fix(agent): drop the empty assistant turn from an approved-action replay

The approved-action replay appends the sealed tool result with no assistant
prose for that round, which produced an assistant message with content "".
Anthropic's Messages API rejects a non-final assistant message with empty
content, so a resumed turn after a tool approval failed before the model saw
the result. A turn carrying neither prose nor reasoning has nothing to say to
any provider, so it is no longer appended. A round with prose, and a
reasoning-only round that DeepSeek thinking mode needs, both still append.

.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.

.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/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

@@ -96,7 +96,11 @@ module.exports = async ({ github, context, core }) => {
 
   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);
-  const screenshotChecked = /- \[x\]\s+\*\*Screenshot or short clip\*\*/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 = [];

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

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:-}

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:-}

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:-}

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`. |

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/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/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/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/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

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/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)
+    )

static/app.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

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

@@ -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;
@@ -6606,7 +6607,7 @@ 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.

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"] == ""

tests/test_auth_event_loop.py

@@ -99,7 +99,12 @@ def test_login_offloads_bcrypt_bearing_calls(monkeypatch):
 
     login = _login_endpoint(auth)
 
-    request = SimpleNamespace(client=SimpleNamespace(host="203.0.113.7"), cookies={})
+    request = SimpleNamespace(
+        client=SimpleNamespace(host="203.0.113.7"),
+        cookies={},
+        url=SimpleNamespace(scheme="http"),
+        headers={},
+    )
     response = MagicMock()
     body = LoginRequest(username="alice", password="hunter2", remember=True)
 

tests/test_auth_policy.py

@@ -242,12 +242,26 @@ def _login_endpoint(auth_manager):
     raise AssertionError("login route not found")
 
 
+def _login_request(scheme="http", forwarded_proto=None):
+    """Stand-in for fastapi.Request carrying the fields login reads: the
+    client host (rate limiter), and the URL scheme plus `X-Forwarded-Proto`
+    (cookie Secure flag)."""
+    headers = {}
+    if forwarded_proto is not None:
+        headers["x-forwarded-proto"] = forwarded_proto
+    return SimpleNamespace(
+        client=SimpleNamespace(host="127.0.0.1"),
+        url=SimpleNamespace(scheme=scheme),
+        headers=headers,
+    )
+
+
 def test_remember_cookie_max_age_matches_token_ttl(tmp_path):
     auth_mod = _auth_module()
     mgr = _make_manager(tmp_path)
     mgr.create_user("alice", "alice-password", is_admin=False)
     endpoint, LoginRequest = _login_endpoint(mgr)
-    request = SimpleNamespace(client=SimpleNamespace(host="127.0.0.1"))
+    request = _login_request()
     response = _CapturingResponse()
     body = LoginRequest(username="alice", password="alice-password", remember=True)
 
@@ -262,7 +276,7 @@ def test_no_remember_omits_cookie_max_age(tmp_path):
     mgr = _make_manager(tmp_path)
     mgr.create_user("bob", "bob-password", is_admin=False)
     endpoint, LoginRequest = _login_endpoint(mgr)
-    request = SimpleNamespace(client=SimpleNamespace(host="127.0.0.1"))
+    request = _login_request()
     response = _CapturingResponse()
     body = LoginRequest(username="bob", password="bob-password", remember=False)
 
@@ -270,3 +284,80 @@ def test_no_remember_omits_cookie_max_age(tmp_path):
 
     # Without "remember", the cookie is a session cookie (no max_age).
     assert "max_age" not in response.cookie_kwargs
+
+
+# ── Session cookie Secure flag ─────────────────────────────────────────
+
+
+def _login_secure_flag(tmp_path, scheme, forwarded_proto=None):
+    """Log a user in over ``scheme`` and return the cookie's Secure flag."""
+    mgr = _make_manager(tmp_path)
+    mgr.create_user("carol", "carol-password", is_admin=False)
+    endpoint, LoginRequest = _login_endpoint(mgr)
+    response = _CapturingResponse()
+    body = LoginRequest(username="carol", password="carol-password")
+    request = _login_request(scheme, forwarded_proto)
+
+    asyncio.run(endpoint(body=body, request=request, response=response))
+
+    return response.cookie_kwargs["secure"]
+
+
+def test_https_login_marks_cookie_secure_without_config(tmp_path, monkeypatch):
+    monkeypatch.delenv("SECURE_COOKIES", raising=False)
+
+    # A session token handed out over HTTPS must not be allowed to travel
+    # back in cleartext just because nobody set an env var.
+    assert _login_secure_flag(tmp_path, "https") is True
+
+
+def test_plain_http_login_leaves_cookie_insecure_without_config(tmp_path, monkeypatch):
+    monkeypatch.delenv("SECURE_COOKIES", raising=False)
+
+    # Marking it Secure here would make the browser drop the cookie and
+    # break login on a plain-HTTP install.
+    assert _login_secure_flag(tmp_path, "http") is False
+
+
+def test_forwarded_proto_https_marks_cookie_secure(tmp_path, monkeypatch):
+    monkeypatch.delenv("SECURE_COOKIES", raising=False)
+
+    # A terminator that is not on an address uvicorn trusts leaves the
+    # connection scheme as http, so the header is the only signal there.
+    assert _login_secure_flag(tmp_path, "http", forwarded_proto="https") is True
+    # Chained proxies send a list; the client-facing hop is the first entry.
+    assert _login_secure_flag(tmp_path, "http", forwarded_proto="https, http") is True
+
+
+def test_forwarded_proto_http_leaves_plain_http_login_insecure(tmp_path, monkeypatch):
+    monkeypatch.delenv("SECURE_COOKIES", raising=False)
+
+    # The header is not a downgrade switch: either signal saying https is
+    # enough, the same test core/middleware.py applies before sending HSTS.
+    assert _login_secure_flag(tmp_path, "http", forwarded_proto="http") is False
+    assert _login_secure_flag(tmp_path, "https", forwarded_proto="http") is True
+
+
+def test_empty_secure_cookies_still_derives_from_scheme(tmp_path, monkeypatch):
+    # docker-compose injects `SECURE_COOKIES=${SECURE_COOKIES:-}`, which sets
+    # the variable to "" when the host has not defined it. Empty means
+    # unconfigured, not "off".
+    monkeypatch.setenv("SECURE_COOKIES", "")
+
+    assert _login_secure_flag(tmp_path, "https") is True
+
+
+def test_secure_cookies_true_forces_secure_on_plain_http(tmp_path, monkeypatch):
+    # The documented knob keeps working for a proxy whose scheme the app
+    # cannot see, e.g. one that is not on a trusted loopback address.
+    monkeypatch.setenv("SECURE_COOKIES", "true")
+
+    assert _login_secure_flag(tmp_path, "http") is True
+
+
+def test_secure_cookies_false_forces_insecure_on_https(tmp_path, monkeypatch):
+    # The escape hatch for an install still answering on both HTTP and
+    # HTTPS: an explicit false wins over the request scheme.
+    monkeypatch.setenv("SECURE_COOKIES", "false")
+
+    assert _login_secure_flag(tmp_path, "https") is False

tests/test_aux_llm_owner_scope.py

@@ -16,7 +16,7 @@ def test_registered_manual_compaction_uses_session_owner_for_utility_endpoint():
 
 
 def test_task_name_generation_uses_owner_scoped_session_endpoint():
-    src = _src("routes/task_routes.py")
+    src = _src("routes/task/task_routes.py")
 
     assert "async def _generate_task_name(prompt: str, owner: Optional[str] = None)" in src
     assert "q = q.filter(DbSession.owner == owner)" in src

tests/test_chat_url_prefetch_failure_context.py

@@ -0,0 +1,177 @@
+"""Automatic chat URL fetch failures must remain visible to the model."""
+
+import logging
+from types import SimpleNamespace
+
+import pytest
+
+import src.chat_processor as chat_processor
+
+
+def _processor():
+    return chat_processor.ChatProcessor(
+        memory_manager=SimpleNamespace(load=lambda owner=None: []),
+        personal_docs_manager=SimpleNamespace(rag_manager=None),
+        skills_manager=None,
+    )
+
+
+def test_failed_url_prefetch_keeps_diagnostics_out_of_model_context(monkeypatch):
+    monkeypatch.setattr(
+        chat_processor,
+        "fetch_webpage_content",
+        lambda url: {
+            "success": False,
+            "error": "NetworkError: connection refused\nattacker supplied detail",
+        },
+    )
+
+    preface, _, _ = _processor().build_context_preface(
+        message=(
+            "Please summarize "
+            "https://example.invalid/report?token=LEAK_URL_MARKER"
+        ),
+        session=SimpleNamespace(endpoint_url="", model="", headers={}),
+        use_web=False,
+        use_rag=False,
+        use_memory=False,
+    )
+
+    failure = next(
+        message
+        for message in preface
+        if (message.get("metadata") or {}).get("source")
+        == "web page fetch failure"
+    )
+    assert failure["metadata"]["trusted"] is False
+    assert "was not read: the page was unavailable" in failure["content"]
+    assert "connection refused" not in failure["content"]
+    assert "attacker supplied detail" not in failure["content"]
+    assert "LEAK_URL_MARKER" not in failure["content"]
+    assert "LEAK_URL_MARKER" not in str(failure["metadata"])
+
+
+def test_failed_url_prefetch_exposes_only_stable_http_status(monkeypatch):
+    monkeypatch.setattr(
+        chat_processor,
+        "fetch_webpage_content",
+        lambda url: {
+            "success": False,
+            "error": "HTTP 403: attacker-controlled response detail",
+        },
+    )
+
+    preface, _, _ = _processor().build_context_preface(
+        message="Read https://example.test/private",
+        session=SimpleNamespace(endpoint_url="", model="", headers={}),
+        use_web=False,
+        use_rag=False,
+        use_memory=False,
+    )
+
+    content = preface[-1]["content"]
+    assert "server returned HTTP 403" in content
+    assert "attacker-controlled" not in content
+
+
+@pytest.mark.parametrize(
+    ("error", "expected_category"),
+    [
+        (
+            "TooLarge: response exceeded configured limit; attacker-controlled detail",
+            "response exceeded the fetch size limit",
+        ),
+        (
+            "Rate limit exceeded: attacker-controlled detail",
+            "request was rate limited",
+        ),
+    ],
+)
+def test_failed_url_prefetch_exposes_only_stable_failure_category(
+    monkeypatch, error, expected_category
+):
+    monkeypatch.setattr(
+        chat_processor,
+        "fetch_webpage_content",
+        lambda url: {"success": False, "error": error},
+    )
+
+    preface, _, _ = _processor().build_context_preface(
+        message="Read https://example.test/report",
+        session=SimpleNamespace(endpoint_url="", model="", headers={}),
+        use_web=False,
+        use_rag=False,
+        use_memory=False,
+    )
+
+    content = preface[-1]["content"]
+    assert expected_category in content
+    assert "attacker-controlled" not in content
+
+
+def test_unexpected_url_prefetch_exception_does_not_abort_context(monkeypatch):
+    def fail(url):
+        raise RuntimeError("socket detail must stay in logs")
+
+    monkeypatch.setattr(chat_processor, "fetch_webpage_content", fail)
+
+    preface, _, _ = _processor().build_context_preface(
+        message="Review https://example.test/report",
+        session=SimpleNamespace(endpoint_url="", model="", headers={}),
+        use_web=False,
+        use_rag=False,
+        use_memory=False,
+    )
+
+    content = preface[-1]["content"]
+    assert "was not read: the page was unavailable" in content
+    assert "socket detail" not in content
+
+
+def test_unexpected_url_prefetch_exception_logs_only_stable_diagnostic(
+    monkeypatch, caplog
+):
+    def fail(url):
+        raise RuntimeError("transport detail LEAK_EXCEPTION_MARKER")
+
+    monkeypatch.setattr(chat_processor, "fetch_webpage_content", fail)
+
+    with caplog.at_level(logging.WARNING, logger=chat_processor.logger.name):
+        _processor().build_context_preface(
+            message=(
+                "Review https://example.test/report?token=LEAK_URL_MARKER"
+            ),
+            session=SimpleNamespace(endpoint_url="", model="", headers={}),
+            use_web=False,
+            use_rag=False,
+            use_memory=False,
+        )
+
+    assert "Automatic URL fetch failed while building context" in caplog.text
+    assert "LEAK_URL_MARKER" not in caplog.text
+    assert "LEAK_EXCEPTION_MARKER" not in caplog.text
+
+
+def test_successful_url_prefetch_keeps_existing_content_shape(monkeypatch):
+    monkeypatch.setattr(
+        chat_processor,
+        "fetch_webpage_content",
+        lambda url: {"success": True, "content": "page body"},
+    )
+
+    preface, _, _ = _processor().build_context_preface(
+        message="Read https://example.test/page",
+        session=SimpleNamespace(endpoint_url="", model="", headers={}),
+        use_web=False,
+        use_rag=False,
+        use_memory=False,
+    )
+
+    page = next(
+        message
+        for message in preface
+        if (message.get("metadata") or {}).get("source")
+        == "web page: https://example.test/page"
+    )
+    assert page["metadata"]["trusted"] is False
+    assert "page body" in page["content"]

tests/test_cookbook_helpers.py

@@ -16,6 +16,7 @@ from routes.cookbook_helpers import (
     _llama_cpp_rebuild_cmd,
     _append_vllm_linux_preflight_lines,
     _local_tooling_path_export,
+    _local_windows_bash_env_prefix,
     _pip_install_attempt,
     _pip_install_fallback_chain,
     _ollama_bind_from_cmd,
@@ -107,6 +108,70 @@ def test_safe_env_prefix_accepts_powershell_activation_path():
     )
 
 
+@pytest.mark.parametrize(
+    ("prefix", "expected"),
+    [
+        ("& 'C:\\Users\\me\\venv\\Scripts\\Activate.ps1'", "source /c/Users/me/venv/Scripts/activate"),
+        (r"& C:\Users\me\venv\Scripts\Activate.ps1", "source /c/Users/me/venv/Scripts/activate"),
+        (
+            r"& C:\Users\me\My Envs\venv\Scripts\Activate.ps1",
+            "source '/c/Users/me/My Envs/venv/Scripts/activate'",
+        ),
+        (
+            "& 'C:\\Users\\me\\My Envs\\venv\\Scripts\\Activate.ps1'",
+            "source '/c/Users/me/My Envs/venv/Scripts/activate'",
+        ),
+        (r"& D:/Envs/venv/Scripts/Activate.ps1", "source /d/Envs/venv/Scripts/activate"),
+    ],
+)
+def test_local_windows_bash_env_prefix_converts_powershell_venv_activation(prefix, expected):
+    converted = _local_windows_bash_env_prefix(prefix)
+
+    assert converted == expected
+    assert _safe_env_prefix(converted).startswith('[ -f "')
+
+
+@pytest.mark.parametrize(
+    "prefix",
+    [
+        None,
+        "",
+        "source /home/me/venv/bin/activate",
+        "conda activate qwen35",
+        'eval "$(conda shell.bash hook)" && conda activate qwen35',
+        r"& \\server\share\venv\Scripts\Activate.ps1",
+    ],
+)
+def test_local_windows_bash_env_prefix_leaves_other_prefixes_unchanged(prefix):
+    assert _local_windows_bash_env_prefix(prefix) == prefix
+
+
+def test_local_windows_bash_env_prefix_handles_long_whitespace_input():
+    prefix = "\t" * 100_000
+
+    assert _local_windows_bash_env_prefix(prefix) == prefix
+
+
+@pytest.mark.parametrize(
+    "relative_path",
+    ["static/js/cookbookRunning.js", "static/js/cookbookDownload.js"],
+)
+def test_primary_windows_venv_emitters_quote_activation_path(relative_path):
+    source = (Path(__file__).resolve().parents[1] / relative_path).read_text(encoding="utf-8")
+
+    assert "'& ' + _psQuote(" in source
+    assert "_psQuote = shared._psQuote;" in source
+
+
+def test_windows_venv_conversion_stays_scoped_to_local_git_bash_runners():
+    source = (Path(__file__).resolve().parents[1] / "routes/cookbook_routes.py").read_text(encoding="utf-8")
+    guarded_conversion = (
+        "_local_windows_bash_env_prefix(req.env_prefix) if local_windows else req.env_prefix"
+    )
+
+    assert source.count(guarded_conversion) == 2
+
+
 def test_validate_local_dir_accepts_external_drive_paths_with_spaces():
     path = "/Volumes/T7 2TB/AI Models/llamacpp"
 

tests/test_docs_query_nondict_rows.py

@@ -9,11 +9,18 @@ class _FakeRag:
 
     def search(self, query, k=5):
         return [
-            {"text": "alpha", "source": "a.txt", "score": 0.9},
+            {
+                "document": "alpha",
+                "metadata": {"source": "a.txt"},
+                "similarity": 0.9,
+            },
             "corrupt-row",
             None,
         ]
 
+    def index_personal_documents(self, directory):
+        return {"indexed_count": 7, "failed_count": 2, "errors": ["bad.pdf"]}
+
 
 def test_query_skips_non_dict_rag_rows():
     # Bypass __init__ (it builds a real RAGManager / Chroma client) and inject
@@ -24,3 +31,65 @@ def test_query_skips_non_dict_rag_rows():
     # old code called r.get(...) on the str/None rows and raised AttributeError.
     assert [c.text for c in out] == ["alpha"]
     assert out[0].source == "a.txt"
+    assert out[0].score == 0.9
+
+
+def test_index_maps_live_vectorrag_result_shape():
+    svc = DocsService.__new__(DocsService)
+    svc.rag = _FakeRag()
+
+    out = asyncio.run(svc.index("/documents"))
+
+    assert out.indexed == 7
+    assert out.failed == 2
+    assert out.errors == ["bad.pdf"]
+
+
+def test_query_normalizes_null_canonical_fields_and_malformed_metadata():
+    class _MalformedRag:
+        def search(self, query, k=5):
+            return [
+                {
+                    "document": None,
+                    "text": "legacy text",
+                    "source": None,
+                    "similarity": None,
+                    "score": 0.4,
+                    "metadata": "not-a-dict",
+                },
+                {
+                    "document": "canonical zero",
+                    "similarity": 0.0,
+                    "metadata": {"source": "nested.txt"},
+                },
+                {
+                    "document": None,
+                    "text": None,
+                    "content": "content fallback",
+                    "similarity": 0.2,
+                    "metadata": ["unexpected"],
+                },
+            ]
+
+    svc = DocsService.__new__(DocsService)
+    svc.rag = _MalformedRag()
+
+    out = asyncio.run(svc.query("query"))
+
+    assert [chunk.text for chunk in out] == [
+        "legacy text",
+        "canonical zero",
+        "content fallback",
+    ]
+
+    assert out[0].source == "unknown"
+    assert out[0].score == 0.4
+    assert out[0].metadata == {}
+
+    assert out[1].source == "nested.txt"
+    assert out[1].score == 0.0
+    assert out[1].metadata == {"source": "nested.txt"}
+
+    assert out[2].source == "unknown"
+    assert out[2].score == 0.2
+    assert out[2].metadata == {}

tests/test_internal_api_base.py

@@ -41,7 +41,7 @@ def test_no_hardcoded_loopback_left_in_call_sites():
         "src/tools/_common.py",
         "src/cookbook_serve_lifecycle.py",
         "src/builtin_actions.py",
-        "routes/task_routes.py",
+        "routes/task/task_routes.py",
     ):
         text = (root / rel).read_text(encoding="utf-8")
         # Allow it only inside comments; flag any code occurrence.

tests/test_issue_description_check.py

@@ -6,10 +6,12 @@ import subprocess
 from pathlib import Path
 
 import pytest
+import yaml
 
 
 _REPO = Path(__file__).resolve().parent.parent
 _CHECKER = _REPO / ".github" / "scripts" / "check-issue-description.js"
+_BUG_TEMPLATE = _REPO / ".github" / "ISSUE_TEMPLATE" / "bug_report.yml"
 _WORKFLOW = _REPO / ".github" / "workflows" / "issue-description-check.yml"
 pytestmark = pytest.mark.skipif(not shutil.which("node"), reason="node not on PATH")
 
@@ -66,11 +68,114 @@ checkIssueDescription({ github, context, core })
     return json.loads(proc.stdout)
 
 
+def _run_bug_issue(revision):
+    body = f"""# Odysseus Revision
+{revision}
+
+# Install Method
+Docker
+
+# Operating System
+Linux
+
+# Steps to Reproduce
+1. Start Odysseus.
+
+# Expected Behaviour
+The application starts normally.
+
+# Actual Behaviour
+The application exits unexpectedly.
+"""
+    harness = r"""
+const checkIssueDescription = require(process.argv[1]);
+const body = process.argv[2];
+const calls = [];
+
+const github = {
+  rest: {
+    issues: {
+      removeLabel: async (params) => calls.push({ method: 'removeLabel', params }),
+      getLabel: async () => ({}),
+      addLabels: async (params) => calls.push({ method: 'addLabels', params }),
+      listComments: async () => ({ data: [] }),
+      createComment: async (params) => calls.push({ method: 'createComment', params }),
+      updateComment: async (params) => calls.push({ method: 'updateComment', params }),
+      deleteComment: async (params) => calls.push({ method: 'deleteComment', params }),
+    },
+  },
+};
+const context = {
+  payload: {
+    action: 'opened',
+    issue: { number: 42, state: 'open', body, labels: [{ name: 'bug' }] },
+  },
+  repo: { owner: 'odysseus-dev', repo: 'odysseus' },
+};
+const core = {
+  warning: (message) => calls.push({ method: 'warning', message }),
+  setFailed: (message) => calls.push({ method: 'setFailed', message }),
+};
+
+checkIssueDescription({ github, context, core })
+  .then(() => process.stdout.write(JSON.stringify(calls)))
+  .catch((error) => {
+    console.error(error);
+    process.exitCode = 1;
+  });
+"""
+    proc = subprocess.run(
+        ["node", "-e", harness, str(_CHECKER), body],
+        capture_output=True,
+        text=True,
+        cwd=str(_REPO),
+        timeout=30,
+    )
+    assert proc.returncode == 0, proc.stderr
+    return json.loads(proc.stdout)
+
+
 def test_workflow_handles_issue_closures():
     workflow = _WORKFLOW.read_text()
     assert "types: [opened, edited, reopened, closed]" in workflow
 
 
+def test_bug_template_requires_exact_revision():
+    template = yaml.safe_load(_BUG_TEMPLATE.read_text())
+    revision = next(item for item in template["body"] if item.get("id") == "revision")
+    assert revision["type"] == "input"
+    assert revision["attributes"]["label"] == "Odysseus Revision"
+    assert "git show -s --abbrev=12 --format='%h (%cs)' HEAD" in revision["attributes"]["description"]
+    assert revision["attributes"]["placeholder"] == "1fef4929cf1d (2026-08-11)"
+    assert revision["validations"]["required"] is True
+
+
+def test_bug_checker_accepts_exact_revision():
+    calls = _run_bug_issue("1fef4929cf1d (2026-08-11)")
+    assert not any(call["method"] in {"createComment", "setFailed"} for call in calls)
+    assert any(
+        call["method"] == "addLabels" and call["params"]["labels"] == ["ready for review"]
+        for call in calls
+    )
+
+
+@pytest.mark.parametrize(
+    "revision",
+    [
+        "",
+        "1fef492",
+        "1fef4929cf1d",
+        "1fef4929cf1d (11 August 2026)",
+        "1fef4929cf1d (2026-08-11) extra",
+    ],
+)
+def test_bug_checker_rejects_missing_or_malformed_revision(revision):
+    calls = _run_bug_issue(revision)
+    comment = next(call["params"]["body"] for call in calls if call["method"] == "createComment")
+    assert "**Odysseus Revision**" in comment
+    assert any(call["method"] == "setFailed" for call in calls)
+
+
 @pytest.mark.parametrize("action", ["closed", "edited"])
 def test_closed_issue_only_drops_ready_for_review(action):
     assert _run_closed_issue(action) == [

tests/test_kokoro_optional_requirements.py

@@ -0,0 +1,40 @@
+from pathlib import Path
+
+import pytest
+from packaging.requirements import Requirement
+
+
+ROOT = Path(__file__).resolve().parents[1]
+
+
+def _optional_requirement(name):
+    lines = (ROOT / "requirements-optional.txt").read_text(encoding="utf-8").splitlines()
+    return Requirement(next(line for line in lines if line.startswith(name)))
+
+
+def test_kokoro_optional_dependency_pins_the_verified_release():
+    kokoro = _optional_requirement("kokoro")
+
+    assert str(kokoro.specifier) == "==0.9.4"
+
+
+@pytest.mark.parametrize(
+    ("python_version", "selected"),
+    (("3.11", True), ("3.12", True), ("3.13", False), ("3.14", False)),
+)
+def test_kokoro_feature_markers_match_supported_python_range(python_version, selected):
+    for name in ("kokoro", "soundfile"):
+        requirement = _optional_requirement(name)
+        assert requirement.marker is not None
+        assert requirement.marker.evaluate({"python_version": python_version}) is selected
+
+
+def test_setup_documents_container_constraint_and_install_command():
+    setup = (ROOT / "docs" / "setup.md").read_text(encoding="utf-8")
+
+    assert "pip install -r requirements-optional.txt" in setup
+    assert "default Docker image currently uses Python 3.14" in setup
+    assert "Python `>=3.10,<3.13`" in setup
+    assert "native Python 3.11 or 3.12" in setup
+    assert "skipped on Python 3.13+" in setup
+    assert "torch build has CUDA" in setup

tests/test_markdown_lazy_lib_loading_js.py

@@ -0,0 +1,510 @@
+"""KaTeX and Mermaid must be vendored and fetched only on first real use.
+
+They used to load from cdn.jsdelivr.net in every <head>, costing ~985 KB on the
+wire per page load, breaking offline installs and announcing each session to a
+third party. These tests pin the replacement contract: one fetch per library,
+never before a formula or a ```mermaid fence actually shows up, and math still
+renders once the library lands.
+"""
+
+import json
+import re
+import shutil
+import subprocess
+import textwrap
+from pathlib import Path
+
+import pytest
+
+_REPO = Path(__file__).resolve().parent.parent
+_HAS_NODE = shutil.which("node") is not None
+
+MERMAID_SRC = "/static/lib/mermaid.min.js"
+KATEX_SRC = "/static/lib/katex/katex.min.js"
+KATEX_CSS = "/static/lib/katex/katex.min.css"
+
+
+@pytest.fixture(scope="module")
+def node_available():
+    if not _HAS_NODE:
+        pytest.skip("node binary not on PATH")
+
+
+def _katex_fonts_block(sw_source: str) -> str:
+    """The literal body of sw.js's KATEX_FONTS array."""
+    match = re.search(r"const KATEX_FONTS = \[(.*?)\]", sw_source, re.S)
+    assert match, "sw.js no longer defines a KATEX_FONTS array"
+    return match.group(1)
+
+
+# A DOM stub small enough to reason about: it records every <script>/<link> the
+# module injects and lets the test decide when each one "loads", which is the
+# only way to observe that a second call reuses the first fetch.
+_HARNESS = r"""
+import fs from 'node:fs';
+import vm from 'node:vm';
+
+// The vendored KaTeX build itself, not a stand-in. A fake renderer that echoes
+// its input cannot tell "a < b" from "a &lt; b" — real KaTeX reads the "&" as
+// an alignment marker and returns a .katex-error span, which is the whole
+// point of the entity tests below. renderToString needs no DOM, so a bare vm
+// context is enough and keeps the library off the harness globals until a test
+// installs it deliberately.
+function loadRealKatex() {
+  const context = { console };
+  context.window = context;
+  context.self = context;
+  context.globalThis = context;
+  vm.createContext(context);
+  vm.runInContext(fs.readFileSync('./static/lib/katex/katex.min.js', 'utf8'), context);
+  if (!context.katex) throw new Error('vendored katex.min.js did not define a katex global');
+  return context.katex;
+}
+
+const injected = { scripts: [], links: [] };
+
+function makeEl(tag) {
+  return {
+    tagName: String(tag).toUpperCase(),
+    _listeners: {},
+    classList: { remove() {} },
+    addEventListener(type, fn) { (this._listeners[type] ||= []).push(fn); },
+    fire(type) { (this._listeners[type] || []).forEach((fn) => fn()); },
+  };
+}
+
+function makeTemplate() {
+  return {
+    _html: '',
+    content: { querySelectorAll() { return []; } },
+    set innerHTML(value) { this._html = value; },
+    get innerHTML() { return this._html; },
+  };
+}
+
+globalThis.window = { location: { origin: 'http://localhost' }, katex: null, mermaid: null };
+globalThis.document = {
+  readyState: 'complete',
+  addEventListener() {},
+  head: {
+    appendChild(el) {
+      if (el.tagName === 'SCRIPT') injected.scripts.push(el);
+      else if (el.tagName === 'LINK') injected.links.push(el);
+      return el;
+    },
+  },
+  createElement(tag) {
+    if (tag === 'template') return makeTemplate();
+    return makeEl(tag);
+  },
+  querySelectorAll() { return []; },
+};
+globalThis.MutationObserver = class { observe() {} };
+
+let source = fs.readFileSync('./static/js/markdown.js', 'utf8');
+source = source.replace(/import uiModule from ['"]\.\/ui\.js['"];/, '');
+source = source.replace(
+  /import \{ splitTableRow \} from ['"]\.\/markdown\/tableRow\.js['"];/,
+  `function splitTableRow(row) {
+    return (row || '').replace(/^\s*\|/, '').replace(/\|\s*$/, '').split('|').map(c => c.trim());
+  }`
+);
+const emojiSource = fs.readFileSync('./static/js/emojiShortcodes.js', 'utf8')
+  .replace(/^export default .*$/m, '')
+  .replace(/export const /g, 'const ')
+  .replace(/export function /g, 'function ');
+source = source.replace(
+  /import \{ replaceEmojiShortcodes, hasEmojiShortcode \} from ['"]\.\/emojiShortcodes\.js['"];/,
+  () => emojiSource
+);
+source = source.replace(
+  /var escapeHtml = uiModule\.esc;/,
+  `var escapeHtml = (value) => String(value ?? '')
+    .replace(/&/g, '&amp;')
+    .replace(/</g, '&lt;')
+    .replace(/>/g, '&gt;')
+    .replace(/"/g, '&quot;')
+    .replace(/'/g, '&#39;');`
+);
+
+const moduleUrl = 'data:text/javascript;base64,' + Buffer.from(source).toString('base64');
+const mod = await import(moduleUrl);
+
+// A container whose querySelectorAll answers from a fixed element list, so a
+// test can hand the renderer exactly the nodes it wants it to see.
+function makeContainer(elements) {
+  return {
+    querySelectorAll(selector) {
+      return (elements[selector] || []).slice();
+    },
+  };
+}
+
+const emit = (value) => console.log(JSON.stringify(value));
+"""
+
+
+def _run_node(body: str, timeout: int = 20):
+    script = _HARNESS + textwrap.dedent(body)
+    result = subprocess.run(
+        ["node", "--input-type=module", "-e", script],
+        cwd=_REPO,
+        capture_output=True,
+        timeout=timeout,
+        text=True,
+    )
+    if result.returncode != 0:
+        raise AssertionError(f"node failed:\nSTDERR:\n{result.stderr}\nSTDOUT:\n{result.stdout}")
+    return json.loads(result.stdout.splitlines()[-1])
+
+
+def test_ensure_mermaid_shares_one_load_between_concurrent_callers(node_available):
+    """Two callers before the library lands must not trigger two fetches."""
+    out = _run_node(
+        """
+        const p1 = mod.ensureMermaid();
+        const p2 = mod.ensureMermaid();
+        const samePromise = p1 === p2;
+        const srcs = injected.scripts.map((s) => s.src);
+
+        let initializeCalls = 0;
+        globalThis.window.mermaid = {
+          initialize() { initializeCalls++; },
+          run() {},
+        };
+        injected.scripts[0].fire('load');
+
+        const [a, b] = await Promise.all([p1, p2]);
+        emit({
+          samePromise,
+          scriptCount: injected.scripts.length,
+          srcs,
+          sameLibrary: a === b && a === globalThis.window.mermaid,
+          initializeCalls,
+        });
+        """
+    )
+    assert out["samePromise"] is True
+    assert out["scriptCount"] == 1
+    assert out["srcs"] == [MERMAID_SRC]
+    assert out["sameLibrary"] is True
+    assert out["initializeCalls"] == 1
+
+
+def test_ensure_mermaid_retries_after_a_failed_load(node_available):
+    """A blocked first fetch must not poison every later diagram."""
+    out = _run_node(
+        """
+        const first = mod.ensureMermaid();
+        injected.scripts[0].fire('error');
+        let firstError = null;
+        try { await first; } catch (e) { firstError = e.message; }
+
+        const second = mod.ensureMermaid();
+        const retried = injected.scripts.length === 2;
+        globalThis.window.mermaid = { initialize() {}, run() {} };
+        injected.scripts[1].fire('load');
+        await second;
+
+        emit({ firstError, retried, differentPromise: first !== second });
+        """
+    )
+    assert MERMAID_SRC in out["firstError"]
+    assert out["retried"] is True
+    assert out["differentPromise"] is True
+
+
+def test_render_mermaid_does_not_fetch_when_no_diagram_is_present(node_available):
+    """The whole point of the lazy load: no fence, no 3.5 MB download."""
+    out = _run_node(
+        """
+        const container = makeContainer({});
+        await mod.renderMermaid(container);
+        emit({ scriptCount: injected.scripts.length });
+        """
+    )
+    assert out["scriptCount"] == 0
+
+
+def test_render_mermaid_fetches_once_a_diagram_is_present(node_available):
+    out = _run_node(
+        """
+        const node = makeEl('pre');
+        node.isConnected = true;
+        const container = makeContainer({ 'pre.mermaid:not([data-processed])': [node] });
+
+        const pending = mod.renderMermaid(container);
+        const srcs = injected.scripts.map((s) => s.src);
+
+        let ranWith = null;
+        globalThis.window.mermaid = {
+          initialize() {},
+          run(opts) { ranWith = opts.nodes.length; },
+        };
+        injected.scripts[0].fire('load');
+        await pending;
+
+        emit({ srcs, ranWith });
+        """
+    )
+    assert out["srcs"] == [MERMAID_SRC]
+    assert out["ranWith"] == 1
+
+
+def test_ensure_katex_loads_script_and_stylesheet_once(node_available):
+    out = _run_node(
+        """
+        const p1 = mod.ensureKatex();
+        const p2 = mod.ensureKatex();
+        const samePromise = p1 === p2;
+        const scriptSrcs = injected.scripts.map((s) => s.src);
+        const linkHrefs = injected.links.map((l) => l.href);
+
+        globalThis.window.katex = { renderToString: (src) => src };
+        injected.scripts[0].fire('load');
+        injected.links[0].fire('load');
+        await Promise.all([p1, p2]);
+
+        emit({ samePromise, scriptSrcs, linkHrefs });
+        """
+    )
+    assert out["samePromise"] is True
+    assert out["scriptSrcs"] == [KATEX_SRC]
+    assert out["linkHrefs"] == [KATEX_CSS]
+
+
+def test_md_to_html_defers_math_when_katex_is_not_loaded_yet(node_available):
+    """Without KaTeX the source is banked verbatim, not dropped or mangled."""
+    out = _run_node(
+        """
+        const html = mod.mdToHtml('Inline $x^2 + y_1$ and\\n\\n$$\\\\frac{a}{b}$$\\n');
+        emit({ html, scriptCount: injected.scripts.length });
+        """
+    )
+    html = out["html"]
+    assert 'class="ody-math-pending" data-display="false"' in html
+    assert 'class="ody-math-pending" data-display="true"' in html
+    # The raw source survives the escaping passes — `y_1` must not become <em>.
+    assert "x^2 + y_1" in html
+    assert "<em>" not in html
+    # Nothing is fetched during the synchronous render itself.
+    assert out["scriptCount"] == 0
+
+
+def test_deferred_math_schedules_a_katex_load(node_available):
+    """Deferring is only safe if the follow-up actually fires.
+
+    An earlier version scheduled this on requestAnimationFrame, which never runs
+    in a headless browser and is throttled to a stop in a background tab — math
+    then sat as plain source text until the tab was focused.
+    """
+    out = _run_node(
+        """
+        mod.mdToHtml('Inline $x^2$ here.');
+        const duringRender = injected.scripts.length;
+        await new Promise((r) => setTimeout(r, 0));
+        emit({ duringRender, scriptSrcs: injected.scripts.map((s) => s.src) });
+        """
+    )
+    assert out["duringRender"] == 0, "the synchronous render must not block on a fetch"
+    assert out["scriptSrcs"] == [KATEX_SRC]
+
+
+def test_entity_math_reaches_katex_as_characters_not_entities(node_available):
+    """"$a &lt; b$" and "$a < b$" must typeset the same, with no parse error.
+
+    mdToHtml escapes the source before the math pass, so a typed "<" arrives at
+    the delimiters as "&lt;" and a typed "&lt;" arrives as "&amp;lt;". KaTeX
+    has no entity syntax and treats the "&" as an alignment marker, so anything
+    still spelled as an entity comes back as a red .katex-error instead of a
+    formula. Both spellings have to be decoded to the character itself, in one
+    pass — decoding "&amp;" first and "&lt;" after would let the second pass eat
+    what the first produced, which is the double-unescape CodeQL flags.
+    """
+    out = _run_node(
+        """
+        const katex = loadRealKatex();
+        globalThis.window.katex = katex;
+        globalThis.katex = katex;
+        emit({
+          entity: mod.mdToHtml('Math: $a &lt; b$ done.'),
+          typed: mod.mdToHtml('Math: $a < b$ done.'),
+          ampersandEntity: mod.mdToHtml('Math: $x &gt; y$ done.'),
+        });
+        """
+    )
+    assert "katex-error" not in out["entity"]
+    assert "katex-error" not in out["typed"]
+    assert "katex-error" not in out["ampersandEntity"]
+    # Same formula, same markup, whichever way the author spelled the operator.
+    assert out["entity"] == out["typed"]
+    assert 'class="katex"' in out["entity"]
+
+
+def test_deferred_entity_math_banks_the_decoded_source(node_available):
+    """The placeholder has to hold the same source the inline path would use.
+
+    renderMath() feeds the span's textContent straight to KaTeX, so an entity
+    left in the bank is a .katex-error that only appears on a cold page — the
+    exact case the lazy load made common.
+    """
+    out = _run_node(
+        """
+        emit({
+          entity: mod.mdToHtml('Math: $a &lt; b$ done.'),
+          typed: mod.mdToHtml('Math: $a < b$ done.'),
+        });
+        """
+    )
+    assert 'class="ody-math-pending"' in out["entity"]
+    assert out["entity"] == out["typed"]
+    # Escaped once for transport, so the span's textContent is "a < b".
+    assert "a &lt; b</span>" in out["entity"]
+
+
+def test_detached_container_math_typesets_with_the_real_renderer(node_available):
+    """The PDF export renders into a container it never attaches to the page.
+
+    mdToHtml defers math to a document-scoped flush, which cannot reach a
+    detached node, so the export has to typeset its own container before
+    handing it to html2pdf. This is that container: pending spans in, real
+    KaTeX markup out, no .katex-error and nothing left pending.
+    """
+    out = _run_node(
+        """
+        const katex = loadRealKatex();
+        const html = mod.mdToHtml('Formula $E = mc^2$ here.');
+
+        const el = makeEl('span');
+        el.textContent = 'E = mc^2';
+        el.getAttribute = (name) => (name === 'data-display' ? 'false' : null);
+        let written = null;
+        Object.defineProperty(el, 'outerHTML', { set(v) { written = v; } });
+        const container = makeContainer({ '.ody-math-pending': [el] });
+
+        const pending = mod.renderMath(container);
+        globalThis.window.katex = katex;
+        injected.scripts[0].fire('load');
+        injected.links[0].fire('load');
+        await pending;
+
+        emit({ html, written });
+        """
+    )
+    # Cold page: mdToHtml could not typeset, so the export HTML starts pending.
+    assert 'class="ody-math-pending"' in out["html"]
+    # After the export's own render pass it is real KaTeX markup.
+    assert 'class="katex"' in out["written"]
+    assert "katex-error" not in out["written"]
+    assert "ody-math-pending" not in out["written"]
+
+
+def test_pdf_export_typesets_its_container_before_html2pdf():
+    """Ordering in a call site, so pin the call site. No node needed."""
+    source = (_REPO / "static/js/document.js").read_text(encoding="utf-8")
+    match = re.search(r"\n  async function exportAsPdf\(\) \{(.*?)\n  \}\n", source, re.S)
+    assert match, "exportAsPdf not found"
+    body = match.group(1)
+
+    render = "await markdownModule.renderMath(container);"
+    assert render in body, "the export never typesets its detached container"
+    assert body.index("container.innerHTML = html;") < body.index(render)
+    assert body.index(render) < body.index("window.html2pdf()")
+
+
+def test_md_to_html_renders_inline_once_katex_is_loaded(node_available):
+    """After the first load mdToHtml goes back to typesetting synchronously."""
+    out = _run_node(
+        """
+        globalThis.window.katex = {
+          renderToString: (src, opts) => `<span class="katex" data-display="${!!(opts && opts.displayMode)}">${src}</span>`,
+        };
+        globalThis.katex = globalThis.window.katex;
+        const html = mod.mdToHtml('Inline $x^2$ here.');
+        emit({ html });
+        """
+    )
+    assert '<span class="katex" data-display="false">x^2</span>' in out["html"]
+    assert "ody-math-pending" not in out["html"]
+
+
+def test_render_math_typesets_deferred_placeholders(node_available):
+    out = _run_node(
+        """
+        const el = makeEl('span');
+        el.textContent = 'x^2';
+        el.getAttribute = (name) => (name === 'data-display' ? 'false' : null);
+        let written = null;
+        Object.defineProperty(el, 'outerHTML', { set(v) { written = v; } });
+
+        const container = makeContainer({ '.ody-math-pending': [el] });
+        const pending = mod.renderMath(container);
+        const scriptSrcs = injected.scripts.map((s) => s.src);
+
+        globalThis.window.katex = {
+          renderToString: (src, opts) => `<span class="katex" data-display="${!!(opts && opts.displayMode)}">${src}</span>`,
+        };
+        injected.scripts[0].fire('load');
+        injected.links[0].fire('load');
+        await pending;
+
+        emit({ scriptSrcs, written });
+        """
+    )
+    assert out["scriptSrcs"] == [KATEX_SRC]
+    assert out["written"] == '<span class="katex" data-display="false">x^2</span>'
+
+
+def test_render_math_does_not_fetch_without_placeholders(node_available):
+    out = _run_node(
+        """
+        await mod.renderMath(makeContainer({}));
+        emit({ scriptCount: injected.scripts.length, linkCount: injected.links.length });
+        """
+    )
+    assert out["scriptCount"] == 0
+    assert out["linkCount"] == 0
+
+
+def test_vendored_assets_exist_and_index_html_has_no_cdn_reference():
+    """Guards the offline/privacy half: no node needed, so it always runs."""
+    for rel in (
+        "static/lib/mermaid.min.js",
+        "static/lib/katex/katex.min.js",
+        "static/lib/katex/katex.min.css",
+    ):
+        path = _REPO / rel
+        assert path.is_file(), f"{rel} is not vendored"
+        assert path.stat().st_size > 1024, f"{rel} looks truncated"
+
+    # KaTeX's stylesheet resolves fonts relative to itself; a missing font
+    # degrades silently to fallback glyphs, so resolve every woff2 the vendored
+    # CSS actually asks for. (.woff/.ttf are listed too but never requested by a
+    # browser that supports woff2, which is what static/fonts/ already assumes.)
+    css_dir = _REPO / "static/lib/katex"
+    css = (css_dir / "katex.min.css").read_text(encoding="utf-8")
+    wanted = sorted(set(re.findall(r"url\((fonts/KaTeX_[\w-]+\.woff2)\)", css)))
+    assert len(wanted) == 20, f"expected 20 woff2 references in the CSS, found {len(wanted)}"
+    missing = [ref for ref in wanted if not (css_dir / ref).is_file()]
+    assert missing == [], f"KaTeX stylesheet references fonts that are not vendored: {missing}"
+
+    # Everything the CSS needs must also survive an offline install: the font
+    # names have to be in KATEX_FONTS and that array has to reach PRECACHE.
+    sw = (_REPO / "static/sw.js").read_text(encoding="utf-8")
+    assert "...KATEX_FONTS," in sw, "KATEX_FONTS is defined but never spread into PRECACHE"
+    precached = {
+        f"fonts/KaTeX_{name}.woff2"
+        for name in re.findall(r"'([\w-]+)',", _katex_fonts_block(sw))
+    }
+    assert precached >= set(wanted), f"not precached: {sorted(set(wanted) - precached)}"
+
+    # The shell must fetch no resource from a third party. Scoped to the tags
+    # that actually load something — an <a href> to an external page is fine,
+    # and the comment explaining the move can keep naming the CDN it left.
+    index = (_REPO / "static/index.html").read_text(encoding="utf-8")
+    remote_loads = re.findall(r"<(?:script|link)\b[^>]*\b(?:src|href)=\"https?://[^\"]+", index)
+    assert remote_loads == [], f"index.html loads remote resources: {remote_loads}"
+
+    sw = (_REPO / "static/sw.js").read_text(encoding="utf-8")
+    assert KATEX_SRC in sw
+    assert KATEX_CSS in sw

tests/test_model_capability_readers.py

@@ -30,6 +30,15 @@ def test_detect_vendor_uses_endpoint_kind_then_host_and_common_local_ports():
     assert detect_vendor("http://localhost:7000/v1") == VENDOR_GENERIC_OPENAI
 
 
+def test_detect_vendor_requires_a_dns_label_boundary():
+    assert detect_vendor("https://api.openai.com./v1") == VENDOR_OPENAI
+    assert detect_vendor("https://notopenai.com/v1") == VENDOR_GENERIC_OPENAI
+    assert detect_vendor("https://fakeopenrouter.ai/v1") == VENDOR_GENERIC_OPENAI
+    assert detect_vendor("https://notgoogleapis.com/v1") == VENDOR_GENERIC_OPENAI
+    assert detect_vendor("https://evilanthropic.com/v1") == VENDOR_GENERIC_OPENAI
+    assert detect_vendor("https://fakeollama.com/v1") == VENDOR_GENERIC_OPENAI
+
+
 def test_generic_openai_reader_keeps_basic_model_payload_unknown():
     records = generic_openai.records_from_payload(
         {

tests/test_model_helper_owner_scope.py

@@ -27,7 +27,7 @@ def test_calendar_quick_parse_resolves_with_owner_scope():
 
 
 def test_task_parse_resolves_with_owner_scope():
-    body = _function_source("routes/task_routes.py", "parse_task")
+    body = _function_source("routes/task/task_routes.py", "parse_task")
     assert "user = _owner(request)" in body
     assert 'resolve_endpoint("utility", owner=user or None)' in body
     assert 'resolve_endpoint("default", owner=user or None)' in body

tests/test_model_routes.py

@@ -1107,6 +1107,47 @@ def test_get_api_models_marks_picker_as_pinned_only(monkeypatch):
     assert by_id["anthropic/claude-sonnet-4"]["is_pinned"] is False
 
 
+def test_get_fresh_api_models_marks_cached_inventory_as_pinned(monkeypatch):
+    ep = _make_endpoint(
+        base_url="https://api.example.test/v1",
+        cached_models=json.dumps(["openai/gpt-image-1", "anthropic/claude-sonnet-4"]),
+        pinned_models=None,
+    )
+    db = _PinnedFakeDb([ep])
+    monkeypatch.setattr(model_routes, "SessionLocal", lambda: db)
+    monkeypatch.setattr(model_routes, "require_admin", lambda request: None)
+    endpoint = _get_route("/api/model-endpoints/{ep_id}/models", "GET")
+
+    result = endpoint("ep1", _PinnedFakeRequest(), SimpleNamespace(headers={}))
+
+    assert [row["id"] for row in result] == [
+        "openai/gpt-image-1",
+        "anthropic/claude-sonnet-4",
+    ]
+    assert all(row["picker_requires_pinning"] is True for row in result)
+    assert all(row["is_pinned"] is True for row in result)
+
+
+def test_get_legacy_api_models_preserves_hidden_selection_as_pinned(monkeypatch):
+    ep = _make_endpoint(
+        base_url="https://api.example.test/v1",
+        cached_models=json.dumps(["m1", "m2", "m3"]),
+        hidden_models=json.dumps(["m2"]),
+        pinned_models=None,
+    )
+    db = _PinnedFakeDb([ep])
+    monkeypatch.setattr(model_routes, "SessionLocal", lambda: db)
+    monkeypatch.setattr(model_routes, "require_admin", lambda request: None)
+    endpoint = _get_route("/api/model-endpoints/{ep_id}/models", "GET")
+
+    result = endpoint("ep1", _PinnedFakeRequest(), SimpleNamespace(headers={}))
+
+    by_id = {row["id"]: row for row in result}
+    assert by_id["m1"]["is_pinned"] is True
+    assert by_id["m2"]["is_pinned"] is False
+    assert by_id["m3"]["is_pinned"] is True
+
+
 def test_reprobe_preserves_pinned_models(monkeypatch):
     ep = _make_endpoint(pinned_models=json.dumps(["deploy-1"]))
     db = _PinnedFakeDb([ep])
@@ -1266,7 +1307,7 @@ def test_list_model_endpoints_returns_key_fingerprint(monkeypatch):
     assert result[1]["api_key_fingerprint"] == ""
 
 
-def test_list_api_endpoint_reports_inventory_count_when_none_pinned(monkeypatch):
+def test_list_api_endpoint_defaults_inventory_to_visible_when_none_pinned(monkeypatch):
     ep = _make_endpoint(
         base_url="https://api.example.test/v1",
         cached_models=json.dumps(["openai/gpt-image-1", "anthropic/claude-sonnet-4"]),
@@ -1279,12 +1320,31 @@ def test_list_api_endpoint_reports_inventory_count_when_none_pinned(monkeypatch)
 
     result = endpoint(_PinnedFakeRequest())
 
-    assert result[0]["models"] == []
+    assert result[0]["models"] == [
+        "openai/gpt-image-1",
+        "anthropic/claude-sonnet-4",
+    ]
+    assert result[0]["pinned_models"] == result[0]["models"]
     assert result[0]["model_count"] == 2
     assert result[0]["picker_requires_pinning"] is True
     assert result[0]["status"] == "online"
 
 
+def test_api_picker_preserves_explicit_empty_selection():
+    ep = _make_endpoint(
+        base_url="https://api.example.test/v1",
+        cached_models=json.dumps(["openai/gpt-image-1", "anthropic/claude-sonnet-4"]),
+        pinned_models=json.dumps([]),
+    )
+
+    models, pinned = model_routes._picker_models_for_endpoint(
+        ep, ep.base_url, ep.endpoint_kind
+    )
+
+    assert models == []
+    assert pinned == []
+
+
 def test_list_api_endpoint_returns_pinned_picker_models(monkeypatch):
     ep = _make_endpoint(
         base_url="https://api.example.test/v1",
@@ -1767,7 +1827,7 @@ def test_api_models_openrouter_uses_pinned_models_not_hidden(monkeypatch):
     assert result["items"][0]["models"] == ["openai/gpt-image-1"]
 
 
-def test_api_models_openrouter_derives_legacy_visible_models(monkeypatch):
+def test_api_models_openrouter_defaults_cached_models_visible(monkeypatch):
     row = _route_ep(
         "openrouter",
         "https://openrouter.ai/api/v1",
@@ -1775,7 +1835,6 @@ def test_api_models_openrouter_derives_legacy_visible_models(monkeypatch):
         pinned_models=None,
         api_key="fake-key",
     )
-    row.hidden_models = json.dumps(["m2"])
     db = _RouteDb([row])
     router = model_routes.setup_model_routes(model_discovery=None)
 
@@ -1788,7 +1847,7 @@ def test_api_models_openrouter_derives_legacy_visible_models(monkeypatch):
     result = _route_endpoint(router, "/api/models")(_route_request())
 
     assert result["items"][0]["endpoint_name"] == "openrouter"
-    assert result["items"][0]["models"] == ["m1", "m3"]
+    assert result["items"][0]["models"] == ["m1", "m2", "m3"]
 
 
 @pytest.mark.asyncio

tests/test_panel_loader_js.py

@@ -0,0 +1,201 @@
+r"""Behaviour of the panel-loader registry in `static/js/panels.js`.
+
+The registry is what keeps a lazily-loaded panel honest: it must import a
+panel's module exactly once no matter how many times the user clicks, it must
+fail loudly on a name nobody registered (rather than handing back `undefined`
+and blowing up somewhere unrelated), and it must not memoise a *failed* load —
+otherwise a panel that failed to load once while offline would stay broken for
+the rest of the session.
+
+`createPanelLoader` is exercised with stub thunks so these tests need no DOM and
+no real panel module graph. The production `loadPanel` is the same function
+applied to the real registry, checked here only for its registered names.
+"""
+
+import json
+import os
+import re
+import shutil
+import subprocess
+import textwrap
+from pathlib import Path
+
+import pytest
+
+_REPO = Path(__file__).resolve().parent.parent
+_PANELS = (_REPO / "static" / "js" / "panels.js").as_posix()
+_HAS_NODE = shutil.which("node") is not None
+
+pytestmark = pytest.mark.skipif(not _HAS_NODE, reason="node binary not on PATH")
+
+
+def _run(js: str) -> str:
+    proc = subprocess.run(
+        ["node", "--input-type=module"],
+        input=js, capture_output=True, text=True, cwd=str(_REPO), timeout=30,
+    )
+    assert proc.returncode == 0, proc.stderr
+    return proc.stdout.strip()
+
+
+def test_repeated_loads_import_once_and_share_one_promise():
+    # A double-click on a rail button calls the loader twice before the first
+    # import has resolved. Both calls must ride the same in-flight import.
+    js = textwrap.dedent(
+        f"""
+        const {{ createPanelLoader }} = await import('{_PANELS}');
+        let calls = 0;
+        const load = createPanelLoader({{
+          demo: () => {{ calls += 1; return Promise.resolve({{ open: () => 'opened' }}); }},
+        }});
+        const a = load('demo');
+        const b = load('demo');
+        const mod = await a;
+        const c = load('demo');          // after it resolved, still cached
+        console.log(JSON.stringify({{
+          calls,
+          sameWhilePending: a === b,
+          sameAfterResolve: a === c,
+          opened: (await c).open(),
+          modOpened: mod.open(),
+        }}));
+        """
+    )
+    out = json.loads(_run(js))
+    assert out["calls"] == 1
+    assert out["sameWhilePending"] is True
+    assert out["sameAfterResolve"] is True
+    assert out["opened"] == "opened"
+    assert out["modOpened"] == "opened"
+
+
+def test_unknown_panel_throws_instead_of_returning_undefined():
+    js = textwrap.dedent(
+        f"""
+        const {{ createPanelLoader }} = await import('{_PANELS}');
+        const load = createPanelLoader({{ demo: () => Promise.resolve({{}}) }});
+        let message = null, returned = 'not-reached';
+        try {{ returned = load('nope'); }} catch (e) {{ message = e.message; }}
+        console.log(JSON.stringify({{ message, returned: String(returned) }}));
+        """
+    )
+    out = json.loads(_run(js))
+    assert out["message"] is not None
+    assert 'nope' in out["message"]
+    assert out["returned"] == "not-reached"
+
+
+def test_a_failed_load_is_not_memoised_so_a_retry_can_succeed():
+    # First open happens offline and the import 404s; the second, back online,
+    # must actually retry rather than replay the cached rejection.
+    js = textwrap.dedent(
+        f"""
+        const {{ createPanelLoader }} = await import('{_PANELS}');
+        let calls = 0;
+        const load = createPanelLoader({{
+          demo: () => {{
+            calls += 1;
+            return calls === 1
+              ? Promise.reject(new Error('offline'))
+              : Promise.resolve({{ open: () => 'opened' }});
+          }},
+        }});
+        let firstError = null;
+        try {{ await load('demo'); }} catch (e) {{ firstError = e.message; }}
+        const second = await load('demo');
+        console.log(JSON.stringify({{ calls, firstError, opened: second.open() }}));
+        """
+    )
+    out = json.loads(_run(js))
+    assert out["firstError"] == "offline"
+    assert out["calls"] == 2
+    assert out["opened"] == "opened"
+
+
+def test_a_thunk_that_throws_synchronously_rejects_rather_than_escaping():
+    js = textwrap.dedent(
+        f"""
+        const {{ createPanelLoader }} = await import('{_PANELS}');
+        const load = createPanelLoader({{ demo: () => {{ throw new Error('boom'); }} }});
+        let sync = null, rejected = null;
+        let p;
+        try {{ p = load('demo'); }} catch (e) {{ sync = e.message; }}
+        try {{ await p; }} catch (e) {{ rejected = e.message; }}
+        console.log(JSON.stringify({{ sync, rejected }}));
+        """
+    )
+    out = json.loads(_run(js))
+    assert out["sync"] is None, "a broken thunk must not throw at call time"
+    assert out["rejected"] == "boom"
+
+
+def test_the_image_editor_is_registered():
+    js = textwrap.dedent(
+        f"""
+        const {{ panelNames }} = await import('{_PANELS}');
+        console.log(JSON.stringify(panelNames()));
+        """
+    )
+    assert "editor" in json.loads(_run(js))
+
+
+# ── Offline coverage ──────────────────────────────────────────────────────
+# A lazily-loaded panel is never fetched during a normal page load, so it only
+# lands in the service-worker cache if sw.js precaches it explicitly. Miss one
+# module and the panel opens fine online and dies offline — the failure mode is
+# invisible until someone is on a plane. This walks the editor's import graph
+# and pins that every file in it is listed in PANEL_PRECACHE, query string
+# included (the SW matches on the full URL).
+
+_SW = _REPO / "static" / "sw.js"
+_JS_DIR = _REPO / "static" / "js"
+
+_STATIC_IMPORT = re.compile(
+    r"""(?:^|\n)\s*(?:import\s+(?:[^;'"()]*?\s+from\s+)?|export\s+(?:\*|\{[^}]*\})\s+from\s+)['"]([^'"]+)['"]"""
+)
+
+
+def _editor_module_graph() -> set[str]:
+    """Every module statically reachable from galleryEditor.js, as '<path>[?query]'."""
+    seen: set[str] = set()
+    stack = ["galleryEditor.js"]
+    while stack:
+        current = stack.pop()
+        if current in seen:
+            continue
+        seen.add(current)
+        source = (_JS_DIR / current.split("?")[0]).read_text(encoding="utf-8")
+        for spec in _STATIC_IMPORT.findall(source):
+            if not spec.startswith("."):
+                continue
+            path, _, query = spec.partition("?")
+            target = os.path.normpath(
+                os.path.join(os.path.dirname(current.split("?")[0]), path)
+            )
+            stack.append(f"{target}?{query}" if query else target)
+    return seen
+
+
+def _panel_precache_entries() -> set[str]:
+    block = re.search(
+        r"const PANEL_PRECACHE = \[(.*?)\];", _SW.read_text(encoding="utf-8"), re.S
+    )
+    assert block, "PANEL_PRECACHE not found in static/sw.js"
+    return set(re.findall(r"'([^']+)'", block.group(1)))
+
+
+def test_every_lazy_editor_module_is_precached_for_offline_use():
+    graph = _editor_module_graph()
+    precached = _panel_precache_entries()
+    # Shared modules (ui.js, spinner.js, ...) are still eagerly loaded by
+    # index.html, so they are cached by the normal page load. Only the part of
+    # the graph that nothing else pulls in needs an explicit entry.
+    lazy_only = {
+        m for m in graph
+        if m == "galleryEditor.js" or m.split("?")[0].startswith("editor/")
+    }
+    missing = {f"/static/js/{m}" for m in lazy_only} - precached
+    assert not missing, (
+        "these editor modules load lazily but are not in PANEL_PRECACHE, so the "
+        f"editor would not open offline: {sorted(missing)}"
+    )

tests/test_pr_description_check.py

@@ -169,6 +169,20 @@ def test_ui_checkbox_without_media_still_needs_visual_evidence():
     assert not any(call["method"] == "setFailed" for call in calls)
 
 
+def test_ticked_screenshot_box_counts_without_the_template_bolding():
+    body = _body(
+        app_ran=True,
+        screenshot=True,
+        media="https://github.com/user-attachments/assets/example",
+    ).replace("**Screenshot or short clip**", "Screenshot or short clip")
+
+    calls = _run_checker(["static/js/example.js"], body)
+
+    assert _added_labels(calls) == {"ready for review"}
+    assert not _comment(calls)
+    assert not any(call["method"] == "setFailed" for call in calls)
+
+
 def test_explicit_not_run_is_honest_but_not_ready():
     calls = _run_checker(
         ["services/example.py"],

tests/test_session_discovery_message_count.py

@@ -0,0 +1,62 @@
+"""Real-SQLite regressions for session discovery with stale derived counts."""
+
+from sqlalchemy import create_engine
+from sqlalchemy.orm import sessionmaker
+from sqlalchemy.pool import NullPool
+
+import core.session_manager as session_manager
+
+
+def _make_manager(db_path, monkeypatch):
+    engine = create_engine(
+        f"sqlite:///{db_path}",
+        connect_args={"check_same_thread": False},
+        poolclass=NullPool,
+    )
+    session_manager.DbSession.metadata.create_all(bind=engine)
+    session_local = sessionmaker(
+        bind=engine,
+        autoflush=False,
+        autocommit=False,
+    )
+    monkeypatch.setattr(session_manager, "SessionLocal", session_local)
+    return session_manager.SessionManager(), session_manager
+
+
+def test_discovery_uses_persisted_rows_and_repairs_cached_count(tmp_path, monkeypatch):
+    from core.models import ChatMessage
+
+    manager, session_manager = _make_manager(
+        tmp_path / "session-discovery.db", monkeypatch
+    )
+    for session_id in ("stale-low", "empty", "stale-high"):
+        manager.create_session(
+            session_id=session_id,
+            name=session_id,
+            endpoint_url="http://example.invalid",
+            model="test-model",
+            rag=False,
+            owner="tester",
+        )
+    manager.add_message("stale-low", ChatMessage("user", "persisted message"))
+
+    db = session_manager.SessionLocal()
+    try:
+        db.query(session_manager.DbSession).filter(
+            session_manager.DbSession.id == "stale-low"
+        ).update({"message_count": 0})
+        db.query(session_manager.DbSession).filter(
+            session_manager.DbSession.id == "stale-high"
+        ).update({"message_count": 7})
+        db.commit()
+    finally:
+        db.close()
+
+    restarted = session_manager.SessionManager()
+
+    assert set(restarted.sessions) == {"stale-low"}
+    assert restarted.sessions["stale-low"].history == []
+    assert restarted.sessions["stale-low"].message_count == 1
+
+    hydrated = restarted.get_session("stale-low")
+    assert [message.content for message in hydrated.history] == ["persisted message"]

tests/test_task_routes_shim.py

@@ -0,0 +1,20 @@
+"""Regression test for the task route shim (slice 2p, #4082/#4071).
+
+The backward-compat shim at ``routes/task_routes.py`` uses ``sys.modules``
+replacement so the legacy import path and the canonical ``routes.task.*``
+path resolve to the *same* module object. This is required because multiple
+tests do ``import routes.task_routes as task_routes`` followed by
+``monkeypatch.setattr(task_routes, "SessionLocal", ...)`` /
+``"get_current_user"``, and test_auth_regressions.py reads
+``task_routes.__file__`` for source introspection.
+"""
+
+import importlib
+
+import routes.task_routes as _shim_task  # noqa: F401
+
+
+def test_legacy_and_canonical_task_module_are_same_object():
+    legacy = importlib.import_module("routes.task_routes")
+    canonical = importlib.import_module("routes.task.task_routes")
+    assert legacy is canonical

tests/test_tidy_research_owner_scope.py

@@ -0,0 +1,159 @@
+"""Owner-scope tests for action_tidy_research.
+
+Broken research files (empty or unparseable JSON) have no readable owner
+stamp. The HTTP path and _find_owned_research_path treat parse failure as
+not-owned, so a regular user's tidy task must not unlink them. Clearing a
+corrupt record is a privileged act instead: admins, and the operator in
+single-user mode (AUTH_ENABLED=false), keep the janitor.
+
+Every case installs a fake AuthManager rather than relying on ambient state.
+Setting AUTH_ENABLED=true alone leaves is_configured False, so the admin check
+these tests exist to pin would never be reached.
+"""
+
+import json
+
+import pytest
+
+from src.builtin_actions import TaskNoop, action_tidy_research
+
+
+@pytest.fixture
+def research_dir(tmp_path, monkeypatch):
+    data_dir = tmp_path / "deep_research"
+    data_dir.mkdir()
+    monkeypatch.setattr("src.builtin_actions.DEEP_RESEARCH_DIR", str(data_dir))
+    return data_dir
+
+
+def _install_auth(monkeypatch, *, is_configured: bool):
+    """Auth enabled, with `root` as the only admin. `is_configured` False models
+    the pre-setup window, where no admin exists yet."""
+
+    import core.auth as core_auth
+
+    configured = is_configured
+
+    class FakeAuthManager:
+        is_configured = configured
+
+        def is_admin(self, user):
+            return user == "root"
+
+    monkeypatch.setenv("AUTH_ENABLED", "true")
+    monkeypatch.setattr(core_auth, "AuthManager", FakeAuthManager)
+
+
+@pytest.fixture
+def configured_auth(monkeypatch):
+    _install_auth(monkeypatch, is_configured=True)
+
+
+def _write(data_dir, name: str, text: str):
+    path = data_dir / f"{name}.json"
+    path.write_text(text, encoding="utf-8")
+    return path
+
+
+def _write_owned(data_dir, name: str, owner: str):
+    return _write(data_dir, name, json.dumps({"owner": owner, "query": name}))
+
+
+@pytest.mark.asyncio
+async def test_configured_non_admin_skips_unparseable_and_empty_files(
+    research_dir, configured_auth
+):
+    alice = _write_owned(research_dir, "alice-ok", "alice")
+    bob = _write_owned(research_dir, "bob-ok", "bob")
+    empty = _write(research_dir, "broken-empty", "")
+    garbage = _write(research_dir, "broken-json", "{not json")
+    alice_txt = alice.read_text(encoding="utf-8")
+    bob_txt = bob.read_text(encoding="utf-8")
+
+    with pytest.raises(TaskNoop, match="not permitted"):
+        await action_tidy_research("alice")
+
+    assert alice.exists()
+    assert alice.read_text(encoding="utf-8") == alice_txt
+    assert bob.exists()
+    assert bob.read_text(encoding="utf-8") == bob_txt
+    assert empty.exists(), "empty file has no owner stamp; non-admin tidy must not unlink it"
+    assert garbage.exists(), "unparseable file has no owner stamp; non-admin tidy must not unlink it"
+
+
+@pytest.mark.asyncio
+async def test_configured_non_admin_skip_reason_does_not_claim_none_broken(
+    research_dir, configured_auth
+):
+    """The skip reason reaches Activity as `skipped -- <reason>`, so it must not
+    report a clean scan over files it declined to look at."""
+    _write(research_dir, "broken-empty", "")
+
+    with pytest.raises(TaskNoop) as excinfo:
+        await action_tidy_research("alice")
+
+    assert "none broken" not in str(excinfo.value)
+
+
+@pytest.mark.asyncio
+async def test_configured_admin_still_removes_broken_files(research_dir, configured_auth):
+    valid = _write_owned(research_dir, "ok", "alice")
+    empty = _write(research_dir, "broken-empty", "   ")
+    garbage = _write(research_dir, "broken-json", "not-json")
+
+    message, ok = await action_tidy_research("root")
+
+    assert ok is True
+    assert "Removed 2" in message
+    assert valid.exists(), "admin sweep clears corrupt records, not valid ones"
+    assert not empty.exists()
+    assert not garbage.exists()
+
+
+@pytest.mark.asyncio
+async def test_pre_setup_denies_owner_who_would_be_admin_once_configured(
+    research_dir, monkeypatch
+):
+    """Auth on, no admin created yet: fail closed even for `root`.
+
+    The owner here is non-empty and would satisfy is_admin, so this fails if the
+    is_configured guard is removed. An empty owner would not discriminate -- it
+    is denied by the `owner and ...` short-circuit either way.
+    """
+    _install_auth(monkeypatch, is_configured=False)
+    empty = _write(research_dir, "broken-empty", "")
+    garbage = _write(research_dir, "broken-json", "{")
+
+    with pytest.raises(TaskNoop, match="not permitted"):
+        await action_tidy_research("root")
+
+    assert empty.exists()
+    assert garbage.exists()
+
+
+@pytest.mark.asyncio
+async def test_empty_owner_does_not_delete_broken_files(research_dir, configured_auth):
+    empty = _write(research_dir, "broken-empty", "")
+    garbage = _write(research_dir, "broken-json", "{")
+
+    with pytest.raises(TaskNoop, match="not permitted"):
+        await action_tidy_research("")
+
+    assert empty.exists()
+    assert garbage.exists()
+
+
+@pytest.mark.asyncio
+async def test_auth_disabled_still_removes_broken_files(research_dir, monkeypatch):
+    monkeypatch.setenv("AUTH_ENABLED", "false")
+    valid = _write_owned(research_dir, "ok", "alice")
+    empty = _write(research_dir, "broken-empty", "   ")
+    garbage = _write(research_dir, "broken-json", "not-json")
+
+    message, ok = await action_tidy_research("alice")
+
+    assert ok is True
+    assert "Removed 2" in message
+    assert valid.exists()
+    assert not empty.exists()
+    assert not garbage.exists()

tests/test_web_fetch_size_caps.py

@@ -6,6 +6,7 @@ notice, per-call override clamped to the hard cap, and a pre-buffer refusal
 when Content-Length already exceeds the hard ceiling.
 """
 import json
+import ipaddress
 from contextlib import contextmanager
 
 import pytest
@@ -90,7 +91,11 @@ class _FakeStream:
 def no_cache(monkeypatch, tmp_path):
     monkeypatch.setattr(content_mod, "CONTENT_CACHE_DIR", tmp_path)
     monkeypatch.setattr(content_mod, "_cache_result", lambda *a, **k: None)
-    monkeypatch.setattr(content_mod, "_public_http_url", lambda u: True)
+    monkeypatch.setattr(
+        content_mod,
+        "_resolve_public_ips",
+        lambda _url: [ipaddress.ip_address("93.184.216.34")],
+    )
 
 
 def _patch_stream(monkeypatch, fake):

tests/test_webhook_trigger_auth_exempt.py

@@ -87,6 +87,7 @@ def test_webhook_trigger_handler_still_validates_token():
     routes_path = os.path.join(
         os.path.dirname(os.path.dirname(os.path.abspath(__file__))),
         "routes",
+        "task",
         "task_routes.py",
     )
     with open(routes_path, encoding="utf-8") as fh: