Squash Odysseus development history

docs/AGENT_TURN_CONTRACT.md

@@ -0,0 +1,75 @@
+# Agent turn contract
+
+Scope: product Agent turns on 7011. Environment-owned native/TUI bridges retain
+their existing execution contract. No model weights or training settings change.
+
+## Boundaries
+
+1. `src/turn_contract.py` classifies capabilities, including explicit compound
+   requests and referential follow-ups. Classification is selection, not permission.
+2. `routes/chat_routes.py` resolves toggles, privileges, global/plan/incognito
+   restrictions, fixture restrictions and available schema inventory before
+   freezing the offered set. Web enabled alone does not select web tools.
+3. `TurnContract` checks `required <= offered <= executable`, stores immutable
+   serialized schema copies, and records unavailable requirements. An unavailable
+   request stops without inference or substitution; unknown actions ask for clarity.
+   Exact account-discovery requests narrow selection to account metadata only;
+   compounds retain their declared family scope. Media operations declare their
+   existing tool dependencies rather than falling back to shell generation.
+4. The agent's prompt/schema route and fallback use that same logical scope.
+   Native versus textual serialization remains model-specific. Answer-only phases
+   can suppress tool calls without granting a different scope.
+   Contract turns preserve the already-compacted conversation and tool-call/result
+   IDs. The standalone specialist prompt's latest-message-only behavior is not used
+   for these product turns. Prompt domains also come from the contract.
+   Accepted in-scope calls retain their model-provided arguments and native IDs;
+   the explicit-intent fallback must not overwrite them with the whole user turn.
+5. The context-bound dispatcher checks membership **and** existing runtime policy,
+   owner restrictions and exact-action approvals. A contract is not authorization
+   to bypass those gates. Contract work bypasses terminating legacy shortcuts.
+6. `_AgentRenderState` explicitly identifies streamed versus canonical output.
+   Later synthesis transfers ownership with turn-scoped replacement. The frontend
+   reconciles visible DOM, not just accumulated strings; tool evidence is retained.
+   Ownership is included in saved metrics and `message_saved` events.
+   History and resume honor replacement scope. Single-capability turns retain
+   canonical output: an always-synthesize trial caused a live notes loop and was
+   reverted. Compound turns cannot terminate after only one capability's result.
+
+## Verification
+
+Use the project's configured Python environment, not an unrelated system Python:
+
+```sh
+python -m pytest -q \
+  tests/test_turn_contract.py tests/test_turn_contract_integration.py \
+  tests/test_agent_turn_contract_boundaries.py tests/test_turn_rendering_js.py \
+  tests/test_contract_prompt_conversation.py tests/test_product_turn_contract_route.py \
+  tests/test_contract_explicit_fallback.py \
+  tests/test_history_resume_rendering_js.py \
+  tests/test_chat_route_tool_policy.py tests/test_tool_policy.py \
+  tests/test_frontend_module_version_parity.py
+node scripts/verify_agent_turn_contract.mjs --max-turns 80 --total-ms 900000
+```
+
+The browser verifier uses `sft_alex_creator` and actual 7011 Agent controls. It
+captures request toggles, SSE contract/tool events, visible output and persisted
+history. Ten families have four initial/follow-up Web-toggle combinations.
+Blocked or unrun cases are not passes. Email requires verified fixture isolation;
+do not enable global fixture mode on the user's live service to make a test pass.
+
+## Remaining limits
+
+- Classification is deterministic and vocabulary-based, not a proof of semantic
+  understanding. Add independent behavior examples for confirmed misses.
+- Schema registration and policy permission do not guarantee a remote provider
+  stays healthy throughout a turn. Runtime failure must remain visible.
+- Separate tool/argument errors, tool-service failures, rendering failures and
+  verifier defects in reports. Do not infer model accuracy from routing alone.
+- Canonical summaries can still ignore presentation constraints such as a
+  requested item count. Do not count those as full functional passes. Forcing an
+  extra model round is not a validated general repair for this deployed model.
+- Keep all imports of a local JS module on the same URL identity. Distinct query
+  versions instantiate separate module state even when source files are identical.
+
+Live baseline and current matrix results are in `reports/agent-turn-contract-*`.
+The implementation is not a claim that every family has passed live verification.

docs/BACKGROUND_TOOL_JOBS.md

@@ -0,0 +1,55 @@
+# Background research → originating chat
+
+Chat `trigger_research` calls carry a **dispatcher-supplied** `origin_chat_id`.
+The research start route verifies chat ownership before registering a durable
+`background_tool_jobs` row and starting the existing research service. Panel
+jobs have no origin and never inject a chat reply.
+
+- Chat default: **2 rounds**, 120-second *soft* research budget. Explicit
+  deeper/Auto rounds regain the normal research time budget. Panel defaults
+  remain unchanged. This is not a guaranteed two-minute wall-clock deadline.
+- A completion callback stores the report and sources. A startup worker also
+  reconciles missed callbacks and research errors/restarts.
+- When the origin has no active foreground/detached run, its model summarizes
+  the report with thinking off and no tools. An outer 75-second deadline also
+  bounds model-slot waits. If synthesis is unavailable, deliver an honest
+  notice plus the report link; preserve the evidence for follow-ups.
+- Message and delivery marker commit in one transaction with a deterministic
+  message ID. Report context is stored in server message metadata and injected
+  as untrusted evidence in regular and compact model history. Long excerpts
+  are explicitly marked; the saved full research report remains accessible.
+- The browser polls owner-scoped `/api/research/chat-jobs/{chat_id}`, appending
+  unseen message IDs only when that chat is current and not streaming. No
+  transcript replacement or forced navigation. Reloaded history deduplicates.
+- Chat uses the existing agent-thread rail and expandable rows. The compact
+  header shows status and a right-aligned BG task label with the shared whirlpool
+  while running; expanding reveals topic, phase/round, source count and report
+  link. Rows update in place, preserving expansion/focus while chat streams.
+  Completed rows remain visible; zero-source runs show a warning, not success.
+  Progress polling excludes reports and internal fields.
+
+Other tools are **not automatically backgrounded**. The durable handoff can be
+reused, but each future producer needs explicit launch/result/permission wiring.
+
+## Verification
+
+```sh
+<configured-path> -q tests/test_background_tool_jobs.py tests/test_research_chat_runtime.py
+node --test tests/backgroundToolJobs.test.mjs
+node scripts/verify_background_delivery_isolation.mjs
+node scripts/verify_background_research_cards.mjs
+node scripts/verify_background_research_chat.mjs
+```
+
+The last script uses disposable `sft_alex_creator` chats and real research/model
+calls, then removes only its own reports/chats. Do not use real-user mutations.
+It checks two-round launch, continued chat, automatic arrival, no transcript
+rebuild/duplicates, reload, and a follow-up. Inspect retained report excerpts
+and generated summary when it fails; do not equate job launch with good research.
+
+Initial live runs verified delivery/navigation/follow-ups but exposed a summary
+attempt-count bug (fixed: helper requires **1 attempt**, not `max_retries=0`).
+A later full run was interrupted by an inference endpoint outage. The corrected
+summary path separately passed a real-model evidence/limitations/citation probe.
+All targeted Python tests passed (441); real DOM isolation checks passed. A clean
+full live run with useful retrieved evidence remains to be recorded.

docs/TYPO_ROUTING_AUDIT_20260909.md

@@ -0,0 +1,73 @@
+# Typo-tolerant tool routing audit
+
+The 9B SFT model was not retrained. This audit targets the earlier harness
+stage that decides which complete tool families the model is allowed to see.
+
+## Method
+
+- Source prompts: real `sft_alex_creator` sessions from `a37dcb3b-...` onward.
+- Labels: recorded single-family tool calls, excluding mixed/ambiguous traces.
+- Variants: deletion, adjacent transposition, duplicated character,
+  keyboard-neighbor substitution, and accidental word split.
+- Split: deterministic SHA-256 assignment before scoring (75% dev, 25% blind).
+- Safety: static routing only; no historical mutation or send action is replayed.
+- Acceptance: at least 95% blind exact-family accuracy and below 1% blind
+  wrong-family authorization. Abstention is measured separately.
+
+## Results
+
+| Router | Dev family supplied | Blind family supplied | Blind exact | Blind wrong-family |
+|---|---:|---:|---:|---:|
+| Previous exact rules | 63.64% | 65.69% | — | — |
+| Conservative fuzzy fallback r4 | 96.31% | 98.31% | 96.62% | 0.00% |
+| Final router + safe-read repair | 98.31% | 98.73% | 97.05% | 0.00% |
+
+The fallback runs only for action/lookup-shaped requests, resolves exactly one
+nearby family term, and abstains on ambiguity. Conceptual questions remain
+tool-free. Complete family schemas are still selected by the immutable turn
+contract; fuzzy matching never chooses an individual tool or its arguments.
+
+Authoritative machine reports:
+
+- `reports/typo-tool-routing-baseline-20260909.json`
+- `reports/typo-tool-routing-fuzzy-r4-20260909.json`
+- `reports/typo-tool-routing-final-20260909.json`
+- `reports/post-followup-agent-80-20260909.json`
+- `reports/post-typo-routing-agent-80-20260909.json`
+- `reports/live-typo-agent-20-20260909.json`
+- `reports/live-typo-unresolved-r3-20260909.json`
+- `reports/live-typo-agent-final-20-20260909.json`
+- `reports/post-typo-safe-read-agent-final-80-20260909.json`
+
+## Live 7011 findings
+
+The post-deployment standard matrix passed 80/80 through the real Agent UI.
+The first read-only typo matrix then attempted 17 of 20 planned turns before
+its total-time limit. Initial Notes, Calendar, Email, Tasks, Documents, and
+Cookbook calls passed. Completed failing turns still had the correct family
+and required tool in `turn_contract.offered`; the 9B model sometimes answered
+without calling that offered tool. Memory and Search also exposed timeouts.
+
+This separates three failure classes:
+
+1. **Tool injection:** addressed by conservative fuzzy family routing; blind
+   exact routing is 96.62% with zero blind wrong-family authorizations.
+2. **Required read execution:** a correctly offered safe list/refresh tool can
+   still be skipped by the model, especially after a typo or on “list those
+   again” follow-ups. This should be handled by the generic deterministic
+   safe-read path, not additional prompt-specific hints.
+3. **Runtime timeout:** Search and one Memory follow-up require loop/backend
+   diagnosis. A timeout is not counted as a model-accuracy or routing result.
+
+The generic safe-read parser and search-family precedence were then repaired.
+The previously unresolved Calendar, Email, Search, and Shell/Files cases passed
+8/8. The complete typo matrix passed 20/20, including initial requests and
+follow-ups for all ten families. The final standard Agent UI compatibility
+matrix passed 80/80 across family, Web-toggle, and follow-up combinations.
+
+The broad routing regression suite passed 458 tests. The model was not
+retrained and no DeepSeek API was used: the measured defect was in harness
+family selection and deterministic safe-read execution, upstream of the
+model. All 1,535 unique labeled historical turns were statically audited to
+mine failure categories. Historical write/send/delete actions were not replayed
+against live data; live verification used the deduplicated read-only matrices.

docs/skills-lifecycle.md

@@ -0,0 +1,26 @@
+# Skills lifecycle
+
+The UI exposes All, Built-in, Approved, and Draft. Draft includes archived
+records so they remain inspectable and recoverable. Built-ins are not audited.
+Approved means published, passing, at the configured confidence threshold,
+and not marked unnecessary. Baseline speed measurements remain evidence, not
+an additional hidden UI approval gate.
+
+Automatic audits process at most eight eligible records at a time, oldest first.
+New records are eligible immediately; inconclusive checks retry after a day;
+failed repairs retry after a week. Passed, duplicate-skipped, and archived records
+are excluded. Existing daily Skills Audit tasks drive this queue. Their quiet
+window deferrals propagate to the scheduler rather than becoming task failures.
+Automatic runs use background model scheduling. Existing self-repair and teacher
+repair stages remain in place; failed candidates remain drafts.
+
+The skill index advertises short descriptions; the agent loads a relevant full
+procedure on demand and applies already-injected procedures directly. Extraction
+prefers verified discoveries and specific workarounds over routine tool usage.
+
+Reference reviewed: NousResearch/hermes-agent, MIT license, commit
+cfdbbb6e35010ace89fbe8243ee82fa4de143e10, cloned to
+<configured-path> In particular tools/skills_tool.py and
+agent/prompt_builder.py use progressive disclosure and task-triggered procedure
+loading. These changes adapt that approach to Odysseus's existing registry;
+no Hermes implementation code was copied.