── docs/                    # Project documentation
 ├── CHANGELOG.md             # Version history
 └── README.md                # This file
 ```
 
 ## Development
 
-For LLM agent development guidelines and architectural decisions, see [AGENTS.md](AGENTS.md).
+For LLM agent development guidelines and architectural decisions, see [CLAUDE.md](CLAUDE.md).
 
 ## Contributing
 
@@ -416,9 +418,9 @@ For LLM agent development guidelines and architectural decisions, see [AGENTS.md
 
 ## Documentation
 
-- **System Philosophy**: [PHILOSOPHY.md](PHILOSOPHY.md) - Vision, goals, and architectural patterns
-- **User Guide**: This file - Installation, usage, and examples
-- **Developer Guidelines**: [AGENTS.md](AGENTS.md) - LLM agent development patterns
+- **System Philosophy**: [docs/philosophy.md](docs/philosophy.md) - Vision, goals, and architectural patterns
+- **Development Roadmap**: [docs/roadmap.md](docs/roadmap.md) - Open work and planned phases
+- **Developer Guidelines**: [CLAUDE.md](CLAUDE.md) - LLM agent development patterns
 - **Version History**: [CHANGELOG.md](CHANGELOG.md) - Changes and releases
 
 ### External References
@@ -432,8 +434,8 @@ For LLM agent development guidelines and architectural decisions, see [AGENTS.md
 
 ## Version
 
-Current version: **1.2.0** - Phase F: Memory System (The Biographer)
+Current version: see [CHANGELOG.md](CHANGELOG.md)
 
 ---
 
-**Note**: Tatlock is a production-ready homelab butler. All household staff use PydanticAI with Ollama for local LLM inference.
+**Note**: Tatlock is a production-ready homelab butler. All household staff use PydanticAI with local Ollama inference (gemma4), with an optional Claude cloud fallback.

ci/secrets.sh

@@ -0,0 +1,50 @@
+#!/usr/bin/env bash
+# Secret scan over the commits about to be pushed.
+#
+# Lives here rather than inside .githooks/pre-push so it can be read, run by
+# hand (`make secrets`), and changed under review. A hook is a trigger; it is
+# not a home for logic. Identical in every repo in this workspace (D-27).
+set -euo pipefail
+
+cd "$(git rev-parse --show-toplevel)"
+
+# A non-login shell — which is what git gives a hook — skips /etc/profile.d
+# and never sees ~/.local/bin, where the gitleaks release tarball lands.
+# Without this the scan reports "not installed" on every push.
+[ -d "$HOME/.local/bin" ] && PATH="$HOME/.local/bin:$PATH"
+
+if ! command -v gitleaks >/dev/null 2>&1; then
+  echo "FAIL secrets — gitleaks not installed, so this check would be a no-op pretending to pass." >&2
+  echo "     https://github.com/gitleaks/gitleaks/releases → ~/.local/bin/gitleaks" >&2
+  exit 1
+fi
+
+# Scan the outgoing range, not full history. History here carries findings
+# that are settled — test fixtures and vendored third-party code — and a gate
+# that fails on something unfixable gets bypassed within a week. What matters
+# is what is about to leave this machine.
+if upstream=$(git rev-parse --abbrev-ref --symbolic-full-name '@{u}' 2>/dev/null); then
+  range="$upstream..HEAD"
+elif git rev-parse --verify --quiet origin/main >/dev/null; then
+  range="origin/main..HEAD"
+else
+  range=""
+fi
+
+if [ -z "$range" ]; then
+  gitleaks dir . --redact --no-banner --exit-code 1 || {
+    echo "FAIL secrets — gitleaks found a credential in the working tree." >&2; exit 1; }
+  exit 0
+fi
+
+[ -n "$(git log --oneline "$range" 2>/dev/null)" ] || exit 0
+
+gitleaks git . --log-opts="$range" --redact --no-banner --exit-code 1 >/dev/null 2>&1 || {
+  echo "FAIL secrets — gitleaks found a credential in the commits being pushed." >&2
+  echo "     inspect (values redacted):  gitleaks git . --log-opts=\"$range\" --redact" >&2
+  echo "     then remove and rotate it, or suppress deliberately:" >&2
+  echo "       inline  '# gitleaks:allow  <reason>'" >&2
+  echo "       or add the fingerprint to .gitleaksignore WITH a reason" >&2
+  exit 1
+}
+echo "  ok   secrets"

docs/claude-integration.md

@@ -0,0 +1,100 @@
+# Claude Integration Plan
+
+## Overview
+
+Tatlock uses a bidirectional Claude architecture:
+- **Scenario A**: Tatlock powered by Claude backend (with Ollama fallback) — **COMPLETE**, then **rolled back to local-first**: Ollama/gemma4 is primary, Claude is retained as fallback (`PREFER_CLOUD_BACKEND=false`)
+- **Scenario B**: Tatlock exposed as MCP server for external Claude instances — **OPEN**
+- **Scenario C**: Offline operation via Ollama — **COMPLETE**
+
+---
+
+## MCP Server (Expose Tools to Claude) — NOT STARTED
+
+Create an MCP server that exposes Tatlock's household tools to external Claude instances.
+
+### New Files
+
+```
+src/mcp/
+├── __init__.py
+├── server.py         # MCP server using mcp Python SDK
+├── tool_adapters.py  # Convert PydanticAI tools → MCP schemas
+├── auth.py           # API key authentication
+└── transport.py      # Streamable HTTP transport
+```
+
+### Docker Stack Addition
+
+```yaml
+tatlock-mcp:
+  image: git.schweitz.net/jpmschweitzer/tatlock:latest
+  command: ["python", "-m", "src.mcp.server"]
+  ports:
+    - "8002:8002"
+  environment:
+    - MCP_AUTH_TOKEN=${MCP_AUTH_TOKEN}
+  networks:
+    - docker-dataplane
+```
+
+### Claude Desktop Configuration
+
+```json
+{
+  "mcpServers": {
+    "tatlock": {
+      "command": "npx",
+      "args": ["mcp-remote", "https://mcp.schweitz.net/sse", "--header", "Authorization: Bearer ${MCP_AUTH_TOKEN}"]
+    }
+  }
+}
+```
+
+### Checklist
+
+- [ ] Create `src/mcp/` module
+- [ ] Tool adapters (PydanticAI → MCP schema)
+- [ ] Authentication middleware
+- [ ] Streamable HTTP transport
+- [ ] Docker stack configuration
+
+---
+
+## Future Phases
+
+- **LiteLLM Gateway** — Unified endpoint for all models, config-driven routing
+- **Multi-Provider** — Add OpenAI, Vertex AI, etc.
+- **Smart Routing** — Context-aware model selection, cost ceiling enforcement
+
+---
+
+## Offline Behavior
+
+| Scenario | Behavior |
+|----------|----------|
+| No API key | Use Ollama exclusively |
+| API unreachable | Use Ollama, log warning |
+| API rate limited | Fallback to Ollama |
+
+| Aspect | Claude | Ollama |
+|--------|--------|--------|
+| Context | 200k tokens | ~8k tokens |
+| Latency | 1-3s (network) | 0.5-1s (local) |
+| Personality | Preserved | Preserved |
+| Tools | All work | All work |
+| Cost | API charges | Free |
+
+---
+
+## Related Repo Handovers
+
+Handover documents created in each repo: `PROJECT_CLAUDIFICATION_HANDOVER.md`
+
+### Open Items
+
+- **library-desk**: Review HybridRAG response size limits, smart_create endpoint, response formats
+- **core-api**: Review list_devices response format, error messages, rate limiting
+- **portainer-core**: Update stack with new env vars, configure secrets, update CONTAINERS.md
+- **webber**: Review content truncation limits, extraction quality
+- **tatlock-ui**: Test streaming with Claude backend, conversation history, tool call display

docs/housekeeper-optimization-findings.md

@@ -0,0 +1,246 @@
+# Housekeeper Agent Optimization Findings
+
+## Background
+
+Research with Gemini identified key issues with mistral-nemo and tool calling:
+- "Pre-computation Hallucination" - model answers before using tools
+- High default temperature (0.7-0.8) causes wandering
+- Model is "chatty and confident" - needs explicit constraints
+
+## Key Recommendations from Gemini Research
+
+1. **Temperature 0.0** for tool-calling agents (deterministic, follows schema)
+2. **Chain of Thought (CoT)** - force step-by-step reasoning
+3. **Negative constraints** - tell model what NOT to do (Nemo responds better)
+4. **Explicit tool descriptions** - verbose docstrings with "never estimate yourself"
+5. **"Strictly tool-based assistant"** pattern - NO internal knowledge claim
+
+---
+
+## Experiment Log
+
+### Baseline (v1.8.6)
+- **Date**: 2025-12-17
+- **Configuration**: Default temperature, improved prompt requiring list_devices first
+- **Results**:
+  - Called list_devices first ✓
+  - Still hallucinated `light.study_desk` despite seeing list with only `light.study` and `light.study_main`
+  - Partial success: turned off `light.study_main`, failed on hallucinated entity
+- **Success rate**: ~50% (1 of 2 study lights controlled correctly)
+
+---
+
+### Experiment 1: Temperature 0.0
+- **Date**: 2025-12-18
+- **Change**: Set `model_settings=ModelSettings(temperature=0.0)` for Housekeeper
+- **Hypothesis**: Deterministic output will force model to use exact entity IDs from tool results
+- **Results**:
+
+  **Study lights test:**
+  - Called `list_devices()` first ✓ (but no domain filter)
+  - Used wrong parameter `device_id` instead of `entity_id` (recovered after validation error)
+  - Only identified `light.studeerlamp` as "study" related (Dutch name)
+  - **Missed `light.study` and `light.study_main`** - didn't match English "study"
+  - Turned off 1 wrong light, missed 2 actual study lights
+
+  **Kitchen lights test:**
+  - Called `list_devices()` first ✓ (no domain filter)
+  - Saw full device list including `light.kitchen`
+  - Used wrong parameter `device_id` instead of `entity_id` (recovered after validation)
+  - After correction, dropped domain prefix: used `kitchen` instead of `light.kitchen`
+  - 404 error - device not found
+
+- **Success rate**: 0% (no target lights successfully controlled)
+- **Observations**:
+  - Temperature 0.0 alone is insufficient
+  - Model consistently confuses `device_id` vs `entity_id` parameter name
+  - After validation error correction, model truncates entity_id (drops domain prefix)
+  - Semantic matching of room names to devices is weak
+  - Model doesn't understand entity_id format: `domain.name`
+
+---
+
+### Experiment 2: Negative Constraints + CoT
+- **Date**: 2025-12-18
+- **Change**: Complete prompt rewrite with:
+  - "You have NO Internal Knowledge" - negative framing
+  - Explicit entity_id format with WRONG/RIGHT examples
+  - Step-by-step process (ALWAYS FOLLOW)
+  - Explicit parameter names section
+  - "What NOT To Do" negative constraints
+- **Hypothesis**: Negative constraints work better with Mistral-Nemo
+- **Results**:
+
+  **Study lights test:**
+  - Called `list_devices(domain="light")` ✓ with domain filter (improvement!)
+  - Still used `device_id` first, recovered to `entity_id` after validation error
+  - After recovery, used correct full format: `light.studeerlamp`
+  - **Still only matched `studeerlamp` not `light.study` or `light.study_main`**
+
+  **Kitchen lights test:**
+  - Called `list_devices(domain="light")` ✓
+  - Called `turn_off(entity_id="light.kitchen")` ✓ correct format!
+  - All 4 kitchen lights turned off (light.kitchen is a group)
+  - **100% success for kitchen!**
+
+- **Success rate**:
+  - Study: 0% (wrong semantic match)
+  - Kitchen: 100% (4/4 lights off)
+  - Combined: ~50% (1 of 2 tests successful)
+- **Observations**:
+  - Domain filter now consistently used ✓
+  - Entity_id format correct after recovery ✓
+  - Semantic matching still fails for "study" → prefers Dutch "studeerlamp" over English "study"
+  - Parameter name confusion persists (`device_id` vs `entity_id`)
+  - Simple room names (kitchen) work; mixed language fails (study/studeerlamp)
+
+---
+
+### Experiment 3: Temperature 0.1 + Explicit Tool Docstrings
+- **Date**: 2025-12-18
+- **Change**:
+  - Temperature 0.1
+  - Updated turn_on/turn_off docstrings with explicit `entity_id=` in examples
+- **Results**:
+  - Still uses `device_id` first, recovers to `entity_id` after validation
+  - Still picks wrong entity (studeerlamp over study)
+- **Success rate**: 0%
+
+---
+
+### Experiment 4: Room Group Priority (with explicit examples)
+- **Date**: 2025-12-18
+- **Change**: Updated prompt with:
+  - Explicit instruction: "Look for EXACT match `light.<room_name>` first!"
+  - Concrete examples: "For 'study lights' → look for `light.study`"
+  - Working example showing `turn_off(entity_id="light.study")`
+- **Hypothesis**: Explicit examples will guide model to use room groups
+- **Results**:
+
+  **Test 1 & 2 (consecutive):**
+  - Called `list_devices(domain="light")` ✓
+  - Device list clearly shows `light.study` at the bottom
+  - First call: `turn_off({"devices":["studeerlamp"]})` - wrong param AND wrong device
+  - After validation error: `turn_off(entity_id="light.studeerlamp")` - correct param, still wrong device
+  - **Completely ignored `light.study` despite prompt explicitly saying to use it**
+
+- **Success rate**: 0% (wrong device controlled)
+- **Observations**:
+  - Model ignores explicit step-by-step instructions in favor of substring matching
+  - Dutch "studeerlamp" contains "studer" which the model prefers over exact "study" match
+  - Even when prompt has a literal example `turn_off(entity_id="light.study")`, model uses `light.studeerlamp`
+  - Positional bias possible - `light.study` appears at end of 21-item list
+  - **Fundamental limitation**: Mistral-Nemo cannot follow explicit matching rules
+
+---
+
+### Experiment 5: Room Groups First (Tool Output Ordering)
+- **Date**: 2025-12-18
+- **Change**: Modified `list_devices` to sort room groups to top of list using HA attributes (`is_hue_group`, `hue_type="room"`)
+- **Hypothesis**: Positional bias - model focuses on items earlier in list
+- **Results**:
+  - Room groups (`light.study`, `light.kitchen`, etc.) now appear first in device list
+  - Combined with improved prompt, model now consistently uses room groups
+  - **70% success rate** (7/10 tests) with default q4 quantization
+
+---
+
+### Experiment 6: Model Quantization (q5_1)
+- **Date**: 2025-12-18
+- **Change**: Upgraded from default Mistral-Nemo quantization (q4) to `mistral-nemo:12b-instruct-2407-q5_1`
+- **Hypothesis**: Higher precision weights improve tool calling accuracy
+- **Results**:
+
+  | Test | Action | Result |
+  |------|--------|--------|
+  | 1 | Turn off study | PASS |
+  | 2 | Turn on study | PASS |
+  | 3 | Toggle study | PASS |
+  | 4 | Turn off kitchen | PASS |
+  | 5 | Turn on kitchen | PASS |
+  | 6 | Toggle kitchen | PASS |
+  | 7 | Turn off bedroom | PASS |
+  | 8 | Turn on bedroom | PASS |
+  | 9 | Turn off living room | PASS |
+  | 10 | Turn on living room | PASS |
+
+- **Success rate**: **100%** (10/10 tests)
+- **Observations**:
+  - q5_1 quantization dramatically improves tool calling accuracy
+  - All room groups correctly identified and used
+  - No parameter confusion (`entity_id` used correctly)
+  - No entity_id truncation issues
+  - Toggle operations now work reliably
+  - Model fits within 10GB VRAM (q6 did not)
+
+---
+
+### Experiment 7: Device List in System Prompt (Context Injection)
+- **Date**: [PENDING]
+- **Change**: Store device list in database (per user/household) and inject into system prompt
+- **Approach**:
+  1. Periodically sync device list from Home Assistant to PostgreSQL
+  2. On each Housekeeper invocation, fetch device list and include in prompt
+  3. Remove need for model to call list_devices() - just match from context
+- **Hypothesis**:
+  - Eliminates tool call step where errors occur
+  - Reduces context size by not returning full device list as tool output
+  - Makes entity matching a language task (in prompt) rather than tool result parsing
+- **Trade-offs**:
+  - Stale data if sync is infrequent
+  - Prompt size increase (but less than tool call response)
+  - Need sync mechanism and storage
+- **Results**: [TO BE RECORDED]
+- **Success rate**: [TO BE RECORDED]
+
+---
+
+## Key Problem Identified (Solved)
+
+The model struggled with:
+1. **Parameter schema adherence** - uses `device_id` when schema requires `entity_id`
+2. **Value preservation** - truncates values after validation errors (drops `light.` prefix)
+3. **Semantic matching** - prefers substring matches ("studeerlamp" contains "studer") over exact matches (`light.study`)
+4. **Following explicit instructions** - ignores step-by-step processes even when examples are provided
+5. **Positional bias** - may not "see" items at the end of long lists
+
+**Solution**: These issues were resolved by:
+1. Using q5_1 quantization instead of default q4 (higher precision weights)
+2. Sorting room groups to top of device list (address positional bias)
+3. Explicit prompt guidance with negative constraints and examples
+
+---
+
+## Potential Next Experiments
+
+### Experiment 5: Room Groups First (List Ordering)
+- **Hypothesis**: Positional bias - model focuses on items earlier in list
+- **Change**: Sort device list to put room groups (entities matching `light.<single_word>`) at the TOP
+- **Effort**: Low - modify list_devices output formatting
+- **Risk**: May affect other use cases where individual devices are needed
+
+### Experiment 6: Simplified Device List Format
+- **Hypothesis**: Markdown formatting adds noise that confuses the model
+- **Change**: Return simple list: `light.study (Study - GROUP), light.study_main (Ceiling light), ...`
+- **Effort**: Low - modify list_devices output
+- **Risk**: Less human-readable responses
+
+---
+
+## Learnings to Apply Elsewhere
+
+1. **Quantization matters** - q5_1 dramatically outperforms q4 for tool calling (100% vs 70%)
+2. **Positional bias is real** - sort important items to top of lists
+3. **Smaller models need simpler workflows** - fewer tool calls, more context injection
+4. **Validation errors don't teach** - model often makes worse mistakes on retry
+5. **Entity IDs are hard** - domain.name format confuses the model
+6. **Consider pre-computation** - move matching logic to code, not LLM
+7. **Use explicit negative constraints** - "NEVER do X" works better than "always do Y"
+
+---
+
+## Notes
+
+- Librarian may need higher temperature for creative synthesis
+- All "action" agents (Housekeeper, future agents) should use low temperature
+- Consider testing with Gemma 2 9B for better function calling (Google, open weights)

docs/philosophy.md

@@ -10,7 +10,7 @@ This document establishes the foundational philosophy and architectural patterns
 - When new architectural insights require rethinking core principles
 
 **When NOT to modify this document**:
-- During implementation of these patterns (use README.md, AGENTS.md, or code comments for technical details)
+- During implementation of these patterns (use README.md, CLAUDE.md, or code comments for technical details)
 - For adding new household members or capabilities within the existing pattern
 - For tactical decisions about specific technologies or tools
 
@@ -270,7 +270,7 @@ The user never directly interacts with the Steward or individual expert agents
 
 **Related Documents**:
 - **README.md**: User-facing documentation and usage guide
-- **AGENTS.md**: LLM agent development guidelines and technical patterns
+- **CLAUDE.md**: LLM agent development guidelines and technical patterns
 - **CHANGELOG.md**: Version history and implemented features
 
 ---

docs/rag_search_implementation.md

@@ -0,0 +1,348 @@
+# Tatlock Integration Guide
+
+Implementation instructions for integrating Library Desk search and content extraction endpoints into the Tatlock project.
+
+## Base Configuration
+
+```
+BASE_URL: http://library-desk:8089  (or your deployment URL)
+AUTH_HEADER: Authorization: Bearer <LIBRARY_API_KEY>
+```
+
+---
+
+## 1. RAG Search Endpoint
+
+**Use case:** Librarian needs to research a topic by searching the web.
+
+### Endpoint
+
+```
+POST /rag/search
+```
+
+### Request
+
+```json
+{
+  "query": "Python async programming best practices",
+  "search_type": "web",
+  "limit": 10,
+  "user": "tatlock-librarian"
+}
+```
+
+| Field | Type | Default | Description |
+|-------|------|---------|-------------|
+| `query` | string | required | Search query (1-500 chars) |
+| `search_type` | enum | `"web"` | `"web"`, `"news"`, or `"images"` |
+| `limit` | int | 10 | Results to return (1-20) |
+| `user` | string | `"default"` | User identifier for tracking |
+
+### Response
+
+```json
+{
+  "query": "Python async programming best practices",
+  "search_type": "web",
+  "results": [
+    {
+      "title": "Async IO in Python: A Complete Walkthrough",
+      "url": "https://realpython.com/async-io-python/",
+      "content": "Full extracted article text via Trafilatura (~2000 chars max)...",
+      "snippet": "Original search engine snippet (150-300 chars)...",
+      "source": "realpython.com",
+      "published_date": "2023-05-15"
+    }
+  ],
+  "total_results": 10,
+  "search_time_ms": 2340,
+  "sources_summary": "## Sources\n- [Async IO in Python](https://realpython.com/async-io-python/)\n- ..."
+}
+```
+
+### Key Fields for Tatlock
+
+| Field | Usage |
+|-------|-------|
+| `results[].content` | Full extracted text - use this for LLM context |
+| `results[].snippet` | Fallback if content extraction failed |
+| `sources_summary` | Pre-formatted markdown for citations |
+
+### Error Handling
+
+| HTTP Code | Meaning | Action |
+|-----------|---------|--------|
+| 400 | Invalid query | Check query length/format |
+| 502 | SearXNG unavailable | Retry with backoff |
+| 504 | Search timeout | Retry or reduce limit |
+| 500 | Internal error | Log and notify |
+
+### Example Usage (Python)
+
+```python
+import httpx
+
+async def search_web(query: str, limit: int = 10) -> dict:
+    async with httpx.AsyncClient() as client:
+        response = await client.post(
+            f"{BASE_URL}/rag/search",
+            headers={"Authorization": f"Bearer {API_KEY}"},
+            json={
+                "query": query,
+                "search_type": "web",
+                "limit": limit,
+                "user": "tatlock-librarian"
+            },
+            timeout=30.0
+        )
+        response.raise_for_status()
+        return response.json()
+
+# Usage
+results = await search_web("machine learning transformers")
+for r in results["results"]:
+    # Prefer full content, fall back to snippet
+    text = r["content"] or r["snippet"]
+    print(f"{r['title']}: {len(text)} chars")
+```
+
+---
+
+## 2. Content Extraction Endpoint
+
+**Use case:** Librarian has a specific URL and needs to read its content.
+
+### Single URL Extraction
+
+```
+POST /content/extract
+```
+
+#### Request
+
+```json
+{
+  "url": "https://example.com/article",
+  "include_metadata": true,
+  "max_length": 2000
+}
+```
+
+#### Response
+
+```json
+{
+  "result": {
+    "url": "https://example.com/article",
+    "title": "Article Title",
+    "content": "Extracted main text content...",
+    "author": "John Doe",
+    "date": "2024-01-15",
+    "language": "en",
+    "success": true,
+    "error": null
+  },
+  "extraction_time_ms": 1250
+}
+```
+
+### Batch URL Extraction
+
+```
+POST /content/extract/batch
+```
+
+#### Request
+
+```json
+{
+  "urls": [
+    "https://example.com/article1",
+    "https://example.com/article2",
+    "https://example.com/article3"
+  ],
+  "include_metadata": true,
+  "max_length": 2000
+}
+```
+
+#### Response
+
+```json
+{
+  "results": [
+    {
+      "url": "https://example.com/article1",
+      "title": "Article 1",
+      "content": "Extracted content...",
+      "success": true,
+      "error": null
+    },
+    {
+      "url": "https://example.com/article2",
+      "title": null,
+      "content": "",
+      "success": false,
+      "error": "Connection timeout"
+    }
+  ],
+  "total_urls": 3,
+  "successful": 2,
+  "failed": 1,
+  "extraction_time_ms": 3500
+}
+```
+
+---
+
+## 3. Error Pattern: Soft Failures
+
+> **Important:** Content extraction uses a **soft failure pattern** - individual URL failures do NOT throw HTTP errors.
+
+### Why Soft Failures?
+
+When extracting content from multiple URLs (batch) or even single URLs:
+- Some sites block bots
+- Some URLs are temporarily down
+- Some pages have no extractable content
+
+Instead of failing the entire request, we return:
+- `success: true/false` per result
+- `error: "reason"` when failed
+- Empty `content: ""` on failure
+
+### Handling Soft Failures
+
+```python
+async def extract_with_fallback(url: str) -> str:
+    response = await client.post(
+        f"{BASE_URL}/content/extract",
+        headers={"Authorization": f"Bearer {API_KEY}"},
+        json={"url": url}
+    )
+    response.raise_for_status()  # Only throws on 4xx/5xx
+
+    data = response.json()
+    result = data["result"]
+
+    if result["success"]:
+        return result["content"]
+    else:
+        # Log the failure, return empty or handle gracefully
+        logger.warning(f"Extraction failed for {url}: {result['error']}")
+        return ""  # Or raise, or use cached version, etc.
+```
+
+### Batch Processing Example
+
+```python
+async def extract_batch_with_stats(urls: list[str]) -> dict:
+    response = await client.post(
+        f"{BASE_URL}/content/extract/batch",
+        headers={"Authorization": f"Bearer {API_KEY}"},
+        json={"urls": urls, "max_length": 3000}
+    )
+    response.raise_for_status()
+
+    data = response.json()
+
+    # Separate successful and failed
+    successful = [r for r in data["results"] if r["success"]]
+    failed = [r for r in data["results"] if not r["success"]]
+
+    if failed:
+        logger.warning(f"{len(failed)} URLs failed extraction:")
+        for f in failed:
+            logger.warning(f"  {f['url']}: {f['error']}")
+
+    return {
+        "contents": {r["url"]: r["content"] for r in successful},
+        "failed_urls": [f["url"] for f in failed],
+        "success_rate": data["successful"] / data["total_urls"]
+    }
+```
+
+---
+
+## 4. Recommended Patterns for Tatlock
+
+### Research Flow
+
+```python
+async def librarian_research(topic: str) -> dict:
+    """
+    Full research flow: search + extract additional context.
+    """
+    # 1. Search for relevant pages
+    search_results = await search_web(topic, limit=10)
+
+    # 2. RAG search already includes extracted content
+    #    Only extract more if you need deeper content
+
+    # 3. Build context for LLM
+    context_parts = []
+    for r in search_results["results"]:
+        content = r["content"] or r["snippet"]
+        if content:
+            context_parts.append(f"## {r['title']}\nSource: {r['url']}\n\n{content}")
+
+    return {
+        "context": "\n\n---\n\n".join(context_parts),
+        "sources": search_results["sources_summary"],
+        "result_count": search_results["total_results"]
+    }
+```
+
+### Reading a Specific Page
+
+```python
+async def librarian_read_page(url: str) -> str:
+    """
+    Read a specific URL the user provided.
+    """
+    response = await client.post(
+        f"{BASE_URL}/content/extract",
+        headers={"Authorization": f"Bearer {API_KEY}"},
+        json={"url": url, "max_length": 5000}  # Longer for deep reads
+    )
+    response.raise_for_status()
+
+    result = response.json()["result"]
+
+    if not result["success"]:
+        raise ValueError(f"Could not read page: {result['error']}")
+
+    # Format for LLM
+    header = f"# {result['title'] or 'Untitled'}\n"
+    if result["author"]:
+        header += f"Author: {result['author']}\n"
+    if result["date"]:
+        header += f"Date: {result['date']}\n"
+
+    return header + "\n" + result["content"]
+```
+
+---
+
+## 5. Rate Limits & Best Practices
+
+| Recommendation | Reason |
+|----------------|--------|
+| Use `limit: 5-10` for searches | More results = longer extraction time |
+| Batch URLs when possible | More efficient than sequential calls |
+| Max 20 URLs per batch | Server limit |
+| Set reasonable timeouts (30s) | Content extraction can be slow |
+| Cache results client-side | Same URL rarely changes content |
+| Use `user` parameter | Helps with debugging and rate limiting |
+
+---
+
+## 6. Quick Reference
+
+| Endpoint | Method | Use Case |
+|----------|--------|----------|
+| `/rag/search` | POST | Search web + get extracted content |
+| `/content/extract` | POST | Read a single URL |
+| `/content/extract/batch` | POST | Read multiple URLs |
+| `/health` | GET | Check service status |

docs/roadmap.md

@@ -0,0 +1,208 @@
+# Tatlock Implementation Roadmap
+
+> **Reference**: See [philosophy.md](philosophy.md) for the target architecture and vision
+
+This document tracks open/planned work. Completed phases have been removed.
+
+## Current State (v2.0.5)
+
+**What we have**:
+- OpenAI-compatible API (Responses API + Chat Completions)
+- Two-tier architecture (Steward → Tatlock)
+- Household staff: Tatlock (Butler), Steward, Librarian, Biographer
+- Core tools: Calculator, Date/Time, Web search (SearXNG)
+- Memory system: Qdrant (vector), Redis (session cache), multi-tenancy via ContextVar
+- Dual backend: Ollama/gemma4 (primary) + Claude (fallback)
+- 439 tests with good coverage
+
+---
+
+## Phase 4: Expert Household Staff — Remaining Agents
+
+**Goal**: Implement remaining domain-specific expert agents
+
+### Planned Agents
+
+1. **The Developer** (Software Development)
+   - Code generation assistance
+   - Debugging support
+   - Documentation generation
+   - Architecture guidance
+
+2. **The Handyman** (System Maintenance)
+   - System status queries
+   - Log analysis
+   - Basic troubleshooting
+   - Infrastructure monitoring
+
+3. **The Secretary** (Scheduling & Organization)
+   - Calendar integration
+   - Task management
+   - Reminder system
+   - Schedule conflict detection
+
+4. **The Housekeeper** (Home Automation)
+   - Home Assistant integration
+   - Device control interface
+   - Status queries
+   - Automation triggers
+
+### Each Agent Includes
+- Specialized prompt and personality
+- Domain-specific tools
+- MCP integration points (where applicable)
+- Integration with Butler orchestration
+
+### Success Criteria
+- [ ] Each agent implemented as separate module
+- [ ] Agents callable via tool framework
+- [ ] Can invoke specialized models (e.g., Codestral for Developer)
+
+---
+
+## Phase 5: Persistence Layer — Database & Multi-Tenancy
+
+**Goal**: Add persistent storage and multi-user support
+
+### Deliverables
+
+1. **PostgreSQL Integration**
+   - Docker compose configuration
+   - Database schema with tenant isolation
+   - Alembic migrations
+   - SQLAlchemy models
+
+2. **Multi-Tenant Architecture**
+   - Tenant identification middleware
+   - Tenant-scoped database sessions
+   - User authentication system
+   - Per-tenant data isolation
+
+3. **Core Data Models**
+   - Users and tenants
+   - Conversations and messages (migrate from in-memory)
+   - Agent interactions log
+   - System configuration and preferences
+
+### Success Criteria
+- [ ] PostgreSQL container running
+- [ ] Multiple users authenticate separately
+- [ ] Each user sees only their own data
+- [ ] Conversations persist across restarts
+- [ ] Database migrations work correctly
+
+---
+
+## Phase 7: MCP (Model Context Protocol) Integration
+
+**Goal**: Enable rich tool integrations via MCP
+
+See also [claude-integration.md](claude-integration.md) for MCP server implementation details.
+
+### Deliverables
+
+1. **MCP Server Framework**
+   - MCP server implementation
+   - Tool registration via MCP
+   - Schema validation
+   - Error handling
+
+2. **MCP Client in Agents**
+   - PydanticAI MCP integration
+   - Tool discovery from MCP servers
+   - Dynamic tool loading
+
+3. **Initial MCP Tools**
+   - File system operations
+   - Database queries
+   - API integrations
+   - System commands
+
+### Success Criteria
+- [ ] MCP server running
+- [ ] Tools exposed via MCP protocol
+- [ ] Agents can discover and use MCP tools
+- [ ] New tools addable without code changes
+- [ ] MCP tools visible in Steward recommendations
+
+---
+
+## Phase 8: Advanced Memory & Context — Remaining Work
+
+**Goal**: Implement sophisticated context management and personalization
+
+### Open Deliverables
+
+1. **Context Management**
+   - Smart context window trimming
+   - Conversation branching
+   - Topic tracking
+
+2. **Personalization**
+   - User preference learning
+   - Interaction pattern analysis
+   - Adaptive responses
+   - Custom agent personalities per user
+
+### Success Criteria
+- [ ] Conversations automatically embedded to Qdrant
+- [ ] Memory improves over time (learning from interactions)
+
+---
+
+## Phase 9: Extended Household Staff
+
+**Goal**: Add specialized agents for additional domains
+
+### Future Agents
+- **The Accountant** — Expense tracking, budgets, financial reports
+- **The Chef** — Meal planning, recipes, nutrition tracking
+- Others as needs emerge
+
+---
+
+## Phase 10: User Experience Refinement
+
+**Goal**: Polish the interaction experience
+
+- Personality tuning and consistency
+- Better progress indicators
+- Response time improvements
+- Streaming smoothness
+
+---
+
+## Phase 11: Production Hardening
+
+**Goal**: Make the system production-ready for homelab deployment
+
+- Complete docker-compose stack
+- Health checks and monitoring
+- Authentication hardening and rate limiting
+- Installation and troubleshooting documentation
+
+---
+
+## Dependencies
+
+```
+Phase 4 (Remaining Agents)
+    ↓
+Phase 5 (Database/Multi-Tenancy) ← Can be deferred
+    ↓
+Phase 7 (MCP) → Phase 8 (Advanced Memory)
+    ↓
+Phase 9 (Extended Staff) → Phase 10 (UX) → Phase 11 (Production)
+```
+
+**Can Be Deferred**: Phase 5 until you need persistence
+**Parallel Opportunities**: Phases 7 and 8 can overlap; 9 and 10 ongoing
+
+---
+
+## Next Steps
+
+1. Implement The Developer agent for code assistance
+2. Add Home Assistant integration for The Housekeeper
+3. Integrate scheduling service for The Secretary
+4. MCP server for external Claude access

docs/steward-routing-findings.md

@@ -0,0 +1,110 @@
+# Steward Routing & Thinking — Findings
+
+**Outcome: no change shipped.** The Steward stays on `gemma4:e2b` with model
+thinking left at its default (on). Every alternative was measured and every one
+loses. This document exists so the experiment is not repeated on the same
+premise.
+
+Run 2026-08-08 with `scripts/benchmark_routing.py` and
+`scripts/fixtures/routing_fixtures.py` (40 labelled queries, one repeat per
+cell, temperature 0.3 as production sends).
+
+---
+
+## The premise was wrong
+
+The experiment was designed around an observation that the Steward pays ~300
+tokens per turn for reasoning that is generated and thrown away: it calls
+`/api/generate`, gemma4 reasons by default, and **no `thinking` field comes back
+in the response**. Disabling thinking therefore looked close to free.
+
+It is not. The reasoning is not discarded — it is emitted inline in `response`,
+and it is what produces a correct `DELEGATE:` line. Those tokens are the work,
+not waste. Suppressing them costs 12.5 points of routing accuracy.
+
+## Results
+
+| config | exact | under | over | tokens | latency | resident | predicted | co-resident with nomic |
+|---|---|---|---|---|---|---|---|---|
+| **e2b, thinking** *(production)* | **97.5%** | 2.5% | 0% | 361 | 5179 ms | 1778 MB | 7.8 GiB | yes |
+| e2b, `think: false` | 85.0% | 12.5% | 5.0% | 48 | 1435 ms | 1778 MB | 7.8 GiB | yes |
+| e4b, thinking | 100% | 0% | 0% | 192 | 4726 ms | 3089 MB | 10.6 GiB | **no** |
+| e4b, `think: false` | 97.5% | 2.5% | 0% | 52 | 2269 ms | 3089 MB | 10.6 GiB | **no** |
+
+`think: true` was also measured and landed within one fixture of the default on
+both models, so production's implicit thinking is the same thing as asking for
+it explicitly. Format compliance was 100% in every cell — a `DELEGATE:` line is
+always emitted.
+
+With 40 fixtures and one repeat, each result is worth 2.5 points, so the
+97.5-vs-100 gaps are single fixtures and inside the noise. The latency and token
+medians (40 calls each) and the e2b `think: false` degradation (6 failures with a
+consistent mechanism) are the parts worth trusting.
+
+## Why each alternative loses
+
+**`think: false` on e2b** — 85% exact, and the failures are not random. All three
+multi-capability fixtures under-route, each missing a second capability. Without
+reasoning the model names one capability and stops decomposing. It is not
+degraded across the board; it specifically stops handling compound requests,
+which is where a user would most notice the Butler quietly doing half the job.
+
+**e4b, either setting** — disqualified by memory, not by quality. Ollama predicts
+**10.6 GiB** for it at 16k context. Maximum available on this card is ~7.9 GiB
+(10.4 free − 2.0 GPU overhead − 0.46 minimum), so e4b *always* exceeds the budget
+and evicts every co-resident before loading. Observed directly: loading it threw
+out both `gemma4:e2b` and `nomic-embed-text`. Losing nomic means Tatlock memory
+and library-desk thrash on every embedding call. Note this is not caused by the
+2 GiB reservation — without it, available would be ~9.7 GiB, still under 10.6.
+
+**Lower `OLLAMA_CONTEXT_LENGTH`** — the obvious way to free headroom, and it does
+not work. Dropping 16384 → 2048, an 8× reduction, moved the prediction only from
+7.8 to 6.7 GiB. The prediction is dominated by weights and batch size, not KV
+cache. It would also truncate the Librarian's retrieved passages and webber's
+code context for a 14% saving that funds nothing.
+
+**Per-request `num_ctx`** — worse. A single request with a different `num_ctx`
+reloads the shared runner, which **drops the `keep_alive: -1` pin** (expiry fell
+from year-2318 to a 2-hour default) and evicts nomic. Three services share this
+Ollama, so mixed context sizes are a thrash generator, and it fails silently.
+
+**`OLLAMA_NUM_PARALLEL > 1`** — never viable here. e2b already predicts 7.8 GiB
+against ~7.9 available, so there is no room for a second slot at any context
+length. It is also set to 1 deliberately, to avoid batch overflow panics.
+
+## What the two axes actually control
+
+They do not interact, which is the useful part:
+
+- **Model choice** governs VRAM and co-residency. e2b 1778 MB, e4b 3089 MB.
+- **Think setting** governs tokens, latency and routing quality — and costs
+  **nothing** in VRAM. Verified: e2b is resident at 1778 MB with `think` unset,
+  true and false alike, because the KV cache is allocated for the full context at
+  load time and `think` is a per-request generation parameter.
+
+So the only real question is whether 313 tokens and 3.7 seconds are worth 12.5
+points of compound-query routing. On a turn that is already three sequential
+Ollama calls, they are.
+
+## Prerequisite: the extraction fix
+
+These numbers are only meaningful because `_extract_capabilities` was fixed first
+(commit `a905363`). It previously substring-matched capability *domains* across
+the Steward's entire response, so ordinary English in the `REASON:` line selected
+agents — "description" contains the housekeeper domain "script", "acknowledge"
+contains "knowledge" and "know".
+
+That made **prose length a routing input**. Benchmarking against it would have
+shown `think: false` improving routing purely because shorter output produces
+fewer accidental substring hits — a thinking policy derived from a parsing
+artefact. The `adversarial` fixture group is regression coverage for exactly this.
+
+## If this is revisited
+
+The constraint is the single 11 GB card, not the model. A second inference host
+(*forge*) removes it entirely, and e4b's 100% routing becomes reachable without
+evicting anything. Re-run then; on this card the answer is settled.
+
+`scripts/benchmark_routing.py` takes `--models`, `--think` and `--repeats`, and
+restores GPU residency on exit — including on SIGTERM, which the first version
+did not.

docs/testing-improvements.md

@@ -0,0 +1,105 @@
+# Testing Improvements for LLM Outputs
+
+## Problem
+
+LLM outputs are non-deterministic. Tests checking for exact string matches fail when the LLM writes "thirty-seven" instead of "37".
+
+## Proposed Solutions
+
+### 1. LLM-as-Judge Pattern
+
+Use a smaller/faster model to evaluate semantic correctness:
+
+```python
+async def llm_judge(output: str, criteria: str) -> bool:
+    """Use LLM to evaluate if output meets criteria."""
+    prompt = f"""
+    Evaluate if this output is correct:
+    Output: {output}
+    Criteria: {criteria}
+    Answer only YES or NO.
+    """
+    result = await judge_model.run(prompt)
+    return "YES" in result.output.upper()
+
+# Usage in test:
+assert await llm_judge(
+    response,
+    "The answer correctly states that sqrt(144) + 25 = 37"
+)
+```
+
+### 2. Fuzzy/Regex Matching
+
+For numeric answers, accept multiple representations:
+
+```python
+import re
+
+def contains_number(text: str, number: int) -> bool:
+    """Check if text contains number in any form."""
+    patterns = [
+        rf'\b{number}\b',  # Digit form
+        number_to_words(number),  # Word form
+    ]
+    return any(re.search(p, text, re.I) for p in patterns)
+
+# Usage:
+assert contains_number(response, 37)  # Matches "37" or "thirty-seven"
+```
+
+### 3. DeepEval Framework
+
+```python
+from deepeval.metrics import AnswerRelevancyMetric
+from deepeval.test_case import LLMTestCase
+
+def test_calculation():
+    test_case = LLMTestCase(
+        input="What is sqrt(144) + 25?",
+        actual_output=response,
+        expected_output="37"
+    )
+    metric = AnswerRelevancyMetric(threshold=0.7)
+    assert metric.measure(test_case)
+```
+
+### 4. pytest-evals Plugin
+
+Minimal pytest plugin for LLM testing with metrics collection.
+
+```bash
+pip install pytest-evals
+```
+
+### 5. Multiple Runs with Threshold
+
+Run flaky tests multiple times and require majority pass:
+
+```python
+@pytest.mark.flaky(reruns=3, reruns_delay=1)
+def test_llm_response():
+    ...
+```
+
+Or custom:
+
+```python
+@pytest.mark.parametrize("run", range(3))
+def test_llm_response(run):
+    ...
+    # Aggregate results across runs
+```
+
+## Resources
+
+- [DeepEval](https://github.com/confident-ai/deepeval) - LLM evaluation framework
+- [pytest-evals](https://github.com/AlmogBaku/pytest-evals) - pytest plugin for LLM evals
+- [LLM Testing Guide 2025](https://www.confident-ai.com/blog/llm-testing-in-2024-top-methods-and-strategies)
+- [Testing LLM Applications - Langfuse](https://langfuse.com/blog/2025-10-21-testing-llm-applications)
+
+## Implementation Priority
+
+1. Add fuzzy number matching helper (quick win)
+2. Evaluate DeepEval for complex output testing
+3. Consider LLM-as-judge for semantic correctness

governance/README.md

@@ -0,0 +1,54 @@
+# Decisions, Questions, Rejected
+
+This directory holds structured planning records that pql parses
+into pql.db. Each record is a `### [DQR]-N: Title` heading inside
+a markdown file. Files live in three per-type subdirectories:
+
+- `decisions/<domain>.md` — confirmed design decisions
+- `questions/<domain>.md` — open questions that may resolve into
+  decisions or rejected proposals
+- `rejected/<domain>.md` — rejected proposals (kept for the audit
+  trail)
+
+The parser infers domain from the filename stem and record type
+from the parent subdirectory.
+
+D-records that propose implementation work link to `initiative`-type
+tickets via `decision_ref`. Run `pql decisions show <id>
+--with-tickets` to inspect implementation status.
+
+## Recommended domains
+
+Start with this canonical set; create files as records land in
+each domain:
+
+- **architecture** — structural commitments (storage, layering,
+  languages, libraries)
+- **process** — team workflow (commits, branches, releases, reviews)
+- **design** — user-facing surface (UX, UI, public APIs)
+- **coding-conventions** — team-internal code shape (style, lint,
+  file layout)
+- **testing** — quality strategy (coverage, layers, gates)
+
+You might also want, project-permitting:
+
+- `accessibility` — if you ship user-facing software
+- `security` — if you handle user data or network surfaces
+- `licensing` — if you release open-source or commercial
+- `documentation` — if user-docs are non-trivial
+- `deployment` — if shipping is non-trivial
+- `performance` — if you have perf budgets / SLOs
+
+<!-- pql:records (auto-generated; do not edit manually) -->
+
+## Decisions
+
+- _(none)_
+
+## Open questions
+
+- _(none)_
+
+## Rejected
+
+- _(none)_

pyproject.toml

@@ -4,17 +4,72 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "tatlock"
-version = "1.2.0"
+version = "2.4.3"
 description = "OpenAI-compatible API with Ollama backend"
 requires-python = ">=3.12"
-dependencies = []
+dependencies = [
+    "fastapi>=0.123,<0.124",
+    "uvicorn[standard]>=0.38,<0.39",
+    "pydantic>=2.11,<2.13",
+    "pydantic-settings>=2.12,<2.13",
+    "pydantic-ai-slim[openai,anthropic]>=1.27,<1.28",
+    # pydantic-ai 1.27 imports the private opentelemetry._events module,
+    # removed in opentelemetry-api 1.44 — cap until pydantic-ai is bumped
+    "opentelemetry-api>=1.30,<1.44",
+    "anthropic>=0.77,<1.0",
+    "httpx>=0.28,<0.29",
+    "sse-starlette>=3.0,<3.1",
+    "python-dotenv>=1.2,<1.3",
+    "starlette>=0.45,<0.46",
+    "redis[hiredis]>=5.2,<6.0",
+    "qdrant-client>=1.12,<2.0",
+    "structlog>=24.1,<25.0",
+]
+
+[project.optional-dependencies]
+dev = [
+    "pytest>=8.3,<8.4",
+    "pytest-asyncio>=0.25,<0.26",
+    "pytest-cov>=6.0,<6.1",
+    "pytest-mock>=3.14,<3.15",
+    "ruff>=0.8,<0.9",
+    "mypy>=1.14,<1.15",
+    "faker>=34.0,<35.0",
+    "coverage[toml]>=7.7,<7.8",
+]
 
 [tool.pytest.ini_options]
+testpaths = ["tests"]
+python_files = ["test_*.py"]
+python_classes = ["Test*"]
+python_functions = ["test_*"]
 asyncio_mode = "auto"
+asyncio_default_fixture_loop_scope = "function"
+cache_dir = ".cache/pytest"
+markers = [
+    "unit: Unit tests",
+    "integration: Integration tests",
+    "slow: Slow running tests",
+    "contract: Wire-level contract tests against live service boundaries",
+]
+addopts = [
+    "--verbose",
+    "--strict-markers",
+    "--tb=short",
+    "--cov=src",
+    "--cov-report=term-missing",
+    "--cov-report=html:build/coverage/html",
+    "--cov-report=xml:build/coverage/coverage.xml",
+    "--cov-branch",
+]
+filterwarnings = [
+    "ignore::DeprecationWarning",
+]
 
 [tool.coverage.run]
 source = ["src"]
 branch = true
+data_file = "build/coverage/.coverage"
 omit = [
     "*/tests/*",
     "*/__pycache__/*",
@@ -37,11 +92,15 @@ exclude_lines = [
 ]
 
 [tool.coverage.html]
-directory = "htmlcov"
+directory = "build/coverage/html"
+
+[tool.coverage.xml]
+output = "build/coverage/coverage.xml"
 
 [tool.ruff]
 line-length = 100
 target-version = "py312"
+cache-dir = ".cache/ruff"
 
 [tool.ruff.lint]
 select = [
@@ -64,6 +123,7 @@ ignore = [
 
 [tool.mypy]
 python_version = "3.12"
+cache_dir = ".cache/mypy"
 warn_return_any = true
 warn_unused_configs = true
 disallow_untyped_defs = true

pytest.ini

@@ -1,28 +0,0 @@
-[pytest]
-testpaths = tests
-python_files = test_*.py
-python_classes = Test*
-python_functions = test_*
-asyncio_mode = auto
-asyncio_default_fixture_loop_scope = function
-
-# Markers
-markers =
-    unit: Unit tests
-    integration: Integration tests
-    slow: Slow running tests
-
-# Coverage options (overridden by pyproject.toml)
-addopts =
-    --verbose
-    --strict-markers
-    --tb=short
-    --cov=src
-    --cov-report=term-missing
-    --cov-report=html
-    --cov-report=xml
-    --cov-branch
-
-# Ignore warnings from dependencies
-filterwarnings =
-    ignore::DeprecationWarning

requirements-dev.txt

@@ -1,25 +0,0 @@
-# Development and Testing Dependencies
-# Install with: pip install -r requirements.txt -r requirements-dev.txt
-
-# Testing Framework
-# Latest pytest with async support
-pytest>=8.3,<8.4
-pytest-asyncio>=0.25,<0.26
-pytest-cov>=6.0,<6.1
-
-# Test client for FastAPI
-httpx>=0.28,<0.29  # Already in requirements.txt but needed for test client
-
-# Code Quality
-# Linting and formatting
-ruff>=0.8,<0.9
-
-# Type checking
-mypy>=1.14,<1.15
-
-# Testing utilities
-pytest-mock>=3.14,<3.15
-faker>=34.0,<35.0
-
-# Coverage reporting
-coverage[toml]>=7.7,<7.8

requirements.txt

@@ -1,55 +0,0 @@
-# Core FastAPI framework and server
-# FastAPI: Modern, fast web framework for building APIs
-# Latest: 0.123.9 (Dec 4, 2025) - No known CVEs
-fastapi>=0.123,<0.124
-
-# ASGI server for running FastAPI
-# Latest: 0.38.0 (Oct 18, 2025) - No known CVEs
-# Note: Old versions had CVE-2020-7694/7695, but 0.38.0 is secure
-uvicorn[standard]>=0.38,<0.39
-
-# Additional dependencies
-# Pydantic for data validation (comes with pydantic-ai but pinning explicitly)
-# Updated to >=2.11 due to ag-ui-protocol dependency requirement
-# Latest: 2.12.4 (Nov 5, 2025) - No known CVEs
-pydantic>=2.11,<2.13
-
-# AI/LLM integration
-# PydanticAI: Agent framework for using Pydantic with LLMs
-# Latest: 1.27.0 (Dec 5, 2025) - No known CVEs
-# Supports Ollama backend out of the box
-pydantic-ai>=1.27,<1.28
-
-# HTTP client for Ollama communication
-# Latest: 0.28.1 - No known CVEs
-httpx>=0.28,<0.29
-
-# Server-Sent Events for streaming responses
-# Required for OpenAI-compatible streaming endpoints
-# Latest: 3.0.2 (Oct 30, 2025) - No known CVEs
-sse-starlette>=3.0,<3.1
-
-# Configuration management
-# Latest: 1.2.1 (Oct 26, 2025) - No known CVEs
-python-dotenv>=1.2,<1.3
-
-# ASGI toolkit (dependency of FastAPI, pinning for security)
-starlette>=0.45,<0.46
-
-# Redis for performance benchmarking and caching
-# Latest: 5.2.1 (Dec 5, 2025) - No known CVEs
-# hiredis: C parser for better performance
-redis[hiredis]>=5.2,<6.0
-
-# Qdrant vector database client for memory storage
-# Latest: 1.12.1 (Dec 2025) - No known CVEs
-qdrant-client>=1.12,<2.0
-
-# Structured logging for observability
-# Latest: 24.4.0 (Aug 22, 2024) - No known CVEs
-structlog>=24.1,<25.0
-
-# Note on version locking strategy:
-# Using >=X.Y,<X.(Y+1) format to lock to minor versions
-# This protects against supply chain attacks while allowing patch updates
-# Update regularly and review changelogs before upgrading minor versions

scripts/benchmark_routing.py

@@ -0,0 +1,235 @@
+"""
+Benchmark Steward routing quality against model and thinking settings.
+
+Talks to Ollama directly. No Tatlock server, no agents, no tools, nothing is
+executed — the mutating fixtures ("turn on the lights", "update the wiki") only
+ever produce a routing decision. That makes this cheap and repeatable, and it
+isolates the question: does the Steward still pick the right capabilities when
+the model reasons less?
+
+The request body is byte-identical to StewardAgent._call_ollama, plus the
+`think` flag under test, so a cell labelled `unset` is exactly what production
+sends today.
+
+Three thinking settings, because "on vs off" hides the interesting case:
+
+    unset   what production sends now. gemma4 reasons by default, and the
+            response carries no `thinking` field, so those tokens are generated
+            and discarded.
+    true    reasoning requested explicitly and returned in `thinking`.
+    false   reasoning suppressed.
+
+Scoring is deliberately asymmetric. A missing capability under-routes and the
+Butler answers without a tool it needed; a spurious one over-routes, and that is
+a real agent call — a stray librarian is a multi-second web search on a query
+that asked for arithmetic. Over-routing is the predicted failure when thinking
+is off, so `forbid` violations are reported separately rather than folded into
+one accuracy number.
+
+Usage:
+    .venv/bin/python scripts/benchmark_routing.py
+    .venv/bin/python scripts/benchmark_routing.py --models gemma4:e2b
+    .venv/bin/python scripts/benchmark_routing.py --think false --repeats 3
+"""
+from __future__ import annotations
+
+import argparse
+import json
+import statistics
+import sys
+import time
+from datetime import UTC, datetime
+from pathlib import Path
+from typing import Any
+
+import httpx
+
+PROJECT_ROOT = Path(__file__).resolve().parents[1]
+sys.path.insert(0, str(PROJECT_ROOT))
+
+from scripts.fixtures.routing_fixtures import FIXTURES  # noqa: E402
+from scripts.ollama_residency import (  # noqa: E402
+    install_sigterm_handler,
+    residency_guard,
+)
+from src.agents.steward.agent import build_steward_prompt  # noqa: E402
+from src.agents.steward.service import _DELEGATE_LINE_RE, _extract_capabilities  # noqa: E402
+from src.core.startup import register_household_members  # noqa: E402
+
+OLLAMA_URL = "http://localhost:11434"
+DEFAULT_MODELS = ["gemma4:e2b", "gemma4:e4b"]
+DEFAULT_THINK = ["unset", "true", "false"]
+RESULTS_DIR = PROJECT_ROOT / "logs"
+
+
+def build_body(model: str, prompt: str, think: str) -> dict[str, Any]:
+    """Mirror StewardAgent._call_ollama exactly, then add the flag under test."""
+    body: dict[str, Any] = {
+        "model": model,
+        "prompt": prompt,
+        "stream": False,
+        "options": {
+            "temperature": 0.3,  # Lower = more consistent
+            "top_p": 0.9,
+        },
+    }
+    if think != "unset":
+        body["think"] = think == "true"
+    return body
+
+
+def call(client: httpx.Client, body: dict[str, Any]) -> dict[str, Any] | None:
+    try:
+        response = client.post(f"{OLLAMA_URL}/api/generate", json=body)
+        response.raise_for_status()
+        return response.json()
+    except Exception as exc:  # noqa: BLE001 - a failed cell must not abort the run
+        print(f"    ! {exc}", file=sys.stderr)
+        return None
+
+
+def score(fixture: dict, found: list[str]) -> dict[str, Any]:
+    expected = set(fixture["expect"])
+    forbidden = set(fixture["forbid"])
+    got = set(found)
+    missing = sorted(expected - got)
+    spurious = sorted(got & forbidden)
+    return {
+        "found": found,
+        "missing": missing,
+        "spurious": spurious,
+        # Exact only when everything expected arrived and nothing forbidden did.
+        "exact": not missing and not spurious,
+        "under_routed": bool(missing),
+        "over_routed": bool(spurious),
+    }
+
+
+def run_cell(client: httpx.Client, model: str, think: str, repeats: int) -> list[dict[str, Any]]:
+    rows: list[dict[str, Any]] = []
+    for fixture in FIXTURES:
+        prompt = build_steward_prompt(fixture["query"], [])
+        body = build_body(model, prompt, think)
+        for rep in range(repeats):
+            started = time.perf_counter()
+            data = call(client, body)
+            elapsed_ms = (time.perf_counter() - started) * 1000
+            if data is None:
+                rows.append({
+                    "id": fixture["id"], "group": fixture["group"], "rep": rep,
+                    "error": True, "exact": False, "under_routed": False, "over_routed": False,
+                })
+                continue
+
+            text = data.get("response", "") or ""
+            found = _extract_capabilities(text)
+            rows.append({
+                "id": fixture["id"],
+                "group": fixture["group"],
+                "rep": rep,
+                "error": False,
+                "latency_ms": round(elapsed_ms, 1),
+                "eval_tokens": data.get("eval_count"),
+                "prompt_tokens": data.get("prompt_eval_count"),
+                # Did the model obey the documented output shape at all?
+                "has_delegate_line": bool(_DELEGATE_LINE_RE.search(text)),
+                # Whether reasoning came back, as opposed to being generated and dropped.
+                "thinking_returned": bool(data.get("thinking")),
+                "response_chars": len(text),
+                **score(fixture, found),
+            })
+    return rows
+
+
+def summarise(rows: list[dict[str, Any]]) -> dict[str, Any]:
+    ok = [r for r in rows if not r["error"]]
+    if not ok:
+        return {"n": 0, "errors": len(rows)}
+    latencies = [r["latency_ms"] for r in ok]
+    tokens = [r["eval_tokens"] for r in ok if r["eval_tokens"] is not None]
+    return {
+        "n": len(ok),
+        "errors": len(rows) - len(ok),
+        "exact_pct": round(100 * sum(r["exact"] for r in ok) / len(ok), 1),
+        "under_routed_pct": round(100 * sum(r["under_routed"] for r in ok) / len(ok), 1),
+        "over_routed_pct": round(100 * sum(r["over_routed"] for r in ok) / len(ok), 1),
+        "format_ok_pct": round(100 * sum(r["has_delegate_line"] for r in ok) / len(ok), 1),
+        "thinking_returned_pct": round(100 * sum(r["thinking_returned"] for r in ok) / len(ok), 1),
+        "latency_ms_median": round(statistics.median(latencies), 1),
+        "latency_ms_mean": round(statistics.fmean(latencies), 1),
+        "eval_tokens_median": round(statistics.median(tokens), 1) if tokens else None,
+        "eval_tokens_total": sum(tokens) if tokens else None,
+    }
+
+
+def main() -> int:
+    install_sigterm_handler()
+
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--models", default=",".join(DEFAULT_MODELS))
+    parser.add_argument("--think", default=",".join(DEFAULT_THINK),
+                        help="comma-separated subset of unset,true,false")
+    parser.add_argument("--repeats", type=int, default=1)
+    parser.add_argument("--timeout", type=float, default=180.0)
+    args = parser.parse_args()
+
+    models = [m.strip() for m in args.models.split(",") if m.strip()]
+    think_modes = [t.strip() for t in args.think.split(",") if t.strip()]
+
+    # build_steward_prompt reads the registry, and the registry is populated at
+    # application startup. Without this the prompt lists no capabilities and every
+    # cell scores zero for reasons that have nothing to do with the model.
+    register_household_members()
+
+    print(f"{len(FIXTURES)} fixtures x {len(models)} models x {len(think_modes)} think "
+          f"x {args.repeats} repeats = {len(FIXTURES) * len(models) * len(think_modes) * args.repeats} calls\n")
+
+    cells: dict[str, Any] = {}
+    # The guard restores production's pinned models however this exits — a
+    # finished run, a failed cell, Ctrl-C or SIGTERM.
+    with residency_guard(models_used=models), httpx.Client(timeout=args.timeout) as client:
+        for model in models:
+            # Absorb the cold load (~36s) outside the measurements.
+            print(f"warming {model} ...", flush=True)
+            call(client, build_body(model, "hi", "false"))
+            for think in think_modes:
+                key = f"{model}|think={think}"
+                print(f"  {key} ...", end=" ", flush=True)
+                started = time.perf_counter()
+                rows = run_cell(client, model, think, args.repeats)
+                summary = summarise(rows)
+                cells[key] = {"summary": summary, "rows": rows}
+                print(f"exact={summary.get('exact_pct')}% "
+                      f"over={summary.get('over_routed_pct')}% "
+                      f"median={summary.get('latency_ms_median')}ms "
+                      f"({time.perf_counter() - started:.0f}s)")
+
+    RESULTS_DIR.mkdir(parents=True, exist_ok=True)
+    stamp = datetime.now(UTC).strftime("%Y%m%dT%H%M%SZ")
+    out = RESULTS_DIR / f"routing-bench-{stamp}.json"
+    out.write_text(json.dumps({
+        "generated_at": datetime.now(UTC).isoformat(),
+        "fixtures": len(FIXTURES),
+        "repeats": args.repeats,
+        "cells": cells,
+    }, indent=2))
+
+    print(f"\n{'cell':28} {'exact':>7} {'under':>7} {'over':>7} {'fmt':>6} {'tok':>7} {'ms':>8}")
+    print("-" * 76)
+    for key, cell in cells.items():
+        s = cell["summary"]
+        print(f"{key:28} {s.get('exact_pct'):>6}% {s.get('under_routed_pct'):>6}% "
+              f"{s.get('over_routed_pct'):>6}% {s.get('format_ok_pct'):>5}% "
+              f"{str(s.get('eval_tokens_median')):>7} {s.get('latency_ms_median'):>8}")
+    print(f"\nwritten to {out}")
+    return 0
+
+
+if __name__ == "__main__":
+    try:
+        sys.exit(main())
+    except KeyboardInterrupt:
+        # The residency guard has already run by the time this is caught;
+        # a traceback here would just bury its output.
+        print("\ninterrupted", file=sys.stderr)
+        sys.exit(130)

scripts/benchmark_steward.py

@@ -268,7 +268,7 @@ async def run_benchmarks(iterations: int = 10, verbose: bool = False):
             print(f"   Max: {overall_max:.3f}s (target: ≤5.0s)")
             print(f"   Avg: {overall_avg:.3f}s (target: ≤1.67s)")
             print(f"\n   Recommendations:")
-            print(f"   - Switch to a faster model (current: mistral-nemo)")
+            print(f"   - Switch to a faster model (current: gemma4:e2b)")
             print(f"   - Reduce system prompt complexity")
             print(f"   - Limit tool calls (currently limited to 3)")
             print(f"   - Consider caching household registry responses")

scripts/benchmark_tool_calling.py

@@ -0,0 +1,559 @@
+"""
+Benchmark tool calling across different Ollama models via Tatlock API.
+
+Sends test prompts through the full Tatlock pipeline (Steward -> Orchestration
+-> Synthesis) and records tool selection accuracy, latency, and response quality.
+
+Between models, swaps OLLAMA_DEFAULT_MODEL in .env and waits for uvicorn
+auto-reload. Requires the server to be running via ./wakeup.sh.
+
+Usage:
+    .venv/bin/python scripts/benchmark_tool_calling.py
+    .venv/bin/python scripts/benchmark_tool_calling.py --models "gemma4:e4b,gemma4:e2b"
+    .venv/bin/python scripts/benchmark_tool_calling.py --iterations 3
+"""
+import argparse
+import asyncio
+import json
+import re
+import statistics
+import sys
+import time
+from dataclasses import dataclass, field
+from pathlib import Path
+
+import httpx
+
+sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
+from scripts.ollama_residency import install_sigterm_handler, residency_guard
+
+# ---------------------------------------------------------------------------
+# Configuration
+# ---------------------------------------------------------------------------
+
+API_BASE = "http://localhost:8777"
+CHAT_URL = f"{API_BASE}/v1/chat/completions"
+HEALTH_URL = f"{API_BASE}/health"
+OLLAMA_URL = "http://localhost:11434"
+ENV_PATH = Path(__file__).parent.parent / ".env"
+
+DEFAULT_MODELS = ["mistral-nemo-large:latest", "gemma4:e4b", "gemma4:e2b"]
+
+# ---------------------------------------------------------------------------
+# Test scenarios
+# ---------------------------------------------------------------------------
+
+
+@dataclass
+class Scenario:
+    name: str
+    prompt: str
+    expected_tool: str | None  # None = no tool expected
+    # Patterns to check in the response text for indirect tool-use evidence
+    success_patterns: list[str] = field(default_factory=list)
+    category: str = "basic"
+
+
+SCENARIOS = [
+    # --- Should call calculate_math ---
+    Scenario(
+        name="Simple arithmetic",
+        prompt="What is 144 divided by 12?",
+        expected_tool="calculate_math",
+        success_patterns=["12"],
+        category="calculator",
+    ),
+    Scenario(
+        name="Square root",
+        prompt="What's the square root of 256?",
+        expected_tool="calculate_math",
+        success_patterns=["16"],
+        category="calculator",
+    ),
+    Scenario(
+        name="Complex math",
+        prompt="Calculate pi times the square of 5",
+        expected_tool="calculate_math",
+        success_patterns=["78.5"],  # pi * 25 ≈ 78.54
+        category="calculator",
+    ),
+    Scenario(
+        name="Word problem",
+        prompt="If I have 3 bags with 17 apples each and I eat 4, how many apples do I have?",
+        expected_tool="calculate_math",
+        success_patterns=["47"],
+        category="calculator",
+    ),
+
+    # --- Should call get_current_time ---
+    Scenario(
+        name="Current date",
+        prompt="What's today's date?",
+        expected_tool="get_current_time",
+        success_patterns=["2026"],  # Should contain current year
+        category="datetime",
+    ),
+    Scenario(
+        name="Current time",
+        prompt="What time is it right now?",
+        expected_tool="get_current_time",
+        success_patterns=[":"],  # Time format contains colons
+        category="datetime",
+    ),
+
+    # --- Should call calculate_date_offset ---
+    Scenario(
+        name="Relative date past",
+        prompt="What was the date 2 weeks ago?",
+        expected_tool="calculate_date_offset",
+        success_patterns=["2026"],
+        category="datetime",
+    ),
+
+    # --- Should call calculate_time_difference ---
+    Scenario(
+        name="Date difference",
+        prompt="How many days between January 1st 2025 and March 15th 2025?",
+        expected_tool="calculate_time_difference",
+        success_patterns=["73", "74"],  # 73 or 74 days
+        category="datetime",
+    ),
+
+    # --- Should NOT call any tool ---
+    Scenario(
+        name="Greeting",
+        prompt="Hello! How are you?",
+        expected_tool=None,
+        success_patterns=["sir"],  # Butler personality
+        category="no_tool",
+    ),
+    Scenario(
+        name="Knowledge question",
+        prompt="What is the capital of France?",
+        expected_tool=None,
+        success_patterns=["Paris"],
+        category="no_tool",
+    ),
+    Scenario(
+        name="Opinion request",
+        prompt="What do you think about rainy days?",
+        expected_tool=None,
+        category="no_tool",
+    ),
+]
+
+# ---------------------------------------------------------------------------
+# Result tracking
+# ---------------------------------------------------------------------------
+
+
+@dataclass
+class RunResult:
+    scenario: str
+    model: str
+    iteration: int
+    latency: float
+    response_text: str
+    has_correct_answer: bool
+    error: str | None = None
+
+
+@dataclass
+class ModelStats:
+    model: str
+    results: list[RunResult] = field(default_factory=list)
+
+    @property
+    def total(self) -> int:
+        return len(self.results)
+
+    @property
+    def errors(self) -> int:
+        return sum(1 for r in self.results if r.error)
+
+    @property
+    def accuracy(self) -> float:
+        valid = [r for r in self.results if not r.error]
+        if not valid:
+            return 0
+        return sum(1 for r in valid if r.has_correct_answer) / len(valid) * 100
+
+    @property
+    def avg_latency(self) -> float:
+        lats = [r.latency for r in self.results if not r.error]
+        return statistics.mean(lats) if lats else 0
+
+    @property
+    def p95_latency(self) -> float:
+        lats = sorted(r.latency for r in self.results if not r.error)
+        if not lats:
+            return 0
+        return lats[min(int(len(lats) * 0.95), len(lats) - 1)]
+
+    @property
+    def max_latency(self) -> float:
+        lats = [r.latency for r in self.results if not r.error]
+        return max(lats) if lats else 0
+
+    def category_accuracy(self, category: str) -> float:
+        cat_scenarios = {s.name for s in SCENARIOS if s.category == category}
+        valid = [r for r in self.results if not r.error and r.scenario in cat_scenarios]
+        if not valid:
+            return 0
+        return sum(1 for r in valid if r.has_correct_answer) / len(valid) * 100
+
+
+# ---------------------------------------------------------------------------
+# .env manipulation
+# ---------------------------------------------------------------------------
+
+
+def swap_model_in_env(model_name: str):
+    """Swap OLLAMA_DEFAULT_MODEL in .env file."""
+    content = ENV_PATH.read_text()
+    content = re.sub(
+        r'^OLLAMA_DEFAULT_MODEL=.*$',
+        f'OLLAMA_DEFAULT_MODEL={model_name}',
+        content,
+        flags=re.MULTILINE,
+    )
+    ENV_PATH.write_text(content)
+    print(f"  .env updated: OLLAMA_DEFAULT_MODEL={model_name}")
+
+
+async def wait_for_server_reload(client: httpx.AsyncClient, timeout: float = 30):
+    """Wait for uvicorn to auto-reload after .env change."""
+    # Give uvicorn a moment to detect the file change
+    await asyncio.sleep(3)
+
+    # Poll health endpoint
+    deadline = time.monotonic() + timeout
+    while time.monotonic() < deadline:
+        try:
+            r = await client.get(HEALTH_URL, timeout=5)
+            if r.status_code == 200:
+                return
+        except Exception:
+            pass
+        await asyncio.sleep(1)
+
+    raise TimeoutError("Server did not come back after reload")
+
+
+async def warm_up_ollama_model(client: httpx.AsyncClient, model_name: str):
+    """Send a throwaway request to load the model into VRAM."""
+    print(f"  Warming up {model_name} in Ollama...", end=" ", flush=True)
+    try:
+        r = await client.post(
+            f"{OLLAMA_URL}/api/generate",
+            json={"model": model_name, "prompt": "hi", "stream": False},
+            timeout=120,
+        )
+        r.raise_for_status()
+        duration = r.json().get("total_duration", 0) / 1e9
+        print(f"OK ({duration:.1f}s)")
+    except Exception as e:
+        print(f"WARN: {e}")
+
+
+# ---------------------------------------------------------------------------
+# Core benchmark logic
+# ---------------------------------------------------------------------------
+
+
+async def run_scenario(
+    client: httpx.AsyncClient,
+    scenario: Scenario,
+    model: str,
+    iteration: int,
+) -> RunResult:
+    """Run a single scenario through the Tatlock API."""
+    payload = {
+        "model": "Tatlock",
+        "messages": [{"role": "user", "content": scenario.prompt}],
+    }
+
+    start = time.monotonic()
+    try:
+        r = await client.post(CHAT_URL, json=payload, timeout=120)
+        latency = time.monotonic() - start
+
+        if r.status_code != 200:
+            return RunResult(
+                scenario=scenario.name,
+                model=model,
+                iteration=iteration,
+                latency=latency,
+                response_text="",
+                has_correct_answer=False,
+                error=f"HTTP {r.status_code}: {r.text[:100]}",
+            )
+
+        data = r.json()
+        response_text = data["choices"][0]["message"]["content"]
+
+        # Check if the response contains expected patterns
+        has_correct = True
+        if scenario.success_patterns:
+            has_correct = any(
+                p.lower() in response_text.lower()
+                for p in scenario.success_patterns
+            )
+
+        return RunResult(
+            scenario=scenario.name,
+            model=model,
+            iteration=iteration,
+            latency=latency,
+            response_text=response_text,
+            has_correct_answer=has_correct,
+        )
+
+    except Exception as e:
+        latency = time.monotonic() - start
+        return RunResult(
+            scenario=scenario.name,
+            model=model,
+            iteration=iteration,
+            latency=latency,
+            response_text="",
+            has_correct_answer=False,
+            error=str(e)[:200],
+        )
+
+
+async def benchmark_model(
+    client: httpx.AsyncClient,
+    model_name: str,
+    iterations: int,
+) -> ModelStats:
+    """Run all scenarios for a single model."""
+    stats = ModelStats(model=model_name)
+
+    print(f"\n{'=' * 70}")
+    print(f"  Model: {model_name}")
+    print(f"{'=' * 70}")
+
+    # Swap model in .env
+    swap_model_in_env(model_name)
+
+    # Warm up model in Ollama BEFORE server reload picks it up
+    await warm_up_ollama_model(client, model_name)
+
+    # Wait for server to reload with new model
+    print("  Waiting for server reload...", end=" ", flush=True)
+    await wait_for_server_reload(client)
+    print("OK")
+
+    # Run a throwaway request through the full pipeline to warm up
+    print("  Warming up pipeline...", end=" ", flush=True)
+    try:
+        await client.post(
+            CHAT_URL,
+            json={"model": "Tatlock", "messages": [{"role": "user", "content": "hi"}]},
+            timeout=120,
+        )
+        print("OK")
+    except Exception as e:
+        print(f"WARN: {e}")
+
+    for iteration in range(iterations):
+        if iterations > 1:
+            print(f"\n  --- Iteration {iteration + 1}/{iterations} ---")
+
+        for scenario in SCENARIOS:
+            result = await run_scenario(client, scenario, model_name, iteration)
+            stats.results.append(result)
+
+            # Display
+            if result.error:
+                print(
+                    f"  [ERR ] {scenario.name:30s} {result.latency:5.1f}s  "
+                    f"{result.error[:60]}"
+                )
+            elif result.has_correct_answer:
+                preview = result.response_text[:60].replace("\n", " ")
+                print(f"  [OK  ] {scenario.name:30s} {result.latency:5.1f}s  {preview}")
+            else:
+                preview = result.response_text[:60].replace("\n", " ")
+                print(f"  [MISS] {scenario.name:30s} {result.latency:5.1f}s  {preview}")
+
+    return stats
+
+
+def print_comparison(all_stats: list[ModelStats]):
+    """Print side-by-side comparison table."""
+    print("\n" + "=" * 80)
+    print("  COMPARISON SUMMARY")
+    print("=" * 80)
+
+    col_width = max(len(s.model) for s in all_stats) + 2
+    label_width = 32
+
+    header = f"{'Metric':<{label_width}}"
+    for s in all_stats:
+        header += f"  {s.model:>{col_width}}"
+    print(f"\n{header}")
+    print("-" * (label_width + (col_width + 2) * len(all_stats)))
+
+    # Answer accuracy
+    row = f"{'Correct answer rate':<{label_width}}"
+    for s in all_stats:
+        row += f"  {s.accuracy:>{col_width - 1}.1f}%"
+    print(row)
+
+    # Latency
+    row = f"{'Avg latency':<{label_width}}"
+    for s in all_stats:
+        row += f"  {s.avg_latency:>{col_width - 1}.1f}s"
+    print(row)
+
+    row = f"{'P95 latency':<{label_width}}"
+    for s in all_stats:
+        row += f"  {s.p95_latency:>{col_width - 1}.1f}s"
+    print(row)
+
+    row = f"{'Max latency':<{label_width}}"
+    for s in all_stats:
+        row += f"  {s.max_latency:>{col_width - 1}.1f}s"
+    print(row)
+
+    # Errors
+    row = f"{'Errors':<{label_width}}"
+    for s in all_stats:
+        row += f"  {s.errors:>{col_width}}"
+    print(row)
+
+    # Per-category
+    categories = sorted(set(sc.category for sc in SCENARIOS))
+    print(f"\n{'Per-category accuracy':<{label_width}}")
+    print("-" * (label_width + (col_width + 2) * len(all_stats)))
+    for cat in categories:
+        row = f"  {cat:<{label_width - 2}}"
+        for s in all_stats:
+            row += f"  {s.category_accuracy(cat):>{col_width - 1}.1f}%"
+        print(row)
+
+    # Mismatches
+    print(f"\n{'Missed answers':<50}")
+    print("-" * 80)
+    any_miss = False
+    for scenario in SCENARIOS:
+        misses = []
+        for s in all_stats:
+            sc_results = [r for r in s.results if r.scenario == scenario.name]
+            fails = [r for r in sc_results if not r.has_correct_answer and not r.error]
+            if fails:
+                preview = fails[0].response_text[:50].replace("\n", " ")
+                misses.append(f"{s.model}: \"{preview}\"")
+        if misses:
+            any_miss = True
+            print(f"  {scenario.name}")
+            for m in misses:
+                print(f"    {m}")
+
+    if not any_miss:
+        print("  (none)")
+
+    print("\n" + "=" * 80)
+
+
+def save_results(all_stats: list[ModelStats], output_path: Path):
+    """Save detailed results to JSON."""
+    data = {}
+    for stats in all_stats:
+        data[stats.model] = {
+            "summary": {
+                "accuracy": stats.accuracy,
+                "avg_latency": round(stats.avg_latency, 2),
+                "p95_latency": round(stats.p95_latency, 2),
+                "max_latency": round(stats.max_latency, 2),
+                "errors": stats.errors,
+                "total_runs": stats.total,
+            },
+            "runs": [
+                {
+                    "scenario": r.scenario,
+                    "iteration": r.iteration,
+                    "latency": round(r.latency, 3),
+                    "has_correct_answer": r.has_correct_answer,
+                    "response_text": r.response_text,
+                    "error": r.error,
+                }
+                for r in stats.results
+            ],
+        }
+
+    output_path.parent.mkdir(parents=True, exist_ok=True)
+    output_path.write_text(json.dumps(data, indent=2))
+    print(f"\nDetailed results saved to: {output_path}")
+
+
+async def main():
+    parser = argparse.ArgumentParser(description="Benchmark tool calling across Ollama models via Tatlock API")
+    parser.add_argument(
+        "--iterations", type=int, default=1,
+        help="Iterations per model (default: 1)",
+    )
+    parser.add_argument(
+        "--models", type=str, default=",".join(DEFAULT_MODELS),
+        help=f"Comma-separated models (default: {','.join(DEFAULT_MODELS)})",
+    )
+    parser.add_argument(
+        "--output", type=str, default="logs/benchmark_results.json",
+        help="JSON output path (default: logs/benchmark_results.json)",
+    )
+    args = parser.parse_args()
+
+    models = [m.strip() for m in args.models.split(",")]
+
+    # Verify server is running
+    async with httpx.AsyncClient() as client:
+        try:
+            r = await client.get(HEALTH_URL, timeout=5)
+            r.raise_for_status()
+            print("Server is running.")
+        except Exception:
+            print("ERROR: Server not running. Start it with ./wakeup.sh first.")
+            return
+
+    print("=" * 70)
+    print("  Tool Calling Benchmark (via Tatlock API)")
+    print("=" * 70)
+    print(f"  Models:     {', '.join(models)}")
+    print(f"  Scenarios:  {len(SCENARIOS)}")
+    print(f"  Iterations: {args.iterations}")
+    print(f"  Total runs: {len(SCENARIOS) * args.iterations * len(models)}")
+
+    # Remember original model to restore after benchmark
+    original_env = ENV_PATH.read_text()
+
+    all_stats = []
+    # Both restores must survive a crash or an interrupt. The .env one especially:
+    # this script rewrites OLLAMA_DEFAULT_MODEL and lets uvicorn reload onto it,
+    # so bailing out mid-run used to leave the *running server* pointed at the
+    # benchmark model — and DEFAULT_MODELS starts at mistral-nemo-large, the 9.2G
+    # model implicated in the 2026-08-07 VRAM outage.
+    install_sigterm_handler()
+    try:
+        with residency_guard(models_used=models):
+            async with httpx.AsyncClient() as client:
+                for model in models:
+                    stats = await benchmark_model(client, model, args.iterations)
+                    all_stats.append(stats)
+    finally:
+        ENV_PATH.write_text(original_env)
+        print("\n  .env restored to original")
+
+    print_comparison(all_stats)
+    save_results(all_stats, Path(args.output))
+
+
+if __name__ == "__main__":
+    try:
+        asyncio.run(main())
+    except KeyboardInterrupt:
+        # .env and GPU residency are both restored by now; do not bury that
+        # output under a traceback.
+        print("\ninterrupted", file=sys.stderr)
+        raise SystemExit(130) from None

scripts/fixtures/routing_fixtures.py

@@ -0,0 +1,159 @@
+"""
+Labelled queries for the Steward routing benchmark.
+
+Each fixture carries both `expect` and `forbid`:
+
+  expect   capabilities that must appear. Missing one is under-routing — the
+           Butler answers without a tool it needed.
+  forbid   capabilities that must not appear. Over-routing is not cosmetic: a
+           spurious librarian is a real multi-second web call, and a spurious
+           housekeeper can actuate hardware.
+
+`forbid` matters more than `expect` here, because over-recommendation is the
+predicted failure when model thinking is disabled and the Steward has less room
+to discriminate.
+
+The `adversarial` group deserves explanation. Until 2026-08-08 the extractor
+substring-matched capability *domains* across the Steward's whole response, so
+ordinary English in its REASON line selected agents: "description" contains the
+housekeeper domain "script", "acknowledge" contains "knowledge" and "know",
+"economy" contains the biographer domain "my". Those queries invite exactly that
+vocabulary. They now serve as an end-to-end regression: routing must depend on
+what the Steward *decided*, not on the words it happened to use while explaining.
+
+Expectations follow the routing rules stated in the Steward prompt itself
+(src/agents/steward/agent.py), not on what a capability could plausibly cover.
+"""
+
+CORE = "tatlock_core"
+LIB = "librarian"
+BIO = "biographer"
+HOUSE = "housekeeper"
+ALL = [CORE, LIB, BIO, HOUSE]
+
+
+def _others(*keep: str) -> list[str]:
+    return [c for c in ALL if c not in keep]
+
+
+FIXTURES: list[dict] = [
+    # --- arithmetic and computation -> tatlock_core --------------------------
+    {"id": "math_add", "group": "math", "query": "What is 61 plus 12?",
+     "expect": [CORE], "forbid": _others(CORE)},
+    {"id": "math_percent", "group": "math", "query": "What is 15% of 240?",
+     "expect": [CORE], "forbid": _others(CORE)},
+    {"id": "math_compound", "group": "math", "query": "If I save 200 a month for 3 years, how much is that?",
+     "expect": [CORE], "forbid": _others(CORE)},
+    {"id": "math_sqrt", "group": "math", "query": "What is the square root of 1764?",
+     "expect": [CORE], "forbid": _others(CORE)},
+
+    # --- date and time -> tatlock_core ---------------------------------------
+    {"id": "time_now", "group": "datetime", "query": "What time is it?",
+     "expect": [CORE], "forbid": _others(CORE)},
+    {"id": "time_date", "group": "datetime", "query": "What is today's date?",
+     "expect": [CORE], "forbid": _others(CORE)},
+    {"id": "time_delta", "group": "datetime", "query": "How many days until Christmas?",
+     "expect": [CORE], "forbid": _others(CORE)},
+
+    # --- personal memory -> biographer ---------------------------------------
+    {"id": "bio_location", "group": "biographer", "query": "Where do I live?",
+     "expect": [BIO], "forbid": [LIB, HOUSE]},
+    {"id": "bio_name", "group": "biographer", "query": "What's my name?",
+     "expect": [BIO], "forbid": [LIB, HOUSE]},
+    {"id": "bio_car", "group": "biographer", "query": "What car do I drive?",
+     "expect": [BIO], "forbid": [LIB, HOUSE]},
+    {"id": "bio_store", "group": "biographer", "query": "Remember that I prefer my coffee black.",
+     "expect": [BIO], "forbid": [LIB, HOUSE]},
+    {"id": "bio_list", "group": "biographer", "query": "What do you know about me?",
+     "expect": [BIO], "forbid": [LIB, HOUSE]},
+    {"id": "bio_forget", "group": "biographer", "query": "Forget my old address.",
+     "expect": [BIO], "forbid": [LIB, HOUSE]},
+
+    # --- research and current information -> librarian ------------------------
+    {"id": "lib_weather", "group": "librarian", "query": "What's the weather in Rotterdam tomorrow?",
+     "expect": [LIB], "forbid": [HOUSE]},
+    {"id": "lib_news", "group": "librarian", "query": "What's in the news today?",
+     "expect": [LIB], "forbid": [HOUSE, BIO]},
+    {"id": "lib_url", "group": "librarian", "query": "Read https://example.com/article and summarise it.",
+     "expect": [LIB], "forbid": [HOUSE, BIO]},
+    {"id": "lib_research", "group": "librarian", "query": "Research how tidal power stations work.",
+     "expect": [LIB], "forbid": [HOUSE, BIO]},
+    {"id": "lib_wiki_create", "group": "librarian", "query": "Create a wiki page about our network topology.",
+     "expect": [LIB], "forbid": [HOUSE, BIO]},
+
+    # --- home automation -> housekeeper --------------------------------------
+    {"id": "house_lights_on", "group": "housekeeper", "query": "Turn on the kitchen lights.",
+     "expect": [HOUSE], "forbid": [LIB, BIO, CORE]},
+    {"id": "house_lights_off", "group": "housekeeper", "query": "Switch off all the lights downstairs.",
+     "expect": [HOUSE], "forbid": [LIB, BIO, CORE]},
+    {"id": "house_thermostat", "group": "housekeeper", "query": "Set the thermostat to 20 degrees.",
+     "expect": [HOUSE], "forbid": [LIB, BIO]},
+    {"id": "house_blinds", "group": "housekeeper", "query": "Close the blinds in the living room.",
+     "expect": [HOUSE], "forbid": [LIB, BIO, CORE]},
+
+    # --- conversational -> nothing at all -------------------------------------
+    # The expensive failure mode: a greeting that triggers a web search.
+    {"id": "chat_greeting", "group": "conversational", "query": "Hello!",
+     "expect": [], "forbid": ALL},
+    {"id": "chat_thanks", "group": "conversational", "query": "Thanks, that's helpful.",
+     "expect": [], "forbid": ALL},
+    {"id": "chat_joke", "group": "conversational", "query": "Tell me a joke.",
+     "expect": [], "forbid": ALL},
+    {"id": "chat_howareyou", "group": "conversational", "query": "How are you doing today?",
+     "expect": [], "forbid": ALL},
+    {"id": "chat_prior_turn", "group": "conversational", "query": "What did I just say?",
+     "expect": [], "forbid": ALL},
+
+    # --- genuinely multi-capability -------------------------------------------
+    {"id": "multi_weather_home", "group": "multi",
+     "query": "What's the weather here, and remember that I like it warm?",
+     "expect": [LIB, BIO], "forbid": []},
+    {"id": "multi_recall_search", "group": "multi",
+     "query": "Look up the best route from my home address to Utrecht.",
+     "expect": [BIO, LIB], "forbid": []},
+    {"id": "multi_math_memory", "group": "multi",
+     "query": "Remember that my budget is 500 euro, then work out 12% of it.",
+     "expect": [BIO, CORE], "forbid": [LIB, HOUSE]},
+
+    # --- adversarial: vocabulary that used to select agents by substring ------
+    # "temperature" is a housekeeper domain, but this is a unit conversion.
+    {"id": "adv_temperature", "group": "adversarial", "query": "Convert 98.6 Fahrenheit to Celsius.",
+     "expect": [CORE], "forbid": [HOUSE, LIB, BIO]},
+    # "description" contains "script"; "discover" contains "cover".
+    {"id": "adv_description", "group": "adversarial",
+     "query": "Give me a short description of what 17 times 23 comes to.",
+     "expect": [CORE], "forbid": [HOUSE, LIB]},
+    # "acknowledge" contains "knowledge" and "know".
+    {"id": "adv_acknowledge", "group": "adversarial",
+     "query": "Just acknowledge this and add 5 and 6 for me.",
+     "expect": [CORE], "forbid": [LIB, BIO]},
+    # "my" appears inside "economy".
+    {"id": "adv_economy", "group": "adversarial",
+     "query": "How many zeros are in one trillion?",
+     "expect": [CORE], "forbid": [BIO, HOUSE]},
+    # "fan" inside "fantastic"; also a climate word without a home-control intent.
+    {"id": "adv_fantastic", "group": "adversarial",
+     "query": "That's fantastic. What is 8 squared?",
+     "expect": [CORE], "forbid": [HOUSE, LIB]},
+    # "home" without any actuation intent.
+    {"id": "adv_home_word", "group": "adversarial", "query": "What time do I usually get home?",
+     "expect": [BIO], "forbid": [HOUSE]},
+    # "search" as ordinary English, not a web-search request.
+    {"id": "adv_search_word", "group": "adversarial",
+     "query": "No need to search anything, just tell me what 9 times 9 is.",
+     "expect": [CORE], "forbid": [LIB]},
+    # "create"/"write" are librarian domains but this is conversational.
+    {"id": "adv_write_word", "group": "adversarial", "query": "Can you write that more simply?",
+     "expect": [], "forbid": [LIB, HOUSE]},
+
+    # --- mutating intents: routing only, nothing is ever executed -------------
+    {"id": "mutate_wiki_update", "group": "mutating", "query": "Update the dossier page with today's findings.",
+     "expect": [LIB], "forbid": [HOUSE, CORE]},
+    {"id": "mutate_scene", "group": "mutating", "query": "Run the movie night scene.",
+     "expect": [HOUSE], "forbid": [LIB, BIO, CORE]},
+]
+
+
+GROUPS = sorted({f["group"] for f in FIXTURES})
+
+assert len({f["id"] for f in FIXTURES}) == len(FIXTURES), "duplicate fixture id"

scripts/ollama_residency.py

@@ -0,0 +1,121 @@
+"""
+Guard production's GPU residency across a benchmark run.
+
+Benchmarks swap models on the card production is serving from. Ollama evicts to
+make room, so a run leaves its own models resident and the production one gone:
+the next voice turn pays a ~36s cold load, and the pin that prevented it is
+silently lost. That happened on 2026-08-08 — a routing benchmark evicted
+gemma4:e2b and left gemma4:e4b behind, and only the monitoring noticing
+`unexpected_models` caught it.
+
+Snapshot before, restore after, and wire the restore to SIGTERM as well as the
+normal path. Python runs `finally` for SIGINT, which arrives as
+KeyboardInterrupt, but the default SIGTERM action terminates outright — so
+`timeout`, a systemd stop or a plain `kill` would skip the guard entirely.
+
+    from scripts.ollama_residency import residency_guard, install_sigterm_handler
+
+    install_sigterm_handler()
+    with residency_guard(models_used=["gemma4:e4b"]):
+        ...
+"""
+from __future__ import annotations
+
+import signal
+from collections.abc import Iterator
+from contextlib import contextmanager
+from datetime import UTC, datetime
+from typing import Any
+
+import httpx
+
+OLLAMA_URL = "http://localhost:11434"
+
+# keep_alive:-1 yields a year-2318 expiry, so "pinned" is simply "expires more
+# than a day out". Matches check-ai-pipeline.sh in system-admin-toj.
+PINNED_THRESHOLD_SECONDS = 86400
+
+
+def install_sigterm_handler() -> None:
+    """Make SIGTERM raise, so `finally` blocks and context managers still run."""
+    def _raise(signum, _frame):
+        raise KeyboardInterrupt(f"signal {signum}")
+
+    signal.signal(signal.SIGTERM, _raise)
+
+
+def snapshot_residency(client: httpx.Client | None = None) -> dict[str, bool]:
+    """Resident models mapped to whether each is pinned."""
+    owns = client is None
+    client = client or httpx.Client(timeout=30)
+    try:
+        data = client.get(f"{OLLAMA_URL}/api/ps", timeout=10).json()
+    except Exception:  # noqa: BLE001 - a missing snapshot must not abort the run
+        return {}
+    finally:
+        if owns:
+            client.close()
+
+    resident: dict[str, bool] = {}
+    now = datetime.now(UTC)
+    for model in data.get("models", []):
+        pinned = False
+        try:
+            expires = datetime.fromisoformat(model.get("expires_at", "").replace("Z", "+00:00"))
+            pinned = (expires - now).total_seconds() > PINNED_THRESHOLD_SECONDS
+        except ValueError:
+            pass
+        resident[model["name"]] = pinned
+    return resident
+
+
+def set_keep_alive(model: str, keep_alive: Any, client: httpx.Client | None = None) -> bool:
+    """Load, unload or pin a model. Embedding models reject /api/generate."""
+    owns = client is None
+    client = client or httpx.Client(timeout=180)
+    payload = {"model": model, "keep_alive": keep_alive}
+    try:
+        for endpoint in ("generate", "embed"):
+            try:
+                response = client.post(f"{OLLAMA_URL}/api/{endpoint}", json=payload, timeout=180)
+            except Exception:  # noqa: BLE001
+                return False
+            if response.status_code == 200:
+                return True
+            if response.status_code == 400 and "does not support generate" in response.text:
+                continue  # embedding-only model; try /api/embed
+            return False
+        return False
+    finally:
+        if owns:
+            client.close()
+
+
+def restore_residency(before: dict[str, bool], used: list[str]) -> None:
+    """Evict what the benchmark loaded, then re-pin what was pinned before."""
+    base = {name.split(":")[0] for name in before}
+    with httpx.Client(timeout=180) as client:
+        for model in used:
+            if model not in before and model.split(":")[0] not in base:
+                print(f"  residency: unloading benchmark model {model}")
+                set_keep_alive(model, 0, client)
+        for name, pinned in before.items():
+            if not pinned:
+                continue
+            ok = set_keep_alive(name, -1, client)
+            print(f"  residency: re-pinned {name}" if ok
+                  else f"  residency: FAILED to re-pin {name} -- run warmup-ollama.sh")
+
+
+@contextmanager
+def residency_guard(models_used: list[str]) -> Iterator[dict[str, bool]]:
+    """Snapshot residency on entry, restore it on exit however that happens."""
+    before = snapshot_residency()
+    pinned = [n for n, p in before.items() if p]
+    print(f"  residency: resident before {sorted(before)}"
+          f"{f' (pinned: {pinned})' if pinned else ''}")
+    try:
+        yield before
+    finally:
+        print("  residency: restoring ...")
+        restore_residency(before, models_used)

scripts/test_housekeeper.sh

@@ -0,0 +1,141 @@
+#!/bin/bash
+# Housekeeper Room Group Detection Test Suite
+# Verifies room groups are controlled by checking actual state changes
+
+API_URL="http://localhost:8777/v1/chat/completions"
+CORE_API="http://localhost:8083"
+RESULTS_FILE="/tmp/housekeeper_test_results.txt"
+
+GREEN='\033[0;32m'
+RED='\033[0;31m'
+YELLOW='\033[1;33m'
+NC='\033[0m'
+
+get_state() {
+    curl -s "$CORE_API/housekeeping/devices/$1" 2>/dev/null | jq -r '.state' 2>/dev/null
+}
+
+echo "=========================================="
+echo "Housekeeper Room Group Test Suite"
+echo "=========================================="
+echo ""
+
+> "$RESULTS_FILE"
+
+run_toggle_test() {
+    local test_num=$1
+    local room=$2
+    local entity="light.$room"
+    local prompt_room="${room//_/ }"
+
+    printf "Test %2d: Toggle %-12s lights ... " "$test_num" "$prompt_room"
+
+    local before=$(get_state "$entity")
+    if [ -z "$before" ] || [ "$before" = "null" ]; then
+        echo -e "${YELLOW}SKIP${NC} (cannot get state)"
+        echo "SKIP|$test_num|Toggle $room|error" >> "$RESULTS_FILE"
+        return
+    fi
+
+    curl -s -X POST "$API_URL" \
+        -H "Content-Type: application/json" \
+        -d "{\"model\": \"tatlock\", \"messages\": [{\"role\": \"user\", \"content\": \"Toggle the $prompt_room lights\"}]}" > /dev/null
+
+    sleep 4
+
+    local after=$(get_state "$entity")
+
+    if [ "$before" != "$after" ]; then
+        echo -e "${GREEN}PASS${NC} ($before -> $after)"
+        echo "PASS|$test_num|Toggle $room|$before->$after" >> "$RESULTS_FILE"
+    else
+        echo -e "${RED}FAIL${NC} (state unchanged: $before)"
+        echo "FAIL|$test_num|Toggle $room|unchanged:$before" >> "$RESULTS_FILE"
+    fi
+}
+
+run_onoff_test() {
+    local test_num=$1
+    local room=$2
+    local action=$3
+    local expected_state=$4
+    # Entity uses underscore, prompt uses space
+    local entity="light.${room//_/ }"
+    entity="light.$room"
+    local prompt_room="${room//_/ }"
+
+    printf "Test %2d: %-8s %-12s lights ... " "$test_num" "$action" "$prompt_room"
+
+    curl -s -X POST "$API_URL" \
+        -H "Content-Type: application/json" \
+        -d "{\"model\": \"tatlock\", \"messages\": [{\"role\": \"user\", \"content\": \"$action the $prompt_room lights\"}]}" > /dev/null
+
+    sleep 4
+
+    local after=$(get_state "$entity")
+
+    if [ "$after" = "$expected_state" ]; then
+        echo -e "${GREEN}PASS${NC} ($after)"
+        echo "PASS|$test_num|$action $room|$after" >> "$RESULTS_FILE"
+    else
+        echo -e "${RED}FAIL${NC} (got $after, expected $expected_state)"
+        echo "FAIL|$test_num|$action $room|got:$after,expected:$expected_state" >> "$RESULTS_FILE"
+    fi
+}
+
+echo "Running tests (~4s each)..."
+echo ""
+
+# Study tests
+run_onoff_test 1 "study" "Turn off" "off"
+run_onoff_test 2 "study" "Turn on" "on"
+run_toggle_test 3 "study"
+
+# Kitchen tests
+run_onoff_test 4 "kitchen" "Turn off" "off"
+run_onoff_test 5 "kitchen" "Turn on" "on"
+run_toggle_test 6 "kitchen"
+
+# Bedroom tests
+run_onoff_test 7 "bedroom" "Turn off" "off"
+run_onoff_test 8 "bedroom" "Turn on" "on"
+
+# Living room tests (entity is light.living_room)
+run_onoff_test 9 "living_room" "Turn off" "off"
+run_onoff_test 10 "living_room" "Turn on" "on"
+
+# Ensure all lights end up ON
+echo ""
+echo "Restoring all lights to ON..."
+for room in "study" "kitchen" "bedroom" "living room"; do
+    curl -s -X POST "$API_URL" \
+        -H "Content-Type: application/json" \
+        -d "{\"model\": \"tatlock\", \"messages\": [{\"role\": \"user\", \"content\": \"Turn on the $room lights\"}]}" > /dev/null
+    sleep 3
+done
+echo "Done."
+
+echo ""
+echo "=========================================="
+echo "Results"
+echo "=========================================="
+
+PASS=$(grep -c "^PASS" "$RESULTS_FILE" 2>/dev/null || echo 0)
+FAIL=$(grep -c "^FAIL" "$RESULTS_FILE" 2>/dev/null || echo 0)
+SKIP=$(grep -c "^SKIP" "$RESULTS_FILE" 2>/dev/null || echo 0)
+TOTAL=$((PASS + FAIL))
+
+echo "Passed:  $PASS"
+echo "Failed:  $FAIL"
+echo "Skipped: $SKIP"
+
+if [ "$TOTAL" -gt 0 ]; then
+    echo ""
+    echo "Success Rate: $((PASS * 100 / TOTAL))% ($PASS/$TOTAL)"
+fi
+
+if [ "$FAIL" -gt 0 ]; then
+    echo ""
+    echo "Failures:"
+    grep "^FAIL" "$RESULTS_FILE"
+fi

src/agents/base.py

@@ -6,7 +6,8 @@ must implement. The interface is designed around the Responses API format.
 """
 
 from abc import ABC, abstractmethod
-from typing import AsyncGenerator, Any
+from collections.abc import AsyncGenerator
+from typing import Any
 
 
 class OutputItem:
@@ -19,12 +20,7 @@ class OutputItem:
     - message: Assistant response message
     """
 
-    def __init__(
-        self,
-        type: str,
-        id: str,
-        **kwargs: Any
-    ):
+    def __init__(self, type: str, id: str, **kwargs: Any):
         self.type = type
         self.id = id
         self.data = kwargs
@@ -40,7 +36,7 @@ class AgentInterface(ABC):
     """
 
     @abstractmethod
-    async def generate_response(
+    def generate_response(
         self,
         messages: list[dict],
         reasoning: dict | None = None,
@@ -48,7 +44,7 @@ class AgentInterface(ABC):
         temperature: float = 1.0,
         max_tokens: int | None = None,
         stop: list[str] | None = None,
-        **kwargs: Any
+        **kwargs: Any,
     ) -> AsyncGenerator[OutputItem, None]:
         """
         Generate streaming response as output items.

src/agents/biographer/__init__.py

@@ -11,6 +11,7 @@ For direct key-based lookups (location, timezone, preferences),
 use the memory_service instead - it's faster and doesn't require LLM.
 The Biographer handles semantic, fuzzy queries.
 """
+
 from src.agents.biographer.agent import (
     get_biographer_agent,
     run_biographer,

src/agents/biographer/agent.py

@@ -7,7 +7,8 @@ A PydanticAI agent that serves as the household's memory keeper:
 - Manages user profile and preferences
 - Forgets information when requested
 """
-from typing import Any, Optional
+
+from typing import Any
 
 from pydantic_ai import Agent
 
@@ -19,7 +20,6 @@ from src.agents.biographer.tools import (
     update_preference,
     update_profile,
 )
-from src.core.config import config
 from src.core.logging_config import get_logger
 
 logger = get_logger(__name__)
@@ -97,24 +97,15 @@ When recalling:
 """
 
 # Lazy initialization to avoid connection issues during imports
-_biographer_agent: Optional[Agent[None, str]] = None
+_biographer_agent: Agent[None, str] | None = None
 
 
 def _create_biographer_agent() -> Agent[None, str]:
     """Create The Biographer PydanticAI agent."""
-    # Import required classes for Ollama configuration
-    from pydantic_ai.models.openai import OpenAIChatModel
-    from pydantic_ai.providers.ollama import OllamaProvider
+    from src.anthropic.model_selector import get_model
 
-    # PydanticAI expects Ollama base URL to end with /v1
-    clean_host = str(config.OLLAMA_HOST).rstrip('/')
-    base_url = f"{clean_host}/v1"
-
-    # Create Ollama model with provider
-    model = OpenAIChatModel(
-        model_name=config.OLLAMA_DEFAULT_MODEL,
-        provider=OllamaProvider(base_url=base_url)
-    )
+    # Get best available model (Claude if available, else Ollama)
+    model = get_model()
 
     agent: Agent[None, str] = Agent(
         model=model,
@@ -134,9 +125,13 @@ def _create_biographer_agent() -> Agent[None, str]:
     # Register management tools
     agent.tool_plain(forget_memory)
 
+    from src.anthropic.model_selector import get_model_info
+
+    model_info = get_model_info()
     logger.info(
         "biographer_agent_created",
-        model=config.OLLAMA_DEFAULT_MODEL,
+        backend=model_info["backend"],
+        model=model_info["model"],
         tool_count=6,
     )
 
@@ -159,7 +154,7 @@ def get_biographer_agent() -> Agent[None, str]:
 async def run_biographer(
     task: str,
     context: str = "",
-    message_history: Optional[list[Any]] = None,
+    message_history: list[Any] | None = None,
 ) -> str:
     """
     Execute a memory task with The Biographer.
@@ -222,7 +217,7 @@ async def run_biographer(
 async def run_biographer_stream(
     task: str,
     context: str = "",
-    message_history: Optional[list[Any]] = None,
+    message_history: list[Any] | None = None,
 ):
     """
     Execute a memory task with streaming output.

src/agents/biographer/capability.py

@@ -4,6 +4,7 @@ Biographer capability registration for the Household Registry.
 Defines The Biographer's capabilities and registers it as a
 household member for coordination by the Steward and Tatlock.
 """
+
 from src.agents.biographer.agent import get_biographer_agent
 from src.agents.biographer.tools import BIOGRAPHER_TOOLS
 from src.core.household_registry import (

src/agents/biographer/tools.py

@@ -10,6 +10,7 @@ These tools enable The Biographer to record and recall the user's story:
 For direct key-based access (get/set profile, preferences),
 use memory_service directly - these tools are for semantic queries.
 """
+
 from src.core.context import get_user
 from src.core.embeddings import get_embedding_client
 from src.core.logging_config import get_logger
@@ -23,9 +24,10 @@ logger = get_logger(__name__)
 # Semantic Recall
 # ============================================================================
 
+
 async def recall_semantic(
     query: str,
-    memory_type: str | None = None,
+    memory_type: str = "",
     limit: int = 5,
 ) -> str:
     """
@@ -64,7 +66,7 @@ async def recall_semantic(
             user=user,
             query_vector=query_vector,
             limit=limit,
-            memory_type=memory_type,
+            memory_type=memory_type if memory_type else None,
         )
 
         if not results:
@@ -108,10 +110,10 @@ async def recall_semantic(
 # Store Memory
 # ============================================================================
 
+
 async def store_insight(
     key: str,
     value: str,
-    keywords: list[str] | None = None,
     importance: float = 0.5,
 ) -> str:
     """
@@ -128,7 +130,6 @@ async def store_insight(
     Args:
         key: Short identifier for the memory (e.g., "car", "employer", "pet")
         value: The actual information to remember
-        keywords: Optional keywords for better search (auto-extracted if not provided)
         importance: How important is this? 0.0 (trivial) to 1.0 (critical)
 
     Returns:
@@ -137,15 +138,12 @@ async def store_insight(
     Examples:
         store_insight("car", "User drives a Tesla Model 3")
         store_insight("employer", "Works at Acme Corp as software engineer", importance=0.8)
-        store_insight("coffee", "Prefers oat milk lattes", keywords=["coffee", "drink", "preference"])
     """
     try:
-        # Auto-generate keywords if not provided
-        if not keywords:
-            keywords = [key]
-            # Extract simple keywords from value
-            words = value.lower().split()
-            keywords.extend([w for w in words if len(w) > 4][:5])
+        # Auto-generate keywords from key and value
+        keywords = [key]
+        words = value.lower().split()
+        keywords.extend([w for w in words if len(w) > 4][:5])
 
         success = await memory_service.store_fact(
             key=key,
@@ -163,7 +161,7 @@ async def store_insight(
                 f"**Keywords:** {', '.join(keywords)}",
                 f"**Importance:** {importance:.1f}",
                 "",
-                "_Memory is now searchable via semantic recall._"
+                "_Memory is now searchable via semantic recall._",
             ]
 
             logger.info(
@@ -221,7 +219,7 @@ async def update_profile(
                 "## Profile Updated",
                 f"**{key}:** {value}",
                 "",
-                "_Profile data is automatically included in context._"
+                "_Profile data is automatically included in context._",
             ]
 
             logger.info(
@@ -276,7 +274,7 @@ async def update_preference(
                 "## Preference Updated",
                 f"**{key}:** {value}",
                 "",
-                "_Preference will be applied to future responses._"
+                "_Preference will be applied to future responses._",
             ]
 
             logger.info(
@@ -298,6 +296,7 @@ async def update_preference(
 # List Memories
 # ============================================================================
 
+
 async def list_memories(
     memory_type: str = "learned_fact",
     limit: int = 20,
@@ -326,7 +325,7 @@ async def list_memories(
 
         # Convert string to MemoryType
         try:
-            mem_type = MemoryType(memory_type)
+            MemoryType(memory_type)  # validated for its ValueError; the value is unused
         except ValueError:
             return f"Invalid memory type '{memory_type}'. Use: user_profile, preference, or learned_fact"
 
@@ -383,6 +382,7 @@ async def list_memories(
 # Forget Memory
 # ============================================================================
 
+
 async def forget_memory(
     key: str,
     memory_type: str = "learned_fact",
@@ -425,7 +425,7 @@ async def forget_memory(
                 f"**Key:** {key}",
                 f"**Type:** {memory_type}",
                 "",
-                "_Memory has been removed._"
+                "_Memory has been removed._",
             ]
 
             logger.info(

src/agents/coordination.py

@@ -1,407 +0,0 @@
-"""
-Multi-agent coordination engine.
-
-Orchestrates delegation from Tatlock to expert agents (Librarian, etc.)
-based on Steward recommendations. Handles:
-- Routing tasks to appropriate agents
-- Parallel and sequential execution
-- Result aggregation
-- Error handling and graceful degradation
-"""
-import asyncio
-import time
-from typing import Any, AsyncGenerator, Optional
-
-from src.agents.librarian import run_librarian, run_librarian_stream
-from src.agents.protocol import (
-    AgentError,
-    AgentRequest,
-    AgentResponse,
-    AgentTimeoutError,
-    AgentUnavailableError,
-    CoordinationResult,
-    DelegationIntent,
-    DelegationReason,
-    ToolCallRecord,
-)
-from src.core.household_registry import get_household_registry
-from src.core.logging_config import get_logger
-
-logger = get_logger(__name__)
-
-
-# Agent execution functions registry
-AGENT_EXECUTORS: dict[str, Any] = {
-    "librarian": run_librarian,
-}
-
-AGENT_STREAM_EXECUTORS: dict[str, Any] = {
-    "librarian": run_librarian_stream,
-}
-
-
-class CoordinationEngine:
-    """
-    Coordinates multi-agent task execution.
-
-    Routes tasks from Tatlock to appropriate expert agents,
-    handles execution, and aggregates results.
-    """
-
-    def __init__(self):
-        """Initialize the coordination engine."""
-        self.registry = get_household_registry()
-        logger.info("coordination_engine_initialized")
-
-    def get_available_agents(self) -> list[str]:
-        """
-        Get list of available expert agents.
-
-        Returns:
-            List of agent names that can accept delegations
-        """
-        available = []
-        for name in self.registry.list_members():
-            member = self.registry.get_member(name)
-            if member and member.agent is not None:
-                available.append(name)
-        return available
-
-    def can_delegate_to(self, agent_name: str) -> bool:
-        """
-        Check if delegation to an agent is possible.
-
-        Args:
-            agent_name: Name of the target agent
-
-        Returns:
-            True if agent is available and can accept tasks
-        """
-        if agent_name not in AGENT_EXECUTORS:
-            return False
-
-        member = self.registry.get_member(agent_name)
-        return member is not None and member.agent is not None
-
-    async def execute_delegation(
-        self,
-        intent: DelegationIntent,
-        context: str = "",
-        message_history: Optional[list[Any]] = None,
-    ) -> AgentResponse:
-        """
-        Execute a single delegation to an expert agent.
-
-        Args:
-            intent: The delegation intent with task details
-            context: Additional context for the agent
-            message_history: Optional conversation history
-
-        Returns:
-            AgentResponse with results
-
-        Raises:
-            AgentUnavailableError: If agent is not available
-            AgentTimeoutError: If execution times out
-            AgentError: For other execution errors
-        """
-        start_time = time.time()
-        agent_name = intent.target_agent
-
-        logger.info(
-            "delegation_started",
-            agent=agent_name,
-            task=intent.task[:100],
-            reason=intent.reason.value,
-        )
-
-        # Check if agent is available
-        if not self.can_delegate_to(agent_name):
-            raise AgentUnavailableError(
-                f"Agent '{agent_name}' is not available for delegation",
-                agent_name=agent_name,
-            )
-
-        # Get the executor
-        executor = AGENT_EXECUTORS.get(agent_name)
-        if not executor:
-            raise AgentUnavailableError(
-                f"No executor found for agent '{agent_name}'",
-                agent_name=agent_name,
-            )
-
-        try:
-            # Build the request
-            request = AgentRequest(
-                task=intent.task,
-                context=context,
-                delegation_reason=intent.reason,
-            )
-
-            # Execute with timeout
-            timeout = request.timeout_seconds or 60
-
-            result = await asyncio.wait_for(
-                executor(
-                    task=request.task,
-                    context=request.context,
-                    message_history=message_history,
-                ),
-                timeout=timeout,
-            )
-
-            duration_ms = int((time.time() - start_time) * 1000)
-
-            logger.info(
-                "delegation_completed",
-                agent=agent_name,
-                duration_ms=duration_ms,
-                output_length=len(result),
-            )
-
-            return AgentResponse(
-                success=True,
-                result=result,
-                reasoning=f"Delegated to {agent_name}: {intent.expected_outcome}",
-                duration_ms=duration_ms,
-            )
-
-        except asyncio.TimeoutError:
-            duration_ms = int((time.time() - start_time) * 1000)
-            logger.error(
-                "delegation_timeout",
-                agent=agent_name,
-                duration_ms=duration_ms,
-            )
-            raise AgentTimeoutError(
-                f"Agent '{agent_name}' timed out after {duration_ms}ms",
-                agent_name=agent_name,
-            )
-
-        except Exception as e:
-            duration_ms = int((time.time() - start_time) * 1000)
-            logger.error(
-                "delegation_error",
-                agent=agent_name,
-                error=str(e),
-                duration_ms=duration_ms,
-                exc_info=True,
-            )
-            return AgentResponse(
-                success=False,
-                result="",
-                error_message=str(e),
-                duration_ms=duration_ms,
-            )
-
-    async def execute_delegation_stream(
-        self,
-        intent: DelegationIntent,
-        context: str = "",
-        message_history: Optional[list[Any]] = None,
-    ) -> AsyncGenerator[str, None]:
-        """
-        Execute a delegation with streaming output.
-
-        Args:
-            intent: The delegation intent with task details
-            context: Additional context for the agent
-            message_history: Optional conversation history
-
-        Yields:
-            Text deltas from the agent
-
-        Raises:
-            AgentUnavailableError: If agent is not available
-        """
-        agent_name = intent.target_agent
-
-        logger.info(
-            "delegation_stream_started",
-            agent=agent_name,
-            task=intent.task[:100],
-        )
-
-        # Check if agent is available
-        if agent_name not in AGENT_STREAM_EXECUTORS:
-            raise AgentUnavailableError(
-                f"Agent '{agent_name}' does not support streaming",
-                agent_name=agent_name,
-            )
-
-        executor = AGENT_STREAM_EXECUTORS[agent_name]
-
-        try:
-            async for delta in executor(
-                task=intent.task,
-                context=context,
-                message_history=message_history,
-            ):
-                yield delta
-
-            logger.info("delegation_stream_completed", agent=agent_name)
-
-        except Exception as e:
-            logger.error(
-                "delegation_stream_error",
-                agent=agent_name,
-                error=str(e),
-                exc_info=True,
-            )
-            yield f"\n\n[Error from {agent_name}: {str(e)}]"
-
-    async def coordinate(
-        self,
-        intents: list[DelegationIntent],
-        context: str = "",
-        message_history: Optional[list[Any]] = None,
-    ) -> CoordinationResult:
-        """
-        Coordinate execution of multiple delegations.
-
-        Handles parallel execution for independent tasks and
-        sequential execution for dependent tasks.
-
-        Args:
-            intents: List of delegation intents to execute
-            context: Shared context for all agents
-            message_history: Optional conversation history
-
-        Returns:
-            CoordinationResult with aggregated results
-        """
-        start_time = time.time()
-        agent_responses: dict[str, AgentResponse] = {}
-        agents_consulted: list[str] = []
-
-        logger.info(
-            "coordination_started",
-            intent_count=len(intents),
-            agents=[i.target_agent for i in intents],
-        )
-
-        # Sort by priority
-        sorted_intents = sorted(intents, key=lambda x: x.priority)
-
-        # Group by dependencies (simple version: sequential for now)
-        # TODO: Implement parallel execution for independent tasks
-        for intent in sorted_intents:
-            try:
-                response = await self.execute_delegation(
-                    intent=intent,
-                    context=context,
-                    message_history=message_history,
-                )
-                agent_responses[intent.target_agent] = response
-                if response.success:
-                    agents_consulted.append(intent.target_agent)
-
-            except AgentError as e:
-                agent_responses[intent.target_agent] = AgentResponse(
-                    success=False,
-                    result="",
-                    error_message=str(e),
-                )
-
-        # Aggregate results
-        successful_results = [
-            r.result for r in agent_responses.values() if r.success and r.result
-        ]
-
-        final_response = "\n\n---\n\n".join(successful_results) if successful_results else ""
-
-        total_duration = int((time.time() - start_time) * 1000)
-
-        logger.info(
-            "coordination_completed",
-            total_duration_ms=total_duration,
-            agents_consulted=agents_consulted,
-            success_count=len(successful_results),
-        )
-
-        return CoordinationResult(
-            final_response=final_response,
-            agent_responses=agent_responses,
-            delegation_intents=intents,
-            total_duration_ms=total_duration,
-            agents_consulted=agents_consulted,
-        )
-
-
-# Global coordination engine instance
-_coordination_engine: Optional[CoordinationEngine] = None
-
-
-def get_coordination_engine() -> CoordinationEngine:
-    """Get the global coordination engine instance."""
-    global _coordination_engine
-    if _coordination_engine is None:
-        _coordination_engine = CoordinationEngine()
-    return _coordination_engine
-
-
-async def delegate_to_librarian(
-    task: str,
-    context: str = "",
-    reason: DelegationReason = DelegationReason.DOMAIN_EXPERTISE,
-    message_history: Optional[list[Any]] = None,
-) -> AgentResponse:
-    """
-    Convenience function to delegate a task to The Librarian.
-
-    Args:
-        task: Research task description
-        context: Additional context
-        reason: Why delegating to Librarian
-        message_history: Optional conversation history
-
-    Returns:
-        AgentResponse with research results
-    """
-    engine = get_coordination_engine()
-
-    intent = DelegationIntent(
-        target_agent="librarian",
-        task=task,
-        reason=reason,
-        expected_outcome="Research findings and relevant information",
-    )
-
-    return await engine.execute_delegation(
-        intent=intent,
-        context=context,
-        message_history=message_history,
-    )
-
-
-async def delegate_to_librarian_stream(
-    task: str,
-    context: str = "",
-    message_history: Optional[list[Any]] = None,
-) -> AsyncGenerator[str, None]:
-    """
-    Convenience function to delegate to Librarian with streaming.
-
-    Args:
-        task: Research task description
-        context: Additional context
-        message_history: Optional conversation history
-
-    Yields:
-        Text deltas from The Librarian
-    """
-    engine = get_coordination_engine()
-
-    intent = DelegationIntent(
-        target_agent="librarian",
-        task=task,
-        reason=DelegationReason.DOMAIN_EXPERTISE,
-        expected_outcome="Research findings",
-    )
-
-    async for delta in engine.execute_delegation_stream(
-        intent=intent,
-        context=context,
-        message_history=message_history,
-    ):
-        yield delta

src/agents/delegation.py

@@ -8,14 +8,189 @@ returns a structured result for synthesis.
 This implements the agent-as-tool pattern recommended by PydanticAI:
 agents call other agents via tool wrappers, keeping each agent focused.
 """
-from dataclasses import dataclass, field
-from typing import Callable, Optional, Any
 
+import asyncio
+from dataclasses import dataclass, field
+from enum import Enum
+
+from src.core.config import config
 from src.core.logging_config import get_logger
+from src.core.tracing import SpanType, trace_span
 
 logger = get_logger(__name__)
 
 
+# =============================================================================
+# Action Types for Think Slug Selection
+# =============================================================================
+
+
+class ActionType(Enum):
+    """
+    Categories of actions for selecting appropriate think messages.
+
+    Each expert has different action types that warrant different
+    butler-perspective messages to the user.
+    """
+
+    RETRIEVE = "retrieve"  # Looking up existing information
+    RESEARCH = "research"  # Conducting new research (web search, etc.)
+    CREATE = "create"  # Creating new content (pages, notes)
+    CONTROL = "control"  # Controlling devices/automations
+    RECORD = "record"  # Recording memories/notes
+
+
+# =============================================================================
+# Household Think Messages (Butler's Perspective)
+# =============================================================================
+
+HOUSEHOLD_THINK_MESSAGES: dict[str, dict[ActionType, dict[str, str]]] = {
+    # Note: No <think> wrappers needed - these go to reasoning_content field
+    "librarian": {
+        ActionType.RETRIEVE: {
+            "start": "Allow me to consult the archives, sir.",
+            "success": "The Librarian has compiled the relevant findings.",
+            "error": "I'm afraid the archives proved difficult to access.",
+        },
+        ActionType.RESEARCH: {
+            "start": "I've dispatched the Librarian to conduct some fresh research.",
+            "success": "The Librarian has returned with findings, sir.",
+            "error": "The research proved inconclusive, I'm afraid.",
+        },
+        ActionType.CREATE: {
+            "start": "I'm having the Librarian prepare a new entry.",
+            "success": "The new material has been properly catalogued, sir.",
+            "error": "I'm afraid there was difficulty filing the entry.",
+        },
+    },
+    "biographer": {
+        ActionType.RETRIEVE: {
+            "start": "Let me consult the household records.",
+            "success": "The Biographer has located the relevant information, sir.",
+            "error": "I'm unable to locate those particular records.",
+        },
+        ActionType.RECORD: {
+            "start": "I've asked the Biographer to take note of this, sir.",
+            "success": "The household records have been updated accordingly.",
+            "error": "I'm afraid there was difficulty recording the entry.",
+        },
+    },
+    "housekeeper": {
+        ActionType.RETRIEVE: {
+            "start": "Allow me to inquire with the household staff.",
+            "success": "The staff reports the current status, sir.",
+            "error": "The household staff is momentarily unavailable, I'm afraid.",
+        },
+        ActionType.CONTROL: {
+            "start": "I'm instructing the household staff now, sir.",
+            "success": "The household has been configured as requested.",
+            "error": "I'm afraid the staff reports an issue with that request.",
+        },
+    },
+}
+
+
+def _detect_action_type(expert: str, task: str) -> ActionType:
+    """
+    Detect action type from expert name and task description.
+
+    Used to select appropriate butler-perspective think messages.
+
+    Args:
+        expert: Name of the expert (librarian, biographer, housekeeper)
+        task: Task description
+
+    Returns:
+        ActionType: Detected action type for message selection
+    """
+    task_lower = task.lower()
+
+    if expert == "librarian":
+        # Web search, URL reading = RESEARCH (fresh external data)
+        if any(w in task_lower for w in ["search", "find", "look up", "research"]):
+            if any(w in task_lower for w in ["web", "online", "internet"]):
+                return ActionType.RESEARCH
+            return ActionType.RETRIEVE
+        if any(w in task_lower for w in ["read", "fetch", "url", "http"]):
+            return ActionType.RESEARCH  # Reading URLs is research
+        if any(w in task_lower for w in ["create", "write", "add", "make", "new"]):
+            return ActionType.CREATE
+        return ActionType.RETRIEVE
+
+    elif expert == "biographer":
+        if any(w in task_lower for w in ["remember", "note", "record", "save", "store"]):
+            return ActionType.RECORD
+        return ActionType.RETRIEVE
+
+    elif expert == "housekeeper":
+        if any(w in task_lower for w in ["turn", "set", "activate", "enable", "disable", "toggle"]):
+            return ActionType.CONTROL
+        return ActionType.RETRIEVE
+
+    return ActionType.RETRIEVE
+
+
+def build_delegation_context(
+    conversation_history: list[dict] | None,
+    max_turns: int = 6,
+    max_chars_per_turn: int = 500,
+) -> str:
+    """
+    Format the most recent conversation turns as delegation context.
+
+    Experts accept a context string but the live paths never passed the
+    in-scope conversation history; this trims it to the last few turns
+    so follow-up questions ("and what about X?") keep their referent.
+
+    Args:
+        conversation_history: Prior messages as {"role", "content"} dicts
+        max_turns: How many trailing turns to include
+        max_chars_per_turn: Truncation limit per turn
+
+    Returns:
+        str: Newline-joined "role: content" lines ("" when no history)
+    """
+    if not conversation_history:
+        return ""
+
+    lines = []
+    for msg in conversation_history[-max_turns:]:
+        if not isinstance(msg, dict):
+            continue
+        role = msg.get("role", "user")
+        content = msg.get("content", "")
+        if isinstance(content, list):
+            # Tolerate structured content parts
+            content = " ".join(
+                part.get("text", "") if isinstance(part, dict) else str(part) for part in content
+            )
+        content = str(content).strip()
+        if content:
+            lines.append(f"{role}: {content[:max_chars_per_turn]}")
+
+    if not lines:
+        return ""
+    return "Recent conversation:\n" + "\n".join(lines)
+
+
+def get_think_message(expert: str, task: str, phase: str) -> str:
+    """
+    Get the appropriate think message for an expert delegation.
+
+    Args:
+        expert: Name of the expert
+        task: Task description (used to detect action type)
+        phase: One of "start", "success", "error"
+
+    Returns:
+        str: Butler-perspective think message
+    """
+    action_type = _detect_action_type(expert, task)
+    expert_messages = HOUSEHOLD_THINK_MESSAGES.get(expert, {})
+    action_messages = expert_messages.get(action_type, expert_messages.get(ActionType.RETRIEVE, {}))
+    return action_messages.get(phase, f"Consulting {expert}...")
+
+
 @dataclass
 class DelegationTask:
     """
@@ -33,19 +208,21 @@ class DelegationTask:
         depends_on: List of task IDs this task depends on
         result: Result from expert after execution
     """
+
     expert_name: str
     task: str
     context: str = ""
     action: str = ""
     priority: int = 0
     depends_on: list[str] = field(default_factory=list)
-    result: Optional[str] = None
+    result: str | None = None
     task_id: str = ""
 
-    def __post_init__(self):
+    def __post_init__(self) -> None:
         """Generate task ID if not provided."""
         if not self.task_id:
             import uuid
+
             self.task_id = f"{self.expert_name}_{uuid.uuid4().hex[:8]}"
 
 
@@ -58,14 +235,17 @@ class DelegationResult:
         expert_name: Which expert handled the task
         task: Original task description
         success: Whether the delegation succeeded
-        output: Expert's response/findings
-        error: Error message if failed
+        output: Expert's response/findings. On failure this holds a
+            curated, user-safe butler sentence (never exception detail)
+        error: Short user-safe error label if failed. Exception detail
+            stays in the logs only
     """
+
     expert_name: str
     task: str
     success: bool
     output: str
-    error: Optional[str] = None
+    error: str | None = None
 
 
 async def delegate_to_librarian(
@@ -112,38 +292,89 @@ async def delegate_to_librarian(
         has_context=bool(context),
     )
 
-    try:
-        # Use run() not run_stream() - avoids Ollama bug
-        output = await run_librarian(task=task, context=context)
+    async with trace_span(
+        "delegate_to_librarian",
+        SpanType.EXPERT,
+        metadata={
+            "expert": "librarian",
+            "task_preview": task[:100],
+            "has_context": bool(context),
+        },
+    ) as span:
+        try:
+            # Use run() not run_stream() - avoids Ollama bug.
+            # One timeout budget for the whole delegation - covers both
+            # live paths (steward direct delegation and streaming), which
+            # previously had no cap at all (SDK default ~600s per LLM call).
+            output = await asyncio.wait_for(
+                run_librarian(task=task, context=context),
+                timeout=config.LIBRARIAN_TIMEOUT,
+            )
 
-        logger.info(
-            "delegation_to_librarian_completed",
-            task=task[:50],
-            output_length=len(output),
-        )
+            logger.info(
+                "delegation_to_librarian_completed",
+                task=task[:50],
+                output_length=len(output),
+            )
 
-        return DelegationResult(
-            expert_name="librarian",
-            task=task,
-            success=True,
-            output=output,
-        )
+            if span:
+                span.metadata["success"] = True
+                span.metadata["output_length"] = len(output)
+                span.details["task"] = task
+                span.details["context"] = context[:500] if context else None
+                span.details["result_preview"] = output[:1000]
 
-    except Exception as e:
-        logger.error(
-            "delegation_to_librarian_error",
-            task=task[:50],
-            error=str(e),
-            exc_info=True,
-        )
+            return DelegationResult(
+                expert_name="librarian",
+                task=task,
+                success=True,
+                output=output,
+            )
 
-        return DelegationResult(
-            expert_name="librarian",
-            task=task,
-            success=False,
-            output="",
-            error=str(e),
-        )
+        except TimeoutError:
+            logger.error(
+                "delegation_to_librarian_timeout",
+                task=task[:50],
+                timeout_seconds=config.LIBRARIAN_TIMEOUT,
+            )
+
+            if span:
+                span.metadata["success"] = False
+                span.details["error"] = f"timed out after {config.LIBRARIAN_TIMEOUT}s"
+
+            return DelegationResult(
+                expert_name="librarian",
+                task=task,
+                success=False,
+                output=(
+                    "I'm afraid the research took longer than expected "
+                    "and had to be abandoned, sir."
+                ),
+                error="The Librarian did not respond within the time budget.",
+            )
+
+        except Exception as e:
+            logger.error(
+                "delegation_to_librarian_error",
+                task=task[:50],
+                error=str(e),
+                exc_info=True,
+            )
+
+            if span:
+                span.metadata["success"] = False
+                span.details["error"] = str(e)
+
+            # Exception detail stays in the logs; the user-facing output
+            # is a curated butler sentence so internals never leak into
+            # synthesis.
+            return DelegationResult(
+                expert_name="librarian",
+                task=task,
+                success=False,
+                output=get_think_message("librarian", task, "error"),
+                error="The Librarian was unable to complete the task.",
+            )
 
 
 async def delegate_to_biographer(
@@ -190,40 +421,159 @@ async def delegate_to_biographer(
         has_context=bool(context),
     )
 
-    try:
-        # Use run() not run_stream() - avoids Ollama bug
-        output = await run_biographer(task=task, context=context)
+    async with trace_span(
+        "delegate_to_biographer",
+        SpanType.EXPERT,
+        metadata={
+            "expert": "biographer",
+            "task_preview": task[:100],
+            "has_context": bool(context),
+        },
+    ) as span:
+        try:
+            # Use run() not run_stream() - avoids Ollama bug
+            output = await run_biographer(task=task, context=context)
 
-        logger.info(
-            "delegation_to_biographer_completed",
-            task=task[:50],
-            output_length=len(output),
-        )
+            logger.info(
+                "delegation_to_biographer_completed",
+                task=task[:50],
+                output_length=len(output),
+            )
 
-        return DelegationResult(
-            expert_name="biographer",
-            task=task,
-            success=True,
-            output=output,
-        )
+            if span:
+                span.metadata["success"] = True
+                span.metadata["output_length"] = len(output)
+                span.details["task"] = task
+                span.details["context"] = context[:500] if context else None
+                span.details["result_preview"] = output[:1000]
 
-    except Exception as e:
-        logger.error(
-            "delegation_to_biographer_error",
-            task=task[:50],
-            error=str(e),
-            exc_info=True,
-        )
+            return DelegationResult(
+                expert_name="biographer",
+                task=task,
+                success=True,
+                output=output,
+            )
 
-        return DelegationResult(
-            expert_name="biographer",
-            task=task,
-            success=False,
-            output="",
-            error=str(e),
-        )
+        except Exception as e:
+            logger.error(
+                "delegation_to_biographer_error",
+                task=task[:50],
+                error=str(e),
+                exc_info=True,
+            )
+
+            if span:
+                span.metadata["success"] = False
+                span.details["error"] = str(e)
+
+            # Exception detail stays in the logs only.
+            return DelegationResult(
+                expert_name="biographer",
+                task=task,
+                success=False,
+                output=get_think_message("biographer", task, "error"),
+                error="The Biographer was unable to complete the task.",
+            )
+
+
+async def delegate_to_housekeeper(
+    task: str,
+    context: str = "",
+) -> DelegationResult:
+    """
+    Delegate a home automation task to The Housekeeper.
+
+    The Housekeeper handles:
+    - Device control (turn on/off, toggle, brightness, color)
+    - Scene activation (movie night, good morning, etc.)
+    - Script execution (automation sequences)
+    - Automation management (enable/disable rules)
+    - Device discovery (list devices by area/type)
+    - State queries (get current state, history)
+
+    Args:
+        task: Clear description of what needs to be done.
+              Include the action verb (turn on, activate, list, etc.)
+              Example: "Turn on the living room lights"
+              Example: "Activate the movie night scene"
+              Example: "What devices are in the bedroom?"
+        context: Additional context from the user's request or
+                 conversation history
+
+    Returns:
+        DelegationResult with The Housekeeper's response
+
+    Example:
+        >>> result = await delegate_to_housekeeper(
+        ...     task="Turn on the bedroom lights at 50% brightness",
+        ...     context="User is getting ready for bed",
+        ... )
+        >>> if result.success:
+        ...     print(result.output)
+    """
+    from src.agents.housekeeper.agent import run_housekeeper
+
+    logger.info(
+        "delegation_to_housekeeper_started",
+        task=task[:100],
+        has_context=bool(context),
+    )
+
+    async with trace_span(
+        "delegate_to_housekeeper",
+        SpanType.EXPERT,
+        metadata={
+            "expert": "housekeeper",
+            "task_preview": task[:100],
+            "has_context": bool(context),
+        },
+    ) as span:
+        try:
+            # Use run() not run_stream() - avoids Ollama bug
+            output = await run_housekeeper(task=task, context=context)
+
+            logger.info(
+                "delegation_to_housekeeper_completed",
+                task=task[:50],
+                output_length=len(output),
+            )
+
+            if span:
+                span.metadata["success"] = True
+                span.metadata["output_length"] = len(output)
+                span.details["task"] = task
+                span.details["context"] = context[:500] if context else None
+                span.details["result_preview"] = output[:1000]
+
+            return DelegationResult(
+                expert_name="housekeeper",
+                task=task,
+                success=True,
+                output=output,
+            )
+
+        except Exception as e:
+            logger.error(
+                "delegation_to_housekeeper_error",
+                task=task[:50],
+                error=str(e),
+                exc_info=True,
+            )
+
+            if span:
+                span.metadata["success"] = False
+                span.details["error"] = str(e)
+
+            # Exception detail stays in the logs only.
+            return DelegationResult(
+                expert_name="housekeeper",
+                task=task,
+                success=False,
+                output=get_think_message("housekeeper", task, "error"),
+                error="The Housekeeper was unable to complete the task.",
+            )
 
 
 # Future expert delegation wrappers will be added here:
-# - delegate_to_home_automation(task, context) -> DelegationResult
 # - delegate_to_developer(task, context) -> DelegationResult
+# - delegate_to_secretary(task, context) -> DelegationResult

src/agents/housekeeper/__init__.py

@@ -0,0 +1,25 @@
+"""
+The Housekeeper - Home Automation Agent.
+
+Provides home automation capabilities through the core-api service,
+which wraps the Home Assistant REST API into LLM-friendly endpoints.
+"""
+
+from src.agents.housekeeper.agent import run_housekeeper, run_housekeeper_stream
+from src.agents.housekeeper.capability import (
+    HOUSEKEEPER_CAPABILITY,
+    register_housekeeper,
+)
+from src.agents.housekeeper.client import CoreAPIClient, get_core_api_client
+
+__all__ = [
+    # Agent entry points
+    "run_housekeeper",
+    "run_housekeeper_stream",
+    # Capability
+    "HOUSEKEEPER_CAPABILITY",
+    "register_housekeeper",
+    # Client
+    "CoreAPIClient",
+    "get_core_api_client",
+]

src/agents/housekeeper/agent.py

@@ -0,0 +1,290 @@
+"""
+The Housekeeper - Expert agent for home automation.
+
+A PydanticAI agent that provides home automation capabilities through
+the core-api service, which wraps Home Assistant REST API, offering:
+- Device discovery and control
+- Scene activation
+- Script execution
+- Automation management
+"""
+
+from typing import Any
+
+from pydantic_ai import Agent
+
+from src.agents.housekeeper.tools import (
+    activate_scene,
+    get_device_state,
+    get_history,
+    list_areas,
+    list_automations,
+    list_devices,
+    list_scenes,
+    list_scripts,
+    run_script,
+    toggle,
+    toggle_automation,
+    turn_off,
+    turn_on,
+)
+from src.core.logging_config import get_logger
+
+logger = get_logger(__name__)
+
+# Housekeeper system prompt - Optimized for Mistral-Nemo function calling
+HOUSEKEEPER_SYSTEM_PROMPT = """You are a strictly tool-based home automation assistant.
+
+## CRITICAL: You Have NO Internal Knowledge
+
+You do NOT know what devices exist. You do NOT know any entity IDs.
+Entity IDs are different in every installation. You MUST discover them using tools.
+
+## Entity ID Format
+
+Entity IDs follow the format: `domain.name`
+Examples: `light.kitchen`, `light.study_main`, `switch.coffee_maker`
+
+The `entity_id` parameter MUST be the COMPLETE value including the domain prefix.
+WRONG: `entity_id="kitchen"`
+RIGHT: `entity_id="light.kitchen"`
+
+## Step-by-Step Process (ALWAYS FOLLOW)
+
+When asked to control devices in a room:
+
+1. THINK: What domain? (light, switch, climate, etc.)
+2. CALL: list_devices(domain="light") to discover available devices
+3. CHECK: Look for EXACT match `light.<room_name>` first!
+   - For "study lights" → look for `light.study` (not light.study_main, not light.studeerlamp)
+   - For "kitchen lights" → look for `light.kitchen` (not light.kitchen_spot_1)
+   - These room groups control ALL lights in that room at once
+   - If found, use ONLY the group (stop looking for individual lights)
+4. FALLBACK: Only if no exact room group exists, find entity_ids containing the room name
+5. CALL: turn_on/turn_off using the EXACT entity_id from step 3 or 4
+
+Example for "Turn off study lights":
+1. Domain is "light"
+2. Call list_devices(domain="light")
+3. Look for room group: `light.study` - FOUND!
+4. Call turn_off(entity_id="light.study")  # This controls all study lights
+
+Example for "Turn off hallway lights" (no room group):
+1. Domain is "light"
+2. Call list_devices(domain="light")
+3. Look for room group: `light.hallway` - NOT FOUND
+4. Find all with "hallway": light.hallway_spot_1, light.hallway_spot_2
+5. Call turn_off for each
+
+## Tool Parameter Names
+
+- turn_on, turn_off, toggle: Use `entity_id` (NOT device_id, NOT id)
+- activate_scene: Use `scene_id`
+- run_script: Use `script_id`
+
+## What NOT To Do
+
+- NEVER guess an entity_id
+- NEVER construct an entity_id from the room name
+- NEVER drop the domain prefix (light., switch., etc.)
+- NEVER use "device_id" - the parameter is called "entity_id"
+- NEVER provide an answer without calling list_devices first
+
+## Response Format
+
+After completing actions, briefly confirm:
+- Which devices were affected (list the entity_ids)
+- Whether each action succeeded or failed
+"""
+
+# Lazy initialization to avoid connection issues during imports
+_housekeeper_agent: Agent[None, str] | None = None
+
+
+def _create_housekeeper_agent() -> Agent[None, str]:
+    """Create the Housekeeper PydanticAI agent."""
+    from src.anthropic.model_selector import get_model
+
+    # Get best available model (Claude if available, else Ollama)
+    model = get_model()
+
+    agent: Agent[None, str] = Agent(
+        model=model,
+        system_prompt=HOUSEKEEPER_SYSTEM_PROMPT,
+        retries=2,
+    )
+
+    # Register discovery tools
+    agent.tool_plain(list_areas)
+    agent.tool_plain(list_devices)
+    agent.tool_plain(get_device_state)
+
+    # Register control tools
+    agent.tool_plain(turn_on)
+    agent.tool_plain(turn_off)
+    agent.tool_plain(toggle)
+
+    # Register scene tools
+    agent.tool_plain(list_scenes)
+    agent.tool_plain(activate_scene)
+
+    # Register script tools
+    agent.tool_plain(list_scripts)
+    agent.tool_plain(run_script)
+
+    # Register automation tools
+    agent.tool_plain(list_automations)
+    agent.tool_plain(toggle_automation)
+
+    # Register history tools
+    agent.tool_plain(get_history)
+
+    from src.anthropic.model_selector import get_model_info
+
+    model_info = get_model_info()
+    logger.info(
+        "housekeeper_agent_created",
+        backend=model_info["backend"],
+        model=model_info["model"],
+        tool_count=13,
+    )
+
+    return agent
+
+
+def get_housekeeper_agent() -> Agent[None, str]:
+    """
+    Get the Housekeeper agent instance (lazy initialization).
+
+    Returns:
+        PydanticAI Agent configured for home automation tasks
+    """
+    global _housekeeper_agent
+    if _housekeeper_agent is None:
+        _housekeeper_agent = _create_housekeeper_agent()
+    return _housekeeper_agent
+
+
+async def run_housekeeper(
+    task: str,
+    context: str = "",
+    message_history: list[Any] | None = None,
+) -> str:
+    """
+    Execute a home automation task with The Housekeeper.
+
+    This is the main entry point for delegating home automation tasks
+    to The Housekeeper from Tatlock or other agents.
+
+    Args:
+        task: The home automation task or request
+        context: Additional context from conversation
+        message_history: Optional conversation history
+
+    Returns:
+        Results and confirmation of actions
+
+    Example:
+        result = await run_housekeeper(
+            task="Turn on the living room lights",
+            context="It's evening",
+        )
+    """
+    agent = get_housekeeper_agent()
+
+    # Build prompt with context if provided
+    prompt = task
+    if context:
+        prompt = f"Context: {context}\n\nTask: {task}"
+
+    logger.info(
+        "housekeeper_task_started",
+        task=task[:100],
+        has_context=bool(context),
+        has_history=bool(message_history),
+    )
+
+    try:
+        # Temperature 0.1 for slight exploration (skipped on Claude backend)
+        from src.anthropic.model_selector import get_sampling_settings
+
+        result = await agent.run(
+            prompt,
+            message_history=message_history,
+            model_settings=get_sampling_settings(0.1),
+        )
+
+        logger.info(
+            "housekeeper_task_completed",
+            task=task[:50],
+            output_length=len(result.output),
+        )
+
+        return result.output
+
+    except Exception as e:
+        logger.error(
+            "housekeeper_task_error",
+            task=task[:50],
+            error=str(e),
+            exc_info=True,
+        )
+        return f"The Housekeeper encountered an error: {str(e)}"
+
+
+async def run_housekeeper_stream(
+    task: str,
+    context: str = "",
+    message_history: list[Any] | None = None,
+):
+    """
+    Execute a home automation task with streaming output.
+
+    Yields text deltas as The Housekeeper generates the response.
+
+    Args:
+        task: The home automation task or request
+        context: Additional context from conversation
+        message_history: Optional conversation history
+
+    Yields:
+        str: Text deltas from the response
+
+    Example:
+        async for delta in run_housekeeper_stream("Turn on the lights"):
+            print(delta, end="", flush=True)
+    """
+    agent = get_housekeeper_agent()
+
+    # Build prompt with context if provided
+    prompt = task
+    if context:
+        prompt = f"Context: {context}\n\nTask: {task}"
+
+    logger.info(
+        "housekeeper_stream_started",
+        task=task[:100],
+    )
+
+    try:
+        # Temperature 0.1 for slight exploration (skipped on Claude backend)
+        from src.anthropic.model_selector import get_sampling_settings
+
+        async with agent.run_stream(
+            prompt,
+            message_history=message_history,
+            model_settings=get_sampling_settings(0.1),
+        ) as response:
+            async for delta in response.stream_text(delta=True):
+                yield delta
+
+        logger.info("housekeeper_stream_completed", task=task[:50])
+
+    except Exception as e:
+        logger.error(
+            "housekeeper_stream_error",
+            task=task[:50],
+            error=str(e),
+            exc_info=True,
+        )
+        yield f"\n\nThe Housekeeper encountered an error: {str(e)}"

src/agents/housekeeper/capability.py

@@ -0,0 +1,91 @@
+"""
+Housekeeper capability registration for the Household Registry.
+
+Defines The Housekeeper's capabilities and registers it as a
+household member for coordination by the Steward and Tatlock.
+"""
+
+from src.agents.housekeeper.agent import get_housekeeper_agent
+from src.agents.housekeeper.tools import HOUSEKEEPER_TOOLS
+from src.core.household_registry import (
+    HouseholdCapability,
+    get_household_registry,
+)
+from src.core.logging_config import get_logger
+
+logger = get_logger(__name__)
+
+
+# The Housekeeper's capability summary for Steward coordination
+HOUSEKEEPER_CAPABILITY = HouseholdCapability(
+    name="housekeeper",
+    role="The Housekeeper",
+    category="automation",
+    description=(
+        "Home automation control: TURN ON/OFF devices, ACTIVATE scenes, "
+        "RUN scripts, LIST devices, MANAGE automations. Controls lights, "
+        "switches, climate, and other smart home devices via Home Assistant."
+    ),
+    domains=[
+        "lights",
+        "switches",
+        "automation",
+        "home",
+        "smart home",
+        "scene",
+        "script",
+        "device",
+        "turn on",
+        "turn off",
+        "temperature",
+        "climate",
+        "fan",
+        "cover",
+        "blinds",
+    ],
+    cost="low",  # Fast local API calls to core-api
+    requires_network=True,  # Needs core-api access
+)
+
+
+def get_housekeeper_capability() -> HouseholdCapability:
+    """Get The Housekeeper's capability definition."""
+    return HOUSEKEEPER_CAPABILITY
+
+
+def register_housekeeper() -> None:
+    """
+    Register The Housekeeper with the Household Registry.
+
+    This makes The Housekeeper available for:
+    - Steward recommendations (via capability summary)
+    - Tatlock delegation (via agent reference)
+    - Tool scoping (via tool list)
+    """
+    registry = get_household_registry()
+
+    # Check if already registered
+    if "housekeeper" in registry:
+        logger.debug("housekeeper_already_registered")
+        return
+
+    registry.register(
+        name="housekeeper",
+        capability=HOUSEKEEPER_CAPABILITY,
+        tools=HOUSEKEEPER_TOOLS,
+        agent=get_housekeeper_agent(),
+    )
+
+    logger.info(
+        "housekeeper_registered",
+        role=HOUSEKEEPER_CAPABILITY.role,
+        domains=HOUSEKEEPER_CAPABILITY.domains,
+        tool_count=len(HOUSEKEEPER_TOOLS),
+    )
+
+
+def unregister_housekeeper() -> None:
+    """Unregister The Housekeeper from the Household Registry."""
+    registry = get_household_registry()
+    registry.unregister("housekeeper")
+    logger.info("housekeeper_unregistered")

src/agents/housekeeper/client.py

@@ -0,0 +1,556 @@
+"""
+HTTP client for the Core-API service.
+
+Provides async methods for home automation operations via Home Assistant.
+Core-API is a separate service that wraps the Home Assistant REST API
+into LLM-friendly endpoints.
+"""
+
+from typing import Any
+
+import httpx
+from pydantic import BaseModel, Field
+
+from src.core.config import config
+from src.core.logging_config import get_logger
+
+logger = get_logger(__name__)
+
+
+# ============================================================================
+# Response Models
+# ============================================================================
+
+
+class Device(BaseModel):
+    """Device from Home Assistant."""
+
+    entity_id: str
+    name: str
+    state: str
+    domain: str
+    area: str | None = None
+    attributes: dict[str, Any] = Field(default_factory=dict)
+
+
+class DeviceState(BaseModel):
+    """Detailed state of a device."""
+
+    entity_id: str
+    state: str
+    attributes: dict[str, Any] = Field(default_factory=dict)
+    last_changed: str | None = None
+    last_updated: str | None = None
+
+
+class Scene(BaseModel):
+    """Scene from Home Assistant."""
+
+    entity_id: str
+    name: str
+    friendly_name: str | None = None
+
+
+class Script(BaseModel):
+    """Script from Home Assistant."""
+
+    entity_id: str
+    name: str
+    description: str | None = None
+    last_triggered: str | None = None
+
+
+class Automation(BaseModel):
+    """Automation from Home Assistant."""
+
+    entity_id: str
+    name: str
+    state: str = "on"
+    last_triggered: str | None = None
+
+
+class HistoryEntry(BaseModel):
+    """History entry for an entity."""
+
+    state: str
+    timestamp: str
+    attributes: dict[str, Any] = Field(default_factory=dict)
+
+
+class ControlResult(BaseModel):
+    """Result of a device control operation."""
+
+    success: bool
+    entity_id: str
+    action: str
+    message: str = ""
+
+
+class Area(BaseModel):
+    """Area/room from Home Assistant."""
+
+    area_id: str
+    name: str
+    device_count: int = 0
+
+
+# ============================================================================
+# Client
+# ============================================================================
+
+
+class CoreAPIClient:
+    """
+    Async HTTP client for Core-API (Home Assistant wrapper).
+
+    Usage:
+        async with CoreAPIClient() as client:
+            devices = await client.list_devices()
+    """
+
+    def __init__(
+        self,
+        base_url: str | None = None,
+        api_key: str | None = None,
+        timeout: int = 30,
+    ):
+        """
+        Initialize the client.
+
+        Args:
+            base_url: Core-API URL (defaults to config)
+            api_key: API key for authentication (defaults to config)
+            timeout: Request timeout in seconds
+        """
+        self.base_url = base_url or str(config.CORE_API_HOST)
+        self.api_key = api_key or config.CORE_API_KEY
+        self.timeout = timeout
+        self._client: httpx.AsyncClient | None = None
+
+    async def __aenter__(self) -> "CoreAPIClient":
+        """Create HTTP client on context entry."""
+        headers = {}
+        if self.api_key:
+            headers["Authorization"] = f"Bearer {self.api_key}"
+
+        self._client = httpx.AsyncClient(
+            base_url=self.base_url,
+            headers=headers,
+            timeout=self.timeout,
+        )
+        return self
+
+    async def __aexit__(self, exc_type: Any, exc_val: Any, exc_tb: Any) -> None:
+        """Close HTTP client on context exit."""
+        if self._client:
+            await self._client.aclose()
+            self._client = None
+
+    def _ensure_client(self) -> httpx.AsyncClient:
+        """Ensure client is initialized."""
+        if self._client is None:
+            raise RuntimeError(
+                "Client not initialized. Use 'async with CoreAPIClient() as client:'"
+            )
+        return self._client
+
+    # ========================================================================
+    # Device Discovery
+    # ========================================================================
+
+    async def list_devices(
+        self,
+        domain: str | None = None,
+        area: str | None = None,
+    ) -> list[Device]:
+        """
+        List devices, optionally filtered by domain or area.
+
+        Args:
+            domain: Filter by domain (light, switch, climate, etc.)
+            area: Filter by area (living_room, bedroom, etc.)
+
+        Returns:
+            List of devices matching filters
+        """
+        client = self._ensure_client()
+
+        params: dict[str, str] = {}
+        if domain:
+            params["domain"] = domain
+        if area:
+            params["area"] = area
+
+        logger.debug("core_api_list_devices", domain=domain, area=area)
+
+        response = await client.get("/housekeeping/devices", params=params or None)
+        response.raise_for_status()
+
+        data = response.json()
+        return [Device(**d) for d in data.get("devices", [])]
+
+    async def list_areas(self) -> list[Area]:
+        """
+        List all areas/rooms in Home Assistant.
+
+        Returns:
+            List of areas with device counts
+        """
+        client = self._ensure_client()
+
+        logger.debug("core_api_list_areas")
+
+        response = await client.get("/housekeeping/areas")
+        response.raise_for_status()
+
+        data = response.json()
+        return [Area(**a) for a in data.get("areas", [])]
+
+    async def get_device_state(self, entity_id: str) -> DeviceState:
+        """
+        Get the current state of a specific device.
+
+        Args:
+            entity_id: Home Assistant entity ID (e.g., light.living_room)
+
+        Returns:
+            Current device state with attributes
+        """
+        client = self._ensure_client()
+
+        logger.debug("core_api_get_state", entity_id=entity_id)
+
+        response = await client.get(f"/housekeeping/devices/{entity_id}")
+        response.raise_for_status()
+
+        return DeviceState(**response.json())
+
+    # ========================================================================
+    # Device Control
+    # ========================================================================
+
+    async def turn_on(
+        self,
+        entity_id: str,
+        brightness: int | None = None,
+        color_temp: int | None = None,
+        rgb_color: tuple[int, int, int] | None = None,
+    ) -> ControlResult:
+        """
+        Turn on a device.
+
+        Args:
+            entity_id: Device to turn on
+            brightness: Optional brightness (0-255) for lights
+            color_temp: Optional color temperature in Kelvin for lights
+            rgb_color: Optional RGB color tuple for lights
+
+        Returns:
+            Result of the operation
+        """
+        client = self._ensure_client()
+
+        payload: dict[str, Any] = {"action": "turn_on"}
+        if brightness is not None:
+            payload["brightness"] = brightness
+        if color_temp is not None:
+            payload["color_temp"] = color_temp
+        if rgb_color is not None:
+            payload["rgb_color"] = list(rgb_color)
+
+        logger.info("core_api_turn_on", entity_id=entity_id, payload=payload)
+
+        response = await client.post(
+            f"/housekeeping/devices/{entity_id}/control",
+            json=payload,
+        )
+        response.raise_for_status()
+
+        data = response.json()
+        return ControlResult(
+            success=data.get("success", True),
+            entity_id=entity_id,
+            action="turn_on",
+            message=data.get("message", ""),
+        )
+
+    async def turn_off(self, entity_id: str) -> ControlResult:
+        """
+        Turn off a device.
+
+        Args:
+            entity_id: Device to turn off
+
+        Returns:
+            Result of the operation
+        """
+        client = self._ensure_client()
+
+        logger.info("core_api_turn_off", entity_id=entity_id)
+
+        response = await client.post(
+            f"/housekeeping/devices/{entity_id}/control",
+            json={"action": "turn_off"},
+        )
+        response.raise_for_status()
+
+        data = response.json()
+        return ControlResult(
+            success=data.get("success", True),
+            entity_id=entity_id,
+            action="turn_off",
+            message=data.get("message", ""),
+        )
+
+    async def toggle(self, entity_id: str) -> ControlResult:
+        """
+        Toggle a device's state.
+
+        Args:
+            entity_id: Device to toggle
+
+        Returns:
+            Result of the operation
+        """
+        client = self._ensure_client()
+
+        logger.info("core_api_toggle", entity_id=entity_id)
+
+        response = await client.post(
+            f"/housekeeping/devices/{entity_id}/control",
+            json={"action": "toggle"},
+        )
+        response.raise_for_status()
+
+        data = response.json()
+        return ControlResult(
+            success=data.get("success", True),
+            entity_id=entity_id,
+            action="toggle",
+            message=data.get("message", ""),
+        )
+
+    # ========================================================================
+    # Scenes
+    # ========================================================================
+
+    async def list_scenes(self) -> list[Scene]:
+        """
+        List all available scenes.
+
+        Returns:
+            List of scenes
+        """
+        client = self._ensure_client()
+
+        logger.debug("core_api_list_scenes")
+
+        response = await client.get("/housekeeping/scenes")
+        response.raise_for_status()
+
+        data = response.json()
+        return [Scene(**s) for s in data.get("scenes", [])]
+
+    async def activate_scene(self, scene_id: str) -> ControlResult:
+        """
+        Activate a scene.
+
+        Args:
+            scene_id: Scene entity ID (e.g., scene.movie_night)
+
+        Returns:
+            Result of the operation
+        """
+        client = self._ensure_client()
+
+        logger.info("core_api_activate_scene", scene_id=scene_id)
+
+        response = await client.post(f"/housekeeping/scenes/{scene_id}/activate")
+        response.raise_for_status()
+
+        data = response.json()
+        return ControlResult(
+            success=data.get("success", True),
+            entity_id=scene_id,
+            action="activate",
+            message=data.get("message", ""),
+        )
+
+    # ========================================================================
+    # Scripts
+    # ========================================================================
+
+    async def list_scripts(self) -> list[Script]:
+        """
+        List all available scripts.
+
+        Returns:
+            List of scripts
+        """
+        client = self._ensure_client()
+
+        logger.debug("core_api_list_scripts")
+
+        response = await client.get("/housekeeping/scripts")
+        response.raise_for_status()
+
+        data = response.json()
+        return [Script(**s) for s in data.get("scripts", [])]
+
+    async def run_script(
+        self,
+        script_id: str,
+        variables: dict[str, Any] | None = None,
+    ) -> ControlResult:
+        """
+        Run a script.
+
+        Args:
+            script_id: Script entity ID (e.g., script.good_morning)
+            variables: Optional variables to pass to the script
+
+        Returns:
+            Result of the operation
+        """
+        client = self._ensure_client()
+
+        payload: dict[str, Any] = {}
+        if variables:
+            payload["variables"] = variables
+
+        logger.info("core_api_run_script", script_id=script_id)
+
+        response = await client.post(
+            f"/housekeeping/scripts/{script_id}/run",
+            json=payload or None,
+        )
+        response.raise_for_status()
+
+        data = response.json()
+        return ControlResult(
+            success=data.get("success", True),
+            entity_id=script_id,
+            action="run",
+            message=data.get("message", ""),
+        )
+
+    # ========================================================================
+    # Automations
+    # ========================================================================
+
+    async def list_automations(self) -> list[Automation]:
+        """
+        List all automations.
+
+        Returns:
+            List of automations with their states
+        """
+        client = self._ensure_client()
+
+        logger.debug("core_api_list_automations")
+
+        response = await client.get("/housekeeping/automations")
+        response.raise_for_status()
+
+        data = response.json()
+        return [Automation(**a) for a in data.get("automations", [])]
+
+    async def toggle_automation(
+        self,
+        automation_id: str,
+        enable: bool,
+    ) -> ControlResult:
+        """
+        Enable or disable an automation.
+
+        Args:
+            automation_id: Automation entity ID
+            enable: True to enable, False to disable
+
+        Returns:
+            Result of the operation
+        """
+        client = self._ensure_client()
+
+        logger.info(
+            "core_api_toggle_automation",
+            automation_id=automation_id,
+            enable=enable,
+        )
+
+        response = await client.post(
+            f"/housekeeping/automations/{automation_id}/toggle",
+            json={"enable": enable},
+        )
+        response.raise_for_status()
+
+        data = response.json()
+        return ControlResult(
+            success=data.get("success", True),
+            entity_id=automation_id,
+            action="enable" if enable else "disable",
+            message=data.get("message", ""),
+        )
+
+    # ========================================================================
+    # History
+    # ========================================================================
+
+    async def get_history(
+        self,
+        entity_id: str,
+        hours: int = 24,
+    ) -> list[HistoryEntry]:
+        """
+        Get history for an entity.
+
+        Args:
+            entity_id: Entity to get history for
+            hours: Number of hours of history (default: 24)
+
+        Returns:
+            List of historical state entries
+        """
+        client = self._ensure_client()
+
+        logger.debug("core_api_get_history", entity_id=entity_id, hours=hours)
+
+        response = await client.get(
+            "/housekeeping/history",
+            params={"entity_id": entity_id, "hours": hours},
+        )
+        response.raise_for_status()
+
+        data = response.json()
+        return [HistoryEntry(**h) for h in data.get("history", [])]
+
+    # ========================================================================
+    # Health Check
+    # ========================================================================
+
+    async def health_check(self) -> bool:
+        """
+        Check if core-api and Home Assistant are healthy.
+
+        Returns:
+            True if healthy, False otherwise
+        """
+        try:
+            client = self._ensure_client()
+            response = await client.get("/housekeeping/health")
+            return response.status_code == 200
+        except Exception as e:
+            logger.warning("core_api_health_check_failed", error=str(e))
+            return False
+
+
+# Global client factory
+async def get_core_api_client() -> CoreAPIClient:
+    """
+    Get a core-api client instance.
+
+    Usage:
+        async with get_core_api_client() as client:
+            devices = await client.list_devices()
+    """
+    return CoreAPIClient()

src/agents/housekeeper/tools.py

@@ -0,0 +1,592 @@
+"""
+Housekeeper tools for PydanticAI agent.
+
+These tools wrap the core-api service and are registered with
+The Housekeeper agent for home automation tasks.
+"""
+
+from src.agents.housekeeper.client import CoreAPIClient
+from src.core.logging_config import get_logger
+
+logger = get_logger(__name__)
+
+
+# ============================================================================
+# Device Discovery
+# ============================================================================
+
+
+async def list_devices(
+    domain: str | None = None,
+    area: str | None = None,
+) -> str:
+    """
+    List available devices in the smart home.
+
+    Use this to discover what devices can be controlled.
+    Can filter by domain (device type) or area (room).
+
+    Args:
+        domain: Device type filter (light, switch, climate, cover, fan, etc.)
+        area: Room/area filter (living_room, bedroom, kitchen, etc.)
+
+    Returns:
+        List of devices with their current states
+
+    Examples:
+        list_devices()  # All devices
+        list_devices(domain="light")  # Only lights
+        list_devices(area="living_room")  # Living room devices
+    """
+    try:
+        async with CoreAPIClient() as client:
+            devices = await client.list_devices(domain=domain, area=area)
+
+            if not devices:
+                filters = []
+                if domain:
+                    filters.append(f"domain={domain}")
+                if area:
+                    filters.append(f"area={area}")
+                filter_str = f" with filters: {', '.join(filters)}" if filters else ""
+                return f"No devices found{filter_str}"
+
+            # Group by domain for readability
+            by_domain: dict[str, list] = {}
+            for device in devices:
+                by_domain.setdefault(device.domain, []).append(device)
+
+            output_parts = ["## Smart Home Devices\n"]
+
+            for dom, dom_devices in sorted(by_domain.items()):
+                output_parts.append(f"### {dom.title()}s")
+
+                # Sort devices: room groups first (using Home Assistant's is_hue_group attribute)
+                def is_room_group(d: object) -> bool:
+                    """Check if device is a room group based on HA attributes."""
+                    attrs = getattr(d, "attributes", {})
+                    # Check for Hue room groups
+                    if attrs.get("is_hue_group") and attrs.get("hue_type") == "room":
+                        return True
+                    # Check for other group indicators (icon or entity_id list)
+                    if "entity_id" in attrs and isinstance(attrs["entity_id"], list):
+                        return True
+                    return False
+
+                sorted_devices = sorted(
+                    dom_devices, key=lambda d: (not is_room_group(d), d.entity_id)
+                )
+
+                for device in sorted_devices:
+                    state_icon = (
+                        "on"
+                        if device.state == "on"
+                        else "off"
+                        if device.state == "off"
+                        else device.state
+                    )
+                    area_str = f" ({device.area})" if device.area else ""
+                    # Mark room groups clearly using actual HA data
+                    group_marker = " [ROOM GROUP]" if is_room_group(device) else ""
+                    output_parts.append(
+                        f"- **{device.name}**{area_str}{group_marker}: {state_icon}"
+                    )
+                    output_parts.append(f"  ID: `{device.entity_id}`")
+                output_parts.append("")
+
+            logger.info("housekeeper_list_devices", count=len(devices))
+            return "\n".join(output_parts)
+
+    except Exception as e:
+        logger.error("housekeeper_list_devices_error", error=str(e))
+        return f"Error listing devices: {str(e)}"
+
+
+async def list_areas() -> str:
+    """
+    List all areas/rooms in the smart home.
+
+    Use this to discover what rooms/areas are configured in Home Assistant.
+    Useful before filtering devices by area.
+
+    Returns:
+        List of areas with device counts
+
+    Examples:
+        list_areas()  # See all rooms/areas
+    """
+    try:
+        async with CoreAPIClient() as client:
+            areas = await client.list_areas()
+
+            if not areas:
+                return "No areas found in Home Assistant"
+
+            output_parts = ["## Smart Home Areas\n"]
+
+            for area in sorted(areas, key=lambda a: a.name):
+                device_str = f" ({area.device_count} devices)" if area.device_count else ""
+                output_parts.append(f"- **{area.name}**{device_str}")
+                output_parts.append(f"  ID: `{area.area_id}`")
+
+            output_parts.append("")
+            output_parts.append(f"*{len(areas)} areas total*")
+
+            logger.info("housekeeper_list_areas", count=len(areas))
+            return "\n".join(output_parts)
+
+    except Exception as e:
+        logger.error("housekeeper_list_areas_error", error=str(e))
+        return f"Error listing areas: {str(e)}"
+
+
+async def get_device_state(entity_id: str) -> str:
+    """
+    Get the current state and attributes of a specific device.
+
+    Use this to check a device's detailed status before or after control.
+
+    Args:
+        entity_id: The device entity ID (e.g., light.living_room, switch.coffee_maker)
+
+    Returns:
+        Detailed device state including all attributes
+
+    Examples:
+        get_device_state("light.living_room")
+        get_device_state("climate.bedroom")
+    """
+    try:
+        async with CoreAPIClient() as client:
+            state = await client.get_device_state(entity_id)
+
+            output_parts = [
+                f"## Device: {entity_id}",
+                f"**State:** {state.state}",
+            ]
+
+            if state.last_changed:
+                output_parts.append(f"**Last Changed:** {state.last_changed}")
+
+            if state.attributes:
+                output_parts.append("\n**Attributes:**")
+                for key, value in state.attributes.items():
+                    if key not in ("friendly_name", "entity_id"):
+                        output_parts.append(f"- {key}: {value}")
+
+            return "\n".join(output_parts)
+
+    except Exception as e:
+        logger.error("housekeeper_get_state_error", error=str(e), entity_id=entity_id)
+        return f"Error getting state for {entity_id}: {str(e)}"
+
+
+# ============================================================================
+# Device Control
+# ============================================================================
+
+
+async def turn_on(
+    entity_id: str,
+    brightness: int | None = None,
+    color_temp: int | None = None,
+) -> str:
+    """
+    Turn on a device. Use the entity_id parameter with the EXACT value from list_devices.
+
+    For lights, can optionally set brightness and color temperature.
+
+    Args:
+        entity_id: The EXACT entity ID from list_devices including domain prefix.
+        brightness: Optional brightness for lights (0-255, where 255 is full brightness)
+        color_temp: Optional color temperature in Kelvin (2700=warm, 6500=cool)
+
+    Returns:
+        Confirmation of the action
+
+    Examples:
+        turn_on(entity_id="light.living_room")
+        turn_on(entity_id="light.bedroom", brightness=128)
+        turn_on(entity_id="switch.coffee_maker")
+    """
+    try:
+        async with CoreAPIClient() as client:
+            result = await client.turn_on(
+                entity_id=entity_id,
+                brightness=brightness,
+                color_temp=color_temp,
+            )
+
+            if result.success:
+                extras = []
+                if brightness is not None:
+                    extras.append(f"brightness {brightness}/255")
+                if color_temp is not None:
+                    extras.append(f"color temp {color_temp}K")
+
+                extra_str = f" ({', '.join(extras)})" if extras else ""
+                return f"Turned on {entity_id}{extra_str}"
+            else:
+                return f"Failed to turn on {entity_id}: {result.message}"
+
+    except Exception as e:
+        logger.error("housekeeper_turn_on_error", error=str(e), entity_id=entity_id)
+        return f"Error turning on {entity_id}: {str(e)}"
+
+
+async def turn_off(entity_id: str) -> str:
+    """
+    Turn off a device. Use the entity_id parameter with the EXACT value from list_devices.
+
+    Args:
+        entity_id: The EXACT entity ID from list_devices including domain prefix.
+
+    Returns:
+        Confirmation of the action
+
+    Examples:
+        turn_off(entity_id="light.living_room")
+        turn_off(entity_id="switch.coffee_maker")
+        turn_off(entity_id="light.kitchen")
+    """
+    try:
+        async with CoreAPIClient() as client:
+            result = await client.turn_off(entity_id=entity_id)
+
+            if result.success:
+                return f"Turned off {entity_id}"
+            else:
+                return f"Failed to turn off {entity_id}: {result.message}"
+
+    except Exception as e:
+        logger.error("housekeeper_turn_off_error", error=str(e), entity_id=entity_id)
+        return f"Error turning off {entity_id}: {str(e)}"
+
+
+async def toggle(entity_id: str) -> str:
+    """
+    Toggle a device's state (on becomes off, off becomes on).
+
+    Use the entity_id parameter with the EXACT value from list_devices.
+
+    Args:
+        entity_id: The EXACT entity ID from list_devices including domain prefix.
+
+    Returns:
+        Confirmation with the new state
+
+    Examples:
+        toggle(entity_id="light.living_room")
+        toggle(entity_id="switch.fan")
+    """
+    try:
+        async with CoreAPIClient() as client:
+            result = await client.toggle(entity_id=entity_id)
+
+            if result.success:
+                return f"Toggled {entity_id}"
+            else:
+                return f"Failed to toggle {entity_id}: {result.message}"
+
+    except Exception as e:
+        logger.error("housekeeper_toggle_error", error=str(e), entity_id=entity_id)
+        return f"Error toggling {entity_id}: {str(e)}"
+
+
+# ============================================================================
+# Scenes
+# ============================================================================
+
+
+async def list_scenes() -> str:
+    """
+    List all available scenes.
+
+    Scenes are pre-configured combinations of device states.
+
+    Returns:
+        List of available scenes
+
+    Examples:
+        list_scenes()
+    """
+    try:
+        async with CoreAPIClient() as client:
+            scenes = await client.list_scenes()
+
+            if not scenes:
+                return "No scenes found"
+
+            output_parts = ["## Available Scenes\n"]
+            for scene in scenes:
+                name = scene.friendly_name or scene.name
+                output_parts.append(f"- **{name}**")
+                output_parts.append(f"  ID: `{scene.entity_id}`")
+
+            logger.info("housekeeper_list_scenes", count=len(scenes))
+            return "\n".join(output_parts)
+
+    except Exception as e:
+        logger.error("housekeeper_list_scenes_error", error=str(e))
+        return f"Error listing scenes: {str(e)}"
+
+
+async def activate_scene(scene_id: str) -> str:
+    """
+    Activate a scene.
+
+    This sets all devices in the scene to their configured states.
+
+    Args:
+        scene_id: Scene entity ID (e.g., scene.movie_night, scene.good_morning)
+
+    Returns:
+        Confirmation of activation
+
+    Examples:
+        activate_scene("scene.movie_night")
+        activate_scene("scene.good_morning")
+    """
+    try:
+        async with CoreAPIClient() as client:
+            result = await client.activate_scene(scene_id=scene_id)
+
+            if result.success:
+                return f"Activated scene: {scene_id}"
+            else:
+                return f"Failed to activate {scene_id}: {result.message}"
+
+    except Exception as e:
+        logger.error("housekeeper_activate_scene_error", error=str(e), scene_id=scene_id)
+        return f"Error activating scene {scene_id}: {str(e)}"
+
+
+# ============================================================================
+# Scripts
+# ============================================================================
+
+
+async def list_scripts() -> str:
+    """
+    List all available automation scripts.
+
+    Scripts are sequences of actions that can be triggered manually.
+
+    Returns:
+        List of available scripts
+
+    Examples:
+        list_scripts()
+    """
+    try:
+        async with CoreAPIClient() as client:
+            scripts = await client.list_scripts()
+
+            if not scripts:
+                return "No scripts found"
+
+            output_parts = ["## Available Scripts\n"]
+            for script in scripts:
+                output_parts.append(f"- **{script.name}**")
+                if script.description:
+                    output_parts.append(f"  {script.description}")
+                output_parts.append(f"  ID: `{script.entity_id}`")
+                if script.last_triggered:
+                    output_parts.append(f"  Last run: {script.last_triggered}")
+
+            logger.info("housekeeper_list_scripts", count=len(scripts))
+            return "\n".join(output_parts)
+
+    except Exception as e:
+        logger.error("housekeeper_list_scripts_error", error=str(e))
+        return f"Error listing scripts: {str(e)}"
+
+
+async def run_script(script_id: str) -> str:
+    """
+    Run an automation script.
+
+    Args:
+        script_id: Script entity ID (e.g., script.good_morning, script.bedtime)
+
+    Returns:
+        Confirmation of execution
+
+    Examples:
+        run_script("script.good_morning")
+        run_script("script.all_lights_off")
+    """
+    try:
+        async with CoreAPIClient() as client:
+            result = await client.run_script(script_id=script_id)
+
+            if result.success:
+                return f"Running script: {script_id}"
+            else:
+                return f"Failed to run {script_id}: {result.message}"
+
+    except Exception as e:
+        logger.error("housekeeper_run_script_error", error=str(e), script_id=script_id)
+        return f"Error running script {script_id}: {str(e)}"
+
+
+# ============================================================================
+# Automations
+# ============================================================================
+
+
+async def list_automations() -> str:
+    """
+    List all automations and their current states.
+
+    Automations are event-triggered rules that run automatically.
+
+    Returns:
+        List of automations with enabled/disabled status
+
+    Examples:
+        list_automations()
+    """
+    try:
+        async with CoreAPIClient() as client:
+            automations = await client.list_automations()
+
+            if not automations:
+                return "No automations found"
+
+            output_parts = ["## Automations\n"]
+
+            # Group by state
+            enabled = [a for a in automations if a.state == "on"]
+            disabled = [a for a in automations if a.state != "on"]
+
+            if enabled:
+                output_parts.append("### Enabled")
+                for auto in enabled:
+                    output_parts.append(f"- **{auto.name}**")
+                    output_parts.append(f"  ID: `{auto.entity_id}`")
+                    if auto.last_triggered:
+                        output_parts.append(f"  Last triggered: {auto.last_triggered}")
+                output_parts.append("")
+
+            if disabled:
+                output_parts.append("### Disabled")
+                for auto in disabled:
+                    output_parts.append(f"- **{auto.name}**")
+                    output_parts.append(f"  ID: `{auto.entity_id}`")
+
+            logger.info("housekeeper_list_automations", count=len(automations))
+            return "\n".join(output_parts)
+
+    except Exception as e:
+        logger.error("housekeeper_list_automations_error", error=str(e))
+        return f"Error listing automations: {str(e)}"
+
+
+async def toggle_automation(automation_id: str, enable: bool) -> str:
+    """
+    Enable or disable an automation.
+
+    Args:
+        automation_id: Automation entity ID
+        enable: True to enable, False to disable
+
+    Returns:
+        Confirmation of the change
+
+    Examples:
+        toggle_automation("automation.morning_lights", enable=True)
+        toggle_automation("automation.vacation_mode", enable=False)
+    """
+    try:
+        async with CoreAPIClient() as client:
+            result = await client.toggle_automation(
+                automation_id=automation_id,
+                enable=enable,
+            )
+
+            action = "Enabled" if enable else "Disabled"
+            if result.success:
+                return f"{action} automation: {automation_id}"
+            else:
+                return f"Failed to {action.lower()} {automation_id}: {result.message}"
+
+    except Exception as e:
+        logger.error(
+            "housekeeper_toggle_automation_error",
+            error=str(e),
+            automation_id=automation_id,
+        )
+        return f"Error toggling automation {automation_id}: {str(e)}"
+
+
+# ============================================================================
+# History
+# ============================================================================
+
+
+async def get_history(entity_id: str, hours: int = 24) -> str:
+    """
+    Get the state history of a device.
+
+    Useful for understanding patterns or troubleshooting.
+
+    Args:
+        entity_id: Device to get history for
+        hours: Number of hours of history (default: 24)
+
+    Returns:
+        List of state changes over the time period
+
+    Examples:
+        get_history("light.living_room")
+        get_history("climate.bedroom", hours=48)
+    """
+    try:
+        async with CoreAPIClient() as client:
+            history = await client.get_history(entity_id=entity_id, hours=hours)
+
+            if not history:
+                return f"No history found for {entity_id} in the last {hours} hours"
+
+            output_parts = [f"## History: {entity_id}", f"*Last {hours} hours*\n"]
+
+            for entry in history[-20:]:  # Show last 20 entries
+                output_parts.append(f"- **{entry.timestamp}**: {entry.state}")
+
+            if len(history) > 20:
+                output_parts.append(f"\n*(showing last 20 of {len(history)} entries)*")
+
+            return "\n".join(output_parts)
+
+    except Exception as e:
+        logger.error("housekeeper_get_history_error", error=str(e), entity_id=entity_id)
+        return f"Error getting history for {entity_id}: {str(e)}"
+
+
+# ============================================================================
+# Tool Collection for Registration
+# ============================================================================
+
+# All tools available to The Housekeeper
+HOUSEKEEPER_TOOLS = [
+    # Discovery
+    list_areas,
+    list_devices,
+    get_device_state,
+    # Control
+    turn_on,
+    turn_off,
+    toggle,
+    # Scenes
+    list_scenes,
+    activate_scene,
+    # Scripts
+    list_scripts,
+    run_script,
+    # Automations
+    list_automations,
+    toggle_automation,
+    # History
+    get_history,
+]

src/agents/librarian/__init__.py

@@ -7,10 +7,10 @@ Connects to the library-desk API to provide:
 - Knowledge graph queries
 - Semantic search
 """
+
 from src.agents.librarian.agent import (
     get_librarian_agent,
     run_librarian,
-    run_librarian_stream,
 )
 from src.agents.librarian.capability import (
     LIBRARIAN_CAPABILITY,
@@ -26,5 +26,4 @@ __all__ = [
     "register_librarian",
     "unregister_librarian",
     "run_librarian",
-    "run_librarian_stream",
 ]

src/agents/librarian/agent.py

@@ -7,10 +7,12 @@ the library-desk API, offering:
 - Wiki and document management
 - Semantic search and knowledge graph exploration
 """
-from typing import Any, Optional
+
+from typing import Any
 
 from pydantic_ai import Agent
 
+from src.agents.librarian.client import library_client_session
 from src.agents.librarian.tools import (
     create_wiki_page,
     explore_knowledge_graph,
@@ -19,12 +21,15 @@ from src.agents.librarian.tools import (
     get_wiki_page,
     hybrid_search,
     list_dossiers,
+    read_url,
+    read_urls_batch,
+    search_web,
     search_wiki,
     semantic_search,
     smart_create_wiki_page,
     update_wiki_page,
 )
-from src.core.config import config
+from src.agents.protocol import AgentError
 from src.core.logging_config import get_logger
 
 logger = get_logger(__name__)
@@ -36,7 +41,14 @@ Your role is to help users find, understand, synthesize, and manage information
 - The personal wiki (Wiki.js) containing documentation and notes
 - The knowledge graph (Neo4j) with entities and relationships
 - Vector embeddings (Qdrant) for semantic search
-- Web search (SearXNG) for current information
+- Paperless documents (📑) - indexed PDFs, scanned documents, invoices, receipts from the user's document archive
+- Volatile cache (⚡) - pre-fetched real-time data for user-relevant locations and items:
+  - weather/forecast: conditions and forecasts for user's configured cities
+  - news: headlines from user's preferred sources
+  - stock/crypto: quotes for user's watched symbols
+  - sun/air_quality: data for user's locations
+  - Note: volatile data may not exist for arbitrary queries - falls back to web search
+- Web search (SearXNG) for current information not available in cache
 
 ## Your Personality
 - Scholarly and thorough in your research
@@ -47,8 +59,23 @@ Your role is to help users find, understand, synthesize, and manage information
 
 ## Your Tools
 
-### Research Tools
-- **hybrid_search**: Your primary research tool - searches all sources at once
+### Web Search & Content Extraction
+- **search_web**: Search the internet for current information (weather, news, facts)
+  - Use for: weather forecasts, current events, recent developments, external facts
+  - Returns extracted content from search results, not just snippets
+- **read_url**: Read and extract content from a specific URL
+  - Use when: user provides a URL or you need to read a specific webpage
+- **read_urls_batch**: Read multiple URLs in parallel (up to 20)
+  - Use for: comparing multiple sources, gathering info from several pages
+
+### Internal Research Tools
+- **hybrid_search**: Your primary research tool - searches ALL sources at once:
+  - Wiki pages (vector similarity)
+  - Knowledge graph (entity relationships)
+  - Paperless documents (📑 indexed PDFs, scans)
+  - Volatile cache (⚡ weather, news, stocks - when available)
+  - Web search (current information)
+  Results are fused and re-ranked by relevance. Volatile data gets priority when fresh.
 - **search_wiki**: Find specific wiki pages by keyword
 - **semantic_search**: Find conceptually similar content
 - **explore_knowledge_graph** / **find_related_entities**: Discover connections
@@ -102,35 +129,58 @@ Your responses are returned to Tatlock (the butler) who will synthesize them int
 - Note any gaps in available information
 - Be concise but thorough - Tatlock will format the final response
 - Structure your findings clearly so they can be easily integrated with other responses
+
+## CRITICAL: Never Fabricate Information
+If a tool fails or you cannot access a data source:
+- Say "I was unable to retrieve [information type]" - be specific about what failed
+- Do NOT provide placeholder, template, or made-up data
+- Do NOT say "Here's what I would have said" or "Here's a sample response"
+- Do NOT invent specific numbers, dates, or facts when the actual data is unavailable
+- It is better to return no information than to return fabricated information
 """
 
+
+# Tool-phase prompt actually used by the agent. The scholarly persona prompt
+# above suppresses tool calling on small local models (gemma4 answers in
+# character - "please provide your request" - without ever calling a tool),
+# the same pathology TATLOCK_ORCHESTRATION_PROMPT fixed for the butler.
+# Tatlock's synthesis phase supplies the user-facing voice, so the research
+# phase only needs tool discipline. Kept: the anti-fabrication rule.
+LIBRARIAN_TASK_PROMPT = """You are The Librarian, the research executor of the \
+Tatlock household. Your only job is to gather accurate findings by calling the \
+provided tools.
+
+- ALWAYS use tools - never answer a research task from memory alone.
+- Research or wiki questions: call hybrid_search first; then search_wiki and \
+get_wiki_page to read specific pages BEFORE summarizing them.
+- Current or external information (weather, news, live facts): call search_web; \
+call read_url when given a specific URL.
+- Wiki writing: smart_create_wiki_page when asked for a page about a topic; \
+create_wiki_page only for user-provided verbatim content; update_wiki_page for \
+edits (search_wiki, then get_wiki_page, then update).
+- Reply with a concise factual summary of what the tools returned, citing page \
+titles and URLs. A later step writes the polished answer, so no personality.
+- NEVER fabricate. If a tool fails or returns nothing, state exactly what you \
+could not retrieve and stop."""
+
 # Lazy initialization to avoid connection issues during imports
-_librarian_agent: Optional[Agent[None, str]] = None
+_librarian_agent: Agent[None, str] | None = None
 
 
 def _create_librarian_agent() -> Agent[None, str]:
     """Create the Librarian PydanticAI agent."""
-    # Import required classes for Ollama configuration
-    from pydantic_ai.models.openai import OpenAIChatModel
-    from pydantic_ai.providers.ollama import OllamaProvider
+    from src.anthropic.model_selector import get_model
 
-    # PydanticAI expects Ollama base URL to end with /v1
-    clean_host = str(config.OLLAMA_HOST).rstrip('/')
-    base_url = f"{clean_host}/v1"
-
-    # Create Ollama model with provider
-    model = OpenAIChatModel(
-        model_name=config.OLLAMA_DEFAULT_MODEL,
-        provider=OllamaProvider(base_url=base_url)
-    )
+    # Get best available model (Claude if available, else Ollama)
+    model = get_model()
 
     agent: Agent[None, str] = Agent(
         model=model,
-        system_prompt=LIBRARIAN_SYSTEM_PROMPT,
+        system_prompt=LIBRARIAN_TASK_PROMPT,
         retries=2,
     )
 
-    # Register research tools
+    # Register research tools (internal knowledge)
     agent.tool_plain(hybrid_search)
     agent.tool_plain(search_wiki)
     agent.tool_plain(semantic_search)
@@ -139,6 +189,11 @@ def _create_librarian_agent() -> Agent[None, str]:
     agent.tool_plain(explore_knowledge_graph)
     agent.tool_plain(find_related_entities)
 
+    # Register web search & content extraction tools
+    agent.tool_plain(search_web)
+    agent.tool_plain(read_url)
+    agent.tool_plain(read_urls_batch)
+
     # Register wiki read tools
     agent.tool_plain(get_wiki_page)
 
@@ -147,10 +202,14 @@ def _create_librarian_agent() -> Agent[None, str]:
     agent.tool_plain(update_wiki_page)
     agent.tool_plain(smart_create_wiki_page)
 
+    from src.anthropic.model_selector import get_model_info
+
+    model_info = get_model_info()
     logger.info(
         "librarian_agent_created",
-        model=config.OLLAMA_DEFAULT_MODEL,
-        tool_count=11,
+        backend=model_info["backend"],
+        model=model_info["model"],
+        tool_count=14,  # 7 research + 3 web + 1 wiki read + 3 wiki write
     )
 
     return agent
@@ -172,7 +231,7 @@ def get_librarian_agent() -> Agent[None, str]:
 async def run_librarian(
     task: str,
     context: str = "",
-    message_history: Optional[list[Any]] = None,
+    message_history: list[Any] | None = None,
 ) -> str:
     """
     Execute a research task with The Librarian.
@@ -188,6 +247,10 @@ async def run_librarian(
     Returns:
         Research results and findings
 
+    Raises:
+        AgentError: If the research task fails. Exception detail is
+            logged here; callers map the failure to a user-safe message.
+
     Example:
         result = await run_librarian(
             task="Find information about Docker networking",
@@ -209,10 +272,12 @@ async def run_librarian(
     )
 
     try:
-        result = await agent.run(
-            prompt,
-            message_history=message_history,
-        )
+        # One shared library-desk connection for all tool calls in this run
+        async with library_client_session():
+            result = await agent.run(
+                prompt,
+                message_history=message_history,
+            )
 
         logger.info(
             "librarian_task_completed",
@@ -223,64 +288,12 @@ async def run_librarian(
         return result.output
 
     except Exception as e:
+        # Full detail stays in the logs; callers receive a structured
+        # failure instead of error text masquerading as research output.
         logger.error(
             "librarian_task_error",
             task=task[:50],
             error=str(e),
             exc_info=True,
         )
-        return f"The Librarian encountered an error: {str(e)}"
-
-
-async def run_librarian_stream(
-    task: str,
-    context: str = "",
-    message_history: Optional[list[Any]] = None,
-):
-    """
-    Execute a research task with streaming output.
-
-    Yields text deltas as The Librarian generates the response.
-
-    Args:
-        task: The research task or question
-        context: Additional context from conversation
-        message_history: Optional conversation history
-
-    Yields:
-        str: Text deltas from the response
-
-    Example:
-        async for delta in run_librarian_stream("Find Docker docs"):
-            print(delta, end="", flush=True)
-    """
-    agent = get_librarian_agent()
-
-    # Build prompt with context if provided
-    prompt = task
-    if context:
-        prompt = f"Context: {context}\n\nTask: {task}"
-
-    logger.info(
-        "librarian_stream_started",
-        task=task[:100],
-    )
-
-    try:
-        async with agent.run_stream(
-            prompt,
-            message_history=message_history,
-        ) as response:
-            async for delta in response.stream_text(delta=True):
-                yield delta
-
-        logger.info("librarian_stream_completed", task=task[:50])
-
-    except Exception as e:
-        logger.error(
-            "librarian_stream_error",
-            task=task[:50],
-            error=str(e),
-            exc_info=True,
-        )
-        yield f"\n\nThe Librarian encountered an error: {str(e)}"
+        raise AgentError("Research task failed", agent_name="librarian") from e

src/agents/librarian/capability.py

@@ -4,6 +4,7 @@ Librarian capability registration for the Household Registry.
 Defines The Librarian's capabilities and registers it as a
 household member for coordination by the Steward and Tatlock.
 """
+
 from src.agents.librarian.agent import get_librarian_agent
 from src.agents.librarian.tools import LIBRARIAN_TOOLS
 from src.core.household_registry import (
@@ -21,10 +22,11 @@ LIBRARIAN_CAPABILITY = HouseholdCapability(
     role="The Librarian",
     category="research",
     description=(
-        "Research and wiki management: can CREATE wiki pages about topics "
+        "Research, web search, and wiki management: can SEARCH the web for current "
+        "information, READ URLs/articles, CREATE wiki pages about topics "
         "(with automatic HybridRAG research), UPDATE existing pages, "
-        "SEARCH wiki/knowledge graph/web, and synthesize information. "
-        "Use for: 'create a page about X', 'update wiki', 'find info on X'"
+        "and synthesize information from multiple sources. "
+        "Use for: 'search for X', 'what is X', 'create a page about X', 'read this URL'"
     ),
     domains=[
         "research",
@@ -33,6 +35,9 @@ LIBRARIAN_CAPABILITY = HouseholdCapability(
         "wiki",
         "documents",
         "search",
+        "web",
+        "url",
+        "internet",
         "synthesis",
         "create",
         "write",

src/agents/librarian/client.py

@@ -7,45 +7,65 @@ Provides async methods for all relevant library-desk endpoints:
 - Vector search
 - Knowledge graph queries
 """
-from typing import Any, Optional
+
+import asyncio
+from collections.abc import AsyncIterator, Awaitable, Callable
+from contextlib import asynccontextmanager
+from contextvars import ContextVar
+from typing import Any
 
 import httpx
 from pydantic import BaseModel, Field
 
 from src.core.config import config
-from src.core.context import get_user
+from src.core.context import apply_tenant_guard, get_user
 from src.core.logging_config import get_logger
 
 logger = get_logger(__name__)
 
+# Retry policy for idempotent/read-only requests (GETs, POST /query/*,
+# POST /rag/search). Writes are never retried.
+_RETRY_ATTEMPTS = 2
+_RETRY_BACKOFF_SECONDS = 0.5
+_RETRYABLE_STATUS_CODES = {502, 503, 504}
+
+# One shared HTTP connection per librarian run (see library_client_session)
+_shared_http_client: ContextVar[httpx.AsyncClient | None] = ContextVar(
+    "library_desk_http_client", default=None
+)
+
 
 # ============================================================================
 # Response Models
 # ============================================================================
 
+
 class WikiPage(BaseModel):
     """Wiki page from library-desk."""
+
     id: int
     path: str
     title: str
-    description: Optional[str] = None
-    content: Optional[str] = None
+    description: str | None = None
+    content: str | None = None
     tags: list[str] = Field(default_factory=list)
-    created_at: Optional[str] = None
-    updated_at: Optional[str] = None
+    created_at: str | None = None
+    updated_at: str | None = None
 
 
 class WikiSearchResult(BaseModel):
     """Search result from wiki search."""
+
     id: int
     path: str
     title: str
-    description: Optional[str] = None
-    locale: Optional[str] = None
+    description: str | None = None
+    locale: str | None = None
 
 
 class VectorSearchResult(BaseModel):
     """Result from semantic vector search."""
+
     page_id: int
     page_path: str
     page_title: str
@@ -56,28 +76,42 @@ class VectorSearchResult(BaseModel):
 
 class HybridSearchResult(BaseModel):
     """Result from HybridRAG search."""
-    source: str  # "vector", "graph", "web"
+
+    source: str  # source_type: "wiki", "web", "volatile", "document"
+    sources: list[str] = Field(
+        default_factory=list
+    )  # legs that found it: "vector", "graph", "web", ...
     title: str
     content: str
-    url: Optional[str] = None
-    score: float
-    page_id: Optional[int] = None
+    url: str | None = None
+    score: float  # rrf_score from the live service
+    page_id: int | None = None
+    related_dossiers: list[dict[str, Any]] = Field(default_factory=list)
     metadata: dict[str, Any] = Field(default_factory=dict)
 
 
 class HybridRAGResponse(BaseModel):
     """Full response from HybridRAG query."""
+
     results: list[HybridSearchResult] = Field(default_factory=list)
     keywords: list[str] = Field(default_factory=list)
     synonyms: list[str] = Field(default_factory=list)
     related_dossiers: list[str] = Field(default_factory=list)
     formatted_context: str = ""
-    search_id: Optional[str] = None
+    search_id: str | None = None
+    source_counts: dict[str, int] = Field(default_factory=dict)
     timing: dict[str, float] = Field(default_factory=dict)
+    # Additive degradation contract - only newer library-desk versions
+    # send these; absence means "no status reported", not "healthy".
+    # Maps each leg (vector/graph/web/volatile/documents) to
+    # "ok" | "failed" | "disabled".
+    source_status: dict[str, str] = Field(default_factory=dict)
+    degraded: bool = False
 
 
 class GraphNode(BaseModel):
     """Node from knowledge graph."""
+
     id: str
     labels: list[str] = Field(default_factory=list)
     properties: dict[str, Any] = Field(default_factory=dict)
@@ -85,12 +119,14 @@ class GraphNode(BaseModel):
 
 class Dossier(BaseModel):
     """A dossier (tag-based collection)."""
+
     name: str
     page_count: int
 
 
 class ResearchSummary(BaseModel):
     """Summary of research performed during smart-create."""
+
     wiki_results: int = 0
     web_results: int = 0
     graph_entities: int = 0
@@ -98,8 +134,54 @@ class ResearchSummary(BaseModel):
     timing_ms: int = 0
 
 
+class WebSearchResult(BaseModel):
+    """Result from web search via /rag/search."""
+
+    title: str
+    url: str
+    content: str = ""  # Full extracted text via Trafilatura
+    snippet: str = ""  # Original search engine snippet
+    source: str = ""  # Domain name
+    published_date: str | None = None
+
+
+class WebSearchResponse(BaseModel):
+    """Response from /rag/search endpoint."""
+
+    query: str
+    search_type: str
+    results: list[WebSearchResult] = Field(default_factory=list)
+    total_results: int = 0
+    search_time_ms: int = 0
+    sources_summary: str = ""  # Pre-formatted markdown citations
+
+
+class ContentExtractionResult(BaseModel):
+    """Result from content extraction."""
+
+    url: str
+    title: str | None = None
+    content: str = ""
+    author: str | None = None
+    date: str | None = None
+    language: str | None = None
+    success: bool = True
+    error: str | None = None
+
+
+class BatchExtractionResponse(BaseModel):
+    """Response from batch content extraction."""
+
+    results: list[ContentExtractionResult] = Field(default_factory=list)
+    total_urls: int = 0
+    successful: int = 0
+    failed: int = 0
+    extraction_time_ms: int = 0
+
+
 class EntityLinking(BaseModel):
     """Entity linking results from smart-create."""
+
     forward_links: int = 0
     backward_links: int = 0
     pages_updated: int = 0
@@ -107,10 +189,11 @@ class EntityLinking(BaseModel):
 
 class SmartCreateResponse(BaseModel):
     """Response from smart-create wiki page endpoint."""
+
     page: WikiPage
     research_summary: ResearchSummary = Field(default_factory=ResearchSummary)
     sources_used: int = 0
-    search_id: Optional[str] = None
+    search_id: str | None = None
     entity_linking: EntityLinking = Field(default_factory=EntityLinking)
 
 
@@ -118,6 +201,7 @@ class SmartCreateResponse(BaseModel):
 # Client
 # ============================================================================
 
+
 class LibraryDeskClient:
     """
     Async HTTP client for Library-Desk API.
@@ -129,9 +213,9 @@ class LibraryDeskClient:
 
     def __init__(
         self,
-        base_url: Optional[str] = None,
-        api_key: Optional[str] = None,
-        timeout: int = 60,
+        base_url: str | None = None,
+        api_key: str | None = None,
+        timeout: int | None = None,
     ):
         """
         Initialize the client.
@@ -140,30 +224,55 @@ class LibraryDeskClient:
             base_url: Library-desk API URL (defaults to config)
             api_key: API key for authentication (defaults to config)
             timeout: Request timeout in seconds
+                (defaults to config.LIBRARY_DESK_TIMEOUT)
         """
         self.base_url = base_url or str(config.LIBRARY_DESK_HOST)
         self.api_key = api_key or config.LIBRARY_DESK_API_KEY
-        self.timeout = timeout
-        self._client: Optional[httpx.AsyncClient] = None
+        self.timeout = timeout if timeout is not None else config.LIBRARY_DESK_TIMEOUT
+        self._client: httpx.AsyncClient | None = None
+        self._owns_client = False
 
-    async def __aenter__(self) -> "LibraryDeskClient":
-        """Create HTTP client on context entry."""
+    def _build_http_client(self) -> httpx.AsyncClient:
+        """Build a configured httpx client."""
         headers = {}
         if self.api_key:
             headers["Authorization"] = f"Bearer {self.api_key}"
 
-        self._client = httpx.AsyncClient(
+        return httpx.AsyncClient(
             base_url=self.base_url,
             headers=headers,
             timeout=self.timeout,
         )
+
+    def _uses_default_target(self) -> bool:
+        """Whether this client targets the configured library-desk instance."""
+        return (
+            self.base_url == str(config.LIBRARY_DESK_HOST)
+            and self.api_key == config.LIBRARY_DESK_API_KEY
+        )
+
+    async def __aenter__(self) -> "LibraryDeskClient":
+        """
+        Acquire an HTTP client on context entry.
+
+        Reuses the run-level shared connection (see library_client_session)
+        when one is active, instead of constructing a new client per call.
+        """
+        shared = _shared_http_client.get()
+        if shared is not None and not shared.is_closed and self._uses_default_target():
+            self._client = shared
+            self._owns_client = False
+        else:
+            self._client = self._build_http_client()
+            self._owns_client = True
         return self
 
     async def __aexit__(self, exc_type: Any, exc_val: Any, exc_tb: Any) -> None:
-        """Close HTTP client on context exit."""
-        if self._client:
+        """Close HTTP client on context exit (only if we own it)."""
+        if self._client and self._owns_client:
             await self._client.aclose()
-            self._client = None
+        self._client = None
+        self._owns_client = False
 
     def _ensure_client(self) -> httpx.AsyncClient:
         """Ensure client is initialized."""
@@ -173,6 +282,69 @@ class LibraryDeskClient:
             )
         return self._client
 
+    def _resolve_user(self, user: str | None) -> str:
+        """
+        Resolve the effective tenant for a request and require it non-empty.
+
+        Library-desk is removing its server-side default user, so every
+        request must carry an explicit tenant (a missing user will 422).
+        An empty tenant is a programming or configuration error - fail
+        loudly here, before any bytes hit the wire.
+
+        Explicit user arguments are stripped and routed through the same
+        tenant guard as context resolution (get_user() already applies
+        it), so a dev environment can never send the production tenant -
+        or a sanitization-collision variant of it - to library-desk.
+        """
+        effective = (user if user is not None else get_user()).strip()
+        if not effective:
+            raise ValueError(
+                "library-desk request requires a non-empty user (tenant); "
+                "got an empty value from the caller or request context"
+            )
+        return apply_tenant_guard(effective)
+
+    async def _request_with_retry(
+        self,
+        send: Callable[[], Awaitable[httpx.Response]],
+        description: str,
+    ) -> httpx.Response:
+        """
+        Send an idempotent/read-only request with a bounded retry.
+
+        Retries once (2 attempts total) with a short backoff on transport
+        errors and retryable 5xx statuses. Only used for GETs and the
+        read-only POST /query/* and /rag/search endpoints - never for
+        wiki writes.
+        """
+        for attempt in range(1, _RETRY_ATTEMPTS + 1):
+            try:
+                response = await send()
+            except httpx.TransportError as e:
+                if attempt >= _RETRY_ATTEMPTS:
+                    raise
+                logger.warning(
+                    "library_desk_retry",
+                    request=description,
+                    error=str(e),
+                    attempt=attempt,
+                )
+            else:
+                if (
+                    response.status_code not in _RETRYABLE_STATUS_CODES
+                    or attempt >= _RETRY_ATTEMPTS
+                ):
+                    return response
+                logger.warning(
+                    "library_desk_retry",
+                    request=description,
+                    status_code=response.status_code,
+                    attempt=attempt,
+                )
+            await asyncio.sleep(_RETRY_BACKOFF_SECONDS * attempt)
+
+        raise RuntimeError("unreachable")  # pragma: no cover
+
     # ========================================================================
     # HybridRAG
     # ========================================================================
@@ -184,33 +356,44 @@ class LibraryDeskClient:
         vector_limit: int = 10,
         graph_limit: int = 10,
         web_limit: int = 5,
+        document_limit: int = 5,
+        volatile_limit: int = 3,
         enable_reranking: bool = True,
         final_result_count: int = 10,
     ) -> HybridRAGResponse:
         """
-        Execute HybridRAG search combining vector, graph, and web results.
+        Execute HybridRAG search combining vector, graph, documents, volatile, and web.
 
         Args:
             query: Search query
             user: User identifier for multi-tenancy (defaults to request context)
-            vector_limit: Max results from vector search
+            vector_limit: Max results from vector search (wiki pages)
             graph_limit: Max results from graph search
-            web_limit: Max results from web search
+            web_limit: Max results from web search (0 to disable)
+            document_limit: Max results from Paperless documents (0 to disable)
+            volatile_limit: Max results from volatile cache (0 to disable)
             enable_reranking: Whether to rerank with LLM
             final_result_count: Number of final results after fusion
 
         Returns:
             HybridRAGResponse with ranked results and context
         """
-        user = user or get_user()
+        user = self._resolve_user(user)
         client = self._ensure_client()
 
+        # The live service requires all limits >= 1 (422 otherwise);
+        # legs are disabled via the enable_* flags, not a zero limit.
         payload = {
             "query": query,
             "config": {
-                "vector_limit": vector_limit,
-                "graph_limit": graph_limit,
-                "web_limit": web_limit,
+                "vector_limit": max(vector_limit, 1),
+                "graph_limit": max(graph_limit, 1),
+                "web_limit": max(web_limit, 1),
+                "document_limit": max(document_limit, 1),
+                "volatile_limit": max(volatile_limit, 1),
+                "enable_documents": document_limit > 0,
+                "enable_volatile": volatile_limit > 0,
+                "enable_web": web_limit > 0,
                 "enable_reranking": enable_reranking,
                 "final_result_count": final_result_count,
             },
@@ -218,36 +401,71 @@ class LibraryDeskClient:
 
         logger.info("library_desk_hybrid_search", query=query, user=user)
 
-        response = await client.post(
-            "/query/hybrid",
-            json=payload,
-            params={"user": user},
+        response = await self._request_with_retry(
+            lambda: client.post(
+                "/query/hybrid",
+                json=payload,
+                params={"user": user},
+            ),
+            "POST /query/hybrid",
         )
         response.raise_for_status()
 
         data = response.json()
 
-        # Parse results
+        # Parse results (live field names: source_type, sources, rrf_score,
+        # related_dossiers; older names kept as fallbacks)
         results = []
         for r in data.get("results", []):
-            results.append(HybridSearchResult(
-                source=r.get("source", "unknown"),
-                title=r.get("title", ""),
-                content=r.get("content", ""),
-                url=r.get("url"),
-                score=r.get("score", 0.0),
-                page_id=r.get("page_id"),
-                metadata=r.get("metadata", {}),
-            ))
+            results.append(
+                HybridSearchResult(
+                    source=r.get("source_type") or r.get("source", "unknown"),
+                    sources=r.get("sources", []),
+                    title=r.get("title", ""),
+                    content=r.get("content", ""),
+                    url=r.get("url"),
+                    score=r.get("rrf_score", r.get("score", 0.0)),
+                    page_id=r.get("page_id"),
+                    related_dossiers=r.get("related_dossiers", []),
+                    metadata=r.get("metadata", {}),
+                )
+            )
+
+        # Handle keywords being either a list or a dict with core_keywords;
+        # the live service nests synonyms inside the keywords dict as a
+        # {term: [synonyms]} map.
+        raw_keywords = data.get("keywords", [])
+        raw_synonyms: Any = data.get("synonyms", [])
+        if isinstance(raw_keywords, dict):
+            keywords = raw_keywords.get("core_keywords", [])
+            raw_synonyms = raw_keywords.get("synonyms", {})
+        else:
+            keywords = raw_keywords
+        if isinstance(raw_synonyms, dict):
+            synonyms = [s for values in raw_synonyms.values() for s in values]
+        else:
+            synonyms = raw_synonyms
+
+        # Aggregate per-result related dossiers into unique top-level titles
+        related_dossiers: list[str] = []
+        for result in results:
+            for dossier in result.related_dossiers:
+                title = dossier.get("title", "")
+                if title and title not in related_dossiers:
+                    related_dossiers.append(title)
 
         return HybridRAGResponse(
             results=results,
-            keywords=data.get("keywords", []),
-            synonyms=data.get("synonyms", []),
-            related_dossiers=data.get("related_dossiers", []),
-            formatted_context=data.get("formatted_context", ""),
+            keywords=keywords,
+            synonyms=synonyms,
+            related_dossiers=related_dossiers,
+            formatted_context=data.get("context", data.get("formatted_context", "")),
             search_id=data.get("search_id"),
+            source_counts=data.get("source_counts", {}),
             timing=data.get("timing", {}),
+            # Additive fields - tolerate absence on older library-desk
+            source_status=data.get("source_status") or {},
+            degraded=bool(data.get("degraded", False)),
         )
 
     # ========================================================================
@@ -271,14 +489,17 @@ class LibraryDeskClient:
         Returns:
             List of matching wiki pages
         """
-        user = user or get_user()
+        user = self._resolve_user(user)
         client = self._ensure_client()
 
         logger.debug("library_desk_wiki_search", query=query, user=user)
 
-        response = await client.get(
-            "/wiki/search",
-            params={"q": query, "user": user, "limit": limit},
+        response = await self._request_with_retry(
+            lambda: client.get(
+                "/wiki/search",
+                params={"q": query, "user": user, "limit": limit},
+            ),
+            "GET /wiki/search",
         )
         response.raise_for_status()
 
@@ -300,12 +521,15 @@ class LibraryDeskClient:
         Returns:
             WikiPage with full content
         """
-        user = user or get_user()
+        user = self._resolve_user(user)
         client = self._ensure_client()
 
-        response = await client.get(
-            f"/wiki/pages/{page_id}",
-            params={"user": user},
+        response = await self._request_with_retry(
+            lambda: client.get(
+                f"/wiki/pages/{page_id}",
+                params={"user": user},
+            ),
+            f"GET /wiki/pages/{page_id}",
         )
         response.raise_for_status()
 
@@ -314,7 +538,7 @@ class LibraryDeskClient:
     async def list_wiki_pages(
         self,
         user: str | None = None,
-        tag: Optional[str] = None,
+        tag: str | None = None,
         limit: int = 50,
     ) -> list[WikiPage]:
         """
@@ -328,14 +552,17 @@ class LibraryDeskClient:
         Returns:
             List of wiki pages
         """
-        user = user or get_user()
+        user = self._resolve_user(user)
         client = self._ensure_client()
 
         params: dict[str, Any] = {"user": user, "limit": limit}
         if tag:
             params["tag"] = tag
 
-        response = await client.get("/wiki/pages", params=params)
+        response = await self._request_with_retry(
+            lambda: client.get("/wiki/pages", params=params),
+            "GET /wiki/pages",
+        )
         response.raise_for_status()
 
         data = response.json()
@@ -348,7 +575,7 @@ class LibraryDeskClient:
         content: str,
         user: str | None = None,
         description: str = "",
-        tags: Optional[list[str]] = None,
+        tags: list[str] | None = None,
     ) -> WikiPage:
         """
         Create a new wiki page.
@@ -364,7 +591,7 @@ class LibraryDeskClient:
         Returns:
             Created WikiPage
         """
-        user = user or get_user()
+        user = self._resolve_user(user)
         client = self._ensure_client()
 
         payload = {
@@ -387,10 +614,10 @@ class LibraryDeskClient:
         self,
         page_id: int,
         user: str | None = None,
-        content: Optional[str] = None,
-        title: Optional[str] = None,
-        tags: Optional[list[str]] = None,
-        description: Optional[str] = None,
+        content: str | None = None,
+        title: str | None = None,
+        tags: list[str] | None = None,
+        description: str | None = None,
     ) -> WikiPage:
         """
         Update an existing wiki page.
@@ -409,7 +636,7 @@ class LibraryDeskClient:
         Returns:
             Updated WikiPage
         """
-        user = user or get_user()
+        user = self._resolve_user(user)
         client = self._ensure_client()
 
         # Build update payload with only provided fields
@@ -443,7 +670,7 @@ class LibraryDeskClient:
         topic: str,
         tags: list[str],
         user: str | None = None,
-        path: Optional[str] = None,
+        path: str | None = None,
         include_web_research: bool = True,
         include_wiki_search: bool = True,
     ) -> SmartCreateResponse:
@@ -467,7 +694,7 @@ class LibraryDeskClient:
         Returns:
             SmartCreateResponse with page and research metadata
         """
-        user = user or get_user()
+        user = self._resolve_user(user)
         client = self._ensure_client()
 
         payload: dict[str, Any] = {
@@ -518,12 +745,15 @@ class LibraryDeskClient:
         Returns:
             List of dossiers with page counts
         """
-        user = user or get_user()
+        user = self._resolve_user(user)
         client = self._ensure_client()
 
-        response = await client.get(
-            "/wiki/dossiers",
-            params={"user": user},
+        response = await self._request_with_retry(
+            lambda: client.get(
+                "/wiki/dossiers",
+                params={"user": user},
+            ),
+            "GET /wiki/dossiers",
         )
         response.raise_for_status()
 
@@ -553,7 +783,7 @@ class LibraryDeskClient:
         Returns:
             List of matching document chunks with scores
         """
-        user = user or get_user()
+        user = self._resolve_user(user)
         client = self._ensure_client()
 
         payload = {
@@ -579,7 +809,7 @@ class LibraryDeskClient:
         self,
         cypher_query: str,
         user: str | None = None,
-        parameters: Optional[dict[str, Any]] = None,
+        parameters: dict[str, Any] | None = None,
     ) -> list[dict[str, Any]]:
         """
         Execute a Cypher query on the knowledge graph.
@@ -594,7 +824,7 @@ class LibraryDeskClient:
         Returns:
             List of result records
         """
-        user = user or get_user()
+        user = self._resolve_user(user)
         client = self._ensure_client()
 
         payload = {
@@ -613,7 +843,7 @@ class LibraryDeskClient:
     async def list_graph_nodes(
         self,
         user: str | None = None,
-        node_type: Optional[str] = None,
+        node_type: str | None = None,
         limit: int = 100,
     ) -> list[GraphNode]:
         """
@@ -627,14 +857,17 @@ class LibraryDeskClient:
         Returns:
             List of graph nodes
         """
-        user = user or get_user()
+        user = self._resolve_user(user)
         client = self._ensure_client()
 
         params: dict[str, Any] = {"user": user, "limit": limit}
         if node_type:
             params["node_type"] = node_type
 
-        response = await client.get("/graph/nodes", params=params)
+        response = await self._request_with_retry(
+            lambda: client.get("/graph/nodes", params=params),
+            "GET /graph/nodes",
+        )
         response.raise_for_status()
 
         data = response.json()
@@ -655,12 +888,15 @@ class LibraryDeskClient:
         Returns:
             Node with relationships and connected nodes
         """
-        user = user or get_user()
+        user = self._resolve_user(user)
         client = self._ensure_client()
 
-        response = await client.get(
-            f"/graph/nodes/{node_id}",
-            params={"user": user},
+        response = await self._request_with_retry(
+            lambda: client.get(
+                f"/graph/nodes/{node_id}",
+                params={"user": user},
+            ),
+            f"GET /graph/nodes/{node_id}",
         )
         response.raise_for_status()
 
@@ -679,12 +915,208 @@ class LibraryDeskClient:
         """
         try:
             client = self._ensure_client()
-            response = await client.get("/health")
+            response = await self._request_with_retry(
+                lambda: client.get("/health"),
+                "GET /health",
+            )
             return response.status_code == 200
         except Exception as e:
             logger.warning("library_desk_health_check_failed", error=str(e))
             return False
 
+    # ========================================================================
+    # RAG Search (Web Search with Content Extraction)
+    # ========================================================================
+
+    async def search_web(
+        self,
+        query: str,
+        user: str | None = None,
+        search_type: str = "web",
+        limit: int = 10,
+    ) -> WebSearchResponse:
+        """
+        Search the web and extract content from results.
+
+        Uses SearXNG for search and Trafilatura for content extraction.
+        Returns both snippets and full extracted text.
+
+        Args:
+            query: Search query (1-500 chars)
+            user: User identifier for tracking
+            search_type: "web", "news", or "images"
+            limit: Number of results (1-20)
+
+        Returns:
+            WebSearchResponse with results and pre-formatted sources
+        """
+        user = self._resolve_user(user)
+        client = self._ensure_client()
+
+        payload = {
+            "query": query,
+            "search_type": search_type,
+            "limit": limit,
+            "user": user,
+        }
+
+        logger.info("library_desk_web_search", query=query, limit=limit)
+
+        response = await self._request_with_retry(
+            lambda: client.post("/rag/search", json=payload),
+            "POST /rag/search",
+        )
+        response.raise_for_status()
+
+        data = response.json()
+
+        results = [
+            WebSearchResult(
+                title=r.get("title", ""),
+                url=r.get("url", ""),
+                content=r.get("content", ""),
+                snippet=r.get("snippet", ""),
+                source=r.get("source", ""),
+                published_date=r.get("published_date"),
+            )
+            for r in data.get("results", [])
+        ]
+
+        return WebSearchResponse(
+            query=data.get("query", query),
+            search_type=data.get("search_type", search_type),
+            results=results,
+            total_results=data.get("total_results", len(results)),
+            search_time_ms=data.get("search_time_ms", 0),
+            sources_summary=data.get("sources_summary", ""),
+        )
+
+    # ========================================================================
+    # Content Extraction
+    # ========================================================================
+
+    async def extract_content(
+        self,
+        url: str,
+        user: str | None = None,
+        include_metadata: bool = True,
+        max_length: int = 5000,
+    ) -> ContentExtractionResult:
+        """
+        Extract main content from a URL.
+
+        Uses Trafilatura for intelligent content extraction,
+        removing boilerplate, ads, and navigation.
+
+        Note: Uses soft failure pattern - check result.success field.
+
+        Args:
+            url: URL to extract content from
+            user: User identifier (defaults to request context)
+            include_metadata: Whether to extract author, date, etc.
+            max_length: Maximum content length
+
+        Returns:
+            ContentExtractionResult (check .success and .error fields)
+        """
+        user = self._resolve_user(user)
+        client = self._ensure_client()
+
+        payload = {
+            "url": url,
+            "include_metadata": include_metadata,
+            "max_length": max_length,
+        }
+
+        logger.debug("library_desk_extract_content", url=url)
+
+        response = await client.post(
+            "/content/extract",
+            json=payload,
+            params={"user": user},
+        )
+        response.raise_for_status()
+
+        data = response.json()
+        result = data.get("result", {})
+
+        return ContentExtractionResult(
+            url=result.get("url", url),
+            title=result.get("title"),
+            content=result.get("content", ""),
+            author=result.get("author"),
+            date=result.get("date"),
+            language=result.get("language"),
+            success=result.get("success", False),
+            error=result.get("error"),
+        )
+
+    async def extract_content_batch(
+        self,
+        urls: list[str],
+        user: str | None = None,
+        include_metadata: bool = True,
+        max_length: int = 2000,
+    ) -> BatchExtractionResponse:
+        """
+        Extract content from multiple URLs in parallel.
+
+        More efficient than sequential calls. Max 20 URLs per batch.
+
+        Note: Uses soft failure pattern - individual failures don't
+        throw errors, check each result's .success field.
+
+        Args:
+            urls: List of URLs to extract (max 20)
+            user: User identifier (defaults to request context)
+            include_metadata: Whether to extract author, date, etc.
+            max_length: Maximum content length per URL
+
+        Returns:
+            BatchExtractionResponse with results and stats
+        """
+        user = self._resolve_user(user)
+        client = self._ensure_client()
+
+        payload = {
+            "urls": urls[:20],  # Server limit
+            "include_metadata": include_metadata,
+            "max_length": max_length,
+        }
+
+        logger.info("library_desk_extract_batch", url_count=len(urls))
+
+        response = await client.post(
+            "/content/extract/batch",
+            json=payload,
+            params={"user": user},
+        )
+        response.raise_for_status()
+
+        data = response.json()
+
+        results = [
+            ContentExtractionResult(
+                url=r.get("url", ""),
+                title=r.get("title"),
+                content=r.get("content", ""),
+                author=r.get("author"),
+                date=r.get("date"),
+                language=r.get("language"),
+                success=r.get("success", False),
+                error=r.get("error"),
+            )
+            for r in data.get("results", [])
+        ]
+
+        return BatchExtractionResponse(
+            results=results,
+            total_urls=data.get("total_urls", len(urls)),
+            successful=data.get("successful", 0),
+            failed=data.get("failed", 0),
+            extraction_time_ms=data.get("extraction_time_ms", 0),
+        )
+
 
 # Global client factory
 async def get_library_client() -> LibraryDeskClient:
@@ -696,3 +1128,30 @@ async def get_library_client() -> LibraryDeskClient:
             results = await client.hybrid_search("query")
     """
     return LibraryDeskClient()
+
+
+@asynccontextmanager
+async def library_client_session() -> AsyncIterator[None]:
+    """
+    Hold ONE shared HTTP connection for the duration of a librarian run.
+
+    While the session is active, every LibraryDeskClient targeting the
+    configured library-desk instance reuses the shared httpx client
+    instead of constructing (and tearing down) a connection per tool
+    call. Nested sessions are no-ops.
+
+    Usage:
+        async with library_client_session():
+            ...  # librarian tools reuse one connection
+    """
+    if _shared_http_client.get() is not None:
+        yield
+        return
+
+    http_client = LibraryDeskClient()._build_http_client()
+    token = _shared_http_client.set(http_client)
+    try:
+        yield
+    finally:
+        _shared_http_client.reset(token)
+        await http_client.aclose()

src/agents/librarian/tools.py

@@ -4,33 +4,133 @@ Librarian tools for PydanticAI agent.
 These tools wrap the library-desk API and are registered with
 The Librarian agent for research and knowledge management tasks.
 """
-from src.agents.librarian.client import LibraryDeskClient
+
+import httpx
+from pydantic_ai import ModelRetry
+
+from src.agents.librarian.client import HybridRAGResponse, LibraryDeskClient
 from src.core.logging_config import get_logger
 
 logger = get_logger(__name__)
 
 
+def _retry_if_transient(e: Exception, what: str) -> None:
+    """
+    Convert transient HTTP errors into ModelRetry so the agent's
+    retry budget (Agent(retries=2)) engages instead of the tool
+    swallowing the failure.
+
+    Only read tools call this - writes are never retried to avoid
+    duplicate wiki pages.
+    """
+    retryable = isinstance(e, httpx.TransportError)
+    if isinstance(e, httpx.HTTPStatusError):
+        status = e.response.status_code
+        retryable = status >= 500 or status == 429
+    if retryable:
+        raise ModelRetry(f"{what} is temporarily unavailable; please retry.") from e
+
+
+# Icons keyed by the values library-desk emits in each result's `sources`
+# list (search legs) and `source_type` (result origin).
+SOURCE_ICONS = {
+    "vector": "📄",
+    "graph": "🔗",
+    "web": "🌐",
+    "document": "📑",
+    "documents": "📑",
+    "volatile": "⚡",
+    "wiki": "📄",
+}
+
+
+def _coverage_note(
+    response: HybridRAGResponse,
+    include_web: bool,
+    include_documents: bool,
+    include_volatile: bool,
+) -> str:
+    """
+    Build a one-line coverage note when the search was degraded or an
+    enabled source leg contributed nothing, so outages stay visible to
+    the model and the user instead of silently narrowing results.
+
+    When the additive source_status/degraded contract is present it is
+    authoritative and used EXCLUSIVELY - no count heuristics. Without
+    it, absence from source_counts is only inferred for the optional
+    legs this request explicitly enabled (web/documents/volatile);
+    the always-on wiki legs (vector/graph) are never inferred, because
+    source_counts only tallies the sources of the final top-N fused
+    results, so their absence is normal ranking behavior, not an outage.
+    """
+    if response.source_status:
+        failed = sorted(leg for leg, status in response.source_status.items() if status == "failed")
+        if failed:
+            return (
+                "⚠️ *Coverage note: results are partial - "
+                f"these sources failed: {', '.join(failed)}.*"
+            )
+        if response.degraded:
+            return (
+                "⚠️ *Coverage note: results are partial - "
+                "one or more sources failed during this search.*"
+            )
+        return ""
+
+    if not response.source_counts:
+        # Older library-desk without per-source reporting - nothing to infer
+        return ""
+
+    # Only legs the request explicitly enabled; never vector/graph (their
+    # absence from the top-N counts is healthy, see docstring)
+    expected = set()
+    if include_web:
+        expected.add("web")
+    if include_documents:
+        expected.add("documents")
+    if include_volatile:
+        expected.add("volatile")
+
+    # Normalize count keys to leg names (document/documents)
+    aliases = {"document": "documents"}
+    reported = {aliases.get(key, key) for key in response.source_counts}
+    missing = sorted(expected - reported)
+    if missing:
+        return (
+            "⚠️ *Coverage note: no results came from: "
+            f"{', '.join(missing)} (source unavailable or nothing found).*"
+        )
+    return ""
+
+
 # ============================================================================
 # HybridRAG Search
 # ============================================================================
 
+
 async def hybrid_search(
     query: str,
     include_web: bool = True,
+    include_documents: bool = True,
+    include_volatile: bool = True,
 ) -> str:
     """
     Search across all knowledge sources using HybridRAG.
 
     This is the primary research tool, combining:
-    - Vector search (semantic similarity over documents)
+    - Vector search (semantic similarity over wiki pages)
     - Knowledge graph (entities and relationships)
+    - Paperless documents (📑 indexed PDFs, scans, invoices)
+    - Volatile cache (⚡ weather, news, stocks - for user's configured items)
     - Web search (current information from SearXNG)
 
-    Results are fused and re-ranked by relevance.
+    Results are fused and re-ranked by relevance. Volatile data gets priority when fresh.
 
     Args:
         query: Natural language research query
         include_web: Whether to include web results (default: True)
+        include_documents: Whether to include Paperless documents (default: True)
+        include_volatile: Whether to include volatile cache data (default: True)
 
     Returns:
         Formatted search results with sources and context
@@ -38,12 +138,16 @@ async def hybrid_search(
     Examples:
         hybrid_search("How does Docker orchestration work with Kubernetes?")
         hybrid_search("What projects use Neo4j?", include_web=False)
+        hybrid_search("Find my electricity invoices", include_web=False, include_volatile=False)
+        hybrid_search("What's the weather in Rotterdam?")  # May hit volatile cache
     """
     try:
         async with LibraryDeskClient() as client:
             response = await client.hybrid_search(
                 query=query,
                 web_limit=5 if include_web else 0,
+                document_limit=5 if include_documents else 0,
+                volatile_limit=3 if include_volatile else 0,
             )
 
             if not response.results:
@@ -58,19 +162,16 @@ async def hybrid_search(
 
             # Add related dossiers
             if response.related_dossiers:
-                output_parts.append(
-                    f"**Related Dossiers:** {', '.join(response.related_dossiers)}"
-                )
+                output_parts.append(f"**Related Dossiers:** {', '.join(response.related_dossiers)}")
 
             output_parts.append("")
 
             # Add results
             for i, result in enumerate(response.results, 1):
-                source_icon = {
-                    "vector": "📄",
-                    "graph": "🔗",
-                    "web": "🌐",
-                }.get(result.source, "•")
+                source_keys = result.sources or [result.source]
+                source_icon = "".join(
+                    dict.fromkeys(SOURCE_ICONS.get(key, "•") for key in source_keys)
+                )
 
                 output_parts.append(
                     f"{i}. {source_icon} **{result.title}** (score: {result.score:.2f})"
@@ -80,23 +181,37 @@ async def hybrid_search(
                 output_parts.append(f"   {result.content[:300]}...")
                 output_parts.append("")
 
+            # Surface degraded coverage so outages are visible downstream
+            coverage_note = _coverage_note(
+                response,
+                include_web=include_web,
+                include_documents=include_documents,
+                include_volatile=include_volatile,
+            )
+            if coverage_note:
+                output_parts.append(coverage_note)
+
             logger.info(
                 "librarian_hybrid_search",
                 query=query,
                 result_count=len(response.results),
+                degraded=response.degraded,
+                source_counts=response.source_counts,
             )
 
             return "\n".join(output_parts)
 
     except Exception as e:
         logger.error("librarian_hybrid_search_error", error=str(e), query=query)
-        return f"Error searching: {str(e)}"
+        _retry_if_transient(e, "The knowledge archive")
+        return "I was unable to search the knowledge archives; the search service did not respond properly."
 
 
 # ============================================================================
 # Wiki Operations
 # ============================================================================
 
+
 async def search_wiki(
     query: str,
     limit: int = 10,
@@ -127,18 +242,22 @@ async def search_wiki(
 
             output_parts = [f"## Wiki Search: {query}\n"]
 
-            for i, page in enumerate(results, 1):
-                output_parts.append(f"{i}. **{page.title}**")
-                output_parts.append(f"   Path: {page.path}")
+            # No ordinal numbering: small models pass the list position to
+            # get_wiki_page instead of the page ID unless the ID is the only
+            # number in sight.
+            for page in results:
+                output_parts.append(f"- **{page.title}** (page_id: {page.id})")
+                output_parts.append(f"  Path: {page.path}")
                 if page.description:
-                    output_parts.append(f"   {page.description}")
+                    output_parts.append(f"  {page.description}")
                 output_parts.append("")
 
             return "\n".join(output_parts)
 
     except Exception as e:
         logger.error("librarian_wiki_search_error", error=str(e))
-        return f"Error searching wiki: {str(e)}"
+        _retry_if_transient(e, "The wiki search")
+        return "I was unable to search the wiki at this time."
 
 
 async def get_wiki_page(
@@ -181,7 +300,8 @@ async def get_wiki_page(
 
     except Exception as e:
         logger.error("librarian_get_page_error", error=str(e), page_id=page_id)
-        return f"Error getting page {page_id}: {str(e)}"
+        _retry_if_transient(e, "The wiki")
+        return f"I was unable to retrieve wiki page {page_id}."
 
 
 async def list_dossiers() -> str:
@@ -207,15 +327,14 @@ async def list_dossiers() -> str:
             output_parts = ["## Research Dossiers\n"]
 
             for dossier in dossiers:
-                output_parts.append(
-                    f"- **{dossier.name}** ({dossier.page_count} pages)"
-                )
+                output_parts.append(f"- **{dossier.name}** ({dossier.page_count} pages)")
 
             return "\n".join(output_parts)
 
     except Exception as e:
         logger.error("librarian_list_dossiers_error", error=str(e))
-        return f"Error listing dossiers: {str(e)}"
+        _retry_if_transient(e, "The dossier index")
+        return "I was unable to retrieve the list of dossiers."
 
 
 async def get_dossier_pages(
@@ -256,13 +375,15 @@ async def get_dossier_pages(
 
     except Exception as e:
         logger.error("librarian_get_dossier_error", error=str(e))
-        return f"Error getting dossier: {str(e)}"
+        _retry_if_transient(e, "The dossier index")
+        return f"I was unable to retrieve the dossier '{dossier_name}'."
 
 
 # ============================================================================
 # Semantic Search
 # ============================================================================
 
+
 async def semantic_search(
     query: str,
     limit: int = 10,
@@ -294,9 +415,7 @@ async def semantic_search(
             output_parts = [f"## Semantic Search: {query}\n"]
 
             for i, result in enumerate(results, 1):
-                output_parts.append(
-                    f"{i}. **{result.page_title}** (score: {result.score:.2f})"
-                )
+                output_parts.append(f"{i}. **{result.page_title}** (score: {result.score:.2f})")
                 output_parts.append(f"   Path: {result.page_path}")
                 output_parts.append(f"   {result.chunk_text[:200]}...")
                 output_parts.append("")
@@ -305,13 +424,15 @@ async def semantic_search(
 
     except Exception as e:
         logger.error("librarian_semantic_search_error", error=str(e))
-        return f"Error in semantic search: {str(e)}"
+        _retry_if_transient(e, "The semantic search")
+        return "I was unable to complete the semantic search."
 
 
 # ============================================================================
 # Knowledge Graph
 # ============================================================================
 
+
 async def explore_knowledge_graph(
     entity_type: str = "Document",
     limit: int = 20,
@@ -358,7 +479,8 @@ async def explore_knowledge_graph(
 
     except Exception as e:
         logger.error("librarian_explore_graph_error", error=str(e))
-        return f"Error exploring knowledge graph: {str(e)}"
+        _retry_if_transient(e, "The knowledge graph")
+        return "I was unable to explore the knowledge graph."
 
 
 async def find_related_entities(
@@ -429,19 +551,262 @@ async def find_related_entities(
 
     except Exception as e:
         logger.error("librarian_find_related_error", error=str(e))
-        return f"Error finding related entities: {str(e)}"
+        _retry_if_transient(e, "The knowledge graph")
+        return f"I was unable to look up entities related to '{entity_name}'."
+
+
+# ============================================================================
+# Web Search & Content Extraction
+# ============================================================================
+
+
+async def search_web(
+    query: str,
+    limit: int = 10,
+    search_type: str = "web",
+) -> str:
+    """
+    Search the web and extract content from results.
+
+    This is the primary tool for finding current information online.
+    Results include both snippets and full extracted text from pages.
+
+    Search types:
+    - "web": General web search (default)
+    - "news": News articles
+    - "images": Image search
+
+    Args:
+        query: Search query (1-500 chars)
+        limit: Number of results (1-20, default: 10)
+        search_type: Type of search ("web", "news", or "images")
+
+    Returns:
+        Formatted search results with sources and extracted content
+
+    Examples:
+        search_web("Python 3.12 new features")
+        search_web("latest tech news", search_type="news", limit=5)
+    """
+    try:
+        async with LibraryDeskClient() as client:
+            response = await client.search_web(
+                query=query,
+                limit=limit,
+                search_type=search_type,
+            )
+
+            if not response.results:
+                return f"No results found for '{query}'"
+
+            output_parts = [f"## Web Search: {query}\n"]
+            output_parts.append(
+                f"*Found {response.total_results} results in {response.search_time_ms}ms*\n"
+            )
+
+            for i, result in enumerate(response.results, 1):
+                output_parts.append(f"### {i}. {result.title}")
+                output_parts.append(f"**Source:** {result.source}")
+                output_parts.append(f"**URL:** {result.url}")
+
+                if result.published_date:
+                    output_parts.append(f"**Date:** {result.published_date}")
+
+                # Use full content if available, otherwise snippet
+                content = result.content or result.snippet
+                if content:
+                    # Truncate for readability
+                    if len(content) > 500:
+                        content = content[:500] + "..."
+                    output_parts.append(f"\n{content}")
+
+                output_parts.append("")
+
+            # Add pre-formatted sources for citations
+            if response.sources_summary:
+                output_parts.append("---")
+                output_parts.append(response.sources_summary)
+
+            logger.info(
+                "librarian_web_search",
+                query=query,
+                result_count=response.total_results,
+                search_type=search_type,
+            )
+
+            return "\n".join(output_parts)
+
+    except Exception as e:
+        logger.error("librarian_web_search_error", error=str(e), query=query)
+        _retry_if_transient(e, "The web search")
+        return "I was unable to search the web at this time."
+
+
+async def read_url(
+    url: str,
+    max_length: int = 5000,
+) -> str:
+    """
+    Read and extract the main content from a URL.
+
+    Use this when you have a specific URL to read, such as:
+    - A link the user provided
+    - A URL from search results you want to read in full
+    - Documentation or article pages
+
+    Extracts the main content, removing ads, navigation, and boilerplate.
+
+    Args:
+        url: The URL to read
+        max_length: Maximum content length (default: 5000)
+
+    Returns:
+        Extracted page content with metadata
+
+    Examples:
+        read_url("https://docs.python.org/3/library/asyncio.html")
+        read_url("https://example.com/article", max_length=10000)
+    """
+    try:
+        async with LibraryDeskClient() as client:
+            result = await client.extract_content(
+                url=url,
+                include_metadata=True,
+                max_length=max_length,
+            )
+
+            if not result.success:
+                return f"Could not read page: {result.error or 'Unknown error'}"
+
+            output_parts = []
+
+            # Header with metadata
+            if result.title:
+                output_parts.append(f"# {result.title}")
+            else:
+                output_parts.append(f"# Content from {url}")
+
+            output_parts.append(f"**URL:** {url}")
+
+            if result.author:
+                output_parts.append(f"**Author:** {result.author}")
+
+            if result.date:
+                output_parts.append(f"**Date:** {result.date}")
+
+            if result.language and result.language != "en":
+                output_parts.append(f"**Language:** {result.language}")
+
+            output_parts.append("")
+
+            # Main content
+            if result.content:
+                output_parts.append(result.content)
+            else:
+                output_parts.append("(No content could be extracted)")
+
+            logger.info(
+                "librarian_read_url",
+                url=url,
+                content_length=len(result.content) if result.content else 0,
+            )
+
+            return "\n".join(output_parts)
+
+    except Exception as e:
+        logger.error("librarian_read_url_error", error=str(e), url=url)
+        _retry_if_transient(e, "Content extraction")
+        return f"I was unable to read the page at {url}."
+
+
+async def read_urls_batch(
+    urls: list[str],
+    max_length: int = 2000,
+) -> str:
+    """
+    Read and extract content from multiple URLs in parallel.
+
+    More efficient than calling read_url multiple times.
+    Max 20 URLs per batch.
+
+    Note: Individual failures don't fail the entire batch -
+    failed URLs are reported but other content is still returned.
+
+    Args:
+        urls: List of URLs to read (max 20)
+        max_length: Maximum content length per URL (default: 2000)
+
+    Returns:
+        Extracted content from all successful URLs with failure report
+
+    Examples:
+        read_urls_batch(["https://example.com/1", "https://example.com/2"])
+    """
+    try:
+        async with LibraryDeskClient() as client:
+            response = await client.extract_content_batch(
+                urls=urls,
+                include_metadata=True,
+                max_length=max_length,
+            )
+
+            output_parts = [
+                "## Batch Content Extraction",
+                f"*Extracted {response.successful}/{response.total_urls} URLs in {response.extraction_time_ms}ms*\n",
+            ]
+
+            # Show successful extractions
+            for result in response.results:
+                if result.success:
+                    title = result.title or result.url
+                    output_parts.append(f"### {title}")
+                    output_parts.append(f"**URL:** {result.url}")
+
+                    if result.content:
+                        # Truncate for readability in batch mode
+                        content = result.content
+                        if len(content) > max_length:
+                            content = content[:max_length] + "..."
+                        output_parts.append(f"\n{content}")
+
+                    output_parts.append("")
+
+            # Report failures
+            failed = [r for r in response.results if not r.success]
+            if failed:
+                output_parts.append("---")
+                output_parts.append("### Failed Extractions")
+                for result in failed:
+                    output_parts.append(f"- {result.url}: {result.error}")
+
+            logger.info(
+                "librarian_read_urls_batch",
+                total=response.total_urls,
+                successful=response.successful,
+                failed=response.failed,
+            )
+
+            return "\n".join(output_parts)
+
+    except Exception as e:
+        logger.error("librarian_read_urls_batch_error", error=str(e))
+        _retry_if_transient(e, "Content extraction")
+        return "I was unable to read the requested pages."
 
 
 # ============================================================================
 # Wiki Write Operations
 # ============================================================================
 
+CLEAR_TAGS_SENTINEL = "__CLEAR__"
+
+
 async def update_wiki_page(
     page_id: int,
-    content: str | None = None,
-    title: str | None = None,
-    tags: list[str] | None = None,
-    description: str | None = None,
+    content: str = "",
+    title: str = "",
+    tags: list[str] = [],  # noqa: B006 - sentinel, never mutated
+    description: str = "",
 ) -> str:
     """
     Update an existing wiki page.
@@ -455,12 +820,17 @@ async def update_wiki_page(
     - Updating tags to organize pages into dossiers
     - Fixing descriptions or titles
 
+    Note: empty values are sentinels for "leave unchanged" (Ollama's
+    OpenAI-compatible API mishandles anyOf[X, null] parameter schemas).
+
     Args:
         page_id: ID of the page to update (get from search_wiki results)
-        content: New markdown content (optional - only if changing content)
-        title: New title (optional - only if renaming)
-        tags: New tag list (optional - replaces existing tags)
-        description: New description (optional)
+        content: New markdown content (empty = leave unchanged)
+        title: New title (empty = leave unchanged)
+        tags: New tag list, replaces existing tags (empty = leave unchanged).
+            To remove ALL tags from a page, pass exactly ["__CLEAR__"]
+            (an empty list means "leave unchanged", not "clear")
+        description: New description (empty = leave unchanged)
 
     Returns:
         Confirmation with updated page details
@@ -468,27 +838,34 @@ async def update_wiki_page(
     Examples:
         update_wiki_page(42, content="# Updated Content\\n\\nNew information here")
         update_wiki_page(42, tags=["projects", "devops"])  # Add to dossiers
+        update_wiki_page(42, tags=["__CLEAR__"])  # Remove all tags
         update_wiki_page(42, description="Updated description")
     """
+    # Empty list = leave unchanged; the explicit clear sentinel sends an
+    # empty tag list to the service, which replaces (clears) all tags.
+    clear_tags = tags == [CLEAR_TAGS_SENTINEL]
+
     try:
         async with LibraryDeskClient() as client:
             page = await client.update_wiki_page(
                 page_id=page_id,
-                content=content,
-                title=title,
-                tags=tags,
-                description=description,
+                content=content if content else None,
+                title=title if title else None,
+                tags=[] if clear_tags else (tags if tags else None),
+                description=description if description else None,
             )
 
             # Build update summary
             updated_fields = []
-            if content is not None:
+            if content:
                 updated_fields.append("content")
-            if title is not None:
+            if title:
                 updated_fields.append("title")
-            if tags is not None:
+            if clear_tags:
+                updated_fields.append("tags (cleared)")
+            elif tags:
                 updated_fields.append("tags")
-            if description is not None:
+            if description:
                 updated_fields.append("description")
 
             output_parts = [
@@ -500,7 +877,9 @@ async def update_wiki_page(
             if page.tags:
                 output_parts.append(f"**Tags:** {', '.join(page.tags)}")
 
-            output_parts.append("\n*Vector embeddings and knowledge graph will be updated automatically.*")
+            output_parts.append(
+                "\n*Vector embeddings and knowledge graph will be updated automatically.*"
+            )
 
             logger.info(
                 "librarian_update_page",
@@ -512,7 +891,7 @@ async def update_wiki_page(
 
     except Exception as e:
         logger.error("librarian_update_page_error", error=str(e), page_id=page_id)
-        return f"Error updating page {page_id}: {str(e)}"
+        return f"I was unable to update wiki page {page_id}."
 
 
 async def create_wiki_page(
@@ -574,7 +953,9 @@ async def create_wiki_page(
             if page.description:
                 output_parts.append(f"**Description:** {page.description}")
 
-            output_parts.append("\n*Vector embeddings and knowledge graph will be updated automatically.*")
+            output_parts.append(
+                "\n*Vector embeddings and knowledge graph will be updated automatically.*"
+            )
 
             logger.info(
                 "librarian_create_page",
@@ -587,13 +968,13 @@ async def create_wiki_page(
 
     except Exception as e:
         logger.error("librarian_create_page_error", error=str(e), title=title)
-        return f"Error creating page: {str(e)}"
+        return f"I was unable to create the page '{title}'."
 
 
 async def smart_create_wiki_page(
     topic: str,
     tags: list[str],
-    path: str | None = None,
+    path: str = "",
     include_web_research: bool = True,
     include_wiki_search: bool = True,
 ) -> str:
@@ -615,7 +996,7 @@ async def smart_create_wiki_page(
     Args:
         topic: The topic to research and create a page about
         tags: List of tags/dossiers for categorization
-        path: Optional custom path (auto-generated from topic if not provided)
+        path: Optional custom path (empty = auto-generated from topic)
         include_web_research: Whether to search the web (default: True)
         include_wiki_search: Whether to search existing wiki (default: True)
 
@@ -631,7 +1012,7 @@ async def smart_create_wiki_page(
             response = await client.smart_create_wiki_page(
                 topic=topic,
                 tags=tags,
-                path=path,
+                path=path if path else None,
                 include_web_research=include_web_research,
                 include_wiki_search=include_wiki_search,
             )
@@ -676,7 +1057,7 @@ async def smart_create_wiki_page(
 
     except Exception as e:
         logger.error("librarian_smart_create_error", error=str(e), topic=topic)
-        return f"Error creating page about '{topic}': {str(e)}"
+        return f"I was unable to create a page about '{topic}'."
 
 
 # ============================================================================
@@ -685,7 +1066,7 @@ async def smart_create_wiki_page(
 
 # All tools available to The Librarian
 LIBRARIAN_TOOLS = [
-    # Research tools
+    # Research tools (internal knowledge)
     hybrid_search,
     search_wiki,
     get_wiki_page,
@@ -694,6 +1075,10 @@ LIBRARIAN_TOOLS = [
     semantic_search,
     explore_knowledge_graph,
     find_related_entities,
+    # Web search & content extraction
+    search_web,
+    read_url,
+    read_urls_batch,
     # Write tools
     create_wiki_page,
     update_wiki_page,

src/agents/lorem_tester.py

@@ -12,16 +12,16 @@ infrastructure is real production code.
 import asyncio
 import random
 import secrets
-from typing import AsyncGenerator, Any
+from collections.abc import AsyncGenerator
+from typing import Any
 
 from src.agents.base import AgentInterface, OutputItem
 from src.core.exceptions import (
-    RateLimitError,
-    ContextLengthError,
     APIError,
+    ContextLengthError,
+    RateLimitError,
 )
 
-
 # Mock lorem ipsum content
 LOREM_PARAGRAPHS = [
     "Lorem ipsum dolor sit amet, consectetur adipiscing elit. Sed do eiusmod tempor incididunt ut labore et dolore magna aliqua.",
@@ -46,33 +46,27 @@ MOCK_TOOLS = [
         "description": "Search the knowledge base for relevant information",
         "parameters": {
             "type": "object",
-            "properties": {
-                "query": {"type": "string", "description": "Search query"}
-            },
-            "required": ["query"]
-        }
+            "properties": {"query": {"type": "string", "description": "Search query"}},
+            "required": ["query"],
+        },
     },
     {
         "name": "calculate",
         "description": "Perform mathematical calculations",
         "parameters": {
             "type": "object",
-            "properties": {
-                "expression": {"type": "string", "description": "Math expression"}
-            },
-            "required": ["expression"]
-        }
+            "properties": {"expression": {"type": "string", "description": "Math expression"}},
+            "required": ["expression"],
+        },
     },
     {
         "name": "get_weather",
         "description": "Get current weather for a location",
         "parameters": {
             "type": "object",
-            "properties": {
-                "location": {"type": "string", "description": "City name"}
-            },
-            "required": ["location"]
-        }
+            "properties": {"location": {"type": "string", "description": "City name"}},
+            "required": ["location"],
+        },
     },
 ]
 
@@ -109,7 +103,7 @@ class LoremTesterAgent(AgentInterface):
         temperature: float = 1.0,
         max_tokens: int | None = None,
         stop: list[str] | None = None,
-        **kwargs: Any
+        **kwargs: Any,
     ) -> AsyncGenerator[OutputItem, None]:
         """
         Generate mock response with reasoning, tools, and content.
@@ -123,8 +117,7 @@ class LoremTesterAgent(AgentInterface):
         # 1. Yield reasoning item if requested
         if reasoning and reasoning.get("summary") == "auto":
             yield await self._create_reasoning_item(
-                messages,
-                effort=reasoning.get("effort", "medium")
+                messages, effort=reasoning.get("effort", "medium")
             )
 
         # 2. Randomly yield function calls if tools available (30% chance)
@@ -150,7 +143,7 @@ class LoremTesterAgent(AgentInterface):
             "reasoning": True,
             "tools": True,
             "vision": False,  # Not yet
-            "audio": False,   # Not yet
+            "audio": False,  # Not yet
         }
 
     # Private helper methods
@@ -179,9 +172,7 @@ class LoremTesterAgent(AgentInterface):
             raise APIError("Invalid tool call: tool 'nonexistent' not found (mock trigger)")
 
     async def _create_reasoning_item(
-        self,
-        messages: list[dict],
-        effort: str = "medium"
+        self, messages: list[dict], effort: str = "medium"
     ) -> OutputItem:
         """Create a reasoning output item with mock thinking steps."""
 
@@ -200,23 +191,17 @@ class LoremTesterAgent(AgentInterface):
         steps = random.sample(REASONING_STEPS, min(num_steps, len(REASONING_STEPS)))
 
         return OutputItem(
-            type="reasoning",
-            id=f"rs_{generate_id()}",
-            summary=steps,
-            status="completed"
+            type="reasoning", id=f"rs_{generate_id()}", summary=steps, status="completed"
         )
 
-    async def _create_tool_calls(
-        self,
-        tools: list[dict]
-    ) -> AsyncGenerator[OutputItem, None]:
+    async def _create_tool_calls(self, tools: list[dict]) -> AsyncGenerator[OutputItem, None]:
         """Create mock function call output items."""
 
         # Randomly select 1-2 tools to "call"
         num_calls = random.randint(1, 2)
         selected_tools = random.sample(
-            MOCK_TOOLS[:min(len(MOCK_TOOLS), len(tools))],
-            min(num_calls, len(MOCK_TOOLS), len(tools))
+            MOCK_TOOLS[: min(len(MOCK_TOOLS), len(tools))],
+            min(num_calls, len(MOCK_TOOLS), len(tools)),
         )
 
         for tool in selected_tools:
@@ -228,7 +213,7 @@ class LoremTesterAgent(AgentInterface):
                 id=f"fc_{generate_id()}",
                 name=tool["name"],
                 arguments=args,
-                status="completed"
+                status="completed",
             )
 
     def _generate_mock_args(self, tool: dict) -> str:
@@ -254,11 +239,7 @@ class LoremTesterAgent(AgentInterface):
             # Generic mock arguments
             return json.dumps({"input": "mock_value"})
 
-    async def _create_message_item(
-        self,
-        messages: list[dict],
-        temperature: float
-    ) -> OutputItem:
+    async def _create_message_item(self, messages: list[dict], temperature: float) -> OutputItem:
         """Create final message output item with lorem ipsum content."""
 
         # Select random lorem ipsum paragraphs
@@ -270,10 +251,6 @@ class LoremTesterAgent(AgentInterface):
             type="message",
             id=f"msg_{generate_id()}",
             role="assistant",
-            content=[{
-                "type": "output_text",
-                "text": content,
-                "annotations": []
-            }],
-            status="completed"
+            content=[{"type": "output_text", "text": content, "annotations": []}],
+            status="completed",
         )

src/agents/orchestration.py

@@ -14,12 +14,13 @@ Supports:
 - Result aggregation from multiple experts
 - Partial failure handling
 """
+
 import asyncio
+from collections.abc import AsyncGenerator
 from dataclasses import dataclass, field
 from enum import Enum
-from typing import AsyncGenerator, Optional, Callable, Any
 
-from src.agents.delegation import DelegationTask, DelegationResult, delegate_to_librarian
+from src.agents.delegation import DelegationResult, DelegationTask, delegate_to_librarian
 from src.core.logging_config import get_logger
 
 logger = get_logger(__name__)
@@ -27,8 +28,9 @@ logger = get_logger(__name__)
 
 class ExecutionMode(str, Enum):
     """Execution mode for multi-expert coordination."""
+
     SEQUENTIAL = "sequential"  # One at a time, in order
-    PARALLEL = "parallel"      # All at once, concurrently
+    PARALLEL = "parallel"  # All at once, concurrently
 
 
 @dataclass
@@ -38,12 +40,13 @@ class OrchestrationContext:
 
     Tracks the user's request, delegation tasks, and results.
     """
+
     user_message: str
     steward_note: str
-    conversation_id: Optional[str] = None
+    conversation_id: str | None = None
 
 
-def parse_delegation_from_steward_note(steward_note: str) -> Optional[DelegationTask]:
+def parse_delegation_from_steward_note(steward_note: str) -> DelegationTask | None:
     """
     Parse a delegation task from Steward's note.
 
@@ -68,9 +71,9 @@ def parse_delegation_from_steward_note(steward_note: str) -> Optional[Delegation
     # Look for DELEGATE: pattern
     # Match: "DELEGATE: expert_name to action description"
     match = re.search(
-        r'DELEGATE:\s*(\w+)\s+to\s+(.+?)(?:\n|REASON:|COMPLEXITY:|CONTEXT:|$)',
+        r"DELEGATE:\s*(\w+)\s+to\s+(.+?)(?:\n|REASON:|COMPLEXITY:|CONTEXT:|$)",
         steward_note,
-        re.IGNORECASE | re.MULTILINE
+        re.IGNORECASE | re.MULTILINE,
     )
 
     if match:
@@ -135,7 +138,7 @@ async def execute_delegation(
 async def orchestrate_with_think_updates(
     user_message: str,
     steward_note: str,
-    delegation_task: Optional[DelegationTask] = None,
+    delegation_task: DelegationTask | None = None,
 ) -> AsyncGenerator[str, None]:
     """
     Orchestrate expert delegation with streaming think updates.
@@ -176,19 +179,19 @@ async def orchestrate_with_think_updates(
     if delegation_task.expert_name == "librarian":
         expert_display_name = "The Librarian"
 
-    yield f"<think>🤝 Consulting {expert_display_name}...</think>\n"
+    yield f"🤝 Consulting {expert_display_name}...\n"
 
     # Execute delegation (uses run() internally)
     result = await execute_delegation(delegation_task)
 
     if result.success:
-        yield f"<think>✅ {expert_display_name} completed research.</think>\n"
+        yield f"✅ {expert_display_name} completed research.\n"
 
         # Yield the expert's findings
         if result.output:
             yield f"\n{result.output}"
     else:
-        yield f"<think>⚠️ {expert_display_name} encountered an issue: {result.error}</think>\n"
+        yield f"⚠️ {expert_display_name} encountered an issue: {result.error}\n"
 
     logger.info(
         "orchestration_complete",
@@ -218,17 +221,21 @@ def extract_delegation_context(
     }
 
     # Extract REASON:
-    reason_match = re.search(r'REASON:\s*(.+?)(?:\n|COMPLEXITY:|CONTEXT:|$)', steward_note, re.IGNORECASE)
+    reason_match = re.search(
+        r"REASON:\s*(.+?)(?:\n|COMPLEXITY:|CONTEXT:|$)", steward_note, re.IGNORECASE
+    )
     if reason_match:
         result["reason"] = reason_match.group(1).strip()
 
     # Extract COMPLEXITY:
-    complexity_match = re.search(r'COMPLEXITY:\s*(.+?)(?:\n|CONTEXT:|$)', steward_note, re.IGNORECASE)
+    complexity_match = re.search(
+        r"COMPLEXITY:\s*(.+?)(?:\n|CONTEXT:|$)", steward_note, re.IGNORECASE
+    )
     if complexity_match:
         result["complexity"] = complexity_match.group(1).strip()
 
     # Extract CONTEXT:
-    context_match = re.search(r'CONTEXT:\s*(.+?)$', steward_note, re.IGNORECASE | re.MULTILINE)
+    context_match = re.search(r"CONTEXT:\s*(.+?)$", steward_note, re.IGNORECASE | re.MULTILINE)
     if context_match:
         result["context"] = context_match.group(1).strip()
 
@@ -239,6 +246,7 @@ def extract_delegation_context(
 # Multi-Expert Coordination
 # ============================================================================
 
+
 @dataclass
 class MultiExpertResult:
     """
@@ -250,6 +258,7 @@ class MultiExpertResult:
         failed_experts: List of expert names that failed
         combined_output: Aggregated output from all successful experts
     """
+
     results: dict[str, DelegationResult] = field(default_factory=dict)
     all_succeeded: bool = True
     failed_experts: list[str] = field(default_factory=list)
@@ -449,12 +458,12 @@ async def orchestrate_multi_expert(
         return
 
     # Stream: Starting multi-expert coordination
-    yield f"<think>🎯 Starting multi-expert coordination ({len(tasks)} tasks, {mode.value})...</think>\n"
+    yield f"🎯 Starting multi-expert coordination ({len(tasks)} tasks, {mode.value})...\n"
 
     if mode == ExecutionMode.PARALLEL:
         # Parallel execution - emit one update then run all at once
         expert_names = ", ".join(_get_display_name(t.expert_name) for t in tasks)
-        yield f"<think>🔄 Consulting in parallel: {expert_names}...</think>\n"
+        yield f"🔄 Consulting in parallel: {expert_names}...\n"
 
         result = await execute_parallel(tasks)
 
@@ -462,9 +471,9 @@ async def orchestrate_multi_expert(
         for expert_name, expert_result in result.results.items():
             display_name = _get_display_name(expert_name)
             if expert_result.success:
-                yield f"<think>✅ {display_name} completed.</think>\n"
+                yield f"✅ {display_name} completed.\n"
             else:
-                yield f"<think>⚠️ {display_name} failed: {expert_result.error}</think>\n"
+                yield f"⚠️ {display_name} failed: {expert_result.error}\n"
 
     else:
         # Sequential execution - emit updates for each task
@@ -472,27 +481,27 @@ async def orchestrate_multi_expert(
 
         for task in tasks:
             display_name = _get_display_name(task.expert_name)
-            yield f"<think>🤝 Consulting {display_name}...</think>\n"
+            yield f"🤝 Consulting {display_name}...\n"
 
             task_result = await execute_delegation(task)
             result.add_result(task_result)
 
             if task_result.success:
-                yield f"<think>✅ {display_name} completed.</think>\n"
+                yield f"✅ {display_name} completed.\n"
             else:
-                yield f"<think>⚠️ {display_name} failed: {task_result.error}</think>\n"
+                yield f"⚠️ {display_name} failed: {task_result.error}\n"
                 if stop_on_failure:
-                    yield "<think>🛑 Stopping due to failure.</think>\n"
+                    yield "🛑 Stopping due to failure.\n"
                     break
 
         result.aggregate_outputs()
 
     # Stream: Summary
     if result.all_succeeded:
-        yield "<think>🎉 All experts completed successfully.</think>\n"
+        yield "🎉 All experts completed successfully.\n"
     else:
         failed_names = ", ".join(_get_display_name(e) for e in result.failed_experts)
-        yield f"<think>⚠️ Some experts failed: {failed_names}</think>\n"
+        yield f"⚠️ Some experts failed: {failed_names}\n"
 
     # Yield combined output
     if result.combined_output:

src/agents/protocol.py

@@ -1,180 +1,11 @@
 """
-Agent communication protocol for multi-agent coordination.
+Agent error protocol.
 
-Defines standardized request/response formats for communication between:
-- Steward (request analysis) → Tatlock (coordination)
-- Tatlock (coordination) → Expert agents (Librarian, Developer, etc.)
+Structured exceptions raised by expert agents (e.g. The Librarian) so
+callers - the delegation wrappers in src/agents/delegation.py - can
+report success=False and map failures to curated user-safe messages
+while exception detail stays in the logs.
 """
-from enum import Enum
-from typing import Any, Optional
-
-from pydantic import BaseModel, Field
-
-
-class DelegationReason(str, Enum):
-    """Why a task is being delegated to an expert agent."""
-    DOMAIN_EXPERTISE = "domain_expertise"  # Expert has specialized knowledge
-    TOOL_ACCESS = "tool_access"  # Expert has required tools
-    RESOURCE_EFFICIENCY = "resource_efficiency"  # Better handled by specialist
-    USER_PREFERENCE = "user_preference"  # User requested specific agent
-
-
-class TaskComplexity(str, Enum):
-    """Complexity estimate for task execution."""
-    SIMPLE = "simple"  # Single tool call, fast
-    MODERATE = "moderate"  # Multiple steps, moderate time
-    COMPLEX = "complex"  # Multi-agent, significant processing
-
-
-class AgentRequest(BaseModel):
-    """
-    Request to an expert agent.
-
-    Contains everything the agent needs to execute a task,
-    including context from the conversation and delegation intent.
-    """
-    task: str = Field(
-        ...,
-        description="Clear description of what the agent should do"
-    )
-    context: str = Field(
-        default="",
-        description="Relevant context from conversation history"
-    )
-    constraints: list[str] = Field(
-        default_factory=list,
-        description="Any constraints or requirements for the task"
-    )
-    delegation_reason: DelegationReason = Field(
-        default=DelegationReason.DOMAIN_EXPERTISE,
-        description="Why this task was delegated to this agent"
-    )
-    user_id: str = Field(
-        default="default",
-        description="User identifier for multi-tenant operations"
-    )
-    max_tokens: Optional[int] = Field(
-        default=None,
-        description="Optional token limit for response"
-    )
-    timeout_seconds: Optional[int] = Field(
-        default=60,
-        description="Maximum time for task completion"
-    )
-
-
-class ToolCallRecord(BaseModel):
-    """Record of a tool call made during execution."""
-    tool_name: str
-    arguments: dict[str, Any]
-    result: str
-    duration_ms: int
-
-
-class AgentResponse(BaseModel):
-    """
-    Response from an expert agent.
-
-    Contains the result, reasoning, and metadata about execution.
-    """
-    success: bool = Field(
-        ...,
-        description="Whether the task completed successfully"
-    )
-    result: str = Field(
-        ...,
-        description="The main output/answer from the agent"
-    )
-    reasoning: str = Field(
-        default="",
-        description="Agent's reasoning process (for transparency)"
-    )
-    tool_calls: list[ToolCallRecord] = Field(
-        default_factory=list,
-        description="Tools called during execution"
-    )
-    confidence: float = Field(
-        default=1.0,
-        ge=0.0,
-        le=1.0,
-        description="Agent's confidence in the result (0.0-1.0)"
-    )
-    sources: list[str] = Field(
-        default_factory=list,
-        description="Sources or references used"
-    )
-    error_message: Optional[str] = Field(
-        default=None,
-        description="Error details if success=False"
-    )
-    duration_ms: int = Field(
-        default=0,
-        description="Total execution time in milliseconds"
-    )
-
-
-class DelegationIntent(BaseModel):
-    """
-    Intent to delegate a task to an expert agent.
-
-    Created by Tatlock when deciding to delegate, based on
-    Steward's recommendations.
-    """
-    target_agent: str = Field(
-        ...,
-        description="Name of the expert agent to delegate to"
-    )
-    task: str = Field(
-        ...,
-        description="Task description for the agent"
-    )
-    reason: DelegationReason = Field(
-        default=DelegationReason.DOMAIN_EXPERTISE,
-        description="Why delegating to this agent"
-    )
-    expected_outcome: str = Field(
-        default="",
-        description="What we expect the agent to provide"
-    )
-    priority: int = Field(
-        default=1,
-        ge=1,
-        le=10,
-        description="Priority (1=highest, 10=lowest)"
-    )
-    depends_on: list[str] = Field(
-        default_factory=list,
-        description="Other delegation IDs this depends on (for sequencing)"
-    )
-
-
-class CoordinationResult(BaseModel):
-    """
-    Result of multi-agent coordination.
-
-    Aggregates results from multiple expert agents into
-    a single coherent response.
-    """
-    final_response: str = Field(
-        ...,
-        description="Synthesized response from all agents"
-    )
-    agent_responses: dict[str, AgentResponse] = Field(
-        default_factory=dict,
-        description="Individual responses keyed by agent name"
-    )
-    delegation_intents: list[DelegationIntent] = Field(
-        default_factory=list,
-        description="All delegations that were executed"
-    )
-    total_duration_ms: int = Field(
-        default=0,
-        description="Total coordination time"
-    )
-    agents_consulted: list[str] = Field(
-        default_factory=list,
-        description="Names of agents that contributed"
-    )
 
 
 class AgentError(Exception):
@@ -184,18 +15,3 @@ class AgentError(Exception):
         self.message = message
         self.agent_name = agent_name
         super().__init__(f"[{agent_name}] {message}")
-
-
-class AgentTimeoutError(AgentError):
-    """Agent execution timed out."""
-    pass
-
-
-class AgentUnavailableError(AgentError):
-    """Agent is not available or registered."""
-    pass
-
-
-class DelegationError(AgentError):
-    """Error during task delegation."""
-    pass

src/agents/registry.py

@@ -8,9 +8,6 @@ It provides a central place to:
 - Check model capabilities
 """
 
-import time
-from typing import Type
-
 from src.agents.base import AgentInterface
 from src.agents.lorem_tester import LoremTesterAgent
 from src.agents.tatlock import TatlockAgent
@@ -63,7 +60,7 @@ class ModelRegistry:
         if model_id not in cls.MODELS:
             raise ModelNotFoundError(model_id)
 
-        agent_class: Type[AgentInterface] = cls.MODELS[model_id]["agent_class"]
+        agent_class: type[AgentInterface] = cls.MODELS[model_id]["agent_class"]
         return agent_class()
 
     @classmethod
@@ -106,14 +103,16 @@ class ModelRegistry:
             agent = cls.get_agent(model_id)
             capabilities = await agent.get_capabilities()
 
-            models.append({
-                "id": model_id,
-                "object": "model",
-                "created": config["created"],
-                "owned_by": config["owned_by"],
-                "capabilities": capabilities,
-                "description": config["description"],
-            })
+            models.append(
+                {
+                    "id": model_id,
+                    "object": "model",
+                    "created": config["created"],
+                    "owned_by": config["owned_by"],
+                    "capabilities": capabilities,
+                    "description": config["description"],
+                }
+            )
 
         return models
 

src/agents/steward/__init__.py

@@ -4,6 +4,7 @@ Steward agent package.
 The Steward analyzes incoming requests and recommends relevant household
 capabilities, creating a two-tier architecture with the Butler.
 """
+
 from .agent import StewardAgent, get_steward_agent
 from .schemas import ConversationContext, StewardRecommendation
 from .service import analyze_request, format_steward_note

src/agents/steward/agent.py

@@ -5,11 +5,13 @@ The Steward analyzes incoming requests, identifies relevant household
 capabilities, and provides focused recommendations to Tatlock (the Butler).
 This creates a two-tier architecture that prevents cognitive overload.
 
-Uses plain text output (not JSON) for reliability with Ollama models.
+Uses plain text output (not JSON) for reliability. Supports both Claude
+(preferred) and Ollama (fallback) backends via direct API calls.
 """
-import httpx
-from typing import Optional
 
+import httpx
+
+from src.anthropic.model_selector import get_model_info, is_claude_available, resolve_backend
 from src.core.config import config
 from src.core.household_registry import get_household_registry
 from src.core.logging_config import get_logger
@@ -27,9 +29,7 @@ def build_steward_prompt(query: str, conversation_history: list[dict]) -> str:
 
     cap_list = []
     for cap in capabilities:
-        cap_list.append(
-            f"• {cap.name} - {cap.description} (domains: {', '.join(cap.domains)})"
-        )
+        cap_list.append(f"• {cap.name} - {cap.description} (domains: {', '.join(cap.domains)})")
     capabilities_text = "\n".join(cap_list)
 
     # Format conversation history if present
@@ -56,13 +56,22 @@ USER QUERY: {query}
 GUIDELINES:
 - Be conservative - only recommend truly necessary capabilities
 - Simple greetings/chat → no capabilities needed (conversational response only)
-- Questions about prior conversation ("what did I say", "my name", "what we discussed") → no capabilities (Tatlock has full history)
+- Questions about prior conversation ("what did I say", "what we discussed") → no capabilities (Tatlock has full history)
 - Math/calculations → tatlock_core
-- Quick web searches → tatlock_core
 - Time/date queries → tatlock_core
+- PERSONAL MEMORY queries → biographer to recall (ALWAYS use for questions about the user themselves):
+  - "where do I live", "what's my location", "my address" → biographer to recall location
+  - "what's my name", "who am I" → biographer to recall name
+  - "what car do I drive", "my vehicle" → biographer to recall car
+  - "what do you know about me", "what have I told you" → biographer to recall or list_memories
+  - "remember that I...", "store that..." → biographer to store_insight
+  - "forget my...", "delete..." → biographer to forget_memory
+  - "my timezone", "my preferences" → biographer to recall preferences
+- Web searches, weather, news, current information → librarian with search_web
+- Read a URL or article → librarian with read_url
 - Wiki creation ("create a page about X", "add X to wiki") → librarian with smart_create
 - Wiki updates ("update the page", "add to dossier") → librarian with update
-- Research queries ("find info", "what do we know about", "search for") → librarian with hybrid_search
+- Research queries about TOPICS (not about the user) → librarian with hybrid_search
 - In-depth research, knowledge synthesis, document lookup → librarian with hybrid_search
 - If conversation history is relevant, note which previous turns matter
 - Assess complexity: simple (1 tool), moderate (2-3 tools), complex (multiple steps)
@@ -74,8 +83,14 @@ COMPLEXITY: [simple/moderate/complex]
 CONTEXT: [any relevant conversation context, or "none"]
 
 EXAMPLES:
+- "DELEGATE: biographer to recall the user's location" (for "where do I live?")
+- "DELEGATE: biographer to recall the user's car" (for "what car do I drive?")
+- "DELEGATE: biographer to list_memories about the user" (for "what do you know about me?")
+- "DELEGATE: biographer to store_insight about user's pet" (for "remember that I have a dog named Max")
+- "DELEGATE: librarian to search_web for tomorrow's weather forecast"
 - "DELEGATE: librarian to create a wiki page about CI/CD pipelines"
-- "DELEGATE: librarian to search for information about Docker networking"
+- "DELEGATE: librarian to hybrid_search for information about Docker networking"
+- "DELEGATE: librarian to read_url https://example.com/article"
 - "DELEGATE: tatlock_core to calculate the result"
 - "DELEGATE: none (conversational response only)"
 
@@ -90,30 +105,82 @@ class StewardAgent:
     Analyzes requests with full conversation context and recommends
     which household capabilities the Butler should use.
 
-    Uses plain text output for reliability with Ollama models.
+    Uses plain text output for reliability. Supports both Claude
+    (preferred) and Ollama (fallback) backends via direct API calls.
     """
 
-    def __init__(self):
-        """Initialize Steward with Ollama model (same as Tatlock for VRAM efficiency)."""
-        self.ollama_host = str(config.OLLAMA_HOST).rstrip('/')
-        self.model_name = config.OLLAMA_DEFAULT_MODEL
-        self.timeout = 30.0  # 30 second timeout for analysis
+    def __init__(self) -> None:
+        """Initialize Steward with backend selection based on availability."""
+        # Ollama config (primary)
+        self.ollama_host = str(config.OLLAMA_HOST).rstrip("/")
+        self.ollama_model = config.OLLAMA_DEFAULT_MODEL
 
+        # Claude config (fallback)
+        self.claude_model = config.ANTHROPIC_MODEL
+        self._anthropic_client = None
+
+        # Determine which backend to use (Ollama-first, Claude when
+        # preferred via config or when Ollama is down)
+        self._use_claude = resolve_backend() == "claude"
+
+        self.timeout = float(config.STEWARD_TIMEOUT)
+
+        model_info = get_model_info()
         logger.info(
             "steward_agent_created",
-            ollama_host=self.ollama_host,
-            model=self.model_name,
+            backend=model_info["backend"],
+            model=model_info["model"],
             timeout=self.timeout,
         )
 
-    async def analyze(
-        self,
-        query: str,
-        conversation_history: Optional[list[dict]] = None
-    ) -> str:
+    def _get_anthropic_client(self):
+        """Get or create Anthropic client (lazy initialization)."""
+        if self._anthropic_client is None:
+            from anthropic import AsyncAnthropic
+
+            self._anthropic_client = AsyncAnthropic(api_key=config.ANTHROPIC_API_KEY)
+        return self._anthropic_client
+
+    async def _call_claude(self, system_prompt: str, user_message: str) -> str:
+        """Call Claude API directly for plain text generation."""
+        client = self._get_anthropic_client()
+
+        # No temperature: rejected by Claude Sonnet 5+ (sampling params deprecated)
+        response = await client.messages.create(
+            model=self.claude_model,
+            max_tokens=1024,
+            system=system_prompt,
+            messages=[{"role": "user", "content": user_message}],
+        )
+
+        return response.content[0].text.strip()
+
+    async def _call_ollama(self, prompt: str) -> str:
+        """Call Ollama API directly for plain text generation."""
+        async with httpx.AsyncClient(timeout=self.timeout) as client:
+            response = await client.post(
+                f"{self.ollama_host}/api/generate",
+                json={
+                    "model": self.ollama_model,
+                    "prompt": prompt,
+                    "stream": False,
+                    "options": {
+                        "temperature": 0.3,  # Lower = more consistent
+                        "top_p": 0.9,
+                    },
+                },
+            )
+
+            response.raise_for_status()
+            result = response.json()
+            return result["response"].strip()
+
+    async def analyze(self, query: str, conversation_history: list[dict] | None = None) -> str:
         """
         Analyze query and return plain text recommendation.
 
+        Uses Claude if available, falls back to Ollama.
+
         Args:
             query: User's query to analyze
             conversation_history: Previous conversation turns
@@ -129,35 +196,61 @@ class StewardAgent:
         history = conversation_history or []
         prompt = build_steward_prompt(query, history)
 
-        logger.debug("steward_calling_ollama", query_preview=query[:100])
+        backend = "claude" if self._use_claude else "ollama"
+        logger.debug(
+            "steward_calling_llm",
+            backend=backend,
+            query_preview=query[:100],
+        )
 
-        # Call Ollama API directly (more reliable than PydanticAI for plain text)
-        async with httpx.AsyncClient(timeout=self.timeout) as client:
-            response = await client.post(
-                f"{self.ollama_host}/api/generate",
-                json={
-                    "model": self.model_name,
-                    "prompt": prompt,
-                    "stream": False,
-                    "options": {
-                        "temperature": 0.3,  # Lower = more consistent
-                        "top_p": 0.9
-                    }
-                }
-            )
-
-            response.raise_for_status()
-            result = response.json()
-
-            analysis_text = result["response"].strip()
+        try:
+            if self._use_claude:
+                # For Claude, split into system + user message
+                # The prompt contains both, but Claude prefers explicit system
+                analysis_text = await self._call_claude(
+                    system_prompt="You are the Steward of the household, advising the Butler (Tatlock) on which capabilities to use. Be concise and specific.",
+                    user_message=prompt,
+                )
+            else:
+                analysis_text = await self._call_ollama(prompt)
 
             logger.debug(
                 "steward_analysis_received",
-                text_preview=analysis_text[:150]
+                backend=backend,
+                text_preview=analysis_text[:150],
             )
 
             return analysis_text
 
+        except Exception as e:
+            # Mid-request fallback: retry on the other backend when possible
+            if self._use_claude:
+                logger.warning(
+                    "steward_claude_fallback",
+                    error=str(e),
+                )
+                analysis_text = await self._call_ollama(prompt)
+                fallback_backend = "ollama_fallback"
+            elif is_claude_available():
+                logger.warning(
+                    "steward_ollama_fallback",
+                    error=str(e),
+                )
+                analysis_text = await self._call_claude(
+                    system_prompt="You are the Steward of the household, advising the Butler (Tatlock) on which capabilities to use. Be concise and specific.",
+                    user_message=prompt,
+                )
+                fallback_backend = "claude_fallback"
+            else:
+                raise
+
+            logger.debug(
+                "steward_analysis_received",
+                backend=fallback_backend,
+                text_preview=analysis_text[:150],
+            )
+            return analysis_text
+
 
 # Global Steward instance
 _steward_agent = None

src/agents/steward/schemas.py

@@ -4,7 +4,8 @@ Steward agent schemas.
 Defines the structured output models for Steward's request analysis
 and capability recommendations.
 """
-from typing import Any, Literal, Optional
+
+from typing import Any, Literal
 
 from pydantic import BaseModel, Field
 
@@ -16,16 +17,16 @@ class ConversationContext(BaseModel):
     The Steward analyzes the full conversation to identify references
     to previous topics, helping the Butler maintain context.
     """
+
     has_previous_context: bool = Field(
         description="Whether the current request references previous conversation turns"
     )
     relevant_turns: list[int] = Field(
         default_factory=list,
-        description="0-indexed turn numbers that are relevant to the current request"
+        description="0-indexed turn numbers that are relevant to the current request",
     )
     context_summary: str = Field(
-        default="",
-        description="Brief summary of relevant context for the Butler"
+        default="", description="Brief summary of relevant context for the Butler"
     )
 
 
@@ -40,25 +41,28 @@ class StewardRecommendation(BaseModel):
     - Conversation context
     - Missing capabilities (if any)
     """
+
     recommended_capabilities: list[str] = Field(
         description="List of household member names to include (e.g., ['tatlock_core'])"
     )
-    reasoning: str = Field(
-        description="Explanation of why these capabilities were recommended"
-    )
+    reasoning: str = Field(description="Explanation of why these capabilities were recommended")
     estimated_complexity: Literal["simple", "moderate", "complex"] = Field(
         description="Complexity assessment: simple (1 tool), moderate (2-3 tools), complex (multiple tools/steps)"
     )
     conversation_context: ConversationContext = Field(
         description="Contextual information from conversation history"
     )
-    missing_capabilities: Optional[str] = Field(
+    missing_capabilities: str | None = Field(
         default=None,
-        description="Description of capabilities that would be helpful but aren't available"
+        description="Description of capabilities that would be helpful but aren't available",
     )
     memory_context: dict[str, Any] = Field(
         default_factory=dict,
-        description="Pre-fetched user context from memory (profile, preferences)"
+        description="Pre-fetched user context from memory (profile, preferences)",
+    )
+    enriched_query: str = Field(
+        default="",
+        description="User query with auto-filled context (location, timezone) when not specified",
     )
 
     def format_for_butler(self) -> str:
@@ -109,6 +113,17 @@ class StewardRecommendation(BaseModel):
                 prefs_str = ", ".join(f"{k}={v}" for k, v in preferences.items())
                 lines.append(f"  • preferences: {prefs_str}")
 
+        # Add delegation instructions when expert agents are recommended
+        delegation_agents = [
+            c for c in self.recommended_capabilities if c in ("biographer", "librarian")
+        ]
+        if delegation_agents:
+            lines.append("-" * 40)
+            lines.append("DELEGATION REQUIRED:")
+            for agent in delegation_agents:
+                lines.append(f'  Call: delegate_to_{agent}(task="[user request]")')
+                lines.append(f'  Or output: [DELEGATE:{agent}] task="[user request]"')
+
         lines.append("=" * 40)
 
         return "\n".join(lines)

src/agents/steward/service.py

@@ -1,54 +1,108 @@
 """
 Steward service layer.
 
-Provides high-level interface for request analysis with logging,
-benchmarking, and error handling.
+Provides high-level interface for request analysis with logging
+and error handling.
 
 Parses plain text recommendations into structured data.
 Includes memory pre-fetch for user context injection.
 """
-import re
-from typing import Any, Optional
 
-from src.core.benchmarks import PerformanceBenchmark, get_benchmark_store
+import re
+from typing import Any
+
 from src.core.household_registry import get_household_registry
 from src.core.logging_config import get_logger, log_operation
 from src.core.memory_service import memory_service
+
 from .agent import get_steward_agent
 from .schemas import ConversationContext, StewardRecommendation
 
 logger = get_logger(__name__)
 
 
+_DELEGATE_LINE_RE = re.compile(r"^[ \t]*DELEGATE:[ \t]*(.+)$", re.IGNORECASE | re.MULTILINE)
+
+
+def _mentions(needle: str, haystack: str) -> bool:
+    """Whole-word containment. Substring matching is what made this go wrong."""
+    return re.search(rf"(?<!\w){re.escape(needle)}(?!\w)", haystack) is not None
+
+
 def _extract_capabilities(text: str) -> list[str]:
     """
-    Extract capability names from Steward's text response.
+    Extract capability names from the Steward's declared delegation.
 
-    Uses keyword matching to find mentioned capabilities.
+    The prompt instructs the Steward to answer in a fixed shape::
+
+        DELEGATE: <capability> to <action> <task>
+        REASON: ...
+        COMPLEXITY: ...
+        CONTEXT: ...
+
+    Only the DELEGATE line states intent; the rest is free prose. An earlier
+    version substring-matched capability *domains* across the whole response,
+    which routed on ordinary English: "description" contains "script" and
+    "discover" contains "cover" (both housekeeper domains), "acknowledge"
+    contains "knowledge" and "know" (librarian, biographer), and "economy"
+    contains "my" (biographer). Any REASON line could therefore summon agents
+    the Steward never asked for, and a spurious librarian is a real
+    multi-second web call.
+
+    It also made prose length a routing input, so anything that shortened the
+    Steward's output — such as disabling model thinking — would look like it had
+    improved routing.
+
+    Resolution is layered, most explicit first:
+      1. a DELEGATE line beginning with a capability name — the documented shape
+      2. a capability named anywhere on a DELEGATE line
+      3. a capability *domain* on a DELEGATE line, for a loosely worded answer
+      4. no DELEGATE line: capability names only, never domains
 
     Args:
         text: Steward's plain text analysis
 
     Returns:
-        List of capability names (e.g., ['tatlock_core'])
+        List of capability names (e.g. ['tatlock_core']), de-duplicated.
     """
-    text_lower = text.lower()
     registry = get_household_registry()
     capabilities = registry.get_all_capabilities()
+    delegate_lines = [line.strip().lower() for line in _DELEGATE_LINE_RE.findall(text or "")]
 
-    found_caps = []
+    found_caps: list[str] = []
 
-    for cap in capabilities:
-        # Check if capability name is mentioned
-        if cap.name.lower() in text_lower:
-            found_caps.append(cap.name)
+    def _add(name: str) -> None:
+        if name not in found_caps:
+            found_caps.append(name)
+
+    if not delegate_lines:
+        # Either the Steward judged no capability necessary — the prompt's
+        # conversational path, whose correct answer is [] — or it ignored the
+        # format. Names only: domain words are ordinary English and would fire
+        # on any prose, which is the bug described above.
+        haystack = (text or "").lower()
+        for cap in capabilities:
+            if _mentions(cap.name.lower(), haystack):
+                _add(cap.name)
+        return found_caps
+
+    for line in delegate_lines:
+        leading = next((c for c in capabilities if line.startswith(c.name.lower())), None)
+        if leading is not None:
+            _add(leading.name)
             continue
 
-        # Check if any domains are mentioned
-        for domain in cap.domains:
-            if domain.lower() in text_lower:
-                found_caps.append(cap.name)
-                break
+        named = [c for c in capabilities if _mentions(c.name.lower(), line)]
+        if named:
+            for cap in named:
+                _add(cap.name)
+            continue
+
+        # Last resort. Scoped to this line, so the REASON and CONTEXT prose that
+        # caused the original misrouting can no longer reach it.
+        for cap in capabilities:
+            if any(_mentions(domain.lower(), line) for domain in cap.domains):
+                _add(cap.name)
 
     return found_caps
 
@@ -74,8 +128,7 @@ def _extract_complexity(text: str) -> str:
 
 
 def _extract_conversation_context(
-    text: str,
-    conversation_history: list[dict]
+    text: str, conversation_history: list[dict]
 ) -> ConversationContext:
     """
     Extract conversation context analysis from text.
@@ -90,13 +143,15 @@ def _extract_conversation_context(
     text_lower = text.lower()
 
     # Check if conversation history is referenced
-    has_context = bool(conversation_history) and any([
-        "previous" in text_lower,
-        "earlier" in text_lower,
-        "context" in text_lower,
-        "turn" in text_lower,
-        "history" in text_lower,
-    ])
+    has_context = bool(conversation_history) and any(
+        [
+            "previous" in text_lower,
+            "earlier" in text_lower,
+            "context" in text_lower,
+            "turn" in text_lower,
+            "history" in text_lower,
+        ]
+    )
 
     # Extract turn numbers if mentioned (e.g., "turn 0", "turn 1")
     relevant_turns = []
@@ -108,21 +163,23 @@ def _extract_conversation_context(
     context_summary = ""
     if has_context:
         # Extract sentence(s) mentioning context
-        sentences = text.split('.')
-        context_sentences = [s for s in sentences if any(
-            word in s.lower() for word in ["previous", "earlier", "context", "history"]
-        )]
+        sentences = text.split(".")
+        context_sentences = [
+            s
+            for s in sentences
+            if any(word in s.lower() for word in ["previous", "earlier", "context", "history"])
+        ]
         if context_sentences:
             context_summary = context_sentences[0].strip()
 
     return ConversationContext(
         has_previous_context=has_context,
         relevant_turns=relevant_turns,
-        context_summary=context_summary
+        context_summary=context_summary,
     )
 
 
-def _extract_missing_capabilities(text: str) -> Optional[str]:
+def _extract_missing_capabilities(text: str) -> str | None:
     """
     Extract missing capability notes from text.
 
@@ -135,20 +192,83 @@ def _extract_missing_capabilities(text: str) -> Optional[str]:
     text_lower = text.lower()
 
     # Look for indicators of missing capabilities
-    if any(word in text_lower for word in [
-        "missing", "unavailable", "not available", "don't have", "doesn't have"
-    ]):
+    if any(
+        word in text_lower
+        for word in ["missing", "unavailable", "not available", "don't have", "doesn't have"]
+    ):
         # Find the sentence mentioning missing capabilities
-        sentences = text.split('.')
+        sentences = text.split(".")
         for sentence in sentences:
-            if any(word in sentence.lower() for word in [
-                "missing", "unavailable", "not available"
-            ]):
+            if any(
+                word in sentence.lower() for word in ["missing", "unavailable", "not available"]
+            ):
                 return sentence.strip()
 
     return None
 
 
+def _build_enriched_query(user_request: str, memory_context: dict[str, Any]) -> str:
+    """
+    Build an enriched query by appending user context when not specified.
+
+    When the user asks location-dependent questions (weather, nearby, etc.)
+    without specifying a location, this appends their known location.
+    Similarly for timezone-dependent queries.
+
+    Args:
+        user_request: The user's original request
+        memory_context: Pre-fetched memory context with profile/preferences
+
+    Returns:
+        str: Query with context appended, or original query if no enrichment needed
+
+    Example:
+        >>> query = _build_enriched_query(
+        ...     "What's the weather?",
+        ...     {"profile": {"location": "Amsterdam", "timezone": "Europe/Amsterdam"}}
+        ... )
+        >>> query
+        "What's the weather?\n\n[User Context: location=Amsterdam, timezone=Europe/Amsterdam]"
+    """
+    if not memory_context:
+        return user_request
+
+    request_lower = user_request.lower()
+    profile = memory_context.get("profile", {})
+    preferences = memory_context.get("preferences", {})
+
+    context_parts = []
+
+    # Check if location is needed and not specified
+    location_keywords = ["weather", "temperature", "forecast", "nearby", "local", "here"]
+    # Use word boundary pattern to avoid false positives like "at" in "what"
+    location_prepositions = [r"\bin\b", r"\bat\b", r"\bnear\b", r"\baround\b", r"\bfor\b"]
+    location_specified = any(re.search(p, request_lower) for p in location_prepositions)
+
+    if any(word in request_lower for word in location_keywords):
+        if not location_specified and profile.get("location"):
+            context_parts.append(f"location={profile['location']}")
+
+    # Check if timezone is needed and not specified
+    time_keywords = ["time", "schedule", "meeting", "appointment", "when", "today", "tomorrow"]
+    timezone_specified = any(word in request_lower for word in ["timezone", "tz", "utc", "gmt"])
+
+    if any(word in request_lower for word in time_keywords):
+        if not timezone_specified and profile.get("timezone"):
+            context_parts.append(f"timezone={profile['timezone']}")
+
+    # Add preferences if relevant
+    if preferences.get("temperature_unit") and "weather" in request_lower:
+        context_parts.append(f"temperature_unit={preferences['temperature_unit']}")
+
+    # Build enriched query
+    if context_parts:
+        context_str = ", ".join(context_parts)
+        return f"{user_request}\n\n[User Context: {context_str}]"
+
+    return user_request
+
+
 async def _prefetch_memory_context(user_request: str) -> dict[str, Any]:
     """
     Pre-fetch user context that might be needed for this request.
@@ -173,30 +293,65 @@ async def _prefetch_memory_context(user_request: str) -> dict[str, Any]:
     profile_keys = []
 
     # Location-related queries
-    if any(word in request_lower for word in [
-        "weather", "temperature", "forecast", "nearby", "local",
-        "directions", "distance", "map", "here"
-    ]):
+    if any(
+        word in request_lower
+        for word in [
+            "weather",
+            "temperature",
+            "forecast",
+            "nearby",
+            "local",
+            "directions",
+            "distance",
+            "map",
+            "here",
+            # Direct location questions
+            "live",
+            "where",
+            "home",
+            "reside",
+            "location",
+            "address",
+        ]
+    ):
         profile_keys.append("location")
 
     # Time-related queries
-    if any(word in request_lower for word in [
-        "time", "schedule", "meeting", "appointment", "reminder",
-        "alarm", "when", "today", "tomorrow"
-    ]):
+    if any(
+        word in request_lower
+        for word in [
+            "time",
+            "schedule",
+            "meeting",
+            "appointment",
+            "reminder",
+            "alarm",
+            "when",
+            "today",
+            "tomorrow",
+        ]
+    ):
         profile_keys.append("timezone")
 
     # Personal queries
-    if any(word in request_lower for word in [
-        "my name", "who am i", "about me"
-    ]):
+    if any(word in request_lower for word in ["my name", "who am i", "about me"]):
         profile_keys.append("name")
 
     # Always fetch preferences if they might affect response format
-    include_preferences = any(word in request_lower for word in [
-        "temperature", "weather", "convert", "unit", "format",
-        "celsius", "fahrenheit", "metric", "imperial"
-    ])
+    include_preferences = any(
+        word in request_lower
+        for word in [
+            "temperature",
+            "weather",
+            "convert",
+            "unit",
+            "format",
+            "celsius",
+            "fahrenheit",
+            "metric",
+            "imperial",
+        ]
+    )
 
     try:
         return await memory_service.prefetch_context(
@@ -215,7 +370,7 @@ async def _prefetch_memory_context(user_request: str) -> dict[str, Any]:
 async def analyze_request(
     user_request: str,
     conversation_history: list[dict],
-    conversation_id: Optional[str] = None,
+    conversation_id: str | None = None,
 ) -> StewardRecommendation:
     """
     Analyze user request with full conversation context.
@@ -223,8 +378,7 @@ async def analyze_request(
     This is the main entry point for Steward analysis. It:
     1. Calls the Steward agent with full conversation history
     2. Logs the operation with timing
-    3. Records performance benchmarks to Redis
-    4. Returns structured recommendations
+    3. Returns structured recommendations
 
     Args:
         user_request: The current user message to analyze
@@ -248,7 +402,7 @@ async def analyze_request(
             "request_preview": user_request[:100],
             "conversation_id": conversation_id,
             "history_length": len(conversation_history),
-        }
+        },
     ) as log_ctx:
         try:
             # Pre-fetch user context from memory (fast, no LLM)
@@ -267,8 +421,7 @@ async def analyze_request(
 
             # Get plain text analysis from Steward
             analysis_text = await steward.analyze(
-                user_request,
-                conversation_history=conversation_history
+                user_request, conversation_history=conversation_history
             )
 
             # Parse plain text into structured recommendation
@@ -277,6 +430,9 @@ async def analyze_request(
             context = _extract_conversation_context(analysis_text, conversation_history)
             missing = _extract_missing_capabilities(analysis_text)
 
+            # Build enriched query with auto-filled context
+            enriched_query = _build_enriched_query(user_request, memory_context)
+
             recommendation = StewardRecommendation(
                 recommended_capabilities=capabilities,
                 reasoning=analysis_text,
@@ -284,6 +440,7 @@ async def analyze_request(
                 conversation_context=context,
                 missing_capabilities=missing,
                 memory_context=memory_context,
+                enriched_query=enriched_query,
             )
 
             # Update log context with results
@@ -299,23 +456,6 @@ async def analyze_request(
                 reasoning=analysis_text[:200],  # First 200 chars
             )
 
-            # Record performance benchmark
-            if log_ctx.get("duration_seconds"):
-                benchmark = PerformanceBenchmark(
-                    operation="steward_analysis",
-                    duration_seconds=log_ctx["duration_seconds"],
-                    success=True,
-                    recommendation_count=len(recommendation.recommended_capabilities),
-                    confidence=None,  # Could add confidence scoring in future
-                    conversation_id=conversation_id,
-                    metadata={
-                        "complexity": recommendation.estimated_complexity,
-                        "has_context": recommendation.conversation_context.has_previous_context,
-                        "missing_capabilities": recommendation.missing_capabilities is not None,
-                    },
-                )
-                await get_benchmark_store().record(benchmark)
-
             return recommendation
 
         except Exception as e:

src/agents/tatlock.py

@@ -6,21 +6,26 @@ The agent embodies a witty, capable British butler personality.
 """
 
 import secrets
-from typing import AsyncGenerator, Any
+from collections.abc import AsyncGenerator
 from dataclasses import dataclass, field
+from typing import Any
 
 from pydantic_ai import Agent, RunContext
 
 from src.agents.base import AgentInterface, OutputItem
 from src.agents.tatlock_core.tools import (
     calculate,
-    get_current_datetime,
     calculate_time_offset,
+    get_current_datetime,
     time_difference,
-    search_web,
 )
-from src.core.config import config
 from src.core.logging_config import get_logger
+from src.core.tracing import (
+    SpanType,
+    add_tool_spans_from_messages,
+    end_span,
+    start_span,
+)
 
 logger = get_logger(__name__)
 
@@ -28,9 +33,10 @@ logger = get_logger(__name__)
 @dataclass
 class ToolCallTracker:
     """Tracks tool calls for reporting to reasoning output."""
+
     calls: list[str] = field(default_factory=list)
 
-    def log_call(self, message: str):
+    def log_call(self, message: str) -> None:
         """Log a tool call."""
         self.calls.append(message)
 
@@ -43,7 +49,16 @@ def generate_id() -> str:
 # System prompt defining Tatlock's personality
 TATLOCK_SYSTEM_PROMPT = """You are Tatlock, a helpful personal assistant with the demeanor of a British butler.
 
-Address users as "sir" and maintain a formal yet personable tone. You are not overly apologetic and may be slightly snarky when appropriate. If an opportunity for a pun presents itself, you cannot resist.
+## Personality
+
+Address users as "sir". Be confident, direct, and efficient - you are an unflappable English butler who gets things done. Dry wit and puns are encouraged.
+
+**CRITICAL - Do NOT:**
+- Apologize unless you genuinely made an error
+- Say "Apologies for any confusion" or "Allow me to rectify" when nothing went wrong
+- Preface successful results with caveats or apologies
+
+When presenting findings: lead with the answer, be concise, skip the preamble.
 
 You coordinate with various household staff (expert agents) to provide comprehensive assistance across:
 - Research and knowledge work
@@ -76,25 +91,62 @@ You have direct access to several permanent tools that you should USE whenever a
    - time_difference: Calculate the time between two dates
    - Use these for ANY date/time queries - never guess at dates or times
 
-3. **Web Search** (search_web): Search for current, volatile, or factual information
-   - Use this for ANY information that might be current, factual, or outside your training data
+3. **Web Search** (via Librarian): For current, volatile, or factual information
+   - Delegate to the Librarian for web searches and research
    - Examples: news, current events, recent developments, specific facts, technical documentation
-   - Always prefer searching over guessing or using potentially outdated knowledge
-   - For extensive research questions, note that this will later be delegated to the librarian
+   - Use: delegate_to_librarian(task="search the web for ...")
 
 ## Tool Usage Guidelines
 
 - **Mathematics**: ALWAYS use the calculator tool, even for simple arithmetic
 - **Dates/Times**: ALWAYS use the date/time tools, never guess or estimate
-- **Current Information**: ALWAYS search for facts, news, or volatile information
-- **Verification**: When facts are important, use search to verify rather than rely on memory alone
+- **Current Information**: Delegate web searches to the Librarian
+- **Verification**: When facts are important, delegate to Librarian for research
 - When you use a tool, explain what you're doing in a butler-appropriate manner
 - Present tool results naturally in your response
 
-Currently in Phase 1 development - expert agent delegation will be added in later phases.
+## Expert Delegation (CRITICAL)
+
+When you see "DELEGATE:" in your instructions, you MUST delegate to the appropriate agent.
+
+**PRIMARY METHOD**: Call the delegation function directly:
+- `delegate_to_librarian(task="...")` for research/wiki tasks
+- `delegate_to_biographer(task="...")` for memory tasks
+
+**FALLBACK METHOD**: If function calling fails, output EXACTLY this format:
+```
+[DELEGATE:biographer] task="Remember that user's name is TestBot"
+```
+or
+```
+[DELEGATE:librarian] task="Search for information about Docker"
+```
+
+**Rules:**
+1. When you see "DELEGATE: biographer" - delegate to biographer
+2. When you see "DELEGATE: librarian" - delegate to librarian
+3. NEVER ask for confirmation - just delegate
+4. NEVER handle delegated tasks yourself
+5. If you cannot call the function, use the [DELEGATE:...] text format EXACTLY
 """
 
 
+# Tool-phase prompt for orchestrate_tool_calls(). The butler personality prompt
+# suppresses tool calling on small local models (gemma4 reasons about the tool,
+# then answers from memory with wrong arithmetic), so the orchestration phase
+# uses a terse operator prompt; synthesize_from_results() applies the persona.
+TATLOCK_ORCHESTRATION_PROMPT = """You are the tool-execution phase of Tatlock, \
+a butler assistant. Your only job is to gather accurate results by calling the \
+provided tools.
+
+- ALWAYS use tools for the task - never answer from memory and never do mental math.
+- Mathematics: call the calculate tool, even for trivial arithmetic.
+- Dates and times: call the date/time tools, never guess.
+- When the instructions say DELEGATE to an agent, call the matching delegate_to_* tool.
+- After the tool results arrive, reply with a one-line factual summary of the results. \
+A later step writes the polished reply, so do not add personality."""
+
+
 class TatlockAgent(AgentInterface):
     """
     Tatlock - The Butler agent using PydanticAI with Ollama.
@@ -103,50 +155,47 @@ class TatlockAgent(AgentInterface):
     currently in Phase 1 (basic LLM integration without expert agents).
     """
 
-    def __init__(self):
-        """Initialize Tatlock configuration (lazy agent creation)."""
-        # Store Ollama configuration
-        self.ollama_host = str(config.OLLAMA_HOST)
-        self.model_name = config.OLLAMA_DEFAULT_MODEL
-        self._agent = None  # Lazy initialization
+    def __init__(self) -> None:
+        """Initialize Tatlock (lazy agent creation)."""
+        # Deps are a ToolCallTracker: every registered tool takes
+        # RunContext[ToolCallTracker], and run() is called with one. Saying so
+        # is what lets the tool registrations below type-check at all.
+        self._agent: Agent[ToolCallTracker, str] | None = None  # Lazy initialization
 
-    def _ensure_agent(self):
+    def _ensure_agent(self) -> None:
         """Ensure the PydanticAI agent is initialized (lazy initialization)."""
         if self._agent is not None:
             return
 
+        from src.anthropic.model_selector import get_model, get_model_info
+
+        model_info = get_model_info()
         logger.info(
             "tatlock_agent_initializing",
-            ollama_host=self.ollama_host,
-            model=self.model_name,
+            backend=model_info["backend"],
+            model=model_info["model"],
         )
 
-        # Import required classes for Ollama configuration
-        from pydantic_ai.models.openai import OpenAIChatModel
-        from pydantic_ai.providers.ollama import OllamaProvider
+        # Get best available model (Claude if available, else Ollama)
+        model = get_model()
 
-        # PydanticAI expects Ollama base URL to end with /v1
-        # Remove trailing slash from ollama_host if present
-        clean_host = self.ollama_host.rstrip('/')
-        base_url = f"{clean_host}/v1"
-
-        # Create Ollama model with provider
-        ollama_model = OpenAIChatModel(
-            model_name=self.model_name,
-            provider=OllamaProvider(base_url=base_url)
-        )
-
-        # Create PydanticAI agent with Ollama model
+        # Create PydanticAI agent
         self._agent = Agent(
-            ollama_model,
+            model,
             system_prompt=TATLOCK_SYSTEM_PROMPT,
+            deps_type=ToolCallTracker,
         )
 
         # Register tools with the agent
         self._register_tools()
 
-    def _register_tools(self):
-        """Register permanent tools with the PydanticAI agent."""
+    def _register_tools(self) -> None:
+        """Register permanent tools with the PydanticAI agent.
+
+        Called only from _ensure_agent, immediately after the agent is built, so
+        the assert documents an invariant rather than guarding a real case.
+        """
+        assert self._agent is not None, "_register_tools called before the agent exists"
 
         # Calculator tool
         @self._agent.tool
@@ -201,7 +250,9 @@ class TatlockAgent(AgentInterface):
 
         # Time difference calculator
         @self._agent.tool
-        def calculate_time_difference(ctx: RunContext[ToolCallTracker], date1_str: str, date2_str: str = "now") -> str:
+        def calculate_time_difference(
+            ctx: RunContext[ToolCallTracker], date1_str: str, date2_str: str = "now"
+        ) -> str:
             """
             Calculate the difference between two dates.
 
@@ -213,31 +264,13 @@ class TatlockAgent(AgentInterface):
                 Human-readable description of the time difference
             """
             if ctx.deps:
-                ctx.deps.log_call(f"🕐 Calculating time difference between {date1_str} and {date2_str}")
+                ctx.deps.log_call(
+                    f"🕐 Calculating time difference between {date1_str} and {date2_str}"
+                )
             return time_difference(date1_str, date2_str)
 
-        # Web search tool
-        @self._agent.tool
-        async def web_search(ctx: RunContext[ToolCallTracker], query: str, num_results: int = 5) -> str:
-            """
-            Search the web using SearXNG for current information.
-
-            Use this tool for ANY information that might be:
-            - Current or time-sensitive (news, events, recent developments)
-            - Factual and verifiable (statistics, technical specs, definitions)
-            - Outside your training data or knowledge cutoff
-
-            Args:
-                query: Search query string
-                num_results: Number of results to return (default: 5, max: 10)
-
-            Returns:
-                Formatted search results with titles, URLs, and snippets
-            """
-            # Log the search query to reasoning output
-            if ctx.deps:
-                ctx.deps.log_call(f"🔍 Searching for: '{query}'")
-            return await search_web(query, num_results)
+        # NOTE: Web search has been moved to The Librarian agent.
+        # Use delegate_to_librarian(task="search web for ...") for web search.
 
     @property
     def agent(self):
@@ -253,7 +286,7 @@ class TatlockAgent(AgentInterface):
         temperature: float = 1.0,
         max_tokens: int | None = None,
         stop: list[str] | None = None,
-        **kwargs: Any
+        **kwargs: Any,
     ) -> AsyncGenerator[OutputItem, None]:
         """
         Generate response using PydanticAI with Ollama.
@@ -287,20 +320,28 @@ class TatlockAgent(AgentInterface):
                     type="message",
                     id=f"msg_{generate_id()}",
                     role="assistant",
-                    content=[{
-                        "type": "output_text",
-                        "text": "I'm afraid I didn't receive a message, sir. How may I assist you?",
-                        "annotations": []
-                    }],
-                    status="completed"
+                    content=[
+                        {
+                            "type": "output_text",
+                            "text": "I'm afraid I didn't receive a message, sir. How may I assist you?",
+                            "annotations": [],
+                        }
+                    ],
+                    status="completed",
                 )
                 return
 
             # Build message history (all messages except the last user message)
             # PydanticAI expects history as list of ModelRequest/ModelResponse objects
-            from pydantic_ai.messages import ModelRequest, ModelResponse, UserPromptPart, TextPart
+            from pydantic_ai.messages import (
+                ModelMessage,
+                ModelRequest,
+                ModelResponse,
+                TextPart,
+                UserPromptPart,
+            )
 
-            message_history = []
+            message_history: list[ModelMessage] = []
             for i, msg in enumerate(messages[:-1]):  # All messages except the last one
                 role = msg.get("role")
                 content = msg.get("content", "")
@@ -312,7 +353,9 @@ class TatlockAgent(AgentInterface):
 
                 # Debug: Check for problematic content
                 if '"' in content or "'" in content:
-                    logger.debug(f"Message {i} ({role}) contains quotes. Content preview: {content[:100]}...")
+                    logger.debug(
+                        f"Message {i} ({role}) contains quotes. Content preview: {content[:100]}..."
+                    )
 
                 # Convert to PydanticAI message format
                 try:
@@ -321,9 +364,7 @@ class TatlockAgent(AgentInterface):
                             ModelRequest(parts=[UserPromptPart(content=content)])
                         )
                     elif role == "assistant":
-                        message_history.append(
-                            ModelResponse(parts=[TextPart(content=content)])
-                        )
+                        message_history.append(ModelResponse(parts=[TextPart(content=content)]))
                 except Exception as e:
                     logger.error(f"Error creating message history item {i}: {e}")
                     logger.error(f"Problematic content: {repr(content)}")
@@ -334,7 +375,9 @@ class TatlockAgent(AgentInterface):
             if message_history:
                 for i, hist_msg in enumerate(message_history):
                     msg_type = type(hist_msg).__name__
-                    content_preview = str(hist_msg.parts[0].content)[:50] if hist_msg.parts else "no parts"
+                    content_preview = (
+                        str(hist_msg.parts[0].content)[:50] if hist_msg.parts else "no parts"
+                    )
                     logger.info(f"  History[{i}]: {msg_type} - {content_preview}...")
 
             # Generate reasoning output if requested
@@ -344,10 +387,10 @@ class TatlockAgent(AgentInterface):
                     id=f"reasoning_{generate_id()}",
                     summary=[
                         "Analyzing your request, sir...",
-                        "Formulating response based on available knowledge..."
+                        "Formulating response based on available knowledge...",
                     ],
                     thinking="",  # PydanticAI doesn't expose internal reasoning yet
-                    status="completed"
+                    status="completed",
                 )
 
             # Create a tool call tracker for this request
@@ -364,7 +407,7 @@ class TatlockAgent(AgentInterface):
             result = await self.agent.run(
                 user_message,
                 message_history=message_history if message_history else None,
-                deps=tracker
+                deps=tracker,
             )
             final_text = result.output
 
@@ -375,7 +418,7 @@ class TatlockAgent(AgentInterface):
                     id=f"reasoning_tools_{generate_id()}",
                     summary=tracker.calls,
                     thinking="",
-                    status="completed"
+                    status="completed",
                 )
 
             # Yield the complete message
@@ -384,12 +427,8 @@ class TatlockAgent(AgentInterface):
                 type="message",
                 id=msg_id,
                 role="assistant",
-                content=[{
-                    "type": "output_text",
-                    "text": final_text,
-                    "annotations": []
-                }],
-                status="completed"
+                content=[{"type": "output_text", "text": final_text, "annotations": []}],
+                status="completed",
             )
 
         except Exception as e:
@@ -398,12 +437,14 @@ class TatlockAgent(AgentInterface):
                 type="message",
                 id=f"msg_{generate_id()}",
                 role="assistant",
-                content=[{
-                    "type": "output_text",
-                    "text": f"My apologies, sir. I encountered an error: {str(e)}",
-                    "annotations": []
-                }],
-                status="failed"
+                content=[
+                    {
+                        "type": "output_text",
+                        "text": f"My apologies, sir. I encountered an error: {str(e)}",
+                        "annotations": [],
+                    }
+                ],
+                status="failed",
             )
 
     async def supports_tools(self) -> bool:
@@ -433,7 +474,7 @@ class TatlockAgent(AgentInterface):
             steward_note: Note from Steward (prepended to request, invisible to user)
             scoped_tools: List of tool definitions from household registry
             message_history: Conversation history in PydanticAI format
-            tool_tracker: Optional tool call tracker for benchmarking
+            tool_tracker: Optional tool call tracker for analysis
 
         Returns:
             str: Tatlock's response text
@@ -447,8 +488,7 @@ class TatlockAgent(AgentInterface):
             ...     tool_tracker=tracker,
             ... )
         """
-        from pydantic_ai.models.openai import OpenAIChatModel
-        from pydantic_ai.providers.ollama import OllamaProvider
+        from src.anthropic.model_selector import get_model
 
         logger.info(
             "tatlock_run_with_scoped_tools",
@@ -459,18 +499,12 @@ class TatlockAgent(AgentInterface):
 
         # Create a fresh agent instance with scoped tools only
         # This ensures Tatlock can ONLY use tools recommended by the Steward
-        clean_host = self.ollama_host.rstrip('/')
-        base_url = f"{clean_host}/v1"
-
-        ollama_model = OpenAIChatModel(
-            model_name=self.model_name,
-            provider=OllamaProvider(base_url=base_url)
-        )
+        model = get_model()
 
         # Create agent with scoped tools
         # Tools from household registry are already PydanticAI Tool objects
         scoped_agent = Agent(
-            ollama_model,
+            model,
             system_prompt=TATLOCK_SYSTEM_PROMPT,
             tools=scoped_tools,  # Pass tools directly to Agent constructor
         )
@@ -479,9 +513,15 @@ class TatlockAgent(AgentInterface):
         enriched_message = f"{steward_note}\n\n{user_message}"
 
         # Convert message history to PydanticAI format
-        from pydantic_ai.messages import ModelRequest, ModelResponse, UserPromptPart, TextPart
+        from pydantic_ai.messages import (
+            ModelMessage,
+            ModelRequest,
+            ModelResponse,
+            TextPart,
+            UserPromptPart,
+        )
 
-        pydantic_history = []
+        pydantic_history: list[ModelMessage] = []
         for msg in message_history:
             role = msg.get("role")
             content = msg.get("content", "")
@@ -490,19 +530,19 @@ class TatlockAgent(AgentInterface):
                 continue
 
             if role == "user":
-                pydantic_history.append(
-                    ModelRequest(parts=[UserPromptPart(content=content)])
-                )
+                pydantic_history.append(ModelRequest(parts=[UserPromptPart(content=content)]))
             elif role == "assistant":
-                pydantic_history.append(
-                    ModelResponse(parts=[TextPart(content=content)])
-                )
+                pydantic_history.append(ModelResponse(parts=[TextPart(content=content)]))
 
         # Run with scoped tools and tracker
+        # Force tool_choice to make LLM actually call tools
+        from src.anthropic.model_selector import get_tool_choice_settings
+
         result = await scoped_agent.run(
             enriched_message,
             message_history=pydantic_history if pydantic_history else None,
-            deps=tool_tracker
+            deps=tool_tracker,
+            model_settings=get_tool_choice_settings(),
         )
 
         logger.info(
@@ -538,8 +578,7 @@ class TatlockAgent(AgentInterface):
         Yields:
             Text chunks from the streaming response
         """
-        from pydantic_ai.models.openai import OpenAIChatModel
-        from pydantic_ai.providers.ollama import OllamaProvider
+        from src.anthropic.model_selector import get_model
 
         logger.info(
             "tatlock_run_with_scoped_tools_stream",
@@ -549,17 +588,11 @@ class TatlockAgent(AgentInterface):
         )
 
         # Create a fresh agent instance with scoped tools only
-        clean_host = self.ollama_host.rstrip('/')
-        base_url = f"{clean_host}/v1"
-
-        ollama_model = OpenAIChatModel(
-            model_name=self.model_name,
-            provider=OllamaProvider(base_url=base_url)
-        )
+        model = get_model()
 
         # Create agent with scoped tools
         scoped_agent = Agent(
-            ollama_model,
+            model,
             system_prompt=TATLOCK_SYSTEM_PROMPT,
             tools=scoped_tools,
         )
@@ -568,9 +601,15 @@ class TatlockAgent(AgentInterface):
         enriched_message = f"{steward_note}\n\n{user_message}"
 
         # Convert message history to PydanticAI format
-        from pydantic_ai.messages import ModelRequest, ModelResponse, UserPromptPart, TextPart
+        from pydantic_ai.messages import (
+            ModelMessage,
+            ModelRequest,
+            ModelResponse,
+            TextPart,
+            UserPromptPart,
+        )
 
-        pydantic_history = []
+        pydantic_history: list[ModelMessage] = []
         for msg in message_history:
             role = msg.get("role")
             content = msg.get("content", "")
@@ -579,13 +618,9 @@ class TatlockAgent(AgentInterface):
                 continue
 
             if role == "user":
-                pydantic_history.append(
-                    ModelRequest(parts=[UserPromptPart(content=content)])
-                )
+                pydantic_history.append(ModelRequest(parts=[UserPromptPart(content=content)]))
             elif role == "assistant":
-                pydantic_history.append(
-                    ModelResponse(parts=[TextPart(content=content)])
-                )
+                pydantic_history.append(ModelResponse(parts=[TextPart(content=content)]))
 
         # Use run() instead of run_stream() to avoid Ollama 400 bug
         # with streaming + tool calls (PydanticAI issues #1292, #2256)
@@ -593,7 +628,7 @@ class TatlockAgent(AgentInterface):
         result = await scoped_agent.run(
             enriched_message,
             message_history=pydantic_history if pydantic_history else None,
-            deps=tool_tracker
+            deps=tool_tracker,
         )
 
         # Stream the final response in chunks to maintain UX
@@ -601,16 +636,288 @@ class TatlockAgent(AgentInterface):
         chunk_size = 50  # characters per chunk
 
         for i in range(0, len(response_text), chunk_size):
-            yield response_text[i:i + chunk_size]
+            yield response_text[i : i + chunk_size]
 
         logger.info("tatlock_scoped_run_complete")
 
+    async def orchestrate_tool_calls(
+        self,
+        user_message: str,
+        steward_note: str,
+        scoped_tools: list[Any],
+        message_history: list[dict],
+        tool_tracker: Any = None,
+    ) -> dict[str, Any]:
+        """
+        Phase 1: Execute tool calls and delegations, return structured results.
+
+        This is the coordination phase where Tatlock orchestrates tool calls
+        and expert delegations. The raw output is captured for Phase 2 synthesis.
+
+        Args:
+            user_message: The user's original message
+            steward_note: Note from Steward (invisible to user)
+            scoped_tools: List of tool definitions from household registry
+            message_history: Conversation history
+            tool_tracker: Optional tool call tracker for analysis
+
+        Returns:
+            dict with:
+                - tools_called: List of tool names that were called
+                - expert_results: Dict mapping expert names to their outputs
+                - tool_outputs: Dict mapping tool names to their outputs
+                - raw_output: The agent's raw text output
+        """
+        from pydantic_ai.messages import (
+            ModelMessage,
+            ModelRequest,
+            ModelResponse,
+            TextPart,
+            ToolCallPart,
+            ToolReturnPart,
+            UserPromptPart,
+        )
+
+        from src.anthropic.model_selector import get_model
+
+        logger.info(
+            "tatlock_orchestrate_tool_calls",
+            user_message_preview=user_message[:100],
+            scoped_tool_count=len(scoped_tools),
+            history_length=len(message_history),
+        )
+
+        # Start tracing span for orchestration phase
+        orchestrate_span = start_span(
+            "tatlock_orchestrate",
+            SpanType.TATLOCK,
+            metadata={
+                "scoped_tool_count": len(scoped_tools),
+                "tool_names": [getattr(t, "__name__", str(t)) for t in scoped_tools[:5]],
+            },
+        )
+
+        # Create a fresh agent instance with scoped tools only
+        model = get_model()
+
+        # Create agent with scoped tools, using the tool-phase prompt
+        scoped_agent = Agent(
+            model,
+            system_prompt=TATLOCK_ORCHESTRATION_PROMPT,
+            tools=scoped_tools,
+        )
+
+        # Prepend Steward's note to the request
+        enriched_message = f"{steward_note}\n\n{user_message}"
+
+        # Convert message history to PydanticAI format
+        pydantic_history: list[ModelMessage] = []
+        for msg in message_history:
+            role = msg.get("role")
+            content = msg.get("content", "")
+
+            if not content or not content.strip():
+                continue
+
+            if role == "user":
+                pydantic_history.append(ModelRequest(parts=[UserPromptPart(content=content)]))
+            elif role == "assistant":
+                pydantic_history.append(ModelResponse(parts=[TextPart(content=content)]))
+
+        # Run with scoped tools and tracker
+        from src.anthropic.model_selector import get_tool_choice_settings
+
+        result = await scoped_agent.run(
+            enriched_message,
+            message_history=pydantic_history if pydantic_history else None,
+            deps=tool_tracker,
+            model_settings=get_tool_choice_settings(),
+        )
+
+        # Extract tool calls and results from the agent's messages
+        tools_called = []
+        expert_results: dict[str, Any] = {}
+        tool_outputs: dict[str, Any] = {}
+
+        # Parse through new messages to find tool calls and returns
+        for msg in result.new_messages():
+            if isinstance(msg, ModelResponse):
+                for part in msg.parts:
+                    if isinstance(part, ToolCallPart):
+                        tools_called.append(part.tool_name)
+            elif isinstance(msg, ModelRequest):
+                for part in msg.parts:
+                    if isinstance(part, ToolReturnPart):
+                        tool_name = part.tool_name
+                        content = part.content
+
+                        # Categorize as expert result or tool output
+                        if tool_name.startswith("delegate_to_"):
+                            expert_name = tool_name.replace("delegate_to_", "")
+                            expert_results[expert_name] = content
+                        else:
+                            tool_outputs[tool_name] = content
+
+        logger.info(
+            "tatlock_orchestration_complete",
+            tools_called=tools_called,
+            expert_count=len(expert_results),
+            tool_output_count=len(tool_outputs),
+        )
+
+        # Add tool-level spans from result messages
+        if orchestrate_span:
+            add_tool_spans_from_messages(result.new_messages(), orchestrate_span)
+
+        # End orchestration span with results
+        end_span(
+            orchestrate_span,
+            metadata_update={
+                "tools_called": tools_called,
+                "expert_count": len(expert_results),
+                "tool_output_count": len(tool_outputs),
+            },
+            details_update={
+                "steward_note_preview": steward_note[:500] if steward_note else None,
+            },
+        )
+
+        return {
+            "tools_called": tools_called,
+            "expert_results": expert_results,
+            "tool_outputs": tool_outputs,
+            "raw_output": result.output,
+        }
+
+    async def synthesize_from_results(
+        self,
+        user_message: str,
+        orchestration_results: dict[str, Any],
+        message_history: list[dict],
+    ) -> str:
+        """
+        Phase 2: Synthesize butler-toned response from gathered results.
+
+        This is the synthesis phase where Tatlock takes the coordination
+        results and produces a properly butler-toned response.
+
+        Args:
+            user_message: The user's original message
+            orchestration_results: Results from orchestrate_tool_calls()
+            message_history: Conversation history
+
+        Returns:
+            str: Butler-toned response synthesized from all results
+        """
+        from pydantic_ai.messages import (
+            ModelMessage,
+            ModelRequest,
+            ModelResponse,
+            TextPart,
+            UserPromptPart,
+        )
+
+        from src.anthropic.model_selector import get_model
+
+        logger.info(
+            "tatlock_synthesize_from_results",
+            user_message_preview=user_message[:100],
+            expert_count=len(orchestration_results.get("expert_results", {})),
+            tool_count=len(orchestration_results.get("tool_outputs", {})),
+        )
+
+        # Start tracing span for synthesis phase
+        synthesize_span = start_span(
+            "tatlock_synthesize",
+            SpanType.TATLOCK,
+            metadata={
+                "expert_count": len(orchestration_results.get("expert_results", {})),
+                "tool_output_count": len(orchestration_results.get("tool_outputs", {})),
+            },
+        )
+
+        # Build synthesis prompt with all available information
+        synthesis_parts = []
+        synthesis_parts.append(f"The user asked: {user_message}")
+        synthesis_parts.append("")
+
+        # Add expert findings if any
+        if orchestration_results.get("expert_results"):
+            synthesis_parts.append("Expert findings:")
+            for expert, result in orchestration_results["expert_results"].items():
+                synthesis_parts.append(f"- {expert.title()}: {result}")
+            synthesis_parts.append("")
+
+        # Add tool outputs if any
+        if orchestration_results.get("tool_outputs"):
+            synthesis_parts.append("Tool results:")
+            for tool, result in orchestration_results["tool_outputs"].items():
+                synthesis_parts.append(f"- {tool}: {result}")
+            synthesis_parts.append("")
+
+        synthesis_parts.append(
+            "Synthesize a response for the user. Be direct and confident. "
+            "Lead with the answer - no apologies, no caveats, no 'mix-ups'. "
+            "Address them as 'sir', be concise, add dry wit if appropriate."
+        )
+
+        synthesis_prompt = "\n".join(synthesis_parts)
+
+        # Create synthesis agent (no tools needed)
+        model = get_model()
+
+        # Synthesis agent uses butler prompt but no tools
+        synthesis_agent = Agent(
+            model,
+            system_prompt=TATLOCK_SYSTEM_PROMPT,
+            # No tools for synthesis phase
+        )
+
+        # Convert message history to PydanticAI format
+        pydantic_history: list[ModelMessage] = []
+        for msg in message_history:
+            role = msg.get("role")
+            content = msg.get("content", "")
+
+            if not content or not content.strip():
+                continue
+
+            if role == "user":
+                pydantic_history.append(ModelRequest(parts=[UserPromptPart(content=content)]))
+            elif role == "assistant":
+                pydantic_history.append(ModelResponse(parts=[TextPart(content=content)]))
+
+        # Run synthesis
+        result = await synthesis_agent.run(
+            synthesis_prompt,
+            message_history=pydantic_history if pydantic_history else None,
+        )
+
+        logger.info(
+            "tatlock_synthesis_complete",
+            response_preview=result.output[:100],
+        )
+
+        # End synthesis span with result
+        end_span(
+            synthesize_span,
+            metadata_update={
+                "response_length": len(result.output),
+            },
+            details_update={
+                "synthesis_prompt": synthesis_prompt[:1000],
+                "response_preview": result.output[:500],
+            },
+        )
+
+        return result.output
+
     async def get_capabilities(self) -> dict:
         """Return current capabilities."""
         return {
-            "streaming": True,   # Streaming implemented
-            "reasoning": True,   # Basic reasoning summaries
-            "tools": True,       # Permanent tools: calculator, date/time, search
-            "vision": False,     # Future
-            "audio": False,      # Future
+            "streaming": True,  # Streaming implemented
+            "reasoning": True,  # Basic reasoning summaries
+            "tools": True,  # Permanent tools: calculator, date/time, search
+            "vision": False,  # Future
+            "audio": False,  # Future
         }

src/agents/tatlock_core/__init__.py

@@ -1,18 +1,19 @@
 """
 Tatlock's core tools package.
 
-Provides calculator, date/time, and web search capabilities.
+Provides calculator and date/time capabilities.
+Web search has been moved to The Librarian agent.
 Organized as a household member with toolset and capability registration.
 """
+
 from .capability import TATLOCK_CORE_CAPABILITY, get_capability
-from .toolset import get_core_tools, tatlock_core_tools
 from .tools import (
     calculate,
     calculate_time_offset,
     get_current_datetime,
-    search_web,
     time_difference,
 )
+from .toolset import get_core_tools, tatlock_core_tools
 
 __all__ = [
     # Tools
@@ -20,7 +21,6 @@ __all__ = [
     "get_current_datetime",
     "calculate_time_offset",
     "time_difference",
-    "search_web",
     # Toolset
     "tatlock_core_tools",
     "get_core_tools",

src/agents/tatlock_core/capability.py

@@ -4,17 +4,17 @@ Household capability definition for Tatlock's core tools.
 Provides the executive summary that the Steward and Butler see
 for coordinating household capabilities.
 """
-from src.core.household_registry import HouseholdCapability
 
+from src.core.household_registry import HouseholdCapability
 
 TATLOCK_CORE_CAPABILITY = HouseholdCapability(
     name="tatlock_core",
     role="Butler's Core Tools",
     category="core",
-    description="Essential tools for computation, date/time operations, and web searches",
-    domains=["computation", "datetime", "information", "research"],
+    description="Essential tools for computation and date/time operations",
+    domains=["computation", "datetime", "math", "calculator"],
     cost="low",
-    requires_network=True,  # For web search
+    requires_network=False,  # Web search moved to Librarian
 )
 
 

src/agents/tatlock_core/tools.py

@@ -6,13 +6,11 @@ These tools are always available to the butler agent:
 - Date/Time toolkit: For current time and time calculations
 - SearXNG search: For searching the web for current information
 """
+
 import math
 import re
 from datetime import datetime, timedelta
 
-import httpx
-
-from src.core.config import config
 from src.core.logging_config import get_logger
 
 logger = get_logger(__name__)
@@ -22,6 +20,7 @@ logger = get_logger(__name__)
 # Calculator Tool
 # ============================================================================
 
+
 def calculate(expression: str) -> str:
     """
     Safely evaluate mathematical expressions.
@@ -50,33 +49,29 @@ def calculate(expression: str) -> str:
         # Create safe namespace with math functions
         safe_dict = {
             # Basic math functions
-            'sqrt': math.sqrt,
-            'pow': math.pow,
-            'abs': abs,
-            'round': round,
-
+            "sqrt": math.sqrt,
+            "pow": math.pow,
+            "abs": abs,
+            "round": round,
             # Trigonometric
-            'sin': math.sin,
-            'cos': math.cos,
-            'tan': math.tan,
-            'asin': math.asin,
-            'acos': math.acos,
-            'atan': math.atan,
-
+            "sin": math.sin,
+            "cos": math.cos,
+            "tan": math.tan,
+            "asin": math.asin,
+            "acos": math.acos,
+            "atan": math.atan,
             # Logarithmic
-            'log': math.log,
-            'log10': math.log10,
-            'log2': math.log2,
-            'exp': math.exp,
-
+            "log": math.log,
+            "log10": math.log10,
+            "log2": math.log2,
+            "exp": math.exp,
             # Other
-            'ceil': math.ceil,
-            'floor': math.floor,
-            'factorial': math.factorial,
-
+            "ceil": math.ceil,
+            "floor": math.floor,
+            "factorial": math.factorial,
             # Constants
-            'pi': math.pi,
-            'e': math.e,
+            "pi": math.pi,
+            "e": math.e,
         }
 
         # Evaluate the expression safely
@@ -101,6 +96,7 @@ def calculate(expression: str) -> str:
 # Date/Time Toolkit
 # ============================================================================
 
+
 def get_current_datetime(format_str: str = "full") -> str:
     """
     Get the current date and time.
@@ -161,7 +157,7 @@ def calculate_time_offset(offset_description: str) -> str:
 
         # Parse the offset description
         # Pattern: "N unit(s) ago/from now"
-        pattern = r'(\d+)\s+(second|minute|hour|day|week|month|year)s?\s+(ago|from\s+now)'
+        pattern = r"(\d+)\s+(second|minute|hour|day|week|month|year)s?\s+(ago|from\s+now)"
         match = re.match(pattern, offset_description.lower().strip())
 
         if not match:
@@ -256,96 +252,5 @@ def time_difference(date1_str: str, date2_str: str = "now") -> str:
         return f"Error calculating time difference: {str(e)}"
 
 
-# ============================================================================
-# SearXNG Search Tool
-# ============================================================================
-
-async def search_web(query: str, num_results: int = 5) -> str:
-    """
-    Search the web using SearXNG.
-
-    Args:
-        query: Search query string
-        num_results: Number of results to return (default: 5, max: 10)
-
-    Returns:
-        Formatted search results as a string with titles, URLs, and snippets
-
-    Examples:
-        search_web("Python async programming") -> "1. Title: ...\n   URL: ...\n   ..."
-    """
-    try:
-        # Limit results
-        num_results = min(num_results, 10)
-
-        # Get SearXNG host with fallback logic
-        searxng_host = str(config.SEARXNG_HOST)
-
-        # Try production host first, fall back to localhost in development
-        hosts_to_try = [searxng_host]
-        if config.ENVIRONMENT.value == "development" and "localhost" not in searxng_host:
-            # Add localhost fallback for development
-            hosts_to_try.append("http://localhost:8087")
-
-        last_error = None
-
-        for host in hosts_to_try:
-            try:
-                logger.debug("searxng_search_attempt", host=host, query=query)
-
-                async with httpx.AsyncClient(timeout=config.SEARXNG_TIMEOUT) as client:
-                    response = await client.get(
-                        f"{host}/search",
-                        params={
-                            "q": query,
-                            "format": "json",
-                            "pageno": 1,
-                        }
-                    )
-
-                    if response.status_code == 200:
-                        data = response.json()
-                        results = data.get("results", [])
-
-                        if not results:
-                            return f"No results found for '{query}'"
-
-                        # Format results
-                        formatted_results = []
-                        for i, result in enumerate(results[:num_results], 1):
-                            title = result.get("title", "No title")
-                            url = result.get("url", "")
-                            content = result.get("content", "No description available")
-
-                            formatted_results.append(
-                                f"{i}. {title}\n"
-                                f"   URL: {url}\n"
-                                f"   {content}\n"
-                            )
-
-                        logger.info(
-                            "searxng_search_success",
-                            host=host,
-                            query=query,
-                            result_count=len(results),
-                        )
-                        return "\n".join(formatted_results)
-                    else:
-                        last_error = f"SearXNG returned status {response.status_code}"
-
-            except httpx.ConnectError:
-                last_error = f"Cannot connect to SearXNG at {host}"
-                logger.warning("searxng_connection_failed", host=host)
-                continue
-            except Exception as e:
-                last_error = str(e)
-                logger.warning("searxng_error", host=host, error=str(e))
-                continue
-
-        # All hosts failed
-        logger.error("searxng_all_hosts_failed", error=last_error)
-        return f"Error searching: {last_error}. Please check that SearXNG is running."
-
-    except Exception as e:
-        logger.error("searxng_unexpected_error", error=str(e), exc_info=True)
-        return f"Error searching: {str(e)}"
+# NOTE: Web search has been moved to The Librarian agent.
+# Use delegate_to_librarian(task="search web for ...") for web search.

src/agents/tatlock_core/toolset.py

@@ -4,11 +4,11 @@ PydanticAI toolset for Tatlock's core tools.
 Converts the core tool functions into PydanticAI tool definitions
 that can be registered with agents and the household registry.
 """
+
 from pydantic_ai.tools import Tool
 
 from . import tools
 
-
 # Create tool definitions for PydanticAI
 calculator_tool = Tool(
     function=tools.calculate,
@@ -55,17 +55,8 @@ time_difference_tool = Tool(
     ),
 )
 
-web_search_tool = Tool(
-    function=tools.search_web,
-    name="search_web",
-    description=(
-        "Search the web using SearXNG for current information. "
-        "Use this to find recent events, current data, or verify facts. "
-        "Returns formatted results with titles, URLs, and snippets. "
-        "Useful for information that may have changed since training data."
-    ),
-    takes_ctx=False,
-)
+# NOTE: Web search has been moved to The Librarian agent.
+# Use delegate_to_librarian(task="search web for ...") for web search.
 
 
 # Combined toolset of all core tools
@@ -74,7 +65,6 @@ tatlock_core_tools = [
     current_datetime_tool,
     time_offset_tool,
     time_difference_tool,
-    web_search_tool,
 ]
 
 

src/agents/tools.py

@@ -4,26 +4,20 @@ Tatlock's permanent tools.
 These tools are always available to the butler agent:
 - Calculator: For all mathematical operations
 - Date/Time toolkit: For current time and time calculations
-- SearXNG search: For searching the web for current information
+
+Note: Web search has been moved to The Librarian agent.
+See src/agents/librarian/tools.py for search_web functionality.
 """
 
-import logging
 import math
 import re
 from datetime import datetime, timedelta
-from typing import Any
-
-import httpx
-
-from src.core.config import config
-
-logger = logging.getLogger(__name__)
-
 
 # ============================================================================
 # Calculator Tool
 # ============================================================================
 
+
 def calculate(expression: str) -> str:
     """
     Safely evaluate mathematical expressions.
@@ -52,33 +46,29 @@ def calculate(expression: str) -> str:
         # Create safe namespace with math functions
         safe_dict = {
             # Basic math functions
-            'sqrt': math.sqrt,
-            'pow': math.pow,
-            'abs': abs,
-            'round': round,
-
+            "sqrt": math.sqrt,
+            "pow": math.pow,
+            "abs": abs,
+            "round": round,
             # Trigonometric
-            'sin': math.sin,
-            'cos': math.cos,
-            'tan': math.tan,
-            'asin': math.asin,
-            'acos': math.acos,
-            'atan': math.atan,
-
+            "sin": math.sin,
+            "cos": math.cos,
+            "tan": math.tan,
+            "asin": math.asin,
+            "acos": math.acos,
+            "atan": math.atan,
             # Logarithmic
-            'log': math.log,
-            'log10': math.log10,
-            'log2': math.log2,
-            'exp': math.exp,
-
+            "log": math.log,
+            "log10": math.log10,
+            "log2": math.log2,
+            "exp": math.exp,
             # Other
-            'ceil': math.ceil,
-            'floor': math.floor,
-            'factorial': math.factorial,
-
+            "ceil": math.ceil,
+            "floor": math.floor,
+            "factorial": math.factorial,
             # Constants
-            'pi': math.pi,
-            'e': math.e,
+            "pi": math.pi,
+            "e": math.e,
         }
 
         # Evaluate the expression safely
@@ -103,6 +93,7 @@ def calculate(expression: str) -> str:
 # Date/Time Toolkit
 # ============================================================================
 
+
 def get_current_datetime(format_str: str = "full") -> str:
     """
     Get the current date and time.
@@ -163,7 +154,7 @@ def calculate_time_offset(offset_description: str) -> str:
 
         # Parse the offset description
         # Pattern: "N unit(s) ago/from now"
-        pattern = r'(\d+)\s+(second|minute|hour|day|week|month|year)s?\s+(ago|from\s+now)'
+        pattern = r"(\d+)\s+(second|minute|hour|day|week|month|year)s?\s+(ago|from\s+now)"
         match = re.match(pattern, offset_description.lower().strip())
 
         if not match:
@@ -256,91 +247,3 @@ def time_difference(date1_str: str, date2_str: str = "now") -> str:
 
     except Exception as e:
         return f"Error calculating time difference: {str(e)}"
-
-
-# ============================================================================
-# SearXNG Search Tool
-# ============================================================================
-
-async def search_web(query: str, num_results: int = 5) -> str:
-    """
-    Search the web using SearXNG.
-
-    Args:
-        query: Search query string
-        num_results: Number of results to return (default: 5, max: 10)
-
-    Returns:
-        Formatted search results as a string with titles, URLs, and snippets
-
-    Examples:
-        search_web("Python async programming") -> "1. Title: ...\n   URL: ...\n   ..."
-    """
-    try:
-        # Limit results
-        num_results = min(num_results, 10)
-
-        # Get SearXNG host with fallback logic
-        searxng_host = str(config.SEARXNG_HOST)
-
-        # Try production host first, fall back to localhost in development
-        hosts_to_try = [searxng_host]
-        if config.ENVIRONMENT.value == "development" and "localhost" not in searxng_host:
-            # Add localhost fallback for development
-            hosts_to_try.append("http://localhost:8087")
-
-        last_error = None
-
-        for host in hosts_to_try:
-            try:
-                logger.info(f"Attempting SearXNG search at {host}")
-
-                async with httpx.AsyncClient(timeout=config.SEARXNG_TIMEOUT) as client:
-                    response = await client.get(
-                        f"{host}/search",
-                        params={
-                            "q": query,
-                            "format": "json",
-                            "pageno": 1,
-                        }
-                    )
-
-                    if response.status_code == 200:
-                        data = response.json()
-                        results = data.get("results", [])
-
-                        if not results:
-                            return f"No results found for '{query}'"
-
-                        # Format results
-                        formatted_results = []
-                        for i, result in enumerate(results[:num_results], 1):
-                            title = result.get("title", "No title")
-                            url = result.get("url", "")
-                            content = result.get("content", "No description available")
-
-                            formatted_results.append(
-                                f"{i}. {title}\n"
-                                f"   URL: {url}\n"
-                                f"   {content}\n"
-                            )
-
-                        return "\n".join(formatted_results)
-                    else:
-                        last_error = f"SearXNG returned status {response.status_code}"
-
-            except httpx.ConnectError:
-                last_error = f"Cannot connect to SearXNG at {host}"
-                logger.warning(f"SearXNG connection failed at {host}, trying next host if available")
-                continue
-            except Exception as e:
-                last_error = str(e)
-                logger.warning(f"SearXNG error at {host}: {e}")
-                continue
-
-        # All hosts failed
-        return f"Error searching: {last_error}. Please check that SearXNG is running."
-
-    except Exception as e:
-        logger.error(f"Unexpected error in search_web: {e}", exc_info=True)
-        return f"Error searching: {str(e)}"

src/anthropic/__init__.py

@@ -0,0 +1,26 @@
+"""
+Anthropic/Claude integration module.
+
+Provides model selection with Ollama as primary backend and Claude
+as the cloud fallback.
+"""
+
+from src.anthropic.model_selector import (
+    check_claude_health,
+    check_ollama_health,
+    get_model,
+    get_tool_choice_settings,
+    is_claude_available,
+    is_ollama_available,
+    resolve_backend,
+)
+
+__all__ = [
+    "check_claude_health",
+    "check_ollama_health",
+    "get_model",
+    "get_tool_choice_settings",
+    "is_claude_available",
+    "is_ollama_available",
+    "resolve_backend",
+]

src/anthropic/model_selector.py

@@ -0,0 +1,293 @@
+"""
+Model selector for Ollama/Claude backend switching.
+
+Provides automatic model selection with Ollama as the primary local backend
+and Claude as the cloud fallback. Claude is used when PREFER_CLOUD_BACKEND
+is enabled, or automatically when Ollama is unavailable at startup.
+
+The Anthropic SDK is imported lazily so a missing or broken `anthropic`
+package degrades to Ollama-only operation instead of crashing the app.
+"""
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING
+
+import httpx
+
+from src.core.config import config
+from src.core.logging_config import get_logger
+
+if TYPE_CHECKING:
+    from pydantic_ai.models.anthropic import AnthropicModel
+    from pydantic_ai.models.openai import OpenAIChatModel
+    from pydantic_ai.settings import ModelSettings
+
+logger = get_logger(__name__)
+
+# Cached health check results (set once at startup)
+_claude_available: bool | None = None
+_ollama_available: bool | None = None
+
+
+async def check_ollama_health() -> bool:
+    """
+    Check if the Ollama server is reachable and has the configured model.
+
+    This should be called once at application startup.
+    The result is cached in `_ollama_available`.
+
+    Returns:
+        True if Ollama is reachable and OLLAMA_DEFAULT_MODEL is pulled.
+    """
+    global _ollama_available
+
+    host = str(config.OLLAMA_HOST).rstrip("/")
+    model = config.OLLAMA_DEFAULT_MODEL
+
+    try:
+        async with httpx.AsyncClient(timeout=5.0) as client:
+            response = await client.get(f"{host}/api/tags")
+            response.raise_for_status()
+            names = [m.get("name", "") for m in response.json().get("models", [])]
+
+        if model in names or f"{model}:latest" in names:
+            _ollama_available = True
+            logger.info(
+                "ollama_health_check_passed",
+                host=host,
+                model=model,
+            )
+            return True
+
+        _ollama_available = False
+        logger.warning(
+            "ollama_health_check_failed",
+            reason="model_not_pulled",
+            host=host,
+            model=model,
+            hint=f"run `ollama pull {model}`",
+        )
+        return False
+
+    except Exception as e:
+        _ollama_available = False
+        logger.warning(
+            "ollama_health_check_failed",
+            reason="server_unreachable",
+            host=host,
+            error=str(e),
+        )
+        return False
+
+
+async def check_claude_health() -> bool:
+    """
+    Check if Claude API is reachable and working.
+
+    This should be called once at application startup.
+    The result is cached in `_claude_available`.
+
+    Returns:
+        True if Claude API is accessible, False otherwise.
+    """
+    global _claude_available
+
+    # No API key configured - Claude not available
+    if not config.ANTHROPIC_API_KEY:
+        logger.info(
+            "claude_health_check_skipped",
+            reason="no_api_key",
+        )
+        _claude_available = False
+        return False
+
+    try:
+        from anthropic import AsyncAnthropic
+
+        client = AsyncAnthropic(api_key=config.ANTHROPIC_API_KEY)
+
+        # Minimal API call to verify connectivity
+        # Using a tiny max_tokens to minimize cost
+        await client.messages.create(
+            model=config.ANTHROPIC_MODEL,
+            max_tokens=1,
+            messages=[{"role": "user", "content": "hi"}],
+        )
+
+        _claude_available = True
+        logger.info(
+            "claude_health_check_passed",
+            model=config.ANTHROPIC_MODEL,
+        )
+        return True
+
+    except Exception as e:
+        _claude_available = False
+        logger.warning(
+            "claude_health_check_failed",
+            error=str(e),
+            model=config.ANTHROPIC_MODEL,
+        )
+        return False
+
+
+def is_claude_available() -> bool:
+    """
+    Check if Claude is available (from cached health check result).
+
+    Returns:
+        True if Claude API was reachable at startup, False otherwise.
+
+    Note:
+        Returns False if health check hasn't been run yet.
+        Call `check_claude_health()` at startup first.
+    """
+    return _claude_available is True
+
+
+def is_ollama_available() -> bool:
+    """
+    Check if Ollama is available (from cached health check result).
+
+    Returns:
+        False only if the startup health check confirmed Ollama is down.
+        Unknown (check not run yet) counts as available so that contexts
+        without lifespan events keep the local-first behavior.
+    """
+    return _ollama_available is not False
+
+
+def resolve_backend(prefer_cloud: bool | None = None) -> str:
+    """
+    Resolve which backend should serve requests.
+
+    Ollama is the primary backend. Claude is used when explicitly
+    preferred via PREFER_CLOUD_BACKEND, or as automatic fallback
+    when the startup health check found Ollama down.
+
+    Args:
+        prefer_cloud: Override config.PREFER_CLOUD_BACKEND for this call.
+
+    Returns:
+        "claude" or "ollama".
+    """
+    use_cloud = prefer_cloud if prefer_cloud is not None else config.PREFER_CLOUD_BACKEND
+
+    if use_cloud and is_claude_available():
+        return "claude"
+
+    if not is_ollama_available() and is_claude_available():
+        logger.warning(
+            "backend_fallback_to_claude",
+            reason="ollama_unavailable",
+        )
+        return "claude"
+
+    return "ollama"
+
+
+def get_model(prefer_cloud: bool | None = None) -> AnthropicModel | OpenAIChatModel:
+    """
+    Get the best available model.
+
+    Returns Ollama unless Claude is preferred (or Ollama is down).
+
+    Args:
+        prefer_cloud: Override config.PREFER_CLOUD_BACKEND for this call.
+                     If None, uses the config value.
+
+    Returns:
+        PydanticAI model instance (OpenAIChatModel or AnthropicModel).
+
+    Example:
+        >>> model = get_model()
+        >>> agent = Agent(model, system_prompt="...")
+    """
+    if resolve_backend(prefer_cloud) == "claude":
+        try:
+            from pydantic_ai.models.anthropic import AnthropicModel
+            from pydantic_ai.providers.anthropic import AnthropicProvider
+
+            logger.debug(
+                "model_selected",
+                backend="claude",
+                model=config.ANTHROPIC_MODEL,
+            )
+            return AnthropicModel(
+                model_name=config.ANTHROPIC_MODEL,
+                provider=AnthropicProvider(api_key=config.ANTHROPIC_API_KEY),
+            )
+        except ImportError as e:
+            logger.error(
+                "claude_backend_import_failed",
+                error=str(e),
+                hint="anthropic package missing or incompatible; using Ollama",
+            )
+
+    from pydantic_ai.models.openai import OpenAIChatModel
+
+    from src.ollama.provider import get_ollama_provider
+
+    logger.debug(
+        "model_selected",
+        backend="ollama",
+        model=config.OLLAMA_DEFAULT_MODEL,
+    )
+    return OpenAIChatModel(
+        model_name=config.OLLAMA_DEFAULT_MODEL,
+        provider=get_ollama_provider(),
+    )
+
+
+def get_tool_choice_settings() -> ModelSettings:
+    """
+    Get model_settings for forcing tool calls on the first request.
+
+    For Claude: PydanticAI handles tool_choice natively, so no extra_body needed.
+    For Ollama: Pass tool_choice="required" via extra_body to force tool calling.
+    """
+    from pydantic_ai.settings import ModelSettings
+
+    if resolve_backend() == "claude":
+        # PydanticAI's Anthropic model handles tool_choice internally
+        return ModelSettings()
+    else:
+        # Ollama needs explicit tool_choice via extra_body
+        return ModelSettings(extra_body={"tool_choice": "required"})
+
+
+def get_sampling_settings(temperature: float) -> ModelSettings:
+    """
+    Get model_settings with a sampling temperature where the backend allows it.
+
+    Ollama accepts a temperature; Claude Sonnet 5+ rejects sampling
+    parameters, so the Claude backend gets empty settings.
+    """
+    from pydantic_ai.settings import ModelSettings
+
+    if resolve_backend() == "claude":
+        return ModelSettings()
+    return ModelSettings(temperature=temperature)
+
+
+def get_model_info() -> dict:
+    """
+    Get information about the current model configuration.
+
+    Useful for health checks and debugging.
+
+    Returns:
+        Dict with backend, model name, and availability info.
+    """
+    backend = resolve_backend()
+
+    return {
+        "backend": backend,
+        "model": config.ANTHROPIC_MODEL if backend == "claude" else config.OLLAMA_DEFAULT_MODEL,
+        "claude_available": is_claude_available(),
+        "claude_configured": bool(config.ANTHROPIC_API_KEY),
+        "ollama_available": is_ollama_available(),
+        "ollama_model": config.OLLAMA_DEFAULT_MODEL,
+        "prefer_cloud": config.PREFER_CLOUD_BACKEND,
+    }

src/chat/router.py

@@ -2,12 +2,13 @@
 Chat completion router.
 OpenAI-compatible /v1/chat/completions endpoint.
 """
+
 import json
 import logging
-from typing import AsyncGenerator
+from collections.abc import AsyncGenerator
 
 from fastapi import APIRouter
-from sse_starlette.sse import EventSourceResponse
+from starlette.responses import StreamingResponse
 
 from src.chat import service
 from src.chat.schemas import (
@@ -22,47 +23,51 @@ router = APIRouter(prefix="/chat", tags=["chat"])
 
 async def _stream_response(
     request: ChatCompletionRequest,
-) -> AsyncGenerator[dict, None]:
+) -> AsyncGenerator[str, None]:
     """
     Generate SSE stream for chat completion.
 
-    EventSourceResponse adds "data: " prefix automatically.
-    We just yield the dict/string content.
+    Yields raw SSE-formatted strings matching OpenAI's format exactly:
+    data: {json}\n\n
     """
     try:
         async for chunk in service.create_chat_completion_stream(request):
-            # Yield dict - EventSourceResponse will format as SSE
-            yield {"data": chunk.model_dump_json()}
+            yield f"data: {chunk.model_dump_json(exclude_unset=True)}\n\n"
 
-        # Send [DONE] message
-        yield {"data": "[DONE]"}
+        yield "data: [DONE]\n\n"
 
     except Exception as e:
         logger.error(f"Error in streaming response: {e}")
-        error_data = {"error": {"message": str(e), "type": "internal_error"}}
-        yield {"data": json.dumps(error_data)}
+        error_data = json.dumps({"error": {"message": str(e), "type": "internal_error"}})
+        yield f"data: {error_data}\n\n"
 
 
 @router.post("/completions", response_model=ChatCompletionResponse)
 async def create_chat_completion(
     request: ChatCompletionRequest,
-) -> ChatCompletionResponse | EventSourceResponse:
+) -> ChatCompletionResponse | StreamingResponse:
     """
     Create chat completion (OpenAI-compatible).
-    
+
     Supports both regular and streaming responses.
-    Currently returns mock lorem ipsum responses.
-    
+
     Args:
         request: Chat completion request
-        
+
     Returns:
         Chat completion response or SSE stream
     """
     logger.info(f"Chat completion request for model: {request.model}")
-    
+
     if request.stream:
         logger.info("Streaming response requested")
-        return EventSourceResponse(_stream_response(request))
-    
+        return StreamingResponse(
+            _stream_response(request),
+            media_type="text/event-stream",
+            headers={
+                "Cache-Control": "no-store",
+                "X-Accel-Buffering": "no",
+            },
+        )
+
     return await service.create_chat_completion(request)

src/chat/schemas.py

@@ -2,6 +2,7 @@
 OpenAI-compatible chat completion schemas.
 Following OpenAI API specification for compatibility.
 """
+
 from typing import Literal
 
 from pydantic import Field
@@ -11,6 +12,7 @@ from src.core.models import CustomBaseModel
 
 class ChatMessage(CustomBaseModel):
     """OpenAI-compatible chat message."""
+
     role: Literal["system", "user", "assistant"]
     content: str
     name: str | None = None
@@ -18,6 +20,7 @@ class ChatMessage(CustomBaseModel):
 
 class ChatCompletionRequest(CustomBaseModel):
     """OpenAI-compatible chat completion request."""
+
     model: str = Field(..., description="Model to use for completion")
     messages: list[ChatMessage] = Field(..., description="List of messages")
     temperature: float | None = Field(default=0.7, ge=0.0, le=2.0)
@@ -29,6 +32,7 @@ class ChatCompletionRequest(CustomBaseModel):
 
 class ChatCompletionChoice(CustomBaseModel):
     """Choice in chat completion response."""
+
     index: int
     message: ChatMessage
     finish_reason: str | None
@@ -36,6 +40,7 @@ class ChatCompletionChoice(CustomBaseModel):
 
 class ChatCompletionUsage(CustomBaseModel):
     """Token usage information."""
+
     prompt_tokens: int
     completion_tokens: int
     total_tokens: int
@@ -43,6 +48,7 @@ class ChatCompletionUsage(CustomBaseModel):
 
 class ChatCompletionResponse(CustomBaseModel):
     """OpenAI-compatible chat completion response."""
+
     id: str
     object: str = "chat.completion"
     created: int
@@ -53,12 +59,15 @@ class ChatCompletionResponse(CustomBaseModel):
 
 class ChatCompletionChunkDelta(CustomBaseModel):
     """Delta in streaming chunk."""
+
     role: str | None = None
     content: str | None = None
+    reasoning_content: str | None = None  # For thinking/reasoning (DeepSeek R1 format)
 
 
 class ChatCompletionChunkChoice(CustomBaseModel):
     """Choice in streaming chunk."""
+
     index: int
     delta: ChatCompletionChunkDelta
     finish_reason: str | None = None
@@ -66,6 +75,7 @@ class ChatCompletionChunkChoice(CustomBaseModel):
 
 class ChatCompletionChunk(CustomBaseModel):
     """OpenAI-compatible streaming chunk."""
+
     id: str
     object: str = "chat.completion.chunk"
     created: int

src/chat/service.py

@@ -4,17 +4,17 @@ Chat completion service.
 Wrapper around Responses API that converts to Chat Completions format.
 Embeds reasoning in <think> tags for Open WebUI compatibility.
 """
-import asyncio
+
 import time
 import uuid
-from typing import AsyncGenerator
+from collections.abc import AsyncGenerator
 
 from src.chat import constants
 from src.chat.schemas import (
+    ChatCompletionChoice,
     ChatCompletionChunk,
     ChatCompletionChunkChoice,
     ChatCompletionChunkDelta,
-    ChatCompletionChoice,
     ChatCompletionRequest,
     ChatCompletionResponse,
     ChatCompletionUsage,
@@ -43,10 +43,7 @@ async def create_chat_completion(
     created_at = int(time.time())
 
     # Convert Chat request to Responses request
-    input_messages = [
-        {"role": msg.role, "content": msg.content}
-        for msg in request.messages
-    ]
+    input_messages = [{"role": msg.role, "content": msg.content} for msg in request.messages]
 
     response_request = ResponseRequest(
         model=request.model,
@@ -54,7 +51,9 @@ async def create_chat_completion(
         reasoning={"effort": "medium", "summary": "auto"},  # Enable reasoning
         temperature=request.temperature or 1.0,
         max_output_tokens=request.max_tokens,
-        stop=request.stop if isinstance(request.stop, list) else ([request.stop] if request.stop else None),
+        stop=request.stop
+        if isinstance(request.stop, list)
+        else ([request.stop] if request.stop else None),
     )
 
     # Call Responses API (will use Steward for Tatlock)
@@ -118,16 +117,13 @@ async def create_chat_completion_stream(
     Yields:
         Chat completion chunks with reasoning as <think> tags
     """
-    from src.responses.streaming import StreamingCoordinator, StreamEventType
+    from src.responses.streaming import StreamEventType, StreamingCoordinator
 
     completion_id = f"chatcmpl-{uuid.uuid4().hex[:24]}"
     created_at = int(time.time())
 
     # Convert Chat request to Responses request
-    input_messages = [
-        {"role": msg.role, "content": msg.content}
-        for msg in request.messages
-    ]
+    input_messages = [{"role": msg.role, "content": msg.content} for msg in request.messages]
 
     response_request = ResponseRequest(
         model=request.model,
@@ -135,7 +131,9 @@ async def create_chat_completion_stream(
         reasoning={"effort": "medium", "summary": "auto"},
         temperature=request.temperature or 1.0,
         max_output_tokens=request.max_tokens,
-        stop=request.stop if isinstance(request.stop, list) else ([request.stop] if request.stop else None),
+        stop=request.stop
+        if isinstance(request.stop, list)
+        else ([request.stop] if request.stop else None),
         stream=True,
     )
 
@@ -163,7 +161,6 @@ async def create_chat_completion_stream(
 
     # Stream from Responses API
     coordinator = StreamingCoordinator()
-    in_reasoning = False
 
     if use_steward:
         stream_generator = coordinator.stream_response_with_steward(response_request)
@@ -172,24 +169,8 @@ async def create_chat_completion_stream(
 
     async for event in stream_generator:
         if event.event == StreamEventType.REASONING_SUMMARY_DELTA:
-            # Start <think> block if needed
-            if not in_reasoning:
-                yield ChatCompletionChunk(
-                    id=completion_id,
-                    object=constants.CHAT_COMPLETION_CHUNK_OBJECT,
-                    created=created_at,
-                    model=request.model,
-                    choices=[
-                        ChatCompletionChunkChoice(
-                            index=0,
-                            delta=ChatCompletionChunkDelta(content="<think>\n"),
-                            finish_reason=None,
-                        )
-                    ],
-                )
-                in_reasoning = True
-
-            # Stream reasoning delta
+            # Stream reasoning via reasoning_content field (DeepSeek R1 format)
+            # Open WebUI renders this as collapsible thinking block
             yield ChatCompletionChunk(
                 id=completion_id,
                 object=constants.CHAT_COMPLETION_CHUNK_OBJECT,
@@ -198,29 +179,15 @@ async def create_chat_completion_stream(
                 choices=[
                     ChatCompletionChunkChoice(
                         index=0,
-                        delta=ChatCompletionChunkDelta(content=event.delta),
+                        delta=ChatCompletionChunkDelta(reasoning_content=event.delta),
                         finish_reason=None,
                     )
                 ],
             )
 
         elif event.event == StreamEventType.REASONING_SUMMARY_DONE:
-            # Close <think> block
-            if in_reasoning:
-                yield ChatCompletionChunk(
-                    id=completion_id,
-                    object=constants.CHAT_COMPLETION_CHUNK_OBJECT,
-                    created=created_at,
-                    model=request.model,
-                    choices=[
-                        ChatCompletionChunkChoice(
-                            index=0,
-                            delta=ChatCompletionChunkDelta(content="</think>\n\n"),
-                            finish_reason=None,
-                        )
-                    ],
-                )
-                in_reasoning = False
+            # Signal end of reasoning block (no content needed)
+            pass  # nothing downstream reads this; the event just ends the block
 
         elif event.event == StreamEventType.OUTPUT_TEXT_DELTA:
             # Stream message content

src/core/benchmarks.py

@@ -1,337 +0,0 @@
-"""
-Performance benchmark storage using Redis.
-
-Tracks operation timing, tool usage, and recommendation accuracy across sessions.
-Provides time-series data for performance analysis and optimization.
-"""
-import json
-from datetime import datetime, timezone
-from typing import Any, Literal, Optional
-
-import redis.asyncio as redis
-from pydantic import BaseModel, Field
-
-from .config import config
-from .logging_config import get_logger
-
-logger = get_logger(__name__)
-
-
-class PerformanceBenchmark(BaseModel):
-    """
-    Performance benchmark record.
-
-    Stores timing and metadata for operations like Steward analysis,
-    tool calls, and agent execution.
-    """
-    timestamp: datetime = Field(default_factory=lambda: datetime.now(timezone.utc))
-    operation: str  # "steward_analysis", "tool_call", "tatlock_execution"
-    duration_seconds: float
-    success: bool
-
-    # Steward-specific fields
-    recommendation_count: Optional[int] = None
-    confidence: Optional[float] = None
-
-    # Tool-specific fields
-    tool_name: Optional[str] = None
-    was_recommended: Optional[bool] = None
-    was_actually_used: Optional[bool] = None
-
-    # Context
-    conversation_id: Optional[str] = None
-    metadata: dict[str, Any] = Field(default_factory=dict)
-
-    def to_redis_dict(self) -> dict[str, Any]:
-        """Convert to dict suitable for Redis storage."""
-        data = self.model_dump()
-        data["timestamp"] = self.timestamp.isoformat()
-        data["metadata"] = json.dumps(self.metadata)
-        return data
-
-    @classmethod
-    def from_redis_dict(cls, data: dict[str, Any]) -> "PerformanceBenchmark":
-        """Reconstruct from Redis dict."""
-        data["timestamp"] = datetime.fromisoformat(data["timestamp"])
-        data["metadata"] = json.loads(data.get("metadata", "{}"))
-        return cls(**data)
-
-
-class BenchmarkStore:
-    """
-    Redis-backed benchmark storage with automatic expiry.
-
-    Stores performance metrics in time-series format with 30-day retention.
-    Provides querying capabilities for analysis and reporting.
-    """
-
-    def __init__(self, redis_client: Optional[redis.Redis] = None):
-        """
-        Initialize benchmark store.
-
-        Args:
-            redis_client: Optional Redis client. If None, creates from config.
-        """
-        self._client = redis_client
-        self._ttl_days = 30  # 30-day retention
-
-    async def _get_client(self) -> redis.Redis:
-        """Get or create Redis client."""
-        if self._client is None:
-            self._client = redis.from_url(
-                config.redis_url,
-                encoding="utf-8",
-                decode_responses=True,
-                socket_timeout=config.REDIS_TIMEOUT,
-                socket_connect_timeout=config.REDIS_TIMEOUT,
-            )
-        return self._client
-
-    async def record(self, benchmark: PerformanceBenchmark) -> None:
-        """
-        Record a performance benchmark.
-
-        Args:
-            benchmark: Performance benchmark to record
-
-        Example:
-            >>> await store.record(PerformanceBenchmark(
-            ...     operation="steward_analysis",
-            ...     duration_seconds=1.23,
-            ...     success=True,
-            ...     recommendation_count=3,
-            ... ))
-        """
-        if not config.ENABLE_BENCHMARKS:
-            return
-
-        try:
-            client = await self._get_client()
-
-            # Generate key: benchmark:{operation}:{timestamp_ms}
-            timestamp_ms = int(benchmark.timestamp.timestamp() * 1000)
-            key = f"benchmark:{benchmark.operation}:{timestamp_ms}"
-
-            # Store as hash
-            await client.hset(key, mapping=benchmark.to_redis_dict())
-
-            # Set expiry
-            await client.expire(key, self._ttl_days * 24 * 60 * 60)
-
-            # Add to sorted set for time-based queries
-            index_key = f"benchmark_index:{benchmark.operation}"
-            await client.zadd(index_key, {key: timestamp_ms})
-            await client.expire(index_key, self._ttl_days * 24 * 60 * 60)
-
-            logger.debug(
-                "benchmark_recorded",
-                operation=benchmark.operation,
-                duration=benchmark.duration_seconds,
-                success=benchmark.success,
-            )
-
-        except Exception as e:
-            logger.warning(
-                "benchmark_recording_failed",
-                error=str(e),
-                operation=benchmark.operation,
-            )
-            # Don't fail the request if benchmarking fails
-
-    async def query(
-        self,
-        operation: str,
-        start_time: Optional[datetime] = None,
-        end_time: Optional[datetime] = None,
-        limit: int = 100,
-    ) -> list[PerformanceBenchmark]:
-        """
-        Query benchmarks by operation and time range.
-
-        Args:
-            operation: Operation name to filter by
-            start_time: Start of time range (inclusive)
-            end_time: End of time range (inclusive)
-            limit: Maximum number of results
-
-        Returns:
-            List of benchmarks matching the query
-
-        Example:
-            >>> from datetime import timedelta
-            >>> now = datetime.now(timezone.utc)
-            >>> yesterday = now - timedelta(days=1)
-            >>> benchmarks = await store.query(
-            ...     "steward_analysis",
-            ...     start_time=yesterday,
-            ...     limit=50
-            ... )
-        """
-        if not config.ENABLE_BENCHMARKS:
-            return []
-
-        try:
-            client = await self._get_client()
-            index_key = f"benchmark_index:{operation}"
-
-            # Convert time range to timestamps
-            min_score = (
-                int(start_time.timestamp() * 1000)
-                if start_time
-                else "-inf"
-            )
-            max_score = (
-                int(end_time.timestamp() * 1000)
-                if end_time
-                else "+inf"
-            )
-
-            # Query sorted set
-            keys = await client.zrevrangebyscore(
-                index_key,
-                max_score,
-                min_score,
-                start=0,
-                num=limit,
-            )
-
-            # Fetch benchmark data
-            benchmarks = []
-            for key in keys:
-                data = await client.hgetall(key)
-                if data:
-                    benchmarks.append(PerformanceBenchmark.from_redis_dict(data))
-
-            return benchmarks
-
-        except Exception as e:
-            logger.error(
-                "benchmark_query_failed",
-                error=str(e),
-                operation=operation,
-            )
-            return []
-
-    async def get_statistics(
-        self,
-        operation: str,
-        start_time: Optional[datetime] = None,
-        end_time: Optional[datetime] = None,
-    ) -> dict[str, Any]:
-        """
-        Get aggregate statistics for an operation.
-
-        Args:
-            operation: Operation name
-            start_time: Start of time range
-            end_time: End of time range
-
-        Returns:
-            Dictionary with statistics (count, avg_duration, success_rate, etc.)
-
-        Example:
-            >>> stats = await store.get_statistics("steward_analysis")
-            >>> print(f"Average duration: {stats['avg_duration']}s")
-            >>> print(f"Success rate: {stats['success_rate']}%")
-        """
-        benchmarks = await self.query(operation, start_time, end_time, limit=1000)
-
-        if not benchmarks:
-            return {
-                "count": 0,
-                "avg_duration": 0.0,
-                "min_duration": 0.0,
-                "max_duration": 0.0,
-                "success_rate": 0.0,
-            }
-
-        durations = [b.duration_seconds for b in benchmarks]
-        successes = sum(1 for b in benchmarks if b.success)
-
-        return {
-            "count": len(benchmarks),
-            "avg_duration": sum(durations) / len(durations),
-            "min_duration": min(durations),
-            "max_duration": max(durations),
-            "success_rate": (successes / len(benchmarks)) * 100,
-            "total_successes": successes,
-            "total_failures": len(benchmarks) - successes,
-        }
-
-    async def get_tool_accuracy(
-        self,
-        start_time: Optional[datetime] = None,
-        end_time: Optional[datetime] = None,
-    ) -> dict[str, Any]:
-        """
-        Analyze tool recommendation accuracy.
-
-        Compares recommended tools vs actually used tools to measure
-        Steward's recommendation precision.
-
-        Args:
-            start_time: Start of time range
-            end_time: End of time range
-
-        Returns:
-            Dictionary with accuracy metrics
-
-        Example:
-            >>> accuracy = await store.get_tool_accuracy()
-            >>> print(f"Precision: {accuracy['precision']}%")
-        """
-        tool_calls = await self.query("tool_call", start_time, end_time, limit=1000)
-
-        if not tool_calls:
-            return {
-                "total_calls": 0,
-                "recommended_and_used": 0,
-                "recommended_not_used": 0,
-                "not_recommended_but_used": 0,
-                "precision": 0.0,
-            }
-
-        recommended_and_used = sum(
-            1 for b in tool_calls
-            if b.was_recommended and b.was_actually_used
-        )
-        not_recommended_but_used = sum(
-            1 for b in tool_calls
-            if not b.was_recommended and b.was_actually_used
-        )
-
-        total_used = sum(1 for b in tool_calls if b.was_actually_used)
-        precision = (
-            (recommended_and_used / total_used * 100) if total_used > 0 else 0.0
-        )
-
-        return {
-            "total_calls": len(tool_calls),
-            "total_used": total_used,
-            "recommended_and_used": recommended_and_used,
-            "not_recommended_but_used": not_recommended_but_used,
-            "precision": precision,
-        }
-
-    async def close(self) -> None:
-        """Close Redis connection."""
-        if self._client:
-            await self._client.aclose()
-            self._client = None
-
-
-# Global benchmark store instance
-_benchmark_store: Optional[BenchmarkStore] = None
-
-
-def get_benchmark_store() -> BenchmarkStore:
-    """
-    Get global benchmark store instance.
-
-    Returns:
-        BenchmarkStore instance
-    """
-    global _benchmark_store
-    if _benchmark_store is None:
-        _benchmark_store = BenchmarkStore()
-    return _benchmark_store

src/core/config.py

@@ -2,13 +2,22 @@
 Global application configuration.
 Following best practice of splitting config across domains.
 """
+
 from enum import Enum
 from functools import lru_cache
 from pathlib import Path
 
-from pydantic import Field, HttpUrl
+from pydantic import Field, HttpUrl, model_validator
 from pydantic_settings import BaseSettings, SettingsConfigDict
 
+# Tenant isolation constants (see docs: tenant-based isolation, no separate
+# test infrastructure). The production tenant owns real data in the shared
+# services (Qdrant/Neo4j/Wiki.js/Redis); everything non-production must run
+# under the reserved test tenant or an explicit test_-prefixed namespace.
+PRODUCTION_TENANT = "jpmschweitzer"
+TEST_TENANT = "llm_tester"
+TEST_TENANT_PREFIX = "test_"
+
 
 def _get_version_from_pyproject() -> str:
     """
@@ -34,6 +43,7 @@ def _get_version_from_pyproject() -> str:
 
 class Environment(str, Enum):
     """Application environment."""
+
     DEVELOPMENT = "development"
     PRODUCTION = "production"
     TESTING = "testing"
@@ -42,135 +52,149 @@ class Environment(str, Enum):
 class Config(BaseSettings):
     """
     Global application configuration.
-    
+
     Loads from environment variables and .env file.
     Domain-specific configs should be in their respective modules.
     """
+
     model_config = SettingsConfigDict(
         env_file=".env",
         env_file_encoding="utf-8",
         case_sensitive=True,
         extra="ignore",
     )
-    
+
     # Application
     APP_NAME: str = "OpenAI-Compatible API"
     APP_VERSION: str = Field(default_factory=_get_version_from_pyproject)
     ENVIRONMENT: Environment = Environment.DEVELOPMENT
     DEBUG: bool = Field(default=False, description="Debug mode")
-    
+
     # API Configuration
     API_HOST: str = Field(default="0.0.0.0", description="API host")
     API_PORT: int = Field(default=8000, description="API port")
     API_PREFIX: str = Field(default="/v1", description="API route prefix")
-    
-    # Ollama Configuration
-    OLLAMA_HOST: HttpUrl = Field(
-        default="http://localhost:11434",
-        description="Ollama server URL"
+
+    # Anthropic Configuration (Claude - cloud fallback)
+    ANTHROPIC_API_KEY: str | None = Field(
+        default=None, description="Anthropic API key for the Claude fallback backend"
     )
-    OLLAMA_DEFAULT_MODEL: str = Field(
-        default="mistral-nemo:latest",
-        description="Default Ollama model"
+    ANTHROPIC_MODEL: str = Field(
+        default="claude-sonnet-5", description="Claude model for the fallback backend"
     )
-    OLLAMA_TIMEOUT: int = Field(
-        default=120,
-        description="Ollama request timeout in seconds"
+    PREFER_CLOUD_BACKEND: bool = Field(
+        default=False, description="Prefer Claude over Ollama (default: local-first)"
+    )
+
+    # Ollama Configuration (local - primary backend)
+    OLLAMA_HOST: HttpUrl = Field(default="http://localhost:11434", description="Ollama server URL")
+    OLLAMA_DEFAULT_MODEL: str = Field(default="gemma4:e2b", description="Default Ollama model")
+    OLLAMA_TIMEOUT: int = Field(default=120, description="Ollama request timeout in seconds")
+    STEWARD_TIMEOUT: int = Field(
+        default=60, description="Steward analysis timeout in seconds (gemma4 needs ~35s warm)"
     )
     STREAM_TIMEOUT: int = Field(
-        default=20,
-        description="Timeout for each streaming turn in seconds"
+        default=20, description="Timeout for each streaming turn in seconds"
     )
 
     # SearXNG Configuration
     SEARXNG_HOST: HttpUrl = Field(
-        default="http://localhost:8087",
-        description="SearXNG server URL"
-    )
-    SEARXNG_TIMEOUT: int = Field(
-        default=30,
-        description="SearXNG request timeout in seconds"
+        default="http://searxng:8080",
+        description="SearXNG server URL (container name; internal port 8080)",
     )
+    SEARXNG_TIMEOUT: int = Field(default=30, description="SearXNG request timeout in seconds")
 
     # Redis Configuration
-    REDIS_HOST: str = Field(
-        default="localhost",
-        description="Redis server host"
-    )
-    REDIS_PORT: int = Field(
-        default=6379,
-        description="Redis server port"
-    )
-    REDIS_DB: int = Field(
-        default=1,
-        description="Redis database number"
-    )
-    REDIS_TIMEOUT: int = Field(
-        default=5,
-        description="Redis connection timeout in seconds"
-    )
+    REDIS_HOST: str = Field(default="localhost", description="Redis server host")
+    REDIS_PORT: int = Field(default=6379, description="Redis server port")
+    REDIS_TIMEOUT: int = Field(default=5, description="Redis connection timeout in seconds")
 
     # Library-Desk Configuration (The Librarian backend)
+    LIBRARIAN_TIMEOUT: int = Field(
+        default=180, description="Total time budget for a librarian delegation in seconds"
+    )
     LIBRARY_DESK_HOST: HttpUrl = Field(
-        default="http://localhost:8089",
-        description="Library-Desk API URL"
+        default="http://library-desk:8089",
+        description="Library-Desk API URL (container name; internal port 8089)",
     )
     LIBRARY_DESK_API_KEY: str = Field(
-        default="",
-        description="API key for Library-Desk authentication"
+        default="", description="API key for Library-Desk authentication"
     )
     LIBRARY_DESK_TIMEOUT: int = Field(
-        default=60,
-        description="Library-Desk request timeout in seconds"
+        default=60, description="Library-Desk request timeout in seconds"
     )
 
+    # Core-API Configuration (The Housekeeper backend)
+    CORE_API_HOST: HttpUrl = Field(
+        default="http://core-api:8083",
+        description="Core-API URL for Home Assistant integration (container name; internal port 8083)",
+    )
+    CORE_API_KEY: str = Field(default="", description="API key for Core-API authentication")
+    CORE_API_TIMEOUT: int = Field(default=30, description="Core-API request timeout in seconds")
+
     # Qdrant Configuration (Memory vector storage)
-    QDRANT_HOST: str = Field(
-        default="localhost",
-        description="Qdrant server host"
-    )
-    QDRANT_PORT: int = Field(
-        default=6333,
-        description="Qdrant server port"
-    )
+    QDRANT_HOST: str = Field(default="localhost", description="Qdrant server host")
+    QDRANT_PORT: int = Field(default=6333, description="Qdrant server port")
     QDRANT_EMBEDDING_DIM: int = Field(
-        default=768,
-        description="Embedding dimension (768 for nomic-embed-text)"
+        default=768, description="Embedding dimension (768 for nomic-embed-text)"
     )
 
     # Ollama Embedding Configuration
     OLLAMA_EMBEDDING_MODEL: str = Field(
-        default="nomic-embed-text",
-        description="Ollama model for embeddings"
+        default="nomic-embed-text", description="Ollama model for embeddings"
     )
 
-    # Redis Memory Database (separate from benchmarks)
-    REDIS_MEMORY_DB: int = Field(
-        default=2,
-        description="Redis database number for memory cache"
-    )
-    REDIS_MEMORY_TTL_HOURS: int = Field(
-        default=24,
-        description="TTL for session context in hours"
-    )
+    # Redis Memory Database
+    REDIS_MEMORY_DB: int = Field(default=1, description="Redis database number for memory cache")
+    REDIS_MEMORY_TTL_HOURS: int = Field(default=24, description="TTL for session context in hours")
 
     # Logging
-    LOG_LEVEL: str = Field(default="INFO", description="Logging level")
-    ENABLE_BENCHMARKS: bool = Field(default=True, description="Enable performance benchmarking")
+    LOG_LEVEL: str | None = Field(
+        default=None, description="Logging level (auto-set based on environment if not specified)"
+    )
+
+    # User Configuration
+    DEFAULT_USER: str | None = Field(
+        default=None,
+        description="Default user for single-user setup (auto-set based on environment if not specified)",
+    )
 
     # CORS
-    CORS_ORIGINS: list[str] = Field(
-        default=["*"],
-        description="Allowed CORS origins"
-    )
+    CORS_ORIGINS: list[str] = Field(default=["*"], description="Allowed CORS origins")
     CORS_ALLOW_CREDENTIALS: bool = True
     CORS_ALLOW_METHODS: list[str] = ["*"]
     CORS_ALLOW_HEADERS: list[str] = ["*"]
 
-    @property
-    def redis_url(self) -> str:
-        """Construct Redis connection URL for benchmarks."""
-        return f"redis://{self.REDIS_HOST}:{self.REDIS_PORT}/{self.REDIS_DB}"
+    @model_validator(mode="after")
+    def _refuse_production_tenant_outside_production(self) -> "Config":
+        """
+        Refuse startup when a non-production environment is explicitly
+        configured with the production tenant.
+
+        This is the hard stop of the tenant isolation guard: a dev/test
+        instance must never be able to read or write the production
+        tenant's data in the shared services.
+
+        The comparison is on the sanitized form: namespaces are derived
+        through sanitize_user_id(), so variants like "JPMSchweitzer" or
+        "jpmschweitzer." collide with the production namespaces and are
+        refused just as loudly.
+        """
+        from src.core.multi_tenancy import sanitize_user_id
+
+        if (
+            self.ENVIRONMENT != Environment.PRODUCTION
+            and self.DEFAULT_USER is not None
+            and sanitize_user_id(self.DEFAULT_USER) == sanitize_user_id(PRODUCTION_TENANT)
+        ):
+            raise ValueError(
+                f"Refusing to start: ENVIRONMENT={self.ENVIRONMENT.value} is "
+                f"explicitly configured with the production tenant "
+                f"'{PRODUCTION_TENANT}'. Non-production environments must use "
+                f"'{TEST_TENANT}' or a '{TEST_TENANT_PREFIX}'-prefixed tenant. "
+                f"Unset DEFAULT_USER or set ENVIRONMENT=production."
+            )
+        return self
 
     @property
     def redis_memory_url(self) -> str:
@@ -192,12 +216,58 @@ class Config(BaseSettings):
         """
         return "json" if self.ENVIRONMENT == Environment.PRODUCTION else "console"
 
+    @property
+    def effective_log_level(self) -> str:
+        """
+        Get effective log level, auto-determining from environment if not set.
+
+        - development: DEBUG (maximum verbosity)
+        - production: WARNING (minimal noise)
+        - testing: INFO
+        """
+        if self.LOG_LEVEL is not None:
+            return self.LOG_LEVEL
+        if self.ENVIRONMENT == Environment.DEVELOPMENT:
+            return "DEBUG"
+        if self.ENVIRONMENT == Environment.PRODUCTION:
+            return "WARNING"
+        return "INFO"
+
+    @property
+    def effective_default_user(self) -> str:
+        """
+        Get effective default user (tenant), enforcing tenant isolation.
+
+        - production: DEFAULT_USER if set, else the production tenant
+        - development/testing: FORCED to the reserved test tenant
+          ("llm_tester") - the only accepted overrides are the test tenant
+          itself or a "test_"-prefixed namespace. Any other DEFAULT_USER
+          value is treated as misconfiguration and ignored.
+        """
+        if self.ENVIRONMENT == Environment.PRODUCTION:
+            return self.DEFAULT_USER or PRODUCTION_TENANT
+
+        if self.DEFAULT_USER is not None and (
+            self.DEFAULT_USER == TEST_TENANT or self.DEFAULT_USER.startswith(TEST_TENANT_PREFIX)
+        ):
+            return self.DEFAULT_USER
+        return TEST_TENANT
+
+    @property
+    def tenant_forced(self) -> bool:
+        """Whether the tenant guard overrode a misconfigured DEFAULT_USER."""
+        return (
+            self.ENVIRONMENT != Environment.PRODUCTION
+            and self.DEFAULT_USER is not None
+            and self.effective_default_user != self.DEFAULT_USER
+        )
+
 
 @lru_cache
 def get_config() -> Config:
     """
     Get cached configuration instance.
-    
+
     Uses lru_cache to ensure config is loaded once and reused.
     """
     return Config()

src/core/context.py

@@ -6,7 +6,7 @@ async calls, eliminating the need to thread user identity through every function
 
 Usage:
     # At request entry (router):
-    token = current_user.set(request.user or "jpmschweitzer")
+    token = current_user.set(request.user or get_default_user())
     try:
         await service.process(request)
     finally:
@@ -16,16 +16,64 @@ Usage:
     from src.core.context import get_user
     user = get_user()  # Returns current request's user
 """
-from contextvars import ContextVar
 
-# Default user for single-user homelab setup
-DEFAULT_USER = "jpmschweitzer"
+from contextvars import ContextVar
+from types import TracebackType
+
+
+def get_default_user() -> str:
+    """
+    Get default user from config (environment-aware).
+
+    - development/testing: llm_tester (isolated test scope)
+    - production: jpmschweitzer (real user)
+    """
+    # Import here to avoid circular dependency
+    from src.core.config import config
+
+    return config.effective_default_user
+
 
 # Request-scoped context variables (async-safe, isolated per request)
-current_user: ContextVar[str] = ContextVar("current_user", default=DEFAULT_USER)
-current_conversation: ContextVar[str | None] = ContextVar(
-    "current_conversation", default=None
-)
+# Note: ContextVar default is evaluated at definition, so we use a sentinel
+# and resolve the real default in get_user()
+_USER_NOT_SET = "__user_not_set__"
+current_user: ContextVar[str] = ContextVar("current_user", default=_USER_NOT_SET)
+current_conversation: ContextVar[str | None] = ContextVar("current_conversation", default=None)
+
+
+def apply_tenant_guard(user: str) -> str:
+    """
+    Enforce tenant isolation at request-context resolution.
+
+    In non-production environments the production tenant must never be
+    the effective user - a request that explicitly asks for it is forced
+    to the reserved test tenant instead (with a loud log line).
+
+    Comparison happens on the *sanitized* form of the user: every local
+    namespace (Qdrant collection, Redis key) is derived through
+    sanitize_user_id(), so any raw variant that collides with the
+    production tenant after sanitization ("JPMSchweitzer",
+    "jpmschweitzer.", " jpmschweitzer", ...) would otherwise resolve to
+    the production namespaces. Those variants are forced too.
+    """
+    # Import here to avoid circular dependency
+    from src.core.config import PRODUCTION_TENANT, TEST_TENANT, Environment, config
+    from src.core.multi_tenancy import sanitize_user_id
+
+    if config.ENVIRONMENT != Environment.PRODUCTION and sanitize_user_id(user) == sanitize_user_id(
+        PRODUCTION_TENANT
+    ):
+        from src.core.logging_config import get_logger
+
+        get_logger(__name__).warning(
+            "tenant_guard_forced",
+            environment=config.ENVIRONMENT.value,
+            requested_tenant=user,
+            forced_tenant=TEST_TENANT,
+        )
+        return TEST_TENANT
+    return user
 
 
 def get_user() -> str:
@@ -34,12 +82,17 @@ def get_user() -> str:
 
     Returns:
         User identifier for the current request.
-        Falls back to DEFAULT_USER if not set.
+        Falls back to environment-aware default if not set.
+        In non-production environments the production tenant is never
+        returned - the tenant guard forces the reserved test tenant.
 
     Example:
-        user = get_user()  # "jpmschweitzer" or whatever was set in router
+        user = get_user()  # "llm_tester" (dev) or "jpmschweitzer" (prod)
     """
-    return current_user.get()
+    user = current_user.get()
+    if user == _USER_NOT_SET:
+        return get_default_user()
+    return apply_tenant_guard(user)
 
 
 def get_conversation_id() -> str | None:
@@ -76,10 +129,10 @@ class RequestContext:
         Initialize request context.
 
         Args:
-            user: User identifier (defaults to DEFAULT_USER if None)
+            user: User identifier (defaults to environment-aware user if None)
             conversation_id: Conversation ID (optional)
         """
-        self.user = user or DEFAULT_USER
+        self.user = user or get_default_user()
         self.conversation_id = conversation_id
         self._user_token = None
         self._conv_token = None
@@ -90,7 +143,12 @@ class RequestContext:
         self._conv_token = current_conversation.set(self.conversation_id)
         return self
 
-    async def __aexit__(self, exc_type, exc_val, exc_tb) -> None:
+    async def __aexit__(
+        self,
+        exc_type: type[BaseException] | None,
+        exc_val: BaseException | None,
+        exc_tb: TracebackType | None,
+    ) -> None:
         """Reset context variables on exit."""
         if self._user_token is not None:
             current_user.reset(self._user_token)
@@ -103,7 +161,12 @@ class RequestContext:
         self._conv_token = current_conversation.set(self.conversation_id)
         return self
 
-    def __exit__(self, exc_type, exc_val, exc_tb) -> None:
+    def __exit__(
+        self,
+        exc_type: type[BaseException] | None,
+        exc_val: BaseException | None,
+        exc_tb: TracebackType | None,
+    ) -> None:
         """Sync context manager exit."""
         if self._user_token is not None:
             current_user.reset(self._user_token)

src/core/embeddings.py

@@ -8,7 +8,8 @@ Provides async embedding operations via Ollama API:
 
 Adapted from library-desk patterns.
 """
-from typing import Optional
+
+from types import TracebackType
 
 import httpx
 
@@ -73,7 +74,12 @@ class OllamaEmbeddingClient:
         await self._get_client()
         return self
 
-    async def __aexit__(self, exc_type, exc_val, exc_tb) -> None:
+    async def __aexit__(
+        self,
+        exc_type: type[BaseException] | None,
+        exc_val: BaseException | None,
+        exc_tb: TracebackType | None,
+    ) -> None:
         """Async context manager exit."""
         await self.close()
 

src/core/exceptions.py

@@ -2,12 +2,13 @@
 Global exception definitions.
 Domain-specific exceptions should be in their respective modules.
 """
+
 from typing import Any
 
 
 class AppException(Exception):
     """Base exception for all application errors."""
-    
+
     def __init__(
         self,
         message: str = "An error occurred",
@@ -22,26 +23,26 @@ class AppException(Exception):
 
 class OllamaConnectionError(AppException):
     """Raised when cannot connect to Ollama service."""
-    
+
     def __init__(self, message: str = "Cannot connect to Ollama service"):
         super().__init__(message=message, status_code=503)
 
 
 class OllamaTimeoutError(AppException):
     """Raised when Ollama request times out."""
-    
+
     def __init__(self, message: str = "Ollama request timed out"):
         super().__init__(message=message, status_code=504)
 
 
 class ModelNotFoundError(AppException):
     """Raised when requested model is not available."""
-    
+
     def __init__(self, model_name: str):
         super().__init__(
             message=f"Model '{model_name}' not found",
             status_code=404,
-            details={"model": model_name}
+            details={"model": model_name},
         )
 
 

src/core/household_registry.py

@@ -5,10 +5,10 @@ Provides centralized registry of household members (agents) with their
 capabilities and tools. Supports two-tier abstraction: executive summaries
 for coordination and full toolsets for execution.
 """
-from typing import Any, Optional
+
+from typing import Any
 
 from pydantic import BaseModel, ConfigDict
-from pydantic_ai import Agent
 
 from .logging_config import get_logger
 
@@ -22,6 +22,7 @@ class HouseholdCapability(BaseModel):
     This is what the Steward and Butler see for coordination.
     High-level description without implementation details.
     """
+
     name: str  # Unique identifier: "tatlock_core", "librarian", "developer"
     role: str  # Display name: "Butler's Core Tools", "The Librarian"
     category: str  # "core", "research", "technical", "automation"
@@ -38,11 +39,12 @@ class HouseholdMember(BaseModel):
     Contains both the executive summary (for coordination) and
     implementation details (tools/agent).
     """
+
     model_config = ConfigDict(arbitrary_types_allowed=True)
 
     capability: HouseholdCapability
     tools: list[Any]  # PydanticAI tool definitions (any type since Tool is a dataclass)
-    agent: Optional[Any] = None  # For expert agents (Phase 4)
+    agent: Any | None = None  # For expert agents (Phase 4)
 
 
 class HouseholdRegistry:
@@ -55,7 +57,7 @@ class HouseholdRegistry:
     3. Agent delegation (Phase 4)
     """
 
-    def __init__(self):
+    def __init__(self) -> None:
         """Initialize empty registry."""
         self._members: dict[str, HouseholdMember] = {}
         logger.info("household_registry_initialized")
@@ -65,7 +67,7 @@ class HouseholdRegistry:
         name: str,
         capability: HouseholdCapability,
         tools: list[Any],
-        agent: Optional[Any] = None,
+        agent: Any | None = None,
     ) -> None:
         """
         Register a household member.
@@ -95,9 +97,7 @@ class HouseholdRegistry:
             ... )
         """
         if name != capability.name:
-            raise ValueError(
-                f"Name mismatch: '{name}' != '{capability.name}'"
-            )
+            raise ValueError(f"Name mismatch: '{name}' != '{capability.name}'")
 
         self._members[name] = HouseholdMember(
             capability=capability,
@@ -132,7 +132,7 @@ class HouseholdRegistry:
                 role=member.capability.role,
             )
 
-    def get_member(self, name: str) -> Optional[HouseholdMember]:
+    def get_member(self, name: str) -> HouseholdMember | None:
         """
         Get full household member specification.
 
@@ -223,13 +223,17 @@ class HouseholdRegistry:
             >>> # Returns: [delegate_to_librarian, calculate, datetime, ...]
             >>> # Instead of: [hybrid_search, search_wiki, create_wiki_page, ... (16 tools)]
         """
-        from src.agents.delegation import delegate_to_librarian
+        from src.agents.delegation import (
+            delegate_to_biographer,
+            delegate_to_housekeeper,
+            delegate_to_librarian,
+        )
 
         # Map of expert names to their delegation wrappers
         delegation_wrappers = {
             "librarian": delegate_to_librarian,
-            # Future: "memory": delegate_to_memory,
-            # Future: "home_automation": delegate_to_home_automation,
+            "biographer": delegate_to_biographer,
+            "housekeeper": delegate_to_housekeeper,
         }
 
         tools = []

src/core/logging_config.py

@@ -4,12 +4,14 @@ Structured logging configuration using structlog.
 Deeply integrates with FastAPI/uvicorn's built-in logging to provide
 seamless structured logs across the entire application stack.
 """
+
 import logging
 import logging.config
 import sys
+from collections.abc import AsyncIterator
 from contextlib import asynccontextmanager
-from datetime import datetime, timezone
-from typing import Any, AsyncIterator
+from datetime import UTC, datetime
+from typing import Any
 
 import structlog
 from structlog.types import EventDict, Processor
@@ -19,7 +21,7 @@ from .config import config
 
 def add_timestamp(logger: Any, method_name: str, event_dict: EventDict) -> EventDict:
     """Add ISO 8601 timestamp to log entries."""
-    event_dict["timestamp"] = datetime.now(timezone.utc).isoformat()
+    event_dict["timestamp"] = datetime.now(UTC).isoformat()
     return event_dict
 
 
@@ -41,11 +43,28 @@ def extract_from_record(logger: Any, method_name: str, event_dict: EventDict) ->
         # Extract custom fields from record
         for key, value in record.__dict__.items():
             if key not in {
-                "name", "msg", "args", "created", "filename", "funcName",
-                "levelname", "levelno", "lineno", "module", "msecs",
-                "message", "pathname", "process", "processName", "relativeCreated",
-                "thread", "threadName", "exc_info", "exc_text", "stack_info",
-                "taskName"
+                "name",
+                "msg",
+                "args",
+                "created",
+                "filename",
+                "funcName",
+                "levelname",
+                "levelno",
+                "lineno",
+                "module",
+                "msecs",
+                "message",
+                "pathname",
+                "process",
+                "processName",
+                "relativeCreated",
+                "thread",
+                "threadName",
+                "exc_info",
+                "exc_text",
+                "stack_info",
+                "taskName",
             }:
                 event_dict[key] = value
 
@@ -122,7 +141,7 @@ def configure_logging() -> None:
     root_logger = logging.getLogger()
     root_logger.handlers.clear()
     root_logger.addHandler(handler)
-    root_logger.setLevel(logging.getLevelName(config.LOG_LEVEL))
+    root_logger.setLevel(logging.getLevelName(config.effective_log_level))
 
     # Configure specific loggers
     for logger_name in [
@@ -135,7 +154,7 @@ def configure_logging() -> None:
         logger = logging.getLogger(logger_name)
         logger.handlers.clear()
         logger.propagate = True
-        logger.setLevel(logging.getLevelName(config.LOG_LEVEL))
+        logger.setLevel(logging.getLevelName(config.effective_log_level))
 
 
 def get_logger(name: str) -> structlog.stdlib.BoundLogger:
@@ -164,7 +183,7 @@ def get_logger(name: str) -> structlog.stdlib.BoundLogger:
 async def log_operation(
     operation: str,
     initial_context: dict[str, Any] | None = None,
-    logger_name: str = "tatlock.operations"
+    logger_name: str = "tatlock.operations",
 ) -> AsyncIterator[dict[str, Any]]:
     """
     Context manager for automatic operation timing and logging.
@@ -187,21 +206,21 @@ async def log_operation(
     context = initial_context or {}
     context["operation"] = operation
 
-    start_time = datetime.now(timezone.utc)
+    start_time = datetime.now(UTC)
     logger.info("operation_started", **context)
 
     try:
         yield context
 
         # Success case
-        duration = (datetime.now(timezone.utc) - start_time).total_seconds()
+        duration = (datetime.now(UTC) - start_time).total_seconds()
         context["duration_seconds"] = duration
         context["success"] = True
         logger.info("operation_completed", **context)
 
     except Exception as e:
         # Error case
-        duration = (datetime.now(timezone.utc) - start_time).total_seconds()
+        duration = (datetime.now(UTC) - start_time).total_seconds()
         context["duration_seconds"] = duration
         context["success"] = False
         context["error"] = str(e)
@@ -228,7 +247,8 @@ def get_uvicorn_log_config() -> dict[str, Any]:
                 "()": structlog.stdlib.ProcessorFormatter,
                 "processors": [
                     structlog.stdlib.ProcessorFormatter.remove_processors_meta,
-                    structlog.processors.JSONRenderer() if config.log_format == "json"
+                    structlog.processors.JSONRenderer()
+                    if config.log_format == "json"
                     else structlog.dev.ConsoleRenderer(colors=True),
                 ],
             },
@@ -241,9 +261,9 @@ def get_uvicorn_log_config() -> dict[str, Any]:
             },
         },
         "loggers": {
-            "uvicorn": {"handlers": ["default"], "level": config.LOG_LEVEL},
-            "uvicorn.error": {"handlers": ["default"], "level": config.LOG_LEVEL},
-            "uvicorn.access": {"handlers": ["default"], "level": config.LOG_LEVEL},
+            "uvicorn": {"handlers": ["default"], "level": config.effective_log_level},
+            "uvicorn.error": {"handlers": ["default"], "level": config.effective_log_level},
+            "uvicorn.access": {"handlers": ["default"], "level": config.effective_log_level},
         },
     }
 

src/core/memory_cache.py

@@ -6,8 +6,9 @@ Provides short-term memory storage with TTL:
 - Recent entities mentioned in conversation
 - User-scoped with conversation isolation
 
-Uses Redis DB 2 (separate from benchmarks in DB 1).
+Uses Redis DB 1.
 """
+
 import json
 from typing import Any
 
@@ -15,7 +16,7 @@ import redis.asyncio as redis
 
 from .config import config
 from .logging_config import get_logger
-from .multi_tenancy import get_session_key, get_entities_key
+from .multi_tenancy import get_entities_key, get_session_key
 
 logger = get_logger(__name__)
 

src/core/memory_service.py

@@ -21,14 +21,14 @@ Usage:
     # Get session context
     ctx = await memory_service.get_session_context(conversation_id)
 """
-from datetime import datetime, timezone
+
+from datetime import UTC, datetime
 from enum import Enum
 from typing import Any
 
 from pydantic import BaseModel, Field
 
-from .config import config
-from .context import get_user, get_conversation_id
+from .context import get_conversation_id, get_user
 from .embeddings import get_embedding_client
 from .logging_config import get_logger
 from .memory_cache import get_memory_cache
@@ -40,22 +40,24 @@ logger = get_logger(__name__)
 
 class MemoryType(str, Enum):
     """Types of memories stored in Qdrant."""
-    USER_PROFILE = "user_profile"      # Name, location, timezone
-    PREFERENCE = "preference"          # Units, language, theme
-    LEARNED_FACT = "learned_fact"      # "My car is a Tesla"
+
+    USER_PROFILE = "user_profile"  # Name, location, timezone
+    PREFERENCE = "preference"  # Units, language, theme
+    LEARNED_FACT = "learned_fact"  # "My car is a Tesla"
 
 
 class MemoryRecord(BaseModel):
     """A memory record stored in Qdrant."""
+
     id: str
     type: MemoryType
-    key: str                           # e.g., "location", "timezone", "car"
-    value: str                         # The actual content
+    key: str  # e.g., "location", "timezone", "car"
+    value: str  # The actual content
     keywords: list[str] = Field(default_factory=list)
-    importance: float = 0.5            # 0.0 - 1.0
-    source: str = "explicit"           # "explicit" | "inferred" | "conversation"
-    created_at: str = Field(default_factory=lambda: datetime.now(timezone.utc).isoformat())
-    updated_at: str = Field(default_factory=lambda: datetime.now(timezone.utc).isoformat())
+    importance: float = 0.5  # 0.0 - 1.0
+    source: str = "explicit"  # "explicit" | "inferred" | "conversation"
+    created_at: str = Field(default_factory=lambda: datetime.now(UTC).isoformat())
+    updated_at: str = Field(default_factory=lambda: datetime.now(UTC).isoformat())
 
 
 class MemoryService:
@@ -72,7 +74,7 @@ class MemoryService:
     - Semantic recall: "What did I mention about X?" → Use Memory Agent
     """
 
-    def __init__(self):
+    def __init__(self) -> None:
         """Initialize memory service with lazy client loading."""
         self._qdrant = None
         self._embedding = None
@@ -528,7 +530,7 @@ class MemoryService:
                 "keywords": keywords,
                 "importance": importance,
                 "source": source,
-                "updated_at": datetime.now(timezone.utc).isoformat(),
+                "updated_at": datetime.now(UTC).isoformat(),
             }
 
             result = await self.qdrant.upsert_memory(

src/core/models.py

@@ -2,6 +2,7 @@
 Custom Pydantic base models for consistent serialization.
 Following best practice of having a global base model.
 """
+
 from datetime import datetime
 from typing import Any
 
@@ -17,12 +18,13 @@ def datetime_to_iso_str(dt: datetime) -> str:
 class CustomBaseModel(BaseModel):
     """
     Custom base model with consistent configuration.
-    
+
     All domain models should inherit from this for:
     - Consistent JSON serialization
     - Timezone-aware datetime handling
     - Alias population support
     """
+
     model_config = ConfigDict(
         json_encoders={datetime: datetime_to_iso_str},
         populate_by_name=True,
@@ -30,14 +32,13 @@ class CustomBaseModel(BaseModel):
         validate_assignment=True,
         arbitrary_types_allowed=True,
     )
-    
+
     def serializable_dict(self, **kwargs: Any) -> dict[str, Any]:
         """
         Return dict with only JSON-serializable fields.
-        
+
         Useful for logging and debugging.
         """
         return jsonable_encoder(
-            self.model_dump(**kwargs),
-            custom_encoder={datetime: datetime_to_iso_str}
+            self.model_dump(**kwargs), custom_encoder={datetime: datetime_to_iso_str}
         )

src/core/multi_tenancy.py

@@ -7,6 +7,7 @@ Provides utilities for user namespace management across:
 
 Adapted from library-desk patterns.
 """
+
 import re
 
 
@@ -39,13 +40,13 @@ def sanitize_user_id(user_id: str) -> str:
     sanitized = sanitized.replace(".", "_")
 
     # Replace any non-alphanumeric characters with underscores
-    sanitized = re.sub(r'[^a-z0-9_]', '_', sanitized)
+    sanitized = re.sub(r"[^a-z0-9_]", "_", sanitized)
 
     # Remove consecutive underscores
-    sanitized = re.sub(r'_+', '_', sanitized)
+    sanitized = re.sub(r"_+", "_", sanitized)
 
     # Remove leading/trailing underscores
-    sanitized = sanitized.strip('_')
+    sanitized = sanitized.strip("_")
 
     return sanitized
 
@@ -141,7 +142,7 @@ def validate_user_id(user_id: str) -> bool:
         return False
 
     # Must contain at least one alphanumeric character
-    if not re.search(r'[a-zA-Z0-9]', user_id):
+    if not re.search(r"[a-zA-Z0-9]", user_id):
         return False
 
     return True

src/core/preprocessing.py

@@ -3,14 +3,16 @@ Request preprocessing pipeline.
 
 Analyzes requests via the Steward and creates scoped toolsets for Tatlock.
 """
+
 from dataclasses import dataclass
 from datetime import datetime
-from typing import Any, Optional
+from typing import Any
 
 from src.agents.steward import analyze_request, format_steward_note
 from src.agents.steward.schemas import StewardRecommendation
 from src.core.household_registry import get_household_registry
 from src.core.logging_config import get_logger
+from src.core.tracing import SpanType, trace_span
 
 logger = get_logger(__name__)
 
@@ -44,6 +46,7 @@ class EnrichedRequest:
         recommendation: Full Steward recommendation
         steward_reasoning: Plain text reasoning for streaming to user
     """
+
     original_request: str
     steward_note: str
     scoped_tools: list[Any]  # PydanticAI tool definitions
@@ -54,7 +57,7 @@ class EnrichedRequest:
 async def preprocess_request(
     user_request: str,
     conversation_history: list[dict],
-    conversation_id: Optional[str] = None,
+    conversation_id: str | None = None,
 ) -> EnrichedRequest:
     """
     Analyze request via Steward and prepare scoped context for Tatlock.
@@ -93,12 +96,34 @@ async def preprocess_request(
         conversation_id=conversation_id,
     )
 
-    # Call Steward with full conversation history
-    recommendation = await analyze_request(
-        enriched_request,
-        conversation_history=conversation_history,
-        conversation_id=conversation_id,
-    )
+    # Call Steward with full conversation history (traced)
+    async with trace_span(
+        "steward_analysis",
+        SpanType.STEWARD,
+        metadata={
+            "request_preview": user_request[:100],
+            "history_length": len(conversation_history),
+        },
+    ) as span:
+        recommendation = await analyze_request(
+            enriched_request,
+            conversation_history=conversation_history,
+            conversation_id=conversation_id,
+        )
+
+        # Update span with results
+        if span:
+            span.metadata.update(
+                {
+                    "recommended_capabilities": recommendation.recommended_capabilities,
+                    "complexity": recommendation.estimated_complexity,
+                    "has_memory_context": bool(recommendation.memory_context),
+                    "has_conversation_context": recommendation.conversation_context.has_previous_context,
+                }
+            )
+            span.details["reasoning"] = recommendation.reasoning
+            if recommendation.enriched_query:
+                span.details["enriched_query"] = recommendation.enriched_query
 
     # Format note for Tatlock (includes conversation context)
     steward_note = await format_steward_note(recommendation)
@@ -107,9 +132,7 @@ async def preprocess_request(
     # Uses agent-as-tool pattern: expert agents get delegation wrappers,
     # core tools are returned directly
     registry = get_household_registry()
-    scoped_tools = registry.get_delegation_tools(
-        recommendation.recommended_capabilities
-    )
+    scoped_tools = registry.get_delegation_tools(recommendation.recommended_capabilities)
 
     logger.info(
         "preprocessing_complete",

src/core/qdrant.py

@@ -8,8 +8,9 @@ Provides async operations for storing and retrieving memory embeddings:
 
 Adapted from library-desk patterns.
 """
+
 from typing import Any
-from uuid import uuid4
+from uuid import NAMESPACE_DNS, uuid4, uuid5
 
 from qdrant_client import QdrantClient
 from qdrant_client.http import models as qdrant_models
@@ -150,15 +151,24 @@ class MemoryQdrantClient:
             ... )
         """
         collection_name = get_memory_collection_name(user)
-        memory_id = memory_id or f"mem_{uuid4().hex[:16]}"
+
+        # Generate deterministic UUID from memory_id (or random if not provided)
+        # Qdrant requires UUID or integer IDs, not arbitrary strings
+        if memory_id:
+            # Deterministic UUID from string - same memory_id = same UUID
+            point_id = str(uuid5(NAMESPACE_DNS, f"{user}:{memory_id}"))
+        else:
+            point_id = str(uuid4())
+            memory_id = point_id  # Use UUID as the memory_id too
 
         try:
             # Ensure collection exists
             await self.ensure_collection(user)
 
-            # Create point
+            # Create point (store original memory_id in payload for reference)
+            payload["memory_id"] = memory_id
             point = qdrant_models.PointStruct(
-                id=memory_id,
+                id=point_id,
                 vector=vector,
                 payload=payload,
             )
@@ -232,14 +242,14 @@ class MemoryQdrantClient:
                     ]
                 )
 
-            # Search
-            results = self._client.search(
+            # Search using new Query API (qdrant-client >= 1.10)
+            results = self._client.query_points(
                 collection_name=collection_name,
-                query_vector=query_vector,
+                query=query_vector,
                 limit=limit,
                 query_filter=query_filter,
                 score_threshold=score_threshold,
-            )
+            ).points
 
             # Format results
             memories = []
@@ -279,11 +289,13 @@ class MemoryQdrantClient:
             Memory data or None if not found
         """
         collection_name = get_memory_collection_name(user)
+        # Convert memory_id to UUID point_id
+        point_id = str(uuid5(NAMESPACE_DNS, f"{user}:{memory_id}"))
 
         try:
             points = self._client.retrieve(
                 collection_name=collection_name,
-                ids=[memory_id],
+                ids=[point_id],
             )
 
             if not points:
@@ -320,12 +332,14 @@ class MemoryQdrantClient:
             True
         """
         collection_name = get_memory_collection_name(user)
+        # Convert memory_id to UUID point_id
+        point_id = str(uuid5(NAMESPACE_DNS, f"{user}:{memory_id}"))
 
         try:
             self._client.delete(
                 collection_name=collection_name,
                 points_selector=qdrant_models.PointIdsList(
-                    points=[memory_id],
+                    points=[point_id],
                 ),
             )
 

src/core/router.py

@@ -1,6 +1,7 @@
 """
 Core router for health and root endpoints.
 """
+
 import logging
 
 from fastapi import APIRouter
@@ -16,7 +17,7 @@ router = APIRouter(tags=["core"])
 async def health_check() -> dict[str, str]:
     """
     Health check endpoint.
-    
+
     Returns:
         Health status
     """
@@ -27,7 +28,7 @@ async def health_check() -> dict[str, str]:
 async def root() -> dict[str, str]:
     """
     Root endpoint.
-    
+
     Returns:
         API information
     """

src/core/startup.py

@@ -5,16 +5,49 @@ Handles initialization of household registry and other startup tasks.
 This module should be called during application startup to register
 all household members.
 """
+
 from src.agents.biographer import register_biographer
+from src.agents.housekeeper import register_housekeeper
 from src.agents.librarian import register_librarian
 from src.agents.tatlock_core import TATLOCK_CORE_CAPABILITY, tatlock_core_tools
+from src.anthropic.model_selector import (
+    check_claude_health,
+    check_ollama_health,
+    get_model_info,
+)
+from src.core.config import Environment, config
 from src.core.household_registry import get_household_registry
 from src.core.logging_config import get_logger
 
 logger = get_logger(__name__)
 
 
-def register_household_members():
+def log_tenant_guard() -> None:
+    """
+    Emit one loud startup log line stating the effective tenant.
+
+    In non-production environments the tenant guard forces the reserved
+    test tenant regardless of DEFAULT_USER misconfiguration - this line
+    makes that override visible at startup.
+    """
+    if config.ENVIRONMENT == Environment.PRODUCTION:
+        logger.info(
+            "tenant_guard_production",
+            environment=config.ENVIRONMENT.value,
+            tenant=config.effective_default_user,
+        )
+        return
+
+    logger.warning(
+        "tenant_guard_active",
+        environment=config.ENVIRONMENT.value,
+        forced_tenant=config.effective_default_user,
+        default_user_overridden=config.tenant_forced,
+        configured_default_user=config.DEFAULT_USER,
+    )
+
+
+def register_household_members() -> None:
     """
     Register all household members with the registry.
 
@@ -64,25 +97,50 @@ def register_household_members():
             error=str(e),
         )
 
+    # Register The Housekeeper (Home Automation)
+    try:
+        register_housekeeper()
+    except Exception as e:
+        # Don't fail startup if Housekeeper registration fails
+        logger.warning(
+            "housekeeper_registration_failed",
+            error=str(e),
+        )
+
     logger.info(
         "household_registration_complete",
         total_members=len(registry),
     )
 
 
-def initialize_application():
+async def initialize_application() -> None:
     """
     Initialize the application.
 
     Performs all startup tasks:
-    1. Register household members
-    2. (Future) Initialize connections
-    3. (Future) Load configuration
+    1. Check Ollama (primary) and Claude (fallback) health for backend selection
+    2. Register household members
+    3. (Future) Initialize connections
 
     This should be called once during application startup.
     """
     logger.info("application_initialization_starting")
 
+    # Tenant isolation guard: state the effective tenant loudly
+    log_tenant_guard()
+
+    # Check backend health: Ollama is primary, Claude is the fallback
+    await check_ollama_health()
+    await check_claude_health()
+    model_info = get_model_info()
+    logger.info(
+        "model_backend_configured",
+        backend=model_info["backend"],
+        model=model_info["model"],
+        ollama_available=model_info["ollama_available"],
+        claude_available=model_info["claude_available"],
+    )
+
     # Register household members
     register_household_members()
 

src/core/tool_tracking.py

@@ -1,13 +1,10 @@
 """
-Tool call tracking and benchmarking.
+Tool call tracking.
 
 Tracks which tools are recommended by the Steward versus which tools
-are actually used by Tatlock, recording benchmarks for analysis.
+are actually used by Tatlock for debugging and analysis.
 """
-from datetime import datetime, timezone
-from typing import Optional
 
-from src.core.benchmarks import PerformanceBenchmark, get_benchmark_store
 from src.core.logging_config import get_logger
 
 logger = get_logger(__name__)
@@ -15,17 +12,13 @@ logger = get_logger(__name__)
 
 class ToolCallTracker:
     """
-    Tracks tool calls for benchmarking and accuracy analysis.
+    Tracks tool calls for accuracy analysis.
 
     Compares Steward's recommendations with Tatlock's actual tool usage
     to measure recommendation accuracy.
     """
 
-    def __init__(
-        self,
-        recommended_capabilities: list[str],
-        conversation_id: Optional[str] = None
-    ):
+    def __init__(self, recommended_capabilities: list[str], conversation_id: str | None = None):
         """
         Initialize tool call tracker.
 
@@ -43,7 +36,21 @@ class ToolCallTracker:
             conversation_id=conversation_id,
         )
 
-    async def track_call(self, tool_name: str, duration: float):
+    def _extract_capability(self, tool_name: str) -> str:
+        """
+        Extract capability name from tool name.
+
+        Tool names like 'delegate_to_librarian' map to capability 'librarian'.
+        """
+        if tool_name.startswith("delegate_to_"):
+            return tool_name.replace("delegate_to_", "")
+        return tool_name
+
+    def log_call(self, message: str) -> None:
+        """Log a tool call message (for UI display)."""
+        logger.debug("tool_call_message", message=message)
+
+    async def track_call(self, tool_name: str, duration: float) -> None:
         """
         Record a tool call with timing.
 
@@ -56,8 +63,9 @@ class ToolCallTracker:
             self.actual_calls[tool_name] = []
         self.actual_calls[tool_name].append(duration)
 
-        # Check if tool was recommended
-        was_recommended = tool_name in self.recommended_capabilities
+        # Check if tool was recommended (normalize tool name to capability)
+        capability = self._extract_capability(tool_name)
+        was_recommended = capability in self.recommended_capabilities
 
         if not was_recommended:
             logger.warning(
@@ -67,23 +75,6 @@ class ToolCallTracker:
                 recommended=list(self.recommended_capabilities),
             )
 
-        # Record benchmark to Redis
-        benchmark = PerformanceBenchmark(
-            timestamp=datetime.now(timezone.utc),
-            operation="tool_call",
-            duration_seconds=duration,
-            success=True,  # If we got here, the call succeeded
-            tool_name=tool_name,
-            was_recommended=was_recommended,
-            was_actually_used=True,
-            conversation_id=self.conversation_id,
-            metadata={
-                "recommended_capabilities": list(self.recommended_capabilities),
-            },
-        )
-
-        await get_benchmark_store().record(benchmark)
-
         logger.debug(
             "tool_call_tracked",
             tool_name=tool_name,
@@ -91,15 +82,17 @@ class ToolCallTracker:
             was_recommended=was_recommended,
         )
 
-    async def finalize(self):
+    async def finalize(self) -> None:
         """
         Finalize tracking and log unused recommended tools.
 
         Called after Tatlock completes its response to identify
         tools that were recommended but never used.
         """
+        # Normalize actual tool names to capabilities for comparison
+        used_capabilities = {self._extract_capability(tool) for tool in self.actual_calls.keys()}
         # Find tools that were recommended but not used
-        unused_tools = self.recommended_capabilities - set(self.actual_calls.keys())
+        unused_tools = self.recommended_capabilities - used_capabilities
 
         if unused_tools:
             logger.info(
@@ -109,24 +102,6 @@ class ToolCallTracker:
                 conversation_id=self.conversation_id,
             )
 
-            # Record benchmarks for unused recommendations
-            for tool_name in unused_tools:
-                benchmark = PerformanceBenchmark(
-                    timestamp=datetime.now(timezone.utc),
-                    operation="tool_call",
-                    duration_seconds=0.0,  # Not used
-                    success=True,
-                    tool_name=tool_name,
-                    was_recommended=True,
-                    was_actually_used=False,
-                    conversation_id=self.conversation_id,
-                    metadata={
-                        "recommended_capabilities": list(self.recommended_capabilities),
-                        "reason": "recommended_but_unused",
-                    },
-                )
-                await get_benchmark_store().record(benchmark)
-
         # Log summary
         total_calls = sum(len(durations) for durations in self.actual_calls.values())
         logger.info(
@@ -145,7 +120,9 @@ class ToolCallTracker:
             Dict with tracking statistics
         """
         total_calls = sum(len(durations) for durations in self.actual_calls.values())
-        unused = self.recommended_capabilities - set(self.actual_calls.keys())
+        # Normalize actual tool names to capabilities for comparison
+        used_capabilities = {self._extract_capability(tool) for tool in self.actual_calls.keys()}
+        unused = self.recommended_capabilities - used_capabilities
 
         return {
             "recommended_capabilities": list(self.recommended_capabilities),
@@ -153,12 +130,8 @@ class ToolCallTracker:
             "tools_unused": list(unused),
             "total_calls": total_calls,
             "accuracy": {
-                "recommended_and_used": len(
-                    self.recommended_capabilities & set(self.actual_calls.keys())
-                ),
+                "recommended_and_used": len(self.recommended_capabilities & used_capabilities),
                 "recommended_but_unused": len(unused),
-                "not_recommended_but_used": len(
-                    set(self.actual_calls.keys()) - self.recommended_capabilities
-                ),
+                "not_recommended_but_used": len(used_capabilities - self.recommended_capabilities),
             },
         }

src/core/tracing.py

@@ -0,0 +1,443 @@
+"""
+Lightweight request tracing for local development.
+
+Captures the full request flow through Tatlock's multi-agent architecture
+as structured JSON traces for debugging and optimization.
+
+Enable via DEBUG=true environment variable.
+
+Traces are written to logs/traces/{trace_id}.json
+View with logs/traces/viewer.html
+"""
+
+import json
+import secrets
+from contextlib import asynccontextmanager
+from contextvars import ContextVar
+from dataclasses import dataclass, field
+from datetime import UTC, datetime
+from enum import Enum
+from pathlib import Path
+from typing import Any
+
+from src.core.logging_config import get_logger
+
+logger = get_logger(__name__)
+
+
+class SpanType(str, Enum):
+    """Types of traced operations."""
+
+    ROUTER = "router"
+    STEWARD = "steward"
+    TATLOCK = "tatlock"
+    EXPERT = "expert"
+    TOOL = "tool"
+
+
+class SpanStatus(str, Enum):
+    """Span completion status."""
+
+    OK = "ok"
+    ERROR = "error"
+
+
+@dataclass
+class Span:
+    """A single traced operation."""
+
+    span_id: str
+    name: str
+    type: SpanType
+    start_time: datetime
+    parent_id: str | None = None
+    end_time: datetime | None = None
+    status: SpanStatus = SpanStatus.OK
+    metadata: dict[str, Any] = field(default_factory=dict)
+    details: dict[str, Any] = field(default_factory=dict)
+    children: list[str] = field(default_factory=list)
+    error: str | None = None
+
+    @property
+    def duration_ms(self) -> float | None:
+        """Calculate duration in milliseconds."""
+        if self.end_time and self.start_time:
+            return (self.end_time - self.start_time).total_seconds() * 1000
+        return None
+
+    def to_dict(self) -> dict[str, Any]:
+        """Convert span to dictionary for JSON serialization."""
+        result = {
+            "span_id": self.span_id,
+            "parent_id": self.parent_id,
+            "name": self.name,
+            "type": self.type.value,
+            "start_time": self.start_time.isoformat(),
+            "end_time": self.end_time.isoformat() if self.end_time else None,
+            "duration_ms": round(self.duration_ms, 2) if self.duration_ms else None,
+            "status": self.status.value,
+            "metadata": self.metadata if self.metadata else None,
+        }
+        # Only include non-empty optional fields
+        if self.details:
+            result["details"] = self.details
+        if self.children:
+            result["children"] = self.children
+        if self.error:
+            result["error"] = self.error
+        return {k: v for k, v in result.items() if v is not None}
+
+
+@dataclass
+class Trace:
+    """Complete trace of a request."""
+
+    trace_id: str
+    conversation_id: str | None
+    user: str
+    timestamp: datetime
+    request: dict[str, Any]
+    spans: list[Span] = field(default_factory=list)
+    response: dict[str, Any] | None = None
+    status: str = "in_progress"
+
+    @property
+    def total_duration_ms(self) -> float | None:
+        """Calculate total trace duration from span timings."""
+        if not self.spans:
+            return None
+        start = min(s.start_time for s in self.spans)
+        ends = [s.end_time for s in self.spans if s.end_time]
+        if not ends:
+            return None
+        end = max(ends)
+        return (end - start).total_seconds() * 1000
+
+    def to_dict(self) -> dict[str, Any]:
+        """Convert trace to dictionary for JSON serialization."""
+        return {
+            "trace_id": self.trace_id,
+            "conversation_id": self.conversation_id,
+            "user": self.user,
+            "timestamp": self.timestamp.isoformat(),
+            "total_duration_ms": round(self.total_duration_ms, 2)
+            if self.total_duration_ms
+            else None,
+            "status": self.status,
+            "request": self.request,
+            "response": self.response,
+            "spans": [s.to_dict() for s in self.spans],
+        }
+
+
+# ContextVar for async-safe trace propagation
+_current_trace: ContextVar[Trace | None] = ContextVar("current_trace", default=None)
+_current_span: ContextVar[Span | None] = ContextVar("current_span", default=None)
+
+
+def tracing_enabled() -> bool:
+    """Check if tracing is enabled (requires DEBUG=true)."""
+    from src.core.config import config
+
+    return config.DEBUG
+
+
+def _generate_id(prefix: str = "") -> str:
+    """Generate unique ID with optional prefix."""
+    return f"{prefix}{secrets.token_hex(8)}"
+
+
+def start_trace(
+    conversation_id: str | None,
+    user: str,
+    request: dict[str, Any],
+) -> Trace | None:
+    """
+    Start a new trace for a request.
+
+    Args:
+        conversation_id: Conversation identifier
+        user: User identifier
+        request: Request data (should include preview and full)
+
+    Returns:
+        Trace object if tracing enabled, None otherwise
+    """
+    if not tracing_enabled():
+        return None
+
+    trace = Trace(
+        trace_id=_generate_id("trace_"),
+        conversation_id=conversation_id,
+        user=user,
+        timestamp=datetime.now(UTC),
+        request=request,
+    )
+    _current_trace.set(trace)
+
+    logger.debug("trace_started", trace_id=trace.trace_id, user=user)
+    return trace
+
+
+def get_current_trace() -> Trace | None:
+    """Get the current trace from context."""
+    return _current_trace.get()
+
+
+def get_current_span() -> Span | None:
+    """Get the current span from context."""
+    return _current_span.get()
+
+
+def start_span(
+    name: str,
+    span_type: SpanType,
+    metadata: dict[str, Any] | None = None,
+    details: dict[str, Any] | None = None,
+) -> Span | None:
+    """
+    Start a new span within the current trace.
+
+    Args:
+        name: Span name (e.g., "steward_analysis")
+        span_type: Type of operation
+        metadata: Quick-access metadata (shown in timeline)
+        details: Expandable details (prompts, full responses)
+
+    Returns:
+        Span object if tracing enabled, None otherwise
+    """
+    trace = get_current_trace()
+    if not trace:
+        return None
+
+    parent = get_current_span()
+    span = Span(
+        span_id=_generate_id("span_"),
+        name=name,
+        type=span_type,
+        start_time=datetime.now(UTC),
+        parent_id=parent.span_id if parent else None,
+        metadata=metadata or {},
+        details=details or {},
+    )
+
+    # Add to parent's children list
+    if parent:
+        parent.children.append(span.span_id)
+
+    trace.spans.append(span)
+    _current_span.set(span)
+
+    logger.debug(
+        "span_started",
+        span_id=span.span_id,
+        name=name,
+        type=span_type.value,
+        parent_id=span.parent_id,
+    )
+    return span
+
+
+def end_span(
+    span: Span | None = None,
+    status: SpanStatus = SpanStatus.OK,
+    metadata_update: dict[str, Any] | None = None,
+    details_update: dict[str, Any] | None = None,
+    error: str | None = None,
+) -> None:
+    """
+    End a span and restore parent as current.
+
+    Args:
+        span: Span to end (defaults to current span)
+        status: Completion status
+        metadata_update: Additional metadata to merge
+        details_update: Additional details to merge
+        error: Error message if failed
+    """
+    if span is None:
+        span = get_current_span()
+    if not span:
+        return
+
+    span.end_time = datetime.now(UTC)
+    span.status = status
+    if error:
+        span.error = error
+        span.status = SpanStatus.ERROR
+    if metadata_update:
+        span.metadata.update(metadata_update)
+    if details_update:
+        span.details.update(details_update)
+
+    # Restore parent span as current
+    trace = get_current_trace()
+    if trace and span.parent_id:
+        parent = next((s for s in trace.spans if s.span_id == span.parent_id), None)
+        _current_span.set(parent)
+    else:
+        _current_span.set(None)
+
+    logger.debug(
+        "span_ended",
+        span_id=span.span_id,
+        duration_ms=span.duration_ms,
+        status=status.value,
+    )
+
+
+def end_trace(
+    response: dict[str, Any] | None = None,
+    status: str = "completed",
+) -> str | None:
+    """
+    End the current trace and write to file.
+
+    Args:
+        response: Response data to include
+        status: Final trace status ("completed" or "error")
+
+    Returns:
+        Path to trace file if written, None otherwise
+    """
+    trace = get_current_trace()
+    if not trace:
+        return None
+
+    trace.response = response
+    trace.status = status
+
+    # Write trace to file
+    trace_path = _write_trace(trace)
+
+    # Clear context
+    _current_trace.set(None)
+    _current_span.set(None)
+
+    logger.info(
+        "trace_completed",
+        trace_id=trace.trace_id,
+        total_duration_ms=round(trace.total_duration_ms, 2) if trace.total_duration_ms else None,
+        span_count=len(trace.spans),
+        path=str(trace_path) if trace_path else None,
+    )
+
+    return str(trace_path) if trace_path else None
+
+
+def _write_trace(trace: Trace) -> Path | None:
+    """Write trace to JSON file."""
+    try:
+        # Ensure traces directory exists
+        traces_dir = Path("logs/traces")
+        traces_dir.mkdir(parents=True, exist_ok=True)
+
+        # Write trace file
+        trace_path = traces_dir / f"{trace.trace_id}.json"
+        with open(trace_path, "w") as f:
+            json.dump(trace.to_dict(), f, indent=2, default=str)
+
+        return trace_path
+
+    except Exception as e:
+        logger.error("trace_write_failed", error=str(e), trace_id=trace.trace_id)
+        return None
+
+
+@asynccontextmanager
+async def trace_span(
+    name: str,
+    span_type: SpanType,
+    metadata: dict[str, Any] | None = None,
+    details: dict[str, Any] | None = None,
+):
+    """
+    Async context manager for tracing a span.
+
+    Automatically handles start/end timing and error capture.
+
+    Usage:
+        async with trace_span("steward_analysis", SpanType.STEWARD) as span:
+            result = await analyze_request(...)
+            if span:
+                span.metadata["result_count"] = len(result)
+
+    Args:
+        name: Span name
+        span_type: Type of operation
+        metadata: Initial metadata
+        details: Initial details (expandable in viewer)
+
+    Yields:
+        Span object or None if tracing disabled
+    """
+    span = start_span(name, span_type, metadata, details)
+    try:
+        yield span
+    except Exception as e:
+        end_span(span, SpanStatus.ERROR, error=str(e))
+        raise
+    else:
+        end_span(span, SpanStatus.OK)
+
+
+def add_tool_spans_from_messages(messages: list[Any], parent_span: Span | None = None) -> None:
+    """
+    Extract tool calls from PydanticAI result messages and add as child spans.
+
+    Call this after an agent.run() to capture tool-level timing retroactively.
+    Note: Since we don't have actual timing, we estimate based on sequence.
+
+    Args:
+        messages: List from result.new_messages()
+        parent_span: Parent span to attach tool spans to
+    """
+    trace = get_current_trace()
+    if not trace or not parent_span:
+        return
+
+    # Import PydanticAI message types
+    try:
+        from pydantic_ai.messages import ModelRequest, ModelResponse, ToolCallPart, ToolReturnPart
+    except ImportError:
+        return
+
+    # Track tool calls and their returns
+    tool_calls: dict[str, dict[str, Any]] = {}
+
+    for msg in messages:
+        if isinstance(msg, ModelResponse):
+            for part in msg.parts:
+                if isinstance(part, ToolCallPart):
+                    tool_calls[part.tool_call_id] = {
+                        "name": part.tool_name,
+                        "args": part.args if hasattr(part, "args") else {},
+                    }
+        elif isinstance(msg, ModelRequest):
+            for part in msg.parts:
+                if isinstance(part, ToolReturnPart):
+                    if part.tool_call_id in tool_calls:
+                        tool_info = tool_calls[part.tool_call_id]
+                        # Create a span for this tool call
+                        span = Span(
+                            span_id=_generate_id("span_"),
+                            name=tool_info["name"],
+                            type=SpanType.TOOL,
+                            start_time=parent_span.start_time,  # Approximate
+                            end_time=parent_span.end_time or datetime.now(UTC),
+                            parent_id=parent_span.span_id,
+                            status=SpanStatus.OK,
+                            metadata={
+                                "tool_name": tool_info["name"],
+                                "args_preview": str(tool_info.get("args", {}))[:100],
+                            },
+                            details={
+                                "args": tool_info.get("args", {}),
+                                "result": part.content[:2000]
+                                if isinstance(part.content, str)
+                                else str(part.content)[:2000],
+                            },
+                        )
+                        parent_span.children.append(span.span_id)
+                        trace.spans.append(span)

src/core/tracing_router.py

@@ -0,0 +1,159 @@
+"""
+Trace viewer router.
+
+Serves the trace viewer UI and trace files when tracing is enabled.
+Only available when DEBUG=true.
+"""
+
+from datetime import UTC
+from pathlib import Path
+from typing import Any
+
+from fastapi import APIRouter, HTTPException
+from fastapi.responses import HTMLResponse, JSONResponse
+
+from src.core.config import config
+from src.core.logging_config import get_logger
+
+logger = get_logger(__name__)
+
+router = APIRouter(prefix="/traces", tags=["traces"])
+
+TRACES_DIR = Path("logs/traces")
+VIEWER_PATH = TRACES_DIR / "viewer.html"
+
+
+def tracing_enabled() -> bool:
+    """Check if tracing is enabled."""
+    return config.DEBUG
+
+
+@router.get("", response_class=HTMLResponse)
+async def get_trace_viewer():
+    """
+    Serve the trace viewer UI.
+
+    Returns the standalone HTML viewer for browsing traces.
+    """
+    if not tracing_enabled():
+        raise HTTPException(status_code=404, detail="Tracing not enabled")
+
+    if not VIEWER_PATH.exists():
+        raise HTTPException(status_code=404, detail="Viewer not found")
+
+    return HTMLResponse(content=VIEWER_PATH.read_text())
+
+
+@router.get("/list")
+async def list_traces(
+    limit: int = 50,
+    since_minutes: int | None = None,
+    status: str | None = None,
+    search: str | None = None,
+):
+    """
+    List available trace files.
+
+    Returns most recent traces first, with basic metadata.
+
+    Args:
+        limit: Maximum number of traces to return (default 50)
+        since_minutes: Only return traces from the last N minutes
+        status: Filter by status (completed, error, streaming)
+        search: Search in request preview text
+    """
+    if not tracing_enabled():
+        raise HTTPException(status_code=404, detail="Tracing not enabled")
+
+    if not TRACES_DIR.exists():
+        return {"traces": [], "total": 0}
+
+    import json
+    from datetime import datetime, timedelta
+
+    # Calculate cutoff time if filtering by time
+    cutoff_time = None
+    if since_minutes:
+        cutoff_time = datetime.now(UTC) - timedelta(minutes=since_minutes)
+
+    # Get all trace files, sorted by modification time (newest first)
+    trace_files = sorted(
+        TRACES_DIR.glob("trace_*.json"),
+        key=lambda p: p.stat().st_mtime,
+        reverse=True,
+    )
+
+    traces: list[dict[str, Any]] = []
+    for path in trace_files:
+        if len(traces) >= limit:
+            break
+
+        try:
+            with open(path) as f:
+                data = json.load(f)
+
+            # Parse timestamp for filtering
+            trace_timestamp = data.get("timestamp")
+            if cutoff_time and trace_timestamp:
+                try:
+                    ts = datetime.fromisoformat(trace_timestamp.replace("Z", "+00:00"))
+                    if ts < cutoff_time:
+                        continue
+                except (ValueError, TypeError):
+                    pass
+
+            # Filter by status
+            trace_status = data.get("status", "")
+            if status and trace_status != status:
+                continue
+
+            # Filter by search text
+            request_preview = data.get("request", {}).get("input_preview", "")
+            if search and search.lower() not in request_preview.lower():
+                continue
+
+            traces.append(
+                {
+                    "trace_id": data.get("trace_id"),
+                    "timestamp": trace_timestamp,
+                    "user": data.get("user"),
+                    "status": trace_status,
+                    "total_duration_ms": data.get("total_duration_ms"),
+                    "span_count": len(data.get("spans", [])),
+                    "request_preview": request_preview[:100],
+                }
+            )
+        except Exception as e:
+            logger.warning("trace_list_parse_error", path=str(path), error=str(e))
+
+    return {"traces": traces, "total": len(traces)}
+
+
+@router.get("/{trace_id}")
+async def get_trace(trace_id: str):
+    """
+    Get a specific trace by ID.
+
+    Returns the full trace JSON.
+    """
+    if not tracing_enabled():
+        raise HTTPException(status_code=404, detail="Tracing not enabled")
+
+    # Sanitize trace_id to prevent path traversal
+    if not trace_id.startswith("trace_") or "/" in trace_id or "\\" in trace_id:
+        raise HTTPException(status_code=400, detail="Invalid trace ID")
+
+    trace_path = TRACES_DIR / f"{trace_id}.json"
+
+    if not trace_path.exists():
+        raise HTTPException(status_code=404, detail="Trace not found")
+
+    try:
+        import json
+
+        with open(trace_path) as f:
+            data = json.load(f)
+        return JSONResponse(content=data)
+    except Exception as e:
+        logger.error("trace_read_error", trace_id=trace_id, error=str(e))
+        raise HTTPException(status_code=500, detail="Failed to read trace") from e

src/main.py

@@ -9,8 +9,9 @@ Main responsibilities:
 - Router registration
 - Lifecycle management
 """
+
+from collections.abc import AsyncGenerator
 from contextlib import asynccontextmanager
-from typing import AsyncGenerator
 
 from fastapi import FastAPI, Request, status
 from fastapi.exceptions import RequestValidationError
@@ -23,6 +24,7 @@ from src.core.exceptions import AppException
 from src.core.logging_config import get_logger
 from src.core.router import router as core_router
 from src.core.startup import initialize_application
+from src.core.tracing_router import router as tracing_router
 from src.models.router import router as models_router
 from src.responses.router import router as responses_router
 
@@ -43,14 +45,16 @@ async def lifespan(app: FastAPI) -> AsyncGenerator[None, None]:
         app_name=config.APP_NAME,
         version=config.APP_VERSION,
         environment=config.ENVIRONMENT.value,
+        prefer_cloud=config.PREFER_CLOUD_BACKEND,
+        anthropic_model=config.ANTHROPIC_MODEL,
         ollama_host=str(config.OLLAMA_HOST),
         ollama_model=config.OLLAMA_DEFAULT_MODEL,
-        redis_url=config.redis_url,
+        redis_url=config.redis_memory_url,
         log_format=config.log_format,
     )
 
-    # Initialize application (register household members, etc.)
-    initialize_application()
+    # Initialize application (check Claude health, register household members, etc.)
+    await initialize_application()
 
     yield
 
@@ -61,7 +65,7 @@ async def lifespan(app: FastAPI) -> AsyncGenerator[None, None]:
 def create_application() -> FastAPI:
     """
     Application factory.
-    
+
     Creates and configures the FastAPI application.
     Following best practice of using factory pattern.
     """
@@ -72,7 +76,7 @@ def create_application() -> FastAPI:
         lifespan=lifespan,
         debug=config.DEBUG,
     )
-    
+
     # Add middleware
     application.add_middleware(
         CORSMiddleware,
@@ -81,26 +85,31 @@ def create_application() -> FastAPI:
         allow_methods=config.CORS_ALLOW_METHODS,
         allow_headers=config.CORS_ALLOW_HEADERS,
     )
-    
+
     # Register exception handlers
     register_exception_handlers(application)
-    
+
     # Include routers
     application.include_router(core_router)  # Health and root endpoints
     application.include_router(chat_router, prefix=config.API_PREFIX)
     application.include_router(models_router, prefix=config.API_PREFIX)
     application.include_router(responses_router, prefix=config.API_PREFIX)  # Responses API
-    
+
+    # Conditionally include tracing router (only in debug mode)
+    if config.DEBUG:
+        application.include_router(tracing_router)
+        logger.info("tracing_router_enabled")
+
     return application
 
 
 def register_exception_handlers(application: FastAPI) -> None:
     """
     Register global exception handlers.
-    
+
     Provides consistent error responses compatible with OpenAI API.
     """
-    
+
     @application.exception_handler(AppException)
     async def app_exception_handler(
         request: Request,