chore(release): align dev version with 1.0.3

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

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

.env.example

@@ -88,8 +88,12 @@ SEARXNG_INSTANCE=http://localhost:8080
 # Keep false for Docker, LAN, reverse proxy, and any shared deployment.
 # LOCALHOST_BYPASS=false
 
-# Mark session cookies Secure. Set true when Odysseus is served through HTTPS
-# by a trusted reverse proxy or private access gateway.
+# Mark session cookies Secure. Left unset, this follows the request scheme:
+# an HTTPS login gets a Secure cookie, a plain-HTTP one does not. Set true to
+# force it on, or false to force it off while you still serve plain HTTP.
+# Upgrading: this used to default to false. Drop a leftover SECURE_COOKIES=false
+# from your .env unless you still need that escape hatch — it keeps HTTPS logins
+# on a non-Secure cookie.
 # SECURE_COOKIES=true
 
 # Optional: pre-seed the first admin password during setup.

.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');
       }

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/atomic_io.py

@@ -30,11 +30,20 @@ def atomic_write_json(path: str, data: Any, *, indent: Optional[int] = None) ->
     """
     os.makedirs(os.path.dirname(path) or ".", exist_ok=True)
     tmp = f"{path}.tmp.{uuid.uuid4().hex}"
-    with open(tmp, "w", encoding="utf-8") as f:
-        json.dump(data, f, indent=indent)
-        f.flush()
-        os.fsync(f.fileno())
-    os.replace(tmp, path)
+
+    try:
+        with open(tmp, "w", encoding="utf-8") as f:
+            json.dump(data, f, indent=indent)
+            f.flush()
+            os.fsync(f.fileno())
+        os.replace(tmp, path)
+    finally:
+        # Directly unlink to avoid a check-then-act race condition.
+        # Swallows FileNotFoundError (on success path) and other cleanup OSErrors.
+        try:
+            os.unlink(tmp)
+        except OSError:
+            pass
 
 
 def atomic_write_text(path: str, text: str) -> None:
@@ -42,8 +51,17 @@ def atomic_write_text(path: str, text: str) -> None:
         raise TypeError("atomic_write_text expects a string")
     os.makedirs(os.path.dirname(path) or ".", exist_ok=True)
     tmp = f"{path}.tmp.{uuid.uuid4().hex}"
-    with open(tmp, "w", encoding="utf-8") as f:
-        f.write(text)
-        f.flush()
-        os.fsync(f.fileno())
-    os.replace(tmp, path)
+
+    try:
+        with open(tmp, "w", encoding="utf-8") as f:
+            f.write(text)
+            f.flush()
+            os.fsync(f.fileno())
+        os.replace(tmp, path)
+    finally:
+        # Directly unlink to avoid a check-then-act race condition.
+        # Swallows FileNotFoundError (on success path) and other cleanup OSErrors.
+        try:
+            os.unlink(tmp)
+        except OSError:
+            pass

core/models.py

@@ -8,6 +8,11 @@ These are simple datacontainers. All persistence is handled by SessionManager.
 from dataclasses import dataclass
 from typing import Dict, List, Any, Optional, TYPE_CHECKING
 
+from src.tool_approval_scopes import (
+    CHAT_SESSION_APPROVAL_CONTEXT_MARKER,
+    CHAT_SESSION_APPROVAL_DECISION,
+)
+
 if TYPE_CHECKING:
     from .session_manager import SessionManager
 
@@ -31,6 +36,35 @@ set_session_manager = set_session_manager_instance
 get_session_manager = get_session_manager_instance
 
 
+def _history_grants_chat_session_approval(
+    history: List["ChatMessage"],
+    session_id: str,
+) -> bool:
+    """Return whether this exact chat has a resolved session-scope grant."""
+
+    expected_session = str(session_id or "")
+    if not expected_session:
+        return False
+    for message in reversed(history or []):
+        metadata = getattr(message, "metadata", None)
+        if not isinstance(metadata, dict):
+            continue
+        tool_events = metadata.get("tool_events")
+        if not isinstance(tool_events, list):
+            continue
+        for event in reversed(tool_events):
+            ask_user = event.get("ask_user") if isinstance(event, dict) else None
+            if not isinstance(ask_user, dict):
+                continue
+            if (
+                ask_user.get("kind") == "tool_approval"
+                and ask_user.get("resolved") == CHAT_SESSION_APPROVAL_DECISION
+                and str(ask_user.get("session_id") or "") == expected_session
+            ):
+                return True
+    return False
+
+
 @dataclass
 class ChatMessage:
     """A single chat message."""
@@ -116,11 +150,27 @@ class Session:
         the model. Display/history-load paths use the raw ``history`` and are
         unaffected.
         """
-        return [
+        messages = [
             msg.to_dict()
             for msg in self.history
             if (msg.metadata or {}).get("source") != "slash"
         ]
+        if not _history_grants_chat_session_approval(self.history, self.id):
+            return messages
+
+        # Keep the grant close to the latest user request so route-neutral
+        # compaction/trimming preserves it. Copy the metadata instead of
+        # mutating the durable transcript object.
+        for index in range(len(messages) - 1, -1, -1):
+            if messages[index].get("role") != "user":
+                continue
+            message = dict(messages[index])
+            metadata = dict(message.get("metadata") or {})
+            metadata[CHAT_SESSION_APPROVAL_CONTEXT_MARKER] = True
+            message["metadata"] = metadata
+            messages[index] = message
+            break
+        return messages
 
     def get(self, key: str, default=None):
         """Dict-like access for compatibility."""

core/session_manager.py

@@ -14,6 +14,8 @@ import logging
 from datetime import datetime, timezone, timedelta
 from typing import Dict, Optional
 
+from sqlalchemy import func
+
 from .database import Session as DbSession, ChatMessage as DbChatMessage, Document as DbDocument, SessionLocal, utcnow_naive
 from .models import Session, ChatMessage
 from src.attachment_refs import persistable_message_content
@@ -92,14 +94,28 @@ class SessionManager:
         try:
             db_sessions = db.query(DbSession).filter(
                 DbSession.archived == False,
-                DbSession.message_count > 0,
+                DbSession.messages.any(),
             ).order_by(DbSession.last_accessed.desc()).limit(100).all()
 
+            # message_count is derived metadata and can drift after interrupted
+            # or legacy writes. Count only the bounded discovery set so startup
+            # remains metadata-only while lazy hydration sees an authoritative
+            # positive count for every discovered non-empty session.
+            message_counts = {}
+            if db_sessions:
+                message_counts = dict(
+                    db.query(DbChatMessage.session_id, func.count(DbChatMessage.id))
+                    .filter(DbChatMessage.session_id.in_([row.id for row in db_sessions]))
+                    .group_by(DbChatMessage.session_id)
+                    .all()
+                )
+
             loaded_count = 0
             for db_session in db_sessions:
                 try:
                     session = self._db_to_session_meta(db_session)
                     if session is not None:
+                        session.message_count = message_counts[db_session.id]
                         self.sessions[db_session.id] = session
                         loaded_count += 1
                 except Exception as e:

docker-compose.gpu-amd.yml

@@ -50,7 +50,7 @@ services:
       - ODYSSEUS_ADMIN_USER=${ODYSSEUS_ADMIN_USER:-admin}
       - ODYSSEUS_ADMIN_PASSWORD=${ODYSSEUS_ADMIN_PASSWORD:-}
       - ALLOWED_ORIGINS=${ALLOWED_ORIGINS:-http://localhost,http://127.0.0.1}
-      - SECURE_COOKIES=${SECURE_COOKIES:-false}
+      - SECURE_COOKIES=${SECURE_COOKIES:-}
       - EMBEDDING_URL=${EMBEDDING_URL:-}
       - EMBEDDING_MODEL=${EMBEDDING_MODEL:-}
       - EMBEDDING_API_KEY=${EMBEDDING_API_KEY:-}

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/chat_helpers.py

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

routes/chat_routes.py

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

routes/cookbook_helpers.py

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

routes/cookbook_routes.py

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

routes/model_routes.py

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

routes/skills_routes.py

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

routes/task/__init__.py

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

scripts/mlx_image_server.py

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

scripts/odysseus-webhook

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

services/docs/service.py

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

services/search/content.py

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

src/agent_loop.py

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

src/chat_processor.py

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

src/constants.py

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

src/model_capability_readers/base.py

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

src/outbound_fetch.py

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

src/tool_approval_scopes.py

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

src/tool_approvals.py

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

src/tool_capabilities.py

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

src/user_time.py

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

static/app.js

@@ -10,8 +10,8 @@ import modelsModule from './js/models.js?v=20260715startupcalm2';
 import ragModule from './js/rag.js';
 import presetsModule from './js/presets.js';
 import searchModule from './js/search.js';
-import chatModule from './js/chat.js?v=20260815toolapproval4';
-import compareModule from './js/compare/index.js?v=20260723compareicon2';
+import chatModule from './js/chat.js?v=20260819approvalcontrol1';
+import compareModule from './js/compare/index.js?v=20260819approvalcontrol1';
 import documentModule from './js/document.js?v=20260815approvalsave1';
 import searchChatModule from './js/search-chat.js';
 import { makeWindowDraggable } from './js/windowDrag.js';
@@ -22,7 +22,7 @@ import {
   settleSessionHydration
 } from './js/startupShell.js';
 import markdownModule from './js/markdown.js';
-import chatRenderer from './js/chatRenderer.js?v=20260815toolapproval4';
+import chatRenderer from './js/chatRenderer.js?v=20260819approvalcontrol1';
 import sessionModule from './js/sessions.js';
 import memoryModule from './js/memory.js?v=20260722memoryloading1';
 import voiceRecorderModule from './js/voiceRecorder.js';

static/index.html

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

static/js/chat.js

@@ -8,8 +8,8 @@
 import Storage from './storage.js';
 import uiModule from './ui.js';
 import sessionModule from './sessions.js';
-import chatRenderer from './chatRenderer.js?v=20260815toolapproval4';
-import chatStream from './chatStream.js?v=20260815approvalsave1';
+import chatRenderer from './chatRenderer.js?v=20260819approvalcontrol1';
+import chatStream from './chatStream.js?v=20260819approvalcontrol1';
 import { addAITTSButton } from './tts-ai.js';
 import markdownModule from './markdown.js';
 import spinnerModule from './spinner.js';
@@ -62,20 +62,18 @@ import { loadPanel } from './panels.js';
   let _contextHeaderBound = false;
   let _pendingToolApproval = null;
 
-  function _submitToolApprovalWhenIdle(approvalId, label) {
+  function _submitToolApprovalWhenIdle(approvalId) {
     if (
       !_pendingToolApproval
       || _pendingToolApproval.approval_id !== approvalId
     ) return;
     if (isStreaming || _sendInFlight) {
-      setTimeout(() => _submitToolApprovalWhenIdle(approvalId, label), 120);
+      setTimeout(() => _submitToolApprovalWhenIdle(approvalId), 120);
       return;
     }
     const input = document.getElementById('message');
     if (input) {
       _pendingToolApproval.draft = input.value || '';
-      input.value = label;
-      input.dispatchEvent(new Event('input', { bubbles: true }));
     }
     const sendButton = document.querySelector('.send-btn');
     if (sendButton) sendButton.click();
@@ -84,16 +82,13 @@ import { loadPanel } from './panels.js';
   document.addEventListener('odysseus:tool-approval', (event) => {
     const detail = event && event.detail ? event.detail : {};
     const decision = String(detail.decision || '').toLowerCase();
-    if (!detail.approval_id || !['approve', 'deny'].includes(decision)) return;
+    if (!detail.approval_id || !['approve', 'approve_task', 'deny'].includes(decision)) return;
     _pendingToolApproval = {
       approval_id: String(detail.approval_id),
       decision,
       document_id: String(detail.document_id || ''),
     };
-    _submitToolApprovalWhenIdle(
-      _pendingToolApproval.approval_id,
-      detail.label || (decision === 'approve' ? 'Allow once' : 'Deny'),
-    );
+    _submitToolApprovalWhenIdle(_pendingToolApproval.approval_id);
   });
 
   function _fmtContextNumber(n) {
@@ -1309,10 +1304,10 @@ import { loadPanel } from './panels.js';
     }
 
     const el = uiModule.el;
-    const msg = el('message').value;
+    const msg = approvalForSend ? '' : el('message').value;
     // Allow empty text when a regen carries over the original message's
     // attachment ids — a photo-only message still has something to send.
-    if (!msg.trim() && !fileHandlerModule.getPendingCount() && !(_pendingRegenAttachments && _pendingRegenAttachments.length)) { _releaseSendFlag(); return; }
+    if (!msg.trim() && !approvalForSend && !fileHandlerModule.getPendingCount() && !(_pendingRegenAttachments && _pendingRegenAttachments.length)) { _releaseSendFlag(); return; }
 
     // --- Slash commands: execute directly without AI (no session needed) ---
     if (!approvalForSend && isCommand(msg.trim())) {
@@ -1590,7 +1585,7 @@ import { loadPanel } from './panels.js';
 
       const userDisplay = _displayOverride || msg;
       _displayOverride = null;
-      const skipBubble = _hideUserBubble;
+      const skipBubble = _hideUserBubble || !!approvalForSend;
       _hideUserBubble = false;
       // Auto-recovery counter: carries across a turn's auto-continues, but resets
       // when the user genuinely sends a new message (so each task gets a fresh cap).
@@ -1833,7 +1828,7 @@ import { loadPanel } from './panels.js';
       if (_inject.suffix) _finalMsgWithInject = _finalMsgWithInject + ' ' + _inject.suffix;
 
       const fd = new FormData();
-      fd.append('message', _finalMsgWithInject);
+      fd.append('message', approvalForSend ? '' : _finalMsgWithInject);
       fd.append('session', streamSessionId);
       if (approvalForSend) {
         fd.append('tool_approval_id', approvalForSend.approval_id);
@@ -2873,7 +2868,7 @@ import { loadPanel } from './panels.js';
                 if (spinner && spinner.element) spinner.destroy();
                 break;
               }
-              if (json.delta || json.type === 'agent_prep' || json.type === 'generated_image' || json.type === 'tool_start' || json.type === 'tool_output' || json.type === 'tool_progress' || json.type === 'agent_step' || json.type === 'loop_breaker_triggered' || json.type === 'intent_nudge_exhausted' || json.type === 'doc_stream_open' || json.type === 'doc_stream_delta' || json.type === 'research_progress') {
+              if (json.delta || json.type === 'agent_prep' || json.type === 'tool_approval_resolved' || json.type === 'generated_image' || json.type === 'tool_start' || json.type === 'tool_output' || json.type === 'tool_progress' || json.type === 'agent_step' || json.type === 'loop_breaker_triggered' || json.type === 'intent_nudge_exhausted' || json.type === 'doc_stream_open' || json.type === 'doc_stream_delta' || json.type === 'research_progress') {
                 clearResponseTimeout();
                 clearProcessingProbe();
                 clearFirstTokenWaitTimers();
@@ -2890,6 +2885,14 @@ import { loadPanel } from './panels.js';
                 }
                 continue;
               }
+              if (json.type === 'tool_approval_resolved') {
+                _cancelThinkingTimer();
+                _removeThinkingSpinner();
+                if (spinner && spinner.element) spinner.destroy();
+                if (!_isBg && roundHolder && roundHolder !== holder) roundHolder.remove();
+                if (!_isBg && holder) holder.remove();
+                continue;
+              }
               if (json.delta) {
                 _cancelThinkingTimer();
                 _removeThinkingSpinner();

static/js/chatRenderer.js

@@ -2327,6 +2327,42 @@ export function removeAskUserCards(root) {
   scope.querySelectorAll('.ask-user-card').forEach((node) => node.remove());
 }
 
+// While a choice card is visible, let plain 1–3 activate the corresponding
+// rendered option. Reuse the option's click path so the question keeps its
+// existing submission semantics. Tool approval cards are excluded: that card
+// exists to make consent deliberate after untrusted context influenced the
+// run, and its first option is the widest grant, so a stray digit must not
+// answer it.
+function _handleAskUserShortcut(event) {
+  if (
+    event.defaultPrevented
+    || event.repeat
+    || event.isComposing
+    || event.ctrlKey
+    || event.altKey
+    || event.metaKey
+    || event.shiftKey
+  ) return;
+  if (!/^[1-3]$/.test(event.key)) return;
+
+  const target = event.target;
+  if (target?.closest?.('input, textarea, select, [contenteditable="true"]')) return;
+
+  const focusedCard = document.activeElement?.closest?.('.ask-user-card') || null;
+  const mainCard = document.querySelector('#chat-history .ask-user-card');
+  const compareCards = document.querySelectorAll('.compare-pane .ask-user-card');
+  const card = focusedCard || mainCard || (compareCards.length === 1 ? compareCards[0] : null);
+  if (!card) return;
+  if (card.dataset.askUserKind === 'tool_approval') return;
+  const option = card.querySelectorAll('.ask-user-option')[Number(event.key) - 1];
+  if (!option || option.disabled) return;
+
+  event.preventDefault();
+  option.click();
+}
+
+document.addEventListener('keydown', _handleAskUserShortcut);
+
 /**
  * Render an ask_user payload as a durable choice card.
  *
@@ -2336,11 +2372,15 @@ export function removeAskUserCards(root) {
  */
 export function renderAskUserCard(payload, options) {
   const aq = payload || {};
+  if (aq.resolved) return null;
   const opts = Array.isArray(aq.options) ? aq.options : [];
-  const chatBox = document.getElementById('chat-history');
+  const renderOptions = options || {};
+  const chatBox = renderOptions.root || document.getElementById('chat-history');
+  const onSubmit = typeof renderOptions.onSubmit === 'function'
+    ? renderOptions.onSubmit
+    : null;
   if (!chatBox || !aq.question || opts.length < 2) return null;
 
-  const renderOptions = options || {};
   removeAskUserCards(chatBox);
 
   const card = document.createElement('div');
@@ -2349,6 +2389,7 @@ export function renderAskUserCard(payload, options) {
   card.tabIndex = -1;
   const multi = !!aq.multi;
   const isToolApproval = aq.kind === 'tool_approval' && !!aq.approval_id;
+  card.dataset.askUserKind = isToolApproval ? 'tool_approval' : 'question';
   const emojiText = (value) => svgifyEmoji(uiModule.esc(String(value)));
 
   const head = document.createElement('div');
@@ -2357,7 +2398,6 @@ export function renderAskUserCard(payload, options) {
   closeBtn.type = 'button';
   closeBtn.className = 'modal-close ask-user-close';
   closeBtn.setAttribute('aria-label', 'Dismiss question');
-  closeBtn.textContent = '×';
   closeBtn.addEventListener('click', () => {
     card.remove();
     const input = uiModule.el('message');
@@ -2400,6 +2440,17 @@ export function renderAskUserCard(payload, options) {
 
   const send = (text) => {
     if (!text) return;
+    if (onSubmit) {
+      const accepted = onSubmit({
+        kind: 'answer',
+        text,
+        label: text,
+        payload: aq,
+        card,
+      });
+      if (accepted !== false) card.remove();
+      return;
+    }
     card.remove();
     const input = uiModule.el('message');
     if (input) input.value = text;
@@ -2433,17 +2484,26 @@ export function renderAskUserCard(payload, options) {
       row.type = 'button';
       row.addEventListener('click', () => {
         if (isToolApproval) {
-          card.remove();
-          document.dispatchEvent(new CustomEvent('odysseus:tool-approval', {
-            detail: {
-              approval_id: aq.approval_id,
-              decision: String((opt && opt.value) || '').toLowerCase(),
-              label,
-              document_id: aq.action && aq.action.document_id
-                ? String(aq.action.document_id)
-                : '',
-            },
-          }));
+          const detail = {
+            approval_id: aq.approval_id,
+            decision: String((opt && opt.value) || '').toLowerCase(),
+            label,
+            document_id: aq.action && aq.action.document_id
+              ? String(aq.action.document_id)
+              : '',
+          };
+          if (onSubmit) {
+            const accepted = onSubmit({
+              kind: 'tool_approval',
+              ...detail,
+              payload: aq,
+              card,
+            });
+            if (accepted !== false) card.remove();
+          } else {
+            card.remove();
+            document.dispatchEvent(new CustomEvent('odysseus:tool-approval', { detail }));
+          }
         } else {
           send(label);
         }
@@ -2628,7 +2688,7 @@ export function addMessage(role, content, modelName, metadata) {
             box.appendChild(threadWrap);
           }
           for (const ev of roundTools) {
-            if (ev.ask_user) pendingAskUser = ev.ask_user;
+            if (ev.ask_user && !ev.ask_user.resolved) pendingAskUser = ev.ask_user;
             const ok = (ev.exit_code === 0 || ev.exit_code == null);
             let outHtml = '';
             if (ev.output && ev.output.trim()) {

static/js/chatStream.js

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

static/js/compare/index.js

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

static/js/compare/stream.js

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

static/js/cookbookDownload.js

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

static/js/cookbookRunning.js

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

static/js/document.js

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

static/js/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/sw.js

@@ -7,7 +7,21 @@
 //   - 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-v378-shared-config-image-editor';
+const CACHE_NAME = 'odysseus-v380-shared-config-image-editor-lazy-katex-mermaid';
+
+// 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:
@@ -73,6 +87,15 @@ 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;

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/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_atomic_io.py

@@ -123,7 +123,7 @@ def test_atomic_write_json_concurrent_writers_do_not_collide(tmp_path):
 
 
 # ---------------------------------------------------------------------------
-# atomic_write_json — failure path: target preserved on serialization error.
+# atomic_write_json — failure paths
 # ---------------------------------------------------------------------------
 def test_atomic_write_json_preserves_target_when_serialization_fails(tmp_path):
     target = tmp_path / "data.json"
@@ -136,6 +136,26 @@ def test_atomic_write_json_preserves_target_when_serialization_fails(tmp_path):
         atomic_write_json(str(target), {"bad": {1, 2, 3}})
 
     assert target.read_text(encoding="utf-8") == before
+    # Temp file should be cleaned up
+    assert _tmp_siblings(tmp_path, "data.json") == []
+
+
+def test_atomic_write_json_preserves_target_when_replace_fails(tmp_path, monkeypatch):
+    target = tmp_path / "data.json"
+    atomic_write_json(str(target), {"existing": "value"})
+    before = target.read_text(encoding="utf-8")
+
+    def boom(src, dst):
+        raise PermissionError("replace failed")
+
+    monkeypatch.setattr(atomic_io.os, "replace", boom)
+
+    with pytest.raises(PermissionError, match="replace failed"):
+        atomic_write_json(str(target), {"new": "content"})
+
+    assert target.read_text(encoding="utf-8") == before
+    # Temp file should be cleaned up
+    assert _tmp_siblings(tmp_path, "data.json") == []
 
 
 # ---------------------------------------------------------------------------
@@ -187,7 +207,7 @@ def test_atomic_write_text_rejects_non_string_before_tmp_file(tmp_path):
 
 
 # ---------------------------------------------------------------------------
-# atomic_write_text — failure path: target preserved when replace fails.
+# atomic_write_text — failure paths
 # ---------------------------------------------------------------------------
 def test_atomic_write_text_preserves_target_when_replace_fails(tmp_path, monkeypatch):
     target = tmp_path / "note.txt"
@@ -195,11 +215,32 @@ def test_atomic_write_text_preserves_target_when_replace_fails(tmp_path, monkeyp
     before = target.read_text(encoding="utf-8")
 
     def boom(src, dst):
-        raise OSError("replace failed")
+        raise PermissionError("replace failed")
 
     monkeypatch.setattr(atomic_io.os, "replace", boom)
 
-    with pytest.raises(OSError):
+    with pytest.raises(PermissionError, match="replace failed"):
         atomic_write_text(str(target), "new content that never lands")
 
     assert target.read_text(encoding="utf-8") == before
+    # Temp file should be cleaned up
+    assert _tmp_siblings(tmp_path, "note.txt") == []
+
+
+def test_cleanup_error_swallows_and_preserves_original_exception(tmp_path, monkeypatch):
+    target = tmp_path / "note.txt"
+    atomic_write_text(str(target), "original content")
+
+    def replace_boom(src, dst):
+        raise PermissionError("replace failed")
+
+    def unlink_boom(path):
+        raise OSError("unlink failed")
+
+    monkeypatch.setattr(atomic_io.os, "replace", replace_boom)
+    monkeypatch.setattr(atomic_io.os, "unlink", unlink_boom)
+
+    # If BOTH the replace fails AND the cleanup unlink fails,
+    # the original replace error should surface, completely swallowing the unlink error.
+    with pytest.raises(PermissionError, match="replace failed"):
+        atomic_write_text(str(target), "new 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_compare_ask_user_routing.py

@@ -0,0 +1,61 @@
+from pathlib import Path
+
+
+def test_compare_renders_ask_user_in_the_originating_pane():
+    root = Path(__file__).resolve().parents[1]
+    stream = (root / "static/js/compare/stream.js").read_text(encoding="utf-8")
+
+    assert "renderAskUserCard" in stream
+    assert "} else if (json.type === 'ask_user') {" in stream
+    assert "root: hist" in stream
+    assert "state._paneSessionIds[paneIdx] === sessionId" in stream
+    assert "streamToPane(paneIdx, sessionId, message, aiMessage, resumeOptions)" in stream
+    assert "handleCompareSubmit" not in stream
+
+
+def test_compare_submits_approval_only_to_the_pane_session():
+    root = Path(__file__).resolve().parents[1]
+    stream = (root / "static/js/compare/stream.js").read_text(encoding="utf-8")
+
+    assert "fd.append('tool_approval_id', opts.toolApproval.approval_id || '');" in stream
+    assert "fd.append('tool_approval_decision', opts.toolApproval.decision || '');" in stream
+    assert "const isApproval = submission.kind === 'tool_approval';" in stream
+    assert "const message = isApproval ? ''" in stream
+    assert "if (!isApproval) _appendPaneMessage(hist, 'user', message);" in stream
+    assert "json.type === 'tool_approval_resolved'" in stream
+    assert "json.decision === 'deny' ? 'Denied.'" in stream
+
+
+def test_compare_continuation_does_not_lose_or_replace_pane_ownership():
+    root = Path(__file__).resolve().parents[1]
+    stream = (root / "static/js/compare/stream.js").read_text(encoding="utf-8")
+    index = (root / "static/js/compare/index.js").read_text(encoding="utf-8")
+
+    assert "if (state._abortControllers[paneIdx] === ac)" in stream
+    assert "if (activeController === originController)" in stream
+    assert "if (activeController) return;" in stream
+    assert "registerStreamActions({ rerollPane, autoPreviewHtml: _autoPreviewHtml, setSendBtn: _setSendBtn });" in index
+    assert "const compareStillStreaming = state._abortControllers.some(Boolean);" in index
+    assert "_setSendBtn(compareStillStreaming ? 'stop' : 'send');" in index
+
+
+def test_compare_restores_the_card_instead_of_dropping_a_timed_out_choice():
+    """The card is removed the moment a choice is accepted.
+
+    If the originating stream still owns the pane when the resume deadline
+    passes, the decision has nowhere to go — so the card has to come back
+    rather than the click vanishing silently.
+    """
+
+    root = Path(__file__).resolve().parents[1]
+    stream = (root / "static/js/compare/stream.js").read_text(encoding="utf-8")
+
+    assert "function _restorePaneAskUserCard(" in stream
+    assert "_restorePaneAskUserCard(paneIdx, sessionId, submission, originController);" in stream
+    assert "submission.payload || {}" in stream
+
+    start = stream.index("function _resumePaneChoiceWhenIdle(")
+    end = stream.index("function _renderPaneAskUserCard(", start)
+    resume = stream[start:end]
+    # The deadline must not fall through to a bare return any more.
+    assert "if (Date.now() - startedAt < 10000) setTimeout(resume, 25);" not in resume

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_foreground_model_routing.py

@@ -344,7 +344,7 @@ async def test_chat_stream_approval_restores_exact_shell_turn_toggle(monkeypatch
 
 
 @pytest.mark.asyncio
-async def test_chat_stream_denial_keeps_originating_run_tainted(monkeypatch):
+async def test_chat_stream_denial_returns_control_resolution(monkeypatch):
     from src.tool_capabilities import capabilities_for_action
 
     captured = {}
@@ -368,11 +368,13 @@ async def test_chat_stream_denial_keeps_originating_run_tainted(monkeypatch):
     )
 
     response = await endpoint(request)
-    async for _ in response.body_iterator:
-        pass
+    chunks = [chunk async for chunk in response.body_iterator]
 
+    event = json.loads(chunks[0][len("data: "):])
+    assert event == {"type": "tool_approval_resolved", "decision": "deny"}
+    assert chunks[-1] == "data: [DONE]\n\n"
+    assert "agent" not in captured
     assert "exact_approval" not in captured
-    assert captured["agent_external_untrusted_context_seen"] is True
     assert chat_routes.tool_approval_store.peek(pending.approval_id) is None
 
 

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_mlx_image_server_security.py

@@ -0,0 +1,140 @@
+"""Pin the mlx_image_server caller-chosen-model + DNS-rebinding regressions.
+
+Background: scripts/mlx_image_server.py used to resolve the model per request
+(``req.model or _args.model``) instead of serving the model the process was
+launched with. ``_is_hidream`` is a substring test and ``_snapshot_path``
+accepts either a local directory or a Hugging Face repo id, so a caller could
+name any directory / repo and the HiDream branch would then run
+``<model>/scripts/hidream_o1/generate_hidream_o1_mlx.py`` under
+``sys.executable``. The server has no auth, and the cookbook binds it to
+``0.0.0.0`` whenever it is serving to a remote host, so that was reachable
+code execution.
+
+The fix pins both request paths to ``_args.model``, matching
+scripts/diffusion_server.py.
+"""
+
+import argparse
+import importlib.util
+import sys
+import types
+from pathlib import Path
+
+import pytest
+
+_SCRIPT = Path(__file__).resolve().parent.parent / "scripts" / "mlx_image_server.py"
+
+_BASE_URL = "http://127.0.0.1"
+
+
+def _load_module():
+    """Fresh import of the server module. Unlike diffusion_server it pulls in no
+    heavy runtime (mlx / torch imports all live inside the request handlers), so
+    the real module is imported rather than AST-extracted."""
+    spec = importlib.util.spec_from_file_location("mlx_image_server_under_test", _SCRIPT)
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)
+    return module
+
+
+@pytest.fixture
+def server(monkeypatch):
+    """Server module launched with a pinned, non-HiDream model."""
+    module = _load_module()
+    module._args = argparse.Namespace(
+        model="mlx-community/pinned-model",
+        host="127.0.0.1",
+        port=8100,
+        steps=0,
+        width=512,
+        height=512,
+        base_model="",
+        lora_style="",
+        lora_paths=[],
+        lora_scales=[],
+        vlm_model="",
+    )
+    return module
+
+
+def _client(module):
+    from fastapi.testclient import TestClient
+
+    return TestClient(module.app, base_url=_BASE_URL)
+
+
+def _plant_hidream_model_dir(tmp_path: Path) -> tuple[Path, Path]:
+    """A directory that satisfies _is_hidream() and carries the script the
+    HiDream branch executes. The script writes a marker so the test can tell
+    whether it ran."""
+    model_dir = tmp_path / "hidream-planted"
+    generator = model_dir / "scripts" / "hidream_o1"
+    generator.mkdir(parents=True)
+    marker = model_dir / "executed.txt"
+    (generator / "generate_hidream_o1_mlx.py").write_text(
+        f"open({str(marker)!r}, 'w').write('ran')\n", encoding="utf-8"
+    )
+    return model_dir, marker
+
+
+def test_generate_does_not_run_code_from_a_caller_named_model_dir(server, tmp_path):
+    """The regression: naming a local directory as the model must not execute
+    the generator script inside it."""
+    model_dir, marker = _plant_hidream_model_dir(tmp_path)
+
+    _client(server).post(
+        "/v1/images/generations",
+        json={"model": str(model_dir), "prompt": "x", "size": "64x64"},
+    )
+
+    assert not marker.exists(), (
+        "code inside the caller-named model directory ran; the request model "
+        "must not select the generator"
+    )
+
+
+def test_generate_does_not_fetch_a_caller_named_repo(server, monkeypatch):
+    """The remote half of the same defect: the caller's string must never reach
+    the Hugging Face downloader."""
+    downloaded = []
+    stub = types.ModuleType("huggingface_hub")
+    stub.snapshot_download = lambda repo: downloaded.append(repo)
+    monkeypatch.setitem(sys.modules, "huggingface_hub", stub)
+
+    _client(server).post(
+        "/v1/images/generations",
+        json={"model": "attacker-account/hidream-anything", "prompt": "x"},
+    )
+
+    assert "attacker-account/hidream-anything" not in downloaded
+
+
+def test_edits_does_not_run_code_from_a_caller_named_model_dir(server, tmp_path):
+    """/v1/images/edits resolved the model the same way and must be pinned too.
+    A "lama" name reaches the inpaint bridge, so the caller's model string is
+    what picks the branch here."""
+    model_dir = tmp_path / "lama-planted"
+    model_dir.mkdir()
+    called = []
+    server._run_inpaint_bridge = lambda *a, **kw: called.append(a)
+    server._run_ddcolor_bridge = lambda *a, **kw: called.append(a)
+
+    resp = _client(server).post(
+        "/v1/images/edits",
+        data={"model": str(model_dir), "prompt": "x"},
+        files={"image": ("i.png", b"not-a-real-png", "image/png")},
+    )
+
+    assert not called, "caller-supplied model selected the edit branch"
+    assert resp.status_code == 422, "pinned non-edit model should be refused"
+
+
+def test_pinned_hidream_model_is_still_served(server, tmp_path, monkeypatch):
+    """Behaviour preservation: pinning must not break a server that was actually
+    launched with a HiDream model."""
+    model_dir, marker = _plant_hidream_model_dir(tmp_path)
+    server._args.model = str(model_dir)
+
+    _client(server).post("/v1/images/generations", json={"model": "", "prompt": "x"})
+
+    assert marker.exists(), "the model this server was launched with must still run"

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_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_tool_approval_frontend_routing.py

@@ -0,0 +1,103 @@
+from pathlib import Path
+
+
+def test_tool_approval_bypasses_polymorphic_send_button_actions():
+    root = Path(__file__).resolve().parents[1]
+    chat = (root / "static/js/chat.js").read_text(encoding="utf-8")
+    stream = (root / "static/js/chatStream.js").read_text(encoding="utf-8")
+
+    # chat.js still defers the sealed approval through a synthetic button click.
+    assert "if (sendButton) sendButton.click();" in chat
+
+    # The capture listener must intercept only that synthetic click and route it
+    # through the chat form submit path, before app.js can reinterpret an empty
+    # composer as New chat or Record voice.
+    assert "if (event.isTrusted) return;" in stream
+    assert "event.stopImmediatePropagation();" in stream
+    assert "chatForm.requestSubmit()" in stream
+    assert "sendButton.dataset.mode = ''" not in stream
+
+
+def test_ask_user_close_button_uses_one_css_glyph():
+    root = Path(__file__).resolve().parents[1]
+    renderer = (root / "static/js/chatRenderer.js").read_text(encoding="utf-8")
+    styles = (root / "static/style.css").read_text(encoding="utf-8")
+
+    assert "closeBtn.className = 'modal-close ask-user-close';" in renderer
+    assert "closeBtn.setAttribute('aria-label', 'Dismiss question');" in renderer
+    assert "closeBtn.textContent = '×';" not in renderer
+    assert ".modal-close::before" in styles
+
+
+def test_ask_user_number_shortcuts_reuse_option_click_path():
+    root = Path(__file__).resolve().parents[1]
+    renderer = (root / "static/js/chatRenderer.js").read_text(encoding="utf-8")
+    start = renderer.index("function _handleAskUserShortcut(event)")
+    end = renderer.index("document.addEventListener('keydown', _handleAskUserShortcut);", start)
+    shortcut = renderer[start:end]
+
+    assert "if (!/^[1-3]$/.test(event.key)) return;" in shortcut
+    assert "event.repeat" in shortcut
+    assert "event.ctrlKey" in shortcut
+    assert "event.altKey" in shortcut
+    assert "event.metaKey" in shortcut
+    assert "event.shiftKey" in shortcut
+    assert "input, textarea, select, [contenteditable=\"true\"]" in shortcut
+    assert "card.querySelectorAll('.ask-user-option')[Number(event.key) - 1]" in shortcut
+    assert "event.preventDefault();" in shortcut
+    assert "option.click();" in shortcut
+
+
+def test_digit_shortcuts_never_answer_a_tool_approval_card():
+    """A stray digit must not grant a scope the user did not deliberately pick."""
+
+    root = Path(__file__).resolve().parents[1]
+    renderer = (root / "static/js/chatRenderer.js").read_text(encoding="utf-8")
+    start = renderer.index("function _handleAskUserShortcut(event)")
+    end = renderer.index("document.addEventListener('keydown', _handleAskUserShortcut);", start)
+    shortcut = renderer[start:end]
+
+    assert "if (card.dataset.askUserKind === 'tool_approval') return;" in shortcut
+    # The renderer has to label the card for that guard to ever fire.
+    assert (
+        "card.dataset.askUserKind = isToolApproval ? 'tool_approval' : 'question';"
+        in renderer
+    )
+
+
+def test_ask_user_renderer_accepts_scoped_root_and_submit_callback():
+    root = Path(__file__).resolve().parents[1]
+    renderer = (root / "static/js/chatRenderer.js").read_text(encoding="utf-8")
+
+    assert "const chatBox = renderOptions.root || document.getElementById('chat-history');" in renderer
+    assert "const onSubmit = typeof renderOptions.onSubmit === 'function'" in renderer
+    assert "kind: 'answer'" in renderer
+    assert "kind: 'tool_approval'" in renderer
+    assert "if (accepted !== false) card.remove();" in renderer
+    assert "document.dispatchEvent(new CustomEvent('odysseus:tool-approval', { detail }))" in renderer
+
+
+def test_every_changed_approval_module_is_cache_busted_together():
+    """A stale module here silently reinterprets the approval click.
+
+    chat.js leaves the composer empty and clicks the polymorphic send button,
+    so a browser that pairs the new chat.js with a cached chatStream.js has no
+    interceptor and lands on the New chat branch instead. The same holds for
+    the compare pane modules, which chatRenderer now shares a keydown listener
+    with.
+    """
+
+    root = Path(__file__).resolve().parents[1]
+    version = "20260819approvalcontrol1"
+    index = (root / "static/index.html").read_text(encoding="utf-8")
+    app = (root / "static/app.js").read_text(encoding="utf-8")
+    chat = (root / "static/js/chat.js").read_text(encoding="utf-8")
+    compare_index = (root / "static/js/compare/index.js").read_text(encoding="utf-8")
+    compare_stream = (root / "static/js/compare/stream.js").read_text(encoding="utf-8")
+
+    assert f"chatStream.js?v={version}" in index
+    assert f"chatStream.js?v={version}" in chat
+    assert f"compare/index.js?v={version}" in app
+    assert f"stream.js?v={version}" in compare_index
+    # One chatRenderer instance, so the ask_user keydown listener binds once.
+    assert f"chatRenderer.js?v={version}" in compare_stream

tests/test_tool_approval_single_action_scope.py

@@ -0,0 +1,89 @@
+"""Callers with no resumable chat keep the original one-use approval scope.
+
+The chat card reuses the wire value ``approve`` for chat-session scope, so any
+caller that still sends ``approve`` meaning "once" has to say so explicitly or
+it silently inherits a run-long gate bypass.
+"""
+
+from pathlib import Path
+
+from src.tool_approval_scopes import ToolApprovalScope
+from src.tool_approvals import ToolApprovalStore
+from src.tool_capabilities import ToolRunSecurityContext, capabilities_for_action
+
+
+def _pending(store: ToolApprovalStore, *, session_id=""):
+    content = "printf exact"
+    return store.create(
+        owner="Alice",
+        session_id=session_id,
+        origin_run_id="run-1",
+        tool_name="bash",
+        content=content,
+        workspace=None,
+        external_untrusted_context_seen=True,
+        capabilities=capabilities_for_action("bash", content),
+    )
+
+
+def test_single_action_grant_leaves_the_gate_armed_behind_the_sealed_action():
+    store = ToolApprovalStore()
+    pending = _pending(store)
+
+    grant = store.consume(
+        pending.approval_id,
+        decision="approve",
+        owner="alice",
+        session_id=None,
+        allow_continuation=False,
+    )
+
+    assert grant is not None
+    assert grant.scope is ToolApprovalScope.SINGLE_ACTION
+    assert grant.allow_remaining_actions is False
+    assert grant.grants_chat_session is False
+
+    resumed = ToolRunSecurityContext(
+        external_untrusted_context_seen=True,
+        approval_gate_bypassed=grant.allow_remaining_actions,
+    )
+    assert resumed.decision_for("bash").allowed is False
+
+
+def test_chat_callers_still_get_the_continuation_scope_they_asked_for():
+    store = ToolApprovalStore()
+    pending = _pending(store, session_id="session-1")
+
+    grant = store.consume(
+        pending.approval_id,
+        decision="approve_task",
+        owner="alice",
+        session_id="session-1",
+    )
+
+    assert grant is not None
+    assert grant.scope is ToolApprovalScope.TASK
+    assert grant.allow_remaining_actions is True
+
+
+def test_deny_is_unaffected_by_the_single_action_flag():
+    store = ToolApprovalStore()
+    pending = _pending(store)
+
+    assert store.consume(
+        pending.approval_id,
+        decision="deny",
+        owner="alice",
+        session_id=None,
+        allow_continuation=False,
+    ) is None
+    assert store.peek(pending.approval_id) is None
+
+
+def test_skill_test_approval_route_opts_out_of_continuation():
+    root = Path(__file__).resolve().parents[1]
+    skills = (root / "routes/skills_routes.py").read_text(encoding="utf-8")
+
+    approve_call = skills.index("exact_approval = tool_approval_store.consume(")
+    end = skills.index(")", skills.index("allow_continuation", approve_call))
+    assert "allow_continuation=False" in skills[approve_call:end]

tests/test_tool_approval_task_scope.py

@@ -0,0 +1,351 @@
+"""Task- and chat-scoped approval continuation coverage for issue #6112."""
+
+import asyncio
+import json
+from dataclasses import replace
+from pathlib import Path
+from types import SimpleNamespace
+
+from core.models import ChatMessage, Session
+from src.tool_approval_scopes import (
+    CHAT_SESSION_APPROVAL_CONTEXT_MARKER,
+    ToolApprovalScope,
+)
+from src.tool_approvals import ExactToolApproval, ToolApprovalStore
+from src.tool_capabilities import ToolRunSecurityContext, capabilities_for_action
+
+
+def _pending(
+    store: ToolApprovalStore,
+    *,
+    selected_tools=None,
+    continuation_query="inspect the project using memory and skills",
+):
+    content = "printf exact"
+    return store.create(
+        owner="Alice",
+        session_id="session-1",
+        origin_run_id="run-1",
+        tool_name="bash",
+        content=content,
+        workspace=None,
+        external_untrusted_context_seen=True,
+        selected_tools=selected_tools,
+        continuation_query=continuation_query,
+        capabilities=capabilities_for_action("bash", content),
+    )
+
+
+def test_card_offers_task_chat_session_and_deny_without_leaking_private_state():
+    pending = _pending(
+        ToolApprovalStore(),
+        selected_tools=["manage_skills", "bash", "manage_skills"],
+    )
+
+    payload = pending.public_payload()
+
+    assert payload["session_id"] == "session-1"
+    assert [option["value"] for option in payload["options"]] == [
+        "approve_task",
+        "approve",
+        "deny",
+    ]
+    assert [option["label"] for option in payload["options"]] == [
+        "Allow for this task",
+        "Allow for this chat session",
+        "Deny",
+    ]
+    serialized = json.dumps(payload, sort_keys=True)
+    assert "Allow once" not in serialized
+    assert "selected_tools" not in serialized
+    assert "continuation_query" not in serialized
+    assert "manage_skills" not in serialized
+    assert "inspect the project" not in serialized
+
+
+def test_allow_for_task_bypasses_only_the_resumed_run_gate():
+    store = ToolApprovalStore()
+    pending = _pending(store, selected_tools=["bash", "manage_skills"])
+    grant = store.consume(
+        pending.approval_id,
+        decision="approve_task",
+        owner="alice",
+        session_id="session-1",
+    )
+
+    assert grant is not None
+    assert grant.scope is ToolApprovalScope.TASK
+    assert grant.allow_remaining_actions is True
+    assert grant.grants_chat_session is False
+    assert grant.pending.continuation_query == (
+        "inspect the project using memory and skills"
+    )
+
+    resumed = ToolRunSecurityContext(
+        external_untrusted_context_seen=True,
+        approval_gate_bypassed=grant.allow_remaining_actions,
+    )
+    assert resumed.decision_for("bash").allowed is True
+
+    # A new ordinary user turn constructs a fresh context and asks again.
+    fresh = ToolRunSecurityContext(external_untrusted_context_seen=True)
+    assert fresh.decision_for("bash").allowed is False
+
+
+def test_allow_for_chat_session_applies_to_later_turns_in_only_that_chat():
+    store = ToolApprovalStore()
+    pending = _pending(store, selected_tools=["bash", "manage_skills"])
+    grant = store.consume(
+        pending.approval_id,
+        decision="approve",
+        owner="alice",
+        session_id="session-1",
+    )
+
+    assert grant is not None
+    assert grant.scope is ToolApprovalScope.CHAT_SESSION
+    assert grant.allow_remaining_actions is True
+    assert grant.grants_chat_session is True
+    assert grant.pending.selected_tools == ("bash", "manage_skills")
+    assert grant.pending.continuation_query.startswith("inspect the project")
+
+    resolved_card = pending.public_payload()
+    resolved_card["resolved"] = "approve"
+    history = [
+        ChatMessage(
+            "assistant",
+            "approval requested",
+            {"tool_events": [{"ask_user": resolved_card}]},
+        ),
+        ChatMessage("user", "continue the work"),
+    ]
+    session = Session(
+        id="session-1",
+        name="Chat",
+        endpoint_url="http://example.invalid",
+        model="test",
+        history=history,
+    )
+
+    messages = session.get_context_messages()
+    assert messages[-1]["metadata"][CHAT_SESSION_APPROVAL_CONTEXT_MARKER] is True
+    assert history[-1].metadata is None
+
+    future_turn = ToolRunSecurityContext(external_untrusted_context_seen=True)
+    future_turn.observe_messages(messages)
+    assert future_turn.approval_gate_bypassed is True
+    assert future_turn.decision_for("bash").allowed is True
+
+    # The persisted card is bound to its original chat id, so a fork/copy does
+    # not inherit the grant merely by copying transcript metadata.
+    other_session = Session(
+        id="session-2",
+        name="Fork",
+        endpoint_url="http://example.invalid",
+        model="test",
+        history=history,
+    )
+    other_messages = other_session.get_context_messages()
+    assert CHAT_SESSION_APPROVAL_CONTEXT_MARKER not in (
+        other_messages[-1].get("metadata") or {}
+    )
+    other_turn = ToolRunSecurityContext(external_untrusted_context_seen=True)
+    other_turn.observe_messages(other_messages)
+    assert other_turn.decision_for("bash").allowed is False
+
+
+def test_deny_executes_nothing_and_grants_no_task_or_chat_scope():
+    store = ToolApprovalStore()
+    pending = _pending(store)
+
+    assert store.consume(
+        pending.approval_id,
+        decision="deny",
+        owner="alice",
+        session_id="session-1",
+    ) is None
+    assert store.peek(pending.approval_id) is None
+
+    denied_card = pending.public_payload()
+    denied_card["resolved"] = "deny"
+    session = Session(
+        id="session-1",
+        name="Chat",
+        endpoint_url="http://example.invalid",
+        model="test",
+        history=[
+            ChatMessage(
+                "assistant",
+                "approval requested",
+                {"tool_events": [{"ask_user": denied_card}]},
+            ),
+            ChatMessage("user", "another request"),
+        ],
+    )
+    messages = session.get_context_messages()
+    assert CHAT_SESSION_APPROVAL_CONTEXT_MARKER not in (
+        messages[-1].get("metadata") or {}
+    )
+
+
+def test_private_continuation_state_is_canonical_bounded_and_digest_bound():
+    selected_tools = ["manage_skills", "bash", "manage_skills", "", 7]
+    selected_tools.extend(f"tool_{index:04d}" for index in range(600))
+    selected_tools.append("x" * 513)
+    pending = _pending(
+        ToolApprovalStore(),
+        selected_tools=selected_tools,
+        continuation_query="  " + ("original request " * 500),
+    )
+    assert pending.selected_tools[:2] == ("bash", "manage_skills")
+    assert len(pending.selected_tools) == 512
+    assert all(len(name) <= 512 for name in pending.selected_tools)
+    assert "x" * 513 not in pending.selected_tools
+    assert pending.continuation_query.startswith("original request")
+    assert len(pending.continuation_query) == 4000
+
+    tampered = replace(
+        pending,
+        selected_tools=("bash", "manage_skills", "send_email"),
+        continuation_query="different request",
+    )
+    grant = ExactToolApproval(tampered)
+    assert grant.matches(
+        owner="alice",
+        session_id="session-1",
+        tool_name="bash",
+        content="printf exact",
+        workspace=None,
+    ) is False
+
+
+def test_consumed_card_resolution_updates_memory_and_persisted_metadata(monkeypatch):
+    from routes import chat_routes
+
+    ask_user = {
+        "kind": "tool_approval",
+        "approval_id": "approval-1",
+        "session_id": "session-1",
+    }
+    metadata = {
+        "_db_id": "message-1",
+        "tool_events": [{"ask_user": ask_user}],
+    }
+    sess = SimpleNamespace(
+        id="session-1",
+        history=[SimpleNamespace(metadata=metadata)],
+    )
+    db_message = SimpleNamespace(meta_data=None)
+
+    class Column:
+        def __eq__(self, value):
+            return value
+
+    class FakeDBMessage:
+        id = Column()
+        session_id = Column()
+
+    class FakeQuery:
+        def filter(self, *conditions):
+            return self
+
+        def first(self):
+            return db_message
+
+    class FakeDB:
+        committed = False
+        rolled_back = False
+        closed = False
+
+        def query(self, model):
+            assert model is FakeDBMessage
+            return FakeQuery()
+
+        def commit(self):
+            self.committed = True
+
+        def rollback(self):
+            self.rolled_back = True
+
+        def close(self):
+            self.closed = True
+
+    db = FakeDB()
+    monkeypatch.setattr(chat_routes, "DBChatMessage", FakeDBMessage)
+    monkeypatch.setattr(chat_routes, "SessionLocal", lambda: db)
+
+    assert chat_routes._mark_tool_approval_resolved(
+        sess,
+        "approval-1",
+        "approve",
+    ) is True
+    assert ask_user["resolved"] == "approve"
+    persisted = json.loads(db_message.meta_data)
+    assert persisted["tool_events"][0]["ask_user"]["resolved"] == "approve"
+    assert "_db_id" not in persisted
+    assert db.committed is True
+    assert db.rolled_back is False
+    assert db.closed is True
+
+
+def test_deny_resolution_stream_is_control_only():
+    from routes.chat_routes import _tool_approval_resolution_stream
+
+    async def collect():
+        return [chunk async for chunk in _tool_approval_resolution_stream("deny")]
+
+    chunks = asyncio.run(collect())
+    assert chunks[-1] == "data: [DONE]\n\n"
+    event = json.loads(chunks[0][len("data: "):])
+    assert event == {"type": "tool_approval_resolved", "decision": "deny"}
+    assert "Denied the" not in "".join(chunks)
+
+
+def test_route_context_agent_frontend_and_cache_bust_wire_the_contract():
+    root = Path(__file__).resolve().parents[1]
+    route = (root / "routes/chat_routes.py").read_text(encoding="utf-8")
+    helpers = (root / "routes/chat_helpers.py").read_text(encoding="utf-8")
+    agent = (root / "src/agent_loop.py").read_text(encoding="utf-8")
+    frontend = (root / "static/js/chat.js").read_text(encoding="utf-8")
+    renderer = (root / "static/js/chatRenderer.js").read_text(encoding="utf-8")
+    app = (root / "static/app.js").read_text(encoding="utf-8")
+    index = (root / "static/index.html").read_text(encoding="utf-8")
+    approvals = (root / "src/tool_approvals.py").read_text(encoding="utf-8")
+    capabilities = (root / "src/tool_capabilities.py").read_text(encoding="utf-8")
+    models = (root / "core/models.py").read_text(encoding="utf-8")
+
+    assert 'decision not in {"approve", "approve_task", "deny"}' in route
+    assert "set(pending_tool_approval.selected_tools)" in route
+    assert "pending_tool_approval.continuation_query" in route
+    assert "persist_user_message=not tool_approval_continuation" in route
+    assert "_mark_tool_approval_resolved(" in route
+    assert "_tool_approval_resolution_stream(decision)" in route
+    assert "Approved the exact" not in route
+    assert "Denied the" not in route
+    assert "continuation_context_message: str | None = None" in helpers
+    assert "persist_user_message: bool = True" in helpers
+    assert "_without_latest_matching_user_message(" not in helpers
+    assert "selected_tools=approval_selected_tools" in agent
+    assert "continuation_query=_retrieval_query or _last_user" in agent
+    assert "approval_gate_bypassed=bool(" in agent
+    assert "['approve', 'approve_task', 'deny']" in frontend
+    assert "input.value = label" not in frontend
+    assert "const msg = approvalForSend ? '' : el('message').value;" in frontend
+    assert "const skipBubble = _hideUserBubble || !!approvalForSend;" in frontend
+    assert "fd.append('message', approvalForSend ? '' : _finalMsgWithInject);" in frontend
+    assert "json.type === 'tool_approval_resolved'" in frontend
+    assert "if (aq.resolved) return null;" in renderer
+    assert "ev.ask_user && !ev.ask_user.resolved" in renderer
+    assert '"label": "Allow once"' not in approvals
+    assert '"label": "Allow for this task"' in approvals
+    assert '"label": "Allow for this chat session"' in approvals
+    assert "scope_for_decision(normalized_decision)" in approvals
+    assert "CHAT_SESSION_APPROVAL_CONTEXT_MARKER" in capabilities
+    assert "CHAT_SESSION_APPROVAL_CONTEXT_MARKER" in models
+
+    version = "20260819approvalcontrol1"
+    assert f"chat.js?v={version}" in app
+    assert f"chat.js?v={version}" in index
+    assert f"chatRenderer.js?v={version}" in frontend
+    assert f"chatRenderer.js?v={version}" in app
+    assert f"chatRenderer.js?v={version}" in index