Squash Odysseus development history

.dockerignore

@@ -15,6 +15,10 @@ build/
 # at runtime — never baked into the image. Mirrored in .gitignore.
 secrets.env
 secrets.env.*
+secrets.env~
+.secrets.env.swp
+.secrets.env.swo
+**/#secrets.env#
 !secrets.env.example
 /data/
 /logs/
@@ -26,6 +30,8 @@ secrets.env.*
 .idea/
 dev-docs/
 docs/
+website/
+assets/branding/
 *.md
 *.db
 *.sqlite

.env.example

@@ -67,6 +67,11 @@ SEARXNG_INSTANCE=http://localhost:8080
 # Auth & Security
 # ============================================================
 
+# Optional backend workspace used automatically by the WebUI when no workspace
+# is saved in the browser. This must be a directory visible to the backend;
+# with host-workspace mapping, a host path is translated before vetting.
+# ODYSSEUS_WORKSPACE_DEFAULT=/workspace/project
+
 # Enable authentication (default: true)
 # AUTH_ENABLED=true
 
@@ -76,12 +81,32 @@ SEARXNG_INSTANCE=http://localhost:8080
 # Change this if another local service already uses 7000 (macOS AirPlay often does).
 # APP_PORT=7000
 
+# Optional HTTP address advertised in companion/mobile pairing codes. Set this
+# when Docker would otherwise advertise a container address or loopback. Use a
+# LAN or Tailscale IPv4 address, a single-label hostname, or an mDNS *.local
+# name that the phone can reach. HTTPS and public hostnames are not supported
+# by the current companion client. Do not include credentials, a path, query,
+# or fragment.
+# COMPANION_BASE_URL=http://192.168.1.50:7000
+
 # Development-only auth bypass for loopback requests.
 # Keep false for Docker, LAN, reverse proxy, and any shared deployment.
 # LOCALHOST_BYPASS=false
 
-# Mark session cookies Secure. Set true when Odysseus is served through HTTPS
-# by a trusted reverse proxy or private access gateway.
+# Skip the external-context exact-approval pause for unattended local agents.
+# Keep false for shared or internet-exposed deployments.
+# ODYSSEUS_UNATTENDED_MODE=false
+
+# Optional post-external-context tool approval gate. Off by default because it
+# can block normal agent work; enable only for deployments that want this fence.
+# ODYSSEUS_TOOL_APPROVAL_GATE=0
+
+# Mark session cookies Secure. Left unset, this follows the request scheme:
+# an HTTPS login gets a Secure cookie, a plain-HTTP one does not. Set true to
+# force it on, or false to force it off while you still serve plain HTTP.
+# Upgrading: this used to default to false. Drop a leftover SECURE_COOKIES=false
+# from your .env unless you still need that escape hatch — it keeps HTTPS logins
+# on a non-Secure cookie.
 # SECURE_COOKIES=true
 
 # Optional: pre-seed the first admin password during setup.
@@ -130,6 +155,42 @@ SEARXNG_INSTANCE=http://localhost:8080
 # FASTEMBED_MODEL=sentence-transformers/all-MiniLM-L6-v2
 # FASTEMBED_CACHE_PATH=  # defaults to ~/.cache/fastembed
 
+# ============================================================
+# Google OAuth2 (Google Workspace / .edu email accounts)
+# ============================================================
+# Required to use the "Connect with Google" OAuth flow in email account setup.
+# Create credentials at: console.cloud.google.com → APIs & Services → Credentials
+#   1. Enable the Gmail API for your project.
+#   2. Configure the OAuth consent screen (User Type: Internal for Workspace orgs).
+#      Add scopes: https://mail.google.com/ and email.
+#   3. Create an OAuth 2.0 Client ID (type: Web application).
+#      Add your redirect URI: http://localhost:7000/api/email/oauth/google/callback
+#      (replace host/port for hosted installs).
+#   4. Copy the Client ID and Client Secret below.
+#
+# GOOGLE_OAUTH_CLIENT_ID=your-client-id.apps.googleusercontent.com
+# GOOGLE_OAUTH_CLIENT_SECRET=replace-with-client-secret
+#
+# Set this explicitly for HTTPS, reverse-proxy, or hosted deployments. The
+# value must exactly match an authorized redirect URI in the Google client.
+# Local HTTP setups may use the callback URL inferred by the application.
+# GOOGLE_OAUTH_REDIRECT_URI=https://your-domain.com/api/email/oauth/google/callback
+
+# Origin the MCP OAuth callback is sent back to, for remote (Streamable HTTP)
+# MCP servers that register it dynamically. Defaults to http://localhost:$APP_PORT,
+# which is right only when you reach Odysseus directly on that port. Set it for
+# HTTPS, reverse-proxy, hosted, and Docker installs — inside the container the
+# app always listens on 7000 and cannot see the host port map, so the default is
+# wrong there whenever APP_PORT is not 7000.
+#
+# Not for Google MCP servers. Those use Desktop App credentials, and Google only
+# accepts loopback redirect URIs for that client type, so a public origin here is
+# rejected with redirect_uri_mismatch. Leave it unset for a Google-only install:
+# the loopback default is what Google wants, and remote users finish through the
+# paste-back page, which never has to load the redirect.
+# https://developers.google.com/identity/protocols/oauth2/native-app
+# OAUTH_REDIRECT_BASE_URL=https://your-domain.com
+
 # ============================================================
 # Misc
 # ============================================================
@@ -168,6 +229,58 @@ SEARXNG_INSTANCE=http://localhost:8080
 # ODYSSEUS_EMAIL_COMPOSE_UPLOAD_MAX_BYTES=26214400       # email compose attachment (25 MB)
 # ODYSSEUS_STT_MAX_AUDIO_BYTES=26214400                  # speech-to-text audio (25 MB)
 # ODYSSEUS_ICS_MAX_BYTES=10485760                        # calendar .ics import (10 MB)
+# ODYSSEUS_TTS_CACHE_MAX_BYTES=524288000                 # TTS cache (500 MB)
+
+# ============================================================
+# Host Docker access (explicit opt-in)
+# ============================================================
+# Default Docker Compose does not mount /var/run/docker.sock. Existing
+# Ollama, vLLM, and other OpenAI-compatible endpoints remain usable without it.
+#
+# Enable this only for intentional Cookbook/local Docker-daemon management.
+# Raw socket access is high-trust and can grant broad control over the host
+# Docker daemon. Set DOCKER_GID to the host docker group's numeric GID.
+# Put these values in .env, or export them before running docker compose.
+# COMPOSE_FILE=docker-compose.yml:docker/host-docker.yml
+# DOCKER_GID=963
+# docker/host-docker.yml sets this inside the container. Keep it paired
+# with the socket overlay; setting it alone is not sufficient.
+# ODYSSEUS_ENABLE_HOST_DOCKER=true
+#
+# Host Docker access can be combined with one GPU overlay:
+# COMPOSE_FILE=docker-compose.yml:docker/gpu.nvidia.yml:docker/host-docker.yml
+# COMPOSE_FILE=docker-compose.yml:docker/gpu.amd.yml:docker/host-docker.yml
+
+# ============================================================
+# Host workspace access (explicit opt-in)
+# ============================================================
+# Docker installs normally see only the container filesystem and /app/data.
+# Enable this when the agent should edit a real host workspace like Codex.
+# This is high-trust: the mounted tree is writable by the Odysseus container.
+# COMPOSE_FILE=docker-compose.yml:docker/host-workspace.yml
+# ODYSSEUS_HOST_WORKSPACE_DIR=/home/you
+# ODYSSEUS_HOST_WORKSPACE_MOUNT=/host/workspace
+#
+# Host workspace access can be combined with host Docker access and GPU overlays:
+# COMPOSE_FILE=docker-compose.yml:docker/host-workspace.yml:docker/host-docker.yml
+
+# ============================================================
+# Host network access (explicit opt-in, Linux Docker)
+# ============================================================
+# Docker bridge networking hides some host/LAN/VPN behavior from the agent:
+# mDNS, some LAN discovery, local VPN/Tailscale state, and host namespace
+# assumptions may differ from native Codex. Enable this only for high-trust
+# local installs where the Odysseus container should share the host network.
+#
+# With host networking, Docker port publishing is disabled and the app listens
+# directly on APP_PORT. The bundled SearXNG/Chroma services stay in Docker and
+# are reached through their host-published loopback ports.
+# COMPOSE_FILE=docker-compose.yml:docker/host-workspace.yml:docker/host-network.yml
+# APP_BIND=127.0.0.1
+# APP_PORT=7000
+# ODYSSEUS_HOST_NETWORK_SEARXNG_INSTANCE=http://127.0.0.1:8080
+# ODYSSEUS_HOST_NETWORK_CHROMADB_HOST=127.0.0.1
+# ODYSSEUS_HOST_NETWORK_CHROMADB_PORT=8100
 
 # ============================================================
 # GPU support (Docker Compose)
@@ -197,3 +310,5 @@ SEARXNG_INSTANCE=http://localhost:8080
 
 # APP_DATA_DIR=./data
 # APP_LOGS_DIR=./logs
+# Maximum serialized layered photo-editor draft size (default: 256 MiB).
+ODYSSEUS_EDITOR_DRAFT_MAX_BYTES=268435456

.gitattributes

@@ -15,6 +15,13 @@ docker/entrypoint.sh text eol=lf
 *.cmd text eol=crlf
 *.bat text eol=crlf
 
+# Vendored third-party bundles in static/lib/ are published minified artifacts
+# and must stay byte-identical to what npm ships — stripping trailing whitespace
+# to satisfy `git diff --check` would desync them from the upstream release. Turn
+# the whitespace check off for that tree instead, and keep the bundles out of
+# GitHub's language statistics.
+static/lib/** -whitespace linguist-vendored
+
 # Binary assets — never normalize.
 *.png  binary
 *.jpg  binary

.github/CODEOWNERS

@@ -6,4 +6,4 @@
 # A per-area ownership map (security/auth, CI, frontend, agent internals, with
 # multiple named owners per line) is being worked out in issue #593; once
 # agreed it replaces this file. Until then, required reviews and the security
-# CI gate (docs/security-ci.md) remain in force via branch protection.
+# CI gate (website/security-ci.md) remain in force via branch protection.

.github/ISSUE_TEMPLATE/bug_report.yml

@@ -6,26 +6,38 @@ body:
   - type: markdown
     attributes:
       value: |
-        **Before submitting:** search [open issues](https://github.com/pewdiepie-archdaemon/odysseus/issues)
-        and [discussions](https://github.com/pewdiepie-archdaemon/odysseus/discussions) first.
+        **Before submitting:** search [open issues](https://github.com/odysseus-dev/odysseus/issues)
+        and [discussions](https://github.com/odysseus-dev/odysseus/discussions) first.
         Duplicate reports slow things down.
 
         For security vulnerabilities, **do not open a public issue** —
-        use [GitHub Security Advisories](https://github.com/pewdiepie-archdaemon/odysseus/security/advisories/new)
-        and read [SECURITY.md](https://github.com/pewdiepie-archdaemon/odysseus/blob/main/SECURITY.md) first.
+        use [GitHub Security Advisories](https://github.com/odysseus-dev/odysseus/security/advisories/new)
+        and read [SECURITY.md](https://github.com/odysseus-dev/odysseus/blob/main/SECURITY.md) first.
 
   - type: checkboxes
     id: prerequisites
     attributes:
       label: Prerequisites
       options:
-        - label: I searched [open issues](https://github.com/pewdiepie-archdaemon/odysseus/issues?q=is%3Aissue+is%3Aopen) and [discussions](https://github.com/pewdiepie-archdaemon/odysseus/discussions) and did not find an existing report of this bug.
+        - label: I searched [open issues](https://github.com/odysseus-dev/odysseus/issues?q=is%3Aissue+is%3Aopen) and [discussions](https://github.com/odysseus-dev/odysseus/discussions) and did not find an existing report of this bug.
           required: true
-        - label: This is **not** a security vulnerability. (Vulnerabilities go to [GitHub Security Advisories](https://github.com/pewdiepie-archdaemon/odysseus/security/advisories/new) — see [SECURITY.md](https://github.com/pewdiepie-archdaemon/odysseus/blob/main/SECURITY.md).)
+        - label: This is **not** a security vulnerability. (Vulnerabilities go to [GitHub Security Advisories](https://github.com/odysseus-dev/odysseus/security/advisories/new) — see [SECURITY.md](https://github.com/odysseus-dev/odysseus/blob/main/SECURITY.md).)
           required: true
         - label: I am running the latest code from the `dev` branch (the default branch you get on clone, where fixes land first) and the bug still reproduces there. Please `git pull` the latest `dev` before filing.
           required: true
 
+  - type: input
+    id: revision
+    attributes:
+      label: Odysseus Revision
+      description: |
+        From the repository root (on the host when using Docker), run
+        `git show -s --abbrev=12 --format='%h (%cs)' HEAD`
+        and paste the output exactly.
+      placeholder: "1fef4929cf1d (2026-08-11)"
+    validations:
+      required: true
+
   - type: dropdown
     id: install-method
     attributes:

.github/ISSUE_TEMPLATE/config.yml

@@ -1,13 +1,13 @@
 blank_issues_enabled: false
 contact_links:
   - name: Question / Need Help
-    url: https://github.com/pewdiepie-archdaemon/odysseus/discussions/categories/q-a
+    url: https://github.com/odysseus-dev/odysseus/discussions/categories/q-a
     about: Ask how-to questions, setup help, and model configuration questions here. Issues are for confirmed bugs and concrete proposals only.
 
   - name: Idea or Suggestion
-    url: https://github.com/pewdiepie-archdaemon/odysseus/discussions/categories/ideas
+    url: https://github.com/odysseus-dev/odysseus/discussions/categories/ideas
     about: Discuss ideas and gauge interest before opening a formal feature request. If there is already a discussion, link it in your feature request.
 
   - name: Security Vulnerability
-    url: https://github.com/pewdiepie-archdaemon/odysseus/security/advisories/new
+    url: https://github.com/odysseus-dev/odysseus/security/advisories/new
     about: Report vulnerabilities privately via GitHub Security Advisories — never as a public issue. Read SECURITY.md before reporting.

.github/ISSUE_TEMPLATE/feature_request.yml

@@ -6,22 +6,22 @@ body:
   - type: markdown
     attributes:
       value: |
-        **Before submitting:** search [open issues](https://github.com/pewdiepie-archdaemon/odysseus/issues)
-        and [discussions](https://github.com/pewdiepie-archdaemon/odysseus/discussions) first.
-        Feature requests that duplicate [ROADMAP.md](https://github.com/pewdiepie-archdaemon/odysseus/blob/main/ROADMAP.md)
+        **Before submitting:** search [open issues](https://github.com/odysseus-dev/odysseus/issues)
+        and [discussions](https://github.com/odysseus-dev/odysseus/discussions) first.
+        Feature requests that duplicate [ROADMAP.md](https://github.com/odysseus-dev/odysseus/blob/main/ROADMAP.md)
         or an existing open issue will be closed as duplicates.
 
         If your idea needs community input before it becomes a concrete proposal,
-        start a [discussion](https://github.com/pewdiepie-archdaemon/odysseus/discussions/categories/ideas) instead.
+        start a [discussion](https://github.com/odysseus-dev/odysseus/discussions/categories/ideas) instead.
 
   - type: checkboxes
     id: prerequisites
     attributes:
       label: Prerequisites
       options:
-        - label: I searched [open issues](https://github.com/pewdiepie-archdaemon/odysseus/issues?q=is%3Aissue+is%3Aopen) and this has not already been proposed.
+        - label: I searched [open issues](https://github.com/odysseus-dev/odysseus/issues?q=is%3Aissue+is%3Aopen) and this has not already been proposed.
           required: true
-        - label: I searched [discussions](https://github.com/pewdiepie-archdaemon/odysseus/discussions) and this is not already being debated there.
+        - label: I searched [discussions](https://github.com/odysseus-dev/odysseus/discussions) and this is not already being debated there.
           required: true
         - label: This is a concrete, actionable proposal — not a vague "it would be nice if..." request.
           required: true

.github/pull_request_template.md

@@ -24,10 +24,11 @@ Fixes #
 
 ## Checklist
 
-- [ ] I searched [open issues](https://github.com/pewdiepie-archdaemon/odysseus/issues) and [open PRs](https://github.com/pewdiepie-archdaemon/odysseus/pulls) — this is not a duplicate.
+- [ ] I searched [open issues](https://github.com/odysseus-dev/odysseus/issues) and [open PRs](https://github.com/odysseus-dev/odysseus/pulls) — this is not a duplicate.
 - [ ] This PR targets `dev`
 - [ ] My changes are limited to the scope described above — no unrelated refactors or whitespace changes mixed in.
 - [ ] I actually ran the app (`docker compose up` or `uvicorn app:app`) and verified the change works end-to-end. Type-checks and unit tests are not enough.
+- [ ] I did not run the app/runtime validation and stated that gap in **How to Test**. Leave this unchecked when the app-run box above is checked.
 
 ## How to Test
 

.github/scripts/check-issue-description.js

@@ -41,6 +41,14 @@ module.exports = async ({ github, context, core }) => {
       break;
 
     case 'bug': {
+      const revisionText = section('Odysseus Revision');
+      if (!/^[0-9a-f]{12} \(\d{4}-\d{2}-\d{2}\)$/i.test(revisionText)) {
+        failures.push(
+          '**Odysseus Revision** — paste the 12-character commit SHA and date, ' +
+          'for example `1fef4929cf1d (2026-08-11)`',
+        );
+      }
+
       if (!section('Install Method')) {
         failures.push('**Install Method** — select how you installed Odysseus');
       }
@@ -153,6 +161,16 @@ module.exports = async ({ github, context, core }) => {
     }
   }
 
+  const LABEL_BAD  = 'needs more info';
+  const LABEL_GOOD = 'ready for review';
+
+  // Closed issues are no longer awaiting review.
+  // This also prevents later edits to closed issues from restoring the label.
+  if (issue.state === 'closed') {
+    await dropLabel(LABEL_GOOD);
+    return;
+  }
+
   // ── Find existing bot comment to update in-place ──────────────────────────
   const MARKER = '<!-- issue-description-check -->';
   const { data: comments } = await github.rest.issues.listComments({
@@ -160,9 +178,6 @@ module.exports = async ({ github, context, core }) => {
   });
   const existing = comments.find(c => c.user.type === 'Bot' && c.body.includes(MARKER));
 
-  const LABEL_BAD  = 'needs more info';
-  const LABEL_GOOD = 'ready for review';
-
   if (failures.length === 0) {
     if (existing) {
       await github.rest.issues.deleteComment({ owner, repo, comment_id: existing.id });

.github/scripts/check-pr-description.js

@@ -21,11 +21,11 @@ module.exports = async ({ github, context, core }) => {
     return strip(m?.[0].replace(new RegExp(`#+\\s+${heading}`, 'i'), '') ?? '');
   }
 
-  const problems = [];
+  const descriptionProblems = [];
 
   // 1. Summary must be filled in.
   if (section('Summary').length < 20) {
-    problems.push('**Summary** is empty or too short — describe what changed and why.');
+    descriptionProblems.push('**Summary** is empty or too short — describe what changed and why.');
   }
 
   // 2. Linked Issue must reference a real issue. Accept a bare #NNN, a closing
@@ -34,18 +34,18 @@ module.exports = async ({ github, context, core }) => {
   const linkedSection = section('Linked Issue');
   const hasIssueRef = /#\d+\b/.test(linkedSection) || /\/issues\/\d+/.test(linkedSection);
   if (!linkedSection || !hasIssueRef) {
-    problems.push('**Linked Issue** — add a reference like `Fixes #NNN`, a bare `#NNN`, or a link to the issue.');
+    descriptionProblems.push('**Linked Issue** — add a reference like `Fixes #NNN`, a bare `#NNN`, or a link to the issue.');
   }
 
   // 3. At least one Type of Change box must be checked.
   const typeBlock = body.match(/##\s+Type of Change[\s\S]*?(?=\n##\s|$)/i)?.[0] ?? '';
   if (!/- \[x\]/i.test(typeBlock)) {
-    problems.push('**Type of Change** — check at least one box.');
+    descriptionProblems.push('**Type of Change** — check at least one box.');
   }
 
   // 4. Duplicate-search checklist item must be checked.
   if (!/- \[x\] I searched/i.test(body)) {
-    problems.push('**Checklist** — check the duplicate-search box to confirm you searched existing issues and PRs.');
+    descriptionProblems.push('**Checklist** — check the duplicate-search box to confirm you searched existing issues and PRs.');
   }
 
   // 5. How to Test must contain enough real detail for a reviewer to act on.
@@ -53,7 +53,83 @@ module.exports = async ({ github, context, core }) => {
   //    code block — so we only require non-trivial content, not a specific shape.
   const howTo = section('How to Test');
   if (howTo.length < 30) {
-    problems.push('**How to Test** — explain how a reviewer can verify this change. Numbered steps, the commands you ran, or a short code block all work — give a sentence or two of real detail (not just "tested locally").');
+    descriptionProblems.push('**How to Test** — explain how a reviewer can verify this change. Numbered steps, the commands you ran, or a short code block all work — give a sentence or two of real detail (not just "tested locally").');
+  }
+
+  // Classify paths from GitHub's API. This workflow runs in the privileged base
+  // context, so it must never check out or execute code from the PR branch.
+  const changedFiles = await github.paginate(github.rest.pulls.listFiles, {
+    owner, repo, pull_number: prNum, per_page: 100,
+  });
+  const changedPaths = changedFiles.map(file => file.filename);
+
+  function isUiSensitivePath(filename) {
+    const path = filename.toLowerCase();
+    return path.startsWith('static/')
+      || path.startsWith('templates/')
+      || /\.(?:html?|css|svg)$/.test(path);
+  }
+
+  function isDocsOnlyPath(filename) {
+    const path = filename.toLowerCase();
+    return /\.(?:md|mdx|rst|adoc|txt)$/.test(path)
+      || (path.startsWith('docs/') && !isUiSensitivePath(path));
+  }
+
+  function isRuntimeSensitivePath(filename) {
+    const path = filename.toLowerCase();
+    if (isUiSensitivePath(path)) return false;
+    if (path.startsWith('tests/') || path.startsWith('.github/')) return false;
+    return /^(?:app\.py|routes\/|services\/|src\/|core\/|mcp_servers\/|scripts\/|docker\/)/.test(path)
+      || /^(?:dockerfile|docker-compose.*\.ya?ml|requirements(?:-optional)?\.txt|pyproject\.toml|setup\.py)$/.test(path)
+      || /\.(?:py|sh|ps1|bat)$/.test(path);
+  }
+
+  let classification = 'tooling';
+  if (changedPaths.some(isUiSensitivePath)) {
+    classification = 'UI-sensitive';
+  } else if (changedPaths.some(isRuntimeSensitivePath)) {
+    classification = 'backend/runtime';
+  } else if (changedPaths.length > 0 && changedPaths.every(isDocsOnlyPath)) {
+    classification = 'docs-only';
+  }
+
+  const appRan = /- \[x\]\s+I actually ran the app\b/i.test(body);
+  const appNotRun = /- \[x\]\s+I did not run the app\/runtime validation\b/i.test(body);
+  // Anchor on the wording, not the template's emphasis: a ticked box the author
+  // retyped without the surrounding ** renders identically on the PR page, so
+  // treating it as unchecked is invisible from their side. Matches the two
+  // attestations above, which already ignore formatting.
+  const screenshotChecked = /- \[x\]\s+[*_]{0,2}Screenshot or short clip[*_]{0,2}/i.test(body);
+  const screenshotSection = section('Screenshots / clips');
+  const hasVisualEvidence = /!\[[^\]]*\]\([^)]+\)|<(?:img|video|source)\b[^>]*(?:src|href)=|https?:\/\/[^\s)]+/i.test(screenshotSection);
+  const evidenceGaps = [];
+  let needsRuntimeValidation = false;
+  let needsVisualEvidence = false;
+
+  if (classification === 'backend/runtime' || classification === 'UI-sensitive') {
+    if (appRan && appNotRun) {
+      needsRuntimeValidation = true;
+      evidenceGaps.push('The app-run and explicit not-run boxes are both checked. Select the one state that is true.');
+    } else if (!appRan) {
+      needsRuntimeValidation = true;
+      if (appNotRun) {
+        evidenceGaps.push('The author explicitly reports that app/runtime validation was not performed.');
+      } else {
+        evidenceGaps.push('App/runtime validation is not author-attested. Check the run box only after running it, or check the explicit not-run box and describe the gap.');
+      }
+    }
+  }
+
+  if (classification === 'UI-sensitive') {
+    if (!screenshotChecked) {
+      needsVisualEvidence = true;
+      evidenceGaps.push('The screenshot/clip checkbox is not checked for this UI-sensitive change.');
+    }
+    if (!hasVisualEvidence) {
+      needsVisualEvidence = true;
+      evidenceGaps.push('The Screenshots / clips section does not contain an actual attachment or link.');
+    }
   }
 
   // ── Comment ──────────────────────────────────────────────────────────────
@@ -62,22 +138,43 @@ module.exports = async ({ github, context, core }) => {
   });
   const existing = comments.find(c => (c.body ?? '').includes(MARKER));
 
-  if (problems.length === 0) {
+  if (descriptionProblems.length === 0 && evidenceGaps.length === 0) {
     if (existing) {
       await github.rest.issues.deleteComment({ owner, repo, comment_id: existing.id });
     }
   } else {
-    const commentBody = [
-      MARKER,
-      '⚠️ **PR description — action needed**',
-      '',
-      'The following required sections are missing or incomplete. Please update the PR description to address them:',
-      '',
-      problems.map(p => `- ${p}`).join('\n'),
+    const commentLines = [MARKER];
+    if (descriptionProblems.length > 0) {
+      commentLines.push(
+        '⚠️ **PR description — action needed**',
+        '',
+        'The following required sections are missing or incomplete. Please update the PR description to address them:',
+        '',
+        descriptionProblems.map(problem => `- ${problem}`).join('\n'),
+      );
+    } else {
+      commentLines.push(
+        '⚠️ **PR description is complete; validation evidence is still outstanding**',
+        '',
+        `Changed-file classification: **${classification}**.`,
+      );
+    }
+    if (evidenceGaps.length > 0) {
+      commentLines.push(
+        '',
+        '**Author-reported runtime / visual state**',
+        '',
+        evidenceGaps.map(gap => `- ${gap}`).join('\n'),
+        '',
+        'Checkboxes are author attestations. GitHub Actions results remain the execution evidence for CI; this check does not prove that a local command ran.',
+      );
+    }
+    commentLines.push(
       '',
       '---',
-      '_This comment is deleted automatically once all sections are complete._',
-    ].join('\n');
+      '_This comment updates automatically when the description or changed files change._',
+    );
+    const commentBody = commentLines.join('\n');
 
     if (existing) {
       await github.rest.issues.updateComment({ owner, repo, comment_id: existing.id, body: commentBody });
@@ -97,34 +194,47 @@ module.exports = async ({ github, context, core }) => {
       return true;
     } catch (e) {
       if (e.status === 404) return false;
+      if (e.status === 403) {
+        core.warning(`Could not inspect label "${name}" — token lacks label read access; skipping.`);
+        return false;
+      }
       throw e;
     }
   }
 
-  async function swapLabel(num, add, remove) {
-    if (await labelExists(add)) {
+  async function setLabel(name, wanted) {
+    if (wanted && await labelExists(name)) {
       try {
-        await github.rest.issues.addLabels({ owner, repo, issue_number: num, labels: [add] });
+        await github.rest.issues.addLabels({ owner, repo, issue_number: prNum, labels: [name] });
       } catch (e) {
         // Fail soft on a token that can't write labels so a label permission
         // problem never masks the actual description verdict.
-        if (e.status !== 403) throw e;
-        core.warning(`Could not add "${add}" — token lacks label write here; skipping.`);
+        if (e.status !== 403 && e.status !== 404) throw e;
+        core.warning(`Could not add "${name}" — label is unavailable or the token lacks label write access; skipping.`);
       }
+    } else if (wanted) {
+      core.warning(`Label "${name}" does not exist in the repo — skipping. Create it once to enable labelling.`);
     } else {
-      core.warning(`Label "${add}" does not exist in the repo — skipping. Create it once to enable labelling.`);
-    }
-    try {
-      await github.rest.issues.removeLabel({ owner, repo, issue_number: num, name: remove });
-    } catch (e) {
-      if (e.status !== 404 && e.status !== 410 && e.status !== 403) throw e;
+      try {
+        await github.rest.issues.removeLabel({ owner, repo, issue_number: prNum, name });
+      } catch (e) {
+        if (e.status !== 404 && e.status !== 410 && e.status !== 403) throw e;
+      }
     }
   }
 
-  if (problems.length === 0) {
-    await swapLabel(prNum, 'ready for review', 'needs work');
-  } else {
-    await swapLabel(prNum, 'needs work', 'ready for review');
-    core.setFailed(`PR description has ${problems.length} issue(s) — see bot comment for details.`);
+  const descriptionComplete = descriptionProblems.length === 0;
+  const evidenceComplete = evidenceGaps.length === 0;
+  const isDraft = Boolean(context.payload.pull_request.draft);
+  await setLabel(
+    'ready for review',
+    descriptionComplete && evidenceComplete && !isDraft,
+  );
+  await setLabel('needs work', !descriptionComplete);
+  await setLabel('needs runtime validation', needsRuntimeValidation);
+  await setLabel('needs visual evidence', needsVisualEvidence);
+
+  if (!descriptionComplete) {
+    core.setFailed(`PR description has ${descriptionProblems.length} issue(s) — see bot comment for details.`);
   }
 };

.github/scripts/focused_test_guidance.py

@@ -0,0 +1,124 @@
+#!/usr/bin/env python3
+"""Report focused pytest guidance for changed paths under tests/."""
+
+from __future__ import annotations
+
+import argparse
+import os
+import shlex
+import subprocess
+import sys
+from collections.abc import Iterable
+from pathlib import PurePosixPath
+
+
+def parse_paths(raw_paths: bytes) -> list[str]:
+    """Decode the NUL-delimited output of ``git diff --name-only -z``."""
+    return [os.fsdecode(path) for path in raw_paths.split(b"\0") if path]
+
+
+def changed_paths_from_merge_base(base_sha: str, head_sha: str) -> list[str]:
+    """Return changed ``tests/`` paths using GitHub PR three-dot semantics.
+
+    GitHub PR changed files are based on the merge base and the PR head, not a
+    direct endpoint diff between the current base branch tip and the PR head.
+    Using the direct endpoint diff can include files changed only on the base
+    branch when the PR branch is stale.
+    """
+    merge_base = subprocess.check_output(
+        ["git", "merge-base", base_sha, head_sha],
+        stderr=subprocess.DEVNULL,
+    ).strip()
+    raw_paths = subprocess.check_output(
+        [
+            "git",
+            "diff",
+            "--name-only",
+            "--diff-filter=ACMRT",
+            "-z",
+            os.fsdecode(merge_base),
+            head_sha,
+            "--",
+            "tests/",
+        ],
+    )
+    return parse_paths(raw_paths)
+
+
+def select_test_paths(paths: Iterable[str]) -> list[str]:
+    """Return unique, repository-relative paths contained by tests/."""
+    selected: set[str] = set()
+    for raw_path in paths:
+        path = PurePosixPath(raw_path)
+        if path.is_absolute() or ".." in path.parts:
+            continue
+        parts = tuple(part for part in path.parts if part != ".")
+        if len(parts) >= 2 and parts[0] == "tests":
+            selected.add(PurePosixPath(*parts).as_posix())
+    return sorted(selected)
+
+
+def is_pytest_file(path: str) -> bool:
+    """Return whether a changed path follows this repository's pytest naming."""
+    name = PurePosixPath(path).name
+    return name.endswith(".py") and (
+        name.startswith("test_") or name.endswith("_test.py")
+    )
+
+
+def pytest_command(paths: Iterable[str]) -> str:
+    """Build a copyable pytest command for changed runnable test files."""
+    command = ["python3", "-m", "pytest", "-q", *paths]
+    return shlex.join(command)
+
+
+def format_report(paths: Iterable[str]) -> str:
+    """Format focused guidance for CI logs and the workflow summary."""
+    changed_paths = select_test_paths(paths)
+    runnable_paths = [path for path in changed_paths if is_pytest_file(path)]
+    lines = ["## Focused test guidance (report-only)", ""]
+    if not changed_paths:
+        lines.append("No changed paths under `tests/`.")
+    else:
+        lines.extend(["Changed paths under `tests/`:", ""])
+        lines.extend(f"- `{path}`" for path in changed_paths)
+    lines.extend(["", "Suggested focused validation:", ""])
+    if runnable_paths:
+        lines.append(f"```sh\n{pytest_command(runnable_paths)}\n```")
+    else:
+        lines.append("No directly runnable pytest files changed.")
+    lines.extend(
+        [
+            "",
+            "This guidance does not infer tests from source changes. "
+            "Existing blocking CI remains the source of truth.",
+        ]
+    )
+    return "\n".join(lines)
+
+
+def _parse_args(argv: list[str]) -> argparse.Namespace:
+    parser = argparse.ArgumentParser(
+        description="Report focused pytest guidance for changed tests/ paths.",
+    )
+    parser.add_argument("--base-sha", help="Pull request base commit SHA.")
+    parser.add_argument("--head-sha", help="Pull request head commit SHA.")
+    return parser.parse_args(argv)
+
+
+def main(argv: list[str] | None = None) -> int:
+    args = _parse_args(sys.argv[1:] if argv is None else argv)
+    if bool(args.base_sha) != bool(args.head_sha):
+        raise SystemExit("--base-sha and --head-sha must be provided together")
+
+    if args.base_sha and args.head_sha:
+        paths = changed_paths_from_merge_base(args.base_sha, args.head_sha)
+    else:
+        paths = parse_paths(sys.stdin.buffer.read())
+
+    print(format_report(paths))
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

.github/workflows/ci.yml

@@ -2,7 +2,7 @@ name: CI
 
 on:
   push:
-    branches: [main]
+    branches: [main, dev]
   pull_request:
 
 # Least privilege: none of the jobs write to the repo.
@@ -15,14 +15,68 @@ concurrency:
   cancel-in-progress: true
 
 jobs:
+  focused-test-guidance:
+    name: Focused test guidance (report-only)
+    if: github.event_name == 'pull_request'
+    runs-on: ubuntu-latest
+    continue-on-error: true
+    steps:
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
+        with:
+          fetch-depth: 0
+          persist-credentials: false
+      - name: Report changed test paths
+        env:
+          BASE_SHA: ${{ github.event.pull_request.base.sha }}
+          HEAD_SHA: ${{ github.event.pull_request.head.sha }}
+        run: |
+          report_file="$RUNNER_TEMP/focused-test-guidance.md"
+          publish_report() {
+            cat "$report_file"
+            if [ -n "${GITHUB_STEP_SUMMARY:-}" ]; then
+              cat "$report_file" >> "$GITHUB_STEP_SUMMARY" || true
+            fi
+            return 0
+          }
+
+          report_unavailable() {
+            {
+              printf '%s\n\n' '## Focused test guidance unavailable (report-only)'
+              printf '%s\n\n' "$1"
+              printf '%s\n' 'Existing blocking CI remains the source of truth.'
+            } > "$report_file"
+            publish_report
+            exit 0
+          }
+
+          if [ -z "$BASE_SHA" ] || [ -z "$HEAD_SHA" ]; then
+            report_unavailable "Pull request base/head metadata is missing."
+          fi
+
+          if ! git cat-file -e "${BASE_SHA}^{commit}" 2>/dev/null; then
+            report_unavailable "The pull request base commit is unavailable locally."
+          fi
+
+          if ! git cat-file -e "${HEAD_SHA}^{commit}" 2>/dev/null; then
+            report_unavailable "The pull request head commit is unavailable locally."
+          fi
+
+          if ! python3 .github/scripts/focused_test_guidance.py \
+            --base-sha "$BASE_SHA" \
+            --head-sha "$HEAD_SHA" > "$report_file"; then
+            report_unavailable "The focused test guidance helper could not produce a report."
+          fi
+
+          publish_report
+
   python-syntax:
     name: Python syntax (compileall)
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10  # v6.0.3
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
         with:
           persist-credentials: false
-      - uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405  # v6.2.0
+      - uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97  # v7.0.0
         with:
           python-version: "3.11"
       # Byte-compile sources — catches syntax errors without installing deps.
@@ -32,10 +86,10 @@ jobs:
     name: JS syntax (node --check)
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10  # v6.0.3
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
         with:
           persist-credentials: false
-      - uses: actions/setup-node@48b55a011bda9f5d6aeb4c2d9c7362e8dae4041e  # v6.4.0
+      - uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020  # v7.0.0
         with:
           node-version: "20"
       # Syntax-check our own JS (skip vendored libs in static/lib).
@@ -49,17 +103,14 @@ jobs:
   python-tests:
     name: Python tests (pytest)
     runs-on: ubuntu-latest
-    # Informational for now: the suite has known flaky / environment-dependent
-    # failures (test isolation + embedding-model assertions). Tracked under the
-    # ROADMAP "fresh install smoke tests" item; make this required once green.
-    continue-on-error: true
+    # Make Python test validation authoritative for the configured scope.
     steps:
-      - uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10  # v6.0.3
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
         with:
           fetch-depth: 0
           persist-credentials: false
 
-      # Detect whether this PR only touches documentation files.
+      # Detect whether this PR only touches repository prose outside the Pages site.
       # If so, skip the expensive pytest run while still reporting a passing check.
       - name: Check for docs-only changes
         id: docs-check
@@ -71,9 +122,10 @@ jobs:
             BASE="${{ github.event.before }}"
             HEAD="${{ github.sha }}"
           fi
-          # List all changed files; if every file matches docs/markdown patterns, skip pytest.
+          # Keep website/ and assets/branding/ out of this bypass: pytest owns
+          # regression guards for their published-file and orphan-asset contracts.
           changed=$(git diff --name-only "$BASE" "$HEAD" 2>/dev/null || git diff --name-only HEAD~1 HEAD)
-          non_docs=$(echo "$changed" | grep -Ev '^(docs/|.*\.md$|\.github/[^/]+\.md$)' || true)
+          non_docs=$(echo "$changed" | grep -Ev '^(docs/|[^/]+\.md$|\.github/[^/]+\.md$)' || true)
           if [ -z "$non_docs" ]; then
             echo "docs_only=true" >> "$GITHUB_OUTPUT"
             echo "Docs-only change detected — skipping pytest."
@@ -81,7 +133,7 @@ jobs:
             echo "docs_only=false" >> "$GITHUB_OUTPUT"
           fi
 
-      - uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405  # v6.2.0
+      - uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97  # v7.0.0
         if: steps.docs-check.outputs.docs_only != 'true'
         with:
           python-version: "3.11"

.github/workflows/codeql.yml

@@ -0,0 +1,41 @@
+name: CodeQL
+
+# Advanced setup so CodeQL also runs on pull requests (including from forks),
+# surfacing findings before merge instead of only after a change lands on dev.
+on:
+  push:
+    branches: [dev, main]
+  pull_request:
+    branches: [dev]
+  schedule:
+    - cron: "17 3 * * 1"
+
+permissions:
+  contents: read
+
+jobs:
+  analyze:
+    name: Analyze (${{ matrix.language }})
+    runs-on: ubuntu-latest
+    permissions:
+      security-events: write
+      actions: read
+      contents: read
+    strategy:
+      fail-fast: false
+      matrix:
+        language: [actions, javascript-typescript, python]
+    steps:
+      - name: Checkout
+        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
+        with:
+          persist-credentials: false
+      - name: Initialize CodeQL
+        uses: github/codeql-action/init@ff2f1c621b7f889edc0d3c761ac2e6a3f8cdb0dd  # v4.37.7
+        with:
+          languages: ${{ matrix.language }}
+          build-mode: none
+      - name: Perform CodeQL Analysis
+        uses: github/codeql-action/analyze@ff2f1c621b7f889edc0d3c761ac2e6a3f8cdb0dd  # v4.37.7
+        with:
+          category: "/language:${{ matrix.language }}"

.github/workflows/container-scan.yml

@@ -37,12 +37,12 @@ jobs:
       contents: read
     steps:
       - name: Checkout repository
-        uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10  # v6.0.3
+        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
         with:
           persist-credentials: false
 
       - name: Lint Dockerfile
-        uses: hadolint/hadolint-action@2332a7b74a6de0dda2e2221d575162eba76ba5e5  # v3.3.0
+        uses: hadolint/hadolint-action@2a66e89f53d0771bb131a7fa31f3136336094aa6  # v3.4.0
         with:
           dockerfile: Dockerfile
           # DL3008: pinning apt package versions is impractical on a -slim base

.github/workflows/container-trivy.yml

@@ -23,12 +23,16 @@ on:
     paths-ignore:
       - '**.md'
       - 'docs/**'
+      - 'website/**'
+      - 'assets/branding/**'
       - '.github/ISSUE_TEMPLATE/**'
   push:
     branches: [main]
     paths-ignore:
       - '**.md'
       - 'docs/**'
+      - 'website/**'
+      - 'assets/branding/**'
       - '.github/ISSUE_TEMPLATE/**'
   workflow_dispatch:
 
@@ -52,17 +56,17 @@ jobs:
       contents: read
     steps:
       - name: Checkout repository
-        uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10  # v6.0.3
+        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
         with:
           persist-credentials: false
 
       - name: Set up Buildx
-        uses: docker/setup-buildx-action@d7f5e7f509e45cec5c76c4d5afdd7de93d0b3df5  # v4.1.0
+        uses: docker/setup-buildx-action@37fe631027851001ddb9b187196cc803df7f5f0e  # v4.3.0
 
       # Build without pushing so a broken Dockerfile is caught here, and the
       # exact image we ship is what gets scanned.
       - name: Build image
-        uses: docker/build-push-action@f9f3042f7e2789586610d6e8b85c8f03e5195baf  # v7.2.0
+        uses: docker/build-push-action@53b7df96c91f9c12dcc8a07bcb9ccacbed38856a  # v7.3.0
         with:
           context: .
           push: false
@@ -93,15 +97,15 @@ jobs:
       security-events: write  # upload SARIF to the Security tab
     steps:
       - name: Checkout repository
-        uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10  # v6.0.3
+        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
         with:
           persist-credentials: false
 
       - name: Set up Buildx
-        uses: docker/setup-buildx-action@d7f5e7f509e45cec5c76c4d5afdd7de93d0b3df5  # v4.1.0
+        uses: docker/setup-buildx-action@37fe631027851001ddb9b187196cc803df7f5f0e  # v4.3.0
 
       - name: Build image
-        uses: docker/build-push-action@f9f3042f7e2789586610d6e8b85c8f03e5195baf  # v7.2.0
+        uses: docker/build-push-action@53b7df96c91f9c12dcc8a07bcb9ccacbed38856a  # v7.3.0
         with:
           context: .
           push: false
@@ -119,7 +123,7 @@ jobs:
           TRIVY_DB_REPOSITORY: ghcr.io/aquasecurity/trivy-db:2
 
       - name: Upload Trivy results
-        uses: github/codeql-action/upload-sarif@8aad20d150bbac5944a9f9d289da16a4b0d87c1e  # v4.36.2
+        uses: github/codeql-action/upload-sarif@ff2f1c621b7f889edc0d3c761ac2e6a3f8cdb0dd  # v4.37.7
         with:
           sarif_file: trivy-results.sarif
           category: trivy-image

.github/workflows/dependency-review.yml

@@ -36,7 +36,7 @@ jobs:
       contents: read
     steps:
       - name: Checkout repository
-        uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10  # v6.0.3
+        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
         with:
           persist-credentials: false
 
@@ -55,12 +55,12 @@ jobs:
       contents: read
     steps:
       - name: Checkout repository
-        uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10  # v6.0.3
+        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
         with:
           persist-credentials: false
 
       - name: Set up Python
-        uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405  # v6.2.0
+        uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97  # v7.0.0
         with:
           python-version: '3.12'
 

.github/workflows/deploy-pages.yml

@@ -0,0 +1,50 @@
+name: Deploy GitHub Pages
+
+on:
+  push:
+    branches: [main]
+    paths:
+      - 'website/**'
+      - '.github/workflows/deploy-pages.yml'
+  workflow_dispatch:
+
+permissions: {}
+
+concurrency:
+  group: pages
+  cancel-in-progress: false
+
+jobs:
+  build:
+    name: Package static site
+    runs-on: ubuntu-latest
+    permissions:
+      contents: read
+      pages: read
+    steps:
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
+        with:
+          persist-credentials: false
+      - uses: actions/configure-pages@45bfe0192ca1faeb007ade9deae92b16b8254a0d  # v6.0.0
+      - uses: actions/jekyll-build-pages@44a6e6beabd48582f863aeeb6cb2151cc1716697  # v1.0.13
+        with:
+          source: website
+          destination: _site
+      - uses: actions/upload-pages-artifact@fc324d3547104276b827a68afc52ff2a11cc49c9  # v5.0.0
+        with:
+          path: _site
+
+  deploy:
+    name: Deploy static site
+    needs: build
+    runs-on: ubuntu-latest
+    permissions:
+      pages: write
+      id-token: write
+    environment:
+      name: github-pages
+      url: ${{ steps.deployment.outputs.page_url }}
+    steps:
+      - name: Deploy to GitHub Pages
+        id: deployment
+        uses: actions/deploy-pages@cd2ce8fcbc39b97be8ca5fce6e763baed58fa128  # v5.0.0

.github/workflows/docker-publish.yml

@@ -14,6 +14,8 @@ on:
     paths-ignore:
       - '**.md'
       - 'docs/**'
+      - 'website/**'
+      - 'assets/branding/**'
       - '.github/ISSUE_TEMPLATE/**'
 
 concurrency:
@@ -45,20 +47,20 @@ jobs:
             arch: arm64
             runner: ubuntu-24.04-arm
     steps:
-      - uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10  # v6.0.3
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
         with:
           persist-credentials: false
       - name: Set up Buildx
-        uses: docker/setup-buildx-action@d7f5e7f509e45cec5c76c4d5afdd7de93d0b3df5  # v4.1.0
+        uses: docker/setup-buildx-action@37fe631027851001ddb9b187196cc803df7f5f0e  # v4.3.0
       - name: Log in to GHCR
-        uses: docker/login-action@650006c6eb7dba73a995cc03b0b2d7f5ca915bee  # v4.2.0
+        uses: docker/login-action@dbcb813823bdd20940b903addbd779551569679f  # v4.6.0
         with:
           registry: ${{ env.REGISTRY }}
           username: ${{ github.actor }}
           password: ${{ secrets.GITHUB_TOKEN }}
       - name: Build and push by digest
         id: build
-        uses: docker/build-push-action@f9f3042f7e2789586610d6e8b85c8f03e5195baf  # v7.2.0
+        uses: docker/build-push-action@53b7df96c91f9c12dcc8a07bcb9ccacbed38856a  # v7.3.0
         with:
           context: .
           platforms: ${{ matrix.platform }}
@@ -86,7 +88,7 @@ jobs:
       contents: read
       packages: write
     steps:
-      - uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10  # v6.0.3
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
         with:
           persist-credentials: false
       - name: Read APP_VERSION + short sha
@@ -103,16 +105,16 @@ jobs:
           pattern: digest-*
           merge-multiple: true
       - name: Set up Buildx
-        uses: docker/setup-buildx-action@d7f5e7f509e45cec5c76c4d5afdd7de93d0b3df5  # v4.1.0
+        uses: docker/setup-buildx-action@37fe631027851001ddb9b187196cc803df7f5f0e  # v4.3.0
       - name: Log in to GHCR
-        uses: docker/login-action@650006c6eb7dba73a995cc03b0b2d7f5ca915bee  # v4.2.0
+        uses: docker/login-action@dbcb813823bdd20940b903addbd779551569679f  # v4.6.0
         with:
           registry: ${{ env.REGISTRY }}
           username: ${{ github.actor }}
           password: ${{ secrets.GITHUB_TOKEN }}
       - name: Compute tags
         id: meta
-        uses: docker/metadata-action@80c7e94dd9b9319bd5eb7a0e0fe9291e23a2a2e9  # v6.1.0
+        uses: docker/metadata-action@dc802804100637a589fabce1cb79ff13a1411302  # v6.2.0
         with:
           images: ${{ env.REGISTRY }}/${{ env.IMAGE_NAME }}
           tags: |

.github/workflows/issue-description-check.yml

@@ -2,7 +2,7 @@ name: ci / issue description check
 
 on:
   issues:
-    types: [opened, edited, reopened]
+    types: [opened, edited, reopened, closed]
 
 permissions:
   issues: write
@@ -14,7 +14,7 @@ jobs:
     # Skip bots (Dependabot, release-drafter, etc.)
     if: ${{ github.event.issue.user.type != 'Bot' }}
     steps:
-      - uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10  # v6.0.3
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
         with:
           sparse-checkout: .github/scripts
           persist-credentials: false

.github/workflows/pr-description-check.yml

@@ -5,7 +5,11 @@ on:
   # works on fork PRs. Safe here: the checkout pins to the base branch (no fork
   # code runs) and the scripts only read context.payload and call the GitHub API.
   pull_request_target:  # zizmor: ignore[dangerous-triggers]
-    types: [opened, edited, synchronize, reopened, ready_for_review]
+    types: [opened, edited, synchronize, reopened, ready_for_review, converted_to_draft]
+
+concurrency:
+  group: pr-description-${{ github.event.pull_request.number }}
+  cancel-in-progress: true
 
 # Default-deny at the workflow level; each job opts into only the scopes it needs.
 # Note: modifying a PR's labels/comments needs pull-requests:write even though the
@@ -23,7 +27,7 @@ jobs:
     # Skip bots: they open PRs programmatically and have their own process.
     if: github.event.pull_request.user.type != 'Bot'
     steps:
-      - uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10  # v6.0.3
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
         with:
           ref: ${{ github.base_ref }}
           sparse-checkout: .github/scripts
@@ -59,12 +63,14 @@ jobs:
 
   check-mergeable:
     name: Flag unmergeable PRs
+    needs: check-description
     runs-on: ubuntu-latest
     permissions:
       pull-requests: write
       issues: write
-    # Skip bots: they open PRs programmatically and have their own process.
-    if: github.event.pull_request.user.type != 'Bot'
+    # Run after description validation failures, but never from an obsolete
+    # workflow run canceled by a newer PR event.
+    if: ${{ !cancelled() && github.event.pull_request.user.type != 'Bot' }}
     steps:
       - uses: actions/github-script@3a2844b7e9c422d3c10d287c895573f7108da1b3  # v9.0.0
         with:

.github/workflows/secret-scan.yml

@@ -35,7 +35,7 @@ jobs:
       contents: read
     steps:
       - name: Checkout repository
-        uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10  # v6.0.3
+        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
         with:
           # Full history so a secret committed in an earlier commit (and later
           # deleted) is still caught -- deletion does not remove it from Git.

.github/workflows/workflow-security.yml

@@ -36,7 +36,7 @@ jobs:
       contents: read
     steps:
       - name: Checkout repository
-        uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10  # v6.0.3
+        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
         with:
           persist-credentials: false
 
@@ -61,12 +61,12 @@ jobs:
       contents: read
     steps:
       - name: Checkout repository
-        uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10  # v6.0.3
+        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
         with:
           persist-credentials: false
 
       - name: Set up Python
-        uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405  # v6.2.0
+        uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97  # v7.0.0
         with:
           python-version: '3.12'
 

.gitignore

@@ -85,6 +85,24 @@ output.txt.txt
 !docs/**/*.gif
 !docs/**/*.webp
 
+# …except shipped website and branding media.
+!website/**/*.jpg
+!website/**/*.jpeg
+!website/**/*.png
+!website/**/*.gif
+!website/**/*.bmp
+!website/**/*.webp
+!website/**/*.tiff
+!website/**/*.pdf
+!assets/branding/**/*.jpg
+!assets/branding/**/*.jpeg
+!assets/branding/**/*.png
+!assets/branding/**/*.gif
+!assets/branding/**/*.bmp
+!assets/branding/**/*.webp
+!assets/branding/**/*.tiff
+!assets/branding/**/*.pdf
+
 # Reports and temp files
 reports/
 tasks/

ACKNOWLEDGMENTS.md

@@ -65,6 +65,16 @@ Vendored in `static/lib/` and served directly:
 | [jsPDF](https://github.com/parallax/jsPDF) (bundled in html2pdf) | PDF generation | MIT |
 | [html2canvas](https://github.com/niklasvh/html2canvas) (bundled in html2pdf) | DOM → canvas rasterization | MIT |
 | [node-qrcode](https://github.com/soldair/node-qrcode) (`qrcode.min.js`) | QR-code rendering (2FA setup) | MIT |
+| [KaTeX](https://github.com/KaTeX/KaTeX) v0.16.22 (`katex/katex.min.{js,css}` + `katex/fonts/*.woff2`) | Math typesetting | MIT ([`licenses/KaTeX-MIT-LICENSE.txt`](licenses/KaTeX-MIT-LICENSE.txt)) |
+| [Mermaid](https://github.com/mermaid-js/mermaid) v11.16.1 (`mermaid.min.js`) | Diagrams from text | MIT ([`licenses/Mermaid-MIT-LICENSE.txt`](licenses/Mermaid-MIT-LICENSE.txt)) |
+
+KaTeX and Mermaid are loaded on first use by `static/js/markdown.js` rather than
+from `index.html`, so a session that renders no math and no diagram never fetches
+either. Only the `.woff2` KaTeX fonts are shipped, matching `static/fonts/`; the
+`.woff` and `.ttf` variants its stylesheet also lists are never requested by a
+browser that supports `woff2`. The bundles are the published npm artifacts,
+unmodified — `.gitattributes` turns the whitespace check off for `static/lib/`
+so they can stay byte-identical to upstream.
 
 ## Front-end libraries loaded at runtime (CDN)
 
@@ -72,8 +82,6 @@ Referenced from `cdn.jsdelivr.net` / `cdnjs.cloudflare.com` at runtime — not v
 
 | Library | Purpose | License |
 |---|---|---|
-| [KaTeX](https://github.com/KaTeX/KaTeX) 0.16.22 | Math typesetting | MIT |
-| [Mermaid](https://github.com/mermaid-js/mermaid) 11 | Diagrams from text | MIT |
 | [Pyodide](https://github.com/pyodide/pyodide) 0.27.5 | In-browser Python runtime | MPL-2.0 |
 | [PDFObject](https://github.com/pipwerks/PDFObject) 2.1.1 | Inline PDF embedding | MIT |
 
@@ -86,6 +94,7 @@ Bundled in `static/fonts/`:
 | [Fira Code](https://github.com/tonsky/FiraCode) | SIL Open Font License 1.1 | Nikita Prokopov & contributors |
 | [Inter](https://github.com/rsms/inter) | SIL Open Font License 1.1 | Rasmus Andersson |
 | [GohuFont](https://font.gohu.org/) (`fonts/custom/GohuFont.ttf`) | WTFPL | Hugo Chargois |
+| [OpenDyslexic](https://opendyslexic.org/) (`fonts/OpenDyslexic-{Regular,Bold}.woff2`) | SIL Open Font License 1.1 ([`licenses/OpenDyslexic-OFL.txt`](licenses/OpenDyslexic-OFL.txt)) | Abbie Gonzalez |
 
 ## Python dependencies
 

CONTRIBUTING.md

@@ -25,7 +25,7 @@ End-users cloning the repo will land on `dev` by default. To run the curated/sta
 Docker is the recommended path for normal testing:
 
 ```bash
-git clone https://github.com/pewdiepie-archdaemon/odysseus.git
+git clone https://github.com/odysseus-dev/odysseus.git
 cd odysseus
 cp .env.example .env
 docker compose up -d --build
@@ -37,7 +37,7 @@ Manual development uses Python 3.11+:
 python3 -m venv venv
 source venv/bin/activate
 pip install -r requirements.txt
-python -m uvicorn app:app --host 0.0.0.0 --port 7000
+python -m uvicorn app:app --host 127.0.0.1 --port 7000
 ```
 
 Windows is not actively tested. Docker on Linux or a Linux/macOS manual install is the safer path for now.

Dockerfile

@@ -1,3 +1,14 @@
+# ---- builder: patch + build wheels for Real-ESRGAN's broken-on-3.14 deps ----
+# basicsr/gfpgan/facexlib read their version via exec()+locals()['__version__'],
+# which raises KeyError on Python 3.13+ (PEP 667). Build patched wheels here so
+# the final image / Cookbook never has to compile the broken sdists. See
+# docker/build-realesrgan-wheels.sh for the full rationale.
+FROM python:3.14-slim AS realesrgan-wheels
+RUN apt-get update && apt-get install -y --no-install-recommends curl \
+    && rm -rf /var/lib/apt/lists/*
+COPY docker/build-realesrgan-wheels.sh /usr/local/bin/build-realesrgan-wheels.sh
+RUN bash /usr/local/bin/build-realesrgan-wheels.sh /wheels
+
 FROM python:3.14-slim
 
 # System deps. tmux is required by Cookbook for background downloads/serves.
@@ -5,7 +16,12 @@ FROM python:3.14-slim
 # downloads, and serves from Docker installs.
 # git/cmake are required when Cookbook builds llama.cpp on first llama.cpp
 # launch inside Docker.
-# nodejs/npm provide npx for the optional built-in Browser MCP server.
+# nodejs/npm provide npx for the built-in Browser MCP server.
+# chromium provides the actual browser binary used by that MCP server.
+# fontconfig + Noto CJK provide real fallback glyphs for multilingual pages;
+# Chromium otherwise renders Chinese/Japanese/Korean labels as empty boxes.
+# iproute2/iputils-ping/net-tools/dnsutils/nmap give Docker-hosted agents the
+# basic network inspection toolkit expected by local LAN/debugging tasks.
 # gosu lets the entrypoint drop privileges cleanly so signals still reach
 # uvicorn directly (no extra shell layer like `su`/`sudo` would add).
 RUN apt-get update && apt-get install -y --no-install-recommends \
@@ -15,11 +31,60 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
     git \
     nodejs \
     npm \
+    chromium \
+    fontconfig \
+    fonts-noto-cjk \
     tmux \
     openssh-client \
+    iproute2 \
+    iputils-ping \
+    net-tools \
+    dnsutils \
+    nmap \
     gosu \
+    libgl1 \
+    libglib2.0-0t64 \
+    libxcb1 \
+    libmagic1 \
     && rm -rf /var/lib/apt/lists/*
 
+# Private browser automation wrapper used by the native `private_browser` tool.
+# Chromium is installed above, so agent-browser can drive the existing browser
+# binary without paying `npx` startup/install overhead on each tool call.
+RUN npm install -g agent-browser@0.35.0 --omit=dev --loglevel=error
+
+# libgl1/libglib2.0-0t64/libxcb1 are runtime shared libs (libGL.so.1,
+# libglib-2.0/libgthread, libxcb.so.1) that opencv-python (cv2) loads. The
+# slim base omits them, so the Cookbook "install realesrgan" path imports cv2
+# and dies with `libxcb.so.1: cannot open shared object file` despite a clean
+# pip install. Using full opencv-python (not -headless) because basicsr/gfpgan/
+# facexlib/realesrgan all depend on the `opencv-python` distribution by name.
+#
+# libmagic1 is the shared lib (libmagic.so.1) that python-magic dlopens for
+# content-based MIME sniffing in src/upload_handler.py. We install both here
+# (libmagic1 + the python-magic wrapper, below) rather than in requirements.txt
+# because python-magic resolves libmagic at import time: where the lib is
+# absent the import can block or raise, so keeping it image-only avoids
+# regressing pip/venv installs on hosts without libmagic. Debian always has the
+# lib here, so the import is instant and detection actually works.
+
+# Docker CLI (client only — daemon stays on the host via the
+# /var/run/docker.sock mount). The Debian `docker.io` package ships
+# dockerd but not the client binary on slim, so grab the static client
+# tarball from download.docker.com instead.
+ARG DOCKER_CLI_VERSION=29.6.2
+RUN ARCH="$(dpkg --print-architecture)" \
+    && case "$ARCH" in \
+         amd64) DARCH=x86_64 ;; \
+         arm64) DARCH=aarch64 ;; \
+         *) echo "unsupported arch $ARCH"; exit 1 ;; \
+       esac \
+    && curl -fsSL "https://download.docker.com/linux/static/stable/${DARCH}/docker-${DOCKER_CLI_VERSION}.tgz" \
+       -o /tmp/docker.tgz \
+    && tar -xzf /tmp/docker.tgz -C /tmp \
+    && install -m 0755 /tmp/docker/docker /usr/local/bin/docker \
+    && rm -rf /tmp/docker /tmp/docker.tgz
+
 WORKDIR /app
 
 # Install Python deps first (layer cache). Optional extras (PyMuPDF AGPL, etc.)
@@ -29,6 +94,20 @@ COPY requirements.txt requirements-optional.txt ./
 RUN pip install --no-cache-dir -r requirements.txt \
     && if [ "$INSTALL_OPTIONAL" = "true" ]; then pip install --no-cache-dir -r requirements-optional.txt; fi
 
+# python-magic powers content-based MIME sniffing in src/upload_handler.py.
+# Image-only (not in requirements.txt) because it needs the libmagic1 system
+# lib installed above; see the apt note near the top of this stage.
+RUN pip install --no-cache-dir python-magic==0.4.27
+
+# Pre-install the patched basicsr/gfpgan/facexlib wheels built in the
+# realesrgan-wheels stage (--no-deps keeps the image lean — torch & friends are
+# pulled only when realesrgan is actually installed). With these dists already
+# satisfied, the Cookbook's plain `pip install realesrgan` resolves them from
+# wheels instead of rebuilding the sdists that fail on Python 3.14.
+COPY --from=realesrgan-wheels /wheels/ /tmp/odysseus-wheels/
+RUN pip install --no-cache-dir --no-deps /tmp/odysseus-wheels/*.whl \
+    && rm -rf /tmp/odysseus-wheels
+
 # Copy app code
 COPY . .
 

HARNESS_VERSION

@@ -0,0 +1 @@
+0.20.5

Odysseus.spec

@@ -0,0 +1,45 @@
+# -*- mode: python ; coding: utf-8 -*-
+
+
+a = Analysis(
+    ['launcher.py'],
+    pathex=[],
+    binaries=[],
+    datas=[('static', 'static'), ('scripts', 'scripts'), ('mcp_servers', 'mcp_servers'), ('services/hwfit/data', 'services/hwfit/data'), ('config', 'config'), ('.env.example', '.env.example')],
+    hiddenimports=[],
+    hookspath=[],
+    hooksconfig={},
+    runtime_hooks=[],
+    excludes=[],
+    noarchive=False,
+    optimize=0,
+)
+pyz = PYZ(a.pure)
+
+exe = EXE(
+    pyz,
+    a.scripts,
+    [],
+    exclude_binaries=True,
+    name='Odysseus',
+    debug=False,
+    bootloader_ignore_signals=False,
+    strip=False,
+    upx=True,
+    console=False,
+    disable_windowed_traceback=False,
+    argv_emulation=False,
+    target_arch=None,
+    codesign_identity=None,
+    entitlements_file=None,
+    icon=['static\\icon.ico'],
+)
+coll = COLLECT(
+    exe,
+    a.binaries,
+    a.datas,
+    strip=False,
+    upx=True,
+    upx_exclude=[],
+    name='Odysseus',
+)

README.md

@@ -1,5 +1,5 @@
 <p align="center">
-  <img src="docs/odysseus-wordmark.png" alt="Odysseus" width="280">
+  <img src="assets/branding/odysseus-wordmark.png" alt="Odysseus" width="238">
 </p>
 
 <p align="center">
@@ -8,7 +8,7 @@
 
 <p align="center">
   <a href="#quick-start">Quick Start</a> ·
-  <a href="docs/setup.md">Setup Guide</a> ·
+  <a href="website/setup.md">Setup Guide</a> ·
   <a href="CONTRIBUTING.md">Contributing</a> ·
   <a href="ROADMAP.md">Roadmap</a>
 </p>
@@ -18,17 +18,17 @@
 </p>
 
 <p align="center">
-  <img src="docs/odysseus.jpg" alt="Odysseus interface">
+  <img src="assets/branding/odysseus-browser.jpg" alt="Odysseus interface">
 </p>
 
 ---
 
 ## Quick Start
 
-> `dev` is the default branch and gets the newest changes first. Use [`main`](https://github.com/pewdiepie-archdaemon/odysseus/tree/main) if you want the more curated branch.
+> `dev` is the default branch and gets the newest changes first. Use [`main`](https://github.com/odysseus-dev/odysseus/tree/main) if you want the more curated branch.
 
 ```bash
-git clone https://github.com/pewdiepie-archdaemon/odysseus.git
+git clone https://github.com/odysseus-dev/odysseus.git
 cd odysseus
 cp .env.example .env
 docker compose up -d --build
@@ -36,7 +36,7 @@ docker compose up -d --build
 
 Open `http://localhost:7000` when the containers are healthy. The first admin password is printed in `docker compose logs odysseus`.
 
-Native installs, GPU notes, Windows/macOS instructions, HTTPS, and configuration live in the [setup guide](docs/setup.md).
+Native installs, GPU notes, Windows/macOS instructions, HTTPS, and configuration live in the [setup guide](website/setup.md).
 
 ## Features
 
@@ -51,7 +51,7 @@ Native installs, GPU notes, Windows/macOS instructions, HTTPS, and configuration
 
 ## Demo
 
-A full hover-to-play tour lives on the landing page: [`docs/index.html`](docs/index.html).
+A full hover-to-play tour lives on the [Odysseus landing page](https://odysseus-dev.github.io/odysseus/). Its source lives under [`website/`](website/).
 
 ## Contributing
 
@@ -59,15 +59,20 @@ Help is welcome. The best entry points are fresh-install testing, provider setup
 
 ## Security
 
-Odysseus is a self-hosted workspace with powerful local tools. Keep auth enabled, keep private data out of Git, and do not expose raw model/service ports publicly. Deployment details are in the [setup guide](docs/setup.md#security-notes).
+Odysseus is a self-hosted workspace with powerful local tools. Keep auth enabled, keep private data out of Git, and do not expose raw model/service ports publicly.
+
+- Keep `AUTH_ENABLED=true` for any network-accessible deployment.
+- Keep `LOCALHOST_BYPASS=false` outside local development.
+
+Deployment details are in the [setup guide](website/setup.md#security-notes).
 
 ## Star History
 
-<a href="https://www.star-history.com/?repos=pewdiepie-archdaemon%2Fodysseus&type=date&legend=top-left">
+<a href="https://star-history.dera.page/#odysseus-dev/odysseus&type=date&legend=top-left">
  <picture>
-   <source media="(prefers-color-scheme: dark)" srcset="https://api.star-history.com/chart?repos=pewdiepie-archdaemon/odysseus&type=date&theme=dark&legend=top-left" />
-   <source media="(prefers-color-scheme: light)" srcset="https://api.star-history.com/chart?repos=pewdiepie-archdaemon/odysseus&type=date&legend=top-left" />
-   <img alt="Star History Chart" src="https://api.star-history.com/chart?repos=pewdiepie-archdaemon/odysseus&type=date&legend=top-left" />
+   <source media="(prefers-color-scheme: dark)" srcset="https://star-history.dera.page/svg?repos=odysseus-dev/odysseus&type=date&theme=dark&legend=top-left" />
+   <source media="(prefers-color-scheme: light)" srcset="https://star-history.dera.page/svg?repos=odysseus-dev/odysseus&type=date&legend=top-left" />
+   <img alt="Star History Chart" src="https://star-history.dera.page/svg?repos=odysseus-dev/odysseus&type=date&legend=top-left" />
  </picture>
 </a>
 

ROADMAP.md

@@ -12,7 +12,6 @@ the codebase, you are probably right to stay away.
   and WSL all need coverage.
 
 - Integration audit: do integrations even work? Confirm what works, what needs setup docs, and what should be removed or hidden. 
-- Self-host troubleshooting cookbook. Document the weird 30-second fixes that otherwise become 30-minute searches: Dovecot cleartext auth for local stacks, ntfy Android Instant Delivery for non-ntfy.sh servers, clipboard limits on plain-HTTP Tailscale URLs, Radicale collection URLs, and similar traps.
 - Cookbook reliability on other computers. This is probably the area most likely to need work across different machines, GPUs, drivers, shells, and Python environments.
 - Cookbook SGLang support across platforms. Make sure SGLang setup/serve works
   predictably on Linux, Windows/WSL, macOS where possible, Docker, and common
@@ -33,6 +32,14 @@ the codebase, you are probably right to stay away.
   before the user request really starts. We need slimmer prompts, better tool
   selection, smaller default tool sets, and clearer guidance for models with
   4k/8k/16k context windows.
+- Local model speculative decoding support. For Odysseus-tuned local models,
+  plan to ship or recommend a small same-tokenizer draft model when the serving
+  backend supports it. Early vLLM testing showed a generic `Qwen3-0.6B` draft
+  beside `Qwen3-8B` can materially reduce wall time, while an unsupported
+  DSpark conversion performed poorly. Treat this as a supported draft-model lane
+  first; keep MTP-specific packaging as future work only when the architecture
+  and runtime support are real. Judge this by time-to-success, tool correctness,
+  grammar, and unchanged target output, not tokens/sec alone.
 - Skill/tool prompt-injection audit. User-editable skills, notes, documents,
   fetched pages, and memories should be treated as untrusted data. Keep testing
   whether models follow malicious instructions from those surfaces.

SECURITY.md

@@ -10,7 +10,7 @@ Security fixes are handled on the default branch until formal releases are cut.
 
 - Keep `AUTH_ENABLED=true` for any network-accessible deployment.
 - Keep `LOCALHOST_BYPASS=false` outside local development.
-- Set `SECURE_COOKIES=true` when Odysseus is served through HTTPS by a trusted reverse proxy or private access gateway.
+- Leave `SECURE_COOKIES` unset unless you need to override it: session cookies are marked `Secure` whenever the request arrives over HTTPS. Set `SECURE_COOKIES=true` to force it on (for a proxy Odysseus cannot see the scheme of), or `SECURE_COOKIES=false` to force it off while you still serve plain HTTP alongside HTTPS.
 - Use HTTPS when exposing the app beyond localhost.
 - Put the authenticated Odysseus web/API entrypoint behind a trusted reverse proxy or private access layer such as Cloudflare Access, Tailscale, or a VPN.
 - Keep ChromaDB, SearXNG, ntfy, Ollama, vLLM, llama.cpp, databases, and raw model/provider APIs internal-only.

THREAT_MODEL.md

@@ -37,7 +37,7 @@ Non-admin defaults are in `core/auth.py:DEFAULT_PRIVILEGES`. Tool enforcement is
 
 - **Sessions:** bcrypt passwords, 7-day session tokens stored atomically in `data/sessions.json` via `core/atomic_io.py`.
 - **2FA:** TOTP with 8 single-use backup codes. Verified after password check, before session issuance.
-- **Reserved usernames:** `internal-tool`, `api`, `demo`, `system` cannot be registered or renamed into. Defined in `core/auth.py:RESERVED_USERNAMES`.
+- **Reserved usernames:** request sentinels and the Default/Local storage owner cannot be registered or renamed into. Defined in `core/auth.py:RESERVED_USERNAMES`.
   - `internal-tool` is security-critical: `core/middleware.py:require_admin` treats any request where `request.state.current_user == "internal-tool"` as the in-process tool loopback and grants admin unconditionally. A real account with that name would silently pass every `require_admin` check.
 - **Orphan sessions:** `validate_token` re-checks that the user record still exists on every call. A deleted user's cookie is dropped on next request rather than continuing to authenticate.
 

app.py

@@ -1,6 +1,20 @@
 # app.py — slim orchestrator
 import mimetypes
 import os
+import sys
+import asyncio
+import time
+import shutil
+import socket
+
+# On Windows, asyncio.create_subprocess_exec/shell require the ProactorEventLoop.
+# When started via `python -m uvicorn` from a terminal, uvicorn sets this
+# automatically. But the VS Code debugger (and other non-uvicorn entrypoints)
+# use the default SelectorEventLoop, which raises NotImplementedError on any
+# subprocess call. Force ProactorEventLoop here so the right loop is always
+# used, regardless of how the process is launched.
+if sys.platform == "win32":
+    asyncio.set_event_loop_policy(asyncio.WindowsProactorEventLoopPolicy())
 
 
 def register_static_mime_types() -> None:
@@ -38,12 +52,12 @@ load_dotenv(encoding="utf-8-sig")
 import asyncio
 import logging
 import secrets
-from datetime import datetime
+from datetime import datetime, timezone
 from typing import Dict
 
 from contextlib import asynccontextmanager
 from fastapi import FastAPI, Request, HTTPException
-from fastapi.responses import JSONResponse, FileResponse, HTMLResponse
+from fastapi.responses import JSONResponse, FileResponse
 from fastapi.middleware.cors import CORSMiddleware
 from fastapi.staticfiles import StaticFiles
 from starlette.middleware.base import BaseHTTPMiddleware
@@ -55,7 +69,13 @@ from core.constants import (
     REQUEST_TIMEOUT, OPENAI_API_KEY, AUTH_FILE,
 )
 from core.database import SessionLocal, ApiToken
-from core.middleware import SecurityHeadersMiddleware, is_cors_preflight
+from core.middleware import (
+    SecurityHeadersMiddleware,
+    get_application_route_path,
+    is_cors_preflight,
+    path_is_route_or_child,
+    with_asgi_root_path,
+)
 from core.auth import AuthManager, normalize_known_username
 from core.exceptions import (
     SessionNotFoundError, InvalidFileUploadError,
@@ -64,8 +84,9 @@ from core.exceptions import (
 
 import bcrypt as _bcrypt
 
-from src.app_helpers import abs_join
+from src.app_helpers import abs_join, serve_html_with_nonce
 from src.generated_images import GENERATED_IMAGE_HEADERS, resolve_generated_image_path
+from src.owner_identity import auth_disabled
 from starlette.responses import RedirectResponse
 
 # ========= LOGGING =========
@@ -113,12 +134,13 @@ app = FastAPI(
 )
 
 # ========= CORS =========
+CORS_ALLOW_METHODS = ["GET", "POST", "PUT", "PATCH", "DELETE"]
 allowed_origins = os.getenv("ALLOWED_ORIGINS", "http://localhost,http://127.0.0.1").split(",")
 app.add_middleware(
     CORSMiddleware,
     allow_origins=allowed_origins,
     allow_credentials=True,
-    allow_methods=["GET", "POST", "PUT", "DELETE"],
+    allow_methods=CORS_ALLOW_METHODS,
     allow_headers=[
         "Accept",
         "Authorization",
@@ -140,7 +162,8 @@ app.add_middleware(
 # model-probe — all served with media_type="text/event-stream") are never
 # compressed or buffered; only complete bodies over minimum_size are. The
 # security-header middleware composes cleanly on top.
-app.add_middleware(GZipMiddleware, minimum_size=1024, compresslevel=6)
+if os.getenv("RESPONSE_COMPRESSION_ENABLED", "true").strip().lower() not in {"0", "false", "no", "off"}:
+    app.add_middleware(GZipMiddleware, minimum_size=1024, compresslevel=6)
 
 # ========= SECURITY HEADERS MIDDLEWARE =========
 app.add_middleware(SecurityHeadersMiddleware)
@@ -185,14 +208,57 @@ class _RequestTimeoutMiddleware(_BaseHTTPMiddleware):
             )
 
 
+class _InteractiveActivityMiddleware(_BaseHTTPMiddleware):
+    async def dispatch(self, request, call_next):
+        from src.interactive_gate import should_track_interactive_request, track_interactive_request
+
+        path = request.url.path or ""
+        if not should_track_interactive_request(path, request.method):
+            return await call_next(request)
+        async def _stop_background():
+            try:
+                await task_scheduler.stop_background_tasks_for_foreground(reason=f"foreground request {request.method} {path}")
+            except Exception:
+                logging.getLogger("app.foreground_gate").debug("foreground task stop failed", exc_info=True)
+        asyncio.create_task(_stop_background())
+        async with track_interactive_request(path, request.method):
+            return await call_next(request)
+
+
+class _SlowRequestLogMiddleware(_BaseHTTPMiddleware):
+    async def dispatch(self, request, call_next):
+        start = time.perf_counter()
+        status = 500
+        try:
+            response = await call_next(request)
+            status = getattr(response, "status_code", 0) or 0
+            return response
+        finally:
+            elapsed = time.perf_counter() - start
+            try:
+                threshold = float(os.getenv("ODYSSEUS_SLOW_REQUEST_LOG_SECONDS", "0.75") or "0.75")
+            except Exception:
+                threshold = 0.75
+            if elapsed >= threshold:
+                logging.getLogger("app.slow_request").warning(
+                    "slow_request method=%s path=%s status=%s elapsed=%.3fs",
+                    request.method,
+                    request.url.path,
+                    status,
+                    elapsed,
+                )
+
+
 app.add_middleware(_RequestTimeoutMiddleware)
+app.add_middleware(_InteractiveActivityMiddleware)
+app.add_middleware(_SlowRequestLogMiddleware)
 
 # ========= AUTH =========
 from routes.auth_routes import setup_auth_routes, SESSION_COOKIE
 
 auth_manager = AuthManager()
 app.state.auth_manager = auth_manager
-AUTH_ENABLED = os.getenv("AUTH_ENABLED", "true").lower() != "false"
+AUTH_ENABLED = not auth_disabled()
 LOCALHOST_BYPASS = os.getenv("LOCALHOST_BYPASS", "false").lower() == "true"
 if LOCALHOST_BYPASS:
     logger.warning("LOCALHOST_BYPASS is enabled, loopback requests bypass authentication. Do not expose this instance to a network.")
@@ -228,7 +294,7 @@ if AUTH_ENABLED:
     def _is_auth_exempt(path: str) -> bool:
         if path in AUTH_EXEMPT_EXACT:
             return True
-        if any(path.startswith(p) for p in AUTH_EXEMPT_PREFIXES):
+        if any(path_is_route_or_child(path, p) for p in AUTH_EXEMPT_PREFIXES):
             return True
         return any(p.match(path) for p in AUTH_EXEMPT_PATTERNS)
 
@@ -299,7 +365,7 @@ if AUTH_ENABLED:
 
     class AuthMiddleware(BaseHTTPMiddleware):
         async def dispatch(self, request: Request, call_next):
-            path = request.url.path
+            path = get_application_route_path(request.scope)
             # A genuine CORS preflight (OPTIONS + Access-Control-Request-Method)
             # carries no credentials by design and must reach CORSMiddleware to be
             # answered. AuthMiddleware is the outermost middleware, so gating the
@@ -316,7 +382,7 @@ if AUTH_ENABLED:
             # (no admin cookie available in that context). Restricted to
             # loopback clients + matching token to keep it locked down.
             try:
-                from core.middleware import INTERNAL_TOOL_HEADER, INTERNAL_TOOL_TOKEN as _ITT
+                from core.middleware import INTERNAL_TOOL_HEADER, INTERNAL_TOOL_TOKEN as _ITT, INTERNAL_TOOL_USER
                 _hdr = request.headers.get(INTERNAL_TOOL_HEADER)
                 if _hdr and secrets.compare_digest(_hdr, _ITT) and _is_trusted_loopback(request):
                     # Impersonation: when the agent's loopback call sets
@@ -328,11 +394,11 @@ if AUTH_ENABLED:
                     if _impersonate and _impersonate in getattr(_auth_mgr, "users", {}):
                         request.state.current_user = _impersonate
                     else:
-                        request.state.current_user = "internal-tool"
+                        request.state.current_user = INTERNAL_TOOL_USER
                     request.state.api_token = False
                     return await call_next(request)
-            except Exception:
-                pass
+            except Exception as _e:
+                logger.warning("Internal tool auth header check failed", exc_info=_e)
             # Allow DIRECT localhost requests (internal service calls from
             # heartbeats etc.). Tunnel/proxy-forwarded requests are excluded by
             # _is_trusted_loopback so LOCALHOST_BYPASS can't be abused over a
@@ -343,7 +409,10 @@ if AUTH_ENABLED:
             if not auth_manager.is_configured:
                 # No users yet — redirect to login for first-time setup
                 if not path.startswith("/api/"):
-                    return RedirectResponse(url="/login", status_code=302)
+                    return RedirectResponse(
+                        url=with_asgi_root_path(request.scope, "/login"),
+                        status_code=302,
+                    )
                 return JSONResponse(status_code=401, content={"error": "Setup required"})
 
             # --- Bearer token auth (API tokens for external integrations) ---
@@ -385,11 +454,10 @@ if AUTH_ENABLED:
                                     _db.close()
                             try:
                                 await _asyncio.to_thread(_do)
-                            except Exception:
-                                pass
+                            except Exception as _e:
+                                logger.debug("Failed to update token last_used_at", exc_info=_e)
                         _asyncio.create_task(_touch_last_used(matched_id))
                         # Keep bearer-token callers out of normal cookie/user
-                        # routes. API-aware routes can read api_token_owner.
                         request.state.current_user = "api"
                         request.state.api_token = True
                         request.state.api_token_id = matched_id
@@ -406,7 +474,10 @@ if AUTH_ENABLED:
             if not auth_manager.validate_token(token):
                 if path.startswith("/api/"):
                     return JSONResponse(status_code=401, content={"error": "Not authenticated"})
-                return RedirectResponse(url="/login", status_code=302)
+                return RedirectResponse(
+                    url=with_asgi_root_path(request.scope, "/login"),
+                    status_code=302,
+                )
 
             # Attach current username to request state for downstream routes
             request.state.current_user = auth_manager.get_username_for_token(token)
@@ -438,7 +509,7 @@ class _RevalidatingStatic(StaticFiles):
         return resp
 
 
-app.mount("/static", _RevalidatingStatic(directory="static"), name="static")
+app.mount("/static", _RevalidatingStatic(directory=STATIC_DIR), name="static")
 
 # ========= GENERATED IMAGES =========
 @app.get("/api/generated-image/{filename}")
@@ -464,8 +535,8 @@ async def serve_generated_image(filename: str, request: Request):
                 _db.close()
     except HTTPException:
         raise
-    except Exception:
-        pass
+    except Exception as _e:
+        logger.warning("Image ownership verification failed for %r", filename, exc_info=_e)
     ext = filename.rsplit('.', 1)[-1].lower()
     mime = {
         "png": "image/png", "jpg": "image/jpeg", "jpeg": "image/jpeg",
@@ -528,6 +599,7 @@ memory_vector     = components.get("memory_vector")
 upload_handler    = components["upload_handler"]
 app.state.upload_handler = upload_handler
 personal_docs_mgr = components["personal_docs_manager"]
+app.state.personal_docs_manager = personal_docs_mgr
 api_key_manager   = components["api_key_manager"]
 preset_manager    = components["preset_manager"]
 chat_processor    = components["chat_processor"]
@@ -571,6 +643,31 @@ webhook_manager = WebhookManager(api_key_manager=api_key_manager)
 auth_router = setup_auth_routes(auth_manager)
 app.include_router(auth_router)
 
+
+@app.post("/api/activity/heartbeat")
+async def activity_heartbeat():
+    from src.interactive_gate import (
+        mark_browser_activity,
+        maybe_stop_background_tasks_for_heartbeat,
+    )
+
+    await mark_browser_activity()
+
+    async def _stop_background():
+        try:
+            await maybe_stop_background_tasks_for_heartbeat(
+                task_scheduler.stop_background_tasks_for_foreground
+            )
+        except Exception:
+            logging.getLogger("app.foreground_gate").debug(
+                "heartbeat task stop failed",
+                exc_info=True,
+            )
+
+    asyncio.create_task(_stop_background())
+    return {"ok": True}
+
+
 # Uploads
 from routes.upload_routes import setup_upload_routes
 upload_router, upload_cleanup_func = setup_upload_routes(upload_handler)
@@ -585,14 +682,20 @@ app.include_router(setup_emoji_routes())
 # Sessions
 from routes.session_routes import setup_session_routes
 session_config = {"REQUEST_TIMEOUT": REQUEST_TIMEOUT, "OPENAI_API_KEY": OPENAI_API_KEY, "SESSIONS_FILE": SESSIONS_FILE}
-app.include_router(setup_session_routes(session_manager, session_config, webhook_manager=webhook_manager))
+app.include_router(setup_session_routes(
+    session_manager,
+    session_config,
+    webhook_manager=webhook_manager,
+    upload_handler=upload_handler,
+    skills_manager=skills_manager,
+))
 
 # Admin Danger Zone wipes (Settings → System → Danger Zone)
-from routes.admin_wipe_routes import setup_admin_wipe_routes
+from routes.admin_wipe.admin_wipe_routes import setup_admin_wipe_routes
 app.include_router(setup_admin_wipe_routes(session_manager))
 
 # Memory
-from routes.memory_routes import setup_memory_routes
+from routes.memory.memory_routes import setup_memory_routes
 memory_router = setup_memory_routes(memory_manager, session_manager, memory_vector=memory_vector)
 app.include_router(memory_router)
 from routes.skills_routes import setup_skills_routes
@@ -609,15 +712,15 @@ app.include_router(setup_chat_routes(
 ))
 
 # Research (background deep-research tasks)
-from routes.research_routes import setup_research_routes
+from routes.research.research_routes import setup_research_routes
 app.include_router(setup_research_routes(research_handler, session_manager=session_manager))
 
 # History
-from routes.history_routes import setup_history_routes
-app.include_router(setup_history_routes(session_manager))
+from routes.history.history_routes import setup_history_routes
+app.include_router(setup_history_routes(session_manager, upload_handler=upload_handler))
 
 # Search
-from routes.search_routes import setup_search_routes
+from routes.search.search_routes import setup_search_routes
 app.include_router(setup_search_routes(config))
 
 # Presets
@@ -629,7 +732,7 @@ from routes.diagnostics_routes import setup_diagnostics_routes
 app.include_router(setup_diagnostics_routes(rag_manager, rag_available, research_handler, memory_vector))
 
 # Cleanup
-from routes.cleanup_routes import setup_cleanup_routes
+from routes.cleanup.cleanup_routes import setup_cleanup_routes
 app.include_router(setup_cleanup_routes(session_manager))
 
 # Personal docs
@@ -664,7 +767,7 @@ app.include_router(setup_stt_routes(stt_service))
 logger.info("STT service initialized (provider managed via settings)")
 
 # Documents (artifacts/canvas)
-from routes.document_routes import setup_document_routes
+from routes.document.document_routes import setup_document_routes
 document_router = setup_document_routes(session_manager, upload_handler)
 app.include_router(document_router)
 
@@ -673,7 +776,7 @@ from routes.signature_routes import setup_signature_routes
 app.include_router(setup_signature_routes())
 
 # Gallery (image library)
-from routes.gallery_routes import setup_gallery_routes
+from routes.gallery.gallery_routes import setup_gallery_routes
 app.include_router(setup_gallery_routes())
 
 # Persisted image-editor drafts (server-backed projects)
@@ -685,7 +788,7 @@ from src.task_scheduler import TaskScheduler
 task_scheduler = TaskScheduler(session_manager)
 from src.event_bus import set_task_scheduler
 set_task_scheduler(task_scheduler)
-from routes.task_routes import setup_task_routes
+from routes.task.task_routes import setup_task_routes
 app.include_router(setup_task_routes(task_scheduler))
 
 from routes.assistant_routes import setup_assistant_routes
@@ -693,7 +796,7 @@ app.include_router(setup_assistant_routes(task_scheduler))
 
 # Calendar (CalDAV)
 from routes.calendar_routes import setup_calendar_routes
-calendar_router = setup_calendar_routes()
+calendar_router = setup_calendar_routes(upload_handler=upload_handler)
 app.include_router(calendar_router)
 
 # Shell (user-facing command execution)
@@ -712,7 +815,7 @@ from routes.hwfit_routes import setup_hwfit_routes
 app.include_router(setup_hwfit_routes())
 
 # Model A/B Comparison
-from routes.compare_routes import setup_compare_routes
+from routes.compare.compare_routes import setup_compare_routes
 app.include_router(setup_compare_routes(session_manager))
 
 # User Preferences
@@ -730,7 +833,7 @@ app.include_router(setup_font_routes())
 # MCP (Model Context Protocol)
 from src.mcp_manager import McpManager
 from src.agent_tools import set_mcp_manager
-from routes.mcp_routes import setup_mcp_routes
+from routes.mcp.mcp_routes import setup_mcp_routes
 
 mcp_manager = McpManager()
 set_mcp_manager(mcp_manager)
@@ -745,7 +848,7 @@ set_ai_rag_manager(rag_manager, personal_docs_mgr)
 logger.info("AI interaction tools initialized (session, memory, RAG, UI control)")
 
 # Webhooks
-from routes.webhook_routes import setup_webhook_routes
+from routes.webhook.webhook_routes import setup_webhook_routes
 app.include_router(setup_webhook_routes(webhook_manager, auth_manager, session_manager, api_key_manager))
 
 # API Tokens
@@ -755,8 +858,8 @@ app.include_router(setup_api_token_routes())
 logger.info("Webhook & API token routes initialized")
 
 # Notes (Google Keep-style notes/todos)
-from routes.note_routes import setup_note_routes
-app.include_router(setup_note_routes(task_scheduler))
+from routes.note.note_routes import setup_note_routes
+app.include_router(setup_note_routes(task_scheduler, upload_handler=upload_handler))
 
 # Email
 from routes.email_routes import setup_email_routes
@@ -777,11 +880,11 @@ app.include_router(setup_codex_routes(
 ))
 app.include_router(setup_claude_routes())
 
-from routes.vault_routes import setup_vault_routes
+from routes.vault.vault_routes import setup_vault_routes
 app.include_router(setup_vault_routes())
 
 # Contacts (CardDAV)
-from routes.contacts_routes import setup_contacts_routes
+from routes.contacts.contacts_routes import setup_contacts_routes
 app.include_router(setup_contacts_routes())
 
 from companion import setup_companion_routes
@@ -789,23 +892,17 @@ app.include_router(setup_companion_routes())
 
 # ========= ROUTES (kept in app.py) =========
 
-def _serve_html_with_nonce(request: Request, file_path: str) -> HTMLResponse:
-    """Read an HTML file and inject the CSP nonce into inline <script> tags."""
-    with open(file_path, "r", encoding="utf-8") as f:
-        html = f.read()
-    nonce = getattr(request.state, "csp_nonce", "")
-    html = html.replace("{{CSP_NONCE}}", nonce)
-    return HTMLResponse(html)
-
 @app.get("/")
 async def serve_index(request: Request):
     static_path = abs_join(BASE_DIR, "static/index.html")
     if os.path.exists(static_path):
-        return _serve_html_with_nonce(request, static_path)
-    root_path = abs_join(BASE_DIR, "index.html")
-    if os.path.exists(root_path):
-        return _serve_html_with_nonce(request, root_path)
-    raise HTTPException(404, "index.html not found")
+        return serve_html_with_nonce(request, static_path)
+    # No static bundle — fall back to a root-level index.html if one is shipped.
+    # If neither exists, serve_html_with_nonce logs it and returns a generic 500:
+    # a missing index.html is a broken deployment (server fault), not a client
+    # "not found". This keeps the app-shell route consistent with the other
+    # bundled-template routes instead of mislabelling the fault as a 404.
+    return serve_html_with_nonce(request, abs_join(BASE_DIR, "index.html"))
 
 @app.get("/notes")
 async def serve_notes(request: Request):
@@ -846,22 +943,54 @@ async def serve_library(request: Request):
 @app.get("/backgrounds")
 async def serve_backgrounds(request: Request):
     """Sandbox page for prototyping background effects. No auth required."""
-    return _serve_html_with_nonce(request, abs_join(BASE_DIR, "static/backgrounds.html"))
+    return serve_html_with_nonce(request, abs_join(BASE_DIR, "static/backgrounds.html"))
 
 @app.get("/login")
 async def serve_login(request: Request):
     if not AUTH_ENABLED:
         return RedirectResponse(url="/", status_code=302)
-    return _serve_html_with_nonce(request, abs_join(BASE_DIR, "static/login.html"))
+    return serve_html_with_nonce(request, abs_join(BASE_DIR, "static/login.html"))
 
 @app.get("/api/version")
 async def get_version():
-    from core.constants import APP_VERSION
-    return {"version": APP_VERSION}
+    from core.constants import APP_BUILD_VERSION, APP_SOURCE_COMMIT, APP_VERSION
+    return {
+        "version": APP_VERSION,
+        "build": APP_BUILD_VERSION,
+        "source_commit": APP_SOURCE_COMMIT,
+    }
 
 @app.get("/api/health")
 async def health_check() -> Dict[str, str]:
-    return {"status": "healthy", "timestamp": datetime.utcnow().isoformat()}
+    return {"status": "healthy", "timestamp": datetime.now(timezone.utc).isoformat()}
+
+@app.post("/api/client-perf")
+async def client_perf(request: Request):
+    """Low-volume frontend timing reports for stalls that happen before SSE logs."""
+    try:
+        data = await request.json()
+    except Exception:
+        data = {}
+    try:
+        kind = str(data.get("type") or "client").replace("\n", " ")[:80]
+        total_ms = float(data.get("total_ms") or 0)
+        stages = data.get("stages") if isinstance(data.get("stages"), list) else []
+        stage_txt = " ".join(
+            f"{str(s.get('name') or '')[:40]}={float(s.get('delta_ms') or 0):.0f}ms"
+            for s in stages[:20]
+            if isinstance(s, dict)
+        )
+        extra = str(data.get("extra") or "").replace("\n", " ")[:200]
+        logging.getLogger("app.client_perf").warning(
+            "client_perf type=%s total=%.0fms %s%s",
+            kind,
+            total_ms,
+            stage_txt,
+            f" extra={extra}" if extra else "",
+        )
+    except Exception:
+        logging.getLogger("app.client_perf").debug("client_perf log failed", exc_info=True)
+    return {"ok": True}
 
 @app.get("/api/ready")
 async def readiness_check() -> JSONResponse:
@@ -889,11 +1018,76 @@ async def runtime_info() -> Dict[str, object]:
         or os.getenv("OLLAMA_URL")
         or ("http://host.docker.internal:11434/v1" if in_docker else "http://127.0.0.1:11434/v1")
     )
+    network_mode = os.getenv("ODYSSEUS_CONTAINER_NETWORK_MODE", "").strip()
+    host_gateway_reachable = False
+    host_gateway_address = ""
+    if in_docker and network_mode != "host":
+        try:
+            resolved = socket.getaddrinfo("host.docker.internal", None)
+            for item in resolved:
+                sockaddr = item[4] if len(item) >= 5 else ()
+                candidate = sockaddr[0] if sockaddr else ""
+                if candidate:
+                    host_gateway_address = str(candidate)
+                    break
+            host_gateway_reachable = True
+        except OSError:
+            host_gateway_reachable = False
+        if not host_gateway_address:
+            host_gateway_address = _docker_default_gateway_ip()
+    container: Dict[str, object] = {
+        "engine": "docker" if in_docker else "",
+        "networkMode": network_mode,
+        "hostAccess": bool(in_docker and network_mode == "host"),
+        "hostGatewayReachable": host_gateway_reachable,
+    }
+    if host_gateway_address:
+        container["hostGatewayAddress"] = host_gateway_address
+    command_names = (
+        "ip",
+        "ss",
+        "arp",
+        "nmap",
+        "ping",
+        "dig",
+        "ssh",
+        "git",
+        "docker",
+    )
+    commands = {name: bool(shutil.which(name)) for name in command_names}
+    capabilities = {
+        "networkInspection": bool(commands["ip"] and (commands["ss"] or commands["arp"])),
+        "lanScan": bool(commands["nmap"]),
+        "dnsLookup": bool(commands["dig"]),
+        "sshClient": bool(commands["ssh"]),
+        "git": bool(commands["git"]),
+        "dockerClient": bool(commands["docker"]),
+    }
     return {
         "in_docker": in_docker,
         "ollama_base_url": ollama_url,
+        "container": container,
+        "commands": commands,
+        "capabilities": capabilities,
     }
 
+
+def _docker_default_gateway_ip() -> str:
+    try:
+        with open("/proc/net/route", "r", encoding="utf-8", errors="ignore") as fh:
+            for line in fh.readlines()[1:]:
+                parts = line.split()
+                if len(parts) < 3 or parts[1] != "00000000":
+                    continue
+                raw = parts[2]
+                if len(raw) != 8:
+                    continue
+                octets = [str(int(raw[i:i + 2], 16)) for i in range(6, -1, -2)]
+                return ".".join(octets)
+    except Exception:
+        return ""
+    return ""
+
 # ========= LIFECYCLE =========
 
 @asynccontextmanager
@@ -933,6 +1127,15 @@ async def _startup_event():
     # GC tasks created with `asyncio.create_task(...)` before they finish.
     _startup_tasks: list[asyncio.Task] = getattr(app.state, "_startup_tasks", [])
     app.state._startup_tasks = _startup_tasks
+    from src.background_tool_jobs import BackgroundToolJobs
+    from routes.chat_routes import _active_streams
+    from src import agent_runs
+    app.state.background_tool_jobs = BackgroundToolJobs(
+        is_busy=lambda sid: sid in _active_streams or agent_runs.is_active(sid),
+        session_manager=session_manager, research_handler=research_handler,
+    )
+    app.state.background_tool_delivery_task = asyncio.create_task(app.state.background_tool_jobs.run())
+    _startup_tasks.append(app.state.background_tool_delivery_task)
     if upload_cleanup_func:
         upload_cleanup_task = asyncio.create_task(upload_cleanup_func())
     # Always-on monitor that auto-continues the agent when a background bash
@@ -951,7 +1154,7 @@ async def _startup_event():
         except BaseException as e:
             logger.warning(f"Built-in MCP registration failed (non-critical): {type(e).__name__}: {e}")
         try:
-            await asyncio.wait_for(mcp_manager.connect_all_enabled(), timeout=20)
+            await mcp_manager.connect_all_enabled()
         except asyncio.TimeoutError:
             logger.warning("User MCP startup timed out (non-critical)")
         except BaseException as e:
@@ -959,57 +1162,70 @@ async def _startup_event():
 
     _startup_tasks.append(asyncio.create_task(_startup_mcp_connections()))
 
-    # Pre-warm the RAG tool index off the request path. Loading the local
-    # embedding model + opening ChromaDB + indexing the built-in tools is a
-    # one-time ~1-3s cost that otherwise lands on the user's FIRST message
-    # (showing up as a big `tool_selection` time). Doing it here makes the
-    # first turn as fast as subsequent ones (warm embed ≈ a few ms).
-    async def _warmup_tool_index():
-        try:
-            from src.tool_index import get_tool_index
-            idx = await asyncio.to_thread(get_tool_index)
-            if idx:
-                await asyncio.to_thread(idx.get_tools_for_query, "warmup", 8)
-                logger.info("[startup] Tool index pre-warmed")
-        except Exception as e:
-            logger.warning(f"Tool index warmup failed (non-critical): {type(e).__name__}: {e}")
+    # Semantic tool selection is part of the agent serving contract. Initialize
+    # it in a background thread by default so startup remains nonblocking while
+    # harness deployments can wait for the explicit readiness state.
+    from src.tool_index import prewarm_tool_index, tool_index_prewarm_enabled
+    if tool_index_prewarm_enabled():
+        async def _warmup_tool_index():
+            status = await asyncio.to_thread(prewarm_tool_index)
+            if status.get("ready"):
+                logger.info(
+                    "[startup] Tool index pre-warmed lanes=%s tools=%s duration_ms=%s",
+                    [lane.get("name") for lane in status.get("lanes", [])],
+                    status.get("builtin_tools"),
+                    status.get("duration_ms"),
+                )
+            else:
+                logger.warning(
+                    "Tool index warmup degraded (non-critical): %s",
+                    status.get("error_type") or status.get("state"),
+                )
 
-    _startup_tasks.append(asyncio.create_task(_warmup_tool_index()))
-    # Warmup: ping all known LLM endpoints to prime connections
-    async def _warmup_endpoints():
-        try:
-            import httpx
-            # model_discovery has no get_endpoints(); that call raised
-            # AttributeError every run and silently disabled warmup/keepalive.
-            # Resolve the /models probe URLs via the real discovery API, off the
-            # event loop since discovery does a blocking port scan.
-            urls = (
-                await asyncio.to_thread(model_discovery.warmup_ping_urls)
-                if model_discovery else []
-            )
-            for url in urls:
-                try:
-                    async with httpx.AsyncClient(timeout=5.0) as client:
-                        await client.get(url)
-                    logger.info(f"Warmup ping OK: {url}")
-                except Exception as e:
-                    logger.debug(f"Warmup ping failed for endpoint: {e}")
-        except Exception as e:
-            logger.debug(f"Warmup ping skipped: {e}")
+        _startup_tasks.append(asyncio.create_task(_warmup_tool_index()))
+    else:
+        logger.info("Tool index prewarm disabled (ODYSSEUS_TOOL_INDEX_PREWARM=0)")
 
-    _startup_tasks.append(asyncio.create_task(_warmup_endpoints()))
-
-    # Keep-alive: ping endpoints every 60 seconds to prevent cold starts
-    async def _keepalive_loop():
-        while True:
+    # Model endpoint pings remain opt-in. They can compete with the first seconds
+    # of UI use on slow or busy machines and are not required for local startup.
+    _startup_warmups_enabled = str(os.getenv("ODYSSEUS_STARTUP_WARMUPS", "")).lower() in {"1", "true", "yes", "on"}
+    if _startup_warmups_enabled:
+        async def _warmup_endpoints():
             try:
-                await asyncio.sleep(60)
-                await _warmup_endpoints()
+                import httpx
+                urls = (
+                    await asyncio.to_thread(model_discovery.warmup_ping_urls)
+                    if model_discovery else []
+                )
+                for url in urls:
+                    try:
+                        async with httpx.AsyncClient(timeout=5.0) as client:
+                            await client.get(url)
+                        logger.info(f"Warmup ping OK: {url}")
+                    except Exception as e:
+                        logger.debug(f"Warmup ping failed for endpoint: {e}")
             except Exception as e:
-                logger.warning(f"Keepalive loop error: {e}")
-                await asyncio.sleep(300)  # Back off on error
+                logger.debug(f"Warmup ping skipped: {e}")
 
-    _startup_tasks.append(asyncio.create_task(_keepalive_loop()))
+        _startup_tasks.append(asyncio.create_task(_warmup_endpoints()))
+    else:
+        logger.info("Model endpoint warmups disabled (set ODYSSEUS_STARTUP_WARMUPS=1 to enable)")
+
+    # Keep-alive is opt-in. The ping path performs model discovery, and when
+    # stale LAN endpoints are configured it can add periodic backend pressure
+    # that delays unrelated UI requests such as Notes/Documents.
+    _keepalive_enabled = str(os.getenv("ODYSSEUS_MODEL_KEEPALIVE", "")).lower() in {"1", "true", "yes", "on"}
+    if _keepalive_enabled:
+        async def _keepalive_loop():
+            while True:
+                try:
+                    await asyncio.sleep(60)
+                    await _warmup_endpoints()
+                except Exception as e:
+                    logger.warning(f"Keepalive loop error: {e}")
+                    await asyncio.sleep(300)  # Back off on error
+
+        _startup_tasks.append(asyncio.create_task(_keepalive_loop()))
 
     async def _ensure_default_tasks():
         # Create/reconcile default automation tasks + personal assistant for every user.
@@ -1061,6 +1277,14 @@ async def _startup_event():
     # Disk-backed skills are not covered by the DB legacy-owner sweep. Repair
     # ownerless or deleted/test-owner SKILL.md files so strict owner filtering
     # does not make an existing library look empty after auth/account changes.
+    try:
+        from services.memory.builtin_skills import install_builtin_skills
+        installed = install_builtin_skills(skills_manager, ())
+        if installed:
+            logger.info("Installed %s built-in skill file(s)", installed)
+    except Exception as e:
+        logger.debug(f"Built-in skill installation skipped: {e}")
+
     try:
         import json as _json
         auth_path = AUTH_FILE
@@ -1106,35 +1330,10 @@ async def _startup_event():
 
     _startup_tasks.append(asyncio.create_task(_null_owner_sweep_loop()))
 
-    # Nightly skill audit — at ~02:00 local, test + judge a batch of the
-    # least-recently-checked skills, auto-fixing/escalating weak ones (never
-    # deletes). Rotates through the library so each night covers different
-    # skills. Gated by the `skill_audit_nightly` setting (default on); hour via
-    # `skill_audit_hour` (default 2), batch size via `skill_audit_batch` (8).
-    async def _skill_audit_nightly_loop():
-        from datetime import timedelta
-        while True:
-            try:
-                from src.settings import get_setting
-                hour = int(get_setting("skill_audit_hour", 2) or 2)
-            except Exception:
-                hour = 2
-            now = datetime.now()
-            nxt = now.replace(hour=hour % 24, minute=0, second=0, microsecond=0)
-            if nxt <= now:
-                nxt += timedelta(days=1)
-            await asyncio.sleep(max(60, (nxt - now).total_seconds()))
-            try:
-                from src.settings import get_setting
-                if not get_setting("skill_audit_nightly", True):
-                    continue
-                batch = int(get_setting("skill_audit_batch", 8) or 8)
-                from routes.skills_routes import run_scheduled_skill_audit
-                await run_scheduled_skill_audit(skills_manager, owner=None, max_skills=batch)
-            except Exception as e:
-                logger.warning(f"Nightly skill audit failed: {e}")
-
-    _startup_tasks.append(asyncio.create_task(_skill_audit_nightly_loop()))
+    # Skills Audit is scheduled per owner by TaskScheduler. Do not also start
+    # an ownerless audit here: its sidecar results cannot be read back through
+    # an authenticated owner's skill namespace, and its model activity can
+    # defer the real per-owner task at the same time of night.
 
     # Cookbook serve lifecycle — kills scheduler-launched serves whose
     # window-end has passed. Paired with the cookbook_serve builtin
@@ -1149,6 +1348,18 @@ async def _startup_event():
 
 async def _shutdown_event():
     logger.info("Application shutting down...")
+    background_delivery = getattr(app.state, 'background_tool_delivery_task', None)
+    if background_delivery:
+        background_delivery.cancel()
+        try:
+            await background_delivery
+        except asyncio.CancelledError:
+            pass
+    try:
+        from src.agent_tools.web_tools import shutdown_private_browser_sessions
+        await shutdown_private_browser_sessions()
+    except Exception as e:
+        logger.warning(f"Private browser shutdown error: {e}")
     if upload_cleanup_task:
         upload_cleanup_task.cancel()
         try:
@@ -1171,3 +1382,12 @@ async def _shutdown_event():
     except Exception as e:
         logger.warning(f"MCP shutdown error: {e}")
     logger.info("Application shutdown complete")
+
+
+if __name__ == "__main__":
+    import uvicorn
+
+    bind_host = os.getenv("APP_BIND", "127.0.0.1")
+    bind_port = int(os.getenv("APP_PORT", "7011"))
+
+    uvicorn.run(app, host=bind_host, port=bind_port, log_level="info")

build-macos-app.sh

@@ -27,13 +27,13 @@ echo "  port:        $PORT"
 rm -rf "$APP"
 mkdir -p "$APP/Contents/MacOS" "$APP/Contents/Resources"
 
-# ── Icon (best effort) — center-crop docs/odysseus.jpg to a square .icns ──
-if [ -f "$REPO_DIR/docs/odysseus.jpg" ] && command -v sips >/dev/null 2>&1; then
+# ── Icon (best effort) — center-crop the branding image to a square .icns ──
+if [ -f "$REPO_DIR/assets/branding/odysseus.jpg" ] && command -v sips >/dev/null 2>&1; then
   TMPIMG="$(mktemp -d)"
   # Center-crop to a square, scale to 512 (sips' icns encoder caps at 512), and
   # let sips emit the .icns directly — more robust across macOS versions than
   # building an .iconset by hand.
-  sips -c 720 720 "$REPO_DIR/docs/odysseus.jpg" --out "$TMPIMG/sq.png" >/dev/null 2>&1 || cp "$REPO_DIR/docs/odysseus.jpg" "$TMPIMG/sq.png"
+  sips -c 720 720 "$REPO_DIR/assets/branding/odysseus.jpg" --out "$TMPIMG/sq.png" >/dev/null 2>&1 || cp "$REPO_DIR/assets/branding/odysseus.jpg" "$TMPIMG/sq.png"
   sips -z 512 512 "$TMPIMG/sq.png" --out "$TMPIMG/icon.png" >/dev/null 2>&1
   if sips -s format icns "$TMPIMG/icon.png" --out "$APP/Contents/Resources/odysseus.icns" >/dev/null 2>&1; then
     echo "  icon:        odysseus.icns"
@@ -42,7 +42,7 @@ if [ -f "$REPO_DIR/docs/odysseus.jpg" ] && command -v sips >/dev/null 2>&1; then
   fi
   rm -rf "$TMPIMG"
 else
-  echo "  icon:        (skipped — no docs/odysseus.jpg)"
+  echo "  icon:        (skipped — no assets/branding/odysseus.jpg)"
 fi
 
 # ── Info.plist ──
@@ -73,6 +73,10 @@ cat > "$APP/Contents/MacOS/$APP_NAME.tmpl" <<'LAUNCHER'
 INSTALL_DIR="__INSTALL_DIR__"
 PORT="__PORT__"
 URL="http://127.0.0.1:${PORT}"
+# uvicorn is started with --port below, but APP_PORT is what the app itself
+# reads when it needs to build a URL for this instance (internal_api_base(),
+# companion pairing, the MCP OAuth callback), so export it as well.
+export APP_PORT="$PORT"
 export PATH="/opt/homebrew/bin:/usr/local/bin:/usr/bin:/bin:$PATH"
 
 UVICORN="$INSTALL_DIR/venv/bin/uvicorn"

build-windows-portable.ps1

@@ -0,0 +1,72 @@
+#Requires -Version 5.1
+<#
+  Build a portable Windows distribution for Odysseus.
+
+  Output layout:
+    dist\Odysseus\Odysseus.exe
+    dist\Odysseus\static\...
+    dist\Odysseus\scripts\...
+    dist\Odysseus\mcp_servers\...
+    dist\Odysseus\services\hwfit\data\...
+
+  The app then keeps using its normal filesystem layout when frozen.
+
+  Usage:
+    powershell -ExecutionPolicy Bypass -File .\build-windows-portable.ps1
+#>
+
+$ErrorActionPreference = "Stop"
+Set-Location -Path $PSScriptRoot
+
+function Write-Step($msg) { Write-Host ""; Write-Host ("==> " + $msg) -ForegroundColor Cyan }
+function Fail($msg) {
+    Write-Host ""
+    Write-Host ("ERROR: " + $msg) -ForegroundColor Red
+    exit 1
+}
+
+Write-Step "Checking for Python"
+$pyExe = $null
+if (Test-Path ".\.venv\Scripts\python.exe") {
+    $pyExe = (Resolve-Path ".\.venv\Scripts\python.exe").Path
+} else {
+    foreach ($c in @("py", "python")) {
+        $cmd = Get-Command $c -ErrorAction SilentlyContinue
+        if ($cmd) { $pyExe = $cmd.Source; break }
+    }
+    if ($pyExe -like "*WindowsApps*python.exe") {
+        $pyCmd = Get-Command py -ErrorAction SilentlyContinue
+        if ($pyCmd) {
+            $pyExe = $pyCmd.Source
+        }
+    }
+}
+if (-not $pyExe) {
+    Fail "Python not found on PATH. Install Python 3.11+ first."
+}
+Write-Host ("Using Python: " + $pyExe)
+
+Write-Step "Installing build dependencies"
+& $pyExe -m pip install --upgrade pip --quiet
+& $pyExe -m pip install -r requirements.txt pyinstaller pystray Pillow
+if ($LASTEXITCODE -ne 0) { Fail "Dependency install failed." }
+
+Write-Step "Building portable exe bundle"
+Remove-Item -Recurse -Force build, dist -ErrorAction SilentlyContinue
+
+$dataArgs = @(
+    "--add-data", "static;static",
+    "--add-data", "scripts;scripts",
+    "--add-data", "mcp_servers;mcp_servers",
+    "--add-data", "services/hwfit/data;services/hwfit/data",
+    "--add-data", "config;config",
+    "--add-data", ".env.example;.env.example"
+)
+
+& $pyExe -m PyInstaller --noconfirm --clean --onedir --noconsole --icon=static/icon.ico --name Odysseus @dataArgs launcher.py
+if ($LASTEXITCODE -ne 0) { Fail "PyInstaller build failed." }
+
+Write-Host ""
+Write-Host "Build complete." -ForegroundColor Green
+Write-Host "Portable app folder: $PSScriptRoot\dist\Odysseus" -ForegroundColor Green
+Write-Host "Distribute the whole folder (or zip it) so static assets and scripts stay with the exe." -ForegroundColor Green

companion/pairing.py

@@ -6,11 +6,14 @@ units so the route layer stays thin and the logic is directly testable.
 
 from __future__ import annotations
 
+import ipaddress
 import json
 import os
+import re
 import secrets
 import socket
 import uuid
+from urllib.parse import urlsplit
 
 import bcrypt
 
@@ -20,6 +23,102 @@ PAIRING_VERSION = 1
 COMPANION_SCOPE = "chat"
 
 
+_COMPANION_IPV4_NETWORKS = tuple(
+    ipaddress.ip_network(cidr)
+    for cidr in (
+        "10.0.0.0/8",
+        "100.64.0.0/10",
+        "127.0.0.0/8",
+        "169.254.0.0/16",
+        "172.16.0.0/12",
+        "192.168.0.0/16",
+    )
+)
+_DNS_LABEL_RE = re.compile(r"[a-z0-9](?:[a-z0-9-]{0,61}[a-z0-9])?\Z")
+
+
+def _valid_companion_client_host(host: str) -> bool:
+    """Match the host forms supported by the current v1 Expo client."""
+    if not host or len(host) > 253 or not host.isascii() or "%" in host:
+        return False
+
+    try:
+        address = ipaddress.ip_address(host)
+    except ValueError:
+        labels = host.split(".")
+        if any(not _DNS_LABEL_RE.fullmatch(label) for label in labels):
+            return False
+        if any(label.startswith("xn--") for label in labels):
+            return False
+        # WHATWG URL parsers treat a decimal or ``0x`` single-label hostname
+        # as an IPv4 number even though Python's strict ``ipaddress`` parser
+        # rejects that spelling.  The v1 client interpolates this host back
+        # into a URL, so accepting e.g. ``134744072`` would make the phone send
+        # its bearer token to public 8.8.8.8.  Keep DNS labels unambiguous.
+        if len(labels) == 1 and (
+            labels[0].isdigit()
+            or re.fullmatch(r"0x[0-9a-f]*", labels[0]) is not None
+        ):
+            return False
+        return len(labels) == 1 or (len(labels) >= 2 and labels[-1] == "local")
+
+    return isinstance(address, ipaddress.IPv4Address) and any(
+        address in network for network in _COMPANION_IPV4_NETWORKS
+    )
+
+
+def parse_companion_base_url(value: str) -> tuple[str, int]:
+    """Validate a v1 companion address and return its legacy (host, port).
+
+    The deployed client understands only HTTP plus a LAN-style host and port.
+    Reject anything outside that exact contract instead of advertising a URL
+    the client would reject, downgrade, or interpret differently.
+    """
+    if not isinstance(value, str) or not value:
+        raise ValueError("COMPANION_BASE_URL must be a canonical HTTP LAN origin")
+    if not value.isascii():
+        raise ValueError("COMPANION_BASE_URL must contain only ASCII characters")
+    if any(
+        ord(char) <= 32 or ord(char) == 127 or char in {"\\", "%"}
+        for char in value
+    ):
+        raise ValueError(
+            "COMPANION_BASE_URL contains a forbidden character"
+        )
+
+    try:
+        parsed = urlsplit(value)
+        port = parsed.port
+    except ValueError as exc:
+        raise ValueError("COMPANION_BASE_URL must be a valid HTTP LAN origin") from exc
+
+    host = parsed.hostname
+    if parsed.scheme.lower() != "http" or not parsed.netloc or not host:
+        raise ValueError("COMPANION_BASE_URL must be a canonical HTTP LAN origin")
+    if parsed.username is not None or parsed.password is not None:
+        raise ValueError("COMPANION_BASE_URL must not contain credentials")
+    if parsed.path or parsed.query or parsed.fragment:
+        raise ValueError("COMPANION_BASE_URL must not contain a path, query, or fragment")
+    if port is not None and not 1 <= port <= 65535:
+        raise ValueError("COMPANION_BASE_URL port must be between 1 and 65535")
+    if not _valid_companion_client_host(host):
+        raise ValueError("COMPANION_BASE_URL host is not supported by companion v1")
+
+    netloc = f"{host}:{port}" if port is not None else host
+    origin = f"http://{netloc}"
+    if value != origin:
+        raise ValueError("COMPANION_BASE_URL must be a canonical HTTP LAN origin")
+    return host, port or 80
+
+
+def configured_companion_origin() -> tuple[str, int] | None:
+    """Return the validated operator-configured v1 address, if any."""
+    value = os.environ.get("COMPANION_BASE_URL")
+    if value is None or value == "":
+        return None
+    return parse_companion_base_url(value)
+
+
 def default_port() -> int:
     """Best guess at the port the server is reachable on. Callers that know the
     real request port should pass it explicitly."""

companion/routes.py

@@ -5,8 +5,9 @@ offers and pair to it, without duplicating any LLM logic.
 
 Auth is enforced globally by AuthMiddleware (app.py), so reaching a handler here
 means the caller is authenticated by either a cookie session or a Bearer `ody_`
-API token. The read endpoints (ping/info/models) accept either; the pairing
-endpoints are admin-cookie only.
+API token. Ping/info accept either credential type, models requires a chat-
+scoped API token for bearer callers, and the pairing endpoints are admin-cookie
+only.
 
 Pairing CSRF posture: minting happens ONLY on POST. The session cookie is
 SameSite=Lax (routes/auth_routes.py), which a browser does not send on a
@@ -18,11 +19,11 @@ on a GET would be unsafe (Lax cookies ride top-level GET navigations), so GET
 
 import html
 
-from fastapi import APIRouter, Request
+from fastapi import APIRouter, HTTPException, Request
 from fastapi.responses import HTMLResponse
 
 from core.middleware import require_admin
-from src.auth_helpers import get_current_user
+from src.auth_helpers import _auth_disabled, get_current_user
 
 from companion import pairing as _pairing
 
@@ -52,6 +53,18 @@ def owner_can_see(row_owner, owner) -> bool:
     return row_owner is None or row_owner == owner
 
 
+def require_models_scope(request: Request) -> None:
+    """Require the companion chat scope for bearer-token model inventory."""
+    if not getattr(request.state, "api_token", False):
+        return
+    scopes = getattr(request.state, "api_token_scopes", None) or []
+    if isinstance(scopes, str):
+        scopes = [scope.strip() for scope in scopes.split(",")]
+    scope_set = {str(scope).strip() for scope in scopes if str(scope).strip()}
+    if _pairing.COMPANION_SCOPE not in scope_set:
+        raise HTTPException(403, "API token requires chat scope")
+
+
 def mint_pairing_token(owner: str, invalidate=None) -> tuple[str, str]:
     """Mint a pairing token AND invalidate the auth middleware's in-memory token
     cache, so the new token is accepted on the very next request without a server
@@ -100,15 +113,22 @@ def setup_companion_routes() -> APIRouter:
         The stock /api/models route scopes to get_current_user, which for a
         bearer token is the sandboxed pseudo-user "api" (owns nothing). Here we
         scope to the token's real owner instead, plus legacy null-owner shared
-        rows -- the same rule as owner_filter. Read-only; never returns api_key
-        material.
+        rows -- the same rule as owner_filter. Explicit auth-disabled mode keeps
+        the stock route's single-user all-endpoints view. Read-only; never
+        returns api_key material.
         """
+        require_models_scope(request)
         import json as _json
 
         from core.database import SessionLocal, ModelEndpoint
         from src.endpoint_resolver import build_chat_url
 
         owner = token_owner(request)
+        single_user_mode = (
+            owner is None
+            and not getattr(request.state, "api_token", False)
+            and _auth_disabled()
+        )
         out = []
         db = SessionLocal()
         try:
@@ -119,7 +139,7 @@ def setup_companion_routes() -> APIRouter:
             if owner:
                 q = q.filter((ModelEndpoint.owner == owner) | (ModelEndpoint.owner == None))  # noqa: E711
             for ep in q.all():
-                if not owner_can_see(ep.owner, owner):
+                if not single_user_mode and not owner_can_see(ep.owner, owner):
                     continue
                 try:
                     model_ids = _json.loads(ep.cached_models) if ep.cached_models else []
@@ -180,19 +200,27 @@ def setup_companion_routes() -> APIRouter:
         the code works immediately, no restart. `?format=json` returns the
         payload for an in-app pairing screen."""
         require_admin(request)
+        try:
+            configured_origin = _pairing.configured_companion_origin()
+        except ValueError as exc:
+            raise HTTPException(500, str(exc)) from None
         owner = get_current_user(request)
         invalidate = getattr(request.app.state, "invalidate_token_cache", None)
         token_id, raw_token = mint_pairing_token(owner, invalidate)
 
-        hosts = _pairing.lan_ip_candidates()
-        host = hosts[0] if hosts else "127.0.0.1"
-        port = request.url.port or _pairing.default_port()
+        if configured_origin:
+            host, port = configured_origin
+            hosts = [host]
+        else:
+            hosts = _pairing.lan_ip_candidates()
+            host = hosts[0] if hosts else "127.0.0.1"
+            port = request.url.port or _pairing.default_port()
         payload = _pairing.pairing_payload(host, port, raw_token)
         qr = _pairing.pairing_qr_png_data_uri(payload)
         qr_ok = bool(qr and qr.startswith("data:image/png;base64,"))
 
         if (request.query_params.get("format") or "").lower() == "json":
-            return {
+            response = {
                 "host": host,
                 "port": port,
                 "token": raw_token,
@@ -201,6 +229,7 @@ def setup_companion_routes() -> APIRouter:
                 "payload": payload,
                 "qr": qr if qr_ok else None,
             }
+            return response
 
         import json as _json
         payload_json = _json.dumps(payload, separators=(",", ":"))

core/atomic_io.py

@@ -15,29 +15,53 @@ from __future__ import annotations
 
 import json
 import os
+import uuid
 from typing import Any, Optional
 
 
 def atomic_write_json(path: str, data: Any, *, indent: Optional[int] = None) -> None:
     """Atomically persist `data` as JSON at `path`.
 
-    The temp file uses the live PID as a suffix so two processes saving the
-    same file (e.g. unit tests) don't collide on the rename target.
+    The temp file uses a random suffix so two concurrent writers saving the
+    same file don't collide on the rename target. A PID suffix does not do
+    this: the PID is constant for the life of a process, so two writers on
+    the same path within one process (or one single-process container, where
+    the PID never changes at all) still race for the same temp file.
     """
     os.makedirs(os.path.dirname(path) or ".", exist_ok=True)
-    tmp = f"{path}.tmp.{os.getpid()}"
-    with open(tmp, "w", encoding="utf-8") as f:
-        json.dump(data, f, indent=indent)
-        f.flush()
-        os.fsync(f.fileno())
-    os.replace(tmp, path)
+    tmp = f"{path}.tmp.{uuid.uuid4().hex}"
+
+    try:
+        with open(tmp, "w", encoding="utf-8") as f:
+            json.dump(data, f, indent=indent)
+            f.flush()
+            os.fsync(f.fileno())
+        os.replace(tmp, path)
+    finally:
+        # Directly unlink to avoid a check-then-act race condition.
+        # Swallows FileNotFoundError (on success path) and other cleanup OSErrors.
+        try:
+            os.unlink(tmp)
+        except OSError:
+            pass
 
 
 def atomic_write_text(path: str, text: str) -> None:
+    if not isinstance(text, str):
+        raise TypeError("atomic_write_text expects a string")
     os.makedirs(os.path.dirname(path) or ".", exist_ok=True)
-    tmp = f"{path}.tmp.{os.getpid()}"
-    with open(tmp, "w", encoding="utf-8") as f:
-        f.write(text)
-        f.flush()
-        os.fsync(f.fileno())
-    os.replace(tmp, path)
+    tmp = f"{path}.tmp.{uuid.uuid4().hex}"
+
+    try:
+        with open(tmp, "w", encoding="utf-8") as f:
+            f.write(text)
+            f.flush()
+            os.fsync(f.fileno())
+        os.replace(tmp, path)
+    finally:
+        # Directly unlink to avoid a check-then-act race condition.
+        # Swallows FileNotFoundError (on success path) and other cleanup OSErrors.
+        try:
+            os.unlink(tmp)
+        except OSError:
+            pass

core/auth.py

@@ -47,25 +47,19 @@ ADMIN_PRIVILEGES["allowed_models_restricted"] = False
 # backwards for this sentinel.
 ADMIN_PRIVILEGES["block_all_models"] = False
 
-from src.constants import AUTH_FILE
+from src.constants import AUTH_FILE, PASSWORD_MIN_LENGTH
+from src.owner_identity import RESERVED_AUTH_USERNAMES
 DEFAULT_AUTH_PATH = AUTH_FILE
 TOKEN_TTL = 60 * 60 * 24 * 7  # 7 days
 
-# Usernames the auth + middleware layer reserve as internal "synthetic owner"
-# sentinels; they must never belong to a real account. The most dangerous is
-# "internal-tool": `core.middleware.require_admin` treats any request whose
-# `current_user == "internal-tool"` as the in-process tool loopback and grants
-# admin, and because the cookie auth path sets `current_user` to the raw
-# username, an account literally named "internal-tool" would be silently
-# treated as an admin by every `require_admin`-gated route. "api" collides with
-# the bearer-token owner-attribution sentinel. "demo"/"system" round out the
-# synthetic-owner set the rest of the codebase already special-cases (see
-# `_SYNTHETIC_OWNERS` in routes/assistant_routes.py and the matching guards in
-# src/task_scheduler.py / routes/research_routes.py) — a real account with one
-# of those names would be denied an assistant and inconsistently owner-scoped.
-# Refuse to create or rename into any of them so the sentinels can't be
-# impersonated. (Keep this in sync with that synthetic-owner set.)
-RESERVED_USERNAMES = frozenset({"internal-tool", "api", "demo", "system"})
+# Usernames the auth + middleware layer reserves for request sentinels and
+# internal storage owners; they must never belong to a real login account.
+# "internal-tool" is the most dangerous because `core.middleware.require_admin`
+# treats it as the in-process tool loopback. "api" collides with bearer-token
+# attribution. "demo"/"system" are synthetic owners already special-cased by
+# scheduler/assistant/research paths. The Default/Local owner is a storage
+# bucket for explicit auth-disabled no-login mode, not a login username.
+RESERVED_USERNAMES = frozenset(RESERVED_AUTH_USERNAMES)
 
 
 def normalize_known_username(users: Dict[str, Any], username: str | None) -> Optional[str]:
@@ -175,16 +169,17 @@ class AuthManager:
                 )
                 old_user = "admin"
             old_hash = self._config["password_hash"]
-            self._config = {
-                "users": {
-                    old_user: {
-                        "password_hash": old_hash,
-                        "created": time.time(),
-                        "is_admin": True,
+            with self._config_lock:
+                self._config = {
+                    "users": {
+                        old_user: {
+                            "password_hash": old_hash,
+                            "created": time.time(),
+                            "is_admin": True,
+                        }
                     }
                 }
-            }
-            self._save()
+                self._save()
             logger.info(f"Migrated single-user auth to multi-user (admin: {old_user})")
 
     def _drop_reserved_loaded_users(self):
@@ -203,8 +198,9 @@ class AuthManager:
                 continue
             normalized[key] = data
         if removed or normalized != users:
-            self._config["users"] = normalized
-            self._save()
+            with self._config_lock:
+                self._config["users"] = normalized
+                self._save()
         if removed:
             logger.warning(
                 "Removed reserved username(s) from auth config: %s",
@@ -243,6 +239,15 @@ class AuthManager:
     def is_configured(self) -> bool:
         return len(self.users) > 0
 
+    def policy(self) -> dict:
+        """Return public auth policy constants for the frontend."""
+        return {
+            "password_min_length": PASSWORD_MIN_LENGTH,
+            "reserved_usernames": sorted(RESERVED_USERNAMES),
+            "signup_enabled": self.signup_enabled,
+            "session_days": TOKEN_TTL // 86400,
+        }
+
     # ------------------------------------------------------------------
     # Account management
     # ------------------------------------------------------------------
@@ -573,16 +578,20 @@ class AuthManager:
             return None
         return self.create_session_trusted(username)
 
-    def create_session_trusted(self, username: str) -> str:
+    def create_session_trusted(self, username: str) -> Optional[str]:
         """Issue a session token for an already-verified user.
         Call only after verify_password (and TOTP if enabled) have passed."""
         username = username.strip().lower()
         token = secrets.token_hex(32)
-        with self._sessions_lock:
-            self._sessions[token] = {
-                "username": username,
-                "expiry": time.time() + TOKEN_TTL,
-            }
+        with self._config_lock:
+            if username not in self.users:
+                logger.warning("Refused to issue session for missing user '%s'", username)
+                return None
+            with self._sessions_lock:
+                self._sessions[token] = {
+                    "username": username,
+                    "expiry": time.time() + TOKEN_TTL,
+                }
         self._save_sessions()
         return token
 

core/exceptions.py

@@ -1,4 +1,4 @@
-# src/exceptions.py
+# core/exceptions.py
 """Custom exceptions for the application."""
 
 class SessionNotFoundError(Exception):

core/log_safety.py

@@ -0,0 +1,27 @@
+"""Helpers for keeping sensitive data out of logs.
+
+Endpoint URLs configured by admins can embed credentials in the userinfo
+(``https://user:pass@host``) or query string (``?api_key=...``). Logging them
+raw leaks those secrets, so route/diagnostic logs run URLs through
+``redact_url`` first. Reconstructing the URL without userinfo/query/fragment
+also doubles as a sanitizer barrier for CodeQL's clear-text-logging query.
+"""
+
+from urllib.parse import urlparse, urlunparse
+
+
+def redact_url(url: str) -> str:
+    """Return a URL safe for logs by removing userinfo and query/fragment.
+
+    Keeps scheme, host, port and path so logs stay useful for debugging.
+    """
+    try:
+        parsed = urlparse(url or "")
+        host = parsed.hostname or ""
+        if ":" in host:  # IPv6 literal — re-bracket so host:port stays unambiguous
+            host = f"[{host}]"
+        if parsed.port:
+            host = f"{host}:{parsed.port}"
+        return urlunparse((parsed.scheme, host, parsed.path, "", "", ""))
+    except Exception:
+        return "<endpoint>"

core/middleware.py

@@ -3,10 +3,14 @@
 
 import os
 import secrets
+from collections.abc import Mapping
 
 from fastapi import HTTPException, Request
 from starlette.middleware.base import BaseHTTPMiddleware
 from starlette.responses import Response
+from starlette.routing import get_route_path
+
+from src.owner_identity import INTERNAL_TOOL_USER, auth_disabled
 
 
 # Per-process token that lets the in-app tool layer hit admin-gated
@@ -17,6 +21,30 @@ INTERNAL_TOOL_TOKEN = os.environ.get("ODYSSEUS_INTERNAL_TOKEN") or secrets.token
 INTERNAL_TOOL_HEADER = "X-Odysseus-Internal-Token"
 
 
+def get_application_route_path(scope: Mapping[str, object]) -> str:
+    """Return the application-relative path used by Starlette routing.
+
+    Uvicorn prefixes ``scope["path"]`` with a configured ASGI ``root_path``;
+    Starlette removes that prefix before matching routes. Middleware policy
+    must use the same path form or a deployment prefix can change which policy
+    applies to an otherwise unchanged application route.
+    """
+    return get_route_path(scope)
+
+
+def with_asgi_root_path(scope: Mapping[str, object], path: str) -> str:
+    """Prefix an application path for a client-facing redirect target."""
+    root_path = scope.get("root_path", "")
+    if not isinstance(root_path, str) or not root_path:
+        return path
+    return f"{root_path.rstrip('/')}{path}"
+
+
+def path_is_route_or_child(path: str, prefix: str) -> bool:
+    """Return whether ``path`` is exactly ``prefix`` or below that route."""
+    return path == prefix or path.startswith(prefix + "/")
+
+
 def is_cors_preflight(method: str, headers) -> bool:
     """True for a genuine CORS preflight: an OPTIONS request carrying the
     Access-Control-Request-Method header. Such requests are credential-less by
@@ -39,13 +67,13 @@ def require_admin(request: Request):
         hdr = request.headers.get(INTERNAL_TOOL_HEADER)
         if hdr and secrets.compare_digest(hdr, INTERNAL_TOOL_TOKEN):
             return
-        if getattr(request.state, "current_user", None) == "internal-tool":
+        if getattr(request.state, "current_user", None) == INTERNAL_TOOL_USER:
             return
     except Exception:
         pass
 
     auth_mgr = getattr(request.app.state, "auth_manager", None)
-    if os.getenv("AUTH_ENABLED", "true").lower() == "false":
+    if auth_disabled():
         return
     if not auth_mgr or not auth_mgr.is_configured:
         raise HTTPException(403, "Admin only")
@@ -65,10 +93,9 @@ class SecurityHeadersMiddleware(BaseHTTPMiddleware):
         response = await call_next(request)
         path = request.url.path
 
-        # Tool render endpoints are served inside iframes — allow framing by self
+        # Tool render endpoints
         is_tool_render = path.startswith("/api/tools/") and path.endswith("/render")
-        # PDF previews are embedded by the in-app document library. Keep the
-        # exception route-scoped so normal app pages remain unframeable.
+        # Document library PDF preview endpoint
         is_document_pdf_preview = path.startswith("/api/document/") and path.endswith("/render-pdf")
         # Visual report pages are self-contained HTML — need inline scripts + external images
         is_report = path.startswith("/api/research/report/")
@@ -95,9 +122,7 @@ class SecurityHeadersMiddleware(BaseHTTPMiddleware):
                 "frame-ancestors 'none'"
             )
         elif is_tool_render:
-            # Tool iframe content: skip all framing headers — the iframe's
-            # sandbox="allow-scripts" attribute provides isolation.
-            # Don't overwrite the route's own restrictive CSP either.
+            # Skip framing headers for tools.
             pass
         elif is_document_pdf_preview:
             response.headers["X-Frame-Options"] = "SAMEORIGIN"
@@ -118,7 +143,7 @@ class SecurityHeadersMiddleware(BaseHTTPMiddleware):
                 f"script-src 'self' 'nonce-{nonce}' https://cdn.jsdelivr.net; "
                 "style-src 'self' 'unsafe-inline' https://cdn.jsdelivr.net; "
                 "font-src 'self' https://cdn.jsdelivr.net; "
-                "img-src 'self' data: blob:; "
+                "img-src 'self' data: blob: https:; "
                 "media-src 'self' blob:; "
                 "connect-src 'self'; "
                 "frame-src 'self'; "

core/models.py

@@ -8,6 +8,11 @@ These are simple datacontainers. All persistence is handled by SessionManager.
 from dataclasses import dataclass
 from typing import Dict, List, Any, Optional, TYPE_CHECKING
 
+from src.tool_approval_scopes import (
+    CHAT_SESSION_APPROVAL_CONTEXT_MARKER,
+    CHAT_SESSION_APPROVAL_DECISION,
+)
+
 if TYPE_CHECKING:
     from .session_manager import SessionManager
 
@@ -31,6 +36,35 @@ set_session_manager = set_session_manager_instance
 get_session_manager = get_session_manager_instance
 
 
+def _history_grants_chat_session_approval(
+    history: List["ChatMessage"],
+    session_id: str,
+) -> bool:
+    """Return whether this exact chat has a resolved session-scope grant."""
+
+    expected_session = str(session_id or "")
+    if not expected_session:
+        return False
+    for message in reversed(history or []):
+        metadata = getattr(message, "metadata", None)
+        if not isinstance(metadata, dict):
+            continue
+        tool_events = metadata.get("tool_events")
+        if not isinstance(tool_events, list):
+            continue
+        for event in reversed(tool_events):
+            ask_user = event.get("ask_user") if isinstance(event, dict) else None
+            if not isinstance(ask_user, dict):
+                continue
+            if (
+                ask_user.get("kind") == "tool_approval"
+                and ask_user.get("resolved") == CHAT_SESSION_APPROVAL_DECISION
+                and str(ask_user.get("session_id") or "") == expected_session
+            ):
+                return True
+    return False
+
+
 @dataclass
 class ChatMessage:
     """A single chat message."""
@@ -74,6 +108,12 @@ class Session:
     owner: Optional[str] = None
     is_important: bool = False
     message_count: int = 0
+    memory_extraction_enabled: bool = True
+    skill_injection_enabled: bool = True
+    thinking_mode: str = "off"
+    temperature_override: Optional[float] = None
+    max_tokens_override: Optional[int] = None
+    cwd: Optional[str] = None
 
     def __post_init__(self):
         if self.headers is None:
@@ -116,11 +156,45 @@ class Session:
         the model. Display/history-load paths use the raw ``history`` and are
         unaffected.
         """
-        return [
+        messages = [
             msg.to_dict()
             for msg in self.history
             if (msg.metadata or {}).get("source") != "slash"
         ]
+        from src.background_tool_jobs import background_result_context
+        messages = [part for message in messages for part in (
+            *background_result_context(message.get('metadata')), message,
+        )]
+        # Resume an interrupted thinking-only response from its actual model
+        # reasoning channel. Restrict this to the latest assistant message so
+        # old traces do not accumulate in context or cause reasoning loops.
+        for index in range(len(messages) - 1, -1, -1):
+            message = messages[index]
+            if message.get("role") != "assistant":
+                continue
+            metadata = message.get("metadata") or {}
+            thinking = str(metadata.get("thinking") or "").strip()
+            if metadata.get("stopped") and thinking:
+                resumed = dict(message)
+                resumed["reasoning_content"] = thinking
+                messages[index] = resumed
+            break
+        if not _history_grants_chat_session_approval(self.history, self.id):
+            return messages
+
+        # Keep the grant close to the latest user request so route-neutral
+        # compaction/trimming preserves it. Copy the metadata instead of
+        # mutating the durable transcript object.
+        for index in range(len(messages) - 1, -1, -1):
+            if messages[index].get("role") != "user":
+                continue
+            message = dict(messages[index])
+            metadata = dict(message.get("metadata") or {})
+            metadata[CHAT_SESSION_APPROVAL_CONTEXT_MARKER] = True
+            message["metadata"] = metadata
+            messages[index] = message
+            break
+        return messages
 
     def get(self, key: str, default=None):
         """Dict-like access for compatibility."""

core/session_manager.py

@@ -14,8 +14,12 @@ import logging
 from datetime import datetime, timezone, timedelta
 from typing import Dict, Optional
 
+from sqlalchemy import func
+
 from .database import Session as DbSession, ChatMessage as DbChatMessage, Document as DbDocument, SessionLocal, utcnow_naive
 from .models import Session, ChatMessage
+from src.attachment_refs import persistable_message_content
+from src.upload_handler import reserve_message_upload_references
 
 # Re-export singleton accessors from models for convenience
 from .models import set_session_manager_instance, get_session_manager_instance
@@ -40,7 +44,18 @@ def _parse_msg_content(raw):
     if isinstance(raw, str) and raw.startswith('[{') and '"type"' in raw:
         try:
             parsed = json.loads(raw)
-            if isinstance(parsed, list) and all(isinstance(p, dict) for p in parsed):
+            # Only treat as serialized multimodal content when EVERY element is
+            # a dict whose "type" is a recognized content-block kind. Otherwise a
+            # plain text message that merely *looks* like a JSON array of objects
+            # (e.g. a user pasting an API schema/sample with a "type" field) was
+            # silently parsed back into a list, destroying the original string.
+            _BLOCK_TYPES = {
+                "text", "image", "image_url", "audio", "input_audio",
+                "input_image", "document", "file",
+            }
+            if (isinstance(parsed, list) and parsed
+                    and all(isinstance(p, dict) and p.get("type") in _BLOCK_TYPES
+                            for p in parsed)):
                 return parsed
         except (json.JSONDecodeError, ValueError):
             pass
@@ -61,6 +76,7 @@ class SessionManager:
     def __init__(self, sessions_file: str = None):
         # sessions_file kept for backward compat, not used
         self.sessions: Dict[str, Session] = {}
+        self.upload_handler = None
         self.load_sessions()
 
     # ------------------------------------------------------------------
@@ -78,14 +94,28 @@ class SessionManager:
         try:
             db_sessions = db.query(DbSession).filter(
                 DbSession.archived == False,
-                DbSession.message_count > 0,
+                DbSession.messages.any(),
             ).order_by(DbSession.last_accessed.desc()).limit(100).all()
 
+            # message_count is derived metadata and can drift after interrupted
+            # or legacy writes. Count only the bounded discovery set so startup
+            # remains metadata-only while lazy hydration sees an authoritative
+            # positive count for every discovered non-empty session.
+            message_counts = {}
+            if db_sessions:
+                message_counts = dict(
+                    db.query(DbChatMessage.session_id, func.count(DbChatMessage.id))
+                    .filter(DbChatMessage.session_id.in_([row.id for row in db_sessions]))
+                    .group_by(DbChatMessage.session_id)
+                    .all()
+                )
+
             loaded_count = 0
             for db_session in db_sessions:
                 try:
                     session = self._db_to_session_meta(db_session)
                     if session is not None:
+                        session.message_count = message_counts[db_session.id]
                         self.sessions[db_session.id] = session
                         loaded_count += 1
                 except Exception as e:
@@ -120,6 +150,12 @@ class SessionManager:
             history=[],
             owner=getattr(db_session, "owner", None),
             is_important=getattr(db_session, "is_important", False) or False,
+            memory_extraction_enabled=getattr(db_session, "memory_extraction_enabled", True) is not False,
+            skill_injection_enabled=getattr(db_session, "skill_injection_enabled", True) is not False,
+            thinking_mode=getattr(db_session, "thinking_mode", "") or "off",
+            temperature_override=getattr(db_session, "temperature_override", None),
+            max_tokens_override=getattr(db_session, "max_tokens_override", None),
+            cwd=getattr(db_session, "cwd", None) or None,
         )
         session.message_count = getattr(db_session, "message_count", 0) or 0
         return session
@@ -178,9 +214,20 @@ class SessionManager:
             history=history,
             owner=getattr(db_session, 'owner', None),
             is_important=getattr(db_session, 'is_important', False) or False,
+            memory_extraction_enabled=getattr(db_session, 'memory_extraction_enabled', True) is not False,
+            skill_injection_enabled=getattr(db_session, 'skill_injection_enabled', True) is not False,
+            thinking_mode=getattr(db_session, "thinking_mode", "") or "off",
+            temperature_override=getattr(db_session, "temperature_override", None),
+            max_tokens_override=getattr(db_session, "max_tokens_override", None),
+            cwd=getattr(db_session, "cwd", None) or None,
         )
 
-        session.message_count = getattr(db_session, 'message_count', len(history))
+        # The rows just loaded are the whole transcript, so they — not the
+        # denormalized sessions.message_count column — are the truth for this
+        # cached object. get_session's hydration gate compares against this
+        # number; seeding it from a drifted column would ask for a reload that
+        # can never close the gap.
+        session.message_count = len(history)
         return session
 
     # ------------------------------------------------------------------
@@ -219,17 +266,26 @@ class SessionManager:
                 logger.warning("Dropping message for deleted session %s", session_id)
                 return
 
+            missing_upload_id = reserve_message_upload_references(
+                getattr(self, "upload_handler", None),
+                getattr(db_session, "owner", None),
+                message.content,
+                message.metadata,
+            )
+            if missing_upload_id:
+                raise ValueError(
+                    f"Referenced upload is no longer available: {missing_upload_id}"
+                )
+
             msg_id = str(uuid.uuid4())
             msg_time = datetime.utcnow()
             if message.metadata is None:
                 message.metadata = {}
             message.metadata.setdefault('timestamp', _message_timestamp_iso(msg_time))
-            # Multimodal content (image/audio attachments) is a list — serialize
-            # to JSON so the Text column can store it.  On reload, _db_to_session
-            # detects the JSON-array prefix and parses it back.
-            _content = message.content
-            if isinstance(_content, list):
-                _content = json.dumps(_content)
+            # Multimodal content may contain provider data URLs for the live
+            # model call. Persist only readable text plus attachment references
+            # so chat_messages/FTS do not duplicate upload bytes.
+            _content = persistable_message_content(message.content, message.metadata)
             db_message = DbChatMessage(
                 id=msg_id,
                 session_id=session_id,
@@ -311,6 +367,28 @@ class SessionManager:
         session = self.get_session(session_id)
         db = SessionLocal()
         try:
+            db_session = db.query(DbSession).filter(DbSession.id == session_id).first()
+            if db_session is None:
+                logger.warning("Cannot replace history for missing session %s", session_id)
+                return False
+
+            # Reserve every incoming attachment before removing any durable
+            # message row. reserve_upload() shares the upload lifecycle lock
+            # with cleanup, so an upload cannot be deleted between this
+            # ownership check/access touch and the replacement transaction.
+            # A failed reservation must leave the existing transcript intact.
+            for message in messages:
+                missing_upload_id = reserve_message_upload_references(
+                    getattr(self, "upload_handler", None),
+                    getattr(db_session, "owner", None),
+                    message.content,
+                    message.metadata,
+                )
+                if missing_upload_id:
+                    raise ValueError(
+                        f"Referenced upload is no longer available: {missing_upload_id}"
+                    )
+
             db.query(DbChatMessage).filter(DbChatMessage.session_id == session_id).delete()
             now = datetime.now(timezone.utc)
             for i, message in enumerate(messages):
@@ -319,15 +397,9 @@ class SessionManager:
                     id=msg_id,
                     session_id=session_id,
                     role=message.role,
-                    # Multimodal content (image/audio attachments) is a list;
-                    # serialize to JSON so the Text column round-trips via
-                    # _parse_msg_content. Storing the raw list let SQLAlchemy
-                    # bind its single-quoted repr, which _parse_msg_content
-                    # cannot parse (it looks for double-quoted "type"), so the
-                    # attachment was destroyed on reload. Mirrors _persist_message.
-                    content=(json.dumps(message.content)
-                             if isinstance(message.content, list)
-                             else message.content),
+                    # Mirrors _persist_message: keep raw media bytes out of the
+                    # persisted transcript and search index.
+                    content=persistable_message_content(message.content, message.metadata),
                     meta_data=json.dumps(message.metadata) if message.metadata else None,
                     timestamp=now + timedelta(microseconds=i),
                 )
@@ -336,12 +408,10 @@ class SessionManager:
                     message.metadata = {}
                 message.metadata["_db_id"] = msg_id
 
-            db_session = db.query(DbSession).filter(DbSession.id == session_id).first()
-            if db_session:
-                db_session.message_count = len(messages)
-                db_session.updated_at = now
-                db_session.last_accessed = now
-                db_session.last_message_at = now
+            db_session.message_count = len(messages)
+            db_session.updated_at = now
+            db_session.last_accessed = now
+            db_session.last_message_at = now
 
             db.commit()
             session.history = list(messages)
@@ -361,30 +431,50 @@ class SessionManager:
     # ------------------------------------------------------------------
 
     def get_session(self, session_id: str) -> Session:
-        """Get a session by ID, loading from DB if needed.
+        """Get a session by ID, loading complete DB history when needed.
 
-        Sessions seeded by `load_sessions` start with empty history. The
-        first read here hydrates them with the message rows.
+        Sessions seeded by ``load_sessions`` start with empty history, and a
+        cached session can also become partially stale. Refresh metadata first,
+        then hydrate whenever the cached transcript is short of the stored rows.
+        Model-send routes enter through this method before building context,
+        while paginated display history reads SQLite directly.
+
+        The gate compares against ``sync_session_metadata``'s reconciled count
+        (the real ``chat_messages`` total), never the denormalized column, so a
+        hydrate always closes the gap and the next read is a cache hit.
         """
         if session_id not in self.sessions:
             self._load_session_from_db(session_id)
-        else:
-            cached = self.sessions[session_id]
-            # Lazy hydrate: metadata-only entries get their messages on first read.
-            if not cached.history and getattr(cached, "message_count", 0) > 0:
-                self._load_session_from_db(session_id)
 
         # Keep model/endpoint metadata fresh. Endpoint deletion can clear the
-        # DB row while a session object is still cached in RAM.
+        # DB row while a session object is still cached in RAM. Refreshing first
+        # also exposes the authoritative message count before completeness is
+        # checked.
         self.sync_session_metadata(session_id)
 
+        cached = self.sessions[session_id]
+        cached_count = len(cached.history or [])
+        stored_count = int(getattr(cached, "message_count", 0) or 0)
+        if cached_count < stored_count:
+            self._load_session_from_db(session_id)
+
         # Update last_accessed
         self._touch_session(session_id)
 
         return self.sessions[session_id]
 
     def sync_session_metadata(self, session_id: str) -> bool:
-        """Refresh non-message session fields from the DB into the cached object."""
+        """Refresh non-message session fields from the DB into the cached object.
+
+        ``message_count`` is reconciled against the real ``chat_messages`` rows
+        rather than copied from the denormalized ``sessions.message_count``
+        column. That column drifts in normal operation — ``_persist_message``
+        swallows a failed insert but ``add_message`` has already appended in
+        memory, so the next successful persist writes rows+1, and a persist for
+        an uncached session writes 0. Hydration keys off this number: a
+        drifted-high column would reload the whole transcript on every warm
+        read, and a drifted-low one would leave the model a truncated one.
+        """
         session = self.sessions.get(session_id)
         if session is None:
             return False
@@ -407,7 +497,12 @@ class SessionManager:
             session.archived = db_session.archived
             session.owner = getattr(db_session, "owner", None)
             session.is_important = getattr(db_session, "is_important", False) or False
-            session.message_count = getattr(db_session, "message_count", session.message_count) or 0
+            session.cwd = getattr(db_session, "cwd", None) or None
+            session.message_count = (
+                db.query(DbChatMessage)
+                .filter(DbChatMessage.session_id == session_id)
+                .count()
+            )
             return True
         except Exception as e:
             logger.error(f"Error syncing session metadata {session_id}: {e}")
@@ -463,9 +558,12 @@ class SessionManager:
         endpoint_url: str,
         model: str,
         rag: bool = False,
-        owner: str = None
+        owner: str = None,
+        cwd: str = None,
+        headers: Optional[Dict[str, str]] = None,
     ) -> Session:
         """Create a new session and save to database."""
+        session_headers = dict(headers or {})
         db = SessionLocal()
         try:
             db_session = DbSession(
@@ -474,8 +572,9 @@ class SessionManager:
                 endpoint_url=endpoint_url,
                 model=model,
                 rag=rag,
-                headers={},
+                headers=session_headers,
                 owner=owner,
+                cwd=cwd or None,
                 created_at=datetime.now(timezone.utc),
                 updated_at=datetime.now(timezone.utc)
             )
@@ -488,8 +587,9 @@ class SessionManager:
                 endpoint_url=endpoint_url,
                 model=model,
                 rag=rag,
-                headers={},
+                headers=session_headers,
                 owner=owner,
+                cwd=cwd or None,
             )
 
             self.sessions[session_id] = session
@@ -506,6 +606,12 @@ class SessionManager:
         """Permanently delete a session and all its messages."""
         db = SessionLocal()
         try:
+            try:
+                from src.session_image_cleanup import cleanup_session_images
+                cleanup_session_images(session_id, db=db)
+            except Exception as e:
+                logger.warning(f"Image cleanup failed while deleting session {session_id}: {e}")
+
             # Detach documents so they survive as orphans in the library
             db.query(DbDocument).filter(DbDocument.session_id == session_id).update(
                 {DbDocument.session_id: None}, synchronize_session=False

docker-compose.gpu-amd.yml

@@ -14,7 +14,7 @@ services:
   odysseus:
     build: .
     ports:
-      - "${APP_BIND:-127.0.0.1}:${APP_PORT:-7000}:7000"
+      - "${APP_BIND:-127.0.0.1}:${APP_PORT:-7011}:7000"
     volumes:
       - ${APP_DATA_DIR:-./data}:/app/data:z
       - ${APP_LOGS_DIR:-./logs}:/app/logs:z
@@ -46,10 +46,11 @@ services:
       - DATABASE_URL=${DATABASE_URL:-sqlite:///./data/app.db}
       - AUTH_ENABLED=${AUTH_ENABLED:-true}
       - LOCALHOST_BYPASS=${LOCALHOST_BYPASS:-false}
+      - COMPANION_BASE_URL=${COMPANION_BASE_URL:-}
       - ODYSSEUS_ADMIN_USER=${ODYSSEUS_ADMIN_USER:-admin}
       - ODYSSEUS_ADMIN_PASSWORD=${ODYSSEUS_ADMIN_PASSWORD:-}
       - ALLOWED_ORIGINS=${ALLOWED_ORIGINS:-http://localhost,http://127.0.0.1}
-      - SECURE_COOKIES=${SECURE_COOKIES:-false}
+      - SECURE_COOKIES=${SECURE_COOKIES:-}
       - EMBEDDING_URL=${EMBEDDING_URL:-}
       - EMBEDDING_MODEL=${EMBEDDING_MODEL:-}
       - EMBEDDING_API_KEY=${EMBEDDING_API_KEY:-}
@@ -58,11 +59,39 @@ services:
       - CLEANUP_INTERVAL_HOURS=${CLEANUP_INTERVAL_HOURS:-24}
       - ODYSSEUS_INPROCESS_POLLERS=${ODYSSEUS_INPROCESS_POLLERS:-1}
       - ODYSSEUS_INPROCESS_TASKS=${ODYSSEUS_INPROCESS_TASKS:-1}
+      - ODYSSEUS_UNATTENDED_MODE=${ODYSSEUS_UNATTENDED_MODE:-false}
+      - ODYSSEUS_QWEN_NATIVE_COMPACT_BUILTINS=${ODYSSEUS_QWEN_NATIVE_COMPACT_BUILTINS:-1}
+      - ODYSSEUS_QWEN_SUPPRESS_LOCAL_CONTEXT=${ODYSSEUS_QWEN_SUPPRESS_LOCAL_CONTEXT:-0}
+      - ODYSSEUS_CAPTURE_MODEL_REQUESTS=${ODYSSEUS_CAPTURE_MODEL_REQUESTS:-0}
+      - ODYSSEUS_MCP_EMAIL_OWNER=${ODYSSEUS_MCP_EMAIL_OWNER:-}
       - ODYSSEUS_SCRIPT_HOST=${ODYSSEUS_SCRIPT_HOST:-localhost}
       - ODYSSEUS_CHAT_UPLOAD_MAX_BYTES=${ODYSSEUS_CHAT_UPLOAD_MAX_BYTES:-10485760}
+      - ODYSSEUS_GALLERY_UPLOAD_MAX_BYTES=${ODYSSEUS_GALLERY_UPLOAD_MAX_BYTES:-104857600}
+      - ODYSSEUS_GALLERY_TRANSFORM_UPLOAD_MAX_BYTES=${ODYSSEUS_GALLERY_TRANSFORM_UPLOAD_MAX_BYTES:-26214400}
+      - ODYSSEUS_MEMORY_IMPORT_MAX_BYTES=${ODYSSEUS_MEMORY_IMPORT_MAX_BYTES:-10485760}
+      - ODYSSEUS_PERSONAL_UPLOAD_MAX_BYTES=${ODYSSEUS_PERSONAL_UPLOAD_MAX_BYTES:-26214400}
+      - ODYSSEUS_EMAIL_COMPOSE_UPLOAD_MAX_BYTES=${ODYSSEUS_EMAIL_COMPOSE_UPLOAD_MAX_BYTES:-26214400}
+      - ODYSSEUS_EDITOR_DRAFT_MAX_BYTES=${ODYSSEUS_EDITOR_DRAFT_MAX_BYTES:-268435456}
+      - ODYSSEUS_STT_MAX_AUDIO_BYTES=${ODYSSEUS_STT_MAX_AUDIO_BYTES:-26214400}
+      - ODYSSEUS_ICS_MAX_BYTES=${ODYSSEUS_ICS_MAX_BYTES:-10485760}
+      - ODYSSEUS_TTS_CACHE_MAX_BYTES=${ODYSSEUS_TTS_CACHE_MAX_BYTES}
+      # Host workspace translation is opt-in. Keep the public compose file
+      # user-neutral; configure these in a local .env or use the host-workspace
+      # overlay with ODYSSEUS_HOST_WORKSPACE_DIR.
+      - ODYSSEUS_WORKSPACE_HOST_ROOT=${ODYSSEUS_WORKSPACE_HOST_ROOT:-}
+      - ODYSSEUS_WORKSPACE_CONTAINER_ROOT=${ODYSSEUS_WORKSPACE_CONTAINER_ROOT:-/workspace}
+      - ODYSSEUS_WORKSPACE_DEFAULT=${ODYSSEUS_WORKSPACE_DEFAULT:-}
       - DATA_BRAVE_API_KEY=${DATA_BRAVE_API_KEY:-}
       - GOOGLE_API_KEY=${GOOGLE_API_KEY:-}
       - GOOGLE_PSE_CX=${GOOGLE_PSE_CX:-}
+      - GOOGLE_OAUTH_CLIENT_ID=${GOOGLE_OAUTH_CLIENT_ID:-}
+      - GOOGLE_OAUTH_CLIENT_SECRET=${GOOGLE_OAUTH_CLIENT_SECRET:-}
+      - GOOGLE_OAUTH_REDIRECT_URI=${GOOGLE_OAUTH_REDIRECT_URI:-}
+      # Externally reachable origin for MCP OAuth callbacks. The container
+      # always listens on 7000 and cannot see the host port map above, so
+      # remote MCP OAuth needs this set whenever the browser reaches
+      # Odysseus on anything other than http://localhost:7000.
+      - OAUTH_REDIRECT_BASE_URL=${OAUTH_REDIRECT_BASE_URL:-}
       - TAVILY_API_KEY=${TAVILY_API_KEY:-}
       - SERPER_API_KEY=${SERPER_API_KEY:-}
       # PUID / PGID — the user/group the container drops to before
@@ -118,12 +147,17 @@ services:
           fi
           sed "s|__SEARXNG_SECRET__|$$secret|g" /tmp/searxng-settings.yml.template > /etc/searxng/settings.yml
         fi
+        # Advisory: a settings file the migration cannot parse or rewrite must
+        # not be what stops searxng from booting. It explains itself on stderr
+        # and we carry on, letting searxng report anything genuinely wrong.
+        /usr/local/searxng/.venv/bin/python /tmp/migrate-searxng-settings.py /etc/searxng/settings.yml || true
         exec /usr/local/searxng/entrypoint.sh
     ports:
       - "127.0.0.1:8080:8080"
     volumes:
       - searxng-data:/etc/searxng
       - ./config/searxng/settings.yml:/tmp/searxng-settings.yml.template:ro,z
+      - ./scripts/migrate_searxng_settings.py:/tmp/migrate-searxng-settings.py:ro,z
     environment:
       - SEARXNG_BASE_URL=http://localhost:8080/
       - SEARXNG_SECRET=${SEARXNG_SECRET:-}

docker-compose.gpu-nvidia.yml

@@ -13,7 +13,7 @@ services:
   odysseus:
     build: .
     ports:
-      - "${APP_BIND:-127.0.0.1}:${APP_PORT:-7000}:7000"
+      - "${APP_BIND:-127.0.0.1}:${APP_PORT:-7011}:7000"
     volumes:
       - ${APP_DATA_DIR:-./data}:/app/data:z
       - ${APP_LOGS_DIR:-./logs}:/app/logs:z
@@ -45,10 +45,11 @@ services:
       - DATABASE_URL=${DATABASE_URL:-sqlite:///./data/app.db}
       - AUTH_ENABLED=${AUTH_ENABLED:-true}
       - LOCALHOST_BYPASS=${LOCALHOST_BYPASS:-false}
+      - COMPANION_BASE_URL=${COMPANION_BASE_URL:-}
       - ODYSSEUS_ADMIN_USER=${ODYSSEUS_ADMIN_USER:-admin}
       - ODYSSEUS_ADMIN_PASSWORD=${ODYSSEUS_ADMIN_PASSWORD:-}
       - ALLOWED_ORIGINS=${ALLOWED_ORIGINS:-http://localhost,http://127.0.0.1}
-      - SECURE_COOKIES=${SECURE_COOKIES:-false}
+      - SECURE_COOKIES=${SECURE_COOKIES:-}
       - EMBEDDING_URL=${EMBEDDING_URL:-}
       - EMBEDDING_MODEL=${EMBEDDING_MODEL:-}
       - EMBEDDING_API_KEY=${EMBEDDING_API_KEY:-}
@@ -57,11 +58,39 @@ services:
       - CLEANUP_INTERVAL_HOURS=${CLEANUP_INTERVAL_HOURS:-24}
       - ODYSSEUS_INPROCESS_POLLERS=${ODYSSEUS_INPROCESS_POLLERS:-1}
       - ODYSSEUS_INPROCESS_TASKS=${ODYSSEUS_INPROCESS_TASKS:-1}
+      - ODYSSEUS_UNATTENDED_MODE=${ODYSSEUS_UNATTENDED_MODE:-false}
+      - ODYSSEUS_QWEN_NATIVE_COMPACT_BUILTINS=${ODYSSEUS_QWEN_NATIVE_COMPACT_BUILTINS:-1}
+      - ODYSSEUS_QWEN_SUPPRESS_LOCAL_CONTEXT=${ODYSSEUS_QWEN_SUPPRESS_LOCAL_CONTEXT:-0}
+      - ODYSSEUS_CAPTURE_MODEL_REQUESTS=${ODYSSEUS_CAPTURE_MODEL_REQUESTS:-0}
+      - ODYSSEUS_MCP_EMAIL_OWNER=${ODYSSEUS_MCP_EMAIL_OWNER:-}
       - ODYSSEUS_SCRIPT_HOST=${ODYSSEUS_SCRIPT_HOST:-localhost}
       - ODYSSEUS_CHAT_UPLOAD_MAX_BYTES=${ODYSSEUS_CHAT_UPLOAD_MAX_BYTES:-10485760}
+      - ODYSSEUS_GALLERY_UPLOAD_MAX_BYTES=${ODYSSEUS_GALLERY_UPLOAD_MAX_BYTES:-104857600}
+      - ODYSSEUS_GALLERY_TRANSFORM_UPLOAD_MAX_BYTES=${ODYSSEUS_GALLERY_TRANSFORM_UPLOAD_MAX_BYTES:-26214400}
+      - ODYSSEUS_MEMORY_IMPORT_MAX_BYTES=${ODYSSEUS_MEMORY_IMPORT_MAX_BYTES:-10485760}
+      - ODYSSEUS_PERSONAL_UPLOAD_MAX_BYTES=${ODYSSEUS_PERSONAL_UPLOAD_MAX_BYTES:-26214400}
+      - ODYSSEUS_EMAIL_COMPOSE_UPLOAD_MAX_BYTES=${ODYSSEUS_EMAIL_COMPOSE_UPLOAD_MAX_BYTES:-26214400}
+      - ODYSSEUS_EDITOR_DRAFT_MAX_BYTES=${ODYSSEUS_EDITOR_DRAFT_MAX_BYTES:-268435456}
+      - ODYSSEUS_STT_MAX_AUDIO_BYTES=${ODYSSEUS_STT_MAX_AUDIO_BYTES:-26214400}
+      - ODYSSEUS_ICS_MAX_BYTES=${ODYSSEUS_ICS_MAX_BYTES:-10485760}
+      - ODYSSEUS_TTS_CACHE_MAX_BYTES=${ODYSSEUS_TTS_CACHE_MAX_BYTES}
+      # Host workspace translation is opt-in. Keep the public compose file
+      # user-neutral; configure these in a local .env or use the host-workspace
+      # overlay with ODYSSEUS_HOST_WORKSPACE_DIR.
+      - ODYSSEUS_WORKSPACE_HOST_ROOT=${ODYSSEUS_WORKSPACE_HOST_ROOT:-}
+      - ODYSSEUS_WORKSPACE_CONTAINER_ROOT=${ODYSSEUS_WORKSPACE_CONTAINER_ROOT:-/workspace}
+      - ODYSSEUS_WORKSPACE_DEFAULT=${ODYSSEUS_WORKSPACE_DEFAULT:-}
       - DATA_BRAVE_API_KEY=${DATA_BRAVE_API_KEY:-}
       - GOOGLE_API_KEY=${GOOGLE_API_KEY:-}
       - GOOGLE_PSE_CX=${GOOGLE_PSE_CX:-}
+      - GOOGLE_OAUTH_CLIENT_ID=${GOOGLE_OAUTH_CLIENT_ID:-}
+      - GOOGLE_OAUTH_CLIENT_SECRET=${GOOGLE_OAUTH_CLIENT_SECRET:-}
+      - GOOGLE_OAUTH_REDIRECT_URI=${GOOGLE_OAUTH_REDIRECT_URI:-}
+      # Externally reachable origin for MCP OAuth callbacks. The container
+      # always listens on 7000 and cannot see the host port map above, so
+      # remote MCP OAuth needs this set whenever the browser reaches
+      # Odysseus on anything other than http://localhost:7000.
+      - OAUTH_REDIRECT_BASE_URL=${OAUTH_REDIRECT_BASE_URL:-}
       - TAVILY_API_KEY=${TAVILY_API_KEY:-}
       - SERPER_API_KEY=${SERPER_API_KEY:-}
       # PUID / PGID — the user/group the container drops to before
@@ -121,12 +150,17 @@ services:
           fi
           sed "s|__SEARXNG_SECRET__|$$secret|g" /tmp/searxng-settings.yml.template > /etc/searxng/settings.yml
         fi
+        # Advisory: a settings file the migration cannot parse or rewrite must
+        # not be what stops searxng from booting. It explains itself on stderr
+        # and we carry on, letting searxng report anything genuinely wrong.
+        /usr/local/searxng/.venv/bin/python /tmp/migrate-searxng-settings.py /etc/searxng/settings.yml || true
         exec /usr/local/searxng/entrypoint.sh
     ports:
       - "127.0.0.1:8080:8080"
     volumes:
       - searxng-data:/etc/searxng
       - ./config/searxng/settings.yml:/tmp/searxng-settings.yml.template:ro,z
+      - ./scripts/migrate_searxng_settings.py:/tmp/migrate-searxng-settings.py:ro,z
     environment:
       - SEARXNG_BASE_URL=http://localhost:8080/
       - SEARXNG_SECRET=${SEARXNG_SECRET:-}

docker-compose.yml

@@ -2,7 +2,7 @@ services:
   odysseus:
     build: .
     ports:
-      - "${APP_BIND:-127.0.0.1}:${APP_PORT:-7000}:7000"
+      - "${APP_BIND:-127.0.0.1}:${APP_PORT:-7011}:7000"
     volumes:
       - ${APP_DATA_DIR:-./data}:/app/data:z
       - ${APP_LOGS_DIR:-./logs}:/app/logs:z
@@ -34,10 +34,11 @@ services:
       - DATABASE_URL=${DATABASE_URL:-sqlite:///./data/app.db}
       - AUTH_ENABLED=${AUTH_ENABLED:-true}
       - LOCALHOST_BYPASS=${LOCALHOST_BYPASS:-false}
+      - COMPANION_BASE_URL=${COMPANION_BASE_URL:-}
       - ODYSSEUS_ADMIN_USER=${ODYSSEUS_ADMIN_USER:-admin}
       - ODYSSEUS_ADMIN_PASSWORD=${ODYSSEUS_ADMIN_PASSWORD:-}
       - ALLOWED_ORIGINS=${ALLOWED_ORIGINS:-http://localhost,http://127.0.0.1}
-      - SECURE_COOKIES=${SECURE_COOKIES:-false}
+      - SECURE_COOKIES=${SECURE_COOKIES:-}
       - EMBEDDING_URL=${EMBEDDING_URL:-}
       - EMBEDDING_MODEL=${EMBEDDING_MODEL:-}
       - EMBEDDING_API_KEY=${EMBEDDING_API_KEY:-}
@@ -46,11 +47,39 @@ services:
       - CLEANUP_INTERVAL_HOURS=${CLEANUP_INTERVAL_HOURS:-24}
       - ODYSSEUS_INPROCESS_POLLERS=${ODYSSEUS_INPROCESS_POLLERS:-1}
       - ODYSSEUS_INPROCESS_TASKS=${ODYSSEUS_INPROCESS_TASKS:-1}
+      - ODYSSEUS_UNATTENDED_MODE=${ODYSSEUS_UNATTENDED_MODE:-false}
+      - ODYSSEUS_QWEN_NATIVE_COMPACT_BUILTINS=${ODYSSEUS_QWEN_NATIVE_COMPACT_BUILTINS:-1}
+      - ODYSSEUS_QWEN_SUPPRESS_LOCAL_CONTEXT=${ODYSSEUS_QWEN_SUPPRESS_LOCAL_CONTEXT:-0}
+      - ODYSSEUS_CAPTURE_MODEL_REQUESTS=${ODYSSEUS_CAPTURE_MODEL_REQUESTS:-0}
+      - ODYSSEUS_MCP_EMAIL_OWNER=${ODYSSEUS_MCP_EMAIL_OWNER:-}
       - ODYSSEUS_SCRIPT_HOST=${ODYSSEUS_SCRIPT_HOST:-localhost}
       - ODYSSEUS_CHAT_UPLOAD_MAX_BYTES=${ODYSSEUS_CHAT_UPLOAD_MAX_BYTES:-10485760}
+      - ODYSSEUS_GALLERY_UPLOAD_MAX_BYTES=${ODYSSEUS_GALLERY_UPLOAD_MAX_BYTES:-104857600}
+      - ODYSSEUS_GALLERY_TRANSFORM_UPLOAD_MAX_BYTES=${ODYSSEUS_GALLERY_TRANSFORM_UPLOAD_MAX_BYTES:-26214400}
+      - ODYSSEUS_MEMORY_IMPORT_MAX_BYTES=${ODYSSEUS_MEMORY_IMPORT_MAX_BYTES:-10485760}
+      - ODYSSEUS_PERSONAL_UPLOAD_MAX_BYTES=${ODYSSEUS_PERSONAL_UPLOAD_MAX_BYTES:-26214400}
+      - ODYSSEUS_EMAIL_COMPOSE_UPLOAD_MAX_BYTES=${ODYSSEUS_EMAIL_COMPOSE_UPLOAD_MAX_BYTES:-26214400}
+      - ODYSSEUS_EDITOR_DRAFT_MAX_BYTES=${ODYSSEUS_EDITOR_DRAFT_MAX_BYTES:-268435456}
+      - ODYSSEUS_STT_MAX_AUDIO_BYTES=${ODYSSEUS_STT_MAX_AUDIO_BYTES:-26214400}
+      - ODYSSEUS_ICS_MAX_BYTES=${ODYSSEUS_ICS_MAX_BYTES:-10485760}
+      - ODYSSEUS_TTS_CACHE_MAX_BYTES=${ODYSSEUS_TTS_CACHE_MAX_BYTES}
+      # Host workspace translation is opt-in. Keep the public compose file
+      # user-neutral; configure these in a local .env or use the host-workspace
+      # overlay with ODYSSEUS_HOST_WORKSPACE_DIR.
+      - ODYSSEUS_WORKSPACE_HOST_ROOT=${ODYSSEUS_WORKSPACE_HOST_ROOT:-}
+      - ODYSSEUS_WORKSPACE_CONTAINER_ROOT=${ODYSSEUS_WORKSPACE_CONTAINER_ROOT:-/workspace}
+      - ODYSSEUS_WORKSPACE_DEFAULT=${ODYSSEUS_WORKSPACE_DEFAULT:-}
       - DATA_BRAVE_API_KEY=${DATA_BRAVE_API_KEY:-}
       - GOOGLE_API_KEY=${GOOGLE_API_KEY:-}
       - GOOGLE_PSE_CX=${GOOGLE_PSE_CX:-}
+      - GOOGLE_OAUTH_CLIENT_ID=${GOOGLE_OAUTH_CLIENT_ID:-}
+      - GOOGLE_OAUTH_CLIENT_SECRET=${GOOGLE_OAUTH_CLIENT_SECRET:-}
+      - GOOGLE_OAUTH_REDIRECT_URI=${GOOGLE_OAUTH_REDIRECT_URI:-}
+      # Externally reachable origin for MCP OAuth callbacks. The container
+      # always listens on 7000 and cannot see the host port map above, so
+      # remote MCP OAuth needs this set whenever the browser reaches
+      # Odysseus on anything other than http://localhost:7000.
+      - OAUTH_REDIRECT_BASE_URL=${OAUTH_REDIRECT_BASE_URL:-}
       - TAVILY_API_KEY=${TAVILY_API_KEY:-}
       - SERPER_API_KEY=${SERPER_API_KEY:-}
       # PUID / PGID — the user/group the container drops to before
@@ -99,12 +128,17 @@ services:
           fi
           sed "s|__SEARXNG_SECRET__|$$secret|g" /tmp/searxng-settings.yml.template > /etc/searxng/settings.yml
         fi
+        # Advisory: a settings file the migration cannot parse or rewrite must
+        # not be what stops searxng from booting. It explains itself on stderr
+        # and we carry on, letting searxng report anything genuinely wrong.
+        /usr/local/searxng/.venv/bin/python /tmp/migrate-searxng-settings.py /etc/searxng/settings.yml || true
         exec /usr/local/searxng/entrypoint.sh
     ports:
       - "127.0.0.1:8080:8080"
     volumes:
       - searxng-data:/etc/searxng
       - ./config/searxng/settings.yml:/tmp/searxng-settings.yml.template:ro,z
+      - ./scripts/migrate_searxng_settings.py:/tmp/migrate-searxng-settings.py:ro,z
     environment:
       - SEARXNG_BASE_URL=http://localhost:8080/
       - SEARXNG_SECRET=${SEARXNG_SECRET:-}

docker/build-realesrgan-wheels.sh

@@ -0,0 +1,70 @@
+#!/usr/bin/env bash
+# Build patched wheels for Real-ESRGAN's unmaintained dependencies.
+#
+# basicsr / gfpgan / facexlib (xinntao, last released 2022) read their version
+# in setup.py with:
+#
+#     exec(compile(f.read(), version_file, 'exec'))
+#     return locals()['__version__']
+#
+# Python 3.13+ implements PEP 667: locals() inside a function returns an
+# independent snapshot that exec() can no longer mutate, so the read raises
+# `KeyError: '__version__'` and the sdist build fails. That is why the Cookbook
+# "install realesrgan" button dies on the python:3.14 image. The packages have
+# no fixed release, so we patch get_version() to exec into an explicit namespace
+# dict (works on every Python) and build wheels from the patched source.
+#
+# Usage: build-realesrgan-wheels.sh [OUTPUT_DIR]   (default: /wheels)
+set -euo pipefail
+
+OUT="${1:-/wheels}"
+mkdir -p "$OUT"
+
+work="$(mktemp -d)"
+trap 'rm -rf "$work"' EXIT
+cd "$work"
+
+# Pinned to the versions Real-ESRGAN 0.3.0 resolves to.
+SPECS="basicsr==1.4.2 gfpgan==1.3.8 facexlib==0.3.0"
+
+for spec in $SPECS; do
+  name="${spec%%==*}"
+  ver="${spec##*==}"
+  # pip download builds metadata (and trips the same bug), so fetch the raw
+  # sdist URL from the PyPI JSON API instead.
+  url="$(python - "$name" "$ver" <<'PY'
+import json, sys, urllib.request
+name, ver = sys.argv[1], sys.argv[2]
+data = json.load(urllib.request.urlopen(f"https://pypi.org/pypi/{name}/{ver}/json"))
+for f in data["urls"]:
+    if f["packagetype"] == "sdist":
+        print(f["url"]); break
+else:
+    sys.exit(f"no sdist found for {name}=={ver}")
+PY
+)"
+  echo ">> fetching ${name} ${ver}: ${url}"
+  curl -fsSL "$url" -o "${name}.tar.gz"
+  tar xzf "${name}.tar.gz"
+done
+
+echo ">> patching get_version()"
+python - <<'PY'
+import pathlib
+old_exec = "exec(compile(f.read(), version_file, 'exec'))"
+new_exec = "_ver_ns = {}\n        exec(compile(f.read(), version_file, 'exec'), _ver_ns)"
+old_ret = "return locals()['__version__']"
+new_ret = "return _ver_ns['__version__']"
+patched = 0
+for setup in pathlib.Path(".").glob("*/setup.py"):
+    s = setup.read_text()
+    if old_exec in s and old_ret in s:
+        setup.write_text(s.replace(old_exec, new_exec).replace(old_ret, new_ret))
+        print("   patched", setup)
+        patched += 1
+assert patched == 3, f"expected to patch 3 setup.py files, patched {patched}"
+PY
+
+echo ">> building wheels into ${OUT}"
+pip wheel --no-deps -w "$OUT" ./basicsr-* ./gfpgan-* ./facexlib-*
+ls -l "$OUT"

docker/entrypoint.sh

@@ -13,6 +13,8 @@ set -e
 
 PUID="${PUID:-1000}"
 PGID="${PGID:-1000}"
+GOSU_BIN="$(command -v gosu)"
+PYTHON_BIN="$(command -v python)"
 
 # Reuse an existing matching group/user if the host's UID/GID already
 # corresponds to one in /etc/passwd (e.g. when the image is rebuilt
@@ -24,26 +26,78 @@ if ! getent passwd "$PUID" >/dev/null 2>&1; then
     useradd -u "$PUID" -g "$PGID" -M -s /bin/sh -d /app odysseus
 fi
 
-# Repair ownership on every writable path the app touches at runtime.
-#
-# Bind-mounted dirs (/app/data, /app/logs) are the obvious ones, but
-# the app ALSO writes inside the image's own source tree at runtime:
-#   - services/cache/{search,content}/*  (search cache LRU)
-#   - services/search_analytics.json
-#   - services/search_engine_error.log
-#   - services/tts cache, etc.
-# These dirs were created as root during `docker build`, so dropping
-# to PUID:PGID would otherwise crash on the first import that tries
-# to mkdir them. Chown the whole /app tree — fast (<1s on this size)
-# and idempotent via the `-not -uid` filter so we only touch files
-# that need fixing.
-for dir in /app /app/data /app/logs; do
+ODY_USER="$(getent passwd "$PUID" | cut -d: -f1)"
+[ -z "$ODY_USER" ] && ODY_USER=odysseus
+
+# Docker-socket group plumbing for the explicit host-Docker overlay. When
+# opted in, the socket is owned by root:<host docker gid>. Add the app user
+# to that group and later call gosu by username so supplementary groups are
+# retained.
+DOCKER_SOCK="${DOCKER_SOCK:-/var/run/docker.sock}"
+if [ "${ODYSSEUS_ENABLE_HOST_DOCKER:-}" = "true" ] && [ -S "$DOCKER_SOCK" ]; then
+    SOCK_GID="$(stat -c '%g' "$DOCKER_SOCK" 2>/dev/null || echo '')"
+    if [ -n "$SOCK_GID" ] && [ "$SOCK_GID" != "0" ]; then
+        if ! getent group "$SOCK_GID" >/dev/null 2>&1; then
+            groupadd -g "$SOCK_GID" docker_host || true
+        fi
+        SOCK_GROUP="$(getent group "$SOCK_GID" | cut -d: -f1)"
+        if [ -n "$SOCK_GROUP" ]; then
+            usermod -aG "$SOCK_GROUP" "$ODY_USER" 2>/dev/null || true
+        fi
+    fi
+fi
+
+mount_root_for() {
+    awk -v target="$1" '$5 == target { print $4; exit }' /proc/self/mountinfo 2>/dev/null || true
+}
+
+is_broad_mount_root() {
+    case "$1" in
+        /|/home|/srv|/var|/usr|/opt|/tmp|/mnt|/media)
+            return 0
+            ;;
+    esac
+    return 1
+}
+
+repair_tree_ownership() {
+    dir="$1"
     if [ -d "$dir" ]; then
-        # `find ... -not -uid` keeps this O(touched-files), not
-        # O(everything), so terabyte-sized maildirs don't slow startup.
-        find "$dir" -not -uid "$PUID" -print0 2>/dev/null \
+        find "$dir" -xdev -not -uid "$PUID" -print0 2>/dev/null \
             | xargs -0 -r chown "$PUID:$PGID" 2>/dev/null || true
     fi
+}
+
+repair_app_tree_ownership() {
+    if [ -d /app ]; then
+        find /app -xdev \
+            \( -path /app/data -o -path /app/logs -o -path /app/.ssh -o -path /app/.cache -o -path /app/.local \) -prune \
+            -o -not -uid "$PUID" -print0 2>/dev/null \
+            | xargs -0 -r chown "$PUID:$PGID" 2>/dev/null || true
+    fi
+}
+
+repair_bind_mount_ownership() {
+    dir="$1"
+    if [ ! -d "$dir" ]; then
+        return
+    fi
+
+    mount_root="$(mount_root_for "$dir")"
+    if is_broad_mount_root "$mount_root"; then
+        echo "Skipping recursive ownership repair for $dir because it maps to broad host path $mount_root" >&2
+        chown "$PUID:$PGID" "$dir" 2>/dev/null || true
+        return
+    fi
+
+    repair_tree_ownership "$dir"
+}
+
+# Repair image-owned writable paths without walking into bind-mounted host
+# trees, then repair the app-owned mount roots separately.
+repair_app_tree_ownership
+for dir in /app/data /app/logs /app/.ssh /app/.cache/huggingface /app/.local; do
+    repair_bind_mount_ownership "$dir"
 done
 
 # Cookbook installs vllm/etc. via `pip install --user`, which pulls
@@ -70,6 +124,7 @@ for cu in \
         break
     fi
 done
+
 # Disable the FlashInfer JIT sampler unconditionally — it is sampler-only
 # and has no impact on the attention path, but requires nvcc + matching
 # CUDA headers at startup. Without this, vLLM crashes with "Could not find
@@ -83,9 +138,9 @@ export PATH="/app/.local/bin:$PATH"
 # Run first-time setup as the app user so data/ files get the right ownership.
 # setup.py is idempotent — skips auth.json / .env if they already exist.
 # || true so a setup failure never prevents the container from starting.
-gosu "$PUID:$PGID" python /app/setup.py || true
+"$GOSU_BIN" "$ODY_USER" "$PYTHON_BIN" /app/setup.py || true
 
 # Drop root and run the actual app. `gosu` is preferred over `su` /
 # `sudo` because it cleans up the process tree (no extra shell layer)
 # so signals (SIGTERM from `docker stop`) reach uvicorn directly.
-exec gosu "$PUID:$PGID" "$@"
+exec "$GOSU_BIN" "$ODY_USER" "$@"

docker/host-docker.yml

@@ -0,0 +1,12 @@
+# High-trust host Docker access. Enable only when local Docker-daemon
+# management from Cookbook is required and you accept that raw socket access
+# grants broad control over the host Docker daemon.
+#   COMPOSE_FILE=docker-compose.yml:docker/host-docker.yml
+#   DOCKER_GID=<numeric host Docker group id>
+services:
+  odysseus:
+    volumes:
+      - /var/run/docker.sock:/var/run/docker.sock
+    group_add: ["${DOCKER_GID:-963}"]
+    environment:
+      - ODYSSEUS_ENABLE_HOST_DOCKER=true

docker/host-network.yml

@@ -0,0 +1,21 @@
+# High-trust host network access. Enable only when the Odysseus agent needs
+# host-native LAN/VPN/mDNS behavior that Docker bridge networking cannot
+# provide. Linux only; Docker Desktop does not provide equivalent host
+# networking semantics.
+#   COMPOSE_FILE=docker-compose.yml:docker/host-workspace.yml:docker/host-network.yml
+#   APP_PORT=7011
+services:
+  odysseus:
+    network_mode: host
+    ports: !reset []
+    environment:
+      - APP_PORT=${APP_PORT:-7011}
+      - APP_BIND=${APP_BIND:-0.0.0.0}
+      - SEARXNG_INSTANCE=${ODYSSEUS_HOST_NETWORK_SEARXNG_INSTANCE:-http://127.0.0.1:8080}
+      - CHROMADB_HOST=${ODYSSEUS_HOST_NETWORK_CHROMADB_HOST:-127.0.0.1}
+      - CHROMADB_PORT=${ODYSSEUS_HOST_NETWORK_CHROMADB_PORT:-8100}
+      - ODYSSEUS_CONTAINER_NETWORK_MODE=host
+    command:
+      - sh
+      - -c
+      - exec uvicorn app:app --host "$${APP_BIND:-0.0.0.0}" --port "$${APP_PORT:-7011}"

docker/host-workspace.yml

@@ -0,0 +1,11 @@
+# High-trust host workspace access. Enable only when the Odysseus agent should
+# work on a host directory outside the container's normal /app/data sandbox.
+#   COMPOSE_FILE=docker-compose.yml:docker/host-workspace.yml
+#   ODYSSEUS_HOST_WORKSPACE_DIR=/absolute/host/path
+#   ODYSSEUS_HOST_WORKSPACE_MOUNT=/host/workspace
+services:
+  odysseus:
+    volumes:
+      - ${ODYSSEUS_HOST_WORKSPACE_DIR:?set ODYSSEUS_HOST_WORKSPACE_DIR}:${ODYSSEUS_HOST_WORKSPACE_MOUNT:-/host/workspace}:rw,z
+    environment:
+      - ODYSSEUS_HOST_WORKSPACE_MOUNT=${ODYSSEUS_HOST_WORKSPACE_MOUNT:-/host/workspace}

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
+/home/pewds/odysseus-cookbook-fresh/.venv/bin/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
+/home/pewds/odysseus-cookbook-fresh/.venv/bin/pytest -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/CLEAN_LOOP_V3_EXPERIMENT_20260909.md

@@ -0,0 +1,99 @@
+# No-RAG clean loop: first diagnostic
+
+## Setup
+
+No live UI, service configuration, or weights changed. The standalone loop sends
+conversation history, native assistant calls and matching tool results directly
+to the served pre-Heretic model. It never rewrites queries, invents calls, swaps
+families, or strips output. Invalid calls return errors. Six executions per turn
+and seven model rounds bound the test.
+
+Both arms use temperature 0, thinking disabled, 768 output tokens, and the
+original tool-work evaluator's `tools_for_mode(..., 'compact_contract_v3')`.
+This matters: the app's plain compact scrubber deletes descriptions, whereas v3
+retains empirically tested micro-hints. Previous plain-compact tests were not
+exact reproductions of the passing benchmark setup.
+
+The 76 tools come from the current app's ten-family inventory, transformed by
+the original v3 builder. This is not a byte-identical frozen 99-tool benchmark
+inventory or proof of training-data identity. The report records schema and
+builder hashes. No schemas are invented for this experiment.
+
+- **Stable:** same compact inventory on every turn, irrespective of spelling.
+- **Routed:** same loop, but existing `requested_capabilities` chooses inventory
+  each turn. This isolates that selector; it is not the complete production RAG
+  or Agent UI path. Other production normalizers are absent in both arms.
+- Private records are synthetic. No real private dispatcher is imported.
+  Only fixture reads and optional public SearXNG calls execute. Other operations
+  return explicit errors, so this does not validate their functionality.
+- Live search sends the exact model query to local SearXNG Bing/Yep, bypassing
+  app query rewriting/filtering. Source results may vary between arms.
+
+## Observations, not a blind score
+
+| Case | Stable compact inventory | Selector arm |
+|---|---|---|
+| `whats the current stock mraket` | Selected `web_search`, query `current stock market` | Offered zero tools; declined live lookup |
+| Exact seeded failed exchange, then `can you look up` | Searched with corrected query | Also searched with corrected query |
+| Summarize search, explicitly no tools | Answered without tools or permission failure | Same |
+| Calendar → email → calendar | Recalled second event at 14:30 | Same |
+| Notes → second note → what does it say | Correct `view` ID and content | Same after fixture correction |
+| Deliberately irrelevant search result | Did not automatically retry | Did not automatically retry |
+| User asks for a better source | Refined and executed another search | Proposed search was not offered and was rejected |
+| Web-disabled lookup | Attempted network access via bash; sandbox rejected it | Invented unsupported current market news without tools |
+
+The initial stable stock answer listed sources, not current index values. It
+does not establish that the market question was fully answered. Its subsequent
+`can you look up` elicited clarification after it had already searched. The
+separate seeded replay removes that differing-history confound.
+
+The first notes fixture incorrectly accepted `get/read`, not the real `view`
+action. Both models selected the correct action, but the fixture rejected it.
+Those six original turns are invalid for execution comparison. A corrected
+six-turn rerun succeeded in both arms; the failed evidence is retained.
+
+Web-off results are a release blocker: removing named web tools alone does not
+enforce network denial across general-purpose tools. The fixture prevented real
+execution, but any UI integration must use the real cross-tool permissions and
+clearly communicate unavailable capabilities. Neither arm is ready for a live
+switch. Source recovery and grounded completion also remain weak.
+
+## What this changes
+
+There is direct evidence that the selector can withhold needed tools, and that
+the model can repair the misspelled query itself when offered the tool. Clean
+history also supports the tested topic switches without synthetic substitutions.
+This supports continuing the clean-path experiment, not retraining or declaring
+the UI fixed. Full inventory is slower in these requests; overlapping runs and
+different source content prevent a controlled latency conclusion.
+
+Next: integrate the clean loop behind a test-only UI profile with real permission
+enforcement and one renderer, preserving the v3 contract. Test live read-only
+follow-ups and explicit Web-off behavior before any rollout. Separately compare
+a generic evidence-check/retry instruction on the weak-result fixture; do not
+manufacture a retry query in the harness.
+
+## Reproduce
+
+Eight boundary tests pass:
+
+```sh
+/home/pewds/odysseus-cookbook-fresh/.venv/bin/pytest -q tests/test_clean_tool_loop.py
+```
+
+Run with a fresh report filename (existing evidence is never overwritten):
+
+```sh
+/home/pewds/odysseus-cookbook-fresh/.venv/bin/python scripts/test_clean_tool_loop.py --live-search --report reports/clean-loop-v3-new-run.json
+```
+
+Evidence:
+
+- `reports/clean-loop-v3-20260909.json`: original 24 turns; notes fixture caveat above.
+- `reports/clean-loop-v3-stock-seeded-20260909.json`: four matched seeded follow-up turns.
+- `reports/clean-loop-v3-notes-fixture-corrected-20260909.json`: corrected six notes turns.
+
+Each report retains model requests, responses, offered inventory and execution
+results. The `completed` status means the request loop finished, **not** that
+the answer passed functional evaluation. These are synthetic/public traces, not
+private user conversations. This test does not measure UI rendering or streaming.

docs/CLEAN_V3_UI_PREVIEW.md

@@ -0,0 +1,220 @@
+# Tools v3 — No-RAG preview
+
+Select this endpoint in the 7011 model picker, with model
+`odysseus-qwen3.5-tools-pre-heretic`. This endpoint owns its complete tool loop
+and enters Agent mode server-side on every turn, including ambiguous follow-ups;
+it does not depend on the legacy per-message intent classifier. Start a new chat
+for an uncontaminated comparison. Enable Web for searches. Clean routing is
+owned by the exact model identity, so both the normal `preheret` endpoint and
+the `cleanv3` alias use this runtime. Every other model remains on legacy RAG.
+
+Endpoint ID: `cleanv3`. Its base URL uses the same inference server's Tailscale
+DNS name, `http://odysseus.tailb895f4.ts.net:18182/v1`, to distinguish it from
+the original IP-address route when existing chats omit endpoint IDs.
+
+## Implementation
+
+- `src/clean_agent_preview.py` is a separate streamed native-tool loop, entered
+  before legacy routing and substitutions. It uses real authenticated tool
+  dispatch, the tool-work `compact_contract_v5` builder, temperature 0,
+  and thinking disabled. No weights change or inference server was started.
+- The offered tool inventory is stable except for permissions/toggles. Safe,
+  explicit personal creates/updates are enabled for notes, tasks, calendar,
+  memory, skills and documents. Destructive operations, shell/code, outbound
+  email, browser interaction, deployment/admin changes and unrelated-family
+  write substitution remain blocked. No tool or argument substitution is
+  applied by the loop.
+- Native calls and matching results persist in `clean_v3_turn` metadata so
+  follow-ups use actual evidence. History retains at most eight complete turns,
+  trimming oldest whole turns for size; individual outputs cap at 8000 chars.
+- Real search still uses the existing search backend and its provider handling;
+  this does not claim that provider quality or every backend transform is fixed.
+- All routing, privileges and default settings outside this exact Odysseus model
+  remain unchanged. The loop has six execution/eight-round limits.
+- Write completion is evidence-bound: affirmative success text is replaced
+  unless a private-write tool succeeded during the turn. Proposed call batches
+  are policy-preflighted atomically, so a batch containing a blocked operation
+  cannot partially execute before denial.
+
+## Verification
+
+399 focused Python tests passed after route integration. Browser runs r1/r2
+accidentally exercised the old loop and are not preview evidence. The runner
+now explicitly asserts `selection_mode=clean_compact_v3_preview`.
+
+`reports/clean-v3-live-ui-r3-20260909.json` confirms the preview route, real notes
+execution, correct repetition from history, successful search and no-tool
+summary, plus visible incremental growth. Its notes assertions were for the
+old routed contract: they prohibited offering web tools even with Web enabled,
+and required another notes call for a verbatim repeat. The updated preview
+checks permit stable offers and accept an exact match to the preceding saved
+answer without re-execution; execution permissions are still asserted.
+
+`reports/clean-v3-live-ui-r4-20260909.json` is the corrected four-turn check,
+including notes with Web off and search with Web on: **4/4 passed**, with the
+preview selection mode explicitly confirmed on every turn.
+These are UI smoke tests, not all-family or factual-answer benchmark scores.
+
+## Disable
+
+Disabling only endpoint `cleanv3` removes the duplicate picker alias; it does
+not disable this model-owned runtime. To roll back the runtime, revert the exact
+model route in `routes/chat_routes.py`. Do not delete weights, adapters, or user
+chats. The v3 schema builder dependency is
+`/home/pewds/odysseus-tool-work/scripts/eval_alltools_unseen_compare.py` and its
+schema-dropout helper; preserve those with this deployment.
+
+## Expanded UI checks — 2026-09-09
+
+24 additional turns completed through the preview: 23 automated passes and one
+checker false alarm. The Cookbook follow-up correctly shortened the previous
+six-server result to the first three requested names without another call. The
+checker required either a fresh call or a verbatim repeat; manual inspection
+confirmed the requested subset. Raw failure evidence is retained, not rescored.
+
+Covered notes/misspellings/second-note selection, calendar/second-event time,
+tasks, documents, memory, skills, Cookbook listing, misspelled search, and Web
+toggle changes. Cross-family flows passed: Germany news → “whats my notes”,
+notes → “seach current stock mraket news”, and calendar → “now show my noes”.
+The model chose `current stock market news` itself. Every completed turn's
+audit confirmed the preview mode. Search source factual accuracy is not graded
+by this suite, and successful reads do not establish mutation coverage.
+
+Email was separately attempted but the test guard stopped it because the
+stable offered inventory exceeded its metadata-only verified scope. Email
+therefore remains unverified in this expanded run; the guard was not weakened.
+No production code, service settings or weights changed during these tests.
+
+Evidence under `reports/`:
+
+- `clean-v3-broader-ui-20260909.json`: 16 turns, 15 automatic passes, Cookbook caveat.
+- `clean-v3-topic-switch-ui-20260909.json`: 6/6 passed.
+- `clean-v3-second-note-ui-20260909.json`: 2/2 passed.
+- `clean-v3-email-notes-ui-20260909.json`: blocked email attempt; notes not run in that file.
+
+## Picker route fix
+
+The previous tests selected sessions through the API, missing a real picker
+bug: local entries were deduplicated by model ID, hiding alternative endpoints
+with the same weights. The picker now uses endpoint+model identity for local
+routes too, displays the endpoint name, and scopes its last-picked send override
+to the current chat. `/api/sessions` returns owner-filtered endpoint identity
+for unambiguous saved URLs, so reload labels do not depend on loading the model
+catalog. Ambiguous identical URLs are not guessed.
+
+The user-authorized chat `ec0683a2-015f-41d7-aa1f-34135c9640cb` was switched to
+`cleanv3` using the authenticated session PATCH API; no messages were inserted
+and no tool actions ran in that chat. Defaults and other chats were unchanged.
+
+The runner's `--picker-route true` starts on the original route, clicks the
+preview in the real picker, sends a greeting, reloads the chat permalink, then
+asks for notes. Early picker/reload reports are incomplete, not passes: their
+label check exposed the unloaded-catalog issue. Focused route/picker/history
+tests: 16 passed.
+
+Final picker test: `reports/clean-v3-picker-reload-r5-20260909.json`, **2/2
+passed**. Real picker click, greeting, permalink reload, and notes follow-up
+all confirmed the preview route. The label survived reload. R4 retained a
+history/DOM mismatch from sending before restored history was ready; the final
+driver explicitly waits for the saved first answer to render before sending.
+This does not claim a general fix for sending during unfinished history loading.
+
+## Native image/VL status
+
+The inference launcher previously set `--limit-mm-per-prompt` to zero images,
+so vLLM rejected attachments before the model saw them. The durable Odysseus
+launcher now permits up to three images per prompt; video remains disabled.
+
+`reports/clean-v3-vl-live-r4-20260909.json` proves the real 7011 attachment
+path, clean compact route, object/color/spatial recognition, permalink reload,
+and ambiguous image follow-up. Those checks pass. Exact OCR of the deterministic
+`ODYSSEUS 42` heading fails in both the untouched Qwen 3.5 9B base and the
+fine-tune, so it remains a base/runtime capability limitation rather than a
+fine-tune regression or harness failure.
+
+The same native path also passes JPEG and lossless WebP transport, object
+recognition, reload, and follow-up grounding. Evidence:
+`reports/clean-v3-vl-jpeg-r1-20260909.json` and
+`reports/clean-v3-vl-webp-r1-20260909.json`. Both remain `partial` only because
+the shared OCR check fails.
+
+## Reversible write check
+
+`scripts/verify_clean_v3_write.mjs` runs against only `sft_alex_creator`. It
+creates one UUID-named note through the real 7011 UI, verifies that exact row,
+requests a destructive bulk deletion, verifies the row still exists, and then
+deletes only its own test row through the authenticated API. The cleanup is
+verified by a 404 lookup.
+
+Final evidence: `reports/clean-v3-write-ui-r8-20260909.json`, **passed**. Both
+turns reported `selection_mode=clean_compact_v3_preview`; creation executed via
+`manage_notes(action=add)`, the destructive action did not execute, and the
+canonical response was “No changes were made.” The earlier r3/r5 files are
+startup/placement failures, while r4/r6/r7 retained genuine intermediate
+harness and verifier failures; none should be interpreted as passes.
+
+## Stateful, search, and email checks
+
+The reversible stateful runner passes all six mutation families in one run:
+calendar, notes, tasks, documents, memory, and skills (**6/6**). Each flow
+creates a UUID-only artifact through the real Agent UI, verifies it by
+owner-scoped API, applies a noun-free correction, verifies persistence, and
+removes only that artifact. A direct database audit found zero active synthetic
+calendar, note, task, or document rows afterward.
+
+The document failure was harness-owned. Compact description dropout left a
+vague free-form `command` field, error envelopes defaulted to exit code 0, and
+the clean loop dropped the active document ID. Compact v5 now exposes only
+required structured `edits`, reports errors truthfully, and executes against
+the request's explicit active document. Fresh document and combined stateful
+runs pass.
+
+Search Web-toggle combinations `00`, `01`, `10`, and `11` pass **8/8** across
+two turns. A web question can no longer silently enable Bash because it says
+“official source”, and an unavailable Web capability exposes no unrelated
+fallback family. The quality suite passes **3/3**: evidence reuse without a
+second call, explicit official-page inspection with `web_fetch`, correction of
+“stock mraket” in actual search arguments, and a truthful unsupported result
+for a synthetic company.
+
+Production-path email reads pass **3/3** through the running email MCP: account
+list, latest inbox list, and referential read of the first result. The report
+retains no account names, addresses, subjects, bodies, prompts, or answers.
+
+Post-fix representative direct/follow-up coverage also passes for every family:
+notes/calendar 4/4, tasks/documents/memory/skills/Cookbook/search/shell 14/14,
+and email 3/3 in its privacy-preserving runner. The combined legacy verifier's
+metadata-only email guard correctly refused its broader stable inventory; that
+stopped report is not counted as a model failure.
+
+Evidence:
+
+- `reports/clean-v3-stateful-all-r3-20260909.json`
+- `reports/clean-v3-stateful-documents-r2-20260909.json`
+- `reports/clean-v3-search-toggle-final-r6-20260909.json`
+- `reports/clean-v3-search-quality-r3-20260909.json`
+- `reports/clean-v3-email-read-r1-20260909.json`
+- `reports/clean-v3-ten-family-tail-postfix-r1-20260909.json`
+
+These checks verify routing, execution, persistence, follow-up, and selected
+answer-quality invariants. They are not yet the sealed all-action ship score.
+
+## Compact v5 and corrected contract evidence
+
+Compact v5 keeps the compact-v3 surface and adds only development-positive
+field hints for Email, Search/Hugging Face quant selection, and Shell/files.
+A Calendar date hint regressed development and was excluded. The Python tool now
+emits one final bare expression, REPL-style, without duplicating explicit
+`print(...)`; this turns otherwise correct computation calls into visible tool
+evidence for all models.
+
+Under frozen scorer `odysseus.contract.v2.5`, development is 327/344 raw
+(95.06%) and 327/336 scorable (97.32%). Sealed blind is 311/344 raw (90.41%)
+and 311/336 scorable (92.56%), with zero reasoning leakage. Calendar, Shell,
+and Tasks remain below the 90% family ship floor, so the model is not yet a
+full benchmark ship candidate.
+
+Fresh post-deploy real-UI evidence passes: stateful flows 6/6, Email 3/3,
+Search quality/recovery 3/3, private browser 3/3, and VL workflow 3/3. The
+Search check accepts a failed attempt only when a later tool succeeds and the
+final answer remains grounded.

docs/REGULAR_MODEL_TOOL_COMPATIBILITY.md

@@ -0,0 +1,125 @@
+# Regular-model tool compatibility
+
+Last verified: 2026-09-09 through the authenticated 7011 Agent UI as
+`sft_alex_creator`.
+
+This is the legacy-RAG track. The exact model
+`odysseus-qwen3.5-tools-pre-heretic` is excluded and remains on its model-owned
+clean compact runtime.
+
+## Current baseline
+
+| Endpoint | Model | Ten-family result | State |
+|---|---|---:|---|
+| DeepSeek | `deepseek-v4-flash` | 10/10 | passed |
+| DeepSeek | `deepseek-v4-pro` | 10/10 | passed |
+| OpenAI | `gpt-5.5` | 10/10 | passed |
+| OpenAI | `gpt-5.6-sol` | 10/10 | passed |
+| OpenAI | `gpt-5.6-terra` | 10/10 | passed |
+| OpenAI | `gpt-5.6-luna` | 10/10 | passed |
+| OpenRouter | `moonshotai/kimi-k3` | 10/10 | passed |
+| OpenRouter | `x-ai/grok-4.5` | 10/10 | passed |
+| OpenRouter | `qwen/qwen3-vl-235b-a22b-instruct` | 10/10 | passed |
+| OpenRouter | `openai/gpt-5-image` | n/a | image generation; chat tools unsupported |
+| Local `100.69.120.65:8062` | `Qwen/Qwen3.5-9B` | not run | endpoint unavailable |
+| Local `100.69.120.65:8062` | `GLM-5.3-Flash-Alis-MLX-4bit` | not run | endpoint unavailable |
+
+The ten-family baseline covers one read-only functional turn each for notes,
+calendar, email accounts, tasks, documents, memory, skills, Cookbook/admin,
+web search, and shell. It verifies the legacy route, expected native tool call,
+execution result, visible UI answer, and absence of reasoning leakage. It is not
+yet a claim that every mutation/action variant, typo, or follow-up passes.
+
+## Typo and follow-up profile
+
+The stricter real-UI profile sends one misspelled read-only request to every
+family, followed immediately by a noun-free reference to the returned result.
+Read-only follow-ups must not call any tool; search follow-ups may either use
+the existing evidence or fetch the prior link. Across the nine chat-capable API
+models, the composited post-repair result is **178/180 turns (98.89%)**:
+
+| Model | Conversation result |
+|---|---:|
+| `deepseek-v4-flash` | 20/20 |
+| `deepseek-v4-pro` | 20/20 |
+| `gpt-5.5` | 20/20 |
+| `gpt-5.6-sol` | 20/20 |
+| `gpt-5.6-terra` | 20/20 |
+| `gpt-5.6-luna` | 18/20 |
+| `moonshotai/kimi-k3` | 20/20 |
+| `x-ai/grok-4.5` | 20/20 |
+| `qwen/qwen3-vl-235b-a22b-instruct` | 20/20 |
+
+Luna's only remaining family miss is a deliberately misspelled Shell request.
+The correct-spelling baseline passes. The harness does not auto-execute a shell
+command to hide that model-owned limitation.
+
+The shared repair recognizes a uniquely misspelled action verb and family noun,
+then seals only declared safe private reads with immutable canonical arguments.
+This repaired Tasks/Documents/Memory and adjacent read families across providers
+without widening mutation or Shell authority. A compact native-tool instruction
+also tells regular API models to map clear typos to a currently offered tool.
+
+Conversation evidence:
+
+- `reports/regular-model-conversation-flash-r3-20260909.json`
+- `reports/regular-model-conversation-remaining-r1-20260909.json`
+- `reports/regular-model-conversation-repair-r1-20260909.json`
+- `reports/regular-model-conversation-shell-r1-20260909.json`
+- `reports/regular-model-conversation-qwen-repair-r1-20260909.json`
+- `reports/regular-model-conversation-qwen-tail-r1-20260909.json`
+- `reports/regular-model-conversation-qwen-search-r1-20260909.json`
+
+Evidence:
+
+- `reports/regular-model-tools-provider-final-r4-20260909.json` — Flash, GPT-5.5, Kimi: 30/30.
+- `reports/regular-model-tools-repair-r3-20260909.json` — Pro and Sol: 20/20; retained Qwen pre-final 9/10 miss.
+- `reports/regular-qwen-vl-full-r4-20260909.json` — Qwen-VL final family-switch run: 10/10.
+- `reports/regular-model-tools-remaining-20260909.json` — Terra, Luna, Grok: 30/30; records unavailable/unsupported models and pre-repair failures.
+- `reports/regular-model-tools-postfix-r1-20260909.json` — post-hardening
+  rerun: nine chat-capable API models passed 90/90 family turns with zero model
+  failures. Its overall status is non-passing only because the two configured
+  local endpoints were offline; the image-only model remains unsupported.
+
+## Family switch and page inspection
+
+The six-turn switch/back flow covers notes → calendar → notes from prior
+evidence → web search → explicit `web_fetch` → calendar from prior evidence.
+All nine API models have a clean 6/6 reproduction (**54/54**). Kimi skipped
+search once in the retained first run and passed a fresh reproduction; that
+variability remains visible instead of being erased.
+
+Evidence:
+
+- `reports/regular-model-switchback-flash-r2-20260909.json`
+- `reports/regular-model-switchback-remaining-r1-20260909.json`
+- `reports/regular-model-switchback-kimi-r1-20260909.json`
+
+## Repair that produced the clean baseline
+
+Regular models no longer inherit up to three stale tool families into every
+explicit new request. Referential follow-ups still resolve from typed recent
+tool evidence, while explicit family switches receive the current family only.
+Safe required reads use `active_capabilities`, so stale offered context cannot
+disable their immutable operation. The stream layer also stops an exact long
+block repeated twice instead of waiting for a provider's full timeout.
+
+The composer no longer treats generic words such as “source”, “system”, “app”,
+or “review” as authority to silently enable Bash. Explicit shell, terminal,
+repository, code-file, and direct coding requests retain workspace
+auto-escalation. This is a shared UI authority fix, not a model-name exception.
+
+Run a bounded subset with:
+
+```sh
+MODELS='deepseek-v4-flash,gpt-5.5' \
+FAMILIES='notes,calendar' WORKERS=2 \
+REPORT_PATH=reports/regular-model-check.json \
+node scripts/verify_regular_model_tools.mjs
+```
+
+Set `PROFILE=conversation` to run the typo plus follow-up profile.
+
+The runner discovers only enabled pinned models (visible cached local models
+when no pins exist), retains no tool outputs or private rows, and deletes only
+the exact sessions it creates.

docs/SEARCH_HARNESS_AUDIT_20260909.md

@@ -0,0 +1,92 @@
+# Search and compact-tool experiment — 2026-09-09
+
+## Decision
+
+Keep the normal routed profile on 7011. The all-tools compact experiment is
+implemented but **disabled**: direct routing success did not translate into a
+working Agent UI. Do not retrain or promote a profile on these measurements.
+
+## Changes
+
+- Short public-web lookups on the target model have an execution budget: two
+  distinct token-normalized searches, one fetch, and up to three browser calls
+  after the two searches. This bounds attempts, not just recovery prose. Existing
+  permissions still apply; this does not make unavailable tools executable.
+- Failed/weak searches reach the model for evaluation and query refinement,
+  instead of the earlier unconditional terminal evidence veto. Some legacy
+  heuristics and official-site shortcuts remain; this is not a completed rewrite.
+- Search providers retain query/engine/date provenance. Unconfigured credentialed
+  fallbacks are skipped. When SearXNG is the sole configured usable provider, Yep
+  on the same instance is an additional fallback.
+- An unavailable warm-only family no longer vetoes an otherwise ordinary reply.
+- The all-tools experiment offers the trained compact inventory subject to
+  permissions. It requires the exact test-owner environment flag and exact model
+  match. The temporary service flag was removed after failed UI testing.
+
+## Evidence and limits
+
+| Measurement | Result | What it establishes |
+|---|---|---|
+| Focused Python regression suite | 401 passed | Covered policy, contract, provider and recovery-budget behavior |
+| Direct family-only compact schemas | 10/10 tool routing | Small public smoke test, not functional or blind accuracy |
+| Direct all-family compact schemas | 10/10 tool routing | Inventory did not break these first calls; roughly 3–4x slower in this run |
+| Full compact Agent UI, revision 3 | All eight turns failed one or more checks | Not suitable for activation; leaks, duplicate/incorrect rendering or missing expected calls |
+| Normal-profile final UI control | Five passed checks, two product failures, one capture error | Not accepted; suite status incomplete |
+| Clean synthetic notes tool-result continuation | Clean answer with both schema sizes | Model can continue correctly on that isolated input, not proof that UI failure is solely harness |
+
+Direct probes used temperature 0; the UI target-model sampling path can cap at
+0.2. Prompts, history and tool-result serialization also differ. Match those
+before attributing UI failures to weights versus harness. Existing UI checks are
+not a grounded factual-answer benchmark. Unit tests are not UI acceptance.
+
+Normal-profile control details: notes initial, both calendar turns, AI search
+initial, and history initial passed the automated checks. Notes follow-up hit a
+Playwright `Network.getResponseBody` capture error and is inconclusive. AI search
+follow-up explicitly requested a summary with no tools, but the contract still
+required `search_browser` and returned a permission failure. The history search
+follow-up failed the visible-leak check. These are separate from source relevance;
+the earlier warm-only fix did not cover classification as an active requirement.
+There is no matched pre-change control establishing a net improvement.
+
+Provider isolation bypassed app relevance filters. Bing general often returned
+broad or unrelated results despite the full query. Google/Mojeek returned no
+results, DDG hit CAPTCHA, and Presearch timed out. Yep returned useful PostgreSQL
+documentation, but was weak or empty for several other questions. Engine health
+and source quality remain unresolved. Fallback cannot help when an earlier weak
+result survives filtering; there is no claim of universal relevance here.
+
+## Reproduce and inspect
+
+- `scripts/audit_search_pipeline.py`: raw provider comparison, no model.
+- `scripts/compare_compact_tool_inventory.py`: read-only model schema comparison;
+  proposed calls are never executed.
+- `scripts/verify_agent_turn_contract.mjs`: real 7011 Agent UI and persisted-history
+  checks. Use the dedicated test account; reports can contain private tool data.
+- `reports/search-provider-isolation.json`, `reports/search-yep-isolation.json`:
+  public provider evidence.
+- `reports/compact-inventory-ablation.json`: direct routing probe.
+- `reports/full-compact-ui-audit-r3.json`: completed rejected UI experiment.
+  Earlier experiment reports include an initialization error and an aborted run;
+  do not combine them into an accuracy score.
+- `reports/routed-control-ui-audit-final.json`: normal-profile control replay;
+  eight attempts, incomplete because of the capture error; failures retained.
+
+Focused suite:
+
+```sh
+/home/pewds/odysseus-cookbook-fresh/.venv/bin/pytest -q tests/test_turn_contract.py tests/test_turn_contract_integration.py tests/test_service_search_provider_guards.py tests/test_web_recovery_budget.py tests/test_tool_policy.py
+```
+
+UI replay (read-only prompts, creates test chats):
+
+```sh
+node scripts/verify_agent_turn_contract.mjs --families notes,calendar,search_ai,search_history --pairs notes:11,calendar:11,search_ai:11,search_history:11 --max-turns 8 --total-ms 360000 --turn-ms 45000 --report reports/routed-control-ui-audit-final.json
+```
+
+## Next discriminating test
+
+Replay the same captured UI request directly, preserving sampling, compact
+schemas, history and tool results. Then change one layer at a time. Separately
+score retrieved-source relevance and supported answers. Replace failing generic
+boundaries only when the replay identifies them; do not add rules for individual
+user phrasings or treat successful tool routing as successful execution.

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
+/home/pewds/hermes-skills-reference. 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.

launch-windows.ps1

@@ -105,6 +105,14 @@ if (-not $pyExe) {
     }
 }
 
+if ($pyExe -like "*WindowsApps*python.exe") {
+    $pyCmd = Get-Command py -ErrorAction SilentlyContinue
+    if ($pyCmd) {
+        $pyExe = $pyCmd.Source
+        $pyArgs = @("-3.11")
+    }
+}
+
 if (-not $pyExe) {
     Fail "Couldn't find Python 3.11+ for Windows setup. Install Python 3.11+ (or open the Python launcher with 'py -3.11') from https://www.python.org/downloads/, then re-run this script."
 }
@@ -155,6 +163,10 @@ if (Test-Path $cudaBase) {
 }
 
 # 7. Start the server (use `python -m uvicorn` - bare `uvicorn` may not be on PATH)
+# -Port only reaches uvicorn as a flag. Everything that builds a URL for this
+# instance - internal_api_base(), companion pairing, the MCP OAuth callback -
+# reads APP_PORT, so set it too or they all assume 7000.
+$env:APP_PORT = $Port
 Write-Step ("Starting Odysseus at http://{0}:{1}" -f $BindHost, $Port)
 Write-Host "Press Ctrl+C to stop."
 Write-Host ""

launcher.py

@@ -0,0 +1,142 @@
+# launcher.py
+"""Dedicated entrypoint for the standalone Windows portable launcher.
+
+Handles:
+- Immediate GUI splash screen creation using tkinter.
+- Suppressing console stream crashes in windowed GUI mode via NullWriter.
+- Spawning system tray icon via pystray and Pillow (lazy-loaded).
+- Auto-opening default browser pointing to the running backend.
+- Launching the FastAPI server (importing and running app.py).
+"""
+import os
+import sys
+import threading
+import time
+import webbrowser
+
+# Define a dummy NullWriter to suppress standard stream crashes (isatty etc.) in GUI mode
+class NullWriter:
+    def write(self, text):
+        pass
+    def flush(self):
+        pass
+    def isatty(self):
+        return False
+
+if sys.stdout is None:
+    sys.stdout = NullWriter()
+if sys.stderr is None:
+    sys.stderr = NullWriter()
+
+
+splash_root = None
+
+# If running from a frozen PyInstaller bundle, launch the splash screen IMMEDIATELY
+if getattr(sys, 'frozen', False):
+    import tkinter as tk
+
+    def show_splash_instantly():
+        global splash_root
+        try:
+            splash_root = tk.Tk()
+            splash_root.title("Odysseus")
+            splash_root.overrideredirect(True)
+            splash_root.configure(bg="#1a1c23")
+
+            # Accented borders
+            splash_root.config(highlightbackground="#e06c75", highlightcolor="#e06c75", highlightthickness=1)
+
+            w, h = 360, 160
+            ws = splash_root.winfo_screenwidth()
+            hs = splash_root.winfo_screenheight()
+            x = (ws - w) // 2
+            y = (hs - h) // 2
+            splash_root.geometry(f"{w}x{h}+{x}+{y}")
+
+            tk.Label(splash_root, text="⛵ Odysseus", font=("Segoe UI", 22, "bold"), bg="#1a1c23", fg="#e06c75").pack(pady=(22, 2))
+            tk.Label(splash_root, text="Launching background services...", font=("Segoe UI", 10), bg="#1a1c23", fg="#d1d4e0").pack(pady=2)
+            tk.Label(splash_root, text="Please wait, this will take a few seconds.", font=("Segoe UI", 8, "italic"), bg="#1a1c23", fg="#5c6370").pack(pady=(12, 0))
+
+            splash_root.attributes("-topmost", True)
+            splash_root.mainloop()
+        except Exception:
+            pass
+
+    # Launch the GUI splash screen immediately on a background thread
+    threading.Thread(target=show_splash_instantly, daemon=True).start()
+
+
+def create_tray_image():
+    # Generate a beautiful 64x64 icon matching Odysseus brand red accent (#e06c75)
+    from PIL import Image, ImageDraw
+    image = Image.new('RGBA', (64, 64), (0, 0, 0, 0))
+    dc = ImageDraw.Draw(image)
+    accent_red = (224, 108, 117, 255)
+    light_red = (224, 108, 117, 150)
+
+    # Draw premium sailing boat
+    dc.polygon([(32, 10), (32, 45), (12, 45)], fill=accent_red)
+    dc.polygon([(32, 18), (32, 45), (48, 45)], fill=light_red)
+    dc.polygon([(8, 48), (56, 48), (44, 56), (20, 56)], fill=accent_red)
+    return image
+
+
+def on_open_browser(icon, item, url):
+    webbrowser.open(url)
+
+
+def on_exit(icon, item):
+    icon.stop()
+    os._exit(0)
+
+
+def setup_system_tray(url):
+    try:
+        import pystray
+        icon_img = create_tray_image()
+        menu = (
+            pystray.MenuItem('Open Odysseus', lambda icon, item: on_open_browser(icon, item, url), default=True),
+            pystray.MenuItem('Exit', on_exit)
+        )
+        tray_icon = pystray.Icon(
+            "Odysseus",
+            icon_img,
+            "Odysseus",
+            menu
+        )
+        tray_icon.run()
+    except Exception:
+        pass
+
+
+def open_browser(url):
+    # Allow uvicorn and app lifecycles to complete warmups
+    time.sleep(3.5)
+
+    # Safely close the splash screen
+    try:
+        global splash_root
+        if splash_root:
+            splash_root.after(0, splash_root.destroy)
+    except Exception:
+        pass
+
+    webbrowser.open(url)
+
+
+if __name__ == "__main__":
+    import uvicorn
+    # Import the FastAPI app from app.py
+    from app import app
+
+    bind_host = os.getenv("APP_BIND", "127.0.0.1")
+    bind_port = int(os.getenv("APP_PORT", "7011"))
+    url = f"http://{bind_host}:{bind_port}"
+
+    if getattr(sys, 'frozen', False):
+        # Start browser manager thread
+        threading.Thread(target=open_browser, args=(url,), daemon=True).start()
+        # Start system tray manager thread
+        threading.Thread(target=setup_system_tray, args=(url,), daemon=True).start()
+
+    uvicorn.run(app, host=bind_host, port=bind_port, log_level="info")

licenses/KaTeX-MIT-LICENSE.txt

@@ -0,0 +1,21 @@
+The MIT License (MIT)
+
+Copyright (c) 2013-2020 Khan Academy and other contributors
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

licenses/Mermaid-MIT-LICENSE.txt

@@ -0,0 +1,21 @@
+The MIT License (MIT)
+
+Copyright (c) 2014 - 2022 Knut Sveidqvist
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

licenses/OpenDyslexic-OFL.txt

@@ -0,0 +1,94 @@
+Copyright (c) 2019-07-29, Abbie Gonzalez (https://abbiecod.es|support@abbiecod.es),
+with Reserved Font Name OpenDyslexic.
+Copyright (c) 12/2012 - 2019
+This Font Software is licensed under the SIL Open Font License, Version 1.1.
+This license is copied below, and is also available with a FAQ at:
+http://scripts.sil.org/OFL
+
+
+-----------------------------------------------------------
+SIL OPEN FONT LICENSE Version 1.1 - 26 February 2007
+-----------------------------------------------------------
+
+PREAMBLE
+The goals of the Open Font License (OFL) are to stimulate worldwide
+development of collaborative font projects, to support the font creation
+efforts of academic and linguistic communities, and to provide a free and
+open framework in which fonts may be shared and improved in partnership
+with others.
+
+The OFL allows the licensed fonts to be used, studied, modified and
+redistributed freely as long as they are not sold by themselves. The
+fonts, including any derivative works, can be bundled, embedded,
+redistributed and/or sold with any software provided that any reserved
+names are not used by derivative works. The fonts and derivatives,
+however, cannot be released under any other type of license. The
+requirement for fonts to remain under this license does not apply
+to any document created using the fonts or their derivatives.
+
+DEFINITIONS
+"Font Software" refers to the set of files released by the Copyright
+Holder(s) under this license and clearly marked as such. This may
+include source files, build scripts and documentation.
+
+"Reserved Font Name" refers to any names specified as such after the
+copyright statement(s).
+
+"Original Version" refers to the collection of Font Software components as
+distributed by the Copyright Holder(s).
+
+"Modified Version" refers to any derivative made by adding to, deleting,
+or substituting -- in part or in whole -- any of the components of the
+Original Version, by changing formats or by porting the Font Software to a
+new environment.
+
+"Author" refers to any designer, engineer, programmer, technical
+writer or other person who contributed to the Font Software.
+
+PERMISSION & CONDITIONS
+Permission is hereby granted, free of charge, to any person obtaining
+a copy of the Font Software, to use, study, copy, merge, embed, modify,
+redistribute, and sell modified and unmodified copies of the Font
+Software, subject to the following conditions:
+
+1) Neither the Font Software nor any of its individual components,
+in Original or Modified Versions, may be sold by itself.
+
+2) Original or Modified Versions of the Font Software may be bundled,
+redistributed and/or sold with any software, provided that each copy
+contains the above copyright notice and this license. These can be
+included either as stand-alone text files, human-readable headers or
+in the appropriate machine-readable metadata fields within text or
+binary files as long as those fields can be easily viewed by the user.
+
+3) No Modified Version of the Font Software may use the Reserved Font
+Name(s) unless explicit written permission is granted by the corresponding
+Copyright Holder. This restriction only applies to the primary font name as
+presented to the users.
+
+4) The name(s) of the Copyright Holder(s) or the Author(s) of the Font
+Software shall not be used to promote, endorse or advertise any
+Modified Version, except to acknowledge the contribution(s) of the
+Copyright Holder(s) and the Author(s) or with their explicit written
+permission.
+
+5) The Font Software, modified or unmodified, in part or in whole,
+must be distributed entirely under this license, and must not be
+distributed under any other license. The requirement for fonts to
+remain under this license does not apply to any document created
+using the Font Software.
+
+TERMINATION
+This license becomes null and void if any of the above conditions are
+not met.
+
+DISCLAIMER
+THE FONT SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO ANY WARRANTIES OF
+MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT
+OF COPYRIGHT, PATENT, TRADEMARK, OR OTHER RIGHT. IN NO EVENT SHALL THE
+COPYRIGHT HOLDER BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
+INCLUDING ANY GENERAL, SPECIAL, INDIRECT, INCIDENTAL, OR CONSEQUENTIAL
+DAMAGES, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+FROM, OUT OF THE USE OR INABILITY TO USE THE FONT SOFTWARE OR FROM
+OTHER DEALINGS IN THE FONT SOFTWARE.

mcp_servers/image_gen_server.py

@@ -73,7 +73,7 @@ async def call_tool(name: str, arguments: dict) -> list[TextContent]:
         if not model_spec:
             for candidate in ("gpt-image-1.5", "gpt-image-1", "dall-e-3"):
                 try:
-                    _resolve_model(candidate)
+                    await asyncio.to_thread(_resolve_model, candidate)
                     model_spec = candidate
                     break
                 except ValueError:
@@ -81,7 +81,13 @@ async def call_tool(name: str, arguments: dict) -> list[TextContent]:
             if not model_spec:
                 return [TextContent(type="text", text="Error: No image model found. Configure one in Admin.")]
 
-        url, model_id, headers = _resolve_model(model_spec)
+        try:
+            url, model_id, headers = await asyncio.to_thread(_resolve_model, model_spec, model_type="image")
+        except ValueError:
+            _lower_model_spec = model_spec.lower()
+            if not any(_name in _lower_model_spec for _name in ("gpt-image", "dall-e")):
+                raise
+            url, model_id, headers = await asyncio.to_thread(_resolve_model, model_spec)
 
         is_gpt_image = "gpt-image" in model_id.lower()
         base_url = url.replace("/chat/completions", "").replace("/v1/messages", "").rstrip("/")

mcp_servers/memory_server.py

@@ -6,6 +6,7 @@ Imports MemoryManager and MemoryVectorStore from the Odysseus codebase.
 """
 
 import asyncio
+import os
 import sys
 import time
 from pathlib import Path
@@ -16,6 +17,8 @@ from mcp.types import Tool, TextContent
 
 sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
 
+from src.memory import MemoryStoreUnreadable
+
 server = Server("memory")
 
 # Late-initialized managers (set during first tool call)
@@ -23,6 +26,71 @@ _memory_manager = None
 _memory_vector = None
 _initialized = False
 
+_OWNER_ENV_KEYS = ("ODYSSEUS_MCP_MEMORY_OWNER", "ODYSSEUS_MEMORY_OWNER")
+_OWNER_SCOPE_ERROR = (
+    "Error: Memory MCP owner is not configured for an owner-scoped memory store. "
+    "Set ODYSSEUS_MCP_MEMORY_OWNER for this server or use the owner-aware native memory tool."
+)
+_UNREADABLE_STORE_ERROR = (
+    "Error: Memory store is temporarily unreadable — nothing was saved. "
+    "Repair or restore memory.json, then retry."
+)
+
+
+def _configured_owner() -> str | None:
+    for key in _OWNER_ENV_KEYS:
+        owner = os.environ.get(key, "").strip()
+        if owner:
+            return owner
+    return None
+
+
+def _entry_owner(entry: dict) -> str | None:
+    owner = entry.get("owner")
+    if owner is None:
+        return None
+    owner_text = str(owner).strip()
+    return owner_text or None
+
+
+def _owner_scoped_store(entries: list[dict]) -> bool:
+    return any(_entry_owner(entry) for entry in entries if isinstance(entry, dict))
+
+
+def _scope_entries(for_update: bool = False) -> tuple[str | None, list[dict], list[dict], str | None]:
+    """Return configured owner, all entries, visible entries, and optional error.
+
+    ``for_update=True`` is for read-modify-write callers. They save the ``all
+    entries`` list back, so an unreadable store must be reported as an error
+    instead of degrading to ``[]`` — otherwise the save writes their one new
+    entry over the whole store (issue #5673).
+    """
+    if for_update:
+        try:
+            entries = _memory_manager.load_all_for_update()
+        except MemoryStoreUnreadable as e:
+            return None, [], [], f"{_UNREADABLE_STORE_ERROR} ({e})"
+    else:
+        entries = _memory_manager.load_all()
+    owner = _configured_owner()
+    if owner is None and _owner_scoped_store(entries):
+        return None, entries, [], _OWNER_SCOPE_ERROR
+    if owner is None:
+        visible = [
+            entry for entry in entries
+            if isinstance(entry, dict) and _entry_owner(entry) is None
+        ]
+    else:
+        visible = [
+            entry for entry in entries
+            if isinstance(entry, dict) and _entry_owner(entry) == owner
+        ]
+    return owner, entries, visible, None
+
+
+def _text_result(text: str) -> list[TextContent]:
+    return [TextContent(type="text", text=text)]
+
 
 def _ensure_init():
     """Lazy-init memory managers on first use."""
@@ -75,24 +143,26 @@ async def list_tools() -> list[Tool]:
 @server.call_tool()
 async def call_tool(name: str, arguments: dict) -> list[TextContent]:
     if name != "manage_memory":
-        return [TextContent(type="text", text=f"Unknown tool: {name}")]
+        return _text_result(f"Unknown tool: {name}")
 
     _ensure_init()
     if not _memory_manager:
-        return [TextContent(type="text", text="Error: Memory manager not available")]
+        return _text_result("Error: Memory manager not available")
 
     action = arguments.get("action", "")
 
     if action == "list":
         category_filter = arguments.get("category", "")
-        memories = _memory_manager.load()
+        _owner, _all_memories, memories, scope_error = _scope_entries()
+        if scope_error:
+            return _text_result(scope_error)
         if category_filter:
             memories = [m for m in memories if m.get("category", "").lower() == category_filter.lower()]
         if not memories:
             msg = "No memories found"
             if category_filter:
                 msg += f" in category '{category_filter}'"
-            return [TextContent(type="text", text=msg + ".")]
+            return _text_result(msg + ".")
 
         lines = [f"Found {len(memories)} memory entries:\n"]
         for m in memories:
@@ -102,15 +172,17 @@ async def call_tool(name: str, arguments: dict) -> list[TextContent]:
             if len(text) > 150:
                 text = text[:150] + "..."
             lines.append(f"- [{cat}] `{mid}` — {text}")
-        return [TextContent(type="text", text="\n".join(lines))]
+        return _text_result("\n".join(lines))
 
     elif action == "add":
         text = arguments.get("text", "")
         category = arguments.get("category", "fact")
         if not text:
-            return [TextContent(type="text", text="Error: Memory text cannot be empty")]
-        entry = _memory_manager.add_entry(text, source="ai_agent", category=category)
-        memories = _memory_manager.load_all()
+            return _text_result("Error: Memory text cannot be empty")
+        owner, memories, _visible, scope_error = _scope_entries(for_update=True)
+        if scope_error:
+            return _text_result(scope_error)
+        entry = _memory_manager.add_entry(text, source="ai_agent", category=category, owner=owner)
         memories.append(entry)
         _memory_manager.save(memories)
         if _memory_vector and _memory_vector.healthy:
@@ -118,25 +190,28 @@ async def call_tool(name: str, arguments: dict) -> list[TextContent]:
                 _memory_vector.add(entry["id"], text)
             except Exception:
                 pass
-        return [TextContent(type="text", text=f"Memory added: [{category}] {text} (id: {entry['id'][:8]})")]
+        return _text_result(f"Memory added: [{category}] {text} (id: {entry['id'][:8]})")
 
     elif action == "edit":
         memory_id = arguments.get("memory_id", "")
         new_text = arguments.get("text", "")
         if not memory_id or not new_text:
-            return [TextContent(type="text", text="Error: edit needs memory_id and text")]
-        memories = _memory_manager.load_all()
-        found = False
+            return _text_result("Error: edit needs memory_id and text")
+        _owner, memories, visible, scope_error = _scope_entries()
+        if scope_error:
+            return _text_result(scope_error)
         full_id = None
-        for m in memories:
+        for m in visible:
             if m.get("id", "").startswith(memory_id):
-                m["text"] = new_text
-                m["timestamp"] = int(time.time())
-                found = True
                 full_id = m["id"]
                 break
-        if not found:
-            return [TextContent(type="text", text=f"Error: Memory '{memory_id}' not found")]
+        if not full_id:
+            return _text_result(f"Error: Memory '{memory_id}' not found")
+        for m in memories:
+            if m.get("id") == full_id:
+                m["text"] = new_text
+                m["timestamp"] = int(time.time())
+                break
         _memory_manager.save(memories)
         if _memory_vector and _memory_vector.healthy and full_id:
             try:
@@ -144,24 +219,26 @@ async def call_tool(name: str, arguments: dict) -> list[TextContent]:
                 _memory_vector.add(full_id, new_text)
             except Exception:
                 pass
-        return [TextContent(type="text", text=f"Memory updated: {new_text}")]
+        return _text_result(f"Memory updated: {new_text}")
 
     elif action == "delete":
         memory_id = arguments.get("memory_id", "")
         if not memory_id:
-            return [TextContent(type="text", text="Error: delete needs memory_id")]
-        memories = _memory_manager.load_all()
+            return _text_result("Error: delete needs memory_id")
+        _owner, memories, visible, scope_error = _scope_entries()
+        if scope_error:
+            return _text_result(scope_error)
         full_id = None
         deleted_text = ""
         deleted_category = ""
-        for m in memories:
+        for m in visible:
             if m.get("id", "").startswith(memory_id):
                 full_id = m["id"]
                 deleted_text = m.get("text", "")
                 deleted_category = m.get("category", "")
                 break
         if not full_id:
-            return [TextContent(type="text", text=f"Error: Memory '{memory_id}' not found")]
+            return _text_result(f"Error: Memory '{memory_id}' not found")
         memories = [m for m in memories if m.get("id") != full_id]
         _memory_manager.save(memories)
         if _memory_vector and _memory_vector.healthy and full_id:
@@ -171,30 +248,32 @@ async def call_tool(name: str, arguments: dict) -> list[TextContent]:
                 pass
         cat = f"[{deleted_category}] " if deleted_category else ""
         snippet = deleted_text if len(deleted_text) <= 120 else deleted_text[:117] + "..."
-        return [TextContent(type="text", text=f"Memory deleted: {cat}{snippet} (id: {memory_id})")]
+        return _text_result(f"Memory deleted: {cat}{snippet} (id: {memory_id})")
 
     elif action == "search":
         query = arguments.get("text", "")
         if not query:
-            return [TextContent(type="text", text="Error: search needs text (query)")]
-        memories = _memory_manager.load()
+            return _text_result("Error: search needs text (query)")
+        _owner, _all_memories, memories, scope_error = _scope_entries()
+        if scope_error:
+            return _text_result(scope_error)
         if hasattr(_memory_manager, 'get_relevant_memories'):
             results = _memory_manager.get_relevant_memories(query, memories, threshold=0.05, max_items=20)
         else:
             query_lower = query.lower()
             results = [m for m in memories if query_lower in m.get("text", "").lower()][:20]
         if not results:
-            return [TextContent(type="text", text=f"No memories found matching '{query}'.")]
+            return _text_result(f"No memories found matching '{query}'.")
         lines = [f"Found {len(results)} matching memories:\n"]
         for m in results:
             cat = m.get("category", "fact")
             mid = m.get("id", "?")[:8]
             text = m.get("text", "")
             lines.append(f"- [{cat}] `{mid}` — {text}")
-        return [TextContent(type="text", text="\n".join(lines))]
+        return _text_result("\n".join(lines))
 
     else:
-        return [TextContent(type="text", text=f"Error: Unknown action '{action}'. Use: list, add, edit, delete, search")]
+        return _text_result(f"Error: Unknown action '{action}'. Use: list, add, edit, delete, search")
 
 
 async def run():

package-lock.json

@@ -4,93 +4,84 @@
   "requires": true,
   "packages": {
     "": {
-      "dependencies": {
-        "@anthropic-ai/sdk": "^0.104.1"
-      },
+      "name": "odysseus",
       "devDependencies": {
-        "@antithesishq/bombadil": "^0.5.0"
-      }
-    },
-    "node_modules/@anthropic-ai/sdk": {
-      "version": "0.104.1",
-      "resolved": "https://registry.npmjs.org/@anthropic-ai/sdk/-/sdk-0.104.1.tgz",
-      "integrity": "sha512-gGACa/+IaiXzRRmF96aOhamoBgapKRBiFWbmmTFP8aMkpaEcuStF+Q61bjo4vPxBM7gqWJNZqsngslRdnLHv0Q==",
-      "license": "MIT",
-      "dependencies": {
-        "json-schema-to-ts": "^3.1.1",
-        "standardwebhooks": "^1.0.0"
-      },
-      "bin": {
-        "anthropic-ai-sdk": "bin/cli"
-      },
-      "peerDependencies": {
-        "zod": "^3.25.0 || ^4.0.0"
-      },
-      "peerDependenciesMeta": {
-        "zod": {
-          "optional": true
-        }
+        "@antithesishq/bombadil": "^0.7.0",
+        "@playwright/test": "^1.62.1"
       }
     },
     "node_modules/@antithesishq/bombadil": {
-      "version": "0.5.0",
-      "resolved": "https://registry.npmjs.org/@antithesishq/bombadil/-/bombadil-0.5.0.tgz",
-      "integrity": "sha512-s0zImmr0iyvSP6QcVLvf40CUiZYIdWBAxiq20uhzujwvfitYa3PGJN652k/pLtVccHM/JrGQxZdvLnihZpltHA==",
+      "version": "0.7.0",
+      "resolved": "https://registry.npmjs.org/@antithesishq/bombadil/-/bombadil-0.7.0.tgz",
+      "integrity": "sha512-alJmnphJ/iUoL5mCsnV3DwtajGy/sEQ3NJJCiMhgjqXshSq2BUtAs0vqdXEiiSkB8HbsOX5CLrAcaogYdwfAJg==",
       "dev": true,
       "license": "MIT",
       "bin": {
         "bombadil": "bin/bombadil.js"
       }
     },
-    "node_modules/@babel/runtime": {
-      "version": "7.29.7",
-      "resolved": "https://registry.npmjs.org/@babel/runtime/-/runtime-7.29.7.tgz",
-      "integrity": "sha512-Nq8OhGWiZIZGV6hLHoyAKLLcJihP/xFeBMGJoUrxTX2psI8dCifzLhZISFb+VWS3wFMRDmCGw5R+dOySCqPLhw==",
-      "license": "MIT",
-      "engines": {
-        "node": ">=6.9.0"
-      }
-    },
-    "node_modules/@stablelib/base64": {
-      "version": "1.0.1",
-      "resolved": "https://registry.npmjs.org/@stablelib/base64/-/base64-1.0.1.tgz",
-      "integrity": "sha512-1bnPQqSxSuc3Ii6MhBysoWCg58j97aUjuCSZrGSmDxNqtytIi0k8utUenAwTZN4V5mXXYGsVUI9zeBqy+jBOSQ==",
-      "license": "MIT"
-    },
-    "node_modules/fast-sha256": {
-      "version": "1.3.0",
-      "resolved": "https://registry.npmjs.org/fast-sha256/-/fast-sha256-1.3.0.tgz",
-      "integrity": "sha512-n11RGP/lrWEFI/bWdygLxhI+pVeo1ZYIVwvvPkW7azl/rOy+F3HYRZ2K5zeE9mmkhQppyv9sQFx0JM9UabnpPQ==",
-      "license": "Unlicense"
-    },
-    "node_modules/json-schema-to-ts": {
-      "version": "3.1.1",
-      "resolved": "https://registry.npmjs.org/json-schema-to-ts/-/json-schema-to-ts-3.1.1.tgz",
-      "integrity": "sha512-+DWg8jCJG2TEnpy7kOm/7/AxaYoaRbjVB4LFZLySZlWn8exGs3A4OLJR966cVvU26N7X9TWxl+Jsw7dzAqKT6g==",
-      "license": "MIT",
+    "node_modules/@playwright/test": {
+      "version": "1.62.1",
+      "resolved": "https://registry.npmjs.org/@playwright/test/-/test-1.62.1.tgz",
+      "integrity": "sha512-DTcUc8qii+cpHvtOwggMtBRMjKZHXYWdw8syRYu2vtzuq4Wxphqq4NfCs5Zt44L6mA8rfDfj+PHnxFc/FeK6mQ==",
+      "dev": true,
+      "license": "Apache-2.0",
       "dependencies": {
-        "@babel/runtime": "^7.18.3",
-        "ts-algebra": "^2.0.0"
+        "playwright": "1.62.1"
+      },
+      "bin": {
+        "playwright": "cli.js"
       },
       "engines": {
-        "node": ">=16"
+        "node": ">=20"
       }
     },
-    "node_modules/standardwebhooks": {
-      "version": "1.0.0",
-      "resolved": "https://registry.npmjs.org/standardwebhooks/-/standardwebhooks-1.0.0.tgz",
-      "integrity": "sha512-BbHGOQK9olHPMvQNHWul6MYlrRTAOKn03rOe4A8O3CLWhNf4YHBqq2HJKKC+sfqpxiBY52pNeesD6jIiLDz8jg==",
+    "node_modules/fsevents": {
+      "version": "2.3.2",
+      "resolved": "https://registry.npmjs.org/fsevents/-/fsevents-2.3.2.tgz",
+      "integrity": "sha512-xiqMQR4xAeHTuB9uWm+fFRcIOgKBMiOBP+eXiyT7jsgVCq1bkVygt00oASowB7EdtpOHaaPgKt812P9ab+DDKA==",
+      "dev": true,
+      "hasInstallScript": true,
       "license": "MIT",
-      "dependencies": {
-        "@stablelib/base64": "^1.0.0",
-        "fast-sha256": "^1.3.0"
+      "optional": true,
+      "os": [
+        "darwin"
+      ],
+      "engines": {
+        "node": "^8.16.0 || ^10.6.0 || >=11.0.0"
       }
     },
-    "node_modules/ts-algebra": {
-      "version": "2.0.0",
-      "resolved": "https://registry.npmjs.org/ts-algebra/-/ts-algebra-2.0.0.tgz",
-      "integrity": "sha512-FPAhNPFMrkwz76P7cdjdmiShwMynZYN6SgOujD1urY4oNm80Ou9oMdmbR45LotcKOXoy7wSmHkRFE6Mxbrhefw==",
-      "license": "MIT"
+    "node_modules/playwright": {
+      "version": "1.62.1",
+      "resolved": "https://registry.npmjs.org/playwright/-/playwright-1.62.1.tgz",
+      "integrity": "sha512-0M+L3LAD8/nm554LOla9Ayx0j0tmFZ0FBcoQ7F1VuVHpM/XpiC8RcDzBQB8W5+hA8L22THxELzeF+2WcUzvcLg==",
+      "dev": true,
+      "license": "Apache-2.0",
+      "dependencies": {
+        "playwright-core": "1.62.1"
+      },
+      "bin": {
+        "playwright": "cli.js"
+      },
+      "engines": {
+        "node": ">=20"
+      },
+      "optionalDependencies": {
+        "fsevents": "2.3.2"
+      }
+    },
+    "node_modules/playwright-core": {
+      "version": "1.62.1",
+      "resolved": "https://registry.npmjs.org/playwright-core/-/playwright-core-1.62.1.tgz",
+      "integrity": "sha512-wPYSwEBJY9GHraISXqyqtx0na0LpO3XEX7jNDhntbex7tzUS7kLnZsOlFruFJB4Hi/rhDMjXGqHewDZ68nYZVw==",
+      "dev": true,
+      "license": "Apache-2.0",
+      "bin": {
+        "playwright-core": "cli.js"
+      },
+      "engines": {
+        "node": ">=20"
+      }
     }
   }
 }

package.json

@@ -1,12 +1,18 @@
 {
+  "name": "odysseus",
+  "private": true,
   "repository": {
     "type": "git",
-    "url": "https://github.com/pewdiepie-archdaemon/odysseus.git"
+    "url": "https://github.com/odysseus-dev/odysseus.git"
+  },
+  "scripts": {
+    "test:photo-editor": "playwright test --config tests/e2e/playwright.config.js",
+    "test:photo-editor:install": "playwright install chromium firefox webkit",
+    "test:photo-editor:firefox": "PHOTO_EDITOR_E2E_BROWSER=firefox playwright test --config tests/e2e/playwright.config.js",
+    "test:photo-editor:webkit": "PHOTO_EDITOR_E2E_BROWSER=webkit playwright test --config tests/e2e/playwright.config.js"
   },
   "devDependencies": {
-    "@antithesishq/bombadil": "^0.5.0"
-  },
-  "dependencies": {
-    "@anthropic-ai/sdk": "^0.104.1"
+    "@antithesishq/bombadil": "^0.7.0",
+    "@playwright/test": "^1.62.1"
   }
 }

plans/ODYSSEUS_TOOL_HARDENING_PLAN.md

@@ -0,0 +1,326 @@
+# Odysseus Tool Runtime Hardening Plan
+
+## Objective
+
+Ship `odysseus-qwen3.5-tools-pre-heretic` with one compact, model-specific tool
+runtime that supports realistic multi-turn use. Keep the existing RAG runtime
+unchanged for every other model. Prove routing, execution, answer quality,
+follow-ups, safety, rendering, latency, and native image/VL understanding through
+the real 7011 Agent UI.
+
+Current evidence is a baseline, not a ship claim:
+
+- Corrected v2.5 + compact-v5 development is 327/344 raw (95.06%) and
+  327/336 scorable (97.32%). Sealed blind is 311/344 raw (90.41%) and
+  311/336 scorable (92.56%), with zero reasoning leakage.
+- Notes, Skills, and Cookbook/admin clear 95% scorable blind. Calendar 87.5%,
+  Shell/files 86.11%, and Tasks 87.5% remain below the 90% family ship floor.
+- Compact-v5 hints improved Email, Search/HF quant, and Shell on development;
+  a Calendar hint regressed and was rejected rather than shipped.
+
+- Ten-family focused baseline: 19/20 functional and 20/20 routing/execution.
+- Typo and cross-family read flows: 26/26 passed.
+- Real use exposed untested write correction and search-to-fetch follow-ups.
+- Email production access, browser interaction, search quality, and broader
+  multi-turn mutations are not yet proven.
+- Nine enabled chat-capable regular API models pass the ten-family read-only
+  legacy-RAG baseline (90/90 combined). Their stricter typo/follow-up profile is
+  178/180 turns: eight models are 20/20 and Luna is 18/20 due only to its
+  misspelled Shell request. One pinned image-generation model is explicitly
+  unsupported and two visible local models are currently offline.
+- Native VL object/spatial recognition and reload follow-up pass. Exact OCR
+  fails equally on the fine-tune and untouched 9B base and remains unresolved.
+  PNG, JPEG, and WebP transport all pass.
+- Reversible create/correct/API-verify/cleanup flows pass 6/6 across every
+  stateful family.
+- Search Web-toggle combinations pass 8/8 and the focused quality suite passes
+  3/3. Production-path email account/inbox/referential reads pass 3/3.
+- The latest regular-model regression is 90/90 across the nine enabled
+  chat-capable API models, with zero failed model turns; two local endpoints
+  remain offline and the image-only model is unsupported.
+- The Epictetus OMLX endpoint was recovered after an unsupported
+  `qwen3_5_mtp` model load wedged the server. Its supported Qwen 27B 4-bit
+  model passes the ten-family real-7011 legacy-RAG smoke 10/10; the unsupported
+  MTP artifact is recorded as a runtime limitation rather than a timeout.
+- Fresh compact-v5 UI regressions pass stateful 6/6, Email 3/3, Search 3/3,
+  private-browser 3/3, and VL workflow 3/3.
+- The exact-model, family-scoped compact runtime now passes 20/20 direct and
+  same-family turns across all ten families on the real 7011 Agent UI. A
+  separate 36/36 robustness run passes misspellings, bounded repeats, browser
+  and news continuation, ambiguous follow-ups, family switchbacks, and a
+  greeting before a tool request.
+- The mobile active-email editor path passes 1/1: `Write reply this email`
+  offers and executes only `update_document`, mutates the open draft, and
+  preserves its reply headers and quoted thread.
+- The active-editor classifier now also covers short mobile wording without a
+  pronoun (`Write reply` / `Draft a reply`) while explicit note, code, file, and
+  new-object requests retain their own families. Whole-draft requests are bound
+  to the sole offered `update_document` writer until one successful write, then
+  tools are removed for the confirmation round. The deployed real-route email
+  regression passes 3/3—including the exact unspecified `Write reply to this
+  email` form—with one write, verified mutation, and preserved reply headers.
+  Clean-v3 now also emits the established `doc_update` event and flattened
+  document metadata on `tool_output`, so a successful database write updates
+  the already-open editor instead of leaving stale UI beside a success message.
+- The client now reuses the existing assistant bubble for `agent_step` round 1
+  instead of replacing it before the first token. A real-7011 sampled
+  greeting-to-Notes conversation passes 2/2 with stable first-round DOM
+  identity; round 2+ remains the only continuation-bubble path.
+- Clean-runtime metrics now expose provider-counted initial injected tokens,
+  all-round input/output, TTFT, tok/s, schema count, agent rounds, and tool-call
+  count. A real 7011 browser run passes 2/2 and visibly renders compact footers
+  plus the full details popup; the sampled Notes turns streamed progressively.
+- The deployed startup bottleneck was an unindexed quadratic transcript-FTS
+  reconciliation. Live-database import fell from about 36 seconds to 0.54
+  seconds; 7011 now answers in about 3 seconds after a controlled restart.
+- A controlled identical-compact comparison already proves the fine-tune's
+  accuracy benefit: 94.48% (325/344) versus the untouched base's 77.91%
+  (268/344). Raw serving speed is effectively tied, so product speed comes
+  from the compact contract and fewer failed/redundant rounds.
+- A fully merged 10,000-row category-repair candidate reached 97.32% scorable
+  development but only 92.26% scorable sealed blind. Calendar (87.5%), Tasks
+  (87.5%), and Shell/files (86.11%) remained below the family floor, so it was
+  rejected and not deployed. Compact-v4/full development A/Bs did not improve
+  Calendar or Tasks over compact-v5; full-schema Shell also fell from 97.22%
+  to 94.44%. This rules out compactness as the primary cause of the remaining
+  blind gaps and supports keeping the compact contract.
+
+## Non-negotiable architecture rules
+
+1. Runtime selection follows exact model identity. The trained Odysseus model
+   uses the clean compact runtime across endpoint aliases; all other models use
+   legacy RAG. Add a regression test for both sides.
+2. Resolve permissions, toggles, and available backends once per turn. Produce
+   one immutable contract satisfying `required ⊆ offered ⊆ executable`.
+3. Never offer a tool that the preview policy will categorically reject. Add a
+   contract self-check covering every offered action/effect combination.
+4. Follow-ups consume typed prior evidence: native call, result, success state,
+   family, and object identifiers. Do not infer continuity from keyword RAG.
+5. Contextual write authority may revise only a recently proven object in the
+   same family. It may not authorize a new object, another family, a destructive
+   action, or an external side effect.
+6. The model chooses tools and valid arguments. The harness validates and
+   executes; it does not silently substitute another family, rewrite arguments,
+   fabricate success, or replace a failed tool with prose claiming completion.
+7. One owner renders each turn: streamed prose or canonical structured output.
+   Never both, and never expose hidden prompts or raw untrusted wrappers.
+8. No exact-prompt production patches. A fix must name the failed layer, add a
+   generic failing invariant test, and cover neighboring cases.
+
+## Failure layers
+
+Every failure is assigned to exactly one primary layer before code changes:
+
+1. **Route:** wrong model runtime or endpoint identity.
+2. **Contract:** required tool absent, forbidden tool present, or toggle drift.
+3. **Model:** wrong/no tool or semantically wrong required arguments despite a
+   correct contract.
+4. **Policy:** valid proposed operation incorrectly allowed or denied.
+5. **Execution:** canonical arguments, backend dispatch, timeout, or result
+   envelope is wrong.
+6. **Evidence:** result is empty, irrelevant, truncated badly, or insufficient.
+7. **Answer:** model misstates or ignores valid tool evidence.
+8. **Rendering:** duplicate, dump-at-end, missing structured output, or stopped
+   stream.
+9. **Performance:** startup, TTFT, tool latency, or oversized context.
+
+Reports store aggregate category, relevant contract/tool metadata, timings, and
+sanitized outputs. Do not copy private hidden benchmark prompts or create a log
+dump that nobody can audit.
+
+## Test matrix
+
+Use the real authenticated 7011 Agent UI and the normal `preheret` picker alias.
+Use `sft_alex_creator` for reversible writes. Never mutate the personal account
+from an automated test.
+
+### A. Every one of the ten families
+
+For calendar, notes, email, tasks, documents, memory, skills, Cookbook/admin,
+search/browser, and shell/files, test:
+
+- direct request;
+- natural misspelling;
+- ambiguous same-family follow-up;
+- switch to another family and back;
+- no-tool greeting before the tool request;
+- requested count/field limit;
+- backend failure rendered truthfully;
+- reload the permalink before a follow-up.
+
+### B. Stateful mutation families
+
+For notes, calendar, tasks, documents, memory, and skills:
+
+- create → verify by API → referential correction → verify;
+- create → list/read → correction → verify;
+- typo correction such as name/date/title without repeating the family noun;
+- correction after one unrelated conversational turn;
+- destructive request is denied atomically;
+- failed write never produces a success claim;
+- cleanup deletes only the UUID-owned test artifact and verifies absence.
+
+### C. Search and browser conversations
+
+- search → summarize existing results without a new call;
+- search → inspect one result with `web_fetch`;
+- poor results → refine query once;
+- insufficient evidence → say so without fabrication;
+- Web toggle combinations `00`, `01`, `10`, and `11` across two turns;
+- private browser open/snapshot/click only after its permission boundary is
+  deliberately enabled and specified; do not smuggle it in via web search.
+
+Grade source relevance, freshness, authority, and whether claims are supported,
+not merely whether `web_search` was called.
+
+### D. Email and shell
+
+- Separate fixture accuracy from production connectivity. A fixture pass cannot
+  promote production email health.
+- Test account listing, inbox listing, reading, and referential follow-up against
+  the configured production-like backend before enabling email actions.
+- Shell remains toggle-gated. Test off/on transitions, canonical raw command
+  dispatch, read-only output, and denial of network/destructive commands.
+
+### E. Rendering and performance
+
+- Assert first visible streamed token, monotonic DOM growth, one final answer,
+  persistence/reload equality, stop behavior, and structured list rendering.
+- Record request preparation, TTFT, tool duration, post-tool TTFT, total time,
+  input/output tokens, and tool-result bytes.
+- Diagnose the 30–40 second 7011 restart separately from inference latency.
+- Bound large calendar/search results before replaying them into later rounds,
+  while preserving IDs and fields needed for follow-ups.
+
+### F. Image/VL recognition
+
+- Attach real PNG, JPEG, and WebP images through the 7011 UI and verify the
+  trained model receives native multimodal message content on its clean route.
+- Test object recognition, visible text/OCR, spatial relationships, charts, and
+  screenshots. Score required facts instead of stylistic wording.
+- Test image → ambiguous follow-up, image → tool request, and tool result → image
+  comparison without requiring the user to attach the same image again.
+- Verify image references survive persistence and permalink reload without raw
+  base64, local paths, or hidden wrappers appearing in chat output.
+- Separate direct model vision from `inspect_media`, browser screenshots, and
+  image generation. The harness must not silently substitute one for another.
+- Compare the fine-tune with its base VL model on the same images to detect
+  whether tool training regressed visual understanding.
+
+### G. Regular-model legacy RAG and tool coverage
+
+- Inventory every enabled non-Odysseus endpoint/model visible in 7011, including
+  its provider, schema mode, native-tool support, context limit, and configured
+  permissions. Do not assume every provider supports the same wire format.
+- Assert that no non-Odysseus model enters the clean-v3 runtime. These models
+  retain the regular RAG/tool loop and are repaired only in that owning path.
+- For each model, test every tool family the effective user policy offers:
+  direct request, misspelling, ambiguous follow-up, family switch, backend
+  failure, and Web/Bash toggle transitions. Record unsupported families as an
+  explicit capability limitation, not a silent pass.
+- Test full schemas versus compact schemas only where both are valid for that
+  model. Store the selected schema mode in every report.
+- Verify provider-native tool calls, textual fallback parsing where required,
+  canonical argument conversion, execution, evidence replay, and rendering.
+- Group fixes by shared legacy-runtime or provider-adapter defect. Do not add
+  model-name prompt exceptions when a transport, schema, or RAG ranking issue is
+  responsible.
+- Maintain a per-model compatibility matrix so adding or changing an endpoint
+  cannot silently regress previously working tools.
+
+## Fix protocol
+
+For each failure:
+
+1. Preserve the raw report and reproduce once on a fresh test session.
+2. Identify the primary failure layer from the taxonomy above.
+3. Add the smallest generic red test at that layer.
+4. Fix the owning module or invariant—not the literal prompt.
+5. Run the focused unit tests, the original scenario, two adjacent scenarios,
+   and the affected family suite.
+6. After a batch of category fixes, rerun the ten-family matrix and legacy-RAG
+   isolation test. Do not rerun training unless the contract and harness are
+   proven correct and failures remain model-owned.
+
+If three failures share a layer, pause case-by-case patching and refactor that
+layer before continuing.
+
+## Execution phases
+
+### Phase 1 — Make the runtime auditable
+
+- Add a sanitized per-turn decision record: model runtime, contract, proposed
+  calls, policy decisions with reason codes, executions, render owner, timings.
+- Add startup/runtime provenance to the UI so a linked chat proves which harness
+  handled it.
+- Add the offered-versus-policy compatibility self-test.
+- Correct stale preview documentation.
+
+### Phase 2 — Build the conversation suite
+
+- Extend the current Playwright verifier with reusable multi-turn scenarios and
+  reversible artifact fixtures.
+- Implement the matrix above, prioritizing search continuations and all
+  stateful corrections because real usage already exposed those gaps.
+- Run independent family groups in parallel, but serialize writes that share a
+  backend or fixture account.
+- Add a small versioned VL fixture set with locally generated, non-private
+  images and deterministic answer keys.
+
+### Phase 3 — Repair by architecture category
+
+- Consolidate model-specific runtime selection in one function.
+- Represent prior successful objects explicitly for referential follow-ups.
+- Align tool capability classification, contract offering, and policy decisions.
+- Standardize tool results into bounded envelopes with source/object IDs.
+- Keep search refinement and evidence sufficiency generic.
+
+### Phase 4 — Accuracy and speed comparison
+
+- Compare the clean fine-tune with the base model using identical compact tools,
+  prompts, toggles, backend state, and semantic scoring.
+- Report functional accuracy, argument accuracy, unsupported success claims,
+  TTFT, total latency, and tokens. Do not compare one model on full schemas and
+  another on compact schemas.
+- Only consider more SFT/RL for failures classified as model-owned after the
+  harness audit.
+
+### Phase 4B — Regular-model repair and verification
+
+- Snapshot the enabled non-Odysseus model inventory.
+- Run the legacy-RAG compatibility matrix in bounded parallel groups, respecting
+  endpoint rate limits and shared backend write serialization.
+- Fix shared harness/provider defects first, then rerun all affected models.
+- Publish separate per-model scores and limitations; do not blend them into the
+  Odysseus fine-tune score.
+
+### Phase 5 — Ship gate
+
+Ship only when:
+
+- every family is at least 90% on sealed functional holdout;
+- overall functional accuracy is at least 95%;
+- realistic follow-up suite is at least 95%, with no repeated failure category;
+- image/VL fixture accuracy does not regress materially from the base model and
+  all attachment/follow-up/persistence flows pass;
+- routing/execution and safety invariants are 100%;
+- all reversible writes are API-verified and cleaned up;
+- search quality and production email are reported separately and honestly;
+- non-Odysseus models demonstrably retain legacy RAG;
+- every enabled regular model has a complete tested-tool compatibility record,
+  and every tool advertised as supported passes its functional checks;
+- no hidden prompt leakage, duplicate rendering, or false success remains;
+- pre-heretic passing weights and merged adapter backups remain recoverable.
+
+## Immediate next batch
+
+1. Expand VL fixtures to charts, screenshots, and image-to-tool turns;
+   investigate the shared base-model OCR limitation without hiding it behind a
+   silent external fallback.
+2. Add deliberately permissioned private-browser open/snapshot/click checks;
+   keep browser interaction unavailable when its boundary is not enabled.
+3. Bring the two configured local regular models online and run their matrix.
+4. Compare fine-tune versus untouched base with identical compact contracts,
+   backend state, prompts, and timing instrumentation.
+5. Run the sealed all-action holdout and prioritize failures by shared
+   layer rather than by prompt.

plans/photo-editor-professional-roadmap.md

@@ -0,0 +1,445 @@
+# Plan: Odysseus Professional Photo Editor
+
+> Source PRD: Conversation goal, "a Photoshop/Photopea clone with Odysseus style"
+
+## Product boundary
+
+Odysseus should provide the editing loop people expect from a professional
+layer-based photo editor without copying Photoshop's visual design or trying to
+match every specialist feature. The target is a dependable browser editor for
+real photo work: direct manipulation, non-destructive layers, precise masking,
+retouching, typography, export, recovery, and optional AI assistance.
+
+The existing quiet Odysseus interface remains the visual language. Dense tools
+are acceptable, but controls should stay restrained, compact, predictable, and
+usable on both desktop and touch devices.
+
+## Existing foundation
+
+The current editor already provides meaningful parts of this product:
+
+- Raster and editable text layers
+- Multi-layer selection, nested groups, clipping, visibility, opacity, and locks
+- Layer, group, and selection masks
+- Marquee, lasso, wand, SAM, Quick Mask, and saved selections
+- Brush, eraser, clone, crop, transform, and text tools
+- Blend modes, adjustment stacks, blur, and several image corrections
+- Rulers, guides, grid, snapping, zooming, and panning
+- Undo/redo history with a memory budget
+- Versioned layered-project serialization, autosave drafts, recovery, and export
+- Optional endpoint-backed inpaint and image-processing tools
+- Desktop and mobile editor layouts with Playwright release-gate coverage
+
+## Architectural decisions
+
+Durable decisions that apply across every phase:
+
+- **Editor ownership**: The editor remains an Odysseus feature. Do not embed a
+  third-party editor or imitate another product's chrome.
+- **Document format**: Continue the versioned Odysseus editor document. Every
+  new persistent capability requires a migration, validation, round-trip test,
+  and corrupt-input recovery behavior.
+- **Layer model**: Grow the document into explicit layer kinds rather than
+  hiding more behavior in raster canvases. The intended kinds are raster, text,
+  shape, adjustment, and placed/smart content.
+- **Non-destructive default**: Preserve source pixels and editable parameters
+  whenever practical. Destructive actions remain available as explicit Apply,
+  Rasterize, or Merge commands.
+- **Interaction engine**: Transform, crop, selections, text frames, masks, and
+  shapes share one pointer-session model for hit testing, pointer capture,
+  modifiers, snapping, cancellation, and undo transactions.
+- **Rendering**: Keep Canvas 2D as the compatibility renderer initially. Move
+  expensive compositing and pixel operations behind renderer/worker boundaries
+  before considering WebGL or WebGPU acceleration.
+- **History**: One continuous gesture creates one undo entry. Preview frames are
+  never separate history entries, and Cancel restores the exact starting state.
+- **Persistence routes**: Continue using `/api/editor-drafts` for layered draft
+  persistence and `/api/gallery` for media-library save/replace operations.
+- **AI boundary**: AI features consume capability-based image endpoints. Core
+  editing never requires a particular model, repository, or provider.
+- **Responsive behavior**: Desktop favors precision; touch targets gain larger
+  invisible hit areas without visually enlarging the whole interface.
+- **Testing**: Every phase adds deterministic geometry/unit tests and at least
+  one complete Playwright workflow covering persistence and undo where relevant.
+- **Incremental architecture**: New behavior leaves the main editor orchestrator
+  through small domain modules. Avoid broad refactors that do not deliver a
+  visible editing improvement in the same phase.
+
+---
+
+## Phase 1: Accurate Transform Frame
+
+**User stories**: I can clearly see and grab the transform frame at any zoom. I
+can resize from corners or sides without grabbing invisible or incorrect areas.
+
+### What to build
+
+Replace the four-corner-only frame with a shared frame geometry model. Render
+four corners, four edge handles, a rotation control, and an optional center
+pivot from the same geometry used for hit testing. Keep handles visually compact
+while providing touch-sized invisible targets. Make the frame stay aligned
+during zoom, pan, viewport resize, and when handles extend outside the image.
+
+### Acceptance criteria
+
+- [x] Eight resize handles, rotation control, and center pivot derive from one geometry result.
+- [x] Drawn handles and hit targets cannot disagree.
+- [x] Handles remain a stable visual size from minimum to maximum zoom.
+- [x] Touch hit targets are at least 40 CSS pixels without oversized visuals.
+- [x] Outside-canvas handles remain interactive and visible when space permits.
+- [x] Hover and active cursors match each handle's current screen direction.
+- [x] Desktop and mobile Playwright tests grab every handle successfully.
+
+---
+
+## Phase 2: Correct Rotated Resize
+
+**User stories**: I can resize a rotated layer naturally. The opposite side or
+corner stays fixed, and the frame follows my pointer rather than drifting.
+
+### What to build
+
+Calculate drag movement in the frame's rotated local coordinate system. Anchor
+the opposite handle in document space and derive the new center from that
+anchor. Support crossing an axis as a deliberate flip instead of clamping to a
+one-pixel box. Apply the same geometry to one layer, multiple layers, and a
+selection transform.
+
+### Acceptance criteria
+
+- [x] Rotated corner and edge drags follow the pointer on the frame's local axes.
+- [x] The opposite anchor remains fixed within a sub-pixel tolerance.
+- [x] Crossing width or height zero produces a predictable horizontal or vertical flip.
+- [x] Shift locks the starting aspect ratio.
+- [x] Alt/Option scales around the transform center.
+- [x] Combined Shift+Alt/Option behavior is deterministic.
+- [x] Rotation snaps to 15-degree increments with Shift and remains smooth otherwise.
+- [x] Geometry tests cover 0, 45, 90, 135, and arbitrary-degree rotations.
+
+---
+
+## Phase 3: Transform Interaction Polish
+
+**User stories**: Transform behaves like a professional tool on mouse, pen, and
+touch. I can see exact values, snap precisely, and never lose a drag at the edge.
+
+### What to build
+
+Use a unified pointer session with pointer capture, live modifiers, and a small
+contextual transform readout. Add accurate rotated-frame interior hit testing,
+keyboard nudging, frame snapping, and clear Apply/Cancel behavior. Keep the
+existing compact Odysseus styling and make the numeric popup a precision surface
+rather than a competing transform implementation.
+
+### Acceptance criteria
+
+- [x] Pointer capture keeps a drag alive outside the canvas and browser viewport.
+- [x] Clicking inside a rotated frame moves it; clicking its empty bounding-box corner does not.
+- [x] Live X, Y, W, H, and angle values stay synchronized with direct manipulation.
+- [x] Arrow keys nudge, Shift+Arrow performs a larger nudge, Enter applies, and Escape cancels.
+- [x] Layer edges, document center/edges, guides, and grid participate in transform snapping.
+- [x] Snap guides clearly identify the active alignment without obscuring the photo.
+- [x] A complete gesture creates exactly one undo step.
+- [x] Touch gestures do not conflict with viewport pinch/pan behavior.
+
+---
+
+## Phase 4: Transform Content Correctness
+
+**User stories**: Transforming layers never unexpectedly damages masks, text,
+group layout, clipping, or image quality. Saving and reopening preserves it.
+
+### What to build
+
+Route raster layers, text layers, linked and unlinked masks, selections, clipped
+layers, and grouped multi-selection through the same transform contract. Keep
+immutable source data during previews and validate the final result through
+undo, cancel, autosave, project download, and reopen.
+
+### Acceptance criteria
+
+- [x] Raster previews are always derived from the session source, never a prior preview.
+- [x] Editable text remains editable after scaling, rotation, and flipping.
+- [x] Linked masks follow the layer while unlinked masks remain in document space.
+- [x] Multi-layer transforms preserve relative centers, order, clipping, and group membership.
+- [x] Transforming a selection changes only the selection mask unless content transform is explicitly chosen.
+- [x] Apply, Cancel, Undo, Redo, autosave reopen, and project-file reopen produce matching pixels and metadata.
+- [x] Large transforms cannot allocate beyond the editor's documented surface budget.
+
+---
+
+## Phase 5: Shared Direct-Manipulation Sessions
+
+**User stories**: Crop, selections, masks, text boxes, and shapes feel consistent
+with Transform instead of each behaving like a separate mini application.
+
+### What to build
+
+Generalize the proven transform pointer session into a reusable interaction
+contract. Migrate crop and selection movement first as a visible tracer bullet,
+including modifiers, snapping, pointer capture, cancel, and one-step history.
+
+### Acceptance criteria
+
+- [x] Transform, crop, and selection movement use the same gesture lifecycle.
+- [x] Tool switching safely commits, cancels, or prompts according to one policy.
+- [x] No stale pointer session can modify a newly selected tool or document.
+- [x] Mouse, pen, and touch event behavior is covered by shared tests.
+- [x] Adding a future frame-based tool does not require another global event stack.
+
+---
+
+## Phase 6: Non-Destructive Placed Layers
+
+**User stories**: I can import an image, resize it repeatedly without cumulative
+quality loss, replace its source, and choose when to rasterize it.
+
+### What to build
+
+Introduce a placed/smart layer kind containing source pixels and persistent
+transform metadata. Import-as-layer uses this kind by default. Rendering applies
+the transform at composite time, while Rasterize produces a normal raster layer.
+
+### Acceptance criteria
+
+- [x] Repeated transforms render from the original source rather than resampling the last result.
+- [x] A placed layer can be replaced while preserving its transform and masks.
+- [x] Rasterize produces a visually matching editable raster layer.
+- [x] Masks, clipping, groups, blend modes, and opacity work with placed layers.
+- [x] Version migration and recovery handle missing or corrupt placed sources.
+- [x] Existing raster projects open without changed output.
+
+---
+
+## Phase 7: Professional Selections And Masks
+
+**User stories**: I can build, inspect, refine, save, transform, and reuse precise
+selections without manually repainting every edge.
+
+### What to build
+
+Unify marquee, lasso, wand, SAM, Quick Mask, and saved selections around one
+selection-mask model. Add explicit replace/add/subtract/intersect modes, feather,
+expand, contract, smooth, border, and a focused refine-edge workflow.
+
+### Acceptance criteria
+
+- [x] Every selection tool supports replace, add, subtract, and intersect modes.
+- [x] Feather, expand, contract, smooth, and border preview before applying.
+- [x] Quick Mask edits the same canonical selection shown by marching ants.
+- [x] Selection-to-layer-mask and layer-mask-to-selection round-trip accurately.
+- [x] Saved selections retain names and pixels across reopen.
+- [x] Edge refinement works without requiring an AI dependency.
+
+---
+
+## Phase 8: Paint And Retouch Workflow
+
+**User stories**: I can paint and retouch photographs with predictable strokes,
+reusable presets, and the controls expected for a mouse, pen, or touch device.
+
+### What to build
+
+Promote brush behavior into a reusable brush engine. Add spacing, smoothing,
+pressure mapping, blend mode, sampled color, presets, and stroke preview. Build
+healing, dodge, and burn as complete retouching paths using that engine.
+
+### Acceptance criteria
+
+- [x] Brush, eraser, clone, masks, and inpaint share spacing and smoothing behavior.
+- [x] Pressure can independently affect size, opacity, or flow when supported.
+- [x] Eyedropper samples composite or active-layer color.
+- [x] Brush presets can be created, named, selected, and deleted.
+- [x] Healing, dodge, and burn create one undo entry per stroke.
+- [x] Long strokes remain smooth without blocking the main interface.
+
+---
+
+## Phase 9: Editable Text And Shapes
+
+**User stories**: I can design labels, cards, and overlays with text and vector
+shapes that remain editable after saving and reopening.
+
+### What to build
+
+Add on-canvas text-frame editing, selection, caret behavior, typography, and
+alignment. Introduce shape layers for rectangle, ellipse, line, and path-backed
+polygons with editable fill, stroke, corners, and transform metadata.
+
+### Acceptance criteria
+
+- [x] Text is edited directly on canvas without immediately rasterizing.
+- [x] Font, size, weight, line height, letter spacing, alignment, and color persist.
+- [x] Rectangle, ellipse, line, and polygon shapes remain editable.
+- [x] Shape fill, stroke, width, and corner radius can be changed after creation.
+- [x] Text and shape layers support masks, clipping, groups, blend modes, and transform.
+- [x] Missing fonts fall back predictably without corrupting the project.
+
+---
+
+## Phase 10: Adjustment Layers And Color
+
+**User stories**: I can correct a photograph non-destructively and return later
+to modify the correction without reconstructing the edit.
+
+### What to build
+
+Promote adjustments into first-class layers with masks and clipping. Deliver
+Levels and Curves first, then exposure, white balance, hue/saturation, color
+balance, selective color, gradients, and channel-aware controls.
+
+### Acceptance criteria
+
+- [ ] Adjustment layers affect content below them and can be clipped or grouped.
+- [ ] Every adjustment has live preview, reset, visibility, opacity, mask, Apply, and Cancel behavior.
+- [ ] Levels includes histogram, input range, gamma, and output range.
+- [ ] Curves supports RGB and channel curves with editable points.
+- [ ] Color results match flattened export and project reopen.
+- [ ] Large previews are throttled or worker-backed and remain cancellable.
+
+---
+
+## Phase 11: Layer Effects And Filters
+
+**User stories**: I can add common visual effects without permanently altering
+the layer and can reorder or disable those effects later.
+
+### What to build
+
+Create an ordered non-destructive filter/effect stack. Begin with Gaussian blur,
+sharpen, shadow, stroke, and color overlay; then add filter masks and reusable
+effect presets.
+
+### Acceptance criteria
+
+- [ ] Effects can be added, reordered, toggled, edited, masked, and removed.
+- [ ] Drop shadow, stroke, color overlay, blur, and sharpen survive project reopen.
+- [ ] Effects render correctly inside groups and clipping stacks.
+- [ ] Apply/rasterize produces a pixel-equivalent raster result.
+- [ ] Expensive filters expose progress and cancellation.
+
+---
+
+## Phase 12: Odysseus Professional Workspace
+
+**User stories**: I can work quickly without fighting floating windows or losing
+the active tool, layer, selection, or document context.
+
+### What to build
+
+Refine the existing shell into a consistent professional workspace: contextual
+tool options, properties inspector, panel persistence, command search, status
+information, multi-document switching, and compact touch sheets. Preserve the
+current Odysseus palette, typography, restrained borders, and frosted surfaces.
+
+### Acceptance criteria
+
+- [ ] Tool options appear in one predictable location and never duplicate popup state.
+- [ ] Panels remember size, collapsed state, and position per device class.
+- [ ] The properties inspector follows the active layer, mask, selection, or tool.
+- [ ] Command search exposes actions and shortcuts without adding toolbar clutter.
+- [ ] Switching documents preserves independent history, zoom, pan, and selection.
+- [ ] Mobile prioritizes canvas area while keeping all commands reachable.
+
+---
+
+## Phase 13: File Interchange And Export
+
+**User stories**: I can bring common assets into Odysseus and export predictable
+results without losing transparency, dimensions, or color intent.
+
+### What to build
+
+Strengthen image import/export first, then add layered interchange where a
+maintained parser makes it safe. Keep Odysseus project files as the lossless
+source of truth and clearly report what an external format cannot preserve.
+
+### Acceptance criteria
+
+- [ ] PNG, JPEG, WebP, and supported modern image imports honor orientation and transparency.
+- [ ] Export exposes format, dimensions, quality, metadata, and transparency choices.
+- [ ] Copy/paste and drag/drop preserve alpha and use placed layers when appropriate.
+- [ ] Layered imports report unsupported features instead of silently flattening them.
+- [ ] Exported pixels are covered by deterministic visual comparisons.
+
+---
+
+## Phase 14: Large-Document Performance And Recovery
+
+**User stories**: Large photos and layered projects remain responsive, autosave
+reliably, and recover after a crash or interrupted network connection.
+
+### What to build
+
+Move serialization, thumbnails, filters, and suitable pixel operations into
+workers. Add dirty-region rendering, reusable surfaces, measurable memory
+budgets, operation cancellation, autosave generations, and recovery diagnostics.
+
+### Acceptance criteria
+
+- [ ] Normal interactions remain responsive on the agreed 4K multi-layer benchmark.
+- [ ] Compositing avoids rebuilding unaffected layers and thumbnails.
+- [ ] History and document surfaces stay within explicit memory limits.
+- [ ] Closing or switching documents cancels stale work safely.
+- [ ] Autosave never lets an older request overwrite newer state.
+- [ ] Recovery can identify the last complete generation and explain skipped data.
+
+---
+
+## Phase 15: Odysseus-Native Assisted Editing
+
+**User stories**: I can use an available local or remote image capability as an
+editing assistant while retaining masks, layers, undo, privacy choices, and
+normal manual controls.
+
+### What to build
+
+Standardize image capability discovery and requests for generation, editing,
+inpainting, segmentation, restoration, and upscaling. Results enter the document
+as named layers with provenance and reusable masks. Add orchestration only after
+the manual operation it assists is dependable.
+
+### Acceptance criteria
+
+- [ ] The UI describes required capabilities rather than model or provider names.
+- [ ] Memory and unrelated chat context are not sent to image endpoints.
+- [ ] Requests show progress, support cancellation, and cannot update a closed document.
+- [ ] Generated results arrive as reversible layers with prompt/settings metadata.
+- [ ] A failed endpoint leaves the source document unchanged and offers a useful retry path.
+- [ ] Manual selection and masking remain available when assisted tools are absent.
+
+---
+
+## Phase 16: Professional Release Gate
+
+**User stories**: I can trust the editor for real work and understand what is
+unsupported before committing an edit.
+
+### What to build
+
+Create a release gate around complete user journeys rather than isolated button
+tests. Cover accessibility, keyboard-only operation, touch, browser differences,
+pixel correctness, persistence, failure recovery, and large-document behavior.
+
+### Acceptance criteria
+
+- [ ] Core workflows pass on current Chromium and Firefox desktop builds.
+- [ ] Mobile workflows pass at representative phone and tablet viewports.
+- [ ] Keyboard-only users can reach every command and escape every modal state.
+- [ ] Transform, masks, text, adjustments, export, and reopen have pixel/metadata regression tests.
+- [ ] No supported action silently flattens or discards editable document data.
+- [ ] The ALPHA badge can be removed based on explicit reliability metrics.
+
+---
+
+## Recommended delivery order
+
+The first four phases are one focused Transform 2.0 program and should ship in
+order. Phases 5 and 6 establish the interaction and document foundations needed
+for the remaining professional tools. After that, phases 7 through 13 can be
+prioritized by user value, while performance and release-gate work continue as
+part of every phase rather than being deferred entirely to the end.
+
+The recommended first milestone is complete when Phases 1 through 4 are live:
+transforming one layer, multiple layers, text, masks, and selections feels
+precise on desktop and mobile and remains correct through undo and reopen.

plans/photo-editor-remaining-scope.md

@@ -0,0 +1,159 @@
+# Photo Editor Remaining Scope
+
+Date: 2026-08-29
+
+## Current verdict
+
+Odysseus is now a credible layered everyday editor, not an editor mockup. The
+first nine roadmap phases are implemented: professional transform geometry,
+shared direct-manipulation sessions, retained placed content, unified
+selections and masks, a reusable brush/retouch engine, and retained text and
+shape layers.
+
+Phase 10 is functionally advanced but not closed. First-class adjustment layers
+now support Levels, Curves, Exposure, White Balance, Brightness/Contrast,
+Hue/Saturation/Lightness, Color Balance, Selective Color, and Gradient Map.
+They participate in clipping, groups, masks, visibility, opacity, history, the
+v14 document format, and flattening. Retained effects have since been added as
+a separate ordered stack with Gaussian Blur, Color Overlay, Drop Shadow, and
+Stroke, including editable colors, visibility, opacity, reorder, rasterize,
+history, persistence, and migration.
+
+Practical readiness estimate:
+
+- Everyday layered photo editing: **about 88%**
+- Dependable professional v1 described by the roadmap: **about 62%**
+- Broad Photoshop/Photopea feature parity: **about 50%**
+
+The remaining gap is dominated by large-document rendering outside the live
+composite path, workspace consolidation, interchange/color policy, and release
+proof rather than basic canvas tools.
+
+## Verification snapshot
+
+- The focused editor unit suite currently passes **31 tests** in Docker.
+- The full photo-editor browser suite currently has **41 passing workflows**;
+  the nested-group selection workflow initially exposed a row-hit regression,
+  which now passes on isolated rerun after the slider-selection fix. The new
+  group-effects workflow also passes.
+- The new adjustment tests exercise deterministic pixel math, nested parameter
+  normalization, retained metadata, undo/redo, clipping, masks, and draft
+  reopen.
+- The latest editor changes have not yet been rebuilt into the live `7011`
+  container.
+
+## Close Phase 10
+
+This is the immediate release slice.
+
+1. Finish the bounded preview path for large documents. Downsampled previews
+   now keep control movement responsive and full resolution is restored for
+   commit/export. Live worker composites now use generation checks, latest-only
+   coalescing, and close/reopen invalidation; extend the same guarantees to
+   remaining preview paths.
+2. Add flattened-export versus reopened-project pixel comparisons for every
+   adjustment family, including groups, clipping, masks, blend mode, and
+   partial opacity.
+3. Validate the color algorithms visually. White Balance and Selective Color
+   are currently deterministic approximations, not color-managed photographic
+   transforms.
+4. Test every adjustment popup on phone and desktop viewports, including tall
+   popups, color inputs, drag, Reset, Apply, Cancel, and Escape.
+5. Decide the migration path for the older per-raster `adjLayers` stack. It can
+   remain readable for compatibility, but new UI should converge on first-class
+   adjustment layers instead of maintaining two competing concepts.
+6. Bump static cache versions, rebuild the live container, and run a short
+   visual smoke test on `7011`.
+
+## Phase 11: Retained effects and filters
+
+The retained-effects slice is implemented for raster/placed/text/shape-compatible
+layer output: Gaussian Blur, Sharpen, Color Overlay, Drop Shadow, and Stroke
+have editable colors/parameters, visibility, opacity, reorder, rasterize,
+history, migration, and reopen support. Effect-specific masks, presets, and
+group-level effects are also implemented and covered by focused browser tests.
+Remaining work is:
+
+1. Extend worker coverage to serialization and remaining preview paths.
+   Thumbnail encoding, retained-effect rasterization, and live composite
+   rendering now use a worker where OffscreenCanvas is available, with
+   synchronous compatibility fallbacks. Generation invalidation, latest-only
+   coalescing, and CPU loop cancellation protect live rendering.
+2. Add explicit group-effect blend/ordering tests for nested groups and
+   non-default blend modes, plus visual comparisons for effect stacks.
+
+Introduce the renderer/worker cancellation boundary here rather than adding
+more synchronous full-canvas filters that Phase 14 must immediately replace.
+
+## Phase 12: Professional workspace
+
+Consolidate fragmented popups into one contextual properties surface. Persist
+panel layout by device class, add command search, expose stable document status,
+and support multiple open documents with independent history, zoom, pan, and
+selection. Mobile should use canvas-first sheets rather than compressed desktop
+panels.
+
+## Phase 13: Interchange and export
+
+Harden orientation, transparency, metadata, and color behavior for PNG, JPEG,
+and WebP first. Add copy/paste and drag/drop through placed layers. Treat
+layered formats as explicit compatibility projects: unsupported PSD/TIFF/HEIC
+features must be reported, never silently discarded. Odysseus project files
+remain the lossless source of truth.
+
+## Phase 14: Performance and recovery
+
+Move remaining preview/pixel paths into workers. Thumbnail encoding,
+autosave serialization, adjustment rendering, and retained-effect rendering
+now have worker-backed paths with compatibility fallbacks. Add
+dirty-region compositing, reusable render surfaces, cancellation tokens,
+operation telemetry, a documented surface/history budget, autosave generations,
+and a checked-in 4K multi-layer benchmark.
+
+This phase is the main architectural risk. Canvas 2D remains a valid
+compatibility renderer, but full-document synchronous passes will not scale to
+professional documents.
+
+## Phase 15: Assisted editing
+
+Normalize generation, editing, inpainting, segmentation, restoration, and
+upscaling behind capability-based endpoints. Keep model/provider names out of
+editor logic. Requests must exclude chat memory, show progress, cancel safely,
+and return named reversible layers with provenance. Manual tools remain fully
+usable without an endpoint.
+
+Much of the endpoint plumbing already exists; the remaining work is consistent
+capability discovery, lifecycle safety, and editor-native result handling.
+
+## Phase 16: Release gate
+
+Run complete user journeys on Chromium and Firefox desktop plus representative
+phone/tablet viewports. Add keyboard-only and accessibility coverage, mixed
+20-edit persistence/export tests, failure recovery, and large-document stress
+tests. No supported operation may silently flatten or discard retained state.
+
+## Architecture debt to control
+
+- `galleryEditor.js` is still a large orchestrator. Continue extracting domain
+  modules as visible features move, without a broad rewrite.
+- Legacy raster adjustment sublayers and first-class adjustment layers overlap.
+  Converge on the first-class model.
+- Pixel effects still rely heavily on synchronous full-canvas work.
+- `static/style.css` carries substantial editor-specific surface area and needs
+  clearer component boundaries before workspace customization expands.
+- The repository worktree contains many unrelated changes. Editor release and
+  merge decisions require a scoped diff or clean integration branch.
+
+## Recommended execution order
+
+1. Close and deploy Phase 10.
+2. Build Phase 11 through a cancellable render boundary.
+3. Consolidate the workspace in Phase 12.
+4. Define color/metadata policy and complete Phase 13.
+5. Finish worker rendering, stress, and recovery in Phase 14.
+6. Normalize assisted editing in Phase 15.
+7. Run the cross-browser professional release gate in Phase 16.
+
+Do not expand into full PSD fidelity, CMYK production, RAW development, 3D, or
+complete Photoshop parity before this critical path passes. Those are separate
+product decisions, not prerequisites for a strong Odysseus editor.

requirements-optional.txt

@@ -1,4 +1,7 @@
 # Optional dependencies — install only if you use the corresponding feature.
+# Local OCR for screenshots, scans, labels, and coordinate-grounded text extraction.
+rapidocr==3.9.2
+onnxruntime>=1.20,<2
 # The app handles their absence gracefully (clear error message on first use).
 #
 # Note: chromadb-client + fastembed moved to requirements.txt — RAG, semantic
@@ -12,6 +15,16 @@
 # GPU-accelerated transcription — it's auto-detected, CPU is used otherwise.
 faster-whisper
 
+# Local text-to-speech via Kokoro-82M for the "local" TTS provider.
+# Kokoro 0.9.4 declares Python >=3.10,<3.13; Odysseus itself requires 3.11+,
+# so pip installs these extras on 3.11-3.12 and deliberately skips them on
+# Python 3.13+ (including the Python 3.14 container image). Kokoro declares
+# torch; the local provider still
+# requires a CUDA-enabled torch build and GPU at runtime. SoundFile is separate
+# in Kokoro's official install instructions and is not a transitive dependency.
+kokoro==0.9.4; python_version >= "3.11" and python_version < "3.13"
+soundfile; python_version >= "3.11" and python_version < "3.13"
+
 # DuckDuckGo as a search provider option.
 # Install if you want DDG in the search-provider dropdown.
 # Alternatives: SearXNG, Brave, Tavily, Serper, Google PSE.
@@ -34,3 +47,6 @@ PyMuPDF
 # [all]/Azure/audio extras (cloud + heavy). Pinned to a release >30 days old per
 # the dependency-age discussion in issue #485.
 markitdown[docx,pptx,xlsx,xls]==0.1.6
+
+# Photoshop PSD opening / flattened previews / layer inspection.
+psd-tools

requirements.txt

@@ -3,10 +3,16 @@ uvicorn
 python-multipart
 python-dotenv
 httpx
+httpcore>=1.0,<2.0
 pydantic>=2.13.4
 pydantic-settings>=2.14.1
 SQLAlchemy
 pypdf
+pypdfium2
+Pillow
+faster-whisper
+PyPDF2
+pdfplumber
 beautifulsoup4
 charset-normalizer
 numpy
@@ -18,6 +24,7 @@ numpy
 chromadb-client
 fastembed
 youtube-transcript-api
+yt-dlp
 # Markdown rendering for research reports (src/visual_report.py).
 # Imported at module-top so it's a hard core dep, not optional.
 markdown
@@ -37,7 +44,10 @@ python-dateutil
 caldav
 cryptography
 bcrypt
-mcp
+# Built-in servers use the v1 low-level Server decorator API. MCP SDK v2 is a
+# breaking rewrite, so keep fresh installs on the maintained v1 line until the
+# servers are migrated together.
+mcp<2
 pyotp
 qrcode[pil]
 croniter

resources/skills/agent/artifact-completion/SKILL.md

@@ -0,0 +1,38 @@
+---
+name: artifact-completion
+description: Create requested artifacts early, iterate from concrete output, and verify final deliverables
+version: 1.0.0
+category: agent
+tags: [artifacts, files, verification, workflow]
+status: published
+confidence: 1.0
+source: builtin
+owner: ""
+created: "2026-08-30T00:00:00Z"
+---
+
+## When to Use
+
+Use when the task requires a file, patch, report, document, image, archive, configuration, or other persistent deliverable rather than only a text answer.
+
+## Procedure
+
+1. Extract the required deliverable path, format, content constraints, and acceptance criteria.
+2. Inspect the source material and existing target without delaying the first valid artifact.
+3. Create a minimal complete version at the required location, then iterate from that concrete output.
+4. Use the format's native parser, renderer, compiler, or test tool to inspect the artifact.
+5. Repair specific validation, content, or presentation failures while preserving correct portions.
+6. Confirm the final path, file type, required content, and usability before reporting completion.
+
+## Pitfalls
+
+- Do not spend the full task budget inspecting without creating the requested output.
+- Do not place the artifact at a convenient path when the task specifies another location.
+- Do not use a filename extension as proof that the file is valid in that format.
+- Do not report completion while placeholders, missing sections, parse errors, or failed checks remain.
+
+## Verification
+
+- The artifact exists at the required path and opens or parses successfully.
+- Required sections, fields, labels, or visual elements are present.
+- Relevant tests, render checks, or validators pass.

resources/skills/agent/terminal-recovery/SKILL.md

@@ -0,0 +1,38 @@
+---
+name: terminal-recovery
+description: Recover from failed terminal commands using evidence-driven diagnosis and bounded retries
+version: 1.0.0
+category: agent
+tags: [terminal, shell, debugging, recovery]
+status: published
+confidence: 1.0
+source: builtin
+owner: ""
+created: "2026-08-30T00:00:00Z"
+---
+
+## When to Use
+
+Use when a command fails, times out, produces incomplete output, or behaves differently from what the task requires.
+
+## Procedure
+
+1. Read the command, exit status, standard output, and standard error before choosing a response.
+2. Confirm the working directory, relevant files, executable availability, permissions, and environment assumptions with minimal read-only probes.
+3. Classify the failure as syntax, missing dependency, wrong path, permissions, resource pressure, timeout, service state, or task logic.
+4. Change one relevant condition and retry the narrowest command that can test the diagnosis.
+5. For a long-running command, use the returned session identifier to poll or provide input instead of launching duplicates.
+6. After recovery, run the original acceptance check and inspect the resulting files or service state.
+
+## Pitfalls
+
+- Do not rerun an unchanged failing command repeatedly.
+- Do not install packages or change global configuration before confirming they are missing and necessary.
+- Do not launch a second server or training job before checking for an existing process and port or device conflicts.
+- Do not treat partial output or a zero exit status as proof that the requested state was produced.
+
+## Verification
+
+- The diagnosed cause is supported by command output or environment state.
+- The corrected command exits as expected.
+- The requested artifact, process, or state passes an independent acceptance check.

resources/skills/agent/tool-discovery/SKILL.md

@@ -0,0 +1,38 @@
+---
+name: tool-discovery
+description: Discover the smallest capable tool set and confirm argument schemas before acting
+version: 1.0.0
+category: agent
+tags: [tools, discovery, routing, schemas]
+status: published
+confidence: 1.0
+source: builtin
+owner: ""
+created: "2026-08-30T00:00:00Z"
+---
+
+## When to Use
+
+Use when a task requires tools whose names, capabilities, or argument shapes are not already clear. This is especially useful when many tools are available or a previous call failed because the wrong tool or parameters were selected.
+
+## Procedure
+
+1. Translate the request into required capabilities such as reading, searching, editing, executing, browsing, or verifying.
+2. Search the tool index for those capabilities and inspect the returned tool descriptions and schemas.
+3. Prefer one direct tool over a chain of indirect tools when it can complete the operation and provide evidence.
+4. Check required parameters, identifiers, path rules, side effects, and approval requirements before calling the tool.
+5. Make a small read-only probe when the environment or target is uncertain.
+6. Execute the selected action, inspect the result, and only broaden the tool search if the result shows a concrete capability gap.
+
+## Pitfalls
+
+- Do not guess tool names or argument keys from memory when the index or schema is available.
+- Do not load unrelated tool groups into context.
+- Do not repeat the same failed call without changing the arguments or strategy.
+- Do not use a broad shell or browser workaround when a scoped native tool already owns the operation.
+
+## Verification
+
+- The chosen tool directly matches the required capability.
+- Required arguments follow the exposed schema.
+- The result contains evidence of the requested effect or a specific error that guides the next step.

resources/skills/agent/verified-state-change/SKILL.md

@@ -0,0 +1,38 @@
+---
+name: verified-state-change
+description: Make scoped state changes with target confirmation, minimal mutation, and read-back verification
+version: 1.0.0
+category: agent
+tags: [state, mutation, verification, safety]
+status: published
+confidence: 1.0
+source: builtin
+owner: ""
+created: "2026-08-30T00:00:00Z"
+---
+
+## When to Use
+
+Use when creating, editing, deleting, moving, sending, scheduling, or otherwise changing persistent state through an application, API, filesystem, or service.
+
+## Procedure
+
+1. Read the current state and identify the target using stable identifiers plus enough content to disambiguate it.
+2. Preserve fields the user did not ask to change and choose the narrowest supported mutation.
+3. For destructive or externally visible actions, confirm that the user's instruction authorizes the exact target and effect.
+4. Perform the mutation once and capture the returned identifier, status, or revision.
+5. Read the target again through an independent list, fetch, status, or content operation.
+6. Compare the observed state with the requested outcome and repair only the specific mismatch.
+
+## Pitfalls
+
+- Do not infer the target from a stale active item when a stable identifier can be fetched.
+- Do not report success from an accepted request alone; asynchronous or partial operations may not have completed.
+- Do not replace an entire object when a field-level update is supported and safer.
+- Do not silently broaden a mutation to adjacent files, records, accounts, or services.
+
+## Verification
+
+- The target identity was confirmed before mutation.
+- A read-back shows the intended values and preserves unrelated state.
+- Any external effect has a concrete status, identifier, or observable result.

resources/skills/communication/action-evidence-synthesis/SKILL.md

@@ -0,0 +1,40 @@
+---
+name: action-evidence-synthesis
+description: "Turn messages, meeting notes, and documents into sourced decisions, actions, dependencies, and risks"
+version: 1.0.0
+category: communication
+tags: [messages, meetings, actions, status, evidence]
+status: published
+confidence: 1.0
+source: builtin
+created: "2026-08-30T00:00:00Z"
+---
+
+## When to Use
+
+Use when information is fragmented across messages, meeting notes, transcripts, or documents and the user needs an action list, status summary, feasibility assessment, or executive brief.
+
+Do not use when the source material is unavailable or when the user only wants a verbatim transcript.
+
+## Procedure
+
+1. Identify the requested scope, audience, time window, and decision to support.
+2. Gather the relevant records in full and preserve stable source identifiers, authors, and timestamps.
+3. Extract explicit decisions, commitments, requests, owners, dates, dependencies, blockers, and changed facts.
+4. Reconcile revisions by preferring the newest authoritative record; keep unresolved conflicts visible instead of guessing.
+5. Separate observed facts from inferred owners, dates, urgency, feasibility, or recommendations, and label every inference as tentative.
+6. Produce the requested format with concise source references beside consequential claims and a final list of open questions.
+
+## Pitfalls
+
+- Do not turn discussion or speculation into a confirmed decision.
+- Do not invent owners or deadlines when none were assigned.
+- Do not silently discard older records that explain a changed commitment.
+- Do not send messages, create tasks, or update calendars unless the user separately authorizes those actions.
+
+## Verification
+
+- Every action has a source, status, and explicit or tentative owner and due date.
+- Conflicting values and revisions are resolved or visibly flagged.
+- The output covers decisions, actions, dependencies, risks, and open questions relevant to the request.
+

resources/skills/communication/reviewable-external-draft/SKILL.md

@@ -0,0 +1,40 @@
+---
+name: reviewable-external-draft
+description: "Reconcile source evidence and prepare an accurate external-facing draft without bypassing review"
+version: 1.0.0
+category: communication
+tags: [drafting, email, messages, review, reconciliation]
+status: published
+confidence: 1.0
+source: builtin
+created: "2026-08-30T00:00:00Z"
+---
+
+## When to Use
+
+Use when preparing a client, customer, partner, leadership, or other external-facing update from internal messages or documents.
+
+Do not use this procedure to send immediately unless the user explicitly authorizes the exact recipient and final content.
+
+## Procedure
+
+1. Confirm the audience, communication channel, requested tone, and whether the user asked for a draft or an immediate send.
+2. Gather the relevant source records and identify the latest values, dates, commitments, and unresolved discrepancies.
+3. Resolve recipient identity through the available contact source and avoid inferring internal versus external status from a display name alone.
+4. Draft only claims supported by the collected evidence; qualify uncertainty and omit internal-only detail that the audience should not receive.
+5. Save or present a reviewable draft through the native draft or document capability.
+6. Report the draft identifier or location plus any reconciliation notes that require human review.
+
+## Pitfalls
+
+- Do not send a draft merely because a send-capable tool is available.
+- Do not copy stale figures when a later correction exists.
+- Do not conceal unresolved discrepancies behind polished prose.
+- Do not expose private internal discussion, credentials, or unrelated personal data.
+
+## Verification
+
+- Recipient identity and communication mode match the request.
+- Dates, figures, status, and commitments map to current source evidence.
+- The result remains reviewable unless an explicit send-now instruction authorized delivery.
+

resources/skills/communication/scheduling-coordination/SKILL.md

@@ -0,0 +1,40 @@
+---
+name: scheduling-coordination
+description: "Coordinate availability, confirmations, calendar changes, and participant notifications with read-back verification"
+version: 1.0.0
+category: communication
+tags: [calendar, scheduling, coordination, availability]
+status: published
+confidence: 1.0
+source: builtin
+created: "2026-08-30T00:00:00Z"
+---
+
+## When to Use
+
+Use when arranging or changing a meeting across multiple participants, calendars, time zones, or communication channels.
+
+Do not create or modify an event when the user asked only for available options or a draft invitation.
+
+## Procedure
+
+1. Extract participants, duration, date range, time zones, location constraints, and required attendees.
+2. Resolve participant identities and inspect the relevant availability using declared calendar and contact capabilities.
+3. Compute candidate intervals in one explicit reference time zone and reject conflicts or insufficient travel buffers.
+4. Present or draft a small set of viable options when confirmation is still required.
+5. After authorization or recorded participant confirmation, create or update the event once with stable attendee identifiers.
+6. Read the event back and verify title, start, end, time zone, attendees, location, and conferencing details before drafting notifications.
+
+## Pitfalls
+
+- Do not overwrite or cancel unrelated events to manufacture availability.
+- Do not mix local times without naming the time zone.
+- Do not treat a proposed time as confirmed.
+- Do not create duplicates when an existing event can be updated safely.
+
+## Verification
+
+- The selected interval satisfies duration, availability, and time-zone constraints.
+- The calendar read-back matches the authorized event details.
+- Notifications describe the same confirmed event and remain drafts unless sending was explicitly authorized.
+

resources/skills/communication/support-triage-and-routing/SKILL.md

@@ -0,0 +1,39 @@
+---
+name: support-triage-and-routing
+description: "Prioritize support requests, identify owners, route internally, and prepare safe customer drafts"
+version: 1.0.0
+category: communication
+tags: [support, triage, urgency, routing, drafts]
+status: published
+confidence: 1.0
+source: builtin
+created: "2026-08-30T00:00:00Z"
+---
+
+## When to Use
+
+Use when reviewing a support backlog, identifying urgent incidents, assigning internal ownership, or drafting customer responses.
+
+Do not use when the request is merely to summarize an unrelated inbox or when sender identity cannot be established safely.
+
+## Procedure
+
+1. Read each in-scope request in full and retain its stable message or ticket identifier.
+2. Resolve whether the sender is internal or external and identify the responsible internal team from available contacts and service ownership data.
+3. Classify urgency from impact and time sensitivity: critical for outage, data loss, security exposure, or imminent contractual breach; high for a blocked user without a workaround; medium for degraded service with a workaround; low for non-blocking inquiries.
+4. Record a concise problem statement, evidence, affected scope, workaround, owner, next action, and response deadline.
+5. Route internally only when the user has authorized operational messaging; prepare external responses as reviewable drafts by default.
+6. Re-read created assignments or drafts and produce an escalation summary grouped by urgency.
+
+## Pitfalls
+
+- Do not infer severity from emotional language alone.
+- Do not expose one customer's data in another customer's response.
+- Do not send externally when the task calls for triage or drafting.
+- Do not mark an issue routed without a stable owner or observable routing result.
+
+## Verification
+
+- Every issue has a stable source identifier, urgency rationale, owner, and next action.
+- Critical and high items have explicit response targets and escalation state.
+- External communication is a draft unless the user explicitly authorized sending.

resources/skills/dev/developer-docs/SKILL.md

@@ -0,0 +1,37 @@
+---
+name: developer-docs
+description: Find, read, and apply authoritative developer documentation during implementation
+version: 1.0.0
+category: dev
+tags: [docs, documentation, api, software-development]
+status: published
+confidence: 1.0
+source: builtin
+owner: ""
+created: "2026-08-18T00:00:00Z"
+---
+
+## When to Use
+
+Use when the user asks how a library, framework, API, protocol, CLI, or SDK works, or when implementation depends on version-specific behavior. Prefer this skill over guessing from memory.
+
+## Procedure
+
+1. Identify the exact product, package, version, and task. Ask one focused clarification only when the target is genuinely ambiguous.
+2. Prefer the vendor's or project's primary documentation, source repository, release notes, and API reference. Use a general search only to locate those sources.
+3. Read the relevant page or reference section, then apply the documented behavior to the user's codebase and active workspace.
+4. Separate documented facts from inference, and call out version or environment assumptions.
+5. For code changes, add a focused regression test for the documented contract and run it before reporting completion.
+
+## Pitfalls
+
+- Do not present search snippets, stale cached knowledge, or a third-party tutorial as authoritative when primary documentation is available.
+- Do not silently mix instructions from different major versions.
+- Do not claim an API or option exists without confirming it in the relevant reference.
+- Do not use web search for a local project task when the active workspace and local tools can answer it.
+
+## Verification
+
+- The cited or retrieved documentation matches the target version.
+- The implementation or answer distinguishes source-backed facts from inference.
+- Any code change has a focused test or a concrete verification command.

resources/skills/general/test-driven-development/SKILL.md

@@ -0,0 +1,40 @@
+---
+name: test-driven-development
+description: Build or fix software with a focused red-green-refactor loop
+version: 1.0.0
+category: general
+tags: [tdd, testing, debugging, red-green-refactor]
+status: published
+confidence: 1.0
+source: builtin
+owner: ""
+created: "2026-08-18T00:00:00Z"
+---
+
+## When to Use
+
+Use when implementing a feature, fixing a bug, or changing behavior where a regression test can define the expected result. Prefer this workflow for parser, routing, agent-loop, and UI behavior changes.
+
+## Procedure
+
+1. Inspect the relevant code, existing tests, and local conventions before editing.
+2. Write the smallest regression test that demonstrates the requested behavior or reproduces the bug.
+3. Run that test and confirm it fails for the expected reason, not because the test setup is broken.
+4. Make the smallest production change that makes the test pass.
+5. Run the focused test again, then run the surrounding module suite.
+6. Review the diff for unrelated changes, brittle assertions, hidden state, and missing error paths.
+7. Report the tests run and any remaining coverage or environment limits.
+
+## Pitfalls
+
+- Do not write a test that only mirrors the implementation; assert the user-visible contract.
+- Do not weaken an assertion just to make a failing test pass.
+- Do not skip the focused failing-test step when the behavior is observable in a local test.
+- Keep network, filesystem, and model calls deterministic with fakes or fixtures unless the integration itself is under test.
+
+## Verification
+
+- The new regression test fails before the fix and passes after it.
+- The relevant focused suite passes.
+- The broader suite passes or its failure is explained with evidence.
+- The final diff contains the test and the production change needed for the same behavior.

resources/skills/media/multimodal-evidence/SKILL.md

@@ -0,0 +1,38 @@
+---
+name: multimodal-evidence
+description: Extract and verify evidence from images, documents, and video without redundant inspection
+version: 1.0.1
+category: media
+tags: [image, video, document, evidence, ocr]
+status: published
+confidence: 1.0
+source: builtin
+owner: ""
+created: "2026-08-30T00:00:00Z"
+---
+
+## When to Use
+
+Use when the answer or requested artifact depends on visual, temporal, tabular, or textual evidence contained in images, documents, or video.
+
+## Procedure
+
+1. Identify the evidence required: objects, text, values, ordering, timestamps, labels, or visual relationships.
+2. Inspect the whole input or a broad representative sample first to establish structure and likely evidence locations.
+3. Narrow to relevant pages, frames, regions, or time intervals and record observations with their locations.
+4. Use the format's native parser for exact text and numbers: for example `python-docx` or ZIP/XML inspection for DOCX, `pdftotext` or a PDF library for PDF, spreadsheet readers for XLSX, and OCR only when the source is image-based. Do not search binary office files with plain `grep` or `cat`.
+5. Resolve conflicts with one targeted reinspection at better scale or a nearby frame rather than repeating the same crop.
+6. Build the answer or artifact from the evidence ledger and perform a final coverage check against every requested item.
+
+## Pitfalls
+
+- Do not infer unseen content from filenames, surrounding text, or a single thumbnail.
+- Do not repeatedly inspect nearly identical regions without a new hypothesis.
+- Do not trust OCR blindly for small labels, punctuation, or numeric values.
+- Do not finalize before checking that every requested item has supporting evidence.
+
+## Verification
+
+- Each factual output can be traced to a page, frame, region, or timestamp.
+- Exact labels and numbers were visually checked after extraction.
+- The final response or artifact covers all requested evidence categories.

resources/skills/research/web-research-fallback/SKILL.md

@@ -0,0 +1,38 @@
+---
+name: web-research-fallback
+description: Research current web information with source-first search and controlled browser fallback
+version: 1.0.0
+category: research
+tags: [web, search, browser, sources, research]
+status: published
+confidence: 1.0
+source: builtin
+owner: ""
+created: "2026-08-30T00:00:00Z"
+---
+
+## When to Use
+
+Use when a task requires current public information, primary sources, multiple pages, or a site that cannot be reliably read from search results alone.
+
+## Procedure
+
+1. Define the facts needed and the preferred primary source for each fact.
+2. Search with a focused query and use result metadata to select likely authoritative pages.
+3. Open the source directly and extract the relevant passage, date, and URL rather than relying on a search snippet.
+4. Use the private browser when the page requires interaction, client-side rendering, navigation, or visual inspection.
+5. If a page fails, try a primary-source alternative or a narrower route before broadening to secondary sources.
+6. Cross-check unstable or consequential claims and distinguish source-backed facts from inference.
+
+## Pitfalls
+
+- Do not treat snippets as evidence for claims not visible on the source page.
+- Do not browse repeatedly without recording what each page established.
+- Do not use a secondary summary when an accessible primary source answers the question.
+- Do not claim freshness without checking publication or update dates.
+
+## Verification
+
+- Each important claim maps to a source that directly supports it.
+- Time-sensitive facts include an observed date or version.
+- Browser interaction produced the needed page state or a documented fallback was used.

routes/admin_wipe/__init__.py

@@ -0,0 +1,5 @@
+"""Admin wipe route domain package (slice 2h, #4082/#4071).
+
+Contains admin_wipe_routes.py, migrated from the flat routes/ directory.
+Backward-compat shim at routes/admin_wipe_routes.py re-exports from here.
+"""

routes/admin_wipe/admin_wipe_routes.py

@@ -0,0 +1,176 @@
+"""Admin Danger Zone — per-category wipes.
+
+Each endpoint is admin-only and truncates exactly one domain so the
+user can selectively reset memory / skills / notes / etc. without
+nuking everything. The catch-all `chats` endpoint mirrors the
+existing /api/sessions/all so the Danger Zone speaks one URL pattern.
+
+URL shape: DELETE /api/admin/wipe/{kind}
+Kinds: chats, memory, skills, notes, tasks, documents, gallery, calendar.
+"""
+
+import json
+import logging
+import os
+import shutil
+from fastapi import APIRouter, HTTPException, Request
+
+from core.middleware import require_admin
+from core.database import (
+    SessionLocal,
+    Session as DbSession,
+    ChatMessage as DbChatMessage,
+    Memory,
+    Note,
+    ScheduledTask,
+    TaskRun,
+    Document,
+    DocumentVersion,
+    GalleryImage,
+    GalleryAlbum,
+    CalendarEvent,
+    CalendarCal,
+)
+from src.constants import DATA_DIR, SKILLS_DIR, SKILLS_FILE, GALLERY_DIR, GALLERY_UPLOADS_DIR
+
+logger = logging.getLogger(__name__)
+
+
+def _wipe_memory_files():
+    """Blank memory.json + drop the per-owner tidy-state sidecar so the
+    next audit doesn't try to diff against gone memories."""
+    for name in ("memory.json", "memory_tidy_state.json"):
+        p = os.path.join(DATA_DIR, name)
+        if not os.path.exists(p):
+            continue
+        try:
+            if name == "memory.json":
+                with open(p, "w", encoding="utf-8") as f:
+                    json.dump([], f)
+            else:
+                os.remove(p)
+        except OSError as e:
+            logger.warning(f"Could not reset {name}: {e}")
+
+
+def _rmtree_quiet(path: str):
+    """rmtree that doesn't crash if the path doesn't exist."""
+    if os.path.isdir(path):
+        try:
+            shutil.rmtree(path)
+        except OSError as e:
+            logger.warning(f"Could not remove {path}: {e}")
+
+
+def setup_admin_wipe_routes(session_manager):
+    """The session_manager is passed in so we can also clear its
+    in-memory cache when wiping chats — without it the DB is empty
+    but the next /api/sessions returns stale entries."""
+    router = APIRouter(prefix="/api/admin")
+
+    @router.delete("/wipe/{kind}")
+    def wipe(kind: str, request: Request):
+        require_admin(request)
+        kind = (kind or "").strip().lower()
+
+        db = SessionLocal()
+        try:
+            if kind == "chats":
+                count = db.query(DbSession).count()
+                db.query(DbChatMessage).delete()
+                db.query(DbSession).delete()
+                db.commit()
+                try:
+                    session_manager.sessions.clear()
+                except Exception:
+                    pass
+                return {"status": "deleted", "kind": kind, "count": count}
+
+            if kind == "memory":
+                count = db.query(Memory).count()
+                db.query(Memory).delete()
+                db.commit()
+                _wipe_memory_files()
+                # Drop the vector store too so semantic search doesn't
+                # return ghosts. Lazy import — chromadb may not be
+                # initialised in every deployment.
+                try:
+                    from src.memory_vector import get_memory_vector_store
+                    mv = get_memory_vector_store()
+                    if mv and hasattr(mv, "clear"):
+                        mv.clear()
+                except Exception as e:
+                    logger.info(f"Memory vector clear skipped: {e}")
+                return {"status": "deleted", "kind": kind, "count": count}
+
+            if kind == "skills":
+                # Skills live as SKILL.md files under data/skills/. Drop
+                # the entire directory; the SkillsManager re-creates the
+                # tree on next write.
+                skills_dir = SKILLS_DIR
+                count = 0
+                if os.path.isdir(skills_dir):
+                    # Count SKILL.md files for the response — quick walk.
+                    for _, _, files in os.walk(skills_dir):
+                        count += sum(1 for f in files if f == "SKILL.md")
+                    _rmtree_quiet(skills_dir)
+                # Legacy fallback file
+                legacy = SKILLS_FILE
+                if os.path.exists(legacy):
+                    try:
+                        os.remove(legacy)
+                    except OSError:
+                        pass
+                return {"status": "deleted", "kind": kind, "count": count}
+
+            if kind == "notes":
+                count = db.query(Note).count()
+                db.query(Note).delete()
+                db.commit()
+                return {"status": "deleted", "kind": kind, "count": count}
+
+            if kind == "tasks":
+                # TaskRun rows reference tasks via FK — clear them first.
+                db.query(TaskRun).delete()
+                count = db.query(ScheduledTask).count()
+                db.query(ScheduledTask).delete()
+                db.commit()
+                return {"status": "deleted", "kind": kind, "count": count}
+
+            if kind == "documents":
+                # DocumentVersion FKs Document — clear children first.
+                db.query(DocumentVersion).delete()
+                count = db.query(Document).count()
+                db.query(Document).delete()
+                db.commit()
+                return {"status": "deleted", "kind": kind, "count": count}
+
+            if kind == "gallery":
+                count = db.query(GalleryImage).count() + db.query(GalleryAlbum).count()
+                db.query(GalleryImage).delete()
+                db.query(GalleryAlbum).delete()
+                db.commit()
+                # Also drop the upload dir so disk doesn't keep orphans.
+                _rmtree_quiet(GALLERY_DIR)
+                _rmtree_quiet(GALLERY_UPLOADS_DIR)
+                return {"status": "deleted", "kind": kind, "count": count}
+
+            if kind == "calendar":
+                # Events FK calendars — clear children first, then both.
+                db.query(CalendarEvent).delete()
+                count = db.query(CalendarCal).count()
+                db.query(CalendarCal).delete()
+                db.commit()
+                return {"status": "deleted", "kind": kind, "count": count}
+
+            raise HTTPException(400, f"Unknown wipe kind: {kind!r}")
+        except HTTPException:
+            raise
+        except Exception as e:
+            db.rollback()
+            logger.exception(f"Wipe {kind} failed")
+            raise HTTPException(500, f"Wipe {kind} failed: {e}")
+        finally:
+            db.close()
+
+    return router

routes/admin_wipe_routes.py

@@ -1,176 +1,17 @@
-"""Admin Danger Zone — per-category wipes.
+"""Backward-compat shim — canonical location is routes/admin_wipe/admin_wipe_routes.py.
 
-Each endpoint is admin-only and truncates exactly one domain so the
-user can selectively reset memory / skills / notes / etc. without
-nuking everything. The catch-all `chats` endpoint mirrors the
-existing /api/sessions/all so the Danger Zone speaks one URL pattern.
-
-URL shape: DELETE /api/admin/wipe/{kind}
-Kinds: chats, memory, skills, notes, tasks, documents, gallery, calendar.
+This module is replaced in ``sys.modules`` by the canonical module object so
+that ``import routes.admin_wipe_routes``, ``from routes.admin_wipe_routes
+import X``, ``importlib.import_module("routes.admin_wipe_routes")``, and the
+``import ... as admin_wipe_routes`` + ``monkeypatch.setattr(admin_wipe_routes,
+"SessionLocal", ...)`` / ``"require_admin"`` pattern used by
+test_admin_wipe_gallery.py all operate on the *same* object the application
+actually uses. Keeps existing import paths working after slice 2h
+(#4082/#4071).
 """
 
-import json
-import logging
-import os
-import shutil
-from fastapi import APIRouter, HTTPException, Request
+import sys as _sys
 
-from core.middleware import require_admin
-from core.database import (
-    SessionLocal,
-    Session as DbSession,
-    ChatMessage as DbChatMessage,
-    Memory,
-    Note,
-    ScheduledTask,
-    TaskRun,
-    Document,
-    DocumentVersion,
-    GalleryImage,
-    GalleryAlbum,
-    CalendarEvent,
-    CalendarCal,
-)
-from src.constants import DATA_DIR, SKILLS_DIR, SKILLS_FILE, GALLERY_DIR, GALLERY_UPLOADS_DIR
+from routes.admin_wipe import admin_wipe_routes as _canonical  # noqa: F401
 
-logger = logging.getLogger(__name__)
-
-
-def _wipe_memory_files():
-    """Blank memory.json + drop the per-owner tidy-state sidecar so the
-    next audit doesn't try to diff against gone memories."""
-    for name in ("memory.json", "memory_tidy_state.json"):
-        p = os.path.join(DATA_DIR, name)
-        if not os.path.exists(p):
-            continue
-        try:
-            if name == "memory.json":
-                with open(p, "w", encoding="utf-8") as f:
-                    json.dump([], f)
-            else:
-                os.remove(p)
-        except OSError as e:
-            logger.warning(f"Could not reset {name}: {e}")
-
-
-def _rmtree_quiet(path: str):
-    """rmtree that doesn't crash if the path doesn't exist."""
-    if os.path.isdir(path):
-        try:
-            shutil.rmtree(path)
-        except OSError as e:
-            logger.warning(f"Could not remove {path}: {e}")
-
-
-def setup_admin_wipe_routes(session_manager):
-    """The session_manager is passed in so we can also clear its
-    in-memory cache when wiping chats — without it the DB is empty
-    but the next /api/sessions returns stale entries."""
-    router = APIRouter(prefix="/api/admin")
-
-    @router.delete("/wipe/{kind}")
-    def wipe(kind: str, request: Request):
-        require_admin(request)
-        kind = (kind or "").strip().lower()
-
-        db = SessionLocal()
-        try:
-            if kind == "chats":
-                count = db.query(DbSession).count()
-                db.query(DbChatMessage).delete()
-                db.query(DbSession).delete()
-                db.commit()
-                try:
-                    session_manager.sessions.clear()
-                except Exception:
-                    pass
-                return {"status": "deleted", "kind": kind, "count": count}
-
-            if kind == "memory":
-                count = db.query(Memory).count()
-                db.query(Memory).delete()
-                db.commit()
-                _wipe_memory_files()
-                # Drop the vector store too so semantic search doesn't
-                # return ghosts. Lazy import — chromadb may not be
-                # initialised in every deployment.
-                try:
-                    from src.memory_vector import get_memory_vector_store
-                    mv = get_memory_vector_store()
-                    if mv and hasattr(mv, "clear"):
-                        mv.clear()
-                except Exception as e:
-                    logger.info(f"Memory vector clear skipped: {e}")
-                return {"status": "deleted", "kind": kind, "count": count}
-
-            if kind == "skills":
-                # Skills live as SKILL.md files under data/skills/. Drop
-                # the entire directory; the SkillsManager re-creates the
-                # tree on next write.
-                skills_dir = SKILLS_DIR
-                count = 0
-                if os.path.isdir(skills_dir):
-                    # Count SKILL.md files for the response — quick walk.
-                    for _, _, files in os.walk(skills_dir):
-                        count += sum(1 for f in files if f == "SKILL.md")
-                    _rmtree_quiet(skills_dir)
-                # Legacy fallback file
-                legacy = SKILLS_FILE
-                if os.path.exists(legacy):
-                    try:
-                        os.remove(legacy)
-                    except OSError:
-                        pass
-                return {"status": "deleted", "kind": kind, "count": count}
-
-            if kind == "notes":
-                count = db.query(Note).count()
-                db.query(Note).delete()
-                db.commit()
-                return {"status": "deleted", "kind": kind, "count": count}
-
-            if kind == "tasks":
-                # TaskRun rows reference tasks via FK — clear them first.
-                db.query(TaskRun).delete()
-                count = db.query(ScheduledTask).count()
-                db.query(ScheduledTask).delete()
-                db.commit()
-                return {"status": "deleted", "kind": kind, "count": count}
-
-            if kind == "documents":
-                # DocumentVersion FKs Document — clear children first.
-                db.query(DocumentVersion).delete()
-                count = db.query(Document).count()
-                db.query(Document).delete()
-                db.commit()
-                return {"status": "deleted", "kind": kind, "count": count}
-
-            if kind == "gallery":
-                count = db.query(GalleryImage).count() + db.query(GalleryAlbum).count()
-                db.query(GalleryImage).delete()
-                db.query(GalleryAlbum).delete()
-                db.commit()
-                # Also drop the upload dir so disk doesn't keep orphans.
-                _rmtree_quiet(GALLERY_DIR)
-                _rmtree_quiet(GALLERY_UPLOADS_DIR)
-                return {"status": "deleted", "kind": kind, "count": count}
-
-            if kind == "calendar":
-                # Events FK calendars — clear children first, then both.
-                db.query(CalendarEvent).delete()
-                count = db.query(CalendarCal).count()
-                db.query(CalendarCal).delete()
-                db.commit()
-                return {"status": "deleted", "kind": kind, "count": count}
-
-            raise HTTPException(400, f"Unknown wipe kind: {kind!r}")
-        except HTTPException:
-            raise
-        except Exception as e:
-            db.rollback()
-            logger.exception(f"Wipe {kind} failed")
-            raise HTTPException(500, f"Wipe {kind} failed: {e}")
-        finally:
-            db.close()
-
-    return router
+_sys.modules[__name__] = _canonical