chore(release): align dev version with 1.0.3

Keep dev version metadata aligned with the current hotfix release while the rolling branch continues toward 1.1.0.

Evidence: the canonical APP_VERSION imports as 1.0.3 and the diff check is clean. This commit changes version metadata only; it does not tag or publish a release.

.env.example

@@ -88,8 +88,12 @@ SEARXNG_INSTANCE=http://localhost:8080
 # 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.
+# 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.
@@ -159,6 +163,21 @@ SEARXNG_INSTANCE=http://localhost:8080
 # 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
 # ============================================================

.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/ISSUE_TEMPLATE/bug_report.yml

@@ -26,6 +26,18 @@ body:
         - 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/pull_request_template.md

@@ -28,6 +28,7 @@ Fixes #
 - [ ] 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');
       }

.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/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
@@ -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:

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 |
 

README.md

@@ -68,11 +68,11 @@ Deployment details are in the [setup guide](docs/setup.md#security-notes).
 
 ## Star History
 
-<a href="https://www.star-history.com/?repos=odysseus-dev%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=odysseus-dev/odysseus&type=date&theme=dark&legend=top-left" />
-   <source media="(prefers-color-scheme: light)" srcset="https://api.star-history.com/chart?repos=odysseus-dev/odysseus&type=date&legend=top-left" />
-   <img alt="Star History Chart" src="https://api.star-history.com/chart?repos=odysseus-dev/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>
 

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.

app.py

@@ -784,7 +784,7 @@ 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

build-macos-app.sh

@@ -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"

core/atomic_io.py

@@ -30,11 +30,20 @@ def atomic_write_json(path: str, data: Any, *, indent: Optional[int] = None) ->
     """
     os.makedirs(os.path.dirname(path) or ".", exist_ok=True)
     tmp = f"{path}.tmp.{uuid.uuid4().hex}"
-    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)
+
+    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:
@@ -42,8 +51,17 @@ def atomic_write_text(path: str, text: str) -> None:
         raise TypeError("atomic_write_text expects a string")
     os.makedirs(os.path.dirname(path) or ".", exist_ok=True)
     tmp = f"{path}.tmp.{uuid.uuid4().hex}"
-    with open(tmp, "w", encoding="utf-8") as f:
-        f.write(text)
-        f.flush()
-        os.fsync(f.fileno())
-    os.replace(tmp, path)
+
+    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/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."""
@@ -116,11 +150,27 @@ 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"
         ]
+        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

@@ -14,6 +14,8 @@ 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
@@ -92,14 +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:

docker-compose.gpu-amd.yml

@@ -50,7 +50,7 @@ services:
       - 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:-}
@@ -75,6 +75,11 @@ services:
       - 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

docker-compose.gpu-nvidia.yml

@@ -49,7 +49,7 @@ services:
       - 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:-}
@@ -74,6 +74,11 @@ services:
       - 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

docker-compose.yml

@@ -38,7 +38,7 @@ services:
       - 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:-}
@@ -63,6 +63,11 @@ services:
       - 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

docs/setup.md

@@ -441,10 +441,19 @@ A grab-bag of small gotchas that otherwise turn into long debugging sessions.
 | Package | Feature unlocked |
 |---------|-----------------|
 | `faster-whisper` | Local speech-to-text (microphone -> text) via the "local" STT provider. |
+| `kokoro`, `soundfile` | Local Kokoro-82M text-to-speech on a CUDA GPU. The pinned Kokoro release supports Odysseus installs on Python 3.11-3.12; these packages are intentionally skipped on Python 3.13+ (including the Python 3.14 container image). |
 | `ddgs` | DuckDuckGo as a search provider option. |
 | `PyMuPDF` | PDF page rendering in the side viewer panel and form-filling. (Note: AGPL-3.0) |
 | `markitdown` | Office/EPUB document text extraction (converts .docx/.xlsx/.pptx/.xls/.epub to Markdown). |
 
+Install the optional set only when you need these features:
+
+```bash
+pip install -r requirements-optional.txt
+```
+
+The default Docker image currently uses Python 3.14, while Kokoro 0.9.4 declares Python `>=3.10,<3.13`. Odysseus itself continues to support Python 3.11+, but this pinned optional local-TTS feature requires a native Python 3.11 or 3.12 environment. Kokoro declares `torch`, but the local provider only activates when that torch build has CUDA and a GPU is visible; install the CUDA build appropriate for your host. Browser and configured endpoint TTS remain available on Python 3.13+ and in the container image.
+
 ### Faster, reproducible installs with uv (optional)
 [uv](https://docs.astral.sh/uv/) works as a drop-in replacement for the
 venv + pip steps in the native install guides, no project changes are needed but this change results in faster installs along with a lockfile for reproducible environments. After [installing `uv`](https://docs.astral.sh/uv/getting-started/installation/), use:
@@ -475,7 +484,7 @@ Odysseus is a self-hosted workspace with powerful local tools: shell access, fil
 
 - Keep `AUTH_ENABLED=true` for any network-accessible deployment.
 - Keep `LOCALHOST_BYPASS=false` outside local development.
-- Use `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. Use `SECURE_COOKIES=true` to force it on for a proxy whose scheme Odysseus cannot see, or `SECURE_COOKIES=false` to force it off while you still serve plain HTTP alongside HTTPS.
 - Do not expose it directly to the public internet without HTTPS and a trusted reverse proxy or private access layer.
 - Keep `.env`, `data/`, `logs/`, databases, uploads, generated media, backups, auth/session files, API keys, and model/provider tokens out of Git and private shares. They are ignored by default.
 - Review `data/auth.json` after first boot: disable open signup unless you intentionally want it, make only your own account admin, and keep demo/test accounts non-admin.
@@ -486,6 +495,14 @@ Odysseus is a self-hosted workspace with powerful local tools: shell access, fil
 - Keep ChromaDB, SearXNG, ntfy, Ollama, vLLM, llama.cpp, databases, and raw model/provider APIs internal-only. Expose only the authenticated Odysseus web/API entrypoint through your trusted proxy or private access layer.
 - Before publishing a fork, run `git status --short` and confirm no private files from `.env`, `data/`, `logs/`, uploads, backups, or local databases are staged.
 
+> **Upgrading an existing install:** `SECURE_COOKIES` used to default to
+> `false`, so an install set up before scheme derivation may still carry
+> `SECURE_COOKIES=false` in its own `.env`. That explicit value stays
+> authoritative, so HTTPS logins keep getting a non-`Secure` session cookie.
+> Pulling this change updates the tracked Compose files, but nothing rewrites
+> your `.env` — drop the line from it unless you deliberately serve plain HTTP
+> alongside HTTPS and want the escape hatch.
+
 ### Private or proxied deployments
 Odysseus serves plain HTTP on its app port. Docker Compose binds Odysseus and the bundled services to `127.0.0.1` by default, so a typical production/private setup is:
 
@@ -494,7 +511,7 @@ Odysseus serves plain HTTP on its app port. Docker Compose binds Odysseus and th
 3. Put the authenticated Odysseus web/API entrypoint behind that layer.
 4. Keep raw service and model ports internal-only.
 
-Cloudflare Access, Tailscale, Caddy, nginx, and Traefik can all fit this pattern; none are required by Odysseus. If your access layer reaches Odysseus on the same host, proxy to `http://127.0.0.1:7000` and keep `AUTH_ENABLED=true`, `LOCALHOST_BYPASS=false`, and `SECURE_COOKIES=true`.
+Cloudflare Access, Tailscale, Caddy, nginx, and Traefik can all fit this pattern; none are required by Odysseus. If your access layer reaches Odysseus on the same host, proxy to `http://127.0.0.1:7000` and keep `AUTH_ENABLED=true` and `LOCALHOST_BYPASS=false`. Any proxy that forwards `X-Forwarded-Proto: https` gets `Secure` session cookies without configuration, so `SECURE_COOKIES` only needs setting when you want to override that — force it on for a proxy that forwards no scheme at all, or off while you still serve plain HTTP.
 `ALLOWED_ORIGINS` lists exact permitted origins for cross-origin browser/API clients; ordinary same-origin reverse-proxy access usually does not need a special CORS entry.
 
 #### Faster over the network: HTTP/2
@@ -582,9 +599,12 @@ Odysseus's own service is unchanged; the proxy runs alongside it. Under Docker,
 run the proxy as another container, or on the host pointing at the published
 port.
 
-**4. Point Odysseus at the new origin** in `.env`, then restart it:
+**4. Point Odysseus at the new origin** in `.env`, then restart it.
+
+A proxy that exposes the HTTPS request scheme to Odysseus needs no `SECURE_COOKIES` setting. Only force it on when the proxy cannot expose that scheme:
 
 ```bash
+# only if the proxy cannot expose the external HTTPS scheme to Odysseus:
 SECURE_COOKIES=true
 # only if you use remote MCP servers with OAuth:
 OAUTH_REDIRECT_BASE_URL=https://odysseus.example.com
@@ -619,10 +639,12 @@ enable it by right-clicking the column headers.
 
 Three things bite when moving an existing install behind TLS:
 
-- Set `SECURE_COOKIES=true` **at the same time** you stop serving plain HTTP,
-  not before. The flag is applied to every login regardless of the scheme the
-  request arrived on, so while an HTTP entrypoint is still reachable the
-  browser will reject the `Secure` cookie there and login will appear to loop.
+- Leave `SECURE_COOKIES` unset when Odysseus can see the external HTTPS scheme;
+  the cookie then follows the request automatically. If your proxy cannot expose
+  that scheme, set `SECURE_COOKIES=true` **at the same time** you stop serving
+  plain HTTP, not before. An explicit `true` applies to every login, so while an
+  HTTP entrypoint is still reachable the browser will reject the `Secure` cookie
+  there and login will appear to loop.
 - `OAUTH_REDIRECT_BASE_URL` defaults to `http://localhost:7000`. Unlike the
   Gmail redirect URI it cannot be derived from a request — it is registered
   with each MCP authorization server up front — so set it to the external
@@ -675,7 +697,7 @@ Key settings:
 | `AUTH_ENABLED` | `true` | Enable/disable login |
 | `LOCALHOST_BYPASS` | `false` | Development-only auth bypass for loopback requests. Keep false for shared/network deployments. |
 | `ALLOWED_ORIGINS` | `http://localhost,http://127.0.0.1` | Comma-separated exact permitted origins for cross-origin browser/API clients. |
-| `SECURE_COOKIES` | `false` | Set true when serving Odysseus through HTTPS at a trusted proxy or private access gateway. |
+| `SECURE_COOKIES` | derived from the request scheme | Marks session cookies `Secure` on HTTPS requests. Set true to force it on, false to force it off. |
 | `DATABASE_URL` | `sqlite:///./data/app.db` | Database connection string |
 | `CHROMADB_HOST` | `localhost` | ChromaDB host for vector memory. Docker overrides this to `chromadb`. |
 | `CHROMADB_PORT` | `8100` | ChromaDB port for manual host runs. Docker overrides this to `8000`. |

launch-windows.ps1

@@ -163,6 +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 ""

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.

requirements-optional.txt

@@ -12,6 +12,16 @@
 # 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.

routes/auth_routes.py

@@ -86,6 +86,33 @@ class SetOpenRegistrationRequest(BaseModel):
 SESSION_COOKIE = "odysseus_session"
 
 
+def _secure_cookie(request: Request) -> bool:
+    """Decide the ``Secure`` attribute of the session cookie.
+
+    ``SECURE_COOKIES`` stays authoritative when it holds an explicit value:
+    ``true`` always marks the cookie Secure (the documented knob for a TLS
+    proxy), ``false`` never does, which is the escape hatch for an install
+    that still answers on plain HTTP alongside HTTPS. Anything else —
+    unset, or the present-but-empty value docker-compose injects for a
+    variable the host has not defined — derives it from the request, so an
+    HTTPS login gets a Secure cookie without any configuration.
+
+    Either the connection scheme or ``X-Forwarded-Proto`` saying https is
+    enough, which is the same test ``core/middleware.py`` applies before it
+    sends HSTS. Uvicorn's proxy-headers middleware already folds that header
+    into the scheme for the proxies it trusts, so reading it here only adds
+    the case of a terminator that is not on a trusted address; the cost is
+    that a client talking to the app directly can set the header and lock
+    its own session out over plain HTTP.
+    """
+    configured = os.getenv("SECURE_COOKIES", "").strip().lower()
+    if configured in ("true", "false"):
+        return configured == "true"
+    # A chained proxy sends a list — the client-facing hop comes first.
+    forwarded_proto = request.headers.get("x-forwarded-proto", "").split(",")[0]
+    return request.url.scheme == "https" or forwarded_proto.strip().lower() == "https"
+
+
 def setup_auth_routes(auth_manager: AuthManager) -> APIRouter:
     router = APIRouter(prefix="/api/auth", tags=["auth"])
 
@@ -159,7 +186,7 @@ def setup_auth_routes(auth_manager: AuthManager) -> APIRouter:
             value=token,
             httponly=True,
             samesite="lax",
-            secure=os.getenv("SECURE_COOKIES", "false").lower() == "true",
+            secure=_secure_cookie(request),
             path="/",
         )
         if body.remember:

routes/chat_helpers.py

@@ -624,6 +624,8 @@ async def build_chat_context(
     agent_mode: bool = False,
     allow_tool_preprocessing: bool = True,
     defer_context_shaping: bool = False,
+    continuation_context_message: str | None = None,
+    persist_user_message: bool = True,
 ) -> ChatContext:
     """Build the full context (preface + messages) for an LLM call.
 
@@ -647,14 +649,14 @@ async def build_chat_context(
     # Add user message to history. Nobody/incognito uses a request-local
     # transcript store instead of session history so stale saved chats cannot
     # bleed into context and the turn is not persisted.
-    if incognito:
+    if persist_user_message and incognito:
         user_meta = {"attachments": preprocessed.attachment_meta} if preprocessed.attachment_meta else None
         _append_incognito_message(session_id, "user", preprocessed.user_content, user_meta)
-    else:
+    elif persist_user_message:
         add_user_message(sess, chat_handler, preprocessed, incognito=False)
 
     # Fire events
-    if not incognito:
+    if persist_user_message and not incognito:
         fire_message_event(request, webhook_manager, session_id, sess, message, compare_mode)
 
     # Resolve owner-scoped prefs/context. Browser requests keep the cookie user;
@@ -666,7 +668,12 @@ async def build_chat_context(
         getattr(chat_handler, "upload_handler", None),
         getattr(sess, "owner", None),
     )
-    casual_low_signal = _is_casual_low_signal(message)
+    context_message = (
+        str(continuation_context_message).strip()
+        if continuation_context_message
+        else message
+    )
+    casual_low_signal = _is_casual_low_signal(context_message)
 
     # Memory enabled?
     mem_enabled = not incognito and not no_memory and uprefs.get("memory_enabled", True)
@@ -703,7 +710,15 @@ async def build_chat_context(
     # Build context preface
     # The stream path uses enhanced_message (with CoT/preprocessing applied),
     # the sync path uses text_for_context.
-    _ctx_msg = preprocessed.enhanced_message if use_enhanced_message else preprocessed.text_for_context
+    _ctx_msg = (
+        context_message
+        if continuation_context_message
+        else (
+            preprocessed.enhanced_message
+            if use_enhanced_message
+            else preprocessed.text_for_context
+        )
+    )
     _preface_kwargs = dict(
         message=_ctx_msg,
         session=sess,

routes/chat_routes.py

@@ -89,6 +89,65 @@ def _stream_failure_status(chunk: str) -> Optional[int]:
     return None
 
 
+def _mark_tool_approval_resolved(sess, approval_id: Any, decision: Any) -> bool:
+    """Persist a consumed approval decision on its existing tool event."""
+
+    approval_key = str(approval_id or "")
+    normalized_decision = str(decision or "").strip().lower()
+    if not approval_key or normalized_decision not in {"approve", "approve_task", "deny"}:
+        return False
+
+    message_id = None
+    resolved_metadata = None
+    for item in reversed(getattr(sess, "history", []) or []):
+        metadata = getattr(item, "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 str(ask_user.get("approval_id") or "") != approval_key:
+                continue
+            ask_user["resolved"] = normalized_decision
+            message_id = metadata.get("_db_id")
+            resolved_metadata = {
+                key: value for key, value in metadata.items() if key != "_db_id"
+            }
+            break
+        if resolved_metadata is not None:
+            break
+
+    if resolved_metadata is None or not message_id:
+        return False
+
+    db = SessionLocal()
+    try:
+        db_message = db.query(DBChatMessage).filter(
+            DBChatMessage.id == message_id,
+            DBChatMessage.session_id == str(getattr(sess, "id", "")),
+        ).first()
+        if db_message is None:
+            return False
+        db_message.meta_data = json.dumps(resolved_metadata)
+        db.commit()
+        return True
+    except Exception:
+        db.rollback()
+        logger.exception("Failed to persist tool approval resolution")
+        return False
+    finally:
+        db.close()
+
+
+async def _tool_approval_resolution_stream(decision: str) -> AsyncGenerator[str, None]:
+    yield f"data: {json.dumps({'type': 'tool_approval_resolved', 'decision': decision})}\n\n"
+    yield "data: [DONE]\n\n"
+
+
 def _chat_candidate_request_factory(
     messages,
     fallback_context_length: int = 0,
@@ -917,6 +976,7 @@ def setup_chat_routes(
         exact_tool_approval = None
         pending_tool_approval = None
         retired_tool_approval_taint = False
+        external_untrusted_context_seen = False
         tool_approval_continuation = False
         # Workspace: confine the agent's file/shell tools to this folder.
         workspace, workspace_rejected = _resolve_request_workspace(
@@ -1050,14 +1110,14 @@ def setup_chat_routes(
             )
 
         try:
-            # Attachment-only sends: skip the message-required check when the
-            # user has attached one or more files (the attachment IS the action).
+            # Attachment-only sends and approval controls may omit message text.
             _has_atts = (
                 bool(body and isinstance(body.get("attachments"), list) and body["attachments"])
                 or bool(form_data.get("attachments"))
             )
             message, session = coerce_message_and_session(
-                body, message, session, session_manager, allow_empty=_has_atts,
+                body, message, session, session_manager,
+                allow_empty=(_has_atts or bool(tool_approval_id)),
             )
             # Verify ownership AFTER coerce (which may resolve a default session)
             # but BEFORE loading. Prevents cross-user session hijack.
@@ -1076,8 +1136,14 @@ def setup_chat_routes(
                         409,
                         "This tool approval is invalid, expired, or belongs to another thread.",
                     )
+                pending_taint = bool(
+                    pending_tool_approval.external_untrusted_context_seen
+                )
+                external_untrusted_context_seen = (
+                    external_untrusted_context_seen or pending_taint
+                )
                 decision = str(tool_approval_decision or "").strip().lower()
-                if decision not in {"approve", "deny"}:
+                if decision not in {"approve", "approve_task", "deny"}:
                     raise HTTPException(400, "Invalid tool approval decision.")
                 if plan_mode:
                     raise HTTPException(
@@ -1091,37 +1157,46 @@ def setup_chat_routes(
                     session_id=session,
                 )
                 tool_approval_continuation = True
-                if decision == "approve" and exact_tool_approval is None:
+                if (
+                    decision in {"approve", "approve_task"}
+                    and exact_tool_approval is None
+                ):
                     raise HTTPException(
                         409,
                         "This tool approval could not be consumed.",
                     )
-                if decision == "approve":
-                    message = (
-                        f"Approved the exact {pending_tool_approval.tool_name} action "
-                        "shown above once."
+                if not _mark_tool_approval_resolved(
+                    sess,
+                    tool_approval_id,
+                    decision,
+                ):
+                    logger.warning(
+                        "Tool approval %s was consumed but its persisted card could not be marked resolved",
+                        tool_approval_id,
                     )
-                    # The sealed server record, not mutable composer state,
-                    # restores the original action workspace.
-                    workspace = pending_tool_approval.workspace or None
-                    workspace_rejected = None
-                    if pending_tool_approval.document_id:
-                        active_doc_id = pending_tool_approval.document_id
-                    # The approval click is the per-turn opt-in for this exact
-                    # sealed action. Restore only the coarse request toggle
-                    # that would otherwise disable it because the synthetic
-                    # "Approved…" message no longer resembles the original
-                    # shell/web request. Current privilege, global-disable,
-                    # incognito, compare, and tool-policy gates still run.
-                    if pending_tool_approval.tool_name == "bash":
-                        allow_bash = "true"
-                    if pending_tool_approval.tool_name in WEB_TOOL_NAMES:
-                        allow_web_search = "true"
-                        _search_enabled = True
-                else:
-                    message = (
-                        f"Denied the {pending_tool_approval.tool_name} action shown above."
+                if decision == "deny":
+                    return StreamingResponse(
+                        _tool_approval_resolution_stream(decision),
+                        media_type="text/event-stream",
                     )
+                # Approval is a control-plane continuation, not a new user turn.
+                # Reuse the sealed interrupted request only for internal context,
+                # retrieval, and policy reconstruction; never persist or display it.
+                message = pending_tool_approval.continuation_query
+                # The sealed server record, not mutable composer state,
+                # restores the original action workspace.
+                workspace = pending_tool_approval.workspace or None
+                workspace_rejected = None
+                if pending_tool_approval.document_id:
+                    active_doc_id = pending_tool_approval.document_id
+                # Restore only the coarse request toggle needed by the exact
+                # sealed action. Current privilege, global-disable, incognito,
+                # compare, and tool-policy gates still run.
+                if pending_tool_approval.tool_name == "bash":
+                    allow_bash = "true"
+                if pending_tool_approval.tool_name in WEB_TOOL_NAMES:
+                    allow_web_search = "true"
+                    _search_enabled = True
                 chat_mode = "agent"
             else:
                 # A normal user message supersedes the card that was waiting
@@ -1132,6 +1207,9 @@ def setup_chat_routes(
                     owner=owner,
                     session_id=session,
                 )
+                external_untrusted_context_seen = (
+                    external_untrusted_context_seen or retired_tool_approval_taint
+                )
             _reconcile_selected_route_from_request(request, sess, session, form_data, owner=owner)
             if _clear_orphaned_session_endpoint(sess, owner=owner):
                 raise HTTPException(400, "Selected model endpoint was removed. Pick another model in Settings.")
@@ -1261,6 +1339,14 @@ def setup_chat_routes(
             agent_mode=(chat_mode == "agent"),
             allow_tool_preprocessing=allow_tool_preprocessing,
             defer_context_shaping=foreground_policy.enabled,
+            continuation_context_message=(
+                pending_tool_approval.continuation_query
+                if exact_tool_approval
+                and pending_tool_approval
+                and pending_tool_approval.continuation_query
+                else None
+            ),
+            persist_user_message=not tool_approval_continuation,
         )
 
         _research_flags = {"do": do_research}  # Mutable container for generator scope
@@ -1663,7 +1749,11 @@ def setup_chat_routes(
                 if foreground_policy.enabled
                 else ctx.messages
             )
-            messages = _ensure_current_request_is_latest_user(context_source, message)
+            messages = (
+                list(context_source)
+                if tool_approval_continuation
+                else _ensure_current_request_is_latest_user(context_source, message)
+            )
 
             # Auto-compact notification
             if ctx.was_compacted:
@@ -2134,7 +2224,10 @@ def setup_chat_routes(
                                     incognito=incognito, compare_mode=compare_mode,
                                     character_name=ctx.preset.character_name,
                                     owner=_user,
-                                    allow_background_extraction=not tool_policy.block_all_tool_calls,
+                                    allow_background_extraction=(
+                                        not tool_policy.block_all_tool_calls
+                                        and not tool_approval_continuation
+                                    ),
                                 )
                             _stream_set(session, status="done")
                             yield chunk
@@ -2223,17 +2316,17 @@ def setup_chat_routes(
                         plan_mode=plan_mode,
                         approved_plan=approved_plan or None,
                         workspace=workspace or None,
+                        relevant_tools=(
+                            set(pending_tool_approval.selected_tools)
+                            if exact_tool_approval
+                            and pending_tool_approval
+                            and pending_tool_approval.selected_tools
+                            else None
+                        ),
                         forced_tools=_forced_tools,
                         uploaded_files=ctx.uploaded_files,
                         defer_context_shaping=_foreground_policy.enabled,
-                        external_untrusted_context_seen=bool(
-                            retired_tool_approval_taint
-                            or (
-                                tool_approval_continuation
-                                and pending_tool_approval
-                                and pending_tool_approval.external_untrusted_context_seen
-                            )
-                        ),
+                        external_untrusted_context_seen=external_untrusted_context_seen,
                         exact_approval=exact_tool_approval,
                     ):
                         if chunk.startswith("data: ") and not chunk.startswith("data: [DONE]"):
@@ -2401,8 +2494,14 @@ def setup_chat_routes(
                                     agent_tool_calls=_agent_tool_calls,
                                     skills_manager=skills_manager,
                                     owner=_user,
-                                    extract_skills=user_requested_agent,
-                                    allow_background_extraction=not tool_policy.block_all_tool_calls,
+                                    extract_skills=(
+                                        user_requested_agent
+                                        and not tool_approval_continuation
+                                    ),
+                                    allow_background_extraction=(
+                                        not tool_policy.block_all_tool_calls
+                                        and not tool_approval_continuation
+                                    ),
                                 )
                             _stream_set(session, status="done")
                             yield chunk

routes/cookbook_helpers.py

@@ -1204,6 +1204,41 @@ def _safe_env_prefix(ep: str | None) -> str | None:
     return f'[ -f "{path}" ] && source "{path}" || true'
 
 
+def _local_windows_bash_env_prefix(ep: str | None) -> str | None:
+    """Convert a frontend PowerShell venv prefix for the local Git Bash runner."""
+    if not ep:
+        return ep
+
+    prefix = ep.strip()
+    if not prefix.startswith("&"):
+        return ep
+
+    raw_path = prefix[1:].lstrip()
+    if not raw_path:
+        return ep
+    if raw_path.startswith("'"):
+        if len(raw_path) < 2 or not raw_path.endswith("'"):
+            return ep
+        quoted_path = raw_path[1:-1]
+        if "'" in quoted_path.replace("''", ""):
+            return ep
+        path = quoted_path.replace("''", "'")
+    else:
+        path = raw_path.rstrip()
+        if "'" in path or '"' in path:
+            return ep
+    if any(c in path for c in "\r\n;&|`$<>"):
+        return ep
+    if not path.replace("\\", "/").casefold().endswith("/scripts/activate.ps1"):
+        return ep
+
+    bash_path = _git_bash_path(path)
+    if "\\" in bash_path:
+        return ep
+    bash_path = bash_path[: -len("Activate.ps1")] + "activate"
+    return "source " + shlex.quote(bash_path)
+
+
 def _ssh_ps(host, script_path, port=None):
     """Build SSH command to run a PowerShell script on a Windows remote."""
     pf = f"-p {port} " if port and port != "22" else ""

routes/cookbook_routes.py

@@ -50,7 +50,7 @@ from routes.cookbook_helpers import (
     _SESSION_ID_RE, _validate_repo_id, _validate_serve_model_id, _validate_include, _validate_token,
     _validate_local_dir, _validate_gpus, _shell_path,
     _ps_squote, _bash_squote, _validate_serve_cmd, _parse_serve_phase, OLLAMA_MISSING_HINT,
-    _safe_env_prefix, _local_tooling_path_export, _append_serve_preflight_exit_lines,
+    _safe_env_prefix, _local_windows_bash_env_prefix, _local_tooling_path_export, _append_serve_preflight_exit_lines,
     _append_serve_exit_code_lines, _append_llama_cpp_linux_accel_build_lines, _cached_model_scan_script,
     load_stored_hf_token,
     _append_vllm_linux_preflight_lines, _ollama_bind_from_cmd, _pip_install_fallback_chain,
@@ -1336,7 +1336,7 @@ def setup_cookbook_routes() -> APIRouter:
             # Local: run hf download in the background (tmux on POSIX, a detached
             # process + logfile on Windows where tmux doesn't exist).
             if req.env_prefix:
-                lines.append(_safe_env_prefix(req.env_prefix))
+                lines.append(_safe_env_prefix(_local_windows_bash_env_prefix(req.env_prefix) if local_windows else req.env_prefix))
             else:
                 lines.append("deactivate 2>/dev/null; hash -r")
             # Show whether the HF token reached this run (masked) — tells a gated
@@ -2166,7 +2166,7 @@ def setup_cookbook_routes() -> APIRouter:
             if req.gpus:
                 runner_lines.append(f"export CUDA_VISIBLE_DEVICES='{req.gpus}'")
             if req.env_prefix:
-                runner_lines.append(_safe_env_prefix(req.env_prefix))
+                runner_lines.append(_safe_env_prefix(_local_windows_bash_env_prefix(req.env_prefix) if local_windows else req.env_prefix))
             else:
                 runner_lines.append("deactivate 2>/dev/null; hash -r")
             _append_venv_nvidia_library_path_lines(runner_lines, cmd=req.cmd)

routes/mcp/mcp_routes.py

@@ -475,7 +475,7 @@ def setup_mcp_routes(mcp_manager: McpManager):
                 return RedirectResponse(auth_url)
             else:
                 # Remote device — show paste-back page
-                return HTMLResponse(_oauth_authorize_page(auth_url, server_id, host, redirect_uri))
+                return HTMLResponse(_oauth_authorize_page(auth_url, server_id, redirect_uri))
         finally:
             db.close()
 
@@ -612,15 +612,13 @@ def setup_mcp_routes(mcp_manager: McpManager):
 def _oauth_authorize_page(
     auth_url: str,
     server_id: str,
-    host: str,
-    redirect_uri: str = "http://localhost:7000/api/mcp/oauth/callback",
+    redirect_uri: str,
 ) -> str:
     """Page with Google sign-in link and URL paste-back form for remote access."""
-    # Escape values interpolated into the page: `host` comes from the request
-    # Host header and `server_id` from the OAuth state — neither is trusted.
+    # Escape values interpolated into the page: `server_id` comes from the OAuth
+    # state and is not trusted.
     auth_url = html.escape(auth_url, quote=True)
     server_id = html.escape(server_id, quote=True)
-    host = html.escape(host, quote=True)
     redirect_uri = html.escape(redirect_uri, quote=True)
     return f"""<!DOCTYPE html>
 <html><head>
@@ -664,7 +662,15 @@ def _oauth_authorize_page(
   </div>
   <a class="auth-link" href="{auth_url}" target="_blank" rel="noopener">Sign in with Google</a>
   <div class="divider"></div>
-  <form method="POST" action="http://{host}/api/mcp/oauth/exchange/{server_id}">
+  <!-- Relative action: the browser resolves it against the origin this page was
+       served from, so the form follows the user through any proxy without the
+       app having to know the scheme or the host. An absolute http:// action is
+       blocked as mixed content on exactly the HTTPS deployments that need
+       paste-back, and request.url.scheme cannot be trusted to spot them —
+       uvicorn only honours X-Forwarded-Proto from a peer in
+       --forwarded-allow-ips, which defaults to 127.0.0.1 and excludes a proxy
+       arriving over the Docker bridge. -->
+  <form method="POST" action="/api/mcp/oauth/exchange/{server_id}">
     <p>Paste the URL from your browser after signing in:</p>
     <input type="text" name="callback_url" placeholder="{redirect_uri}?code=..." required>
     <br><button type="submit">Connect</button>

routes/model_routes.py

@@ -1351,14 +1351,14 @@ def _legacy_visible_api_models(ep) -> List[str]:
 def _picker_models_for_endpoint(ep, base_url: str, kind: str):
     """Return model IDs that should appear in the picker for an endpoint.
 
-    API providers expose remote inventory from /v1/models. Treat that cache as
-    inventory, not approval: only manually pinned API models should appear in
-    the picker. Local/self-hosted endpoints keep the older hide-list behavior.
+    API providers expose remote inventory from /v1/models. Default to that
+    visible inventory until an explicit pinned-model allow-list is saved.
+    Local/self-hosted endpoints keep the older hide-list behavior.
     """
     pinned = _normalize_model_ids(getattr(ep, "pinned_models", None))
     if _picker_requires_pinning(base_url, kind):
         if not _has_explicit_pinned_models(ep):
-            pinned = _legacy_visible_api_models(ep) if _hidden_model_ids(ep) else []
+            pinned = _legacy_visible_api_models(ep)
         return pinned, pinned
     return _visible_models(
         _cached_model_ids(ep),
@@ -2342,9 +2342,7 @@ def setup_model_routes(model_discovery):
                 else:
                     response.headers["X-Model-Refresh-Status"] = "failed"
                     response.headers["X-Model-Refresh-Warning"] = "Model refresh failed or returned no models; kept cached models."
-            pinned = _normalize_model_ids(getattr(ep, "pinned_models", None))
-            if picker_requires_pinning and not _has_explicit_pinned_models(ep):
-                pinned = _legacy_visible_api_models(ep)
+            _, pinned = _picker_models_for_endpoint(ep, base, kind)
             pinned_set = set(pinned)
             return [
                 {

routes/skills_routes.py

@@ -1603,6 +1603,9 @@ def setup_skills_routes(skills_manager: SkillsManager) -> APIRouter:
             decision=decision,
             owner=user,
             session_id=None,
+            # The button here says "Allow once" and there is no chat to carry a
+            # scope into, so the gate must re-arm behind the sealed action.
+            allow_continuation=False,
         )
 
         if decision == "approve" and exact_approval is None:

routes/task/__init__.py

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

scripts/mlx_image_server.py

@@ -327,7 +327,12 @@ def list_models():
 
 @app.post("/v1/images/generations")
 def generate(req: ImageRequest):
-    model = req.model or _args.model
+    # The served model is the one this process was launched with. `req.model`
+    # is accepted for OpenAI wire compatibility and ignored, matching
+    # scripts/diffusion_server.py: honouring it would let a caller point the
+    # generator at any local directory or Hugging Face repo, and the HiDream
+    # branch runs a python script from inside that directory.
+    model = _args.model
     width, height = _size(req.size)
     out_images = []
     count = max(1, min(int(req.n or 1), 4))
@@ -393,7 +398,7 @@ async def edit_image(
     size: str = Form("1024x1024"),
     response_format: str = Form("b64_json"),
 ):
-    active_model = model or _args.model
+    active_model = _args.model  # pinned; see generate()
     if _is_lama_inpaint(active_model) or _is_ddcolor(active_model):
         image_raw = await image.read()
         mask_raw = await mask.read() if mask is not None else None

scripts/odysseus-webhook

@@ -2,7 +2,7 @@
 """odysseus-webhook — shell wrapper for scheduled-task webhook tokens.
 
 Tasks in the scheduled-task system can carry a `webhook_token`. Any
-HTTP POST to `/api/webhook/<token>` fires the task. This CLI lists,
+HTTP POST to `/api/tasks/<task-id>/webhook/<token>` fires the task. This CLI lists,
 rotates, and revokes those tokens.
 
     odysseus-webhook list                       # tasks that have a token
@@ -21,6 +21,7 @@ quiet_logs()
 
 import argparse, json, logging, os, secrets, sys
 from pathlib import Path
+from urllib.parse import quote
 
 try:
     from core.database import SessionLocal, ScheduledTask
@@ -53,6 +54,14 @@ def _summary(t: "ScheduledTask", reveal: bool = False) -> dict:
     }
 
 
+def _task_webhook_url(base: str, task_id: str, token: str) -> str:
+    """Build the live task-route URL without leaking ids into path syntax."""
+    root = (base or "http://localhost:7000").rstrip("/")
+    task_part = quote(str(task_id), safe="")
+    token_part = quote(str(token), safe="")
+    return f"{root}/api/tasks/{task_part}/webhook/{token_part}"
+
+
 def cmd_list(args):
     db = SessionLocal()
     try:
@@ -109,8 +118,7 @@ def cmd_url(args):
             fail(f"no task with id {args.id!r}")
         if not t.webhook_token:
             fail(f"task {args.id!r} has no webhook token (rotate one first)")
-        base = (args.base or "http://localhost:7000").rstrip("/")
-        url = f"{base}/api/webhook/{t.webhook_token}"
+        url = _task_webhook_url(args.base, t.id, t.webhook_token)
         emit({
             "task_id": t.id,
             "name": t.name,

services/docs/service.py

@@ -50,16 +50,46 @@ class DocsService:
             List of DocChunk objects
         """
         results = self.rag.search(query, k=top_k)
-        return [
-            DocChunk(
-                text=r.get("text", r.get("content", "")),
-                source=r.get("source", r.get("metadata", {}).get("source", "unknown")),
-                score=r.get("score", 0.0),
-                metadata=r.get("metadata"),
+        chunks = []
+
+        for result in results:
+            if not isinstance(result, dict):
+                continue
+
+            metadata = result.get("metadata")
+            if not isinstance(metadata, dict):
+                metadata = {}
+
+            text = result.get("document")
+            if text is None:
+                text = result.get("text")
+            if text is None:
+                text = result.get("content")
+            if text is None:
+                text = ""
+
+            source = result.get("source")
+            if source is None:
+                source = metadata.get("source")
+            if source is None:
+                source = "unknown"
+
+            score = result.get("similarity")
+            if score is None:
+                score = result.get("score")
+            if score is None:
+                score = 0.0
+
+            chunks.append(
+                DocChunk(
+                    text=text,
+                    source=source,
+                    score=score,
+                    metadata=metadata,
+                )
             )
-            for r in results
-            if isinstance(r, dict)
-        ]
+
+        return chunks
 
     async def index(self, directory: str) -> IndexResult:
         """
@@ -73,8 +103,8 @@ class DocsService:
         """
         result = self.rag.index_personal_documents(directory)
         return IndexResult(
-            indexed=result.get("indexed", 0),
-            failed=result.get("failed", 0),
+            indexed=result.get("indexed_count", result.get("indexed", 0)),
+            failed=result.get("failed_count", result.get("failed", 0)),
             errors=result.get("errors", []),
         )
 

services/search/content.py

@@ -2,22 +2,18 @@
 
 import copy
 import io
-import ipaddress
 import json
 import os
 import re
 import logging
-import socket
-import ssl
 from datetime import datetime, timedelta
-from typing import Iterable, List, cast
-from urllib.parse import urljoin, urlparse
+from typing import List
 
 import httpx
-import httpcore
 from bs4 import BeautifulSoup
 
 from src.constants import WEB_FETCH_SOFT_MAX_BYTES, WEB_FETCH_HARD_MAX_BYTES, WEB_FETCH_USER_AGENT
+from src import outbound_fetch as _outbound_fetch
 
 from .analytics import RateLimitError, error_logger
 from .cache import (
@@ -29,336 +25,40 @@ from .cache import (
 
 logger = logging.getLogger(__name__)
 
-_PRIVATE_NETWORKS = (
-    ipaddress.ip_network("0.0.0.0/8"),
-    ipaddress.ip_network("10.0.0.0/8"),
-    ipaddress.ip_network("127.0.0.0/8"),
-    ipaddress.ip_network("169.254.0.0/16"),
-    ipaddress.ip_network("172.16.0.0/12"),
-    ipaddress.ip_network("192.168.0.0/16"),
-    ipaddress.ip_network("::1/128"),
-    ipaddress.ip_network("fc00::/7"),
-    ipaddress.ip_network("fe80::/10"),
-)
+def _is_private_address(addr):
+    return _outbound_fetch._is_private_address(addr)
 
 
-def _is_private_address(addr: ipaddress._BaseAddress) -> bool:
-    if isinstance(addr, ipaddress.IPv6Address) and addr.ipv4_mapped is not None:
-        addr = addr.ipv4_mapped
-    return (
-        addr.is_private
-        or addr.is_loopback
-        or addr.is_link_local
-        or addr.is_reserved
-        or addr.is_multicast
-        or addr.is_unspecified
-        or any(addr in net for net in _PRIVATE_NETWORKS)
+def _resolve_hostname_ips(hostname):
+    return _outbound_fetch._resolve_hostname_ips(hostname)
+
+
+def _public_http_url(url):
+    return _outbound_fetch._public_http_url(url, resolver=_resolve_hostname_ips)
+
+
+def _resolve_public_ips(url):
+    return _outbound_fetch._resolve_public_ips(url, resolver=_resolve_hostname_ips)
+
+
+_PinnedBackend = _outbound_fetch._PinnedBackend
+_PinnedTransport = _outbound_fetch._PinnedTransport
+BodyTooLargeError = _outbound_fetch.BodyTooLargeError
+_CappedFetch = _outbound_fetch._CappedFetch
+
+
+def _get_public_url(url, headers, timeout, max_redirects=5, max_bytes=None):
+    return _outbound_fetch._get_public_url(
+        url,
+        headers=headers,
+        timeout=timeout,
+        max_redirects=max_redirects,
+        max_bytes=max_bytes,
+        resolve_public_ips=_resolve_public_ips,
+        transport_factory=_PinnedTransport,
     )
 
 
-def _resolve_hostname_ips(hostname: str) -> list[ipaddress._BaseAddress]:
-    try:
-        infos = socket.getaddrinfo(hostname, None)
-    except Exception:
-        return []
-    out = []
-    for info in infos:
-        try:
-            out.append(ipaddress.ip_address(info[4][0]))
-        except Exception:
-            continue
-    return out
-
-
-def _public_http_url(url: str) -> bool:
-    try:
-        parsed = urlparse(url)
-        if parsed.scheme not in ("http", "https"):
-            return False
-        host = (parsed.hostname or "").strip()
-        if not host:
-            return False
-        lower = host.lower()
-        if lower in ("localhost", "metadata", "metadata.google.internal"):
-            return False
-        if lower.endswith((".local", ".localhost", ".internal", ".lan", ".intranet")):
-            return False
-        try:
-            return not _is_private_address(ipaddress.ip_address(host))
-        except ValueError:
-            pass
-        addrs = _resolve_hostname_ips(host)
-        return bool(addrs) and not any(_is_private_address(a) for a in addrs)
-    except Exception:
-        return False
-
-
-def _resolve_public_ips(url: str) -> list[ipaddress._BaseAddress]:
-    parsed = urlparse(url)
-    if parsed.scheme not in ("http", "https") or not parsed.hostname:
-        raise httpx.RequestError(f"Blocked non-public URL: {url}")
-    host = (parsed.hostname or "").strip().lower()
-    if host in ("localhost", "metadata", "metadata.google.internal"):
-        raise httpx.RequestError(f"Blocked non-public hostname: {host}")
-    try:
-        ip = ipaddress.ip_address(host)
-        if _is_private_address(ip):
-            raise httpx.RequestError(f"Blocked non-public IP literal: {host}")
-        return [ip]
-    except httpx.RequestError:
-        raise
-    except ValueError:
-        pass
-    addrs = _resolve_hostname_ips(host)
-    if not addrs or any(_is_private_address(a) for a in addrs):
-        raise httpx.RequestError(f"Blocked non-public URL: {url}")
-    return addrs
-
-
-class _PinnedBackend(httpcore.NetworkBackend):
-    """Network backend that connects to a pre-resolved IP.
-
-    httpcore derives the TLS SNI and the ``Host`` header from the URL's
-    origin, not from the host argument passed to ``connect_tcp``. So
-    routing the TCP connect to a resolved IP while leaving the URL
-    untouched keeps SNI / vhost behaviour correct and closes the
-    DNS-rebinding TOCTOU between the SSRF check and the connect.
-    """
-
-    def __init__(self, ip: ipaddress._BaseAddress):
-        self._ip = str(ip)
-        self._real = httpcore.SyncBackend()
-
-    def connect_tcp(
-        self,
-        host: str,
-        port: int,
-        timeout: float | None = None,
-        local_address: str | None = None,
-        socket_options=None,
-    ):
-        return self._real.connect_tcp(
-            self._ip, port, timeout, local_address, socket_options
-        )
-
-    def connect_unix_socket(self, path, timeout=None, socket_options=None):
-        return self._real.connect_unix_socket(path, timeout, socket_options)
-
-    def sleep(self, seconds: float) -> None:
-        return self._real.sleep(seconds)
-
-
-# Map httpcore exception classes to their httpx equivalents. Built
-# once at import time from the public exception classes; avoids any
-# import of httpx's private transport machinery. httpcore's
-# ``ConnectionNotAvailable`` is a pool-internal signal (the pool will
-# close and retry on its own) — we never expect to see it surface to
-# a transport caller, so it has no httpx counterpart here.
-_HTTPCORE_TO_HTTPX_EXC = {
-    httpcore.ConnectError: httpx.ConnectError,
-    httpcore.ConnectTimeout: httpx.ConnectTimeout,
-    httpcore.LocalProtocolError: httpx.LocalProtocolError,
-    httpcore.NetworkError: httpx.NetworkError,
-    httpcore.PoolTimeout: httpx.PoolTimeout,
-    httpcore.ProtocolError: httpx.ProtocolError,
-    httpcore.ProxyError: httpx.ProxyError,
-    httpcore.ReadError: httpx.ReadError,
-    httpcore.ReadTimeout: httpx.ReadTimeout,
-    httpcore.RemoteProtocolError: httpx.RemoteProtocolError,
-    httpcore.TimeoutException: httpx.TimeoutException,
-    httpcore.UnsupportedProtocol: httpx.UnsupportedProtocol,
-    httpcore.WriteError: httpx.WriteError,
-    httpcore.WriteTimeout: httpx.WriteTimeout,
-}
-
-
-class _PinnedTransport(httpx.BaseTransport):
-    """Transport that pins every TCP connect to a pre-resolved IP.
-
-    Uses only the public ``httpcore`` and ``httpx`` APIs — no
-    subclassing of ``httpx.HTTPTransport``, no reads of private
-    ``httpcore.ConnectionPool`` attributes, no imports from
-    ``httpx private transport internals``. The URL is passed through unchanged so SNI
-    / vhost work as if httpx had been given the hostname directly;
-    only the TCP destination is pinned, closing the DNS-rebinding
-    TOCTOU between the SSRF check and the connect.
-    """
-
-    def __init__(self, ip: ipaddress._BaseAddress, *, http2: bool = False):
-        self._pool = httpcore.ConnectionPool(
-            ssl_context=ssl.create_default_context(),
-            http1=True,
-            http2=http2,
-            network_backend=_PinnedBackend(ip),
-        )
-
-    def __enter__(self):
-        self._pool.__enter__()
-        return self
-
-    def __exit__(self, exc_type=None, exc_value=None, traceback=None) -> None:
-        self._pool.__exit__(exc_type, exc_value, traceback)
-
-    def handle_request(self, request: httpx.Request) -> httpx.Response:
-        httpcore_req = httpcore.Request(
-            method=request.method,
-            url=httpcore.URL(
-                scheme=request.url.raw_scheme,
-                host=request.url.raw_host,
-                port=request.url.port,
-                target=request.url.raw_path,
-            ),
-            headers=request.headers.raw,
-            content=request.stream,
-            extensions=request.extensions,
-        )
-        try:
-            httpcore_resp = self._pool.handle_request(httpcore_req)
-            # Eager materialisation matches the original
-            # ``response.text`` usage in fetch_webpage_content. The
-            # sync pool's stream is a plain Iterable[bytes] despite
-            # the httpcore type hint unioning the async variant.
-            content = b"".join(cast(Iterable[bytes], httpcore_resp.stream))
-        except Exception as exc:
-            mapped = _HTTPCORE_TO_HTTPX_EXC.get(type(exc))
-            if mapped is not None:
-                raise mapped(str(exc)) from exc
-            raise
-
-        return httpx.Response(
-            status_code=httpcore_resp.status,
-            headers=httpcore_resp.headers,
-            content=content,
-            extensions=httpcore_resp.extensions,
-        )
-
-    def close(self) -> None:
-        self._pool.close()
-
-class BodyTooLargeError(Exception):
-    """The server declared a body larger than the hard fetch ceiling."""
-
-    def __init__(self, url: str, declared_bytes: int):
-        self.url = url
-        self.declared_bytes = declared_bytes
-        super().__init__(
-            f"response body is {declared_bytes:,} bytes, over the "
-            f"{WEB_FETCH_HARD_MAX_BYTES:,}-byte hard cap"
-        )
-
-
-class _CappedFetch:
-    """Result of a size-capped streaming GET.
-
-    Carries just what fetch_webpage_content needs from an httpx.Response,
-    plus the cap bookkeeping: the (possibly truncated) body, whether the
-    cap cut it short, and the size the server declared via Content-Length
-    (wire bytes; None when absent).
-    """
-
-    __slots__ = ("status_code", "headers", "content", "truncated",
-                 "declared_bytes", "encoding", "url")
-
-    def __init__(self, status_code, headers, content, truncated,
-                 declared_bytes, encoding, url):
-        self.status_code = status_code
-        self.headers = headers
-        self.content = content
-        self.truncated = truncated
-        self.declared_bytes = declared_bytes
-        self.encoding = encoding
-        self.url = url
-
-    @property
-    def text(self) -> str:
-        return self.content.decode(self.encoding or "utf-8", errors="replace")
-
-    def raise_for_status(self):
-        if self.status_code >= 400:
-            request = httpx.Request("GET", self.url)
-            raise httpx.HTTPStatusError(
-                f"HTTP {self.status_code} for {self.url}",
-                request=request,
-                response=httpx.Response(self.status_code, request=request),
-            )
-
-
-def _get_public_url(url: str, headers: dict, timeout: int, max_redirects: int = 5,
-                    max_bytes: int = None) -> "_CappedFetch":
-    """Capped streaming GET with SSRF-guarded, DNS-pinned manual redirects.
-
-    Each hop is resolved once, validated as public, and then the actual TCP
-    connection is pinned to that resolved IP. The request URL is left unchanged
-    so Host and TLS SNI keep the original hostname.
-    """
-    cap = min(max_bytes or WEB_FETCH_SOFT_MAX_BYTES, WEB_FETCH_HARD_MAX_BYTES)
-    current = url
-    for _ in range(max_redirects + 1):
-        ips = _resolve_public_ips(current)
-
-        # Force identity transfer-encoding. With gzip/deflate the wire bytes
-        # and Content-Length can be a small fraction of the decoded body, so a
-        # tiny compressed response could pass the hard-cap preflight and then
-        # expand past the ceiling in one decoded chunk before the streamed cap
-        # below can slice it.
-        req_headers = dict(headers or {})
-        req_headers["Accept-Encoding"] = "identity"
-
-        with httpx.Client(
-            headers=req_headers,
-            timeout=timeout,
-            follow_redirects=False,
-            transport=_PinnedTransport(ips[0]),
-        ) as client:
-            with client.stream("GET", current) as response:
-                if response.status_code in (301, 302, 303, 307, 308):
-                    location = response.headers.get("location")
-                    if not location:
-                        return _CappedFetch(response.status_code, response.headers, b"",
-                                            False, None, response.encoding, str(response.url))
-                    current = urljoin(str(response.url), location)
-                    continue
-
-                # A server can ignore the identity request and still return a
-                # compressed body; httpx.iter_bytes would then decode it, and a
-                # tiny gzip can balloon into one decoded chunk far past the cap.
-                # Refuse compressed Content-Encoding so the streamed cap stays
-                # a real memory bound.
-                enc = (response.headers.get("content-encoding") or "").strip().lower()
-                if enc and enc != "identity":
-                    raise httpx.RequestError(
-                        f"Refusing compressed response (Content-Encoding: {enc}) after "
-                        "requesting identity: cannot bound decoded body size",
-                        request=httpx.Request("GET", current),
-                    )
-
-                declared = None
-                raw_len = response.headers.get("content-length")
-                if raw_len and raw_len.isdigit():
-                    declared = int(raw_len)
-
-                if declared is not None and declared > WEB_FETCH_HARD_MAX_BYTES:
-                    raise BodyTooLargeError(current, declared)
-
-                chunks = []
-                read = 0
-                truncated = False
-                for chunk in response.iter_bytes():
-                    read += len(chunk)
-                    if read > cap:
-                        keep = cap - (read - len(chunk))
-                        if keep > 0:
-                            chunks.append(chunk[:keep])
-                        truncated = True
-                        break
-                    chunks.append(chunk)
-
-                return _CappedFetch(response.status_code, response.headers,
-                                    b"".join(chunks), truncated, declared,
-                                    response.encoding, str(response.url))
-
-    raise httpx.RequestError("Too many redirects", request=httpx.Request("GET", current))
-
 # PDF extraction (optional dependency)
 try:
     from pdfminer.high_level import extract_text as pdf_extract_text

src/agent_loop.py

@@ -3081,10 +3081,17 @@ def _append_tool_results(
             messages.append(result_message)
     else:
         tool_output_text = "\n\n".join(tool_results)
-        msg = {"role": "assistant", "content": round_response}
-        if round_reasoning:
-            msg["reasoning_content"] = round_reasoning
-        messages.append(msg)
+        # An approved-action replay injects the sealed tool result with no
+        # assistant prose for that round, which used to append an assistant turn
+        # whose content was "". Anthropic's Messages API rejects a non-final
+        # assistant message with empty content (HTTP 400), so the resumed turn
+        # died before the model saw the result. A turn carrying neither prose nor
+        # reasoning has nothing to say to any provider, so skip it entirely.
+        if round_response.strip() or round_reasoning:
+            msg = {"role": "assistant", "content": round_response}
+            if round_reasoning:
+                msg["reasoning_content"] = round_reasoning
+            messages.append(msg)
         # Tool output (shell/python stdout, file reads, fetched pages, email
         # bodies, MCP results) is sourced from outside the server. Wrap it as
         # untrusted data so prompt-injection inside a tool result is treated as
@@ -3460,7 +3467,10 @@ async def stream_agent_loop(
                 and exact_approval.pending.external_untrusted_context_seen
             )
             or messages_contain_external_untrusted_context(messages)
-        )
+        ),
+        approval_gate_bypassed=bool(
+            exact_approval and exact_approval.allow_remaining_actions
+        ),
     )
     mcp_mgr = get_mcp_manager()
     prep_timings: Dict[str, float] = {}
@@ -5698,6 +5708,16 @@ async def stream_agent_loop(
                         "policy": "exact_tool_approval_target",
                     }
                 else:
+                    # The approval click becomes a synthetic user turn. Seal the
+                    # actual server-selected candidates now so that continuation
+                    # does not lose memory, skills, MCP, documents, or other
+                    # ToolIndex/RAG-selected tools by classifying that synthetic text.
+                    approval_selected_tools = set(_relevant_tools or ())
+                    approval_selected_tools.update(
+                        name for name in _tool_names_sent if name
+                    )
+                    approval_selected_tools.add(block.tool_type)
+                    approval_selected_tools.difference_update(disabled_tools)
                     pending_approval = tool_approval_store.create(
                         owner=owner,
                         session_id=session_id,
@@ -5725,6 +5745,8 @@ async def stream_agent_loop(
                         external_untrusted_context_seen=(
                             run_security.external_untrusted_context_seen
                         ),
+                        selected_tools=approval_selected_tools,
+                        continuation_query=_retrieval_query or _last_user,
                         capabilities=capabilities_for_action(
                             block.tool_type,
                             block.content,

src/builtin_actions.py

@@ -810,13 +810,27 @@ async def action_tidy_research(owner: str, **kwargs) -> Tuple[str, bool]:
 
     Research history lives entirely in data/deep_research/<id>.json and is NOT
     backed by chat-session rows — so a file must never be deleted just because
-    no chat session matches its id. Only prune files that fail to load."""
+    no chat session matches its id. Only prune files that fail to load.
+
+    A broken file has no readable owner stamp, so it cannot be matched against
+    `owner`. Clearing one is privileged: admins and the single-user operator
+    (AUTH_ENABLED=false) may, a regular user may not, and neither may anyone
+    during the pre-setup window before an admin exists.
+    """
     try:
         from pathlib import Path
         import json as _json
+        from src.tool_security import owner_is_admin_or_single_user
         research_dir = Path(DEEP_RESEARCH_DIR)
         if not research_dir.exists():
             raise TaskNoop("no research directory")
+        if not owner_is_admin_or_single_user(owner):
+            # Return before the glob rather than filtering inside the loop: the
+            # loop reports "none broken" off an empty `removed`, which reaches
+            # Activity as a false report to a user whose files it skipped, and a
+            # regular user need not read every owner's file to learn it may
+            # delete none of them.
+            raise TaskNoop("not permitted to remove unattributable research files")
         files = list(research_dir.glob("*.json"))
         removed = []
         for p in files:

src/chat_processor.py

@@ -462,7 +462,14 @@ class ChatProcessor:
         skip_url_fetch = len(message) > 2000 or len(non_yt_urls) > 3
         if not skip_url_fetch:
             for url in non_yt_urls:
-                result = fetch_webpage_content(url)
+                try:
+                    result = fetch_webpage_content(url)
+                except Exception:
+                    # The URL and exception can both contain signed-query
+                    # credentials or response-controlled text. Keep the log
+                    # diagnostic stable as well as the model-facing context.
+                    logger.warning("Automatic URL fetch failed while building context")
+                    result = {"success": False, "error": ""}
                 if result.get('success'):
                     content = result.get('content', '')[:10000]
                     preface.append(untrusted_context_message(
@@ -470,6 +477,24 @@ class ChatProcessor:
                         f"Content from {url}:\n\n{content}",
                         provenance_origin="external",
                     ))
+                else:
+                    # A failed automatic URL fetch is context too. Never pass
+                    # exception text or response-controlled diagnostics back to
+                    # the model: reduce the result to a small transport-owned
+                    # status and explicitly state that the page was not read.
+                    error = str(result.get("error") or "")
+                    status = "the page was unavailable"
+                    status_match = re.match(r"^HTTP\s+(\d{3})\b", error)
+                    if status_match:
+                        status = f"the server returned HTTP {status_match.group(1)}"
+                    elif error.startswith("TooLarge:"):
+                        status = "the response exceeded the fetch size limit"
+                    elif error.startswith("Rate limit"):
+                        status = "the request was rate limited"
+                    preface.append(untrusted_context_message(
+                        "web page fetch failure",
+                        f"A linked page was not read: {status}.",
+                    ))
 
         # Skills index — progressive disclosure. Only injected when the
         # model has the `manage_skills` tool available (agent_mode), and

src/constants.py

@@ -4,7 +4,7 @@ import os
 
 from src.runtime_paths import get_app_root, get_default_data_dir
 
-APP_VERSION = "1.0.2"
+APP_VERSION = "1.0.3"
 
 # Base paths
 BASE_DIR = os.path.join(get_app_root(), "")

src/mcp_oauth.py

@@ -15,18 +15,32 @@ from urllib.parse import urlparse, parse_qs
 
 logger = logging.getLogger(__name__)
 
+
+def _resolve_redirect_base() -> str:
+    """Origin the browser is sent back to after authorizing.
+
+    Falls back to the port the app binds natively (APP_PORT, read the same way
+    by app.py and launcher.py) rather than a fixed 7000: the macOS launcher
+    defaults to 7860, and a callback on the wrong port reaches nothing. The
+    hostname stays `localhost` rather than internal_api_base()'s 127.0.0.1 —
+    this URI is registered with the authorization server (via DCR, or by hand
+    for Google clients), so changing the host invalidates registrations that
+    already exist.
+    """
+    return (
+        os.environ.get("OAUTH_REDIRECT_BASE_URL")
+        or os.environ.get("APP_PUBLIC_URL")
+        or f"http://localhost:{os.environ.get('APP_PORT', '7000')}"
+    ).rstrip("/")
+
+
 # OAuth redirect URI registered with every authorization server via DCR. Loopback
 # is allowed for native/desktop clients (RFC 8252); remote users finish via the
-# paste-back flow. Deployments not reachable at http://localhost:7000 (custom
-# port, reverse proxy, or public domain) must set OAUTH_REDIRECT_BASE_URL (or
-# APP_PUBLIC_URL) to their externally reachable origin so the redirect lands back
-# on Odysseus. APP_PORT is intentionally not used: it is only the Docker host
-# port-map; the app always listens on 7000 inside the container.
-_REDIRECT_BASE = (
-    os.environ.get("OAUTH_REDIRECT_BASE_URL")
-    or os.environ.get("APP_PUBLIC_URL")
-    or "http://localhost:7000"
-).rstrip("/")
+# paste-back flow. Deployments whose externally reachable origin differs from the
+# port Odysseus binds — reverse proxy, public domain, or Docker, whose host port
+# map is invisible inside the container — must set OAUTH_REDIRECT_BASE_URL (or
+# APP_PUBLIC_URL), otherwise the redirect never lands back on Odysseus.
+_REDIRECT_BASE = _resolve_redirect_base()
 REDIRECT_URI = f"{_REDIRECT_BASE}/api/mcp/oauth/callback"
 
 # How long the background connect waits for the user to authorize before giving up.

src/model_capability_readers/base.py

@@ -290,17 +290,22 @@ def detect_vendor(base_url: Any = "", endpoint_kind: Any = "") -> str:
         return kind_map[kind]
 
     parsed = urlparse(compact_str(base_url))
-    host = (parsed.hostname or "").lower()
+    host = (parsed.hostname or "").lower().rstrip(".")
     port = parsed.port
-    if host.endswith("openrouter.ai"):
+
+    def host_matches(domain: str) -> bool:
+        domain = domain.lower().rstrip(".")
+        return host == domain or host.endswith(f".{domain}")
+
+    if host_matches("openrouter.ai"):
         return VENDOR_OPENROUTER
-    if host.endswith("openai.com"):
+    if host_matches("openai.com"):
         return VENDOR_OPENAI
-    if host.endswith("anthropic.com"):
+    if host_matches("anthropic.com"):
         return VENDOR_ANTHROPIC
-    if host.endswith("googleapis.com"):
+    if host_matches("googleapis.com"):
         return VENDOR_GOOGLE
-    if host.endswith("ollama.com") or port == 11434:
+    if host_matches("ollama.com") or port == 11434:
         return VENDOR_OLLAMA
     if port == 1234:
         return VENDOR_LMSTUDIO

src/outbound_fetch.py

@@ -0,0 +1,354 @@
+"""SSRF-guarded synchronous HTTP fetching primitives.
+
+This module owns outbound URL classification, one-resolution-per-hop DNS
+pinning, redirects, and response-body budgets.  It deliberately has no search
+or content-extraction dependencies so callers outside search can reuse the
+same transport boundary.
+"""
+
+from __future__ import annotations
+
+import ipaddress
+import socket
+import ssl
+from typing import Callable, Iterable, cast
+from urllib.parse import urljoin, urlparse
+
+import httpcore
+import httpx
+
+from src.constants import WEB_FETCH_HARD_MAX_BYTES, WEB_FETCH_SOFT_MAX_BYTES
+
+
+_PRIVATE_NETWORKS = (
+    ipaddress.ip_network("0.0.0.0/8"),
+    ipaddress.ip_network("10.0.0.0/8"),
+    ipaddress.ip_network("127.0.0.0/8"),
+    ipaddress.ip_network("169.254.0.0/16"),
+    ipaddress.ip_network("172.16.0.0/12"),
+    ipaddress.ip_network("192.168.0.0/16"),
+    ipaddress.ip_network("::1/128"),
+    ipaddress.ip_network("fc00::/7"),
+    ipaddress.ip_network("fe80::/10"),
+)
+
+
+def _is_private_address(addr: ipaddress._BaseAddress) -> bool:
+    if isinstance(addr, ipaddress.IPv6Address) and addr.ipv4_mapped is not None:
+        addr = addr.ipv4_mapped
+    return (
+        addr.is_private
+        or addr.is_loopback
+        or addr.is_link_local
+        or addr.is_reserved
+        or addr.is_multicast
+        or addr.is_unspecified
+        or any(addr in net for net in _PRIVATE_NETWORKS)
+    )
+
+
+def _resolve_hostname_ips(hostname: str) -> list[ipaddress._BaseAddress]:
+    try:
+        infos = socket.getaddrinfo(hostname, None)
+    except Exception:
+        return []
+    out = []
+    for info in infos:
+        try:
+            out.append(ipaddress.ip_address(info[4][0]))
+        except Exception:
+            continue
+    return out
+
+
+def _public_http_url(
+    url: str,
+    *,
+    resolver: Callable[[str], list[ipaddress._BaseAddress]] | None = None,
+) -> bool:
+    resolver = resolver or _resolve_hostname_ips
+    try:
+        parsed = urlparse(url)
+        if parsed.scheme not in ("http", "https"):
+            return False
+        host = (parsed.hostname or "").strip()
+        if not host:
+            return False
+        lower = host.lower()
+        if lower in ("localhost", "metadata", "metadata.google.internal"):
+            return False
+        if lower.endswith((".local", ".localhost", ".internal", ".lan", ".intranet")):
+            return False
+        try:
+            return not _is_private_address(ipaddress.ip_address(host))
+        except ValueError:
+            pass
+        addrs = resolver(host)
+        return bool(addrs) and not any(_is_private_address(a) for a in addrs)
+    except Exception:
+        return False
+
+
+def _resolve_public_ips(
+    url: str,
+    *,
+    resolver: Callable[[str], list[ipaddress._BaseAddress]] | None = None,
+) -> list[ipaddress._BaseAddress]:
+    resolver = resolver or _resolve_hostname_ips
+    parsed = urlparse(url)
+    if parsed.scheme not in ("http", "https") or not parsed.hostname:
+        raise httpx.RequestError(f"Blocked non-public URL: {url}")
+    host = (parsed.hostname or "").strip().lower()
+    if host in ("localhost", "metadata", "metadata.google.internal"):
+        raise httpx.RequestError(f"Blocked non-public hostname: {host}")
+    try:
+        ip = ipaddress.ip_address(host)
+        if _is_private_address(ip):
+            raise httpx.RequestError(f"Blocked non-public IP literal: {host}")
+        return [ip]
+    except httpx.RequestError:
+        raise
+    except ValueError:
+        pass
+    addrs = resolver(host)
+    if not addrs or any(_is_private_address(a) for a in addrs):
+        raise httpx.RequestError(f"Blocked non-public URL: {url}")
+    return addrs
+
+
+class _PinnedBackend(httpcore.NetworkBackend):
+    """Network backend that connects to a pre-resolved IP."""
+
+    def __init__(self, ip: ipaddress._BaseAddress):
+        self._ip = str(ip)
+        self._real = httpcore.SyncBackend()
+
+    def connect_tcp(
+        self,
+        host: str,
+        port: int,
+        timeout: float | None = None,
+        local_address: str | None = None,
+        socket_options=None,
+    ):
+        return self._real.connect_tcp(
+            self._ip, port, timeout, local_address, socket_options
+        )
+
+    def connect_unix_socket(self, path, timeout=None, socket_options=None):
+        return self._real.connect_unix_socket(path, timeout, socket_options)
+
+    def sleep(self, seconds: float) -> None:
+        return self._real.sleep(seconds)
+
+
+_HTTPCORE_TO_HTTPX_EXC = {
+    httpcore.ConnectError: httpx.ConnectError,
+    httpcore.ConnectTimeout: httpx.ConnectTimeout,
+    httpcore.LocalProtocolError: httpx.LocalProtocolError,
+    httpcore.NetworkError: httpx.NetworkError,
+    httpcore.PoolTimeout: httpx.PoolTimeout,
+    httpcore.ProtocolError: httpx.ProtocolError,
+    httpcore.ProxyError: httpx.ProxyError,
+    httpcore.ReadError: httpx.ReadError,
+    httpcore.ReadTimeout: httpx.ReadTimeout,
+    httpcore.RemoteProtocolError: httpx.RemoteProtocolError,
+    httpcore.TimeoutException: httpx.TimeoutException,
+    httpcore.UnsupportedProtocol: httpx.UnsupportedProtocol,
+    httpcore.WriteError: httpx.WriteError,
+    httpcore.WriteTimeout: httpx.WriteTimeout,
+}
+
+
+class _PinnedTransport(httpx.BaseTransport):
+    """Transport that pins every TCP connect to a pre-resolved IP."""
+
+    def __init__(self, ip: ipaddress._BaseAddress, *, http2: bool = False):
+        self._pool = httpcore.ConnectionPool(
+            ssl_context=ssl.create_default_context(),
+            http1=True,
+            http2=http2,
+            network_backend=_PinnedBackend(ip),
+        )
+
+    def __enter__(self):
+        self._pool.__enter__()
+        return self
+
+    def __exit__(self, exc_type=None, exc_value=None, traceback=None) -> None:
+        self._pool.__exit__(exc_type, exc_value, traceback)
+
+    def handle_request(self, request: httpx.Request) -> httpx.Response:
+        httpcore_req = httpcore.Request(
+            method=request.method,
+            url=httpcore.URL(
+                scheme=request.url.raw_scheme,
+                host=request.url.raw_host,
+                port=request.url.port,
+                target=request.url.raw_path,
+            ),
+            headers=request.headers.raw,
+            content=request.stream,
+            extensions=request.extensions,
+        )
+        try:
+            httpcore_resp = self._pool.handle_request(httpcore_req)
+            content = b"".join(cast(Iterable[bytes], httpcore_resp.stream))
+        except Exception as exc:
+            mapped = _HTTPCORE_TO_HTTPX_EXC.get(type(exc))
+            if mapped is not None:
+                raise mapped(str(exc)) from exc
+            raise
+
+        return httpx.Response(
+            status_code=httpcore_resp.status,
+            headers=httpcore_resp.headers,
+            content=content,
+            extensions=httpcore_resp.extensions,
+        )
+
+    def close(self) -> None:
+        self._pool.close()
+
+
+class BodyTooLargeError(Exception):
+    """The server declared a body larger than the hard fetch ceiling."""
+
+    def __init__(self, url: str, declared_bytes: int):
+        self.url = url
+        self.declared_bytes = declared_bytes
+        super().__init__(
+            f"response body is {declared_bytes:,} bytes, over the "
+            f"{WEB_FETCH_HARD_MAX_BYTES:,}-byte hard cap"
+        )
+
+
+class _CappedFetch:
+    """Result of a size-capped streaming GET."""
+
+    __slots__ = (
+        "status_code",
+        "headers",
+        "content",
+        "truncated",
+        "declared_bytes",
+        "encoding",
+        "url",
+    )
+
+    def __init__(
+        self,
+        status_code,
+        headers,
+        content,
+        truncated,
+        declared_bytes,
+        encoding,
+        url,
+    ):
+        self.status_code = status_code
+        self.headers = headers
+        self.content = content
+        self.truncated = truncated
+        self.declared_bytes = declared_bytes
+        self.encoding = encoding
+        self.url = url
+
+    @property
+    def text(self) -> str:
+        return self.content.decode(self.encoding or "utf-8", errors="replace")
+
+    def raise_for_status(self):
+        if self.status_code >= 400:
+            request = httpx.Request("GET", self.url)
+            raise httpx.HTTPStatusError(
+                f"HTTP {self.status_code} for {self.url}",
+                request=request,
+                response=httpx.Response(self.status_code, request=request),
+            )
+
+
+def _get_public_url(
+    url: str,
+    headers: dict,
+    timeout: int,
+    max_redirects: int = 5,
+    max_bytes: int | None = None,
+    *,
+    resolve_public_ips: Callable[[str], list[ipaddress._BaseAddress]] | None = None,
+    transport_factory: Callable[[ipaddress._BaseAddress], httpx.BaseTransport] | None = None,
+) -> _CappedFetch:
+    """Capped streaming GET with SSRF-guarded, DNS-pinned redirects."""
+    resolve_public_ips = resolve_public_ips or _resolve_public_ips
+    transport_factory = transport_factory or _PinnedTransport
+    cap = min(max_bytes or WEB_FETCH_SOFT_MAX_BYTES, WEB_FETCH_HARD_MAX_BYTES)
+    current = url
+    for _ in range(max_redirects + 1):
+        ips = resolve_public_ips(current)
+        req_headers = dict(headers or {})
+        req_headers["Accept-Encoding"] = "identity"
+
+        with httpx.Client(
+            headers=req_headers,
+            timeout=timeout,
+            follow_redirects=False,
+            transport=transport_factory(ips[0]),
+        ) as client:
+            with client.stream("GET", current) as response:
+                if response.status_code in (301, 302, 303, 307, 308):
+                    location = response.headers.get("location")
+                    if not location:
+                        return _CappedFetch(
+                            response.status_code,
+                            response.headers,
+                            b"",
+                            False,
+                            None,
+                            response.encoding,
+                            str(response.url),
+                        )
+                    current = urljoin(str(response.url), location)
+                    continue
+
+                enc = (response.headers.get("content-encoding") or "").strip().lower()
+                if enc and enc != "identity":
+                    raise httpx.RequestError(
+                        f"Refusing compressed response (Content-Encoding: {enc}) after "
+                        "requesting identity: cannot bound decoded body size",
+                        request=httpx.Request("GET", current),
+                    )
+
+                declared = None
+                raw_len = response.headers.get("content-length")
+                if raw_len and raw_len.isdigit():
+                    declared = int(raw_len)
+
+                if declared is not None and declared > WEB_FETCH_HARD_MAX_BYTES:
+                    raise BodyTooLargeError(current, declared)
+
+                chunks = []
+                read = 0
+                truncated = False
+                for chunk in response.iter_bytes():
+                    read += len(chunk)
+                    if read > cap:
+                        keep = cap - (read - len(chunk))
+                        if keep > 0:
+                            chunks.append(chunk[:keep])
+                        truncated = True
+                        break
+                    chunks.append(chunk)
+
+                return _CappedFetch(
+                    response.status_code,
+                    response.headers,
+                    b"".join(chunks),
+                    truncated,
+                    declared,
+                    response.encoding,
+                    str(response.url),
+                )
+
+    raise httpx.RequestError(
+        "Too many redirects", request=httpx.Request("GET", current)
+    )

src/tool_approval_scopes.py

@@ -0,0 +1,35 @@
+"""Shared wire values and scope markers for tool approval continuations."""
+
+from __future__ import annotations
+
+from enum import Enum
+
+
+# Keep the existing wire values so the current route and no-build frontend do
+# not need a second protocol migration. ``approve`` no longer means one action;
+# it now selects chat-session scope.
+TASK_APPROVAL_DECISION = "approve_task"
+CHAT_SESSION_APPROVAL_DECISION = "approve"
+DENY_APPROVAL_DECISION = "deny"
+
+# Session.get_context_messages() adds this server-owned marker only when the
+# session history contains a matching, resolved chat-session approval.
+CHAT_SESSION_APPROVAL_CONTEXT_MARKER = "_tool_approval_chat_session_granted"
+
+
+class ToolApprovalScope(str, Enum):
+    # Surfaces without a resumable chat (the skill tester, unattended audits)
+    # keep the original one-use meaning: the sealed action runs and the gate
+    # re-arms immediately for anything after it.
+    SINGLE_ACTION = "single_action"
+    TASK = "task"
+    CHAT_SESSION = "chat_session"
+
+
+def scope_for_decision(decision: object) -> ToolApprovalScope | None:
+    normalized = str(decision or "").strip().lower()
+    if normalized == TASK_APPROVAL_DECISION:
+        return ToolApprovalScope.TASK
+    if normalized == CHAT_SESSION_APPROVAL_DECISION:
+        return ToolApprovalScope.CHAT_SESSION
+    return None

src/tool_approvals.py

@@ -1,8 +1,9 @@
-"""Opaque, exact, one-use approvals for tainted model-requested actions.
+"""Opaque exact-action approvals with explicit task and chat scopes.
 
-The model may propose an action after untrusted context, but only the server
-stores and later executes the exact approved tool input.  Browser-visible
-fields are display copies, never authority.
+The server still seals and claims the first displayed action exactly once. The
+selected scope then bypasses only the automatic post-external-context approval
+gate for the rest of the resumed task or chat session. Browser-visible fields
+are display copies, never authority.
 """
 
 from __future__ import annotations
@@ -16,6 +17,13 @@ import time
 from dataclasses import dataclass, field
 from typing import Any
 
+from src.tool_approval_scopes import (
+    CHAT_SESSION_APPROVAL_DECISION,
+    DENY_APPROVAL_DECISION,
+    TASK_APPROVAL_DECISION,
+    ToolApprovalScope,
+    scope_for_decision,
+)
 from src.tool_capabilities import ToolCapabilities, capabilities_for_action
 
 
@@ -33,6 +41,51 @@ def _normalized_workspace(workspace: Any) -> str:
     return os.path.realpath(os.path.expanduser(workspace))
 
 
+_MAX_APPROVAL_SELECTED_TOOLS = 512
+_MAX_APPROVAL_TOOL_NAME_CHARS = 512
+_MAX_APPROVAL_CONTINUATION_QUERY_CHARS = 4000
+
+
+def _normalized_selected_tools(
+    selected_tools: Any,
+    *,
+    required_tool: Any = None,
+) -> tuple[str, ...]:
+    if isinstance(selected_tools, str):
+        selected_tools = (selected_tools,)
+    try:
+        values = selected_tools or ()
+        names = {
+            name.strip()
+            for name in values
+            if (
+                isinstance(name, str)
+                and name.strip()
+                and len(name.strip()) <= _MAX_APPROVAL_TOOL_NAME_CHARS
+            )
+        }
+        required_name = str(required_tool or "").strip()
+        if required_name and len(required_name) <= _MAX_APPROVAL_TOOL_NAME_CHARS:
+            names.add(required_name)
+        ordered = sorted(names)
+        if len(ordered) <= _MAX_APPROVAL_SELECTED_TOOLS:
+            return tuple(ordered)
+        kept = ordered[:_MAX_APPROVAL_SELECTED_TOOLS]
+        if required_name and required_name in names and required_name not in kept:
+            kept[-1] = required_name
+            kept.sort()
+        return tuple(kept)
+    except TypeError:
+        return ()
+
+
+def _normalized_continuation_query(value: Any) -> str:
+    # The query is server-derived from the interrupted run and already lives in
+    # session history. Keep the pending copy bounded because approvals are held
+    # in memory until consumed or expired.
+    return str(value or "").strip()[:_MAX_APPROVAL_CONTINUATION_QUERY_CHARS]
+
+
 def _canonical_digest(payload: dict[str, Any]) -> str:
     encoded = json.dumps(
         payload,
@@ -60,6 +113,8 @@ def _binding_payload(
     document_version: Any,
     document_digest: Any,
     external_untrusted_context_seen: bool,
+    selected_tools: Any,
+    continuation_query: Any,
     effects: tuple[str, ...],
     result_integrity: str,
 ) -> dict[str, Any]:
@@ -76,6 +131,10 @@ def _binding_payload(
         ),
         "document_digest": str(document_digest or "").strip().lower(),
         "external_untrusted_context_seen": bool(external_untrusted_context_seen),
+        "selected_tools": list(
+            _normalized_selected_tools(selected_tools, required_tool=tool_name)
+        ),
+        "continuation_query": _normalized_continuation_query(continuation_query),
         "effects": list(effects),
         "result_integrity": str(result_integrity),
     }
@@ -99,26 +158,47 @@ class PendingToolApproval:
     digest: str
     created_at: float
     expires_at: float
+    # Server-only continuation state. Both fields are digest-bound and never
+    # exposed in the browser payload.
+    selected_tools: tuple[str, ...] = ()
+    continuation_query: str = ""
 
     def public_payload(self, *, reason: str | None = None) -> dict[str, Any]:
         return {
             "kind": "tool_approval",
             "approval_id": self.approval_id,
-            "question": "Allow this exact action once?",
+            # The browser already owns this chat id. Persisting it with the
+            # resolved card lets history-derived session grants remain bound to
+            # this exact chat and prevents inheritance by a forked session.
+            "session_id": self.session_id,
+            "question": "Allow this task to continue?",
             "description": reason or (
-                "Untrusted context influenced this run, so this action needs "
-                "your explicit approval."
+                "Untrusted context influenced this run, so continuing with "
+                "otherwise-gated actions needs your explicit approval."
             ),
             "options": [
                 {
-                    "label": "Allow once",
-                    "value": "approve",
-                    "description": "Execute only the sealed action shown here.",
+                    "label": "Allow for this task",
+                    "value": TASK_APPROVAL_DECISION,
+                    "description": (
+                        "Execute the sealed action and allow every otherwise-gated "
+                        "action needed to finish this request. Current tool, account, "
+                        "workspace, and sandbox restrictions still apply."
+                    ),
+                },
+                {
+                    "label": "Allow for this chat session",
+                    "value": CHAT_SESSION_APPROVAL_DECISION,
+                    "description": (
+                        "Execute the sealed action and stop asking at this gate for "
+                        "later requests in this chat. Current tool, account, workspace, "
+                        "and sandbox restrictions still apply."
+                    ),
                 },
                 {
                     "label": "Deny",
-                    "value": "deny",
-                    "description": "Do not execute it.",
+                    "value": DENY_APPROVAL_DECISION,
+                    "description": "Do not execute the proposed action.",
                 },
             ],
             "action": {
@@ -137,12 +217,23 @@ class PendingToolApproval:
 
 @dataclass
 class ExactToolApproval:
-    """A consumed grant that the dispatcher can claim exactly once."""
+    """A consumed exact first action plus an explicit continuation scope."""
 
     pending: PendingToolApproval
+    scope: ToolApprovalScope = ToolApprovalScope.TASK
+    # The seam consumed by agent_loop. Both chat-card allow choices cover the
+    # complete resumed task, because one-action scope there immediately
+    # re-entered the same gate on the next round. Callers with no resumable
+    # chat still get SINGLE_ACTION, which leaves the gate armed behind the
+    # sealed action.
+    allow_remaining_actions: bool = True
     _claimed: bool = field(default=False, init=False, repr=False)
     _lock: threading.Lock = field(default_factory=threading.Lock, init=False, repr=False)
 
+    @property
+    def grants_chat_session(self) -> bool:
+        return self.scope is ToolApprovalScope.CHAT_SESSION
+
     def _matches_unlocked(
         self,
         *,
@@ -175,6 +266,8 @@ class ExactToolApproval:
             external_untrusted_context_seen=(
                 self.pending.external_untrusted_context_seen
             ),
+            selected_tools=self.pending.selected_tools,
+            continuation_query=self.pending.continuation_query,
             effects=effects,
             result_integrity=result_integrity,
         )
@@ -255,6 +348,8 @@ class ToolApprovalStore:
         document_id: Any = None,
         document_version: Any = None,
         document_digest: Any = None,
+        selected_tools: Any = None,
+        continuation_query: Any = None,
         external_untrusted_context_seen: bool,
         capabilities: ToolCapabilities,
     ) -> PendingToolApproval:
@@ -272,6 +367,8 @@ class ToolApprovalStore:
             document_version=document_version,
             document_digest=document_digest,
             external_untrusted_context_seen=external_untrusted_context_seen,
+            selected_tools=selected_tools,
+            continuation_query=continuation_query,
             effects=effects,
             result_integrity=result_integrity,
         )
@@ -294,6 +391,8 @@ class ToolApprovalStore:
             digest=_canonical_digest(payload),
             created_at=now,
             expires_at=now + self._ttl_seconds,
+            selected_tools=tuple(payload["selected_tools"]),
+            continuation_query=payload["continuation_query"],
         )
         with self._lock:
             self._purge_expired_locked(now)
@@ -331,7 +430,17 @@ class ToolApprovalStore:
         decision: Any,
         owner: Any,
         session_id: Any,
+        allow_continuation: bool = True,
     ) -> ExactToolApproval | None:
+        """Consume a pending approval.
+
+        ``allow_continuation`` is the caller's assertion that it owns a
+        resumable conversation the granted scope can apply to. Callers without
+        one (the skill tester, unattended audits) pass ``False`` and get the
+        original one-use grant, so a button labelled "Allow once" cannot widen
+        into a run-long bypass just because the chat card reuses the same wire
+        value.
+        """
         now = time.time()
         with self._lock:
             self._purge_expired_locked(now)
@@ -348,9 +457,21 @@ class ToolApprovalStore:
                 # another owner's pending action.
                 return None
             self._pending.pop(approval_key, None)
-        if str(decision or "").strip().lower() != "approve":
+        normalized_decision = str(decision or "").strip().lower()
+        scope = scope_for_decision(normalized_decision)
+        if scope is None:
             return None
-        return ExactToolApproval(pending)
+        if not allow_continuation:
+            return ExactToolApproval(
+                pending,
+                scope=ToolApprovalScope.SINGLE_ACTION,
+                allow_remaining_actions=False,
+            )
+        return ExactToolApproval(
+            pending,
+            scope=scope,
+            allow_remaining_actions=True,
+        )
 
     def peek(self, approval_id: Any) -> PendingToolApproval | None:
         now = time.time()

src/tool_capabilities.py

@@ -14,6 +14,7 @@ from enum import Enum
 from types import MappingProxyType
 from typing import Any, Iterable, Mapping
 
+from src.tool_approval_scopes import CHAT_SESSION_APPROVAL_CONTEXT_MARKER
 from src.tool_security import BUILTIN_EMAIL_TOOLS
 
 
@@ -618,13 +619,30 @@ class ToolRunSecurityContext:
     external_untrusted_context_seen: bool = False
     external_sources: list[str] = field(default_factory=list)
     run_id: str = field(default_factory=lambda: uuid.uuid4().hex)
+    # Task-scope approval sets this for the resumed in-memory run. Chat-scope
+    # approval is projected from the server-owned session history marker below.
+    # The bypass affects only this automatic gate; current tool policy, ownership,
+    # workspace confinement, and execution/sandbox restrictions still apply.
+    approval_gate_bypassed: bool = False
 
     def observe_messages(self, messages: Iterable[dict]) -> None:
-        """Promote any server-labelled untrusted prompt context into the gate."""
-        if messages_contain_external_untrusted_context(messages):
+        """Apply server-owned chat scope and promote untrusted prompt context."""
+        message_list = list(messages or ())
+        if any(
+            isinstance(message, dict)
+            and isinstance(message.get("metadata"), dict)
+            and message["metadata"].get(
+                CHAT_SESSION_APPROVAL_CONTEXT_MARKER
+            ) is True
+            for message in message_list
+        ):
+            self.approval_gate_bypassed = True
+        if messages_contain_external_untrusted_context(message_list):
             self.external_untrusted_context_seen = True
 
     def decision_for(self, tool_name: Any, content: Any = None) -> ToolGateDecision:
+        if self.approval_gate_bypassed:
+            return ToolGateDecision(True)
         if not self.external_untrusted_context_seen:
             return ToolGateDecision(True)
         capabilities = capabilities_for_action(tool_name, content)

src/user_time.py

@@ -8,7 +8,7 @@ from __future__ import annotations
 
 import re
 from contextvars import ContextVar
-from datetime import datetime, timedelta, timezone
+from datetime import datetime, timedelta, timezone, tzinfo
 from typing import Dict, Optional
 
 
@@ -65,19 +65,31 @@ def format_utc_offset(offset_min: Optional[int]) -> str:
     return f"{sign}{hours:02d}:{minutes:02d}"
 
 
-def user_timezone() -> timezone:
-    """Return the best known user timezone as a fixed-offset tzinfo."""
+def _zoneinfo_from_name():
+    """Return ZoneInfo for the request's IANA name, or None if missing/invalid."""
+    name = get_user_tz_name()
+    if not name:
+        return None
+    try:
+        from zoneinfo import ZoneInfo
+        return ZoneInfo(name)
+    except Exception:
+        return None
+
+
+def user_timezone() -> tzinfo:
+    """Return the best known user timezone.
+
+    A valid IANA name wins over x-tz-offset. The offset is a fixed number and
+    can disagree with the name (wrong sign, stale client); the name carries DST.
+    """
+    zone = _zoneinfo_from_name()
+    if zone is not None:
+        return zone
     offset = get_user_tz_offset()
-    if offset is None:
-        name = get_user_tz_name()
-        if name:
-            try:
-                from zoneinfo import ZoneInfo
-                return ZoneInfo(name)
-            except Exception:
-                pass
-        return datetime.now().astimezone().tzinfo or timezone.utc
-    return timezone(timedelta(minutes=offset))
+    if offset is not None:
+        return timezone(timedelta(minutes=offset))
+    return datetime.now().astimezone().tzinfo or timezone.utc
 
 
 def now_user_local(now_utc: Optional[datetime] = None) -> datetime:
@@ -100,14 +112,13 @@ def _clock_label(dt: datetime) -> str:
 
 def timezone_label(dt: Optional[datetime] = None) -> str:
     """Return a concise display label such as Australia/Brisbane, UTC+10:00."""
-    offset = get_user_tz_offset()
-    if offset is None:
-        if dt is None:
-            dt = datetime.now().astimezone()
-        offset = int((dt.utcoffset() or timedelta()).total_seconds() // 60)
+    if dt is None:
+        dt = now_user_local()
+    offset = int((dt.utcoffset() or timedelta()).total_seconds() // 60)
     offset_label = f"UTC{format_utc_offset(offset)}"
-    name = get_user_tz_name()
-    return f"{name}, {offset_label}" if name else offset_label
+    if _zoneinfo_from_name() is not None:
+        return f"{get_user_tz_name()}, {offset_label}"
+    return offset_label
 
 
 def current_datetime_prompt(now_utc: Optional[datetime] = None) -> str:

start-macos.sh

@@ -34,6 +34,10 @@ fi
 # values (APP_PORT / APP_BIND), then built-in defaults.
 PORT="${ODYSSEUS_PORT:-${APP_PORT:-7860}}"   # 7860, not 7000 — macOS AirPlay Receiver holds 7000.
 HOST="${ODYSSEUS_HOST:-${APP_BIND:-127.0.0.1}}" # Set APP_BIND=0.0.0.0 in .env for LAN/Tailscale access.
+# The port only reaches uvicorn as a flag, so export it too: everything that
+# builds a URL for this instance — internal_api_base(), the companion pairing
+# code, the MCP OAuth callback — reads APP_PORT and would otherwise assume 7000.
+export APP_PORT="$PORT"
 PROBE_HOST="$HOST"
 if [ "$PROBE_HOST" = "0.0.0.0" ] || [ "$PROBE_HOST" = "::" ]; then
     PROBE_HOST="127.0.0.1"

static/app.js

@@ -10,8 +10,8 @@ import modelsModule from './js/models.js?v=20260715startupcalm2';
 import ragModule from './js/rag.js';
 import presetsModule from './js/presets.js';
 import searchModule from './js/search.js';
-import chatModule from './js/chat.js?v=20260815toolapproval4';
-import compareModule from './js/compare/index.js?v=20260723compareicon2';
+import chatModule from './js/chat.js?v=20260819approvalcontrol1';
+import compareModule from './js/compare/index.js?v=20260819approvalcontrol1';
 import documentModule from './js/document.js?v=20260815approvalsave1';
 import searchChatModule from './js/search-chat.js';
 import { makeWindowDraggable } from './js/windowDrag.js';
@@ -22,7 +22,7 @@ import {
   settleSessionHydration
 } from './js/startupShell.js';
 import markdownModule from './js/markdown.js';
-import chatRenderer from './js/chatRenderer.js?v=20260815toolapproval4';
+import chatRenderer from './js/chatRenderer.js?v=20260819approvalcontrol1';
 import sessionModule from './js/sessions.js';
 import memoryModule from './js/memory.js?v=20260722memoryloading1';
 import voiceRecorderModule from './js/voiceRecorder.js';
@@ -50,6 +50,7 @@ import * as researchPanelModule from './js/research/panel.js?v=20260630researcht
 import ttsModule from './js/tts-ai.js';
 import spinnerModule from './js/spinner.js';
 import { initKeyboardShortcuts } from './js/keyboard-shortcuts.js';
+import { getSettings } from './js/appConfig.js';
 import { initSidebarLayout, syncRailSide } from './js/sidebar-layout.js?v=20260715startupclean';
 import { initSectionCollapse, initSectionDrag } from './js/section-management.js';
 
@@ -1518,13 +1519,11 @@ function initializeEventListeners() {
     })
     .catch(() => {});
 
-  // Hide Gallery when image generation is disabled in settings
-  const _prefetchedSettings = sessionStorage.getItem('ody-prefetch-settings');
-  sessionStorage.removeItem('ody-prefetch-settings');
-  window._initSettingsReady = (_prefetchedSettings
-    ? Promise.resolve(JSON.parse(_prefetchedSettings))
-    : fetch(`${API_BASE}/api/auth/settings`, { credentials: 'same-origin' }).then(r => r.json())
-  ).then(settings => {
+  // Hide Gallery when image generation is disabled in settings.
+  // getSettings() consumes the login prefetch itself, so every other module
+  // that asks for settings this load gets the same snapshot without a request.
+  window._initSettingsReady = getSettings()
+    .then(settings => {
       // NOTE: image_gen_enabled only governs *generating* images in chat — the
       // tool is blocked server-side (chat_routes / agent_loop). The Gallery
       // holds uploads and past images too, so it stays visible regardless;
@@ -3705,7 +3704,7 @@ function startOdysseusApp() {
   modelsModule.init(API_BASE);
   ragModule.init(API_BASE);
   presetsModule.init(API_BASE);
-  searchModule.init(API_BASE);
+  searchModule.init();
   chatModule.init(API_BASE);
   chatModule.initListeners();
   groupModule.init(API_BASE);

static/index.html

@@ -231,23 +231,11 @@
       }
     }
   </style>
-  <!-- KaTeX CSS is loaded with media="print" so it doesn't block render,
-       then flipped to "all" via JS after load. Mermaid init runs once the
-       library finishes loading. Both hooks are wired via addEventListener
-       below (inline onload= attrs are blocked by CSP script-src-attr). -->
-  <link id="katex-css" rel="stylesheet" href="https://cdn.jsdelivr.net/npm/katex@0.16.22/dist/katex.min.css" media="print">
-  <script async src="https://cdn.jsdelivr.net/npm/katex@0.16.22/dist/katex.min.js"></script>
-  <script id="mermaid-script" async src="https://cdn.jsdelivr.net/npm/mermaid@11/dist/mermaid.min.js"></script>
-  <script nonce="{{CSP_NONCE}}">
-    (function(){
-      var k = document.getElementById('katex-css');
-      if (k) k.addEventListener('load', function(){ k.media = 'all'; }, { once: true });
-      var m = document.getElementById('mermaid-script');
-      if (m) m.addEventListener('load', function(){
-        if (window.odysseusInitMermaid) window.odysseusInitMermaid();
-      }, { once: true });
-    })();
-  </script>
+  <!-- KaTeX and Mermaid are vendored in /static/lib and pulled in by
+       static/js/markdown.js the first time a page actually renders math or a
+       ```mermaid fence. They used to load here from cdn.jsdelivr.net on every
+       page load, which cost ~985 KB on the wire, broke offline installs, and
+       announced every session to a third party. -->
   <!-- Preload the two faces first paint actually uses: Fira Code 400 and 600,
        the app font and the weight the sidebar and header text render at. They
        are declared in style.css, so without a hint they are only discovered
@@ -2584,10 +2572,10 @@
 <script type="module" src="/static/js/tts-ai.js"></script>
 <script type="module" src="/static/js/document.js?v=20260815approvalsave1"></script>
 <script type="module" src="/static/js/gallery.js?v=20260708match1"></script>
-<script type="module" src="/static/js/chatRenderer.js?v=20260815toolapproval4"></script>
+<script type="module" src="/static/js/chatRenderer.js?v=20260819approvalcontrol1"></script>
 <script type="module" src="/static/js/codeRunner.js"></script>
-<script type="module" src="/static/js/chatStream.js?v=20260815approvalsave1"></script>
-  <script type="module" src="/static/js/chat.js?v=20260815toolapproval4"></script>
+<script type="module" src="/static/js/chatStream.js?v=20260819approvalcontrol1"></script>
+  <script type="module" src="/static/js/chat.js?v=20260819approvalcontrol1"></script>
 <script type="module" src="/static/js/cookbook.js"></script>
 <script src="/static/js/cookbookSchedule.js"></script>
 <script type="module" src="/static/js/search-chat.js"></script>

static/js/admin.js

@@ -6,6 +6,7 @@ import settingsModule from './settings.js';
 import { providerLogo, providerLogoFromUrl } from './providers.js';
 import { sortModelObjects } from './modelSort.js';
 import { PROVIDER_DEVICE_FLOWS, formatDeviceFlowError, runProviderDeviceFlow } from './providerDeviceFlow.js';
+import { getSettings, getTools, invalidateSettings, invalidateTools } from './appConfig.js';
 
 let initialized = false;
 let modalEl = null;
@@ -345,8 +346,7 @@ function initSignupToggle() {
 
 function initShareDefaultsToggle() {
   const toggle = el('adm-shareDefaultsToggle');
-  fetch('/api/auth/settings', { credentials: 'same-origin' })
-    .then(r => r.json())
+  getSettings()
     .then(d => { toggle.checked = !!d.share_defaults_with_users; })
     .catch(e => console.warn('Settings fetch failed:', e));
   toggle.addEventListener('change', async () => {
@@ -361,6 +361,9 @@ function initShareDefaultsToggle() {
       toggle.checked = !!data.share_defaults_with_users;
     } catch (e) {
       toggle.checked = !toggle.checked;
+    } finally {
+      // Drop the shared snapshot: it still says what this toggle used to be.
+      invalidateSettings();
     }
   });
 }
@@ -1893,8 +1896,16 @@ async function loadBuiltinTools() {
   const list = el('adm-builtin-tools-list');
   if (!list) return;
   try {
-    const res = await fetch('/api/tools', { credentials: 'same-origin' });
-    const data = await res.json();
+    // This panel is an editor, and its save posts the whole disabled list
+    // rebuilt from the checkboxes below. So it has to render authoritative
+    // state: a snapshot that went stale out of band (the manage_settings tool,
+    // another tab) would be re-posted wholesale on the next unrelated toggle
+    // and would silently undo the newer state. refreshAll() calls this on every
+    // panel open, so drop the shared entry and refill it. The startup read that
+    // chatRenderer.js shares is unaffected; this panel just never edits a cache,
+    // which is the same rule the settings panel follows by reading directly.
+    invalidateTools();
+    const data = await getTools();
     const tools = data.tools || [];
     if (!tools.length) { list.innerHTML = '<div class="admin-empty">No tools found</div>'; return; }
 
@@ -1968,17 +1979,50 @@ async function loadBuiltinTools() {
       });
     });
 
-    // Helper: save disabled tools + update counters
-    async function _saveToolState() {
-      const allChecks = list.querySelectorAll('input[data-tool-id]');
-      const disabled = [];
-      allChecks.forEach(c => { if (!c.checked) disabled.push(c.dataset.toolId); });
-      await fetch('/api/tools', {
-        method: 'POST',
-        headers: { 'Content-Type': 'application/json' },
-        body: JSON.stringify({ disabled }),
-        credentials: 'same-origin',
-      });
+    // Merge only the user's intended changes onto authoritative server state.
+    // /api/tools replaces the full disabled list, so rebuilding it from this
+    // panel's DOM can undo a change made by another tab or manage_settings
+    // after the panel was opened.
+    async function _saveToolState(changes) {
+      invalidateTools();
+      const latest = await getTools();
+      const state = new Map(
+        (latest.tools || []).map(t => [t.id, !!t.enabled])
+      );
+
+      for (const change of changes) {
+        if (state.has(change.id)) {
+          state.set(change.id, !!change.enabled);
+        }
+      }
+
+      const disabled = Array.from(state.entries())
+        .filter(([, enabled]) => !enabled)
+        .map(([id]) => id);
+
+      try {
+        const res = await fetch('/api/tools', {
+          method: 'POST',
+          headers: { 'Content-Type': 'application/json' },
+          body: JSON.stringify({ disabled }),
+          credentials: 'same-origin',
+        });
+        if (!res.ok) throw new Error(`Failed to update tools (${res.status})`);
+
+        // Bring the still-open editor forward to the same merged snapshot so an
+        // out-of-band change is visible instead of leaving stale checkboxes.
+        list.querySelectorAll('input[data-tool-id]').forEach(c => {
+          if (state.has(c.dataset.toolId)) {
+            c.checked = state.get(c.dataset.toolId);
+          }
+        });
+        list.querySelectorAll('.admin-tool-category').forEach(_updateCatCounter);
+      } finally {
+        // This route persists disabled_tools into the settings store
+        // (routes/model_routes.py), so both snapshots are now stale.
+        invalidateTools();
+        invalidateSettings();
+      }
     }
     function _updateCatCounter(catEl) {
       if (!catEl) return;
@@ -1993,7 +2037,9 @@ async function loadBuiltinTools() {
     // Wire individual tool toggles
     list.querySelectorAll('input[data-tool-id]').forEach(chk => {
       chk.addEventListener('change', async () => {
-        await _saveToolState();
+        await _saveToolState([
+          { id: chk.dataset.toolId, enabled: chk.checked },
+        ]);
         _updateCatCounter(chk.closest('.admin-tool-category'));
       });
     });
@@ -2004,8 +2050,10 @@ async function loadBuiltinTools() {
         const catEl = chk.closest('.admin-tool-category');
         if (!catEl) return;
         const checked = chk.checked;
+        const changes = Array.from(catEl.querySelectorAll('input[data-tool-id]'))
+          .map(c => ({ id: c.dataset.toolId, enabled: checked }));
         catEl.querySelectorAll('input[data-tool-id]').forEach(c => { c.checked = checked; });
-        await _saveToolState();
+        await _saveToolState(changes);
         _updateCatCounter(catEl);
       });
     });

static/js/appConfig.js

@@ -0,0 +1,86 @@
+// static/js/appConfig.js
+//
+// One shared, invalidatable cache for the two config endpoints that every
+// module wants at startup.
+//
+// Before this, /api/auth/settings was fetched independently by six modules and
+// /api/tools by three, none of them aware of the others — 4 and 3 requests on a
+// single cold load. Worse than the requests: each caller could observe a
+// different snapshot of the same object, and chatRenderer.js is imported under
+// three different ?v= query strings, so it is three separate module instances
+// each issuing its own /api/tools fetch. Caching here fixes both, because the
+// cache lives in one module every instance imports by the same specifier.
+//
+// URLs are bare paths on purpose. The callers that used `${API_BASE}/api/...`
+// resolved to the identical URL — API_BASE is `window.location.origin`
+// (app.js) — so nothing about the request changes for them.
+//
+// WRITERS MUST INVALIDATE. Anything that POSTs /api/auth/settings calls
+// invalidateSettings(); anything that POSTs /api/tools calls invalidateTools()
+// *and* invalidateSettings(), because that route persists `disabled_tools`
+// into the same settings store (routes/model_routes.py). Miss one and the UI
+// serves a stale settings object for the rest of the session, which is worse
+// than the duplicate fetches this replaces.
+//
+// The resolved object is shared by reference, so treat it as read-only: copy
+// before mutating (`{ ...await getSettings() }`).
+
+// Written by login.html immediately before it redirects to '/', so the first
+// load after a login can skip the request entirely. Consumed once per page
+// load, by whichever module asks for settings first.
+const PREFETCH_KEY = 'ody-prefetch-settings';
+
+const _URLS = { settings: '/api/auth/settings', tools: '/api/tools' };
+const _cache = { settings: null, tools: null };
+
+function _readPrefetchedSettings() {
+  try {
+    const raw = sessionStorage.getItem(PREFETCH_KEY);
+    if (!raw) return null;
+    sessionStorage.removeItem(PREFETCH_KEY);
+    return JSON.parse(raw);
+  } catch (_) {
+    return null;
+  }
+}
+
+// A rejected promise must not stay in the slot. Plain `??=` memoisation would
+// keep it, so one transient blip during boot would leave keybinds, TTS and the
+// search provider on their defaults for the whole session with no retry. Clear
+// the slot on failure — unless a later invalidate/refetch already replaced it —
+// and rethrow, so every caller's existing .catch() still runs exactly as before.
+function _get(key) {
+  if (_cache[key]) return _cache[key];
+  const pending = fetch(_URLS[key], { credentials: 'same-origin' })
+    .then(r => r.json())
+    .catch(err => {
+      if (_cache[key] === pending) _cache[key] = null;
+      throw err;
+    });
+  _cache[key] = pending;
+  return pending;
+}
+
+/** GET /api/auth/settings, once per page load (or once per invalidation). */
+export function getSettings() {
+  if (!_cache.settings) {
+    const prefetched = _readPrefetchedSettings();
+    if (prefetched) _cache.settings = Promise.resolve(prefetched);
+  }
+  return _get('settings');
+}
+
+/** GET /api/tools, once per page load (or once per invalidation). */
+export function getTools() {
+  return _get('tools');
+}
+
+/** Call after any write that can change settings. */
+export function invalidateSettings() {
+  _cache.settings = null;
+}
+
+/** Call after any write that can change the tool enable/disable state. */
+export function invalidateTools() {
+  _cache.tools = null;
+}

static/js/chat.js

@@ -8,8 +8,8 @@
 import Storage from './storage.js';
 import uiModule from './ui.js';
 import sessionModule from './sessions.js';
-import chatRenderer from './chatRenderer.js?v=20260815toolapproval4';
-import chatStream from './chatStream.js?v=20260815approvalsave1';
+import chatRenderer from './chatRenderer.js?v=20260819approvalcontrol1';
+import chatStream from './chatStream.js?v=20260819approvalcontrol1';
 import { addAITTSButton } from './tts-ai.js';
 import markdownModule from './markdown.js';
 import spinnerModule from './spinner.js';
@@ -35,6 +35,7 @@ import {
   inheritModelRouteState,
 } from './chatModelProvenance.js';
 import { createTerminalStreamError, isRecoverableStreamError } from './chatStreamErrors.js';
+import { loadPanel } from './panels.js';
 
   const RESEARCH_TIMEOUT_MS = 360000;
   const DEFAULT_TIMEOUT_MS = 120000;
@@ -61,20 +62,18 @@ import { createTerminalStreamError, isRecoverableStreamError } from './chatStrea
   let _contextHeaderBound = false;
   let _pendingToolApproval = null;
 
-  function _submitToolApprovalWhenIdle(approvalId, label) {
+  function _submitToolApprovalWhenIdle(approvalId) {
     if (
       !_pendingToolApproval
       || _pendingToolApproval.approval_id !== approvalId
     ) return;
     if (isStreaming || _sendInFlight) {
-      setTimeout(() => _submitToolApprovalWhenIdle(approvalId, label), 120);
+      setTimeout(() => _submitToolApprovalWhenIdle(approvalId), 120);
       return;
     }
     const input = document.getElementById('message');
     if (input) {
       _pendingToolApproval.draft = input.value || '';
-      input.value = label;
-      input.dispatchEvent(new Event('input', { bubbles: true }));
     }
     const sendButton = document.querySelector('.send-btn');
     if (sendButton) sendButton.click();
@@ -83,16 +82,13 @@ import { createTerminalStreamError, isRecoverableStreamError } from './chatStrea
   document.addEventListener('odysseus:tool-approval', (event) => {
     const detail = event && event.detail ? event.detail : {};
     const decision = String(detail.decision || '').toLowerCase();
-    if (!detail.approval_id || !['approve', 'deny'].includes(decision)) return;
+    if (!detail.approval_id || !['approve', 'approve_task', 'deny'].includes(decision)) return;
     _pendingToolApproval = {
       approval_id: String(detail.approval_id),
       decision,
       document_id: String(detail.document_id || ''),
     };
-    _submitToolApprovalWhenIdle(
-      _pendingToolApproval.approval_id,
-      detail.label || (decision === 'approve' ? 'Allow once' : 'Deny'),
-    );
+    _submitToolApprovalWhenIdle(_pendingToolApproval.approval_id);
   });
 
   function _fmtContextNumber(n) {
@@ -1308,10 +1304,10 @@ import { createTerminalStreamError, isRecoverableStreamError } from './chatStrea
     }
 
     const el = uiModule.el;
-    const msg = el('message').value;
+    const msg = approvalForSend ? '' : el('message').value;
     // Allow empty text when a regen carries over the original message's
     // attachment ids — a photo-only message still has something to send.
-    if (!msg.trim() && !fileHandlerModule.getPendingCount() && !(_pendingRegenAttachments && _pendingRegenAttachments.length)) { _releaseSendFlag(); return; }
+    if (!msg.trim() && !approvalForSend && !fileHandlerModule.getPendingCount() && !(_pendingRegenAttachments && _pendingRegenAttachments.length)) { _releaseSendFlag(); return; }
 
     // --- Slash commands: execute directly without AI (no session needed) ---
     if (!approvalForSend && isCommand(msg.trim())) {
@@ -1589,7 +1585,7 @@ import { createTerminalStreamError, isRecoverableStreamError } from './chatStrea
 
       const userDisplay = _displayOverride || msg;
       _displayOverride = null;
-      const skipBubble = _hideUserBubble;
+      const skipBubble = _hideUserBubble || !!approvalForSend;
       _hideUserBubble = false;
       // Auto-recovery counter: carries across a turn's auto-continues, but resets
       // when the user genuinely sends a new message (so each task gets a fresh cap).
@@ -1832,7 +1828,7 @@ import { createTerminalStreamError, isRecoverableStreamError } from './chatStrea
       if (_inject.suffix) _finalMsgWithInject = _finalMsgWithInject + ' ' + _inject.suffix;
 
       const fd = new FormData();
-      fd.append('message', _finalMsgWithInject);
+      fd.append('message', approvalForSend ? '' : _finalMsgWithInject);
       fd.append('session', streamSessionId);
       if (approvalForSend) {
         fd.append('tool_approval_id', approvalForSend.approval_id);
@@ -2872,7 +2868,7 @@ import { createTerminalStreamError, isRecoverableStreamError } from './chatStrea
                 if (spinner && spinner.element) spinner.destroy();
                 break;
               }
-              if (json.delta || json.type === 'agent_prep' || json.type === 'generated_image' || json.type === 'tool_start' || json.type === 'tool_output' || json.type === 'tool_progress' || json.type === 'agent_step' || json.type === 'loop_breaker_triggered' || json.type === 'intent_nudge_exhausted' || json.type === 'doc_stream_open' || json.type === 'doc_stream_delta' || json.type === 'research_progress') {
+              if (json.delta || json.type === 'agent_prep' || json.type === 'tool_approval_resolved' || json.type === 'generated_image' || json.type === 'tool_start' || json.type === 'tool_output' || json.type === 'tool_progress' || json.type === 'agent_step' || json.type === 'loop_breaker_triggered' || json.type === 'intent_nudge_exhausted' || json.type === 'doc_stream_open' || json.type === 'doc_stream_delta' || json.type === 'research_progress') {
                 clearResponseTimeout();
                 clearProcessingProbe();
                 clearFirstTokenWaitTimers();
@@ -2889,6 +2885,14 @@ import { createTerminalStreamError, isRecoverableStreamError } from './chatStrea
                 }
                 continue;
               }
+              if (json.type === 'tool_approval_resolved') {
+                _cancelThinkingTimer();
+                _removeThinkingSpinner();
+                if (spinner && spinner.element) spinner.destroy();
+                if (!_isBg && roundHolder && roundHolder !== holder) roundHolder.remove();
+                if (!_isBg && holder) holder.remove();
+                continue;
+              }
               if (json.delta) {
                 _cancelThinkingTimer();
                 _removeThinkingSpinner();
@@ -6606,7 +6610,7 @@ import { createTerminalStreamError, isRecoverableStreamError } from './chatStrea
     // Images → Gallery editor.
     if (isImage) {
       try {
-        const gx = await import('./galleryEditor.js');
+        const gx = await loadPanel('editor');
         if (gx.openEditor) { gx.openEditor(url, id, null, name); return; }
       } catch (e) { console.warn('gallery open failed', e); }
       window.open(url, '_blank');

static/js/chatRenderer.js

@@ -9,7 +9,9 @@ import { providerLogo, providerLabel } from './providers.js';
 import settingsModule from './settings.js';
 import spinnerModule from './spinner.js';
 import { bindMenuDismiss } from './escMenuStack.js';
+import { loadPanel } from './panels.js';
 import { matchModelKey } from './model/matchKey.js';
+import { getTools } from './appConfig.js';
 
 const SEARCH_ICON = '<svg viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2"><circle cx="11" cy="11" r="8"/><path d="M21 21l-4.35-4.35"/></svg>';
 const REPORT_ICON = '<svg width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><path d="M14 2H6a2 2 0 0 0-2 2v16a2 2 0 0 0 2 2h12a2 2 0 0 0 2-2V8z"/><polyline points="14 2 14 8 20 8"/><line x1="16" y1="13" x2="8" y2="13"/><line x1="16" y1="17" x2="8" y2="17"/><line x1="10" y1="9" x2="8" y2="9"/></svg>';
@@ -445,8 +447,12 @@ function stripExecutedFence(match, tag, inline, body) {
 
 async function loadExecFenceRegex() {
   try {
-    const res = await fetch('/api/tools', { credentials: 'same-origin' });
-    const data = await res.json();
+    // Shared with admin.js, and — more to the point — with the other copies of
+    // this module: chatRenderer.js is imported under three different ?v= query
+    // strings, so it is instantiated three times per load and used to issue
+    // three identical /api/tools requests. appConfig.js is imported by one
+    // specifier from all of them, so they now share a single fetch.
+    const data = await getTools();
     const tags = (data.tools || [])
       .map((t) => t.id)
       .filter((id) => id && !EXEC_FENCE_NON_TOOL.has(id));
@@ -1548,7 +1554,7 @@ export function buildImageBubble(imageUrl, prompt, model, size, quality, imageId
     try {
       const [galleryMod, editorMod] = await Promise.all([
         import('./gallery.js'),
-        import('./galleryEditor.js'),
+        loadPanel('editor'),
       ]);
       // Ensure the Gallery modal is open so the editor has a container
       // to render into; switch its tabs to the Edit tab.
@@ -2321,6 +2327,42 @@ export function removeAskUserCards(root) {
   scope.querySelectorAll('.ask-user-card').forEach((node) => node.remove());
 }
 
+// While a choice card is visible, let plain 1–3 activate the corresponding
+// rendered option. Reuse the option's click path so the question keeps its
+// existing submission semantics. Tool approval cards are excluded: that card
+// exists to make consent deliberate after untrusted context influenced the
+// run, and its first option is the widest grant, so a stray digit must not
+// answer it.
+function _handleAskUserShortcut(event) {
+  if (
+    event.defaultPrevented
+    || event.repeat
+    || event.isComposing
+    || event.ctrlKey
+    || event.altKey
+    || event.metaKey
+    || event.shiftKey
+  ) return;
+  if (!/^[1-3]$/.test(event.key)) return;
+
+  const target = event.target;
+  if (target?.closest?.('input, textarea, select, [contenteditable="true"]')) return;
+
+  const focusedCard = document.activeElement?.closest?.('.ask-user-card') || null;
+  const mainCard = document.querySelector('#chat-history .ask-user-card');
+  const compareCards = document.querySelectorAll('.compare-pane .ask-user-card');
+  const card = focusedCard || mainCard || (compareCards.length === 1 ? compareCards[0] : null);
+  if (!card) return;
+  if (card.dataset.askUserKind === 'tool_approval') return;
+  const option = card.querySelectorAll('.ask-user-option')[Number(event.key) - 1];
+  if (!option || option.disabled) return;
+
+  event.preventDefault();
+  option.click();
+}
+
+document.addEventListener('keydown', _handleAskUserShortcut);
+
 /**
  * Render an ask_user payload as a durable choice card.
  *
@@ -2330,11 +2372,15 @@ export function removeAskUserCards(root) {
  */
 export function renderAskUserCard(payload, options) {
   const aq = payload || {};
+  if (aq.resolved) return null;
   const opts = Array.isArray(aq.options) ? aq.options : [];
-  const chatBox = document.getElementById('chat-history');
+  const renderOptions = options || {};
+  const chatBox = renderOptions.root || document.getElementById('chat-history');
+  const onSubmit = typeof renderOptions.onSubmit === 'function'
+    ? renderOptions.onSubmit
+    : null;
   if (!chatBox || !aq.question || opts.length < 2) return null;
 
-  const renderOptions = options || {};
   removeAskUserCards(chatBox);
 
   const card = document.createElement('div');
@@ -2343,6 +2389,7 @@ export function renderAskUserCard(payload, options) {
   card.tabIndex = -1;
   const multi = !!aq.multi;
   const isToolApproval = aq.kind === 'tool_approval' && !!aq.approval_id;
+  card.dataset.askUserKind = isToolApproval ? 'tool_approval' : 'question';
   const emojiText = (value) => svgifyEmoji(uiModule.esc(String(value)));
 
   const head = document.createElement('div');
@@ -2351,7 +2398,6 @@ export function renderAskUserCard(payload, options) {
   closeBtn.type = 'button';
   closeBtn.className = 'modal-close ask-user-close';
   closeBtn.setAttribute('aria-label', 'Dismiss question');
-  closeBtn.textContent = '×';
   closeBtn.addEventListener('click', () => {
     card.remove();
     const input = uiModule.el('message');
@@ -2394,6 +2440,17 @@ export function renderAskUserCard(payload, options) {
 
   const send = (text) => {
     if (!text) return;
+    if (onSubmit) {
+      const accepted = onSubmit({
+        kind: 'answer',
+        text,
+        label: text,
+        payload: aq,
+        card,
+      });
+      if (accepted !== false) card.remove();
+      return;
+    }
     card.remove();
     const input = uiModule.el('message');
     if (input) input.value = text;
@@ -2427,17 +2484,26 @@ export function renderAskUserCard(payload, options) {
       row.type = 'button';
       row.addEventListener('click', () => {
         if (isToolApproval) {
-          card.remove();
-          document.dispatchEvent(new CustomEvent('odysseus:tool-approval', {
-            detail: {
-              approval_id: aq.approval_id,
-              decision: String((opt && opt.value) || '').toLowerCase(),
-              label,
-              document_id: aq.action && aq.action.document_id
-                ? String(aq.action.document_id)
-                : '',
-            },
-          }));
+          const detail = {
+            approval_id: aq.approval_id,
+            decision: String((opt && opt.value) || '').toLowerCase(),
+            label,
+            document_id: aq.action && aq.action.document_id
+              ? String(aq.action.document_id)
+              : '',
+          };
+          if (onSubmit) {
+            const accepted = onSubmit({
+              kind: 'tool_approval',
+              ...detail,
+              payload: aq,
+              card,
+            });
+            if (accepted !== false) card.remove();
+          } else {
+            card.remove();
+            document.dispatchEvent(new CustomEvent('odysseus:tool-approval', { detail }));
+          }
         } else {
           send(label);
         }
@@ -2622,7 +2688,7 @@ export function addMessage(role, content, modelName, metadata) {
             box.appendChild(threadWrap);
           }
           for (const ev of roundTools) {
-            if (ev.ask_user) pendingAskUser = ev.ask_user;
+            if (ev.ask_user && !ev.ask_user.resolved) pendingAskUser = ev.ask_user;
             const ok = (ev.exit_code === 0 || ev.exit_code == null);
             let outHtml = '';
             if (ev.output && ev.output.trim()) {

static/js/chatStream.js

@@ -9,6 +9,35 @@ import markdownModule from './markdown.js';
 import sessionModule from './sessions.js';
 import documentModule from './document.js?v=20260815approvalsave1';
 
+// Tool approvals are control-plane submits for the current chat. chat.js
+// deliberately leaves the composer untouched, then programmatically clicks the
+// shared send button after it records the sealed approval id/decision. That
+// button is polymorphic: with an empty composer it can mean New chat or Record
+// voice instead of Send. Intercept only the programmatic approval click and
+// route it through the form submit path, which already reaches chat.js directly.
+document.addEventListener('odysseus:tool-approval', () => {
+  const sendButton = document.querySelector('.send-btn');
+  const chatForm = document.getElementById('chat-form');
+  if (!sendButton || !chatForm) return;
+
+  const interceptApprovalClick = (event) => {
+    // A real user click must retain the normal send/new-chat/STT behavior.
+    if (event.isTrusted) return;
+    sendButton.removeEventListener('click', interceptApprovalClick, true);
+    event.preventDefault();
+    event.stopImmediatePropagation();
+    if (chatForm.requestSubmit) chatForm.requestSubmit();
+    else chatForm.dispatchEvent(new Event('submit', { bubbles: true, cancelable: true }));
+  };
+
+  sendButton.addEventListener('click', interceptApprovalClick, true);
+  // Fail-safe cleanup if the approval continuation never reaches its deferred
+  // synthetic click (for example because the surrounding view is torn down).
+  setTimeout(() => {
+    sendButton.removeEventListener('click', interceptApprovalClick, true);
+  }, 60000);
+}, true);
+
 /**
  * Handle a ui_control SSE event — AI-driven UI manipulation.
  * Extracted from the duplicated ui_control + tool_output.ui_event handlers.

static/js/compare/index.js

@@ -21,7 +21,7 @@ import { EVAL_PROMPTS, WAVE_FRAMES,
 import { fetchModels, _persistSelections, _modelDisplayNames, getExcludedModels, setExcludedModels } from './models.js';
 import { showModelSelector, disableToolToggles, restoreToolToggles, _syncToolbarIndicator } from './selector.js?v=20260723compareicon2';
 import { _checkUnprobed, _clearProbeWaves } from './probe.js';
-import { streamToPane, _renderSearchResults, _runSynthForPane, _formatMs, registerStreamActions } from './stream.js';
+import { streamToPane, _renderSearchResults, _runSynthForPane, _formatMs, registerStreamActions } from './stream.js?v=20260819approvalcontrol1';
 import {
   stopAll, stopPane, rerollPane, shufflePanePositions, resetCompare,
   _addPane, _removePane, toggleExpandPane, togglePanePreview, copyPaneResponse,
@@ -1006,11 +1006,16 @@ async function _executeCompare(message) {
     console.error('Compare error:', err);
     if (uiModule) uiModule.showError('Compare failed: ' + err.message);
   } finally {
-    state._streaming = false;
-    _setSendBtn('send');
-    // Re-enable header buttons
-    document.querySelectorAll('#compare-shuffle-btn, #compare-check-btn, #compare-add-btn').forEach(b => {
-      b.disabled = false; b.style.opacity = '0.7'; b.style.pointerEvents = '';
+    // A pane may have started its own ask_user/approval continuation while the
+    // original all-pane Promise was settling. Keep Compare busy until every
+    // pane-owned controller is gone instead of exposing a second broadcast send.
+    const compareStillStreaming = state._abortControllers.some(Boolean);
+    state._streaming = compareStillStreaming;
+    _setSendBtn(compareStillStreaming ? 'stop' : 'send');
+    document.querySelectorAll('#compare-shuffle-btn, #compare-check-btn, #compare-add-btn').forEach((button) => {
+      button.disabled = compareStillStreaming;
+      button.style.opacity = compareStillStreaming ? '0.25' : '0.7';
+      button.style.pointerEvents = compareStillStreaming ? 'none' : '';
     });
   }
 }
@@ -1514,7 +1519,7 @@ async function showShufflePoolEditor() {
 // ────────────────────────────────────────────────────────────────────────────
 
 registerCompareActions({ stopAll, resetCompare });
-registerStreamActions({ rerollPane, autoPreviewHtml: _autoPreviewHtml });
+registerStreamActions({ rerollPane, autoPreviewHtml: _autoPreviewHtml, setSendBtn: _setSendBtn });
 registerPaneActions({ setSendBtn: _setSendBtn, deactivate, streamToPane, renderSearchResults: _renderSearchResults, fetchModels });
 
 // ────────────────────────────────────────────────────────────────────────────

static/js/compare/stream.js

@@ -1,7 +1,7 @@
 // compare/stream.js — SSE streaming to panes
 import state from './state.js';
 import { addFinishBadge } from './vote.js';
-import { getModelCost, safeDisplayImageSrc } from '../chatRenderer.js';
+import { getModelCost, renderAskUserCard, safeDisplayImageSrc } from '../chatRenderer.js?v=20260819approvalcontrol1';
 import markdownModule from '../markdown.js';
 import spinnerModule from '../spinner.js';
 import uiModule from '../ui.js';
@@ -24,11 +24,157 @@ function _safeHttpHref(raw) {
 // ── Lazy-registered functions from compare.js (avoids circular deps) ──
 let _rerollPane = null;
 let _autoPreviewHtml = null;
+let _setSendBtn = null;
 
 /** Register external functions that live in compare.js. */
-function registerStreamActions({ rerollPane, autoPreviewHtml }) {
+function registerStreamActions({ rerollPane, autoPreviewHtml, setSendBtn }) {
   _rerollPane = rerollPane;
   _autoPreviewHtml = autoPreviewHtml;
+  _setSendBtn = setSendBtn;
+}
+
+function _paneSessionIsCurrent(paneIdx, sessionId) {
+  return Boolean(
+    state.isActive
+    && state._paneSessionIds[paneIdx] === sessionId
+    && document.getElementById('cmp-history-' + paneIdx)
+  );
+}
+
+function _setCompareBusy(active) {
+  state._streaming = Boolean(active);
+  if (_setSendBtn) _setSendBtn(active ? 'stop' : 'send');
+  document.querySelectorAll('#compare-shuffle-btn, #compare-check-btn, #compare-add-btn').forEach((button) => {
+    button.disabled = Boolean(active);
+    button.style.opacity = active ? '0.25' : '0.7';
+    button.style.pointerEvents = active ? 'none' : '';
+  });
+}
+
+function _syncCompareBusyFromPanes() {
+  _setCompareBusy((state._abortControllers || []).some(Boolean));
+}
+
+function _appendPaneMessage(hist, role, text) {
+  const message = document.createElement('div');
+  message.className = 'msg ' + (role === 'user' ? 'msg-user' : 'msg-ai');
+  const roleEl = document.createElement('div');
+  roleEl.className = 'role';
+  roleEl.textContent = role === 'user' ? 'You' : 'AI';
+  const body = document.createElement('div');
+  body.className = 'body';
+  body.textContent = text || '';
+  message.appendChild(roleEl);
+  message.appendChild(body);
+  hist.appendChild(message);
+  return message;
+}
+
+function _createPaneContinuationMessage(hist) {
+  const message = _appendPaneMessage(hist, 'assistant', '');
+  const body = message.querySelector('.body');
+  if (spinnerModule) {
+    const spinner = spinnerModule.create('Continuing...', 'right');
+    body.appendChild(spinner.createElement());
+    spinner.start();
+    message._spinner = spinner;
+  }
+  return message;
+}
+
+function _restorePaneAskUserCard(paneIdx, sessionId, submission, originController) {
+  const hist = document.getElementById('cmp-history-' + paneIdx);
+  const restored = _renderPaneAskUserCard(
+    paneIdx,
+    sessionId,
+    submission.payload || {},
+    hist,
+    null,
+    originController,
+  );
+  if (uiModule) {
+    uiModule.showError(
+      restored
+        ? 'This pane is still streaming — choose again once it settles.'
+        : 'Compare pane is still streaming; the choice was not sent.',
+    );
+  }
+  return restored;
+}
+
+function _resumePaneChoiceWhenIdle(paneIdx, sessionId, originController, submission) {
+  if (!_paneSessionIsCurrent(paneIdx, sessionId)) return false;
+
+  const startedAt = Date.now();
+  const resume = () => {
+    if (!_paneSessionIsCurrent(paneIdx, sessionId)) return;
+    const activeController = state._abortControllers[paneIdx];
+    if (activeController === originController) {
+      if (Date.now() - startedAt < 10000) {
+        setTimeout(resume, 25);
+        return;
+      }
+      // The originating stream never released the pane. The card was already
+      // removed when the choice was accepted, so put it back rather than
+      // swallowing a decision the user made.
+      _restorePaneAskUserCard(paneIdx, sessionId, submission, originController);
+      return;
+    }
+    // A reroll/model replacement already owns this pane. Never send the stale
+    // choice into that replacement stream or session UI.
+    if (activeController) return;
+
+    const hist = document.getElementById('cmp-history-' + paneIdx);
+    if (!hist) return;
+    hist.querySelectorAll('.ask-user-card').forEach((card) => card.remove());
+
+    const isApproval = submission.kind === 'tool_approval';
+    const message = isApproval ? '' : String(submission.text || submission.label || '');
+    if (!isApproval) _appendPaneMessage(hist, 'user', message);
+    const aiMessage = _createPaneContinuationMessage(hist);
+    hist.scrollTop = hist.scrollHeight;
+
+    const resumeOptions = { skipBadge: true };
+    if (isApproval) {
+      resumeOptions.toolApproval = {
+        approval_id: String(submission.approval_id || ''),
+        decision: String(submission.decision || '').toLowerCase(),
+      };
+    }
+
+    _setCompareBusy(true);
+    streamToPane(paneIdx, sessionId, message, aiMessage, resumeOptions)
+      .catch((error) => {
+        console.error('Compare pane continuation failed:', error);
+        if (uiModule) uiModule.showError('Compare continuation failed: ' + error.message);
+      })
+      .finally(_syncCompareBusyFromPanes);
+  };
+
+  setTimeout(resume, 0);
+  return true;
+}
+
+function _renderPaneAskUserCard(paneIdx, sessionId, payload, hist, aiMsgEl, originController) {
+  if (!hist || !hist.isConnected || !_paneSessionIsCurrent(paneIdx, sessionId)) return null;
+  if (aiMsgEl && aiMsgEl._spinner) {
+    if (aiMsgEl._spinner.element) aiMsgEl._spinner.destroy();
+    aiMsgEl._spinner = null;
+  }
+  const card = renderAskUserCard(payload, {
+    root: hist,
+    onSubmit: (submission) => _resumePaneChoiceWhenIdle(
+      paneIdx,
+      sessionId,
+      originController,
+      submission,
+    ),
+  });
+  if (card) {
+    card.dataset.comparePane = String(paneIdx);
+    card.dataset.compareSession = String(sessionId);
+  }
+  return card;
 }
 
 /** Format milliseconds as human-readable duration (e.g. "120ms", "1.23s", "4.5s"). */
@@ -164,6 +310,7 @@ async function streamToPane(paneIdx, sessionId, message, aiMsgEl, opts) {
   let metrics = null;
   let timedOut = false;
   let streamOk = false;
+  let awaitingChoice = false;
   let currentToolBlock = null;  // track active agent tool block
   // Idle timeout — abort only if no data is received for this many seconds.
   // Long generations (SVG, big code) are fine as long as the stream stays
@@ -219,6 +366,10 @@ async function streamToPane(paneIdx, sessionId, message, aiMsgEl, opts) {
     const fd = new FormData();
     fd.append('message', message);
     fd.append('session', sessionId);
+    if (opts.toolApproval) {
+      fd.append('tool_approval_id', opts.toolApproval.approval_id || '');
+      fd.append('tool_approval_decision', opts.toolApproval.decision || '');
+    }
 
     // Compare mode determines what tools/features are enabled
     const isAgent = state._compareMode === 'agent';
@@ -322,6 +473,36 @@ async function streamToPane(paneIdx, sessionId, message, aiMsgEl, opts) {
               }
             }
 
+          // ── Pane-local question / approval selector ──
+          } else if (json.type === 'ask_user') {
+            awaitingChoice = true;
+            _renderPaneAskUserCard(
+              paneIdx,
+              sessionId,
+              json.data || {},
+              hist,
+              aiMsgEl,
+              ac,
+            );
+            if (hist) hist.scrollTop = hist.scrollHeight;
+
+          // Deny ends as a tiny resolution-only stream, so replace the
+          // continuation spinner with an explicit pane-local result.
+          } else if (json.type === 'tool_approval_resolved') {
+            if (aiMsgEl._spinner) {
+              if (aiMsgEl._spinner.element) aiMsgEl._spinner.destroy();
+              aiMsgEl._spinner = null;
+            }
+            accumulated = json.decision === 'deny' ? 'Denied.' : 'Approval recorded.';
+            let target = aiMsgEl._textEl;
+            if (!target) {
+              target = document.createElement('div');
+              target.className = 'compare-text-content';
+              aiBody.appendChild(target);
+              aiMsgEl._textEl = target;
+            }
+            target.textContent = accumulated;
+
           // ── Tool start (bash, web search agent tool) ──
           } else if (json.type === 'tool_start') {
             // Finalize any accumulated text before the tool block
@@ -640,19 +821,21 @@ async function streamToPane(paneIdx, sessionId, message, aiMsgEl, opts) {
       // TTFT removed from the header per user request — just show total time.
       _timerEl.textContent = _formatMs(_totalMs);
     }
-    state._abortControllers[paneIdx] = null;
+    if (state._abortControllers[paneIdx] === ac) {
+      state._abortControllers[paneIdx] = null;
+    }
     // Hide stop button, show response action buttons
     const _paneElFinal = document.querySelector(`.compare-pane[data-pane="${paneIdx}"]`);
     if (_paneElFinal) {
       const _stopBtnFinal = _paneElFinal.querySelector('.pane-stop-btn');
       if (_stopBtnFinal) _stopBtnFinal.style.display = 'none';
-      if (accumulated.trim()) {
+      if (!awaitingChoice && accumulated.trim()) {
         _paneElFinal.querySelectorAll('.pane-needs-response').forEach(b => b.style.display = '');
       }
     }
     state._paneMetrics[paneIdx] = metrics;
     state._paneElapsed[paneIdx] = _totalMs;
-    if (!opts.skipBadge) {
+    if (!opts.skipBadge && !awaitingChoice) {
       if (streamOk) {
         state._finishOrder++;
         if (state._parallel) {
@@ -682,7 +865,7 @@ async function streamToPane(paneIdx, sessionId, message, aiMsgEl, opts) {
       }
     }
     // Auto-grade against expected answer — stamps ✓ or ✗ on the pane header.
-    if (streamOk && state._expectedAnswer) {
+    if (streamOk && !awaitingChoice && state._expectedAnswer) {
       _stampGradeBadge(paneIdx, accumulated, state._expectedAnswer);
     }
     // Show copy/reroll buttons now that response exists

static/js/cookbookDownload.js

@@ -15,6 +15,7 @@ let _getPlatform;
 let _serverByVal;
 let _isWindows;
 let _buildEnvPrefix;
+let _psQuote;
 let _buildServeCmd;
 let _detectBackend;
 let _detectToolParser;
@@ -538,7 +539,7 @@ export async function _runModelDownload(panel, model, backend, hostOverride) {
   if (srv.downloadDir) payload.local_dir = srv.downloadDir;
   if (isWin) {
     if (env === 'venv' && envPath) {
-      payload.env_prefix = '& ' + (envPath.endsWith('\\Scripts\\Activate.ps1') ? envPath : envPath + '\\Scripts\\Activate.ps1');
+      payload.env_prefix = '& ' + _psQuote(envPath.endsWith('\\Scripts\\Activate.ps1') ? envPath : envPath + '\\Scripts\\Activate.ps1');
     } else if (env === 'conda' && envPath) {
       payload.env_prefix = 'conda activate ' + envPath;
     }
@@ -652,6 +653,7 @@ export function initDownload(shared) {
   _serverByVal = shared._serverByVal;
   _isWindows = shared._isWindows;
   _buildEnvPrefix = shared._buildEnvPrefix;
+  _psQuote = shared._psQuote;
   _buildServeCmd = shared._buildServeCmd;
   _detectBackend = shared._detectBackend;
   _detectToolParser = shared._detectToolParser;

static/js/cookbookRunning.js

@@ -338,6 +338,7 @@ let _sshPrefix;
 let _getPlatform;
 let _isWindows;
 let _buildEnvPrefix;
+let _psQuote;
 let _loadPresets;
 let _savePresets;
 let _copyText;
@@ -1971,7 +1972,7 @@ export async function _launchServeTask(shortName, repo, cmd, fields, hostOverrid
   let envPrefix = '';
   if (_isWindows()) {
     if (_envState.env === 'venv' && _envState.envPath) {
-      envPrefix = '& ' + (_envState.envPath.endsWith('\\Scripts\\Activate.ps1') ? _envState.envPath : _envState.envPath + '\\Scripts\\Activate.ps1');
+      envPrefix = '& ' + _psQuote(_envState.envPath.endsWith('\\Scripts\\Activate.ps1') ? _envState.envPath : _envState.envPath + '\\Scripts\\Activate.ps1');
     } else if (_envState.env === 'conda' && _envState.envPath) {
       envPrefix = 'conda activate ' + _envState.envPath;
     }
@@ -4402,6 +4403,7 @@ export function initRunning(shared) {
   _getPlatform = shared._getPlatform;
   _isWindows = shared._isWindows;
   _buildEnvPrefix = shared._buildEnvPrefix;
+  _psQuote = shared._psQuote;
   _loadPresets = shared._loadPresets;
   _savePresets = shared._savePresets;
   _copyText = shared._copyText;

static/js/document.js

@@ -9738,6 +9738,11 @@ import { bindMenuDismiss, dismissOrRemove } from './escMenuStack.js';
     const container = document.createElement('div');
     container.style.cssText = 'padding:20px;font-family:sans-serif;font-size:12px;color:#000;background:#fff;line-height:1.6;';
     container.innerHTML = html;
+    // This container is detached, so the document-scoped flush mdToHtml
+    // schedules never sees it. Typeset the deferred math before html2pdf
+    // rasterises, or the PDF gets raw formula source. renderMath() returns
+    // immediately, without loading KaTeX, when there is nothing pending.
+    await markdownModule.renderMath(container);
     const baseName = _getExportBaseName();
     window.html2pdf().set({
       margin: 10,

static/js/emailLibrary.js

@@ -23,6 +23,7 @@ import {
   _tryFoldHintSig, _foldSignature, _SIG_ICON, _QUOTE_ICON,
 } from './emailLibrary/signatureFold.js';
 import { state } from './emailLibrary/state.js';
+import { getSettings } from './appConfig.js';
 import { collapseSidebarToRail } from './modalSnap.js';
 import { emailApiUrl } from './emailShared.js';
 import { bindMenuDismiss, dismissOrRemove } from './escMenuStack.js';
@@ -993,8 +994,7 @@ function _syncEmailReminderBellVisibility(enabled) {
 
 async function _loadEmailReminderBellVisibility() {
   try {
-    const res = await fetch('/api/auth/settings', { credentials: 'same-origin' });
-    const settings = await res.json();
+    const settings = await getSettings();
     _syncEmailReminderBellVisibility(settings.reminder_channel === 'email');
   } catch (_) {
     _syncEmailReminderBellVisibility(false);

static/js/gallery.js

@@ -3,7 +3,7 @@
  */
 
 import uiModule from './ui.js';
-import { openEditor, closeEditor, isEditorOpen } from './galleryEditor.js?v=20260708match1';
+import { loadPanel } from './panels.js';
 import spinnerModule from './spinner.js';
 import { makeWindowDraggable } from './windowDrag.js';
 import { bindMenuDismiss, dismissOrRemove } from './escMenuStack.js';
@@ -15,6 +15,54 @@ const API_BASE = window.location.origin;
 let _open = false;
 let _galleryResizeHandler = null;
 
+// ── Image editor, loaded on first use ──
+// galleryEditor.js plus everything under js/editor/ is 54 modules / 576 KB.
+// It used to be a static import here, so every page load paid for it even
+// though most sessions never touch the Edit tab. The wrappers below keep the
+// three call shapes the rest of this file already uses.
+//
+// closeEditor() and isEditorOpen() stay synchronous on purpose: if the module
+// was never loaded there is no edit session to close, and none can be open.
+let _editorMod = null;
+let _editorLoading = false;
+
+async function _loadEditor() {
+  _editorLoading = true;
+  try {
+    _editorMod = await loadPanel('editor');
+    return _editorMod;
+  } finally {
+    _editorLoading = false;
+  }
+}
+
+async function openEditor(...args) {
+  let mod = _editorMod;
+  if (!mod) {
+    try {
+      mod = await _loadEditor();
+    } catch (e) {
+      // Previously unreachable — a static import either loaded or the whole
+      // page failed. Now it can fail on its own (offline before the panel was
+      // ever cached), so say so instead of doing nothing.
+      console.error('[gallery] image editor failed to load', e);
+      uiModule?.showError?.('Failed to load the image editor');
+      return;
+    }
+  }
+  return mod.openEditor(...args);
+}
+
+function closeEditor(...args) {
+  return _editorMod ? _editorMod.closeEditor(...args) : undefined;
+}
+
+// True while the module is still in flight as well — the gallery-close paths
+// use this to refuse to tear the container down under an edit that is opening.
+function isEditorOpen() {
+  return _editorLoading || (_editorMod ? _editorMod.isEditorOpen() : false);
+}
+
 // Auto-refresh gallery when new image is generated
 window.addEventListener('gallery-refresh', (e) => {
   if (e?.detail?.source === 'chat-upload' && _sort !== 'recent') {

static/js/keyboard-shortcuts.js

@@ -3,6 +3,7 @@
 // ============================================
 
 import { IS_MAC, isAltGrEvent } from './platform.js';
+import { getSettings } from './appConfig.js';
 
 const _defaultKeybinds = {
   search: 'ctrl+k', toggle_sidebar: 'ctrl+alt+b', new_session: 'ctrl+alt+n',
@@ -56,8 +57,7 @@ export function initKeyboardShortcuts(modules) {
   window._odysseusKeybinds = { ..._defaultKeybinds };
 
   // Load saved keybinds
-  fetch('/api/auth/settings', { credentials: 'same-origin' })
-    .then(r => r.json())
+  getSettings()
     .then(s => { if (s.keybinds) window._odysseusKeybinds = { ..._defaultKeybinds, ...s.keybinds }; })
     .catch(() => {});
 

static/js/markdown.js

@@ -10,6 +10,127 @@ import { replaceEmojiShortcodes, hasEmojiShortcode } from './emojiShortcodes.js'
 
 var escapeHtml = uiModule.esc;
 
+// Mermaid and KaTeX are vendored under /static/lib and fetched on first use.
+// Loading them from <head> cost every session ~985 KB on the wire even though
+// most chats never contain a diagram or a formula. Both loaders memoise the
+// *promise* rather than the resolved library, so concurrent callers share one
+// fetch and a double trigger cannot start two loads. A failed load clears the
+// memo so the next diagram/formula retries instead of being poisoned forever.
+const MERMAID_SRC = '/static/lib/mermaid.min.js';
+const KATEX_SRC = '/static/lib/katex/katex.min.js';
+const KATEX_CSS = '/static/lib/katex/katex.min.css';
+// Marks math emitted before KaTeX finished loading; renderMath() swaps these
+// for typeset output. The source stays as readable text inside the span, so a
+// load that never completes degrades to plain text rather than to nothing.
+const MATH_PENDING_CLASS = 'ody-math-pending';
+
+// KaTeX has no entity syntax: it reads a bare "&" as an alignment marker and
+// errors out on anything that is not a valid column break, so "a &lt; b" comes
+// back as a red .katex-error instead of a formula. mdToHtml escapes the whole
+// string before the math pass, which leaves two spellings of the same
+// character at the delimiters — a typed "<" arrives as "&lt;", while a typed
+// "&lt;" arrives as "&amp;lt;" — and both have to reach KaTeX as "<".
+//
+// One alternation, longest form first, so nothing this writes is scanned
+// again. Chained .replace() calls cannot do it: unescaping "&amp;" first lets
+// the next pass eat the "&lt;" it just produced (the double-unescape CodeQL
+// flags), and unescaping it last leaves the entity spelling intact and breaks
+// the render. The code-block pass upstream keeps its chained order on purpose
+// — Markdown does not decode entities inside code, so "&lt;" there is meant to
+// stay visible.
+const MATH_SOURCE_ENTITY_RE = /&amp;(?:lt|gt|amp|quot|#39);|&lt;|&gt;|&amp;/g;
+const MATH_SOURCE_ENTITIES = {
+  '&amp;lt;': '<',
+  '&amp;gt;': '>',
+  '&amp;amp;': '&',
+  '&amp;quot;': '"',
+  '&amp;#39;': "'",
+  '&lt;': '<',
+  '&gt;': '>',
+  '&amp;': '&',
+};
+
+function decodeMathSource(text) {
+  return String(text).replace(MATH_SOURCE_ENTITY_RE, (entity) => MATH_SOURCE_ENTITIES[entity]);
+}
+
+let _mermaidPromise = null;
+let _katexPromise = null;
+let _mathFlushScheduled = false;
+
+function _loadScript(src) {
+  return new Promise((resolve, reject) => {
+    const script = document.createElement('script');
+    script.src = src;
+    script.addEventListener('load', () => resolve(), { once: true });
+    script.addEventListener('error', () => reject(new Error('Failed to load ' + src)), { once: true });
+    document.head.appendChild(script);
+  });
+}
+
+function _loadStylesheet(href) {
+  // Resolves either way: without the stylesheet KaTeX still produces correct
+  // markup, just unstyled, which beats failing the whole math render.
+  return new Promise((resolve) => {
+    const link = document.createElement('link');
+    link.rel = 'stylesheet';
+    link.href = href;
+    link.addEventListener('load', () => resolve(), { once: true });
+    link.addEventListener('error', () => resolve(), { once: true });
+    document.head.appendChild(link);
+  });
+}
+
+/**
+ * Load Mermaid on first use and initialize it once.
+ */
+export function ensureMermaid() {
+  return (_mermaidPromise ??= _loadScript(MERMAID_SRC)
+    .then(() => {
+      if (!window.mermaid) throw new Error('mermaid global missing after load');
+      window.mermaid.initialize({ startOnLoad: false, theme: 'dark', securityLevel: 'loose' });
+      return window.mermaid;
+    })
+    .catch((err) => {
+      _mermaidPromise = null;
+      throw err;
+    }));
+}
+
+/**
+ * Load KaTeX (script + stylesheet) on first use.
+ */
+export function ensureKatex() {
+  return (_katexPromise ??= Promise.all([_loadScript(KATEX_SRC), _loadStylesheet(KATEX_CSS)])
+    .then(() => {
+      if (!window.katex) throw new Error('katex global missing after load');
+      return window.katex;
+    })
+    .catch((err) => {
+      _katexPromise = null;
+      throw err;
+    }));
+}
+
+// mdToHtml() is synchronous and its callers insert the returned string into the
+// DOM themselves, so the placeholders are usually not attached yet when this
+// fires. Loading first and scanning afterwards covers that gap: by the time
+// KaTeX is in, the caller's innerHTML assignment has long since happened.
+//
+// setTimeout, not requestAnimationFrame: this has nothing to do with paint, and
+// rAF is throttled to a stop in a background tab (and never fires at all in a
+// headless browser), which would leave math untypeset until the tab is focused.
+function _scheduleMathFlush() {
+  if (_mathFlushScheduled) return;
+  _mathFlushScheduled = true;
+  setTimeout(() => {
+    _mathFlushScheduled = false;
+    ensureKatex()
+      .then(() => renderMath(document))
+      .catch((e) => console.warn('KaTeX load error:', e));
+  }, 0);
+}
+
 function safeLinkUrl(rawUrl) {
   const url = String(rawUrl || '').trim();
   if (url.startsWith('#')) {
@@ -631,49 +752,45 @@ export function mdToHtml(src, opts) {
 
   // KaTeX math rendering (after code blocks are extracted, so math in code is safe)
   const mathBlocks = [];
-  if (window.katex) {
-    // Display math: \[ ... \]  — GPT-style delimiter (gpt-5.x, Claude, etc.).
-    // Handle before $$/$ so all common delimiters render.
-    s = s.replace(/\\\[([\s\S]*?)\\\]/g, (match, math) => {
-      try {
-        const raw = math.replace(/&amp;/g, '&').replace(/&lt;/g, '<').replace(/&gt;/g, '>');
-        const placeholder = `___MATH_BLOCK_${mathBlocks.length}___`;
-        mathBlocks.push(katex.renderToString(raw.trim(), { displayMode: true, throwOnError: false }));
-        return placeholder;
-      } catch (e) { return match; }
-    });
-    // Inline math: \( ... \)  — GPT-style inline delimiter. Single-line only
-    // ([^\n]) so a stray escaped paren in prose can't swallow across lines.
-    s = s.replace(/\\\(([^\n]*?)\\\)/g, (match, math) => {
-      try {
-        const raw = math.replace(/&amp;/g, '&').replace(/&lt;/g, '<').replace(/&gt;/g, '>');
-        const placeholder = `___MATH_BLOCK_${mathBlocks.length}___`;
-        mathBlocks.push(katex.renderToString(raw.trim(), { displayMode: false, throwOnError: false }));
-        return placeholder;
-      } catch (e) { return match; }
-    });
-    // Display math: $$...$$
-    s = s.replace(/\$\$([\s\S]*?)\$\$/g, (match, math) => {
-      try {
-        const raw = math.replace(/&amp;/g, '&').replace(/&lt;/g, '<').replace(/&gt;/g, '>');
-        const placeholder = `___MATH_BLOCK_${mathBlocks.length}___`;
-        mathBlocks.push(katex.renderToString(raw.trim(), { displayMode: true, throwOnError: false }));
-        return placeholder;
-      } catch (e) { return match; }
-    });
-    // Inline math: $...$ — single line only, and Pandoc-style delimiter rules so
-    // currency doesn't render as math ("$5 to $10"): the opening $ must be
-    // immediately followed by a non-space, the closing $ must be immediately
-    // preceded by a non-space and not followed by a digit.
-    s = s.replace(/(?<![\$\d])\$(?!\$)(?=\S)([^\$\n]+?)(?<=\S)\$(?!\$|\d)/g, (match, math) => {
-      try {
-        const raw = math.replace(/&amp;/g, '&').replace(/&lt;/g, '<').replace(/&gt;/g, '>');
-        const placeholder = `___MATH_BLOCK_${mathBlocks.length}___`;
-        mathBlocks.push(katex.renderToString(raw.trim(), { displayMode: false, throwOnError: false }));
-        return placeholder;
-      } catch (e) { return match; }
-    });
-  }
+  let sawPendingMath = false;
+
+  // Typeset straight away when KaTeX is already in, otherwise bank the source in
+  // an inert placeholder for renderMath() to swap once the library lands.
+  const pushMath = (math, displayMode) => {
+    const raw = decodeMathSource(math).trim();
+    const placeholder = `___MATH_BLOCK_${mathBlocks.length}___`;
+    if (window.katex) {
+      mathBlocks.push(katex.renderToString(raw, { displayMode, throwOnError: false }));
+    } else {
+      sawPendingMath = true;
+      mathBlocks.push(`<span class="${MATH_PENDING_CLASS}" data-display="${displayMode}">${escapeHtml(raw)}</span>`);
+    }
+    return placeholder;
+  };
+
+  // Display math: \[ ... \]  — GPT-style delimiter (gpt-5.x, Claude, etc.).
+  // Handle before $$/$ so all common delimiters render.
+  s = s.replace(/\\\[([\s\S]*?)\\\]/g, (match, math) => {
+    try { return pushMath(math, true); } catch (e) { return match; }
+  });
+  // Inline math: \( ... \)  — GPT-style inline delimiter. Single-line only
+  // ([^\n]) so a stray escaped paren in prose can't swallow across lines.
+  s = s.replace(/\\\(([^\n]*?)\\\)/g, (match, math) => {
+    try { return pushMath(math, false); } catch (e) { return match; }
+  });
+  // Display math: $$...$$
+  s = s.replace(/\$\$([\s\S]*?)\$\$/g, (match, math) => {
+    try { return pushMath(math, true); } catch (e) { return match; }
+  });
+  // Inline math: $...$ — single line only, and Pandoc-style delimiter rules so
+  // currency doesn't render as math ("$5 to $10"): the opening $ must be
+  // immediately followed by a non-space, the closing $ must be immediately
+  // preceded by a non-space and not followed by a digit.
+  s = s.replace(/(?<![\$\d])\$(?!\$)(?=\S)([^\$\n]+?)(?<=\S)\$(?!\$|\d)/g, (match, math) => {
+    try { return pushMath(math, false); } catch (e) { return match; }
+  });
+
+  if (sawPendingMath) _scheduleMathFlush();
 
   // Handle pipe tables
   s = s.replace(/(?:^|\n)([^\n]*\|[^\n]*\|[^\n]*)(?:\n([^\n]*\|[^\n]*\|[^\n]*))*/g, (table) => {
@@ -826,19 +943,47 @@ export function renderContent(content) {
 }
 
 /**
- * Initialize any unprocessed Mermaid diagrams in a container (or whole document)
+ * Initialize any unprocessed Mermaid diagrams in a container (or whole document).
+ * Returns a promise so callers can await the (lazy) library load if they need to.
  */
 export function renderMermaid(container) {
-  if (!window.mermaid) return;
-  initMermaid();
   const target = container || document;
-  const pending = target.querySelectorAll('pre.mermaid:not([data-processed])');
-  if (pending.length === 0) return;
-  try {
-    window.mermaid.run({ nodes: pending });
-  } catch (e) {
-    console.warn('Mermaid render error:', e);
-  }
+  if (!target || typeof target.querySelectorAll !== 'function') return Promise.resolve();
+  // Cheap pre-check: no fence on the page means Mermaid is never fetched.
+  if (target.querySelectorAll('pre.mermaid:not([data-processed])').length === 0) return Promise.resolve();
+  return ensureMermaid()
+    .then((mermaid) => {
+      // Re-query after the load: during streaming the renderer replaces the
+      // message body repeatedly, so the nodes seen before the fetch are stale.
+      const nodes = [...target.querySelectorAll('pre.mermaid:not([data-processed])')]
+        .filter((node) => node.isConnected);
+      if (nodes.length === 0) return;
+      return mermaid.run({ nodes });
+    })
+    .catch((e) => { console.warn('Mermaid render error:', e); });
+}
+
+/**
+ * Typeset any math that mdToHtml() had to defer because KaTeX was not loaded
+ * yet. Once KaTeX is in, mdToHtml() renders inline and this finds nothing.
+ */
+export function renderMath(container) {
+  const target = container || document;
+  if (!target || typeof target.querySelectorAll !== 'function') return Promise.resolve();
+  if (target.querySelectorAll('.' + MATH_PENDING_CLASS).length === 0) return Promise.resolve();
+  return ensureKatex()
+    .then((katex) => {
+      target.querySelectorAll('.' + MATH_PENDING_CLASS).forEach((el) => {
+        const displayMode = el.getAttribute('data-display') === 'true';
+        try {
+          el.outerHTML = katex.renderToString(el.textContent || '', { displayMode, throwOnError: false });
+        } catch (e) {
+          // Leave the source visible — readable, just not typeset.
+          el.classList.remove(MATH_PENDING_CLASS);
+        }
+      });
+    })
+    .catch((e) => { console.warn('KaTeX render error:', e); });
 }
 
 const markdownModule = {
@@ -853,20 +998,14 @@ const markdownModule = {
   extractThinkingBlocks,
   normalizeThinkingMarkup,
   startsWithReasoningPrefix,
-  renderMermaid
+  renderMermaid,
+  renderMath,
+  ensureMermaid,
+  ensureKatex
 };
 
 export default markdownModule;
 
-// Mermaid is loaded async so it cannot delay the app shell.
-function initMermaid() {
-  if (!window.mermaid || window.__odysseusMermaidReady) return;
-  window.mermaid.initialize({ startOnLoad: false, theme: 'dark', securityLevel: 'loose' });
-  window.__odysseusMermaidReady = true;
-}
-window.odysseusInitMermaid = initMermaid;
-initMermaid();
-
 // Persist which thinking sections were expanded across page refreshes.
 // IDs are render-generated (Date.now-based) so we key by a stable hash of
 // the inner text content instead — same content reproduces the same hash on

static/js/panels.js

@@ -0,0 +1,53 @@
+/**
+ * Panel loader registry — imports the modules behind a feature panel the
+ * first time that panel is actually used.
+ *
+ * The panels already populate themselves on open (each fetches its own data).
+ * They were just not *loaded* on demand: every one of them sat on the critical
+ * path of every page load, opened or not.
+ *
+ * A panel only belongs here once it has been checked for import-time side
+ * effects that something outside the panel depends on at startup. Entries get
+ * added one panel at a time, not in bulk.
+ *
+ * Note the service worker still precaches these modules (PANEL_PRECACHE in
+ * sw.js) — they are off the critical path, not off the offline manifest.
+ */
+
+const LOADERS = {
+  editor: () => import('./galleryEditor.js'),
+};
+
+/**
+ * Build a memoising loader over a name -> import-thunk map. Exported so the
+ * behaviour can be tested without pulling a real panel's module graph in.
+ */
+export function createPanelLoader(loaders) {
+  const cache = new Map();
+  return function load(name) {
+    const cached = cache.get(name);
+    if (cached) return cached;
+    const loader = loaders[name];
+    if (!loader) throw new Error(`loadPanel: unknown panel "${name}"`);
+    // A failed load (offline, 404, syntax error) is not memoised — caching the
+    // rejection would leave the panel broken for the rest of the session even
+    // after the network came back.
+    const pending = Promise.resolve().then(loader).catch((err) => {
+      cache.delete(name);
+      throw err;
+    });
+    cache.set(name, pending);
+    return pending;
+  };
+}
+
+/**
+ * Load a panel's module, once. Returns the same promise on every call for the
+ * same panel; throws for a name that is not registered.
+ */
+export const loadPanel = createPanelLoader(LOADERS);
+
+/** Registered panel names — the registry is the list, not a second copy of it. */
+export function panelNames() {
+  return Object.keys(LOADERS);
+}

static/js/search.js

@@ -4,20 +4,21 @@
  * Search settings management — reads active provider from admin settings.
  */
 
-let API_BASE = '';
+import { getSettings, invalidateSettings } from './appConfig.js';
+
 let _provider = 'searxng';
 let _loaded = false;
 
-export function init(apiBase) {
-  API_BASE = apiBase;
+// No API base parameter any more: the settings request lives in appConfig.js and
+// resolves against the document origin, which is exactly what API_BASE held.
+export function init() {
   // Fetch provider on init so it's ready when chat needs it
   _fetchProvider();
 }
 
 async function _fetchProvider() {
   try {
-    const res = await fetch((API_BASE || '') + '/api/auth/settings', { credentials: 'same-origin' });
-    const s = await res.json();
+    const s = await getSettings();
     _provider = s.search_provider || 'searxng';
     _loaded = true;
   } catch (e) { /* keep default */ }
@@ -39,6 +40,9 @@ export function getProviderLabel() {
 
 /** Re-fetch after admin saves new settings */
 export function refresh() {
+  // Drop the shared snapshot first: the point of this call is to observe the
+  // settings that were just written, so it must not be served from cache.
+  invalidateSettings();
   _fetchProvider();
 }
 

static/js/settings.js

@@ -26,11 +26,37 @@ import { sortModelIds } from './modelSort.js';
 import { providerLogo } from './providers.js';
 import { isAltGrEvent } from './platform.js';
 import { bindMenuDismiss } from './escMenuStack.js';
+import { invalidateSettings } from './appConfig.js';
 
 let initialized = false;
 let modalEl = null;
 let _authPolicy = { password_min_length: 8 };
 
+/**
+ * POST a settings patch, then drop the shared snapshot in appConfig.js.
+ *
+ * Every write in this file goes through here so no save path can forget the
+ * invalidation — a stale settings object served for the rest of the session is
+ * a worse bug than the duplicate fetches the cache removes. The invalidation is
+ * in a `finally` because a request that throws on the way back may still have
+ * been applied server-side.
+ *
+ * Reads in this file deliberately stay direct fetches: this panel is the writer
+ * and edits what it reads, so it must see the authoritative state, not a cache.
+ */
+async function _postSettings(body) {
+  try {
+    return await fetch('/api/auth/settings', {
+      method: 'POST',
+      credentials: 'same-origin',
+      headers: { 'Content-Type': 'application/json' },
+      body: JSON.stringify(body),
+    });
+  } finally {
+    invalidateSettings();
+  }
+}
+
 const el = byId;
 function esc(s) { return uiModule.esc(s); }
 function safeRasterDataUrl(raw) {
@@ -276,10 +302,7 @@ function _bindFallbackWidget(opts) {
     var body = {};
     body[settingKey] = clean;
     try {
-      await fetch('/api/auth/settings', { method: 'POST', credentials: 'same-origin',
-        headers: { 'Content-Type': 'application/json' },
-        body: JSON.stringify(body)
-      });
+      await _postSettings(body);
     } catch (e) { console.warn('[fallback] save failed for ' + settingKey, e); }
   }
 
@@ -389,12 +412,9 @@ async function initDefaultChat() {
 
   async function saveDefault() {
     try {
-      await fetch('/api/auth/settings', { method: 'POST', credentials: 'same-origin',
-        headers: { 'Content-Type': 'application/json' },
-        body: JSON.stringify({
-          default_endpoint_id: epSel.value,
-          default_model: modelSel.value
-        })
+      await _postSettings({
+        default_endpoint_id: epSel.value,
+        default_model: modelSel.value
       });
       msg.textContent = 'Saved'; msg.style.color = 'var(--fg)';
       setTimeout(function() { msg.textContent = ''; }, 2000);
@@ -449,12 +469,9 @@ async function initUtilityModel() {
   // no toggle, "—" means "unset, use chat").
   async function saveUtility() {
     try {
-      await fetch('/api/auth/settings', { method: 'POST', credentials: 'same-origin',
-        headers: { 'Content-Type': 'application/json' },
-        body: JSON.stringify({
-          utility_endpoint_id: epSel.value || '',
-          utility_model: modelSel.value || ''
-        })
+      await _postSettings({
+        utility_endpoint_id: epSel.value || '',
+        utility_model: modelSel.value || ''
       });
       msg.textContent = 'Saved'; msg.style.color = 'var(--fg)';
       setTimeout(function() { msg.textContent = ''; }, 1500);
@@ -547,10 +564,7 @@ async function initTeacherModel() {
         spec = ep ? (modelSel.value + '@' + ep.name) : modelSel.value;
       }
       var enabled = enabledToggle ? !!enabledToggle.checked : false;
-      await fetch('/api/auth/settings', { method: 'POST', credentials: 'same-origin',
-        headers: { 'Content-Type': 'application/json' },
-        body: JSON.stringify({ teacher_enabled: enabled, teacher_model: spec })
-      });
+      await _postSettings({ teacher_enabled: enabled, teacher_model: spec });
       msg.textContent = enabled ? (spec ? 'Saved' : 'Pick an endpoint + model') : 'Disabled';
       msg.style.color = enabled && !spec ? 'var(--red)' : 'var(--fg)';
       setTimeout(function() { msg.textContent = ''; }, 2000);
@@ -625,8 +639,7 @@ async function initImageSettings() {
 
   async function saveSettings() {
     try {
-      const res = await fetch('/api/auth/settings', { method: 'POST', credentials: 'same-origin', headers: { 'Content-Type': 'application/json' },
-        body: JSON.stringify({ image_gen_enabled: enabledToggle ? enabledToggle.checked : false, image_model: modelSel.value, image_quality: qualSel.value }) });
+      const res = await _postSettings({ image_gen_enabled: enabledToggle ? enabledToggle.checked : false, image_model: modelSel.value, image_quality: qualSel.value });
       if (!res.ok) throw new Error(await res.text().catch(() => `HTTP ${res.status}`));
       msg.textContent = 'Saved'; msg.style.color = 'var(--fg)'; setTimeout(() => { msg.textContent = ''; }, 2000);
     } catch (e) { msg.textContent = 'Failed to save'; msg.style.color = 'var(--red)'; }
@@ -700,8 +713,7 @@ async function initVisionSettings() {
 
   async function saveSettings() {
     try {
-      await fetch('/api/auth/settings', { method: 'POST', credentials: 'same-origin', headers: { 'Content-Type': 'application/json' },
-        body: JSON.stringify({ vision_enabled: enabledToggle ? enabledToggle.checked : true, vision_model: vlSel.value }) });
+      await _postSettings({ vision_enabled: enabledToggle ? enabledToggle.checked : true, vision_model: vlSel.value });
       msg.textContent = 'Saved'; msg.style.color = 'var(--fg)'; setTimeout(() => { msg.textContent = ''; }, 2000);
     } catch (e) { msg.textContent = 'Failed to save'; msg.style.color = 'var(--red)'; }
   }
@@ -782,8 +794,7 @@ async function initTtsSettings() {
 
   async function saveTTS() {
     try {
-      await fetch('/api/auth/settings', { method: 'POST', credentials: 'same-origin', headers: { 'Content-Type': 'application/json' },
-        body: JSON.stringify({ tts_enabled: ttsEnabledToggle ? ttsEnabledToggle.checked : true, tts_provider: provSel.value, tts_model: getModel() || 'tts-1', tts_voice: getVoice() || 'alloy', tts_speed: speedSelect.value || '1' }) });
+      await _postSettings({ tts_enabled: ttsEnabledToggle ? ttsEnabledToggle.checked : true, tts_provider: provSel.value, tts_model: getModel() || 'tts-1', tts_voice: getVoice() || 'alloy', tts_speed: speedSelect.value || '1' });
       ttsMsg.textContent = 'Saved'; ttsMsg.style.color = 'var(--fg)'; setTimeout(() => { ttsMsg.textContent = ''; }, 2000);
       if (window.aiTTSManager) window.aiTTSManager.checkAvailability();
     } catch (e) { ttsMsg.textContent = 'Failed to save'; ttsMsg.style.color = 'var(--red)'; }
@@ -944,9 +955,7 @@ async function initSttSettings() {
   async function saveSTT() {
     try {
       var enabled = sttEnabledToggle ? sttEnabledToggle.checked : false;
-      await fetch('/api/auth/settings', { method: 'POST', credentials: 'same-origin',
-        headers: { 'Content-Type': 'application/json' },
-        body: JSON.stringify({ stt_enabled: enabled, stt_provider: provSel.value, stt_model: getModel() || 'base', stt_language: langInput.value.trim() }) });
+      await _postSettings({ stt_enabled: enabled, stt_provider: provSel.value, stt_model: getModel() || 'base', stt_language: langInput.value.trim() });
       sttMsg.textContent = 'Saved'; sttMsg.style.color = 'var(--fg)'; setTimeout(() => { sttMsg.textContent = ''; }, 2000);
       // Notify voiceRecorder of effective provider and update send button icon
       if (window.voiceRecorderModule) window.voiceRecorderModule._sttProvider = effectiveProvider();
@@ -1102,10 +1111,7 @@ async function initSearchSettings() {
         payload[kf] = keyInput.value.trim();
         _settings[kf] = keyInput.value.trim();
       }
-      await fetch('/api/auth/settings', { method: 'POST', credentials: 'same-origin',
-        headers: { 'Content-Type': 'application/json' },
-        body: JSON.stringify(payload)
-      });
+      await _postSettings(payload);
       msg.textContent = 'Saved'; msg.style.color = 'var(--fg)';
       setTimeout(refreshStatus, 2000);
       if (searchModule && searchModule.refresh) searchModule.refresh();
@@ -1257,11 +1263,7 @@ async function initSearchSettings() {
   async function _saveFallbackChain(chain) {
     _settings.search_fallback_chain = chain;
     try {
-      await fetch('/api/auth/settings', {
-        method: 'POST', credentials: 'same-origin',
-        headers: { 'Content-Type': 'application/json' },
-        body: JSON.stringify({ search_fallback_chain: chain }),
-      });
+      await _postSettings({ search_fallback_chain: chain });
       msg.textContent = 'Saved'; msg.style.color = 'var(--fg)';
       setTimeout(refreshStatus, 2000);
     } catch (e) { msg.textContent = 'Failed to save'; msg.style.color = 'var(--red)'; }
@@ -1425,10 +1427,7 @@ async function initResearchSettings() {
       }
     }
     try {
-      await fetch('/api/auth/settings', { method: 'POST', credentials: 'same-origin',
-        headers: { 'Content-Type': 'application/json' },
-        body: JSON.stringify(payload)
-      });
+      await _postSettings(payload);
       msg.textContent = 'Saved'; msg.style.color = 'var(--fg)';
       setTimeout(showStatus, 2000);
     } catch (e) { msg.textContent = 'Failed to save'; msg.style.color = 'var(--red)'; }
@@ -1492,10 +1491,7 @@ async function initResearchSearchSettings() {
 
   async function saveResearchSearch() {
     try {
-      await fetch('/api/auth/settings', { method: 'POST', credentials: 'same-origin',
-        headers: { 'Content-Type': 'application/json' },
-        body: JSON.stringify({ research_search_provider: searchSel.value })
-      });
+      await _postSettings({ research_search_provider: searchSel.value });
       msg.textContent = 'Saved'; msg.style.color = 'var(--fg)';
       setTimeout(function() { msg.textContent = ''; }, 2000);
     } catch (e) { msg.textContent = 'Failed to save'; msg.style.color = 'var(--red)'; }
@@ -1537,10 +1533,7 @@ async function initAgentSettings() {
     if (rounds != null) payload.agent_max_rounds = rounds;
     if (supInput) payload.agent_supervisor_ladder = !!supInput.checked;
     try {
-      await fetch('/api/auth/settings', { method: 'POST', credentials: 'same-origin',
-        headers: { 'Content-Type': 'application/json' },
-        body: JSON.stringify(payload)
-      });
+      await _postSettings(payload);
       msg.textContent = (tools > 0 ? 'Limit: ' + tools + ' tool calls' : 'Unlimited tool calls') +
         (rounds != null ? ' · ' + rounds + ' steps/message' : '') +
         (supInput && supInput.checked ? ' · supervisor on' : '');
@@ -1935,11 +1928,7 @@ async function initShortcuts() {
 
   async function saveKeybinds() {
     try {
-      await fetch('/api/auth/settings', {
-        method: 'POST', credentials: 'same-origin',
-        headers: { 'Content-Type': 'application/json' },
-        body: JSON.stringify({ keybinds }),
-      });
+      await _postSettings({ keybinds });
       // Update global keybinds so they take effect immediately
       window._odysseusKeybinds = keybinds;
       if (uiModule && uiModule.showToast) uiModule.showToast('Shortcut saved');
@@ -2232,11 +2221,7 @@ async function initReminderSettings() {
       pubDebounce = setTimeout(async () => {
         try {
           const val = pubUrlIn.value.trim().replace(/\/+$/, '');
-          await fetch('/api/auth/settings', {
-            method: 'POST', credentials: 'same-origin',
-            headers: { 'Content-Type': 'application/json' },
-            body: JSON.stringify({ app_public_url: val }),
-          });
+          await _postSettings({ app_public_url: val });
           if (pubUrlMsg) {
             pubUrlMsg.textContent = val ? 'Saved' : 'Cleared (deep-links disabled)';
             pubUrlMsg.style.color = 'var(--green,#50fa7b)';
@@ -2534,12 +2519,7 @@ async function initReminderSettings() {
 
   async function save(patch) {
     try {
-      await fetch('/api/auth/settings', {
-        method: 'POST',
-        credentials: 'same-origin',
-        headers: { 'Content-Type': 'application/json' },
-        body: JSON.stringify(patch),
-      });
+      await _postSettings(patch);
     } catch (e) { console.warn('Failed to save reminder settings', e); }
   }
 

static/js/slashCommands.js

@@ -22,6 +22,7 @@ import settingsModule from './settings.js';
 import cookbookModule from './cookbook.js';
 import { EVAL_PROMPTS } from './compare/index.js';
 import { PROVIDER_DEVICE_FLOWS, formatDeviceFlowError, runProviderDeviceFlow } from './providerDeviceFlow.js';
+import { getSettings } from './appConfig.js';
 
 // ── Module state ──────────────────────────────────────────────────────
 
@@ -5220,8 +5221,7 @@ async function _cmdShortcuts(args, ctx) {
   };
 
   try {
-    const res = await fetch(`${API_BASE}/api/auth/settings`, { credentials: 'same-origin' });
-    const settings = await res.json();
+    const settings = await getSettings();
     if (settings.keybinds) {
       keybinds = { ...keybinds, ...settings.keybinds };
     }

static/js/tasks.js

@@ -10,6 +10,7 @@ import { topPortalZ } from './toolWindowZOrder.js';
 import { sortModelIds } from './modelSort.js';
 import { ordinalSuffix } from './util/ordinal.js';
 import { bindMenuDismiss, dismissOrRemove } from './escMenuStack.js';
+import { getSettings, invalidateSettings } from './appConfig.js';
 
 const API_BASE = window.location.origin;
 let _open = false;
@@ -214,31 +215,31 @@ async function _fetchActions() {
   return _builtinActions;
 }
 
-let _urgentEmailSettings = null;
 async function _fetchUrgentEmailSettings() {
-  if (_urgentEmailSettings) return _urgentEmailSettings;
   try {
-    const res = await fetch('/api/auth/settings', { credentials: 'same-origin' });
-    _urgentEmailSettings = await res.json();
+    return await getSettings();
   } catch (e) {
-    _urgentEmailSettings = { urgent_email_prompt: '' };
+    return { urgent_email_prompt: '' };
   }
-  return _urgentEmailSettings;
 }
 
 async function _saveUrgentEmailSettings(prompt) {
-  _urgentEmailSettings = {
-    ...(_urgentEmailSettings || {}),
-    urgent_email_prompt: prompt || '',
-  };
-  await fetch('/api/auth/settings', {
-    method: 'POST',
-    credentials: 'same-origin',
-    headers: { 'Content-Type': 'application/json' },
-    body: JSON.stringify({
-      urgent_email_prompt: prompt || '',
-    }),
-  });
+  try {
+    await fetch('/api/auth/settings', {
+      method: 'POST',
+      credentials: 'same-origin',
+      headers: { 'Content-Type': 'application/json' },
+      body: JSON.stringify({
+        urgent_email_prompt: prompt || '',
+      }),
+    });
+  } finally {
+    // The shared snapshot still carries the old prompt — drop it so the next
+    // read (here or in any other module) sees what was just written. In a
+    // `finally` because a request that throws on the way back may still have
+    // been applied.
+    invalidateSettings();
+  }
 }
 
 const _EMAIL_ACCOUNT_ACTIONS = new Set([

static/js/tts-ai.js

@@ -1,6 +1,8 @@
 // static/js/tts-ai.js
 // AI Text-to-Speech Module — supports server TTS and browser Web Speech API
 
+import { getSettings } from './appConfig.js';
+
 class AITTSManager {
     constructor() {
         this.currentAudio = null;
@@ -30,10 +32,11 @@ class AITTSManager {
 
     async checkAvailability() {
         try {
-            // Check user setting first — if TTS is disabled in settings, don't show buttons
+            // Check user setting first — if TTS is disabled in settings, don't show buttons.
+            // settings.js re-calls this right after saving TTS settings; it invalidates
+            // the shared cache before doing so, so this still sees the new value.
             try {
-                const settingsRes = await fetch('/api/auth/settings', { credentials: 'same-origin' });
-                const settings = await settingsRes.json();
+                const settings = await getSettings();
                 if (settings.tts_enabled === false) {
                     this.available = false;
                     this._provider = 'disabled';

static/sw.js

@@ -7,16 +7,42 @@
 //   - Other static assets (images/fonts/libs): cache-first with bg refresh.
 //   - API / non-GET: never cached.
 // Bump CACHE_NAME whenever the precache list or SW logic changes.
-const CACHE_NAME = 'odysseus-v376-settings-title-icons';
+const CACHE_NAME = 'odysseus-v380-shared-config-image-editor-lazy-katex-mermaid';
 
-// Core shell precached on install so repeat opens are instant without any
-// network wait. Keep this list in sync with the <script type="module"> tags
-// and <link rel="stylesheet"> in index.html.
+// KaTeX resolves these from its own stylesheet, so caching the CSS without them
+// gives offline math fallback glyphs instead of proper typesetting.
+const KATEX_FONTS = [
+  'AMS-Regular', 'Caligraphic-Bold', 'Caligraphic-Regular',
+  'Fraktur-Bold', 'Fraktur-Regular',
+  'Main-Bold', 'Main-BoldItalic', 'Main-Italic', 'Main-Regular',
+  'Math-BoldItalic', 'Math-Italic',
+  'SansSerif-Bold', 'SansSerif-Italic', 'SansSerif-Regular',
+  'Script-Regular',
+  'Size1-Regular', 'Size2-Regular', 'Size3-Regular', 'Size4-Regular',
+  'Typewriter-Regular',
+].map(name => `/static/lib/katex/fonts/KaTeX_${name}.woff2`);
+
+
+// Two lists, two jobs — they are no longer the same set and must not be
+// "resynced" back into one:
+//
+//   PRECACHE       = the app shell. Mirrors the <script type="module"> tags
+//                    and <link rel="stylesheet"> in index.html — i.e. what
+//                    loads before first paint.
+//   PANEL_PRECACHE = modules that index.html deliberately does NOT load,
+//                    because js/panels.js imports them on first use. They are
+//                    off the critical path, not out of the offline manifest:
+//                    without them here, a panel the user never opened while
+//                    online could not open offline at all.
+//
+// Both are fetched at install time, in the background. Entries must match the
+// exact URL the browser requests, query string included.
 const PRECACHE = [
   '/',
   '/static/style.css',
   '/static/app.js',
   '/static/js/storage.js',
+  '/static/js/appConfig.js',
   '/static/js/ui.js',
   '/static/js/markdown.js',
   '/static/js/dragSort.js',
@@ -61,6 +87,75 @@ const PRECACHE = [
   '/static/js/sidebar-layout.js',
   '/static/js/section-management.js',
   '/static/lib/highlight.min.js',
+  // Math turns up in ordinary answers and KaTeX is small, so precaching it and
+  // its fonts keeps formulas typeset offline. Mermaid is deliberately NOT
+  // precached: at 3.5 MB it would re-download on every CACHE_NAME bump, a poor
+  // trade for a library most sessions never touch. The cache-first rule below
+  // picks it up the first time a diagram renders, which is also when it starts
+  // mattering offline.
+  '/static/lib/katex/katex.min.js',
+  '/static/lib/katex/katex.min.css',
+  ...KATEX_FONTS,
+];
+
+// Lazily-imported panel modules (js/panels.js). Not in index.html by design;
+// precached so the panel still opens with no network.
+const PANEL_PRECACHE = [
+  // Image editor — galleryEditor.js and its js/editor/ graph.
+  '/static/js/galleryEditor.js',
+  '/static/js/editor/ai-inpaint.js?v=20260708match1',
+  '/static/js/editor/ai-models.js',
+  '/static/js/editor/ai-rembg.js',
+  '/static/js/editor/ai-tool-runner.js',
+  '/static/js/editor/ai-tools-misc.js',
+  '/static/js/editor/build/controls.js?v=20260708match1',
+  '/static/js/editor/build/popups.js',
+  '/static/js/editor/build/right-panel.js',
+  '/static/js/editor/build/toolbar.js?v=20260708sam3',
+  '/static/js/editor/build/topbar.js',
+  '/static/js/editor/build/transform-popup.js',
+  '/static/js/editor/canvas-coords.js',
+  '/static/js/editor/canvas-events.js',
+  '/static/js/editor/canvas-transforms.js',
+  '/static/js/editor/checkerboard.js',
+  '/static/js/editor/clipboard-and-drop.js',
+  '/static/js/editor/composite-helpers.js',
+  '/static/js/editor/filters/blur.js',
+  '/static/js/editor/filters/edge-feather.js',
+  '/static/js/editor/fx/adj-popup.js',
+  '/static/js/editor/fx/filter-string.js',
+  '/static/js/editor/fx/histogram.js',
+  '/static/js/editor/fx/pixel-pass.js',
+  '/static/js/editor/harmonize-masks.js',
+  '/static/js/editor/history-panel.js',
+  '/static/js/editor/keyboard-shortcuts.js',
+  '/static/js/editor/layer-helpers.js',
+  '/static/js/editor/layer-panel.js',
+  '/static/js/editor/mask-utils.js',
+  '/static/js/editor/shortcuts-popover.js',
+  '/static/js/editor/slider-ux.js',
+  '/static/js/editor/snap.js',
+  '/static/js/editor/state.js',
+  '/static/js/editor/stroke-pipeline.js',
+  '/static/js/editor/stroke-tool-sliders.js',
+  '/static/js/editor/tools/clone.js',
+  '/static/js/editor/tools/crop.js',
+  '/static/js/editor/tools/flood-fill.js',
+  '/static/js/editor/tools/lasso-mask.js',
+  '/static/js/editor/tools/lasso.js',
+  '/static/js/editor/tools/move.js',
+  '/static/js/editor/tools/stroke.js',
+  '/static/js/editor/tools/transform-drag.js',
+  '/static/js/editor/tools/transform-handles.js',
+  '/static/js/editor/tools/transform-session.js',
+  '/static/js/editor/tools/wand.js',
+  '/static/js/editor/wire-import.js',
+  '/static/js/editor/wire-inpaint-controls.js?v=20260708match1',
+  '/static/js/editor/wire-merge-buttons.js',
+  '/static/js/editor/wire-selection-controls.js',
+  '/static/js/editor/wire-topbar-menus.js',
+  '/static/js/editor/wire-topbar-overflow.js',
+  '/static/js/editor/wire-topbar.js',
 ];
 
 self.addEventListener('install', (e) => {
@@ -69,7 +164,7 @@ self.addEventListener('install', (e) => {
       // addAll is atomic — if any item fails, none are cached. Use individual
       // puts so a single 404 can't block the whole install.
       Promise.all(
-        PRECACHE.map(url =>
+        [...PRECACHE, ...PANEL_PRECACHE].map(url =>
           fetch(url, { cache: 'reload' })
             .then(res => res.ok ? cache.put(url, res) : null)
             .catch(() => null)

tests/cli/test_webhook_cli_mask.py

@@ -10,3 +10,28 @@ def test_mask_token_handles_short_values(monkeypatch):
     assert cli._mask_token("short") == "***"
     assert cli._mask_token("abcdef1234567890") == "abcdef…7890"
     assert cli._mask_token("short", reveal=True) == "short"
+
+
+def test_task_webhook_url_matches_live_route_and_escapes_path_parts(monkeypatch):
+    make_core_db_stub(monkeypatch, models=["ScheduledTask"])
+    cli = load_script("odysseus-webhook")
+
+    url = cli._task_webhook_url(
+        "https://ody.example/",
+        "task/with space",
+        "token/with space",
+    )
+
+    assert url == (
+        "https://ody.example/api/tasks/task%2Fwith%20space/"
+        "webhook/token%2Fwith%20space"
+    )
+
+
+def test_default_task_webhook_url_uses_current_task_prefix(monkeypatch):
+    make_core_db_stub(monkeypatch, models=["ScheduledTask"])
+    cli = load_script("odysseus-webhook")
+
+    assert cli._task_webhook_url(None, "task-1", "secret") == (
+        "http://localhost:7000/api/tasks/task-1/webhook/secret"
+    )

tests/helpers/test_settings_shell_coordinator.mjs

@@ -817,6 +817,12 @@ const STUBS = new Map([
       bindMenuDismiss() {},
     },
   ],
+  [
+    path.join(JS, 'appConfig.js'),
+    {
+      invalidateSettings() {},
+    },
+  ],
   [
     path.join(JS, 'windowDrag.js'),
     {

tests/test_app_config_shared_fetch_js.py

@@ -0,0 +1,375 @@
+"""Pin the shared config cache in static/js/appConfig.js.
+
+Background: /api/auth/settings was fetched independently by six modules and
+/api/tools by three (chatRenderer.js is imported under three different ?v=
+query strings, so it is three separate module instances) — 4 and 3 requests on
+one cold load. Beyond the redundant work, each caller could observe a different
+snapshot of the same object. appConfig.js gives them one promise each.
+
+The two properties that matter are opposites, so both are tested here:
+concurrent and later callers must NOT refetch, and a caller after a write MUST
+see the new value — which only holds if every writer invalidates. The last test
+is a source scan that checks exactly that, since a forgotten invalidation
+serves a stale settings object for the rest of the session, which is worse than
+the duplicate fetches this replaces.
+
+Driven through `node --input-type=module` so the real module runs, same idiom as
+test_esc_menu_stack_js.py. The module source is inlined rather than imported by
+path because the repo has no `"type": "module"` in package.json; appConfig.js
+has no imports of its own, so inlining is exact. `fetch` and `sessionStorage`
+are stubbed, so nothing here touches the network or depends on timing.
+"""
+import json
+import re
+import shutil
+import subprocess
+from pathlib import Path
+
+import pytest
+
+_REPO = Path(__file__).resolve().parent.parent
+_MODULE = _REPO / "static" / "js" / "appConfig.js"
+_HAS_NODE = shutil.which("node") is not None
+_SRC = _MODULE.read_text(encoding="utf-8") if _MODULE.exists() else ""
+
+# Browser stand-ins. Every fetch is recorded and resolved from a queue the test
+# controls, so "how many requests went out" is an exact count, not a guess.
+_STUBS = r"""
+const calls = [];
+let responses = [];
+globalThis.__queue = (fn) => { responses.push(fn); };
+globalThis.fetch = (url, opts) => {
+  calls.push([url, opts]);
+  const next = responses.shift();
+  if (!next) throw new Error('unexpected fetch: ' + url);
+  return next();
+};
+globalThis.__calls = () => calls;
+globalThis.__json = (value) => () => Promise.resolve({ json: () => Promise.resolve(value) });
+globalThis.__fail = (msg) => () => Promise.reject(new Error(msg));
+
+const store = new Map();
+globalThis.sessionStorage = {
+  getItem: (k) => (store.has(k) ? store.get(k) : null),
+  setItem: (k, v) => { store.set(k, String(v)); },
+  removeItem: (k) => { store.delete(k); },
+};
+globalThis.__seedPrefetch = (value) => {
+  store.set('ody-prefetch-settings', JSON.stringify(value));
+};
+globalThis.__prefetchLeft = () => store.has('ody-prefetch-settings');
+"""
+
+
+def _run(body: str) -> str:
+    js = _STUBS + "\n" + _SRC + "\n" + body
+    proc = subprocess.run(
+        ["node", "--input-type=module"],
+        input=js, capture_output=True, text=True, encoding="utf-8",
+        cwd=str(_REPO), timeout=30,
+    )
+    assert proc.returncode == 0, proc.stderr
+    return proc.stdout.strip()
+
+
+@pytest.mark.skipif(not _HAS_NODE, reason="node binary not on PATH")
+def test_concurrent_callers_share_one_request():
+    # The startup case: several modules ask before the first response lands.
+    body = """
+    __queue(__json({ tts_enabled: true }));
+    const [a, b, c] = await Promise.all([getSettings(), getSettings(), getSettings()]);
+    console.log(JSON.stringify({
+      requests: __calls().length,
+      url: __calls()[0][0],
+      credentials: __calls()[0][1].credentials,
+      sameObject: a === b && b === c,
+      value: a.tts_enabled,
+    }));
+    """
+    assert json.loads(_run(body)) == {
+        "requests": 1,
+        "url": "/api/auth/settings",
+        "credentials": "same-origin",
+        "sameObject": True,
+        "value": True,
+    }
+
+
+@pytest.mark.skipif(not _HAS_NODE, reason="node binary not on PATH")
+def test_later_caller_reuses_the_resolved_snapshot():
+    # A panel opened long after boot must not re-request.
+    body = """
+    __queue(__json({ search_provider: 'brave' }));
+    const first = await getSettings();
+    const second = await getSettings();
+    console.log(JSON.stringify({ requests: __calls().length, sameObject: first === second }));
+    """
+    assert json.loads(_run(body)) == {"requests": 1, "sameObject": True}
+
+
+@pytest.mark.skipif(not _HAS_NODE, reason="node binary not on PATH")
+def test_invalidate_forces_the_next_read_to_refetch():
+    # The write path: save a setting, then read it back and see the new value.
+    body = """
+    __queue(__json({ tts_enabled: true }));
+    __queue(__json({ tts_enabled: false }));
+    const before = await getSettings();
+    invalidateSettings();
+    const after = await getSettings();
+    console.log(JSON.stringify({
+      requests: __calls().length,
+      before: before.tts_enabled,
+      after: after.tts_enabled,
+    }));
+    """
+    assert json.loads(_run(body)) == {"requests": 2, "before": True, "after": False}
+
+
+@pytest.mark.skipif(not _HAS_NODE, reason="node binary not on PATH")
+def test_a_failed_fetch_does_not_poison_the_cache():
+    # Plain `??=` memoisation would keep the rejected promise, so one blip at
+    # boot would leave keybinds/TTS/search on defaults for the whole session.
+    body = """
+    __queue(__fail('offline'));
+    __queue(__json({ tts_enabled: true }));
+    let rejected = false;
+    try { await getSettings(); } catch (e) { rejected = e.message === 'offline'; }
+    const retry = await getSettings();
+    console.log(JSON.stringify({
+      rejected,
+      requests: __calls().length,
+      recovered: retry.tts_enabled,
+    }));
+    """
+    assert json.loads(_run(body)) == {"rejected": True, "requests": 2, "recovered": True}
+
+
+@pytest.mark.skipif(not _HAS_NODE, reason="node binary not on PATH")
+def test_settings_and_tools_are_independent_slots():
+    body = """
+    __queue(__json({ tts_enabled: true }));
+    __queue(__json({ tools: [{ id: 'web_search' }] }));
+    await getSettings();
+    const tools = await getTools();
+    const toolsUrl = __calls()[1][0];
+    invalidateSettings();          // must not drop the tools snapshot
+    const toolsAgain = await getTools();
+    console.log(JSON.stringify({
+      requests: __calls().length,
+      toolsUrl,
+      sameObject: tools === toolsAgain,
+    }));
+    """
+    assert json.loads(_run(body)) == {
+        "requests": 2,
+        "toolsUrl": "/api/tools",
+        "sameObject": True,
+    }
+
+
+@pytest.mark.skipif(not _HAS_NODE, reason="node binary not on PATH")
+def test_login_prefetch_is_used_once_and_then_consumed():
+    # login.html stashes a settings snapshot in sessionStorage just before it
+    # redirects, so the first load after a login should issue no request at all.
+    body = """
+    __seedPrefetch({ tts_enabled: false, from: 'prefetch' });
+    __queue(__json({ tts_enabled: true, from: 'network' }));
+    const first = await getSettings();
+    const consumed = !__prefetchLeft();
+    invalidateSettings();
+    const second = await getSettings();
+    console.log(JSON.stringify({
+      requests: __calls().length,
+      first: first.from,
+      consumed,
+      second: second.from,
+    }));
+    """
+    assert json.loads(_run(body)) == {
+        "requests": 1,
+        "first": "prefetch",
+        "consumed": True,
+        "second": "network",
+    }
+
+
+# ── Writers must invalidate ─────────────────────────────────────────────────
+
+_WRITE_ENDPOINTS = {
+    "/api/auth/settings": "invalidateSettings",
+    "/api/tools": "invalidateTools",
+}
+# login.html is the pre-app login page: it has no module graph and its only call
+# is the prefetch GET, so it is not a writer and cannot import appConfig.js.
+_SCANNED = [_REPO / "static" / "app.js"] + sorted((_REPO / "static" / "js").rglob("*.js"))
+
+
+def _post_sites(source: str, endpoint: str):
+    """Yield the 1-based line of every fetch() to `endpoint` that is a POST."""
+    for m in re.finditer(re.escape(f"'{endpoint}'"), source):
+        window = source[m.start():m.start() + 240]
+        if re.search(r"method:\s*'POST'", window):
+            yield source[:m.start()].count("\n") + 1
+
+
+def test_every_settings_writer_invalidates_the_shared_cache():
+    """A POST that skips the invalidation serves a stale object for the session.
+
+    Checked by source scan rather than at runtime: the failure mode is a call
+    site that was never wired up, which no unit test of the cache itself can
+    see. `appConfig.js` itself is skipped — it is the cache, not a writer.
+    """
+    missing = []
+    for path in _SCANNED:
+        if path.name == "appConfig.js":
+            continue
+        source = path.read_text(encoding="utf-8")
+        for endpoint, invalidator in _WRITE_ENDPOINTS.items():
+            for line in _post_sites(source, endpoint):
+                # The invalidation belongs in the same function as the POST;
+                # accept it anywhere in the surrounding 20 lines either way.
+                lines = source.splitlines()
+                near = "\n".join(lines[max(0, line - 20):line + 20])
+                if invalidator + "(" not in near:
+                    missing.append(f"{path.relative_to(_REPO)}:{line} POST {endpoint}")
+    assert not missing, (
+        "POST sites with no nearby cache invalidation — the UI will serve a "
+        "stale snapshot after these writes:\n  " + "\n  ".join(missing)
+    )
+
+
+@pytest.mark.skipif(not _HAS_NODE, reason="node binary not on PATH")
+def test_out_of_band_tool_change_is_not_undone_by_an_unrelated_panel_save():
+    """Admin > Tools must render authoritative state, not the startup snapshot.
+
+    The save posts the whole disabled-tool list rebuilt from the checkboxes, so
+    a stale render turns any unrelated toggle into a lost update: a tool
+    disabled out of band (manage_settings, another tab) comes back enabled.
+    refreshAll() calls loadBuiltinTools() on every panel open, which is why the
+    editor drops the shared entry before reading it.
+    """
+    body = """
+    // Boot: chatRenderer.js reads the tool list for the exec-fence regex.
+    __queue(__json({ tools: [{ id: 'web_search', enabled: true }, { id: 'shell', enabled: true }] }));
+    const boot = await getTools();
+
+    // Out of band, this page hearing nothing about it: web_search is disabled.
+    __queue(__json({ tools: [{ id: 'web_search', enabled: false }, { id: 'shell', enabled: true }] }));
+
+    // Admin > Tools opens. loadBuiltinTools() invalidates, then reads.
+    invalidateTools();
+    const panel = await getTools();
+
+    // The user toggles one unrelated tool off. The save posts every unchecked
+    // box, so the list is only right if the render was authoritative.
+    const post = (snapshot) => {
+      const boxes = snapshot.tools.map(t => ({ id: t.id, checked: t.enabled }));
+      boxes.find(b => b.id === 'shell').checked = false;
+      return boxes.filter(b => !b.checked).map(b => b.id);
+    };
+
+    console.log(JSON.stringify({
+      requests: __calls().length,
+      posted: post(panel),
+      postedFromStaleSnapshot: post(boot),
+    }));
+    """
+    assert json.loads(_run(body)) == {
+        "requests": 2,
+        # web_search stays disabled, which is the point.
+        "posted": ["web_search", "shell"],
+        # What the page-lifetime snapshot would have posted: web_search silently
+        # re-enabled by a toggle that had nothing to do with it.
+        "postedFromStaleSnapshot": ["shell"],
+    }
+
+
+@pytest.mark.skipif(not _HAS_NODE, reason="node binary not on PATH")
+def test_out_of_band_tool_change_after_panel_open_is_preserved_on_save():
+    """Saving must merge the user's edit onto a fresh authoritative snapshot."""
+    body = """
+    // Panel opens while both tools are enabled.
+    __queue(__json({ tools: [
+      { id: 'web_search', enabled: true },
+      { id: 'shell', enabled: true },
+    ] }));
+    invalidateTools();
+    const panel = await getTools();
+
+    // Another tab disables web_search after this panel has already rendered.
+    __queue(__json({ tools: [
+      { id: 'web_search', enabled: false },
+      { id: 'shell', enabled: true },
+    ] }));
+
+    // The user only disables shell. Saving refreshes the authoritative state
+    // first and applies that one intended change on top of it.
+    invalidateTools();
+    const latest = await getTools();
+    const state = new Map(latest.tools.map(t => [t.id, !!t.enabled]));
+    state.set('shell', false);
+
+    const disabled = Array.from(state.entries())
+      .filter(([, enabled]) => !enabled)
+      .map(([id]) => id);
+
+    console.log(JSON.stringify({
+      panelWebSearchEnabled: panel.tools.find(t => t.id === 'web_search').enabled,
+      requests: __calls().length,
+      disabled,
+    }));
+    """
+    assert json.loads(_run(body)) == {
+        "panelWebSearchEnabled": True,
+        "requests": 2,
+        "disabled": ["web_search", "shell"],
+    }
+
+
+def test_admin_tool_save_refreshes_before_full_state_post():
+    """Pin the lost-update guard in the Admin Tools full-list writer."""
+    source = (_REPO / "static" / "js" / "admin.js").read_text(encoding="utf-8")
+    match = re.search(
+        r"async function _saveToolState\(changes\) \{(.*?)\n    \}\n"
+        r"    function _updateCatCounter",
+        source,
+        re.S,
+    )
+    assert match, "_saveToolState(changes) not found in static/js/admin.js"
+
+    body = match.group(1)
+    invalidate = body.find("invalidateTools()")
+    refresh = body.find("getTools()")
+    post = body.find("fetch('/api/tools'")
+
+    assert -1 not in (invalidate, refresh, post)
+    assert invalidate < refresh < post, (
+        "Admin Tools must invalidate and refresh authoritative tool state before "
+        "posting the endpoint's full disabled-tools replacement list"
+    )
+    assert "for (const change of changes)" in body
+
+
+def test_the_admin_tools_editor_does_not_read_a_cached_snapshot():
+    """Pin the invalidate-before-read in loadBuiltinTools().
+
+    A source scan because the failure is an ordering in a call site, not
+    behaviour of the cache: getTools() is doing exactly its job either way.
+    """
+    source = (_REPO / "static" / "js" / "admin.js").read_text(encoding="utf-8")
+    match = re.search(r"\nasync function loadBuiltinTools\(\) \{\n(.*?)\n\}\n", source, re.S)
+    assert match, "loadBuiltinTools() not found in static/js/admin.js"
+    body = match.group(1)
+    read = body.find("getTools(")
+    assert read != -1, "loadBuiltinTools() no longer reads the shared tool cache"
+    assert "invalidateTools(" in body[:read], (
+        "loadBuiltinTools() reads the shared /api/tools snapshot without dropping "
+        "it first, so a reopened panel can render tool state that changed out of "
+        "band and re-post it on the next unrelated toggle"
+    )
+
+
+def test_appconfig_is_precached_by_the_service_worker():
+    """PRECACHE is hand-maintained; a module missing from it breaks offline."""
+    sw = (_REPO / "static" / "sw.js").read_text(encoding="utf-8")
+    assert "'/static/js/appConfig.js'" in sw

tests/test_approved_replay_message_shape.py

@@ -0,0 +1,100 @@
+"""Regression coverage for the assistant turn an approved-action replay appends.
+
+Anthropic's Messages API rejects a non-final assistant message whose content is
+empty, so the resumed turn after a tool approval used to fail before the model
+ever saw the sealed result. The replay injects its result with no assistant
+prose for that round, which is the only path that produced such a turn.
+"""
+
+import src.llm_core as llm_core
+from src.agent_loop import _append_tool_results
+
+_RESULT = "bash: ok\nhello"
+_RECORD = {
+    "tool_name": "bash",
+    "content": "printf hello",
+    "result": {"output": "hello", "exit_code": 0},
+    "text": _RESULT,
+}
+
+
+def _replay_messages(round_response="", round_reasoning=""):
+    """Mirror the approved-action injection in stream_agent_loop."""
+    messages = [
+        {"role": "system", "content": "system preface"},
+        {"role": "user", "content": "run the command and summarise it"},
+    ]
+    _append_tool_results(
+        messages,
+        round_response,
+        [],
+        [_RESULT],
+        [_RESULT],
+        False,
+        0,
+        round_reasoning=round_reasoning,
+        tool_result_records=[_RECORD],
+    )
+    return messages
+
+
+def _empty_assistant_turns(messages):
+    return [
+        index
+        for index, message in enumerate(messages)
+        if message.get("role") == "assistant"
+        and not str(message.get("content") or "").strip()
+        and not message.get("tool_calls")
+    ]
+
+
+def test_replay_appends_no_empty_assistant_turn():
+    assert _empty_assistant_turns(_replay_messages()) == []
+
+
+def test_replay_payload_is_a_single_non_empty_user_turn_for_anthropic():
+    # Asserting the whole sequence rather than scanning a slice: once the empty
+    # spacer is gone the payload is one message, so a "no offenders in
+    # chat[:-1]" check would pass without inspecting anything.
+    sanitized = llm_core._sanitize_llm_messages(_replay_messages())
+    payload = llm_core._build_anthropic_payload(
+        "claude-sonnet-5", sanitized, 0.2, 512
+    )
+    chat = payload["messages"]
+    assert [message["role"] for message in chat] == ["user"]
+    assert all(str(message.get("content") or "").strip() for message in chat)
+
+
+def test_replay_merges_tool_output_after_the_request_with_its_fence_intact():
+    # Dropping the empty assistant turn leaves two adjacent user messages, which
+    # the sanitizer merges. Pin that shape: the tool output must still sit
+    # behind its untrusted fence and must not precede the operator's request.
+    sanitized = llm_core._sanitize_llm_messages(_replay_messages())
+    user_turns = [m for m in sanitized if m.get("role") == "user"]
+    assert len(user_turns) == 1
+    merged = user_turns[0]["content"]
+    assert merged.index("run the command") < merged.index("UNTRUSTED SOURCE DATA")
+    assert merged.index("UNTRUSTED SOURCE DATA") < merged.index("hello")
+
+
+def test_a_round_with_prose_still_appends_its_assistant_turn():
+    messages = _replay_messages(round_response="running that now")
+    assistants = [m for m in messages if m.get("role") == "assistant"]
+    assert [m["content"] for m in assistants] == ["running that now"]
+
+
+def test_reasoning_only_round_keeps_its_carrier_and_its_known_empty_content():
+    """Characterises the one empty-content turn this change deliberately leaves.
+
+    A round with reasoning and no prose still appends its carrier, and that
+    carrier's content is still "", which Anthropic would still reject. Dropping
+    it would lose the reasoning DeepSeek thinking mode requires on the next
+    request, and the approval replay never takes this path because it passes no
+    reasoning. Left alone on purpose; this test makes the gap visible instead of
+    silent, and should be updated by whoever closes it.
+    """
+    messages = _replay_messages(round_reasoning="thinking about it")
+    assistants = [m for m in messages if m.get("role") == "assistant"]
+    assert len(assistants) == 1
+    assert assistants[0].get("reasoning_content") == "thinking about it"
+    assert assistants[0]["content"] == ""