docs(style): improve roadmap and contributing Markdown layout

The handbook says substantial features need Product Owner approval
before implementation, but not how a proposal gets there, what counts
as substantial, or what happens when no answer comes. Add the full
workflow: triage within 7 days, a decision brief, four product labels,
a lazy-consensus timeout that only applies to big-but-reversible work,
and definitions for every judgement call (small, big, reversible, new
direction, fits ROADMAP, objection, recorded decision).

The Feature Request form gets three optional fields for the brief, and
the PR template, CONTRIBUTING and triage point at the workflow.

.env.example

@@ -105,6 +105,17 @@ SEARXNG_INSTANCE=http://localhost:8080
 # trusted single-user setups that accept injected content reaching side-effect tools.
 # ODYSSEUS_TOOL_APPROVAL_GATE=1
 
+# Comma-separated list of exact qualified MCP read and write tool names
+# on host-local servers. Only explicitly configured stdio MCP servers qualify.
+# Declarations do not override built-in tool effects. Declared reads remain
+# untrusted (EXTERNAL_UNTRUSTED) to prevent prompt injections from reaching shell
+# or network tools, but allow continuation to declared non-destructive write tools
+# on the exact same stdio server without repeated approvals. Stdio transport does
+# not guarantee network isolation. Authorization must still originate from the
+# trusted request. Automatic same-store writes carry residual prompt-injection risk.
+# ODYSSEUS_LOCAL_MCP_READ_TOOLS=mcp__notes__search_notes,mcp__notes__read_note
+# ODYSSEUS_LOCAL_MCP_WRITE_TOOLS=mcp__notes__create_note,mcp__notes__update_note
+
 # 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.

.github/ISSUE_TEMPLATE/feature_request.yml

@@ -14,6 +14,8 @@ body:
         If your idea needs community input before it becomes a concrete proposal,
         start a [discussion](https://github.com/odysseus-dev/odysseus/discussions/categories/ideas) instead.
 
+        How a proposal is triaged and decided: [product and feature proposals](https://github.com/odysseus-dev/odysseus/blob/dev/docs/maintainers/product-and-features.md#feature-proposal-workflow).
+
   - type: checkboxes
     id: prerequisites
     attributes:
@@ -79,6 +81,32 @@ body:
       label: Prior Art / Related Issues
       description: Link any related issues, discussions, or external references that informed this proposal.
 
+  - type: dropdown
+    id: scope
+    attributes:
+      label: Size and reversibility
+      description: Your best guess; the triager decides. Definitions are in the maintainer handbook (product-and-features.md).
+      multiple: true
+      options:
+        - "Small, reversible change within ROADMAP"
+        - "Big (more than one PR, or changes existing default behaviour)"
+        - "New direction (new area, multi-user, data leaving the machine, ...)"
+        - "Hard to undo (migration, deleted data, changed contract or permissions)"
+        - "Not sure"
+
+  - type: textarea
+    id: decision_questions
+    attributes:
+      label: Questions for the Product Owner
+      description: If this needs a product decision, the exact yes/no questions — one per independent part.
+      placeholder: "1. Should notes be shareable between accounts on one instance? (yes/no)"
+
+  - type: textarea
+    id: evidence
+    attributes:
+      label: Screenshots, recording or demo
+      description: For anything that renders, attach screenshots or a short recording. Link a demo branch if you have one.
+
   - type: dropdown
     id: willing_to_implement
     attributes:

.github/pull_request_template.md

@@ -13,6 +13,9 @@
 
      Private maintainer-preview PRs may instead use:
      N/A — maintainer integration work
+
+     New features: the linked issue carries `product: approved` or
+     `product: maintainer decision` (docs/maintainers/product-and-features.md).
 -->
 
 Fixes #

.github/workflows/ci.yml

@@ -179,6 +179,7 @@ jobs:
           uv pip install --system --require-hashes --only-binary=:all: --no-deps -r requirements-bootstrap.txt
           python scripts/check_ca_trust.py --python
           uv pip install --system -r requirements.txt
+          uv pip install --system "playwright==1.62.0"
         if: steps.docs-check.outputs.docs_only != 'true'
       - uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020  # v7.0.0
         if: steps.docs-check.outputs.docs_only != 'true'
@@ -256,3 +257,27 @@ jobs:
         env:
           PYTEST_SHARD: ${{ matrix.shard }}
         run: python -m pytest -q -rs --durations=25 --shard "$PYTEST_SHARD"
+
+      - name: Playwright mobile layout checks (shard 1/4 only)
+        if: steps.docs-check.outputs.docs_only != 'true' && matrix.shard == '1/4'
+        run: |
+          ODYSSEUS_TEST_PYTHON=python ODYSSEUS_LAYOUT_BROWSER=chromium npx --no-install playwright test \
+            --config tests/e2e/ios-layout.config.cjs \
+            --reporter=line
+
+  benchmark-gfm-security:
+    name: Benchmark GFM rendering security
+    runs-on: ubuntu-24.04
+    steps:
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
+        with:
+          persist-credentials: false
+      - uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97  # v7.0.0
+        with:
+          python-version: "3.11"
+      # Test-only renderer, pinned independently of application requirements.
+      - name: Install GitHub Flavored Markdown renderer
+        run: |
+          python -m pip install --disable-pip-version-check --only-binary=:all: "cmarkgfm==2025.10.22"
+      - name: Verify rendered benchmark report security
+        run: python tests/test_audit_c5_benchmark_evidence.py --gfm

.gitignore

@@ -53,6 +53,8 @@ services/data/
 
 # Test capture artifacts (browser session dumps may contain personal content)
 .playwright-mcp/
+test-results/
+playwright-report/
 
 # OS
 .DS_Store

.gitleaks.toml

@@ -35,3 +35,15 @@ regexTarget = "secret"
 regexes = [
   '''^7abc5b334a6c3c52dd3eba15d8bc516da728a67f14fb7f56d0bb5992d89a16ca$''',
 ]
+
+[[allowlists]]
+description = "Synthetic dummy API key in trace artifact sanitizer test fixture"
+targetRules = ["generic-api-key"]
+condition = "AND"
+paths = [
+  '''(?:^|/)tests/test_trace_artifact_sanitizer\.py$''',
+]
+regexTarget = "secret"
+regexes = [
+  '''^sk-123456789$''',
+]

CONTRIBUTING.md

@@ -1,28 +1,50 @@
 # Contributing to Odysseus
 
-Thanks for helping. The project is moving quickly, so the best contributions are focused, easy to review, and easy to test.
+Thanks for helping improve Odysseus. The best contributions are **focused, reviewable, and backed by meaningful validation**.
+
+> [!IMPORTANT]
+> **Public pull requests target `dev`, not `main`.** Search existing work before implementing, and follow the current [PR template](.github/pull_request_template.md). Runtime changes require relevant execution evidence, not just a green CI badge.
+
+## At a glance
+
+| Topic | What contributors should do |
+| --- | --- |
+| **Branch** | Target `dev`; `main` is the curated release branch |
+| **Scope** | One coherent change per PR; coordinate larger modularization work through [#4071](https://github.com/odysseus-dev/odysseus/issues/4071) |
+| **Validation** | Document commands, results, runtime checks, and anything not tested |
+| **Visual changes** | Run the UI, match the existing visual language, and attach screenshots or clips |
+| **Review** | Follow the [risk-based maintainer policy](docs/maintainers/changes-and-review.md) and applicable repository protections |
+| **Security** | Report vulnerabilities privately using [SECURITY.md](SECURITY.md) |
+
+**Jump to:** [Branches](#branch-model) · [Getting started](#before-you-start) · [Setup](#setup) · [Checks](#running-checks) · [Pull requests](#pull-requests) · [Modularization](#structural-changes-and-modularization) · [Visual changes](#style-and-visual-changes) · [Code conventions](#code-conventions) · [Issue reports](#issue-reports)
+
+---
 
 ## Branch model
 
-Odysseus has two branches:
+| Branch | Purpose |
+| --- | --- |
+| **`dev`** | Default integration branch and target for every public PR. Work can be in flux. |
+| **`main`** | Curated release branch. A [Release Maintainer](docs/maintainers/testing-and-releases.md#release-ownership) promotes a verified `dev` commit. |
 
-- **`dev`** — where all PRs land. Things can be in flux here; the merge button gets used freely.
+If your PR accidentally targets `main`, use **Edit** on GitHub to change its base to `dev`. Changing the base does not itself require a rebase.
-- **`main`** — what users run. Curated and tested by the release's [Release Maintainer](docs/maintainers/testing-and-releases.md#release-ownership). Fast-forwarded to a stable `dev` commit at each release.
 
-**Open your PR against `dev`, not `main`.** The GitHub "base" dropdown defaults to `dev`. If you opened a PR against `main` by accident, click "Edit" on the PR and change the base — no rebase needed.
+End users cloning the repository land on `dev` by default. To use the curated release branch, run `git checkout main` after cloning.
-
-End-users cloning the repo will land on `dev` by default. To run the curated/stable version: `git checkout main` after clone.
 
 ## Before You Start
 
-- Search existing issues and pull requests before opening a new one.
+1. Search [open issues](https://github.com/odysseus-dev/odysseus/issues) and [PRs](https://github.com/odysseus-dev/odysseus/pulls) to avoid duplicate or conflicting work.
-- Prefer one bug fix or feature per pull request.
+2. Prefer one bug fix, feature, or coherent refactor per PR. Avoid broad rewrites, formatting-only churn, and unrelated file moves.
-- Avoid broad rewrites, formatting-only changes, or moving many files unless the issue is specifically about structure.
+3. For a feature, open a **Feature Request** first. Big, new-direction, or difficult-to-reverse changes require the documented [product decision process](docs/maintainers/product-and-features.md#feature-proposal-workflow).
-- If you want to work on a large feature, open an issue first and describe the approach.
+4. For structural work, coordinate the intended module boundary with maintainers and the [modularization tracker #4071](https://github.com/odysseus-dev/odysseus/issues/4071).
+
+> [!NOTE]
+> **Coding agents:** If you use an LLM coding agent, open an issue describing the problem before submitting its PR. Bulk agent-generated submissions that ignore the project's contribution format or visual style may be closed without review, even when the underlying fix is correct.
 
 ## Setup
 
-Docker is the recommended path for normal testing:
+<details open>
+<summary><strong>Docker (recommended for normal testing)</strong></summary>
 
 ```bash
 git clone https://github.com/odysseus-dev/odysseus.git
@@ -31,7 +53,10 @@ cp .env.example .env
 docker compose up -d --build
 ```
 
-Manual development uses Python 3.11+:
+</details>
+
+<details>
+<summary><strong>Manual Python development (Python 3.11+)</strong></summary>
 
 ```bash
 python3 -m venv venv
@@ -40,11 +65,14 @@ pip install -r requirements.txt
 python -m uvicorn app:app --host 127.0.0.1 --port 7000
 ```
 
-Windows is not actively tested. Docker on Linux or a Linux/macOS manual install is the safer path for now.
+</details>
+
+> [!NOTE]
+> Native Windows is not actively tested. Docker on Linux or a Linux/macOS manual installation is the safer path for now. Document the environment actually used for testing.
 
 ## Running Checks
 
-Run the smallest relevant checks for your change:
+Run relevant checks for the affected area. Examples from the existing contributor workflow:
 
 ```bash
 python -m pytest
@@ -60,78 +88,94 @@ docker compose up -d --build
 docker compose logs --tail=120 odysseus
 ```
 
-Mention what you ran in the pull request description. If you could not run a check, say so.
+> [!IMPORTANT]
+> **Report what actually ran.** Include commands, results, and relevant manual/runtime verification in the PR description. If a check was not run, mark it as **not run** and explain the limitation. Passing syntax checks and unit tests do not automatically prove changed behavior works in the running application.
+
+See the [testing and releases handbook](docs/maintainers/testing-and-releases.md) for risk-based validation and isolation requirements.
 
 ## Maintainer governance and review guidance
 
-For repository roles, product decisions, risk-based approvals, issue triage, independent security reviews, CodeQL findings, and release handoffs, see the [Odysseus maintainer handbook](docs/maintainers/README.md). This complements the contributor requirements below and does not change the branch model or current PR template.
+The [Odysseus maintainer handbook](docs/maintainers/README.md) is authoritative for repository roles, product decisions, risk-based approvals, issue triage, security review, CodeQL findings, and release handoffs.
+
+> [!NOTE]
+> The handbook complements this contributor guide. It does not override the branch model, live GitHub protection rules, or the PR template.
 
 ## Pull Requests
 
-Good pull requests usually include:
+Keep PRs **cohesive**. Small, focused changes are the default. A larger, coordinated module decomposition may be appropriate under the criteria below, but unrelated cleanup, formatting, refactors, and behavior changes must not be mixed.
 
-- A short explanation of the bug or feature.
+| Include in the PR | Why it matters |
-- The files or areas changed.
+| --- | --- |
-- Manual test steps or automated test results from running the actual app, not just the test suite.
+| **Summary and scope** | Explain the problem, intended behavior, and areas changed |
-- Screenshots or short recordings for UI changes.
+| **Linked issue** | Connect the PR to the existing report or approved work, for example `Fixes #123` or `Part of #4071` |
-- Links to related issues, for example `Fixes #123`.
+| **Validation evidence** | Show manual test steps or automated results that exercise the actual behavior |
+| **UI evidence, when applicable** | Attach screenshots or short recordings from the running application |
+| **Limitations** | Clearly identify tests or environments that were not exercised |
 
-Please keep PRs small. Large PRs that mix unrelated cleanup, formatting, refactors, and behavior changes are much harder to review.
+For exact required sections and attestations, follow the current [PR template](.github/pull_request_template.md).
 
-> **Auto-generated PRs.** If you are running an LLM agent (Devin, Cursor, OpenHands, Claude Code, etc.) against this repo: please open an issue describing the problem first instead of opening a PR directly. Bulk agent-generated PRs that don't match the project's visual style or contribution format will be closed without review, even when the underlying fix is correct.
+### Structural changes and modularization
+
+**One coherent responsibility is the unit of review.** A larger PR can be appropriate when extracting that responsibility from an oversized module is safer or simpler than splitting the work into many conflicting or interdependent PRs.
+
+Before implementation and review:
+
+- **Coordinate:** Read existing issues/PRs and agree on the boundary with maintainers. Use [#4071](https://github.com/odysseus-dev/odysseus/issues/4071) for codebase modularization.
+- **Constrain scope:** Keep one subsystem or responsibility per PR. Avoid unrelated features, formatting, and opportunistic cleanup.
+- **Preserve contracts:** Maintain observable behavior, imports, public interfaces, persistence contracts, and runtime invariants unless an approved issue explicitly scopes a change.
+- **Validate behavior:** Include targeted regression tests and runtime checks. Frontend work needs browser validation and screenshots or recordings when rendering may change.
+- **Explain tradeoffs:** Describe extraction boundaries, compatibility, validation results, and known limitations in the PR.
+
+> [!TIP]
+> **Large does not mean broad.** There is no arbitrary line-count limit for a well-scoped extraction. Correctness, reviewability, and integration risk determine whether the scope is acceptable.
+
+> [!CAUTION]
+> **Temporary Harness V2 coordination boundary:** Unrelated changes to `src/agent_loop.py` and closely related execution, authorization, orchestration, and context contracts are temporarily frozen while Harness V2 is developed and validated. Coordinate such work with maintainers. Other modules remain open, and urgent bug or security reports are welcome.
+
+These contributions remain subject to the [risk-based review policy](docs/maintainers/changes-and-review.md), current repository protections, and PR template.
 
 ## Style and visual changes
 
-Odysseus has an intentional visual style. PRs that ignore it will be closed without merge, no matter how correct the underlying code is.
+Odysseus has an intentional visual language. Changes that ignore it can be closed without merge even when their underlying code is correct.
 
-Before submitting any change that affects what the app looks like — buttons, icons, fonts, colors, spacing, layout, CSS, HTML, SVG, or any `static/js/` module that draws to the DOM — please:
+For any change affecting buttons, icons, typography, colors, spacing, layout, CSS, HTML, SVG, or JavaScript that renders into the DOM:
 
-1. **Run the app locally** and view the change in a browser. Type-checks and unit tests are not enough.
+1. **Run the application locally** and inspect the change in a browser. Type checks and unit tests alone are insufficient.
-2. **Attach a screenshot or short clip** of the change in the running app. Add a mobile screenshot too if the change affects mobile.
+2. **Attach a screenshot or short clip** from the running app, plus a mobile screenshot when mobile behavior changes.
-3. **Match the existing visual language.** Specifically:
+3. **Reuse the existing design system:**
-   - Reuse existing CSS variables (`--red`, `--fg`, `--bg`, `--card`, `--border`, …). Do not introduce new color values, font sizes, or spacing units.
+   - Prefer existing CSS tokens (`--red`, `--fg`, `--bg`, `--card`, `--border`, etc.). Do not introduce parallel colors, type scales, or spacing systems.
-   - Reuse existing button, input, card, and border classes. Don't invent parallel styling for similar widgets.
+   - Reuse existing button, input, card, and border classes rather than creating duplicate patterns.
-   - **No Unicode emoji in UI or code.** Use inline SVG (matching the monochrome icon style already in `static/index.html`) or plain text.
+   - Do **not** add Unicode emoji in UI or code. Use inline SVG consistent with `static/index.html` or plain text.
-   - Monospaced font (`Fira Code`) for primary UI text. Don't override.
+   - Preserve the existing primary monospaced font (`Fira Code`).
-   - Dark theme is the default; any light-mode work goes through the existing theme system, not hard-coded.
+   - Keep dark mode as the default; route any light-mode work through the existing theme system.
-4. **Don't add parallel components.** If a similar widget already exists in the app, extend it instead of writing a new one.
+4. **Extend existing components** instead of introducing a parallel component for the same purpose.
 
-If you are unsure whether a change is "visual," it is. Default to attaching a screenshot.
+> [!IMPORTANT]
+> If you are unsure whether a change is visual, assume that it is and provide a screenshot. Visual evidence does not replace behavioral validation.
 
 ## Code conventions
 
-Don't hardcode values that the project already exposes through a constant or a helper. Hardcoded literals drift out of sync, break on non-default deployments, and reintroduce bugs we've already fixed.
+**Use existing constants and helpers.** Hardcoded duplicates drift, break non-default deployments, and can reintroduce previously fixed issues.
 
-- **Filesystem paths:** never build writable paths from `Path(__file__)...` into the source tree, hardcode `/app/...`, or use a relative `"data/..."` string. Every persisted file and directory has a named constant in `src/constants.py` (for example `AUTH_FILE`, `USER_PREFS_FILE`, `SETTINGS_FILE`, `TTS_CACHE_DIR`, `CHROMA_DIR`). Import and use that named constant; do not re-derive the path locally with `os.path.join(DATA_DIR, "x.json")` or `DATA_DIR / "x.json"`. `DATA_DIR` is the single place that reads `ODYSSEUS_DATA_DIR`, so use it directly only for dynamic paths that have no fixed name (for example per-owner files). If a data file or directory has no constant yet, add one to `src/constants.py`. The source tree is read-only in Docker and `/app/...` does not exist on native runs; guard directory creation so an unwritable path degrades gracefully instead of crashing at import.
+- **Filesystem paths:** Do not derive writable paths from `Path(__file__)...` into the source tree, hardcode `/app/...`, or use a relative `"data/..."` string. Persisted files and directories have named constants in `src/constants.py`, such as `AUTH_FILE`, `USER_PREFS_FILE`, `SETTINGS_FILE`, `TTS_CACHE_DIR`, and `CHROMA_DIR`. Import those names rather than deriving paths with `os.path.join(DATA_DIR, "x.json")` or `DATA_DIR / "x.json"`. `DATA_DIR` reads `ODYSSEUS_DATA_DIR` and is intended directly only for dynamic paths without fixed names, such as per-owner files. Add a constant when a fixed path does not yet have one. The source tree is read-only in Docker, and `/app/...` does not exist on native runs. Guard directory creation so unwritable paths degrade gracefully rather than crashing at import.
-- **Internal API / loopback URLs:** don't hardcode `http://localhost:7000`. Use `internal_api_base()` from `src.constants` (it honors `ODYSSEUS_INTERNAL_BASE` / `APP_PORT`).
+- **Internal API and loopback URLs:** Do not hardcode `http://localhost:7000`. Use `internal_api_base()` from `src.constants`, which honors `ODYSSEUS_INTERNAL_BASE` and `APP_PORT`.
-- **Ports, limits, model lists, and similar:** reuse the existing constant if one exists; if it doesn't and the value is used in more than one place, add a constant rather than copying the literal.
+- **Ports, limits, and model lists:** Reuse existing constants. When a value is shared by multiple locations and has no constant, add one in `src/constants.py` instead of duplicating a literal. `core/constants.py` is a compatibility re-export, not another source of truth.
 
-If you need a value that has no constant or helper yet, add it to `src/constants.py` (the single source of truth for paths and config; `core/constants.py` only re-exports it for backward compatibility) and import it, rather than repeating a literal across files.
+**Commits:** Use [Conventional Commits](https://www.conventionalcommits.org) with `type(scope): summary`, for example `fix(search): ...`, `feat(notes): ...`, or `docs(contributing): ...`. Common types include `fix`, `feat`, `refactor`, `docs`, `test`, `chore`, and `ci`. Keep the subject short and imperative, with the motivation in the body when useful.
-
-**Commits:** use [Conventional Commits](https://www.conventionalcommits.org), `type(scope): summary` (e.g. `fix(search): ...`, `feat(notes): ...`, `docs(contributing): ...`). Common types: `fix`, `feat`, `refactor`, `docs`, `test`, `chore`, `ci`. Keep the subject short and imperative; put the "why" in the body when it isn't obvious.
 
 ## Issue Reports
 
-For bugs, include:
+| Bug reports should include | Model-serving reports should additionally include |
+| --- | --- |
+| Installation method (Docker, Python, WSL, etc.) | Backend (Ollama, vLLM, SGLang, llama.cpp, LM Studio, etc.) |
+| OS, browser, and device when relevant | Model name |
+| Exact reproduction steps | GPU/CPU and operating system |
+| Expected versus actual behavior | Cookbook task logs or backend/server logs |
+| Relevant logs, screenshots, or terminal output | Relevant backend configuration |
 
-- Install method: Docker, manual Python, WSL, etc.
+Issues containing only "help", "does not work", or an unexplained screenshot may be closed as not actionable. Start with a reproducible problem or use [Discussions](https://github.com/odysseus-dev/odysseus/discussions) for setup questions.
-- OS, browser, and device if relevant.
-- Exact steps to reproduce.
-- Expected behavior and actual behavior.
-- Logs, screenshots, or terminal output.
-
-For model-serving issues, include:
-
-- Backend: Ollama, vLLM, SGLang, llama.cpp, LM Studio, etc.
-- Model name.
-- GPU/CPU and operating system.
-- Cookbook task logs or server logs.
-
-Issues with only "help", "does not work", or a screenshot without context may be closed as not actionable.
 
 ## Security
 
-Do not post secrets, API keys, private logs, personal documents, or public IPs in issues or pull requests.
+> [!CAUTION]
-
+> **Do not publish secrets.** Keep API keys, sensitive logs, personal documents, and public IP addresses out of public issues and PRs. Report vulnerabilities privately through [SECURITY.md](SECURITY.md).
-For security reports, follow [SECURITY.md](SECURITY.md).
-

Dockerfile

@@ -1,9 +1,22 @@
+# Both stages need the fixed native decoder, including curl in the builder and
+# distro Node in the runtime. A cached Python base may predate the Debian fix.
+FROM python:3.14.9-slim AS brotli-base
+COPY scripts/check_brotli.py /usr/local/bin/check-brotli.py
+COPY scripts/check_ca_trust.py /usr/local/bin/check-ca-trust.py
+# Check inherited trust before the first HTTPS package request, retaining Debian's
+# signed archive verification. Refresh only the explicit native dependency.
+RUN python /usr/local/bin/check-ca-trust.py --system \
+    && sed -i 's|http://deb.debian.org/|https://deb.debian.org/|g' /etc/apt/sources.list.d/debian.sources \
+    && apt-get update && apt-get install -y --no-install-recommends libbrotli1 \
+    && python /usr/local/bin/check-brotli.py \
+    && rm -rf /var/lib/apt/lists/*
+
 # ---- builder: patch + build wheels for Real-ESRGAN's broken-on-3.14 deps ----
 # basicsr/gfpgan/facexlib read their version via exec()+locals()['__version__'],
 # which raises KeyError on Python 3.13+ (PEP 667). Build patched wheels here so
 # the final image / Cookbook never has to compile the broken sdists. See
 # docker/build-realesrgan-wheels.sh for the full rationale.
-FROM python:3.14.9-slim AS realesrgan-wheels
+FROM brotli-base AS realesrgan-wheels
 RUN apt-get update && apt-get install -y --no-install-recommends curl ca-certificates \
     && rm -rf /var/lib/apt/lists/*
 COPY scripts/check_ca_trust.py /usr/local/bin/check-ca-trust.py
@@ -14,7 +27,7 @@ RUN python -m pip install --no-cache-dir --require-hashes --only-binary=:all: --
 COPY docker/build-realesrgan-wheels.sh /usr/local/bin/build-realesrgan-wheels.sh
 RUN bash /usr/local/bin/build-realesrgan-wheels.sh /wheels
 
-FROM python:3.14.9-slim
+FROM brotli-base
 
 # System deps. tmux is required by Cookbook for background downloads/serves.
 # openssh-client is required for Cookbook remote server tests, setup, probes,
@@ -52,6 +65,7 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
     libglib2.0-0t64 \
     libxcb1 \
     libmagic1 \
+    && python /usr/local/bin/check-brotli.py --node /usr/bin/node \
     && rm -rf /var/lib/apt/lists/*
 
 # Refresh both Python trust bundles before installing application dependencies.
@@ -132,7 +146,8 @@ RUN pip install --no-cache-dir --no-deps /tmp/odysseus-wheels/*.whl \
 # Check gRPC's effective defaults after all Python dependencies are installed.
 # The application minimum fixes its bundled roots; the checked override also
 # keeps the effective Docker selection consistent with system trust.
-RUN python /usr/local/bin/check-ca-trust.py --system --python --grpc
+RUN python /usr/local/bin/check-ca-trust.py --system --python --grpc \
+    && python /usr/local/bin/check-brotli.py --node /usr/bin/node
 
 # Copy app code
 COPY . .

ROADMAP.md

@@ -1,87 +1,117 @@
 # Roadmap / Help Wanted
 
-Odysseus is on a voyage, but not home yet. It works great for me (lol), but this ship is moving fast and feedback/help would be appreciated! (I don't know what I'm doing, help).
+Odysseus is an evolving, self-hosted AI workspace. This roadmap highlights where improvements and contributions are most useful, from reliability and maintainability to product capabilities.
 
-If you see weird CSS, strange layout behavior, or a suspiciously murky corner of
+> [!NOTE]
-the codebase, you are probably right to stay away.
+> **Direction, not a release schedule.** This page is not a complete issue backlog or a promise to deliver every proposal. Implementation scope and priority are coordinated through existing issues, pull requests, and the project's [feature proposal process](docs/maintainers/product-and-features.md).
 
-## High Priority
+## At a glance
 
-- SQUASH BUGS
+| Area | Direction | Coordination |
-- Fresh install smoke tests on Linux, macOS, and Windows. Docker, native Python,
+| --- | --- | --- |
-  and WSL all need coverage.
+| **Stabilization** | Reliability, CI, fresh installs, security, and integrations | [Stabilization & Modularization milestone](https://github.com/odysseus-dev/odysseus/milestone/1) |
+| **Modularization** | Clearer module ownership, safer changes, and contributor onboarding | [Canonical tracker #4071](https://github.com/odysseus-dev/odysseus/issues/4071) |
+| **Agent runtime** | Active Harness V2 development, with a narrow temporary change freeze | [Coordination in #4071](https://github.com/odysseus-dev/odysseus/issues/4071) |
+| **Product improvements** | Local models, research, frontend experience, and operations | [Issues](https://github.com/odysseus-dev/odysseus/issues) and [feature proposals](docs/maintainers/product-and-features.md) |
 
-- Integration audit: do integrations even work? Confirm what works, what needs setup docs, and what should be removed or hidden. 
+**Jump to:** [Stabilization](#stabilization-and-reliability) · [Modularization](#modularization-and-maintainability) · [AI and local models](#ai-local-models-and-research) · [Frontend](#frontend-and-user-experience) · [Backend](#backend-and-operations) · [Contributing](#contributing)
-- Cookbook reliability on other computers. This is probably the area most likely to need work across different machines, GPUs, drivers, shells, and Python environments.
-- Cookbook SGLang support across platforms. Make sure SGLang setup/serve works
-  predictably on Linux, Windows/WSL, macOS where possible, Docker, and common
-  NVIDIA/AMD hardware paths.
-- Deep Research model presets by hardware. Recommend approved model/parameter
-  profiles for small, medium, and large local setups so people with different
-  hardware can use Deep Research without guessing. Surface this either in Deep
-  Research settings or as a Cookbook scan/dropdown suggestion.
-- Cookbook model scan/download ranking. Prioritize newer architectures and
-  better hardware-fit models instead of scoring everything almost the same.
-  Ranking should account for architecture age, quant format, VRAM/RAM fit,
-  backend support, vision/mmproj requirements, and likely serve reliability.
-- Cookbook error feedback and logging. Failed downloads, dependency installs,
-  preflights, and serve jobs should show the actual command/output/error in the
-  UI, with copyable logs and clear next steps instead of just "crashed".
-- Agent prompt/context bloat. Agent mode is too heavy for smaller local models:
-  tool schemas, skills, memory, documents, and instructions can eat the context
-  before the user request really starts. We need slimmer prompts, better tool
-  selection, smaller default tool sets, and clearer guidance for models with
-  4k/8k/16k context windows.
-- Local model speculative decoding support. For Odysseus-tuned local models,
-  plan to ship or recommend a small same-tokenizer draft model when the serving
-  backend supports it. Early vLLM testing showed a generic `Qwen3-0.6B` draft
-  beside `Qwen3-8B` can materially reduce wall time, while an unsupported
-  DSpark conversion performed poorly. Treat this as a supported draft-model lane
-  first; keep MTP-specific packaging as future work only when the architecture
-  and runtime support are real. Judge this by time-to-success, tool correctness,
-  grammar, and unchanged target output, not tokens/sec alone.
-- Skill/tool prompt-injection audit. User-editable skills, notes, documents,
-  fetched pages, and memories should be treated as untrusted data. Keep testing
-  whether models follow malicious instructions from those surfaces.
-- Better degraded-state reporting for ChromaDB, SearXNG, email, ntfy, and provider probes.
-- Email performance audit. Fetching, searching, opening, deleting, and sending
-  email can feel slow, especially over IMAP/SMTP providers with high latency.
-  Need someone who knows mail performance to profile the current flow, identify
-  whether the bottleneck is IMAP folder select/fetch, cache invalidation,
-  attachment/body loading, SMTP handshakes, or frontend refresh behavior, then
-  propose safer caching/prefetch/batching without breaking multi-account state.
-- Provider setup/probing audit for Anthropic, Gemini, Groq, xAI, OpenRouter, OpenAI, and DeepSeek.
 
-## Refactor Targets
+---
-- CSS cleanup. `static/style.css` basically Calypso's island atm.
-- Tour core helper. The onboarding tours have too much copy-pasted scaffolding; promote a shared `tour-core.js` helper before adding more tours.
-- Modal/window positioning cleanup. Some window controls have improved, but the
-  underlying popup/dropdown/fixed-position behavior is still too fragile.
-- Mobile media override discoverability. A lot of "CSS did not move" bugs are mobile `@media` overrides of the same selector; comments or linting around desktop/mobile paired rules would help.
-- Dead code pass for old routes, stale feature flags, and unused UI states.
 
-## Frontend
+## Stabilization and reliability
 
-- Expand the Editor for quicker, more robust everyday use. Better file/document
+**Priority:** Make Odysseus dependable across installations, platforms, providers, and supported hardware.
-  handling, smoother window behavior, clearer save/export flows, stronger image
-  editing affordances, and fewer brittle edge cases.
-- Better AI integration for Notes and Todos. Notes should be easier for the
-  agent to read, update, summarize, and turn into actions. Todos should be
-  assignable to an agent from the UI, possibly through a button, task action,
-  or dedicated skill/tool flow.
-- Mobile gallery/editor polish. Easier to launch/download inpaint model or any missing pieces.
-- Accessibility pass: keyboard navigation, focus states, contrast, reduced motion.
-- Improve empty states and error messages on fresh installs.
-- Tighten first-run setup, hints, and tours so they do not repeat or fight each other.
-- Vendor CDN assets eventually for a more fully self-hosted/offline mode.
 
-## Backend
+| Workstream | What needs improvement | Tracking |
+| --- | --- | --- |
+| **CI and validation** | Exact-revision validation, meaningful regression coverage, and release confidence | [#5932](https://github.com/odysseus-dev/odysseus/issues/5932), [#7109](https://github.com/odysseus-dev/odysseus/issues/7109) |
+| **Test quality** | Deterministic tests, isolation, focused execution, and maintainable test structure | [#2523](https://github.com/odysseus-dev/odysseus/issues/2523), [testing standard](tests/TESTING_STANDARD.md) |
+| **Fresh installations** | Smoke tests for Linux, macOS, Windows, WSL, and Docker; baseline startup and graceful degradation | [#3968](https://github.com/odysseus-dev/odysseus/issues/3968) |
+| **Email integration** | Remaining Google OAuth regressions, account isolation, and cross-provider reliability | [#5637](https://github.com/odysseus-dev/odysseus/issues/5637) |
+| **Storage integrity** | Upload payload growth, indexing, and attachment/file lifecycle | [#4889](https://github.com/odysseus-dev/odysseus/issues/4889) |
 
-- More tests around endpoint probing and provider setup.
+Additional reliability priorities:
-- Better task scheduler defaults and visibility.
-- Backup/restore guide and helper flow for `data/`.
-- Security hardening around admin-only tools and clear docs for their risk.
 
-## Not The Focus Right Now
+- **Cookbook:** Improve model installation, dependency management, hardware detection, and model serving across operating systems, GPUs, drivers, shells, Python environments, and supported backends.
+- **Security:** Keep strengthening authorization, administrative tools, agent tool execution, prompt-injection boundaries, filesystem access, and untrusted-content handling.
+- **Integrations and degraded states:** Audit supported providers and optional services, including ChromaDB, SearXNG, email, and ntfy. Make failures observable and nonessential outages predictable.
 
-I prob shouldnt add more themes.
+> [!TIP]
+> Start with an existing, reproducible issue. The [Stabilization & Modularization milestone](https://github.com/odysseus-dev/odysseus/milestone/1) is a curated work queue, not an invitation to implement every item in this roadmap without coordination.
+
+## Modularization and maintainability
+
+The goal is **clearer responsibility boundaries and safer development**, not smaller files for their own sake. Changes should preserve behavior, compatibility, and relevant runtime contracts.
+
+| Area | Current foundation | Remaining opportunities |
+| --- | --- | --- |
+| **CSS** | The former `static/style.css` has been split into ordered files under `static/css/` | Improve ownership while preserving the cascade, themes, and responsive layouts |
+| **Tools** | Substantial tool-domain extraction has landed | Clarify remaining ownership and preserve compatibility through the existing tool interfaces |
+| **Frontend** | Existing feature modules remain in use | Decompose large document, gallery, settings, calendar, and other UI modules around coherent behavior |
+| **Backend** | Domain-specific routes and services already exist | Identify safe boundaries in Cookbook routes, persistence, and other tightly coupled modules |
+| **Tests** | Taxonomy, focused execution, and structural improvements exist | Align fixtures, helpers, and regression coverage with implementation boundaries |
+| **UI infrastructure** | Targeted cleanup is underway | Consolidate tour scaffolding, modal/window positioning, responsive overrides, and obsolete paths |
+
+**Canonical coordination:** [Codebase modularization and contributor onboarding, #4071](https://github.com/odysseus-dev/odysseus/issues/4071).
+
+> [!IMPORTANT]
+> **Cohesive work is more important than a small line count.** A larger PR may be appropriate for one well-defined decomposition when several smaller PRs would create avoidable conflicts or integration overhead. Coordinate the boundary, avoid unrelated work, and supply regression and runtime evidence. See [CONTRIBUTING.md](CONTRIBUTING.md#structural-changes-and-modularization).
+
+### Agent runtime and Harness V2
+
+Harness V2 is under active development.
+
+> [!CAUTION]
+> **Temporary agent-runtime freeze.** Unrelated changes to `src/agent_loop.py` and closely related execution, orchestration, authorization, and context contracts must be coordinated with maintainers while Harness V2 is developed and validated.
+>
+> This does **not** freeze unrelated application modules. Urgent bug reports and security issues remain welcome. No release date is implied.
+
+## AI, local models, and research
+
+These are **areas of product interest**, not commitments to a specific implementation or delivery order. Follow normal issue triage and product-decision requirements.
+
+| Direction | Opportunity |
+| --- | --- |
+| **Smaller-model agents** | Reduce unnecessary prompt, tool-schema, skill, memory, and document context overhead; improve tool selection for constrained context windows |
+| **Cookbook and SGLang** | Improve installation and serving across relevant hardware, operating systems, and deployment methods |
+| **Model discovery** | Rank models by architecture, quantization, RAM/VRAM fit, backend support, multimodal requirements, and expected reliability |
+| **Deep Research presets** | Explore suitable model and parameter configurations for small, medium, and large local systems |
+| **Actionable diagnostics** | Expose command output, logs, and next steps for failed downloads, dependencies, preflights, and serving jobs |
+| **Speculative decoding** | Investigate compatible draft-model configurations and evaluate end-to-end correctness and time to success |
+| **Provider setup** | Improve endpoint discovery, capability checks, and setup for supported local and hosted providers |
+
+## Frontend and user experience
+
+- **Editors and files:** Improve document/file editing, window behavior, save/export flows, and image-editing reliability.
+- **Notes and Todos:** Expand agent-assisted reading, updates, summarization, and turning information into tasks.
+- **Mobile:** Improve gallery/editor usability, model setup, and responsive behavior.
+- **Accessibility:** Strengthen keyboard navigation, focus states, contrast, and reduced-motion support.
+- **First run:** Improve setup, hints, empty states, error messages, and onboarding tours without duplicate guidance.
+- **Self-hosting:** Reduce unnecessary external asset dependencies where practical for offline operation.
+
+## Backend and operations
+
+- **Email performance:** Measure fetching, searching, body and attachment loading, caching, sending, and frontend refresh behavior before optimizing.
+- **Provider reliability:** Expand regression coverage for setup, endpoint probing, and degraded-service behavior.
+- **Task scheduling:** Improve defaults, reliability, and operational visibility.
+- **Data recovery:** Improve backup and restore guidance and workflows for persistent data.
+- **Administrative security:** Review privileged capabilities and sensitive integrations against their authorization boundaries.
+
+## Not a current priority
+
+- Additional themes without a demonstrated product need.
+- Broad rewrites or framework migrations without clear, measurable value.
+- Changes undertaken solely to reduce file counts.
+- Uncoordinated changes to temporarily frozen agent-runtime contracts.
+
+---
+
+## Contributing
+
+1. Read [CONTRIBUTING.md](CONTRIBUTING.md) and search [existing issues](https://github.com/odysseus-dev/odysseus/issues) and [PRs](https://github.com/odysseus-dev/odysseus/pulls).
+2. Coordinate modularization through [#4071](https://github.com/odysseus-dev/odysseus/issues/4071) and active stabilization work through the [milestone](https://github.com/odysseus-dev/odysseus/milestone/1).
+3. Use the repository's issue and PR templates and target **`dev`**.
+4. Include relevant tests, runtime evidence, and screenshots or recordings for visual changes.
+
+> [!NOTE]
+> Product-direction or difficult-to-reverse changes follow the [feature proposal process](docs/maintainers/product-and-features.md). Report security vulnerabilities privately through [SECURITY.md](SECURITY.md), never in a public issue.

THREAT_MODEL.md

@@ -64,12 +64,13 @@ External content that reaches the LLM is treated as untrusted via `src/prompt_se
 
 `src/tool_capabilities.py` carries a second layer: once untrusted content has entered a run, `ToolRunSecurityContext.decision_for()` blocks tools that execute code, mutate state, or cause external side effects until the user authorises the action separately.
 
-It is on unless `ODYSSEUS_TOOL_APPROVAL_GATE` is set to a falsy value (`0`/`false`/`no`/`off`). Request authority narrows which tool families a turn may use, but within an admitted family it does not bind the exact action: a request to read email admits `send_email` and `delete_email`, and agent processes inherit the host network. Until those are covered by their own boundaries, this gate is what stops an injected instruction from reaching them. Turning it off lets a successful injection reach `bash`, `host_shell`, `send_email` and `delete_email` without a separate confirmation.
+It is on unless `ODYSSEUS_TOOL_APPROVAL_GATE` is set to a falsy value (`0`/`false`/`no`/`off`). The admin-only `tool_approval_mode` setting (Settings → Agent Tools, `off` = "Off (YOLO)") is the same knob exposed in the UI: with `off`, every run sets `approval_gate_bypassed` and this gate stops asking, exactly as if the env var were off. Request authority narrows which tool families a turn may use, but within an admitted family it does not bind the exact action: a request to read email admits `send_email` and `delete_email`, and agent processes inherit the host network. Until those are covered by their own boundaries, this gate is what stops an injected instruction from reaching them. Turning it off — by either knob — lets a successful injection reach `bash`, `host_shell`, `send_email` and `delete_email` without a separate confirmation.
 
-Two exemptions apply even when the gate is on, both deliberate:
+Three exemptions apply even when the gate is on, all deliberate:
 
 - Sources in `_CONTROL_PLANE_CONTEXT_SOURCES` (skills, runtime descriptors, the open editor document, the open email, uploaded files) are treated as control-plane metadata and still permit read-only tools.
 - A TUI run that advertises a host shell bridge and declares `unattended_mode` exempts the local execution set in `TUI_CLIENT_TOOL_NAMES`. Personal, network and deployment-local tools are never exempted.
+- Narrow continuation for local MCP tools declared via `ODYSSEUS_LOCAL_MCP_READ_TOOLS` and `ODYSSEUS_LOCAL_MCP_WRITE_TOOLS`: Only explicitly configured stdio MCP servers qualify; remote HTTP and SSE servers fail closed. Declarations never override built-in tool effects, so built-in email, browser, memory, image generation, and other built-in servers cannot receive continuation exceptions. Declared read tools remain untrusted (`EXTERNAL_UNTRUSTED`) to ensure that prompt injections in local data cannot reach shell, network, or unrelated tools without approval. Stdio transport does not guarantee network isolation, as local processes may still reach the host network. However, declared non-destructive write tools on the exact same stdio MCP server are permitted to continue without an approval prompt if the only untrusted context in the run was produced by declared local reads on that same server, without any explicit error or external taint. Genuinely external content (web search, web fetch, emails, external message context) blocks these writes. Authorization must still originate from the trusted request: request authority must permit the tool and retain the sealed backend resource grant. The residual same-store write risk (the agent writing to the same local store after reading a local prompt injection) remains subject to the user's initial request authority.
 
 ## Security Headers
 

core/auth.py

@@ -47,7 +47,7 @@ ADMIN_PRIVILEGES["allowed_models_restricted"] = False
 # backwards for this sentinel.
 ADMIN_PRIVILEGES["block_all_models"] = False
 
-from src.constants import AUTH_FILE, PASSWORD_MIN_LENGTH
+from src.constants import AUTH_FILE, PASSWORD_MIN_LENGTH, USERNAME_MAX_LENGTH
 from src.owner_identity import RESERVED_AUTH_USERNAMES
 DEFAULT_AUTH_PATH = AUTH_FILE
 TOKEN_TTL = 60 * 60 * 24 * 7  # 7 days
@@ -62,6 +62,24 @@ TOKEN_TTL = 60 * 60 * 24 * 7  # 7 days
 RESERVED_USERNAMES = frozenset(RESERVED_AUTH_USERNAMES)
 
 
+def username_error(username: str | None) -> Optional[str]:
+    """Why a new username is not acceptable, or None when it is.
+
+    Admin routes carry the name as a URL path segment, where "/" never
+    reaches the handler, so such an account could not be managed afterwards.
+    """
+    name = (username or "").strip()
+    if not name:
+        return "Username is required"
+    if len(name) > USERNAME_MAX_LENGTH:
+        return f"Username must be at most {USERNAME_MAX_LENGTH} characters"
+    if "/" in name or "\\" in name or any(not c.isprintable() for c in name):
+        return "Username must not contain slashes or control characters"
+    if name in (".", ".."):
+        return "Username is not allowed"
+    return None
+
+
 def normalize_known_username(users: Dict[str, Any], username: str | None) -> Optional[str]:
     """Return a normalized username only when it exists in the auth user map."""
     key = str(username or "").strip().lower()
@@ -396,6 +414,15 @@ class AuthManager:
             current = self.get_privileges(username)
             for k, v in privileges.items():
                 if k in DEFAULT_PRIVILEGES:
+                    default = DEFAULT_PRIVILEGES[k]
+                    if isinstance(default, bool):
+                        ok = isinstance(v, bool)
+                    elif isinstance(default, int):
+                        ok = isinstance(v, int) and not isinstance(v, bool)
+                    else:
+                        ok = isinstance(v, list)
+                    if not ok:
+                        raise ValueError(f"Invalid value for privilege '{k}'")
                     current[k] = v
             self._config["users"][username]["privileges"] = current
             self._save()
@@ -436,6 +463,17 @@ class AuthManager:
                 admin_count = sum(1 for d in self.users.values() if d.get("is_admin"))
                 if admin_count <= 1:
                     return SetAdminResult.LAST_ADMIN
+                # Token creation is admin-only and a demoted user can neither
+                # list nor revoke their tokens, so purge them with the role.
+                # Done before the flag flips: if the token store fails the
+                # exception propagates and the user is still an admin.
+                from core.database import get_db_session, ApiToken
+                with get_db_session() as db:
+                    removed_tokens = db.query(ApiToken).filter(ApiToken.owner == username).delete()
+                if removed_tokens:
+                    logger.info(
+                        f"Revoked {removed_tokens} API token(s) owned by demoted admin '{username}'"
+                    )
             # Write order matters for lock-free readers: get_privileges()
             # reads without _config_lock and trusts is_admin, so the admin
             # flag must be flipped while the stored map is safe to expose —

core/database.py

@@ -568,6 +568,7 @@ class ModelEndpoint(TimestampMixin, Base):
     hidden_models = Column(Text, nullable=True)    # JSON list of model IDs that failed probing
     cached_models = Column(Text, nullable=True)    # JSON list of last-known model IDs (avoids probe on list)
     pinned_models = Column(Text, nullable=True)    # JSON list of admin-pinned model IDs (manual, may not appear in /v1/models)
+    cookbook_serve_session_id = Column(String, nullable=True)  # Current Cookbook serve identity
     model_type = Column(String, nullable=True, default="llm")  # "llm" or "image"
     # auto = classify by URL; local = self-hosted server; api/proxy = external
     # OpenAI-compatible API even when reachable through a private/tailnet IP.
@@ -1378,6 +1379,26 @@ def _migrate_add_pinned_models_column():
         except Exception:
             pass
 
+
+def _migrate_add_cookbook_serve_session_id_column():
+    """Bind Cookbook cleanup to the serve session that owns the endpoint."""
+    from sqlalchemy import inspect
+
+    with engine.begin() as conn:
+        columns = {
+            column["name"]
+            for column in inspect(conn).get_columns("model_endpoints")
+        }
+
+        if "cookbook_serve_session_id" not in columns:
+            conn.execute(
+                text(
+                    "ALTER TABLE model_endpoints "
+                    "ADD COLUMN cookbook_serve_session_id VARCHAR"
+                )
+            )
+
+
 def _migrate_add_notes_sort_order():
     """Add sort_order, image_url, repeat columns to notes if they don't exist."""
     import sqlite3
@@ -2390,6 +2411,7 @@ def init_db():
     _migrate_add_hidden_models_column()
     _migrate_add_cached_models_column()
     _migrate_add_pinned_models_column()
+    _migrate_add_cookbook_serve_session_id_column()
     _migrate_add_notes_sort_order()
     _migrate_add_model_type_column()
     _migrate_add_model_endpoint_refresh_columns()

docker-compose.gpu-amd.yml

@@ -73,6 +73,8 @@ services:
       - ODYSSEUS_CAPTURE_MODEL_REQUESTS=${ODYSSEUS_CAPTURE_MODEL_REQUESTS:-0}
       - ODYSSEUS_MCP_EMAIL_OWNER=${ODYSSEUS_MCP_EMAIL_OWNER:-}
       - ODYSSEUS_SCRIPT_HOST=${ODYSSEUS_SCRIPT_HOST:-localhost}
+      - ODYSSEUS_LOCAL_MCP_READ_TOOLS=${ODYSSEUS_LOCAL_MCP_READ_TOOLS:-}
+      - ODYSSEUS_LOCAL_MCP_WRITE_TOOLS=${ODYSSEUS_LOCAL_MCP_WRITE_TOOLS:-}
       - ODYSSEUS_CHAT_UPLOAD_MAX_BYTES=${ODYSSEUS_CHAT_UPLOAD_MAX_BYTES:-10485760}
       - ODYSSEUS_GALLERY_UPLOAD_MAX_BYTES=${ODYSSEUS_GALLERY_UPLOAD_MAX_BYTES:-104857600}
       - ODYSSEUS_GALLERY_TRANSFORM_UPLOAD_MAX_BYTES=${ODYSSEUS_GALLERY_TRANSFORM_UPLOAD_MAX_BYTES:-26214400}
@@ -82,13 +84,15 @@ services:
       - ODYSSEUS_EDITOR_DRAFT_MAX_BYTES=${ODYSSEUS_EDITOR_DRAFT_MAX_BYTES:-268435456}
       - ODYSSEUS_STT_MAX_AUDIO_BYTES=${ODYSSEUS_STT_MAX_AUDIO_BYTES:-26214400}
       - ODYSSEUS_ICS_MAX_BYTES=${ODYSSEUS_ICS_MAX_BYTES:-10485760}
-      - ODYSSEUS_TTS_CACHE_MAX_BYTES=${ODYSSEUS_TTS_CACHE_MAX_BYTES}
+      - ODYSSEUS_TTS_CACHE_MAX_BYTES=${ODYSSEUS_TTS_CACHE_MAX_BYTES:-524288000}
       # Host workspace translation is opt-in. Keep the public compose file
       # user-neutral; configure these in a local .env or use the host-workspace
       # overlay with ODYSSEUS_HOST_WORKSPACE_DIR.
       - ODYSSEUS_WORKSPACE_HOST_ROOT=${ODYSSEUS_WORKSPACE_HOST_ROOT:-}
       - ODYSSEUS_WORKSPACE_CONTAINER_ROOT=${ODYSSEUS_WORKSPACE_CONTAINER_ROOT:-/workspace}
       - ODYSSEUS_WORKSPACE_DEFAULT=${ODYSSEUS_WORKSPACE_DEFAULT:-}
+      # Local first-token watchdog in seconds; 0 disables it. Empty keeps the code default (60s).
+      - ODYSSEUS_FIRST_TOKEN_TIMEOUT=${ODYSSEUS_FIRST_TOKEN_TIMEOUT:-}
       - DATA_BRAVE_API_KEY=${DATA_BRAVE_API_KEY:-}
       - GOOGLE_API_KEY=${GOOGLE_API_KEY:-}
       - GOOGLE_PSE_CX=${GOOGLE_PSE_CX:-}
@@ -165,7 +169,7 @@ services:
         /usr/local/searxng/.venv/bin/python /tmp/migrate-searxng-settings.py /etc/searxng/settings.yml || true
         exec /usr/local/searxng/entrypoint.sh
     ports:
-      - "127.0.0.1:8080:8080"
+      - "127.0.0.1:${SEARXNG_HOST_PORT:-8080}:8080"
     volumes:
       - searxng-data:/etc/searxng
       - ./config/searxng/settings.yml:/tmp/searxng-settings.yml.template:ro,z

docker-compose.gpu-nvidia.yml

@@ -72,6 +72,8 @@ services:
       - ODYSSEUS_CAPTURE_MODEL_REQUESTS=${ODYSSEUS_CAPTURE_MODEL_REQUESTS:-0}
       - ODYSSEUS_MCP_EMAIL_OWNER=${ODYSSEUS_MCP_EMAIL_OWNER:-}
       - ODYSSEUS_SCRIPT_HOST=${ODYSSEUS_SCRIPT_HOST:-localhost}
+      - ODYSSEUS_LOCAL_MCP_READ_TOOLS=${ODYSSEUS_LOCAL_MCP_READ_TOOLS:-}
+      - ODYSSEUS_LOCAL_MCP_WRITE_TOOLS=${ODYSSEUS_LOCAL_MCP_WRITE_TOOLS:-}
       - ODYSSEUS_CHAT_UPLOAD_MAX_BYTES=${ODYSSEUS_CHAT_UPLOAD_MAX_BYTES:-10485760}
       - ODYSSEUS_GALLERY_UPLOAD_MAX_BYTES=${ODYSSEUS_GALLERY_UPLOAD_MAX_BYTES:-104857600}
       - ODYSSEUS_GALLERY_TRANSFORM_UPLOAD_MAX_BYTES=${ODYSSEUS_GALLERY_TRANSFORM_UPLOAD_MAX_BYTES:-26214400}
@@ -81,13 +83,15 @@ services:
       - ODYSSEUS_EDITOR_DRAFT_MAX_BYTES=${ODYSSEUS_EDITOR_DRAFT_MAX_BYTES:-268435456}
       - ODYSSEUS_STT_MAX_AUDIO_BYTES=${ODYSSEUS_STT_MAX_AUDIO_BYTES:-26214400}
       - ODYSSEUS_ICS_MAX_BYTES=${ODYSSEUS_ICS_MAX_BYTES:-10485760}
-      - ODYSSEUS_TTS_CACHE_MAX_BYTES=${ODYSSEUS_TTS_CACHE_MAX_BYTES}
+      - ODYSSEUS_TTS_CACHE_MAX_BYTES=${ODYSSEUS_TTS_CACHE_MAX_BYTES:-524288000}
       # Host workspace translation is opt-in. Keep the public compose file
       # user-neutral; configure these in a local .env or use the host-workspace
       # overlay with ODYSSEUS_HOST_WORKSPACE_DIR.
       - ODYSSEUS_WORKSPACE_HOST_ROOT=${ODYSSEUS_WORKSPACE_HOST_ROOT:-}
       - ODYSSEUS_WORKSPACE_CONTAINER_ROOT=${ODYSSEUS_WORKSPACE_CONTAINER_ROOT:-/workspace}
       - ODYSSEUS_WORKSPACE_DEFAULT=${ODYSSEUS_WORKSPACE_DEFAULT:-}
+      # Local first-token watchdog in seconds; 0 disables it. Empty keeps the code default (60s).
+      - ODYSSEUS_FIRST_TOKEN_TIMEOUT=${ODYSSEUS_FIRST_TOKEN_TIMEOUT:-}
       - DATA_BRAVE_API_KEY=${DATA_BRAVE_API_KEY:-}
       - GOOGLE_API_KEY=${GOOGLE_API_KEY:-}
       - GOOGLE_PSE_CX=${GOOGLE_PSE_CX:-}
@@ -168,7 +172,7 @@ services:
         /usr/local/searxng/.venv/bin/python /tmp/migrate-searxng-settings.py /etc/searxng/settings.yml || true
         exec /usr/local/searxng/entrypoint.sh
     ports:
-      - "127.0.0.1:8080:8080"
+      - "127.0.0.1:${SEARXNG_HOST_PORT:-8080}:8080"
     volumes:
       - searxng-data:/etc/searxng
       - ./config/searxng/settings.yml:/tmp/searxng-settings.yml.template:ro,z

docker-compose.yml

@@ -61,6 +61,8 @@ services:
       - ODYSSEUS_CAPTURE_MODEL_REQUESTS=${ODYSSEUS_CAPTURE_MODEL_REQUESTS:-0}
       - ODYSSEUS_MCP_EMAIL_OWNER=${ODYSSEUS_MCP_EMAIL_OWNER:-}
       - ODYSSEUS_SCRIPT_HOST=${ODYSSEUS_SCRIPT_HOST:-localhost}
+      - ODYSSEUS_LOCAL_MCP_READ_TOOLS=${ODYSSEUS_LOCAL_MCP_READ_TOOLS:-}
+      - ODYSSEUS_LOCAL_MCP_WRITE_TOOLS=${ODYSSEUS_LOCAL_MCP_WRITE_TOOLS:-}
       - ODYSSEUS_CHAT_UPLOAD_MAX_BYTES=${ODYSSEUS_CHAT_UPLOAD_MAX_BYTES:-10485760}
       - ODYSSEUS_GALLERY_UPLOAD_MAX_BYTES=${ODYSSEUS_GALLERY_UPLOAD_MAX_BYTES:-104857600}
       - ODYSSEUS_GALLERY_TRANSFORM_UPLOAD_MAX_BYTES=${ODYSSEUS_GALLERY_TRANSFORM_UPLOAD_MAX_BYTES:-26214400}
@@ -70,13 +72,15 @@ services:
       - ODYSSEUS_EDITOR_DRAFT_MAX_BYTES=${ODYSSEUS_EDITOR_DRAFT_MAX_BYTES:-268435456}
       - ODYSSEUS_STT_MAX_AUDIO_BYTES=${ODYSSEUS_STT_MAX_AUDIO_BYTES:-26214400}
       - ODYSSEUS_ICS_MAX_BYTES=${ODYSSEUS_ICS_MAX_BYTES:-10485760}
-      - ODYSSEUS_TTS_CACHE_MAX_BYTES=${ODYSSEUS_TTS_CACHE_MAX_BYTES}
+      - ODYSSEUS_TTS_CACHE_MAX_BYTES=${ODYSSEUS_TTS_CACHE_MAX_BYTES:-524288000}
       # Host workspace translation is opt-in. Keep the public compose file
       # user-neutral; configure these in a local .env or use the host-workspace
       # overlay with ODYSSEUS_HOST_WORKSPACE_DIR.
       - ODYSSEUS_WORKSPACE_HOST_ROOT=${ODYSSEUS_WORKSPACE_HOST_ROOT:-}
       - ODYSSEUS_WORKSPACE_CONTAINER_ROOT=${ODYSSEUS_WORKSPACE_CONTAINER_ROOT:-/workspace}
       - ODYSSEUS_WORKSPACE_DEFAULT=${ODYSSEUS_WORKSPACE_DEFAULT:-}
+      # Local first-token watchdog in seconds; 0 disables it. Empty keeps the code default (60s).
+      - ODYSSEUS_FIRST_TOKEN_TIMEOUT=${ODYSSEUS_FIRST_TOKEN_TIMEOUT:-}
       - DATA_BRAVE_API_KEY=${DATA_BRAVE_API_KEY:-}
       - GOOGLE_API_KEY=${GOOGLE_API_KEY:-}
       - GOOGLE_PSE_CX=${GOOGLE_PSE_CX:-}
@@ -146,7 +150,7 @@ services:
         /usr/local/searxng/.venv/bin/python /tmp/migrate-searxng-settings.py /etc/searxng/settings.yml || true
         exec /usr/local/searxng/entrypoint.sh
     ports:
-      - "127.0.0.1:8080:8080"
+      - "127.0.0.1:${SEARXNG_HOST_PORT:-8080}:8080"
     volumes:
       - searxng-data:/etc/searxng
       - ./config/searxng/settings.yml:/tmp/searxng-settings.yml.template:ro,z

docker/build-realesrgan-wheels.sh

@@ -31,21 +31,43 @@ for spec in $SPECS; do
   name="${spec%%==*}"
   ver="${spec##*==}"
   # pip download builds metadata (and trips the same bug), so fetch the raw
-  # sdist URL from the PyPI JSON API instead.
+  # sdist URL from the PyPI JSON API instead.  Use curl with retries to
-  url="$(python - "$name" "$ver" <<'PY'
+  # survive transient TLS / network failures that urllib cannot recover from.
-import json, sys, urllib.request
+  metadata="${name}-${ver}.json"
-name, ver = sys.argv[1], sys.argv[2]
+  curl --fail --silent --show-error --location \
-data = json.load(urllib.request.urlopen(f"https://pypi.org/pypi/{name}/{ver}/json"))
+    --retry 8 \
-for f in data["urls"]:
+    --retry-delay 3 \
-    if f["packagetype"] == "sdist":
+    --retry-max-time 120 \
-        print(f["url"]); break
+    --retry-all-errors \
+    "https://pypi.org/pypi/${name}/${ver}/json" \
+    -o "$metadata"
+
+  url="$(python - "$metadata" <<'PY'
+import json, sys
+
+with open(sys.argv[1], encoding="utf-8") as f:
+    data = json.load(f)
+
+for item in data["urls"]:
+    if item["packagetype"] == "sdist":
+        print(item["url"])
+        break
 else:
-    sys.exit(f"no sdist found for {name}=={ver}")
+    raise SystemExit("no sdist found")
 PY
   )"
+  rm -f "$metadata"
+
   echo ">> fetching ${name} ${ver}: ${url}"
-  curl -fsSL "$url" -o "${name}.tar.gz"
+  curl --fail --silent --show-error --location \
+    --retry 8 \
+    --retry-delay 3 \
+    --retry-max-time 120 \
+    --retry-all-errors \
+    "$url" \
+    -o "${name}.tar.gz"
   tar xzf "${name}.tar.gz"
+  rm -f "${name}.tar.gz"
 done
 
 echo ">> patching get_version()"

docs/QUICKEDIT_COVERAGE.md

@@ -0,0 +1,255 @@
+# Quick Edit Coverage Audit
+
+Audit of how much of the PhotoCraft editor is reachable through **Quick Edit**
+(the natural-language editor agent). Built against build **0.20.44** (HARNESS
+bridge fan-out fix `4dd11826`) and client commit `e78b5a96`.
+
+Quick Edit executes in two stages:
+
+1. **Plan** — `static/js/editor/editor-agent.js` posts `{prompt, state,
+   screenshot}` to `/api/editor/agent`. HARNESS `routes/editor_agent_routes.py`
+   plans with the shared style guide (`src/editor_style_guide.py`) plus an LLM,
+   validates every step against its own `OP_CATALOG`, then executes.
+2. **Execute** — only `ai_edit_image` / `generate_image` run server-side; every
+   other op is delivered to the page over the editor bridge (SSE) and run by the
+   client control surface (`window.geEditorControl.invoke`, catalog in
+   `static/js/galleryEditor.js`). A local fallback planner lives in
+   `static/js/editor/nl-edit.js`.
+
+There are **three** catalogs that must agree:
+
+| Catalog | File | Owner |
+|---|---|---|
+| Client control ops | `static/js/galleryEditor.js` | PHOTOSHOP |
+| Server `OP_CATALOG` | `routes/editor_agent_routes.py` | HARNESS |
+| Style guide / planner | `src/editor_style_guide.py` + `nl-edit.js` | HARNESS + PHOTOSHOP |
+
+Legend: **C** = in client catalog · **S** = in server `OP_CATALOG` (incl.
+`NEW_OP_NAMES`) · **P** = the planner/style guide emits it for some phrasing.
+A capability is **usable from Quick Edit** only when the planner can emit it
+(**P**) and the op exists on the side that runs it.
+
+---
+
+## 1. Menu bar (12 menus)
+
+| Capability | Trigger phrasing | Op | C | S | P | Status |
+|---|---|---|---|---|---|---|
+| File → Import | "open this image" | `document.importImage` | ✔ | ✔ | partial | OK |
+| File → Load project | "open my project" | `document.openPsd` | ✔ | ✘ | ✘ | **Server gap** |
+| File → Save / Save copy / Export PSD | "save as psd" | `document.savePsd` | ✔ | ✔ | ✔ | OK |
+| File → Export | "export png" | `export.png` | ✔ | ✔ | ✔ | OK |
+| File → Save project | — | (project store) | n/a | n/a | ✘ | Out of scope (not an edit op) |
+| Edit → Undo/Redo/History | "undo" | `history.undo` / `redo` | ✔ | ✔ | ✔ | OK |
+| Edit → Free Transform | "free transform" | `transform.start` | ✔ | ✔ | ✔ | OK |
+| Edit → Transform → Rotate 90/180 | "rotate this layer 90" | `layer.rotate` | ✔ | ✔ | ✔ | OK (whole image vs layer routed by the fast path) |
+| Edit → Transform → Flip H/V | "flip this layer" | `layer.flip` | ✔ | ✔ | ✔ | OK (whole image vs layer routed by the fast path) |
+| Image → Fill | "fill the whole image with white" | — | ✘ | ✘ | ✘ | **No op** |
+| Image → Canvas Size | "make the canvas wider" | `document.resizeCanvas` | ✔ | ✔ | partial | OK |
+| Image → Image Size | "resize the image to 1200px" | `document.resizeCanvas` | ✔ | ✔ | partial | OK |
+| Image → Rotate 90/180 (whole image) | "rotate the whole image 90" | `document.rotate`/all-layers | ✘ | ✘ | ✘ | **No control op** |
+| Image → Flip H/V (whole image) | "flip the whole canvas" | `document.flip`/all-layers | ✘ | ✘ | ✘ | **No control op** |
+| Image → Mode RGB/Grayscale | "make it black and white" | `adjustments.apply` black-white | ✔ | ✔ | ✔ | OK (mode via adjustment) |
+| Image → Adjustments → Invert | "invert the colours" | `adjustments.apply` invert | ✔ | ✔ | ✔ | OK (fast path colour-inverts the active layer; "invert the selection" stays Select ▸ Inverse) |
+| Image → Mode CMYK/Lab, Bits 16/32 | — | — | ✘ | ✘ | ✘ | Not implemented (disabled in UI) |
+| Layer → New | "add a layer" | `layers.addRaster` | ✔ | ✔ | partial | OK |
+| Layer → Duplicate | "duplicate the layer" | `layers.duplicate` | ✔ | ✔ | partial | OK |
+| Layer → Group | "group the layers" | `layers.group` | ✔ | ✔ | ✘ | **Server OK, planner gap** |
+| Layer → Rasterize | "rasterize the text" | `layers.rasterize` | ✔ | ✔ | ✔ | OK |
+| Layer → Delete | "delete the layer" | `layers.delete` | ✔ | ✔ | partial | OK |
+| Layer → Merge Down | "merge down" | `layers.mergeDown` | ✔ | ✘ | ✘ | **Server gap** |
+| Layer → Merge All / Flatten | "flatten the image" | `layers.flatten` | ✔ | ✘ | ✘ | **Server gap** |
+| Layer → Select All | "select all layers" | (multi-select) | ✔ | ✔ | ✔ | OK (fast path selects every layer) |
+| Layer → Combine (unite/subtract/…) | "combine the shapes" | `_combineSelectedShapes` | ✔ | ✔ | ✔ | OK (fast path selects the shape layers and combines) |
+| Layer → Fill | "fill the layer with blue" | `selection.fill` / flood | partial | ✘ | ✘ | **Gap** |
+| Layer → Layer Mask (reveal/hide/apply/invert/feather) | "add a layer mask" | `layers.addMaskFromSelection` | ✔ | ✔ | ✔ | OK |
+| Layer → Opacity | "set the layer opacity to 50%" | `layers.setOpacity` | ✔ | ✔ | partial | OK (fast path sets the active layer's opacity) |
+| Layer → Blend Mode | "set the blend mode to multiply" | `layers.setBlend` | ✔ | ✔ | partial | OK (fast path sets the active layer's blend mode) |
+| Type → Add | "add a title" | `text.add` | ✔ | ✔ | ✔ | OK |
+| Type → Character | "make the text bigger" | `text.edit` | ✔ | ✘ | ✘ | **Server gap + planner gap** |
+| Type → Paragraph | "centre the text" | `text.layout` | ✔ | ✔ | partial | OK |
+| Type → Warp | "warp the text" | `_rerenderTextLayer` (warp) | ✔ | ✔ | ✔ | OK |
+| Select → All/Deselect/Invert | "invert the selection" | `selection.*` | ✔ | ✔ | ✔ | OK |
+| Select → Reselect | "reselect" | `_reselectSelection` | ✔ | ✔ | ✔ | OK |
+| Select → Transform | "transform the selection" | `transform.start` | ✔ | ✔ | partial | OK |
+| Select → Refine | "refine the edge" | `selection.feather/expand` | ✔ | ✘ | ✘ | **Server gap** |
+| Select → Color Range | "select the blue sky" | `selection.colorRange` | ✔ | ✔ | partial | OK |
+| Select → Select and Mask | "select the subject" | `selection.subject` | ✔ | ✔ | ✔ | OK |
+| Select → Save | "save the selection" | `_saveNamedSelection` | ✔ | ✔ | ✔ | OK |
+| Select → Load | "load the selection as sky" | `_loadNamedSelection` | ✔ | ✔ | ✔ | OK |
+| Filter → every filter (30+) | "pixelate the image" | `filters.apply` | ✔ | ✔ | ✔ | OK (fast path covers pixelate/emboss/twirl/pinch/spherize/ripple/solarize/find-edges/crystallize/noise/despeckle/unsharp + blur/sharpen/vignette) |
+| Filter → Blur → Motion / Zoom | "motion blur" / "zoom blur" | `filters.apply` (blur-motion / blur-zoom) | ✔ | ✔ | ✔ | OK (fast path maps motion blur and zoom blur to their Filter-menu variants) |
+| Filter → Layer effects (drop shadow, stroke, colour overlay, outer glow) | "add a drop shadow" | `layers.addStyle` | ✔ | ✔ | ✔ | OK (fast path adds drop shadow / stroke / outer glow / colour & gradient overlay; colour parsed from the prompt) |
+| View → Rulers/Grid/Snap/Clear guides | "show the grid" | `#ge-view-menu` | ✔ | ✔ | ✔ | OK |
+| Window → panels | — | n/a | n/a | n/a | ✘ | UI-only |
+| 3D / Plugins | — | — | ✘ | ✘ | ✘ | Not implemented (disabled in UI) |
+
+## 2. Tools
+
+Toolbar tools are interactive, not bridge ops, so coverage is about whether the
+**effect** of a tool is reachable by prompt.
+
+| Tool | NL-reachable? | Notes |
+|---|---|---|
+| Move / Transform | `transform.*` | OK (transform.start/apply/cancel) |
+| Hand / Zoom | — | View only |
+| Crop | `document.crop` + format guard | OK |
+| Text / Shape | `text.add`, `text.layout` | Shape has no add-shape op — **gap** |
+| Brush / Eraser / Dodge / Burn / Smudge / Clone / Heal / Sharpen | — | Interactive painting; no NL op — **gap** (out of scope for planner) |
+| Gradient | — | **No op** |
+| **Paint Bucket** | `fill.flood` | **Server gap + planner gap** (op exists client-side only) |
+| Eyedropper | — | UI only |
+| Marquee / Lasso / Wand / Quick Select / Object Select / Pen / Path | `selection.rectangle/ellipse/polygon/subject/…` | OK for region/reference selects; lasso/pen paths have `selection.polygon` only |
+| SAM / Inpaint / Rembg | `selection.subject`, `ai.removeObject`, `ai.removeBackground` | OK |
+| **Pixel brush + grid** | — | UI state only; no op — **gap** (acceptable: it is a brush mode, not an edit op) |
+
+## 3. Right-click context menu
+
+| Item | Backing op | Reachable by prompt? |
+|---|---|---|
+| Brush size / preset / Pixel Brush | `state` UI | No (interactive) |
+| Select All / Deselect / Inverse / Feather / Expand / Contract / Border / Smooth | `selection.*` | Yes |
+| Transform Selection / Refine Edge | `transform.start`, `selection.feather` | Partial |
+| Apply / Cancel Crop | `document.crop` | Yes |
+| Commit / Cancel / Free Transform | `transform.apply/cancel/start` | Yes |
+| Add Text Layer / Rasterize Text | `text.add`, `layers.rasterize` | Yes (rasterize OK) |
+| Copy Color Hex / Swap Colors | `state.color` | OK (fast path swaps / resets the foreground & background swatches) |
+| Fill Selection w/ Foreground | `selection.fill` | OK (fast path fills the selection/canvas with the foreground or background colour) |
+| Cut / Copy / Paste | `selection.copy/paste` | **Gap** (no planner intent) |
+| Rotate 90/180, Flip H/V | `layer.rotate/flip` | **Server+planner gap** (same as Edit menu) |
+| Undo / Redo | `history.*` | Yes |
+
+## 4. Adjustments
+
+`adjustments.apply` / `adjustments.applyMasked` cover all adjustment types
+(brightness-contrast, white-balance, shadows-highlights, vibrance, clarity,
+vignette, black-white, …). The style guide emits several. **Status: OK**
+(planner covers the common ones; the long tail is reachable only via explicit
+wording).
+
+## 5. Filters
+
+`filters.apply` / `filters.applyMasked` run every filter the UI lists. The
+planner only emits **sharpen**, **blur-gaussian**, and **vignette** via intents.
+**Gap (planner):** no intent for noise, pixelate, stylize, distort, render,
+unsharp, sharpen-edges, colour-overlay/drop-shadow/stroke.
+**Gap (HARNESS):** `filters.apply`/`applyMasked` are in the server catalog, so
+the fix is style-guide intents, not catalog.
+
+## 6. Layer styles
+
+`layers.addStyle {type, params}` exists in both catalogs. **Planner never emits
+it**, so "add a drop shadow to the title", "add a stroke", "colour overlay" fail.
+**Gap (planner).**
+
+## 7. Selection ops
+
+`selection.rectangle/ellipse/polygon/fill/delete/copy/paste/feather/expand/
+contract/selectAll/deselect/invert/colorRange/subject/samText/face/sky/
+background/hasMask/maskBounds/exportMask/regionLuminance`. **Status: mostly OK.**
+Missing: reselect, save-selection.
+
+## 8. Document ops
+
+| Capability | Op | C | S | P |
+|---|---|---|---|---|
+| Crop | `document.crop` | ✔ | ✔ | ✔ |
+| Resize canvas / image | `document.resizeCanvas` | ✔ | ✔ | partial |
+| **Extend canvas** | `document.extend` | ✔ | ✘ | ✘ |
+| Save PSD | `document.savePsd` | ✔ | ✔ | ✔ |
+| Export layers | `document.exportLayers` | ✔ | ✘ | ✘ |
+| Duplicate / Create / Switch / List | `document.*` | ✔ | ✘ | partial |
+
+## 9. AI edit & generate modes
+
+| Mode | Op | Notes |
+|---|---|---|
+| Generative fill | `ai.generativeFill` | OK |
+| Remove object | `ai.removeObject` | OK |
+| Harmonize | `ai.harmonize` | OK |
+| Upscale | `ai.generativeUpscale` | OK |
+| Generate image | `ai.generateImage` / server `generate_image` | OK |
+| Expand | `ai.generativeExpand` | OK (content-bounds prefill, SW v760) |
+| Remove background | `ai.removeBackground` | OK |
+| Instruction edit / scene replace | server `ai_edit_image` | OK (HARNESS) |
+
+---
+
+## 10. Gap list
+
+### Client-side (fix in this repo — PHOTOSHOP-owned) — **ALL FIXED**
+1. ✅ **`fill.flood` not in the NL path** — added to `nl-edit.js OP_CATALOG` +
+   `localPlan` fill intent ("fill … with <colour>"; falls back to `selection.fill`
+   when a region is selected).
+2. ✅ **`layer.flip` / `layer.rotate` not in the NL path** — added to
+   `nl-edit.js OP_CATALOG` + "this layer" intent (whole-image phrasing routes to
+   `document.flip`/`document.rotate`).
+3. ✅ **`text.edit` not in the NL path** — added catalog entry + intent for font
+   size/colour ("make the text bigger/smaller", "make it red").
+4. ✅ **`document.extend` not in the NL path** — added catalog entry + intent for
+   "extend the canvas to 16:9" (derives target size from the document).
+5. ✅ **`layers.align`** — added planner intent ("align … to the centre/left/…").
+6. ✅ **`layers.addStyle`** — added planner trigger ("drop shadow", "stroke",
+   "colour overlay").
+7. ✅ **Whole-image rotate/flip has no control op** — added `document.rotate` /
+   `document.flip` control ops (`galleryEditor.js`) and catalog/intent entries.
+8. ✅ **`selection.copy/paste` intents** — added "copy the selection"/"paste".
+
+### Server-side (file for HARNESS in the passover)
+1. `routes/editor_agent_routes.py OP_CATALOG` is missing ops the client runs:
+   `layer.flip`, `layer.rotate`, `text.edit`, `fill.flood`, `layers.align`,
+   `layers.rasterize`, `layers.mergeDown`, `layers.flatten`, `document.extend`,
+   `document.exportLayers`, `document.*` (list/duplicate/create/switch/openPsd),
+   `mask.*`, `selection.*` (polygon/feather/expand/contract/copy/paste/hasMask/
+   face), `transform.*` are partly present.
+2. `src/editor_style_guide.py` has no intents for: layer transform
+   (rotate/flip active layer), whole-document rotate/flip, text editing, align,
+   layer styles (drop shadow/stroke/overlay), flood fill, copy/paste, noise/
+   pixelate/stylize/distort/render filters.
+
+## 11. Coverage summary
+
+Counting distinct user-facing capabilities in the toolbar + 12 menus + right-click
+(~150 including the filter/effect long tail):
+
+- **Have a client op:** ~85%
+- **Client op also in server OP_CATALOG:** ~75%
+- **Planner can emit for natural phrasing:** ~72% (was ~55%; client-side
+  planner intents for fill, layer/whole-image transform, text edit, align,
+  layer styles, copy/paste and extend are now implemented)
+- **Usable from Quick Edit end-to-end today:** **45/45 prompts (100%)** measured
+  live on build **0.20.45** with `tests/e2e/photo-editor/quickedit-coverage.js`
+  (`/tmp/opencode/qe-coverage45/quickedit-coverage.json`). HARNESS build 0.20.45
+  (`4a6932a2`) added the 12 missing structural ops to the server `OP_CATALOG` +
+  style guide v5, so every audited prompt now plans to the exact op.
+
+The gap was dominated by the **planner**, not by missing client functionality;
+that planner gap is now closed end-to-end (30/45 on 0.20.44 -> 45/45 on 0.20.45).
+
+## 12. Test prompts (~40)
+
+See `docs/QUICKEDIT_PROMPTS.md` and the runner
+`tests/e2e/photo-editor/quickedit-coverage.js` for per-prompt pass/fail.
+
+### Live run (build 0.20.45) — 45/45 PASS
+
+All 45 prompts plan to the exact expected op. HARNESS 0.20.45 routes these
+structural ops server-side: `fill.flood`, `text.edit`, `layers.align`,
+`layer.rotate`/`layer.flip`, `document.rotate`/`document.flip`,
+`selection.copy`/`selection.paste`, `layers.flatten`, `layers.mergeDown` (plus
+the previously passing set). Results:
+`/tmp/opencode/qe-coverage45/quickedit-coverage.json`.
+
+### Previous run (build 0.20.44) — 30/45 PASS
+
+Kept for provenance. Failures were all the server planner emitting a generic
+`adjustments.apply`/`transform.apply` or a wrong op; ids 1, 2, 3, 4, 8, 9, 10,
+11, 12, 13, 14, 19, 20, 40, 41. All now fixed in 0.20.45.
+
+
+## Provenance
+
+Recon inventory captured from `static/js/editor/build/topbar.js`,
+`static/js/editor/build/toolbar.js`, `static/js/editor/canvas-context-menu.js`,
+`static/js/galleryEditor.js` (`installEditorControl`), `static/js/editor/nl-edit.js`,
+`routes/editor_agent_routes.py`, and `src/editor_style_guide.py`.

docs/maintainers/README.md

@@ -9,6 +9,7 @@ This handbook records the project's maintainer governance and gives reviewers a
 | Area | Working rule |
 |---|---|
 | Product direction | Pewds is the Product Owner |
+| New feature | Feature Request issue; big, new-direction or hard-to-undo features need a recorded Product Owner decision ([workflow](product-and-features.md#feature-proposal-workflow)) |
 | Ordinary, reversible maintainer PR | Self-merge allowed after appropriate validation and applicable repository checks |
 | External contribution | Maintainer review required |
 | High-risk change | At least one **independent, qualified** maintainer review |
@@ -22,7 +23,7 @@ These are governance rules, not guarantees that a GitHub permission or branch-pr
 ## Start here
 
 - [Roles and access](roles-and-access.md): responsibility, permissions, and continuity.
-- [Product and feature proposals](product-and-features.md): product approval versus technical ownership.
+- [Product and feature proposals](product-and-features.md): product approval versus technical ownership, and the feature proposal workflow.
 - [Changes and review](changes-and-review.md): review thresholds, self-merge, contributor corrections.
 - [Practical review playbook](review-playbook.md): risk, provenance, CI, CodeQL, independent audit, and handoff evidence.
 - [Triage](triage.md): turning incoming reports into actionable work.

docs/maintainers/product-and-features.md

@@ -2,13 +2,15 @@
 
 Pewds ([@pewdiepie-archdaemon](https://github.com/pewdiepie-archdaemon)) is the Product Owner for Odysseus and has final authority over product direction. Maintainers own technical feasibility, implementation, and safety.
 
+This page has two parts: **which changes need product approval**, and **the workflow every feature proposal follows**, from issue to merged PR. The terms in bold are defined in [Definitions](#definitions); when a case is unclear, the definitions decide, not intuition.
+
 ## Which changes require product approval?
 
-**Product Owner approval before implementation:** substantial new features, changes to product direction, and product changes that are hard to reverse.
+**Product Owner approval before implementation:** features that are **big**, set a **new direction**, or are **hard to undo**.
 
-**No separate product approval by default:** bug fixes, security fixes, maintenance, tests, CI, documentation, internal refactors, and small, reversible product changes within existing direction.
+**No separate product approval by default:** bug fixes, security fixes, maintenance, tests, CI, documentation, internal refactors, and features that are **small**, **reversible**, and **fit ROADMAP**.
 
-Maintainers may make those smaller product decisions. Pewds can ask for changes or a rollback; the maintainer owns safe rollback or recovery.
+Maintainers may make those smaller product decisions. Pewds can ask for changes or a rollback; the maintainer who decided owns the rollback or recovery.
 
 ### Urgent security and reliability fixes
 
@@ -25,19 +27,171 @@ Product Owner proposal or implementation
 
 A Product Owner proposal is already product-approved; it is not automatically technically approved or exempt from high-risk review.
 
-## Community or maintainer initiated
+## Feature proposal workflow
 
-Community proposals start as a Feature Request issue. Check [ROADMAP.md](../../ROADMAP.md) and existing issues first; take early ideas to [Discussions](https://github.com/odysseus-dev/odysseus/discussions/categories/ideas).
+Every community or maintainer feature proposal follows these steps. The issue is the record at every step: labels show where a proposal stands, and comments hold the decisions.
 
-```text
+```mermaid
-Feature proposal
+flowchart TD
-    -> Product Owner accepts / rejects, where approval is required
+    A["0. Idea (optional)<br/>Discussions › Ideas"] -.-> B["1. Feature Request issue"]
-    -> technical feasibility
+    B --> C{"2. Triage<br/>within 7 days"}
-    -> plan and implementation
+    C -- "duplicate / not actionable" --> X["Closed with a reason"]
-    -> risk-based review and validation
+    C -- "small + reversible + fits ROADMAP" --> M["product: maintainer decision"]
+    C -- "big, new direction or hard to undo" --> N["product: needs decision"]
+    N --> D["3. Decision brief"]
+    D --> E{"4. Product Owner decides"}
+    E -- yes --> P["product: approved"]
+    E -- no --> R["product: declined<br/>closed as not planned"]
+    E -- "no answer in 14 days" --> L{"Only reason was size?"}
+    L -- yes --> M
+    L -- "no: new direction or hard to undo" --> N
+    M --> F["5. Implementation PRs<br/>Part of #N"]
+    P --> F
+    F --> G["6. Review and merge"]
+    G --> H["7. Close the issue"]
 ```
 
-Record the decision in an issue, proposal, work item, or PR so future maintainers can discover it. Do not rely on a chat-only decision.
+### Step 0 — Idea (optional)
+
+Half-formed ideas go to [Discussions › Ideas](https://github.com/odysseus-dev/odysseus/discussions/categories/ideas) to gauge interest. Nothing is decided there. A proposal is ready for step 1 when it is **actionable**.
+
+### Step 1 — Feature Request issue
+
+**Who:** the proposer.
+
+Open a [Feature Request](https://github.com/odysseus-dev/odysseus/issues/new?template=feature_request.yml) after checking [ROADMAP.md](../../ROADMAP.md) and open issues. Besides the required fields, the form asks three optional questions that make triage faster:
+
+- whether the proposer thinks it is big, a new direction, or hard to undo;
+- the yes/no questions they want the Product Owner to answer;
+- screenshots, a recording, or a demo branch.
+
+A proposal with several independent parts may be one issue, provided each part is listed separately so it can be decided separately.
+
+### Step 2 — Triage
+
+**Who:** a **triager**. **When:** within 7 days of the issue being opened.
+
+The triager does one of three things and records why in a comment:
+
+| Outcome | When | Label | Next |
+|---|---|---|---|
+| Close | **Duplicate** or not **actionable** | — (closed as duplicate or not planned, with a link) | — |
+| Maintainer decision | **Small** *and* **reversible** *and* **fits ROADMAP** | `product: maintainer decision` | Step 5 |
+| Needs the Product Owner | **Big**, *or* **new direction**, *or* **hard to undo** | `product: needs decision` | Step 3 |
+
+The triage comment names the ROADMAP line the proposal fits, or says that it fits none, and which of big / new direction / hard to undo applies. When the triager is unsure, the proposal goes to the Product Owner.
+
+### Step 3 — Decision brief
+
+**Who:** the proposer writes it; the triager checks it is complete before the clock starts.
+
+The brief goes at the top of the issue body (edit the issue, or post it as the first comment and link it from the body) and contains:
+
+1. **Summary:** what changes for users, in three to five lines.
+2. **Questions:** the exact yes/no questions, one per independent part.
+3. **Evidence:** screenshots or a recording for anything that renders; a demo branch if one exists.
+4. **What cannot be undone:** migrations, deleted data, changed contracts, or "nothing".
+5. **For big designs only:** architecture, migration and rollback, and the planned PR slices. A long design may live in a proposal document under `/design` and be linked from the brief.
+
+When the brief is complete, the triager mentions the Product Owner on the issue. This is **day 0** for the timeout.
+
+### Step 4 — Product Owner decision
+
+**Who:** the Product Owner. **When:** on the issue at any time, or at the next **queue review**.
+
+| Answer | Label | Recorded as |
+|---|---|---|
+| Yes, possibly with conditions | `product: approved` | A **recorded decision** listing any conditions |
+| No | `product: declined` | A recorded decision; the issue is closed as not planned |
+| A question | stays `product: needs decision` | The clock pauses until the proposer answers, then restarts at day 0 |
+| Different answers per part | the label for the approved part(s) | The recorded decision lists each part and its answer |
+
+A decision given outside GitHub (Matrix, a call) counts once a maintainer has recorded it on the issue. A declined feature can be maintained in a contributor's fork, and the issue can be reopened only with **materially new** information.
+
+#### When there is no answer: lazy consensus
+
+| Day | What happens |
+|---|---|
+| 0 | Brief complete; Product Owner mentioned on the issue |
+| 7 | One reminder from a maintainer, and the issue is put on the next queue review |
+| 14 | If there has been no **objection**: |
+
+- If the **only** reason the proposal needed the Product Owner was that it is **big** — it is **reversible** and **fits ROADMAP** — a maintainer may decide. They record "decided by lazy consensus on YYYY-MM-DD", apply `product: maintainer decision`, and own any rollback Pewds later asks for.
+- If it is a **new direction** or **hard to undo**, it keeps waiting. Silence never approves a one-way change or a change of direction. It is raised at each queue review.
+- A proposal still waiting after **90 days** is closed as not planned with a comment saying it expired without a decision. It can be reopened by anyone with the same brief.
+
+### Step 5 — Implementation
+
+**Who:** the proposer or any contributor.
+
+Feature PRs link the issue with `Part of #N` (or `Fixes #N` for the last slice) and stay within what was decided. A maintainer does not merge a feature PR whose issue carries neither `product: approved` nor `product: maintainer decision`, and checks any conditions in the recorded decision.
+
+If the work grows beyond what was decided — a new part, or a part that becomes hard to undo — that part goes back to step 2.
+
+A maintainer's own small, reversible feature that fits ROADMAP may skip the separate issue: the PR description states why it qualifies, and the PR is the record.
+
+### Step 6 — Review and merge
+
+Unchanged: [changes and review](changes-and-review.md). Product approval is not technical approval; high-risk changes still need an independent qualified review.
+
+### Step 7 — Close
+
+The issue is closed when the last PR lands. If the feature was a ROADMAP item, the same PR or a follow-up updates ROADMAP.md.
+
+## Labels
+
+| Label | Meaning | Applied by |
+|---|---|---|
+| `product: needs decision` | Waiting for the Product Owner | Triager |
+| `product: approved` | The Product Owner said yes (see the recorded decision for scope and conditions) | Product Owner or the maintainer recording the decision |
+| `product: maintainer decision` | Small, reversible and on ROADMAP, or decided by lazy consensus | Triager, or the maintainer applying lazy consensus |
+| `product: declined` | The Product Owner said no | Product Owner or the maintainer recording the decision |
+
+Only maintainers and the Product Owner set or change these labels. Exactly one of them is on a triaged feature issue at any time.
+
+## Definitions
+
+**Feature.** A change that adds or alters user-visible behavior, a setting, or a public contract (HTTP API, MCP tool, plugin interface, file format). Restoring documented or clearly intended behavior is a bug fix, not a feature.
+
+**Small.** All of the following are true:
+
+- it ships as one PR of ordinary size — as a guide, no more than about 600 changed lines excluding tests, generated files and lockfiles;
+- it touches one area of the app;
+- it adds an option or a capability without changing what existing users see or get by default;
+- it adds no new dependency, service, or background process.
+
+Anything else is **big**.
+
+**Reversible.** Reverting the PR(s) fully restores the previous state for every user, with nothing left behind: no data migrated, rewritten or deleted; no stored format that other code or users now depend on; no public contract that integrations have started to use; no security boundary moved. If any of those is not true, the change is **hard to undo**.
+
+**Hard to undo.** Not reversible. Typical cases: data migrations, deleting or purging data, new persisted formats, new or changed authentication, permission or sharing models, and removing or renaming a public API, MCP tool, setting or file location.
+
+**New direction.** The change takes Odysseus somewhere it has not committed to. Any of:
+
+- a new top-level area of the app (a new panel, tab or module);
+- multi-user or collaboration features where the app was single-user;
+- data or requests leaving the user's machine by default, including telemetry and new external services;
+- dropping or adding a supported platform, install method or model backend;
+- anything listed under "Not The Focus Right Now" in ROADMAP.md;
+- reversing a previous Product Owner decision.
+
+**Fits ROADMAP.** The triager can point to a specific ROADMAP.md item that the proposal implements or directly extends, and the proposal is not listed under "Not The Focus Right Now". "It is in the same area as an item" is not enough.
+
+**Actionable.** The issue states a concrete problem and a concrete change in behavior, specific enough to write acceptance criteria from.
+
+**Duplicate.** Same user-facing outcome as an open issue or open PR, or a request for a ROADMAP item that adds nothing to it. Close it with a link, and move any new information to the original. A proposal that offers to *implement* a ROADMAP item, with a concrete design, is not a duplicate: it is triaged normally and usually **fits ROADMAP**.
+
+**Triager.** Any maintainer other than the proposer. A maintainer's own proposal is triaged by another maintainer.
+
+**Objection.** A comment on the issue — or one recorded there from another channel — by the Product Owner or a maintainer, asking to stop, wait, or change the scope. A maintainer's objection stops lazy consensus for that proposal; it then waits for the Product Owner like a new direction does. Questions and suggestions that do not ask to stop are not objections.
+
+**Recorded decision.** A comment on the issue that states the decision, who made it, the date, where it was made if not on GitHub, the parts it covers, and any conditions — together with the matching label change.
+
+**Queue review.** Every 14 days, a maintainer brings the [`product: needs decision` list](https://github.com/odysseus-dev/odysseus/issues?q=is%3Aissue+is%3Aopen+label%3A%22product%3A+needs+decision%22) to the Product Owner and records each answer on its issue. The maintainer who runs it is agreed among maintainers and noted here: *to be decided*.
+
+**Days.** Calendar days, counted from the timestamp of the event on the issue (label applied, brief completed, reminder posted).
+
+**Materially new.** Information the original decision did not have: a different design, new evidence of demand, or a changed ROADMAP.
 
 ## Decision boundaries
 
@@ -45,4 +199,4 @@ The Product Owner decides **what belongs in Odysseus**. Maintainers decide **how
 
 A rejected or out-of-scope feature can be maintained independently in a contributor's fork.
 
-If the Product Owner is unavailable, work requiring prior product approval waits unless authority was explicitly delegated. Smaller decisions and urgent fixes follow the rules above.
+If the Product Owner is unavailable, work that needs product approval follows the timeout above. It never proceeds on silence when it is a new direction or hard to undo. Smaller decisions and urgent fixes follow the rules above.

docs/maintainers/triage.md

@@ -26,6 +26,8 @@ Record what is known and distinguish it from what remains unverified:
 - Priority, ownership, and relevant domain expertise.
 - Remaining unknowns, risks, and safe recovery or workaround when available.
 
+Feature requests follow the [feature proposal workflow](product-and-features.md#feature-proposal-workflow): triage within 7 days, and a product label on every triaged feature issue.
+
 Small and obvious fixes may go straight to a focused PR, consistent with [CONTRIBUTING.md](../../CONTRIBUTING.md) and [changes and review](changes-and-review.md).
 
 ## Practical priority language

frontend/src/features/notes/ui/cardInteractions.js

@@ -112,7 +112,7 @@ export function createCardInteractions({
     if (state.selectMode) {
       body.querySelectorAll('.note-card').forEach(card => {
         card.addEventListener('click', (e) => {
-          if (e.target.closest('.note-card-cb')) return; // checkbox handles itself
+          if (e.target.closest('.note-card-cb, a')) return; // links open without toggling selection
           e.stopPropagation();
           tapToEditOrSelect(card);
         });

frontend/src/features/notes/ui/editor.js

@@ -51,7 +51,7 @@ export function createEditor({
       <input type="hidden" class="note-form-due" value="${note?.due_date || ''}" />
       <input type="hidden" class="note-form-repeat" value="${note?.repeat || 'none'}" />
     </div>
-    ${currentImageUrl && type !== 'draw' ? `<div class="note-form-image-wrap"><img class="note-form-image" src="${_esc(currentImageUrl)}" draggable="false" /><button class="note-form-image-rm" title="Remove">&times;</button></div>` : ''}
+    ${currentImageUrl && type !== 'draw' ? `<div class="note-form-image-wrap"><img class="note-form-image" src="${_esc(currentImageUrl)}" alt="" draggable="false" /><button class="note-form-image-rm" title="Remove" aria-label="Remove">&times;</button></div>` : ''}
     <div class="note-form-body">
       ${type === 'note'
           ? `<textarea class="note-form-content" placeholder="Take a note..." rows="4">${_esc(note?.content || '')}</textarea>`
@@ -297,7 +297,7 @@ export function createEditor({
           form.querySelector('.note-form-image-wrap')?.remove();
           const wrap = document.createElement('div');
           wrap.className = 'note-form-image-wrap';
-          wrap.innerHTML = `<img class="note-form-image" draggable="false" /><button class="note-form-image-rm" title="Remove">&times;</button>`;
+          wrap.innerHTML = `<img class="note-form-image" alt="" draggable="false" /><button class="note-form-image-rm" title="Remove" aria-label="Remove">&times;</button>`;
           // Insert AFTER the whole header (a flex-row), not after the
           // title input itself : otherwise the image lands as a sibling
           // of the title inside the header and flex puts them side-by-side.

frontend/src/features/notes/ui/text.js

@@ -31,7 +31,7 @@ export function createText({
         url = url.slice(0, -1);
       }
       const href = url.startsWith('www.') ? `https://${url}` : url;
-      return `<a href="${_attrEsc(href)}" class="note-link" target="_blank" rel="noopener noreferrer" onclick="event.stopPropagation()">${url}</a>` + (url !== m ? m.slice(url.length) : '');
+      return `<a href="${_attrEsc(href)}" class="note-link" target="_blank" rel="noopener noreferrer">${url}</a>` + (url !== m ? m.slice(url.length) : '');
     });
   }
 

mcp_servers/email_server.py

@@ -33,6 +33,7 @@ from mcp.types import Tool, TextContent
 
 sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
 
+from src.email_mime import decode_text_part as _decode_text_part
 server = Server("email")
 EMAIL_SOCKET_TIMEOUT = float(os.environ.get("EMAIL_SOCKET_TIMEOUT", "20"))
 from src.constants import DATA_DIR as _DATA_DIR, APP_DB, EMAIL_CACHE_DB, SETTINGS_FILE as _SETTINGS_FILE, MAIL_ATTACHMENTS_DIR
@@ -923,29 +924,25 @@ def _extract_text(msg):
     """Extract plain text body from email message."""
     if msg.is_multipart():
         text_parts = []
+        html_parts = []
         for part in msg.walk():
             ct = part.get_content_type()
             cd = str(part.get("Content-Disposition", ""))
             if ct == "text/plain" and "attachment" not in cd:
-                payload = part.get_payload(decode=True)
+                text = _decode_text_part(part)
-                if payload:
+                if text:
-                    charset = part.get_content_charset() or "utf-8"
+                    text_parts.append(text)
-                    text_parts.append(payload.decode(charset, errors="replace"))
+            elif ct == "text/html" and not html_parts and "attachment" not in cd:
-            elif ct == "text/html" and not text_parts and "attachment" not in cd:
+                raw_html = _decode_text_part(part)
-                payload = part.get_payload(decode=True)
+                if raw_html:
-                if payload:
-                    charset = part.get_content_charset() or "utf-8"
-                    raw_html = payload.decode(charset, errors="replace")
                     text = re.sub(r"<br\s*/?>", "\n", raw_html, flags=re.I)
                     text = re.sub(r"<[^>]+>", "", text)
                     text = html.unescape(text)
-                    text_parts.append(text.strip())
+                    html_parts.append(text.strip())
-        return "\n".join(text_parts)
+        return "\n".join(text_parts or html_parts)
     else:
-        payload = msg.get_payload(decode=True)
+        text = _decode_text_part(msg)
-        if payload:
+        if text:
-            charset = msg.get_content_charset() or "utf-8"
-            text = payload.decode(charset, errors="replace")
             if msg.get_content_type() == "text/html":
                 text = re.sub(r"<br\s*/?>", "\n", text, flags=re.I)
                 text = re.sub(r"</(?:p|div|li|tr|h[1-6])\s*>", "\n", text, flags=re.I)

mcp_servers/image_gen_server.py

@@ -33,7 +33,7 @@ async def list_tools() -> list[Tool]:
                     "prompt": {"type": "string", "description": "Image description prompt"},
                     "model": {"type": "string", "description": "Model name (auto-detects if omitted)"},
                     "size": {"type": "string", "description": "Image size (default 1024x1024)"},
-                    "quality": {"type": "string", "description": "Quality: low, medium, high, auto (default medium)"},
+                    "quality": {"type": "string", "description": "Quality: low, medium, high, auto (default configured quality, otherwise medium)"},
                 },
                 "required": ["prompt"],
             },
@@ -66,7 +66,7 @@ async def call_tool(name: str, arguments: dict) -> list[TextContent]:
 
         if not model_spec:
             model_spec = _settings.get("image_model", "")
-        if quality == "medium" and _settings.get("image_quality"):
+        if "quality" not in arguments and _settings.get("image_quality"):
             quality = _settings["image_quality"]
 
         # Auto-detect best available image model

package.json

@@ -9,7 +9,11 @@
     "test:photo-editor": "playwright test --config tests/e2e/playwright.config.js",
     "test:photo-editor:install": "playwright install chromium firefox webkit",
     "test:photo-editor:firefox": "PHOTO_EDITOR_E2E_BROWSER=firefox playwright test --config tests/e2e/playwright.config.js",
-    "test:photo-editor:webkit": "PHOTO_EDITOR_E2E_BROWSER=webkit playwright test --config tests/e2e/playwright.config.js"
+    "test:photo-editor:webkit": "PHOTO_EDITOR_E2E_BROWSER=webkit playwright test --config tests/e2e/playwright.config.js",
+    "test:ios-layout": "playwright test --config tests/e2e/ios-layout.config.cjs",
+    "test:ios-layout:chromium": "ODYSSEUS_LAYOUT_BROWSER=chromium playwright test --config tests/e2e/ios-layout.config.cjs",
+    "test:ios-layout:webkit": "ODYSSEUS_LAYOUT_BROWSER=webkit playwright test --config tests/e2e/ios-layout.config.cjs",
+    "test:ios-layout:install": "playwright install chromium webkit"
   },
   "devDependencies": {
     "@antithesishq/bombadil": "^0.7.0",

requirements-dev.txt

@@ -2,3 +2,4 @@
 -r requirements.txt
 # psutil lets `-n auto` use physical cores instead of logical CPU threads.
 pytest-xdist[psutil]>=3.8,<4
+playwright==1.62.0

routes/auth_routes.py

@@ -13,11 +13,12 @@ import re
 from pathlib import Path
 
 from core.atomic_io import atomic_write_json, atomic_write_text
-from core.auth import AuthManager, RESERVED_USERNAMES, SetAdminResult, TOKEN_TTL
+from core.auth import AuthManager, RESERVED_USERNAMES, SetAdminResult, TOKEN_TTL, username_error
 from src.constants import DEEP_RESEARCH_DIR, MEMORY_FILE, PASSWORD_MIN_LENGTH, SKILLS_DIR
 from src.auth_helpers import _auth_disabled
 from src.rate_limiter import RateLimiter
 from src.settings_scrub import scrub_settings
+from src import tool_capabilities
 from src.settings import (
     load_settings as _load_settings,
     save_settings as _save_settings,
@@ -154,8 +155,9 @@ def setup_auth_routes(auth_manager: AuthManager) -> APIRouter:
             raise HTTPException(400, "Already configured")
         if len(body.password) < PASSWORD_MIN_LENGTH:
             raise HTTPException(400, f"Password must be at least {PASSWORD_MIN_LENGTH} characters")
-        if len(body.username.strip()) < 1:
+        name_error = username_error(body.username)
-            raise HTTPException(400, "Username is required")
+        if name_error:
+            raise HTTPException(400, name_error)
         if body.username.lower() in RESERVED_USERNAMES:
             raise HTTPException(403, "Username is reserved")
         ok = await asyncio.to_thread(auth_manager.setup, body.username, body.password)
@@ -174,8 +176,9 @@ def setup_auth_routes(auth_manager: AuthManager) -> APIRouter:
             raise HTTPException(403, "Registration is disabled. Ask an admin for an account.")
         if len(body.password) < PASSWORD_MIN_LENGTH:
             raise HTTPException(400, f"Password must be at least {PASSWORD_MIN_LENGTH} characters")
-        if len(body.username.strip()) < 1:
+        name_error = username_error(body.username)
-            raise HTTPException(400, "Username is required")
+        if name_error:
+            raise HTTPException(400, name_error)
         if body.username.lower() in RESERVED_USERNAMES:
             raise HTTPException(403, "Username is reserved")
         ok = await asyncio.to_thread(auth_manager.create_user, body.username, body.password, is_admin=False)
@@ -333,8 +336,9 @@ def setup_auth_routes(auth_manager: AuthManager) -> APIRouter:
             raise HTTPException(403, "Admin only")
         if len(body.password) < PASSWORD_MIN_LENGTH:
             raise HTTPException(400, f"Password must be at least {PASSWORD_MIN_LENGTH} characters")
-        if len(body.username.strip()) < 1:
+        name_error = username_error(body.username)
-            raise HTTPException(400, "Username is required")
+        if name_error:
+            raise HTTPException(400, name_error)
         if body.username.lower() in RESERVED_USERNAMES:
             raise HTTPException(403, "Username is reserved")
         ok = auth_manager.create_user(body.username, body.password, body.is_admin)
@@ -361,8 +365,16 @@ def setup_auth_routes(auth_manager: AuthManager) -> APIRouter:
         user = _get_current_user(request)
         if not user or not auth_manager.is_admin(user):
             raise HTTPException(403, "Admin only")
+        try:
             body = await request.json()
+        except ValueError:
+            raise HTTPException(400, "Request body must be a JSON object")
+        if not isinstance(body, dict):
+            raise HTTPException(400, "Request body must be a JSON object")
+        try:
             ok = auth_manager.set_privileges(username, body)
+        except ValueError as e:
+            raise HTTPException(400, str(e))
         if not ok:
             raise HTTPException(404, "User not found or is admin")
         return {"ok": True, "privileges": auth_manager.get_privileges(username)}
@@ -376,6 +388,9 @@ def setup_auth_routes(auth_manager: AuthManager) -> APIRouter:
         new_username = (body.username or "").strip().lower()
         if not new_username:
             raise HTTPException(400, "Username required")
+        name_error = username_error(new_username)
+        if name_error:
+            raise HTTPException(400, name_error)
         if old_username == new_username:
             return {"ok": True, "username": new_username, "renamed_self": old_username == user}
         if old_username not in auth_manager.users:
@@ -653,7 +668,18 @@ def setup_auth_routes(auth_manager: AuthManager) -> APIRouter:
         user = _get_current_user(request)
         if not user or not auth_manager.is_admin(user):
             raise HTTPException(403, "Admin only")
+        try:
             result = auth_manager.set_admin(username, body.is_admin, user)
+        finally:
+            # A demotion purges the user's ApiToken rows; dirty the bearer
+            # cache so already-cached tokens stop authenticating.
+            if not body.is_admin:
+                invalidator = getattr(
+                    getattr(getattr(request, "app", None), "state", None),
+                    "invalidate_token_cache", None,
+                )
+                if callable(invalidator):
+                    invalidator()
         if result is SetAdminResult.USER_NOT_FOUND:
             raise HTTPException(404, "User not found")
         if result is SetAdminResult.NOT_AUTHORIZED:
@@ -753,6 +779,11 @@ def setup_auth_routes(auth_manager: AuthManager) -> APIRouter:
         user = _get_current_user(request)
         settings = without_retired_settings(_load_settings())
         if _auth_disabled() or (user and auth_manager.is_admin(user)):
+            # Deployment-level truth for the Tool approval row: a falsy
+            # ODYSSEUS_TOOL_APPROVAL_GATE disables the approval gate for every
+            # run regardless of the stored tool_approval_mode, so surface it to
+            # admins instead of letting the row mislead them the other way.
+            settings["tool_approval_gate_env_disabled"] = not tool_capabilities.TOOL_APPROVAL_GATE_ENABLED
             return settings
         return scrub_settings(settings)
 
@@ -769,6 +800,9 @@ def setup_auth_routes(auth_manager: AuthManager) -> APIRouter:
             "agent_max_tool_calls": (0, 1000),  # 0 = unlimited
             "auto_compact_threshold_percent": (50, 95),
         }
+        _VALID_VALUES = {
+            "tool_approval_mode": {"ask", "off"},
+        }
         for key in DEFAULT_SETTINGS:
             if key in RETIRED_SETTING_KEYS:
                 continue
@@ -782,6 +816,15 @@ def setup_auth_routes(auth_manager: AuthManager) -> APIRouter:
                 except (TypeError, ValueError):
                     raise HTTPException(400, f"{key} must be an integer")
                 val = max(lo, min(val, hi))
+
+            if key in _VALID_VALUES:
+                allowed = _VALID_VALUES[key]
+                if not isinstance(val, str) or val not in allowed:
+                    raise HTTPException(
+                        400,
+                        f"{key} must be one of {sorted(allowed)}",
+                    )
+
             current[key] = val
         _save_settings(current)
         return without_retired_settings(current)
@@ -887,6 +930,8 @@ def setup_auth_routes(auth_manager: AuthManager) -> APIRouter:
         """Create a new integration (admin only)."""
         _require_admin_user(request)
         body = await request.json()
+        if not isinstance(body, dict):
+            raise HTTPException(400, "Request body must be a JSON object")
         item = add_integration(body)
         return {"ok": True, "integration": mask_integration_secret(item)}
 
@@ -895,6 +940,8 @@ def setup_auth_routes(auth_manager: AuthManager) -> APIRouter:
         """Update an existing integration (admin only)."""
         _require_admin_user(request)
         body = await request.json()
+        if not isinstance(body, dict):
+            raise HTTPException(400, "Request body must be a JSON object")
         item = update_integration(integration_id, body)
         if not item:
             raise HTTPException(404, "Integration not found")

routes/calendar_routes.py

@@ -80,6 +80,18 @@ def _require_user(request: Request) -> str:
     return FALLBACK_OWNER
 
 
+async def _json_object_body(request: Request) -> dict:
+    """Parse a JSON object body. An unreadable body counts as empty; a body
+    that is valid JSON but not an object (a list, string or number) is a 400."""
+    try:
+        body = await request.json()
+    except Exception:
+        return {}
+    if not isinstance(body, dict):
+        raise HTTPException(400, "Request body must be a JSON object")
+    return body
+
+
 def _get_or_404_calendar(db, cal_id: str, owner: str) -> CalendarCal:
     cal = db.query(CalendarCal).filter(CalendarCal.id == cal_id).first()
     if not cal:
@@ -994,10 +1006,7 @@ def setup_calendar_routes(upload_handler=None) -> APIRouter:
     async def save_config(request: Request):
         """Legacy single-account endpoint — upserts the first account."""
         owner = _require_user(request)
-        try:
+        body = await _json_object_body(request)
-            body = await request.json()
-        except Exception:
-            body = {}
         accounts = _get_caldav_accounts(owner)
         if not (body.get("url") or "").strip():
             _save_caldav_accounts(owner, [])
@@ -1052,10 +1061,7 @@ def setup_calendar_routes(upload_handler=None) -> APIRouter:
         """Add a new CalDAV account."""
         import uuid as _uuid
         owner = _require_user(request)
-        try:
+        body = await _json_object_body(request)
-            body = await request.json()
-        except Exception:
-            body = {}
         from src.caldav_sync import validate_caldav_url
         try:
             url = validate_caldav_url(body.get("url", ""))
@@ -1080,10 +1086,7 @@ def setup_calendar_routes(upload_handler=None) -> APIRouter:
     async def update_caldav_account(account_id: str, request: Request):
         """Update an existing CalDAV account by id."""
         owner = _require_user(request)
-        try:
+        body = await _json_object_body(request)
-            body = await request.json()
-        except Exception:
-            body = {}
         accounts = _get_caldav_accounts(owner)
         idx = next((i for i, a in enumerate(accounts) if a.get("id") == account_id), None)
         if idx is None:
@@ -1124,10 +1127,7 @@ def setup_calendar_routes(upload_handler=None) -> APIRouter:
         test an already-saved account. Falls back to the first saved account
         when nothing is provided."""
         owner = _require_user(request)
-        try:
+        body = await _json_object_body(request)
-            body = await request.json()
-        except Exception:
-            body = {}
         url = (body.get("url") or "").strip()
         user = (body.get("username") or "").strip()
         pw = body.get("password") or ""

routes/chat_helpers.py

@@ -696,11 +696,10 @@ async def auto_name_session(session_manager, sess):
             logger.info(f"Auto-named session {sess.id} deterministically (no authorized model endpoint): {title}")
             return
 
-        # max_tokens big enough that reasoning models (Minimax M2,
+        # A configured effort is the user's explicit choice. Otherwise thinking
-        # DeepSeek R1, QwQ, etc.) have headroom for <think>…</think>
+        # stays off, because reasoning would use up this small token budget and
-        # plus the actual title — 200 used to clip them mid-reasoning
+        # leave no title.
-        # so strip_think left an empty string and no rename happened.
+        title_effort = effort_for_call(t_url, t_model, owner, sess)
-        # Timeout matches: 60s gives slow local reasoners room to finish.
         try:
             title = await llm_call_async(
                 t_url,
@@ -713,7 +712,8 @@ async def auto_name_session(session_manager, sess):
                 max_tokens=64,
                 headers=t_headers,
                 timeout=15,
-                reasoning_effort=effort_for_call(t_url, t_model, owner, sess),
+                reasoning_effort=title_effort,
+                thinking_mode=None if title_effort else "off",
             )
         except Exception as e:
             logger.warning("[auto-name] non-spec title call failed for %s: %s", sess.id, e)
@@ -1276,6 +1276,7 @@ async def build_chat_context(
         message=_ctx_msg,
         session=sess,
         use_web=use_web_val and not skip_web,
+        use_rag=use_rag_val,
         use_memory=mem_enabled,
         time_filter=time_filter,
         preset_system_prompt=preset.system_prompt,
@@ -1286,8 +1287,6 @@ async def build_chat_context(
         incognito=incognito,
         use_skills=skills_enabled,
     )
-    if use_rag is not None or is_research_spinoff or casual_low_signal:
-        _preface_kwargs["use_rag"] = use_rag_val
     preface, rag_sources, web_sources = chat_processor.build_context_preface(**_preface_kwargs)
 
     # Capture used memories immediately

routes/chat_routes.py

@@ -1901,7 +1901,7 @@ def _ts_or_zero(value) -> float:
 
 
 def _recover_empty_session_model(sess, session_id: str, owner: str | None = None) -> bool:
-    """Re-populate sess.model from the matching endpoint's cached models.
+    """Recover sess.model from the endpoint's visible cached and pinned models.
 
     Covers the window between endpoint setup and the first chat send: the
     picker showed a model in the dropdown but the session record never got
@@ -1965,6 +1965,9 @@ def _recover_empty_session_model(sess, session_id: str, owner: str | None = None
                 visible = _visible_models(cached, getattr(ep, "hidden_models", None))
             except Exception:
                 visible = cached
+        if not is_chatgpt_subscription:
+            from src.endpoint_resolver import _endpoint_enabled_models
+            visible = _endpoint_enabled_models(ep)
         if current_model and current_model in {str(item).strip() for item in visible}:
             return False
         if is_chatgpt_subscription:
@@ -3923,6 +3926,19 @@ def setup_chat_routes(
             if ctx.preprocessed.attachment_meta:
                 yield f"data: {json.dumps({'type': 'attachments', 'data': ctx.preprocessed.attachment_meta})}\n\n"
 
+            # Give the page the saved user message's id. Only the reply's id
+            # was sent back, so a message sent in the open page had none and
+            # a later regenerate/edit of it could not cut the chat by id.
+            try:
+                _last_user_msg = next((m for m in reversed(getattr(sess, "history", None) or [])
+                                       if getattr(m, "role", None) == "user"), None)
+                _user_db_id = ((getattr(_last_user_msg, "metadata", None) or {}).get("_db_id")
+                               if _last_user_msg is not None else None)
+                if _user_db_id:
+                    yield f"data: {json.dumps({'type': 'user_message_saved', 'id': _user_db_id})}\n\n"
+            except Exception:
+                pass
+
             # Announce any docs auto-created during preprocess (e.g. fillable
             # PDF → editable markdown) so the editor pane switches to them
             # before the model starts streaming.

routes/cookbook/__init__.py

@@ -0,0 +1 @@
+"""Cookbook route domain package (#4071)."""

routes/cookbook/endpoint_registration.py

@@ -0,0 +1,284 @@
+"""Auto-register Cookbook-served models as model endpoints."""
+
+import json
+import logging
+import uuid
+
+from routes.cookbook_helpers import ServeRequest
+
+logger = logging.getLogger(__name__)
+
+
+def _auto_register_image_endpoint(
+    req: ServeRequest,
+    remote: str | None,
+    serve_session_id: str | None = None,
+) -> str | None:
+    """Register a diffusion model as an image endpoint so it appears in the model selector."""
+    import re
+    from core.database import SessionLocal, ModelEndpoint
+    from routes.model_routes import _normalize_model_ids
+    from src.settings import load_settings, save_settings
+
+    # Parse port from command (--port NNNN), default 8100 for diffusion_server
+    port_match = re.search(r'--port\s+(\d+)', req.cmd)
+    port = int(port_match.group(1)) if port_match else 8100
+
+    # Determine host
+    if remote:
+        # SSH alias — use as hostname (Tailscale resolves it later)
+        host = remote.split("@")[-1] if "@" in remote else remote
+    else:
+        host = "localhost"
+
+    base_url = f"http://{host}:{port}/v1"
+
+    # Friendly display name from repo_id
+    short_name = req.repo_id.split("/")[-1] if "/" in req.repo_id else req.repo_id
+    display_name = f"{short_name} (image)"
+    pinned_models = [req.repo_id] if req.repo_id else []
+
+    db = SessionLocal()
+    try:
+        # Check for existing endpoint with same base_url — update it
+        existing = db.query(ModelEndpoint).filter(ModelEndpoint.base_url == base_url).first()
+        if existing:
+            existing.is_enabled = True
+            existing.model_type = "image"
+            # Follow the newly served model unless the user renamed the endpoint.
+            pinned_names = {f"{m.split('/')[-1]} (image)" for m in _normalize_model_ids(existing.pinned_models)}
+            if existing.name in pinned_names:
+                existing.name = display_name
+            existing.endpoint_kind = "local"
+            existing.model_refresh_mode = "manual"
+            existing.cookbook_serve_session_id = serve_session_id
+            if pinned_models:
+                existing.cached_models = json.dumps(pinned_models)
+                existing.pinned_models = json.dumps(pinned_models)
+            db.commit()
+            settings = load_settings()
+            if settings.get("image_gen_enabled") is not True:
+                settings["image_gen_enabled"] = True
+                save_settings(settings)
+            logger.info(f"Updated existing image endpoint: {base_url}")
+            return existing.id
+
+        ep_id = f"img-{uuid.uuid4().hex[:8]}"
+        ep = ModelEndpoint(
+            id=ep_id,
+            name=display_name,
+            base_url=base_url,
+            api_key=None,
+            is_enabled=True,
+            model_type="image",
+            endpoint_kind="local",
+            model_refresh_mode="manual",
+            cached_models=json.dumps(pinned_models) if pinned_models else None,
+            pinned_models=json.dumps(pinned_models) if pinned_models else None,
+            cookbook_serve_session_id=serve_session_id,
+        )
+        db.add(ep)
+        db.commit()
+        settings = load_settings()
+        settings["image_gen_enabled"] = True
+        if not settings.get("image_model"):
+            settings["image_model"] = req.repo_id
+        save_settings(settings)
+        logger.info(f"Auto-registered image endpoint: {display_name} @ {base_url}")
+        return ep_id
+    except Exception as e:
+        logger.error(f"Failed to auto-register image endpoint: {e}")
+        db.rollback()
+        return None
+    finally:
+        db.close()
+
+
+def _auto_register_llm_endpoint(
+    req: ServeRequest,
+    remote: str | None,
+    serve_session_id: str | None = None,
+) -> str | None:
+    """Register a freshly-served LLM as a model endpoint so it appears in the
+    model picker without a manual /setup step — the text-model sibling of
+    _auto_register_image_endpoint.
+
+    Cookbook serve commands launch an OpenAI-compatible server (llama.cpp's
+    llama-server, vLLM, SGLang, or Ollama) on a known port. We point an
+    endpoint at that server's /v1; the picker auto-discovers the model id by
+    probing /v1/models and dims the endpoint until the server is reachable,
+    so registering immediately (before the server finishes loading) is safe.
+    """
+    logger.info(
+        f"_auto_register_llm_endpoint: ENTRY repo_id={req.repo_id!r} "
+        f"remote={remote!r} cmd_prefix={req.cmd[:80]!r}"
+    )
+    import re
+    from core.database import SessionLocal, ModelEndpoint
+    from routes.model_routes import _normalize_model_ids
+
+    # Port: ordered fallbacks so we match whatever the user actually
+    # asked for, not a hardcoded default:
+    #   1. explicit `--port N`  (vllm / sglang / llama-server)
+    #   2. `OLLAMA_HOST=host:port`  (the way Ollama specifies its bind)
+    #   3. fallback by backend (11434 ollama / 8080 llama.cpp)
+    # Previously the OLLAMA_HOST form was silently ignored and we
+    # registered every Ollama endpoint at 11434 — even if the user
+    # set OLLAMA_HOST=0.0.0.0:11435 to avoid colliding with an
+    # existing systemd Ollama, the registered endpoint pointed at
+    # the OLD port and showed as offline.
+    port_match = re.search(r'--port\s+(\d+)', req.cmd)
+    # Host length is capped at the longest DNS name; unbounded, repeated
+    # assignments make this scan quadratic and stall the event loop.
+    ollama_host_match = re.search(r'OLLAMA_HOST=[^\s]{0,253}?:(\d+)', req.cmd)
+    if port_match:
+        port = int(port_match.group(1))
+    elif ollama_host_match:
+        port = int(ollama_host_match.group(1))
+    elif "ollama" in req.cmd:
+        port = 11434
+    else:
+        port = 8080  # llama.cpp's llama-server default — the Apple Silicon path
+
+    # Determine host. The cookbook tmux for `local=true` serves runs INSIDE
+    # the odysseus container — so the right URL for the in-container
+    # backend to reach it is `localhost`, NOT `host.docker.internal`
+    # (the latter points at the docker HOST, which doesn't have a server
+    # on that port). The previous host.docker.internal fallback only made
+    # sense for /setup-added external services like systemd Ollama on the
+    # host — and those go through manual setup, not this auto-register
+    # code path. For remote serves we still use the SSH host alias.
+    if remote:
+        host = remote.split("@")[-1] if "@" in remote else remote
+    elif re.search(r"\bdocker\s+exec\s+(?:ollama-rocm|ollama-test)\b", req.cmd or ""):
+        host = "host.docker.internal"
+    else:
+        host = "localhost"
+
+    base_url = f"http://{host}:{port}/v1"
+
+    short_name = req.repo_id.split("/")[-1] if "/" in req.repo_id else req.repo_id
+    display_name = short_name or "Local model"
+    is_mlx_deepseek_v4 = (
+        "mlx_lm.server" in (req.cmd or "")
+        and "deepseek-v4" in ((req.repo_id or "") + " " + (req.cmd or "")).lower()
+    )
+    mlx_shim_model_id = ""
+    if is_mlx_deepseek_v4 and short_name:
+        home_match = re.search(r"((?:/Users|/home)/[^/\s'\"]+)", req.cmd or "")
+        remote_home = home_match.group(1) if home_match else ""
+        if remote_home:
+            mlx_shim_model_id = f"{remote_home}/.cache/odysseus/mlx-shims/{short_name}"
+
+    # If the serve command opts models into OpenAI tool-calling, record it so
+    # agent_loop trusts emitted tool_calls instead of the name heuristic.
+    is_ollama_endpoint = "ollama" in (req.cmd or "").lower()
+    supports_tools = True if "--enable-auto-tool-choice" in req.cmd else None
+    # Pin the model the user launched for every Cookbook-created LLM
+    # endpoint, not just Ollama. Some OpenAI-compatible servers report a
+    # deployment alias from /v1/models, and a stale server can answer on the
+    # same port while the new launch failed. Keeping the requested model id
+    # pinned makes the picker reflect the actual launch intent.
+    launched_model = mlx_shim_model_id or req.repo_id
+    pinned_models = [launched_model] if launched_model else []
+
+    # Launch registration must not delete other endpoints.
+    # Matching names, model pins, and failed probes do not prove that
+    # another serve has stopped. Lifecycle cleanup owns termination.
+    db = SessionLocal()
+    try:
+        # Reuse an endpoint already pointed at this URL instead of duplicating.
+        existing = db.query(ModelEndpoint).filter(ModelEndpoint.base_url == base_url).first()
+        if existing:
+            existing.is_enabled = True
+            existing.model_type = "llm"
+            # Follow the newly served model unless the user renamed the endpoint.
+            pinned_names = {m.split("/")[-1] for m in _normalize_model_ids(existing.pinned_models)}
+            if existing.name in pinned_names:
+                existing.name = display_name
+            existing.endpoint_kind = "local"
+            existing.model_refresh_mode = "auto"
+            # A reused endpoint belongs to the newest serve session.
+            existing.cookbook_serve_session_id = serve_session_id
+            if launched_model:
+                # Pins represent model selection, not launch history.
+                existing_pins = _normalize_model_ids(
+                    existing.pinned_models
+                )
+                if launched_model not in existing_pins:
+                    existing_pins.append(launched_model)
+                existing.pinned_models = json.dumps(existing_pins)
+            if is_ollama_endpoint:
+                existing.endpoint_kind = "ollama"
+                if pinned_models:
+                    existing.cached_models = json.dumps(pinned_models)
+            if supports_tools is not None:
+                existing.supports_tools = supports_tools
+            db.commit()
+            logger.info(f"Updated existing local model endpoint: {base_url}")
+            # Re-probe so cached_models matches what the server actually
+            # serves right now (the URL may have stayed the same but the
+            # model behind it changed across launches).
+            try:
+                if mlx_shim_model_id:
+                    existing.cached_models = json.dumps([mlx_shim_model_id])
+                    existing.pinned_models = json.dumps([mlx_shim_model_id])
+                    db.commit()
+                else:
+                    from routes.model_routes import _probe_endpoint
+                    import json as _json2
+                    probed = _probe_endpoint(base_url, existing.api_key, timeout=5)
+                    if probed:
+                        existing.cached_models = _json2.dumps(probed)
+                        db.commit()
+            except Exception as _pe:
+                logger.warning(f"Re-probe failed for {base_url}: {_pe!r}")
+            return existing.id
+
+        ep_id = f"local-{uuid.uuid4().hex[:8]}"
+        ep = ModelEndpoint(
+            id=ep_id,
+            name=display_name,
+            base_url=base_url,
+            api_key=None,
+            is_enabled=True,
+            model_type="llm",
+            endpoint_kind="ollama" if is_ollama_endpoint else "local",
+            model_refresh_mode="auto",
+            cached_models=json.dumps(pinned_models) if pinned_models else None,
+            pinned_models=json.dumps(pinned_models) if pinned_models else None,
+            supports_tools=supports_tools,
+            cookbook_serve_session_id=serve_session_id,
+        )
+        db.add(ep)
+        db.commit()
+        logger.info(f"Auto-registered local model endpoint: {display_name} @ {base_url}")
+        # Probe /v1/models NOW and write cached_models so the chat
+        # picker actually shows the model on the next /api/models
+        # call. Without this immediate probe, the endpoint has empty
+        # cached_models until the next background refresh fires (up
+        # to a minute later) and the picker shows nothing — even
+        # though the endpoint is in the DB and the server is up.
+        try:
+            if mlx_shim_model_id:
+                ep.cached_models = json.dumps([mlx_shim_model_id])
+                ep.pinned_models = json.dumps([mlx_shim_model_id])
+                db.commit()
+                logger.info(f"Auto-register: pinned MLX DeepSeek-V4 shim model @ {base_url}")
+            else:
+                from routes.model_routes import _probe_endpoint
+                import json as _json2
+                probed = _probe_endpoint(base_url, None, timeout=5)
+                if probed:
+                    ep.cached_models = _json2.dumps(probed)
+                    db.commit()
+                    logger.info(f"Auto-register: probed {len(probed)} models @ {base_url}")
+        except Exception as _pe:
+            logger.warning(f"Auto-register: probe-after-create failed for {base_url}: {_pe!r}")
+        return ep_id
+    except Exception as e:
+        logger.error(f"Failed to auto-register local model endpoint: {e}")
+        db.rollback()
+        return None
+    finally:
+        db.close()

routes/cookbook_helpers.py

@@ -14,6 +14,7 @@ from fastapi import HTTPException
 from pydantic import BaseModel
 
 from routes._validators import validate_remote_host, validate_ssh_port
+from routes.cookbook_output import HF_PARTIAL_BLOB_HELPER
 from core.platform_compat import _ssh_exec_argv
 
 logger = logging.getLogger(__name__)
@@ -378,6 +379,7 @@ def _cached_model_scan_script(model_dirs: list[str] | None = None, add_hf_cache:
     """
     lines = [
         "import json, os, re, shutil, subprocess, urllib.request",
+        *HF_PARTIAL_BLOB_HELPER.splitlines(),
         "models = []",
         "seen = set()",
         "BLOCKED_ROOTS = ('/sys', '/proc', '/dev', '/run', '/var/run')",
@@ -436,8 +438,10 @@ def _cached_model_scan_script(model_dirs: list[str] | None = None, add_hf_cache:
         "        sz, nf, ic = 0, 0, False",
         "        if os.path.isdir(blobs):",
         "            for f in os.scandir(blobs):",
+        "                if f.name.endswith('.incomplete'):",
+        "                    if not is_partial_blob(f.path): continue",
+        "                    ic = True",
         "                if f.is_file(): nf += 1; sz += f.stat().st_size",
-        "                if f.name.endswith('.incomplete'): ic = True",
         "        snap = os.path.join(cache, d, 'snapshots')",
         "        def snapshot_size():",
         "            total, count, incomplete = 0, 0, False",
@@ -724,8 +728,13 @@ def _check_serve_binary(seg: str) -> None:
     if not tokens:
         return
     env_re = re.compile(r"^[A-Za-z_][A-Za-z0-9_]*=")
-    first = next((t for t in tokens if not env_re.match(t)), "")
+    idx = next((i for i, t in enumerate(tokens) if not env_re.match(t)), len(tokens))
+    first = tokens[idx] if idx < len(tokens) else ""
     base = os.path.basename(first)
+    # The vLLM/SGLang Dependencies recipes install with `uv pip install …
+    # --torch-backend auto`. Allow uv for that subcommand only, not `uv run`.
+    if base == "uv" and tokens[idx + 1:idx + 3] == ["pip", "install"]:
+        return
     if base not in _SERVE_CMD_ALLOWLIST:
         raise HTTPException(
             400,
@@ -1030,6 +1039,166 @@ def _append_llama_cpp_linux_accel_build_lines(runner_lines: list[str]) -> None:
     runner_lines.append('  fi  # end _odysseus_have_prebuilt guard')
 
 
+ODYSSEUS_LLAMA_SHIM_MARKER = "# Auto-generated by Odysseus Cookbook: a `llama-server` lookalike"
+
+
+def _append_llama_server_resolution_definition_lines(runner_lines: list[str]) -> None:
+    """Define the helper shell function that discovers, reconciles and prioritizes llama-server."""
+    marker = f"'{_bash_squote(ODYSSEUS_LLAMA_SHIM_MARKER)}'"
+    runner_lines.append('_odysseus_resolve_llama_server() {')
+    runner_lines.append('  _odysseus_llama_bin=""')
+    runner_lines.append(f'  local _marker={marker}')
+    runner_lines.append('  local _saved_ifs="$IFS"')
+    runner_lines.append("  IFS=':'")
+    runner_lines.append('  local -a _parts')
+    runner_lines.append('  read -r -a _parts <<< "$PATH"')
+    runner_lines.append('  IFS="$_saved_ifs"')
+    runner_lines.append('  local _dir _cand')
+    runner_lines.append('  for _dir in "${_parts[@]}"; do')
+    runner_lines.append('    [ -n "$_dir" ] || continue')
+    runner_lines.append('    _cand="$_dir/llama-server"')
+    runner_lines.append('    [ -x "$_cand" ] || continue')
+    runner_lines.append('    if [ -f "$_cand" ] && [ ! -L "$_cand" ] && grep -Fxq "$_marker" "$_cand" 2>/dev/null; then')
+    runner_lines.append('      continue')
+    runner_lines.append('    fi')
+    runner_lines.append('    if "$_cand" --version >/dev/null 2>&1; then')
+    runner_lines.append('      _odysseus_llama_bin="$_cand"')
+    runner_lines.append('      break')
+    runner_lines.append('    fi')
+    runner_lines.append('  done')
+    runner_lines.append('  if [ -f "$HOME/bin/llama-server" ] && [ ! -L "$HOME/bin/llama-server" ] && grep -Fxq "$_marker" "$HOME/bin/llama-server" 2>/dev/null; then')
+    runner_lines.append('    if [ -n "$_odysseus_llama_bin" ]; then')
+    runner_lines.append('      echo "[odysseus] Native llama-server detected — retiring Odysseus compatibility shim in ~/bin"')
+    runner_lines.append('      rm -f "$HOME/bin/llama-server"')
+    runner_lines.append('      hash -r')
+    runner_lines.append('    elif python3 -c "import llama_cpp.server" 2>/dev/null; then')
+    runner_lines.append('      _odysseus_llama_bin="$HOME/bin/llama-server"')
+    runner_lines.append('    else')
+    runner_lines.append('      echo "[odysseus] Removing stale Odysseus llama-server shim (python3 -m llama_cpp.server unavailable)"')
+    runner_lines.append('      rm -f "$HOME/bin/llama-server"')
+    runner_lines.append('      hash -r')
+    runner_lines.append('    fi')
+    runner_lines.append('  fi')
+    runner_lines.append('  if [ -n "$_odysseus_llama_bin" ]; then')
+    runner_lines.append('    local _cand_dir')
+    runner_lines.append('    _cand_dir="$(dirname "$_odysseus_llama_bin")"')
+    runner_lines.append('    if [ "${PATH%%:*}" != "$_cand_dir" ]; then')
+    runner_lines.append('      export PATH="$_cand_dir:$PATH"')
+    runner_lines.append('    fi')
+    runner_lines.append('    hash -r')
+    runner_lines.append('    return 0')
+    runner_lines.append('  fi')
+    runner_lines.append('  return 1')
+    runner_lines.append('}')
+
+
+def _append_llama_server_shim_definition_lines(runner_lines: list[str]) -> None:
+    """Define the helper shell function that generates the Odysseus compatibility shim."""
+    _append_llama_server_resolution_definition_lines(runner_lines)
+    marker = f"'{_bash_squote(ODYSSEUS_LLAMA_SHIM_MARKER)}'"
+    runner_lines.append('_odysseus_create_llama_shim() {')
+    runner_lines.append('  local _target="$HOME/bin/llama-server"')
+    runner_lines.append(f'  local _marker={marker}')
+    runner_lines.append('  if [ -e "$_target" ] || [ -L "$_target" ]; then')
+    runner_lines.append('    if [ -f "$_target" ] && [ ! -L "$_target" ] && grep -Fxq "$_marker" "$_target" 2>/dev/null; then')
+    runner_lines.append('      : # existing Odysseus shim — safe to replace')
+    runner_lines.append('    else')
+    runner_lines.append('      echo "[odysseus] WARNING: $_target already exists and is not an Odysseus-managed shim; leaving untouched." >&2')
+    runner_lines.append('      return 1')
+    runner_lines.append('    fi')
+    runner_lines.append('  fi')
+    runner_lines.append('  mkdir -p "$HOME/bin" || return 1')
+    runner_lines.append('  local _tmp')
+    runner_lines.append('  _tmp="$(mktemp "$HOME/bin/.llama-server.odysseus.tmp.XXXXXX")" || {')
+    runner_lines.append('    echo "[odysseus] ERROR: unable to create secure temporary llama-server shim in $HOME/bin" >&2')
+    runner_lines.append('    return 1')
+    runner_lines.append('  }')
+    runner_lines.append("  if ! cat > \"$_tmp\" <<'_ODY_LLAMA_SHIM_EOF'")
+    runner_lines.append('#!/usr/bin/env bash')
+    runner_lines.append(ODYSSEUS_LLAMA_SHIM_MARKER)
+    runner_lines.append('# that translates the native CLI to `python -m llama_cpp.server`.')
+    runner_lines.append('# Lets cookbook-generated launch commands run unchanged on hosts')
+    runner_lines.append('# where only the pip llama-cpp-python package is installed.')
+    runner_lines.append('ARGS=()')
+    runner_lines.append('while [ $# -gt 0 ]; do')
+    runner_lines.append('  case "$1" in')
+    runner_lines.append('    -ngl|--gpu-layers|--n-gpu-layers) ARGS+=(--n_gpu_layers "$2"); shift 2 ;;')
+    runner_lines.append('    -c|--ctx-size) ARGS+=(--n_ctx "$2"); shift 2 ;;')
+    runner_lines.append('    -b|--batch-size) ARGS+=(--n_batch "$2"); shift 2 ;;')
+    runner_lines.append('    -ub|--ubatch-size) shift 2 ;;  # llama-cpp-python has no separate ubatch')
+    runner_lines.append('    --flash-attn) ARGS+=(--flash_attn true); shift 2 ;;')
+    runner_lines.append('    --cache-type-k) ARGS+=(--type_k "$2"); shift 2 ;;')
+    runner_lines.append('    --cache-type-v) ARGS+=(--type_v "$2"); shift 2 ;;')
+    runner_lines.append('    --n-cpu-moe) ARGS+=(--n_cpu_moe "$2"); shift 2 ;;')
+    runner_lines.append('    --mmproj) ARGS+=(--clip_model_path "$2"); shift 2 ;;')
+    runner_lines.append('    --image-max-tokens) shift 2 ;;  # native-only')
+    runner_lines.append('    --no-mmap) ARGS+=(--no_mmap true); shift ;;')
+    runner_lines.append('    --no-warmup) shift ;;  # native-only')
+    runner_lines.append('    --chat-template) ARGS+=(--chat_format "$2"); shift 2 ;;')
+    runner_lines.append('    --fit|--split-mode|--tensor-split|--main-gpu|--parallel) shift 2 ;;  # native-only')
+    runner_lines.append('    --mlock) ARGS+=(--use_mlock true); shift ;;')
+    runner_lines.append('    *) ARGS+=("$1"); shift ;;')
+    runner_lines.append('  esac')
+    runner_lines.append('done')
+    runner_lines.append('exec python3 -m llama_cpp.server "${ARGS[@]}"')
+    runner_lines.append('_ODY_LLAMA_SHIM_EOF')
+    runner_lines.append('  then')
+    runner_lines.append('    rm -f -- "$_tmp"')
+    runner_lines.append('    return 1')
+    runner_lines.append('  fi')
+    runner_lines.append('  chmod +x "$_tmp" || { rm -f -- "$_tmp"; return 1; }')
+    runner_lines.append('  mv -f -- "$_tmp" "$_target" || { rm -f -- "$_tmp"; return 1; }')
+    runner_lines.append('  hash -r')
+    runner_lines.append('  echo "[odysseus] Created llama-server shim → python -m llama_cpp.server (no native binary needed)"')
+    runner_lines.append('}')
+
+
+def _append_llama_server_shim_reconciliation_lines(runner_lines: list[str]) -> None:
+    """Reconcile an existing ~/bin/llama-server if it is owned by Odysseus Cookbook.
+
+    Invariants:
+    - Never touch an arbitrary user-owned executable or symlink.
+    - If ~/bin/llama-server is an Odysseus generated shim:
+        - If a native llama-server is present and healthy elsewhere on PATH, remove the shim
+          so the native binary is used.
+        - If python3 -m llama_cpp.server is unavailable, remove the stale shim
+          and hash -r so the fallback/build path is triggered.
+        - If no healthy native server exists and llama_cpp.server is working, preserve it.
+    - If a healthy native candidate exists on PATH (even later in PATH), prioritize it and
+      ensure the launched executable is the validated candidate.
+    """
+    runner_lines.append('[ -r ~/.config/odysseus-llama-cpp-env ] && . ~/.config/odysseus-llama-cpp-env')
+    runner_lines.append('if ! declare -F _odysseus_resolve_llama_server >/dev/null 2>&1; then')
+    _append_llama_server_resolution_definition_lines(runner_lines)
+    runner_lines.append('fi')
+    runner_lines.append('_odysseus_resolve_llama_server || true')
+
+
+def _append_llama_server_finalization_lines(runner_lines: list[str], platform_hint: str = "") -> None:
+    """Reconcile post-install launcher availability and enforce fail-closed invariant.
+
+    Guarantees that if python3 -m llama_cpp.server is usable but the `llama-server`
+    command is unresolvable, the compatibility launcher is created. If no usable
+    launcher exists after all build/install attempts, sets ODYSSEUS_PREFLIGHT_EXIT=127.
+    """
+    err_msg = (
+        f"ERROR: llama.cpp serving is not available on {platform_hint} after install attempts."
+        if platform_hint
+        else "ERROR: llama.cpp serving is not available after install/build attempts."
+    )
+    runner_lines.append('if ! declare -F _odysseus_resolve_llama_server >/dev/null 2>&1; then')
+    _append_llama_server_resolution_definition_lines(runner_lines)
+    runner_lines.append('fi')
+    runner_lines.append('if [ -z "$_odysseus_llama_bin" ] && python3 -c "import llama_cpp.server" 2>/dev/null; then')
+    runner_lines.append('  _odysseus_create_llama_shim')
+    runner_lines.append('  _odysseus_resolve_llama_server || true')
+    runner_lines.append('fi')
+    runner_lines.append('if [ -z "$_odysseus_llama_bin" ]; then')
+    runner_lines.append(f'  echo "{err_msg}"')
+    runner_lines.append('  ODYSSEUS_PREFLIGHT_EXIT=127')
+    runner_lines.append('fi')
+
+
 def _llama_cpp_rebuild_cmd(update_source: bool = False) -> str:
     """Shell command that clears the Cookbook-managed llama.cpp build.
 
@@ -1327,6 +1496,20 @@ def _diagnose_serve_output(text: str) -> dict | None:
             "Port is already in use.",
             [{"label": "retry on port 8001", "op": "replace", "flag": "--port", "value": "8001"}],
         ),
+        (
+            # PyPI torch/vLLM wheels now target CUDA 13, which needs driver 580+.
+            # torch: "The NVIDIA driver on your system is too old (found version 12080)";
+            # vLLM kernels: "CUDA driver version is insufficient for CUDA runtime version".
+            r"NVIDIA driver on your system is too old|CUDA driver version is insufficient",
+            "The installed PyTorch/vLLM build needs a newer NVIDIA driver than this server has.",
+            [
+                {
+                    "label": "install builds for the driver's CUDA version, e.g. vLLM's +cu129 wheel from its GitHub release with uv pip install … --torch-backend cu129",
+                    "op": "manual",
+                },
+                {"label": "or update the NVIDIA driver (CUDA 13 builds need driver 580 or newer)", "op": "manual"},
+            ],
+        ),
         (
             r"No CUDA GPUs are available|no GPU.*found|CUDA_VISIBLE_DEVICES.*invalid",
             "No GPUs are visible to the serve process.",

routes/cookbook_output.py

@@ -8,6 +8,20 @@ import re
 
 _FETCHING_ZERO_FILES_RE = re.compile(r"Fetching\s+0\s+files", re.IGNORECASE)
 
+# Shared source embedded in local/remote inventory and status probes. A blank
+# retry marker beside its completed, nonempty content-addressed blob is stale;
+# a blank marker alone may be a download that has not received its first byte.
+HF_PARTIAL_BLOB_HELPER = """def is_partial_blob(path):
+    if not path.endswith('.incomplete'):
+        return False
+    try:
+        completed = path[:-len('.incomplete')]
+        return (os.path.getsize(path) != 0 or not os.path.isfile(completed)
+                or os.path.getsize(completed) == 0)
+    except OSError:
+        return True
+"""
+
 # Probe scripts for the dead-session download check, run as
 # `python3 -c <PROBE> <repo_id> <cache_root>` (locally or over SSH).
 # cache_root is the task's custom download dir, '' for the default HF cache.
@@ -15,7 +29,7 @@ _FETCHING_ZERO_FILES_RE = re.compile(r"Fetching\s+0\s+files", re.IGNORECASE)
 # HF_HOME=<local_dir>, so that task's cache lives under <local_dir>/hub, and
 # the probe process's own environment knows nothing about it.
 HF_CACHE_COMPLETE_PROBE = (
-    "import os,sys;"
+    "import os,sys\n" + HF_PARTIAL_BLOB_HELPER +
     "repo=sys.argv[1];"
     "root=os.path.expanduser(sys.argv[2]) if len(sys.argv)>2 and sys.argv[2] else '';"
     "base=os.path.join(root,'hub') if root else (os.environ.get('HUGGINGFACE_HUB_CACHE') or os.path.join(os.environ.get('HF_HOME', os.path.expanduser('~/.cache/huggingface')), 'hub'));"
@@ -24,18 +38,18 @@ HF_CACHE_COMPLETE_PROBE = (
     "ok=os.path.isdir(snap) and any(os.path.isdir(os.path.join(snap,x)) and os.listdir(os.path.join(snap,x)) for x in os.listdir(snap));"
     "inc=False;"
     "blobs=os.path.join(d,'blobs');"
-    "inc=os.path.isdir(blobs) and any(x.endswith('.incomplete') for x in os.listdir(blobs));"
+    "inc=os.path.isdir(blobs) and any(is_partial_blob(os.path.join(blobs,x)) for x in os.listdir(blobs));"
     "sys.exit(0 if ok and not inc else 1)"
 )
 
 HF_CACHE_INCOMPLETE_PROBE = (
-    "import os,sys;"
+    "import os,sys\n" + HF_PARTIAL_BLOB_HELPER +
     "repo=sys.argv[1];"
     "root=os.path.expanduser(sys.argv[2]) if len(sys.argv)>2 and sys.argv[2] else '';"
     "base=os.path.join(root,'hub') if root else (os.environ.get('HUGGINGFACE_HUB_CACHE') or os.path.join(os.environ.get('HF_HOME', os.path.expanduser('~/.cache/huggingface')), 'hub'));"
     "d=os.path.join(base,'models--'+repo.replace('/','--'));"
     "blobs=os.path.join(d,'blobs');"
-    "inc=os.path.isdir(blobs) and any(x.endswith('.incomplete') for x in os.listdir(blobs));"
+    "inc=os.path.isdir(blobs) and any(is_partial_blob(os.path.join(blobs,x)) for x in os.listdir(blobs));"
     "sys.exit(0 if inc else 1)"
 )
 

routes/cookbook_routes.py

@@ -43,6 +43,9 @@ from routes.cookbook_output import (
     error_aware_output_tail, classify_dead_download,
     HF_CACHE_COMPLETE_PROBE, HF_CACHE_INCOMPLETE_PROBE,
 )
+from routes.cookbook.endpoint_registration import (
+    _auto_register_image_endpoint, _auto_register_llm_endpoint,
+)
 
 logger = logging.getLogger(__name__)
 
@@ -57,9 +60,11 @@ from routes.cookbook_helpers import (
     _pip_install_no_cache, _user_shell_path_bootstrap, _venv_safe_local_pip_install_cmd,
     _diagnose_serve_output, run_ssh_command_async,
     _ollama_bind_from_cmd, _pip_install_fallback_chain, _pip_install_no_cache,
-    _user_shell_path_bootstrap, _venv_safe_local_pip_install_cmd,
     _append_pip_install_runner_lines, _pip_install_command_without_break_system_packages,
     _normalize_llama_cpp_python_cache_types,
+    _append_llama_server_shim_definition_lines,
+    _append_llama_server_shim_reconciliation_lines,
+    _append_llama_server_finalization_lines,
     ModelDownloadRequest, ServeRequest,
 )
 
@@ -466,175 +471,6 @@ def setup_cookbook_routes() -> APIRouter:
                     task["payload"].pop("hf_token", None)
         return state
 
-    def _diagnose_serve_output(text: str) -> dict | None:
-        """Server-side mirror of the Cookbook UI's common serve diagnoses.
-
-        The browser uses cookbook-diagnosis.js for clickable fixes. This gives
-        the agent/tool path the same structured signal so it can retry with an
-        adjusted command instead of guessing from raw tmux output.
-        """
-        if not text:
-            return None
-        tail = text[-6000:]
-        patterns = [
-            (
-                r"No available memory for the cache blocks|Available KV cache memory:.*-",
-                "No GPU memory left for KV cache after loading model.",
-                [
-                    {"label": "retry with GPU memory utilization 0.95", "op": "replace", "flag": "--gpu-memory-utilization", "value": "0.95"},
-                    {"label": "retry with context 2048", "op": "replace", "flag": "--max-model-len", "value": "2048"},
-                ],
-            ),
-            (
-                r"CUDA out of memory|torch\.cuda\.OutOfMemoryError|CUDA error: out of memory|warming up sampler|max_num_seqs.*gpu_memory_utilization",
-                "GPU ran out of memory during startup or warmup.",
-                [
-                    {"label": "retry with context 4096", "op": "replace", "flag": "--max-model-len", "value": "4096"},
-                    {"label": "retry with GPU memory utilization 0.80", "op": "replace", "flag": "--gpu-memory-utilization", "value": "0.80"},
-                    {"label": "retry with --enforce-eager", "op": "append", "arg": "--enforce-eager"},
-                ],
-            ),
-            (
-                r"not divisib|must be divisible|attention heads.*divisible",
-                "Tensor parallel size is incompatible with the model.",
-                [
-                    {"label": "retry with tensor parallel size 1", "op": "replace", "flag": "--tensor-parallel-size", "value": "1"},
-                    {"label": "retry with tensor parallel size 2", "op": "replace", "flag": "--tensor-parallel-size", "value": "2"},
-                ],
-            ),
-            (
-                r"KV cache.*too (small|large)|max_model_len.*exceeds|maximum.*context",
-                "Context length is too large for available GPU memory.",
-                [
-                    {"label": "retry with context 8192", "op": "replace", "flag": "--max-model-len", "value": "8192"},
-                    {"label": "retry with context 4096", "op": "replace", "flag": "--max-model-len", "value": "4096"},
-                ],
-            ),
-            (
-                r"enable-auto-tool-choice requires --tool-call-parser",
-                "Auto tool choice requires an explicit tool call parser.",
-                [{"label": "retry with Hermes tool parser", "op": "append", "arg": "--tool-call-parser hermes"}],
-            ),
-            (
-                r"Please pass.*trust.remote.code=True|contains custom code which must be executed to correctly load|does not recognize this architecture|model type.*but Transformers does not",
-                "Model requires custom code or newer model support.",
-                [{"label": "retry with --trust-remote-code", "op": "append", "arg": "--trust-remote-code"}],
-            ),
-            (
-                r"Either a revision or a version must be specified|transformers\.integrations\.hub_kernels|kernels/layer",
-                "vLLM/Transformers kernel package mismatch.",
-                [{"label": "update vLLM, Transformers, and kernels on this server", "op": "dependency", "package": "vllm transformers kernels"}],
-            ),
-            (
-                r"Address already in use|bind.*address.*in use",
-                "Port is already in use.",
-                [{"label": "retry on port 8001", "op": "replace", "flag": "--port", "value": "8001"}],
-            ),
-            (
-                r"No CUDA GPUs are available|no GPU.*found|CUDA_VISIBLE_DEVICES.*invalid",
-                "No GPUs are visible to the serve process.",
-                [{"label": "clear Cookbook GPU selection or choose available GPUs", "op": "settings", "field": "gpus", "value": ""}],
-            ),
-            (
-                r"Failed to infer device type|NVML Shared Library Not Found|No module named 'amdsmi'|platform is not available",
-                "vLLM could not find a supported GPU (CUDA or ROCm). "
-                "This machine may have integrated or unsupported graphics only.",
-                [
-                    {"label": "switch to llama.cpp (CPU/Metal, works without a discrete GPU)", "op": "manual"},
-                    {"label": "switch to Ollama (CPU/Metal, works without a discrete GPU)", "op": "manual"},
-                ],
-            ),
-            (
-                r"vllm.*command not found|No module named vllm|ERROR: vLLM is not installed",
-                "vLLM is not installed or not in PATH on this server.",
-                [{"label": "install vLLM in Cookbook Dependencies", "op": "dependency", "package": "vllm"}],
-            ),
-            (
-                r"sgl_kernel[\s\S]*(Python\.h|libnuma\.so\.1|common_ops|libnvrtc\.so)|"
-                r"(Python\.h|libnuma\.so\.1|common_ops|libnvrtc\.so)[\s\S]*sgl_kernel|"
-                r"Could not load any common_ops library|"
-                r"Please ensure sgl_kernel is properly installed",
-                "SGLang native kernel/runtime is missing or mismatched on this server.",
-                [
-                    {"label": "repair sglang-kernel in this Python environment", "op": "dependency", "package": "sglang-kernel"},
-                    {"label": "if libnvrtc is still missing, install the matching CUDA/NVRTC runtime on this host", "op": "manual"},
-                ],
-            ),
-            (
-                r"sglang.*command not found|No module named sglang|SGLang is not installed",
-                "SGLang is not installed or not in PATH on this server.",
-                [{"label": "install SGLang in Cookbook Dependencies", "op": "dependency", "package": "sglang[all]"}],
-            ),
-            (
-                r"No module named ['\"]?mlx_lm|mlx_lm.*command not found|MLX is not installed|MLX LM is not installed",
-                "MLX LM is not installed on this server.",
-                [{"label": "install mlx-lm in Cookbook Dependencies", "op": "dependency", "package": "mlx-lm"}],
-            ),
-            (
-                r"OmniGen2Pipeline|module diffusers has no attribute .*Pipeline|custom_pipeline=.*failed",
-                "This image model uses a custom Diffusers pipeline that the launch environment does not know yet.",
-                [{"label": "update Diffusers image dependencies", "op": "dependency", "package": "diffusers transformers accelerate"}],
-            ),
-            (
-                r"Unable to quantize model of type <class ['\"]mlx_lm\.models\.switch_layers\.QuantizedSwitchLinear['\"]>|QuantizedSwitchLinear",
-                "MLX-LM tried to quantize an already-quantized DeepSeek switch layer.",
-                [
-                    {"label": "relaunch from the cached local Hugging Face snapshot path on this Mac", "op": "manual"},
-                    {"label": "Odysseus now rewrites MLX repo-id launches to a cached snapshot when one exists", "op": "manual"},
-                ],
-            ),
-            # System build deps come BEFORE the generic llama.cpp catch-all
-            # so cmake / build-essential / git missing → a specific OS-package
-            # remediation instead of "install llama-cpp-python[server]" (which
-            # itself fails to compile when cmake is absent).
-            (
-                r"cmake: command not found|cmake.*not found.*[Cc]ould not",
-                "cmake is required to build llama.cpp from source but isn't installed on this server.",
-                [{"label": "install build deps for llama.cpp (apt: cmake build-essential git / pacman: cmake base-devel git / dnf: cmake gcc-c++ make git / brew: cmake git)", "op": "dependency", "package": "llama-cpp-python[server]"}],
-            ),
-            (
-                r"^(make|g\+\+|gcc): command not found|Could not find C\+\+ compiler",
-                "A C/C++ compiler (build-essential) is required to build llama.cpp from source.",
-                [{"label": "install build deps for llama.cpp on this server", "op": "dependency", "package": "llama-cpp-python[server]"}],
-            ),
-            (
-                r"^git: command not found",
-                "git is required to clone the llama.cpp source tree.",
-                [{"label": "install build deps for llama.cpp on this server", "op": "dependency", "package": "llama-cpp-python[server]"}],
-            ),
-            (
-                r"llama-server.*command not found|llama\.cpp.*not found|No module named.*llama_cpp|No module named 'starlette_context'",
-                "llama.cpp / llama-cpp-python dependencies are missing.",
-                [{"label": "install llama.cpp dependencies or llama-cpp-python[server]", "op": "dependency", "package": "llama-cpp-python[server]"}],
-            ),
-            (
-                r"No GGUF found on this host|no \.gguf file|No GGUF file found",
-                "No GGUF file found for this model on this host. The llama.cpp backend needs a .gguf file.",
-                [{"label": "download a GGUF build of this model (repo name usually ends in -GGUF, file like Q4_K_M.gguf)", "op": "manual"}],
-            ),
-            (
-                r"No module named 'torch'|No module named torch|No module named 'torchvision'|No module named torchvision|No module named 'diffusers'|No module named diffusers|No module named 'scipy'|No module named scipy|install scipy if you want to use beta sigmas|requires the Torchvision library",
-                "Diffusion serving requires PyTorch, Torchvision, Diffusers, Accelerate, and SciPy.",
-                [{"label": "install Diffusers image deps in Cookbook Dependencies", "op": "dependency", "package": "diffusers[torch] torchvision accelerate scipy python-multipart"}],
-            ),
-            (
-                r"403 Forbidden|401 Unauthorized|Access to model.*is restricted|gated repo|not in the authorized list|awaiting a review",
-                "Model access is gated or unauthorized.",
-                [{"label": "set HF token and request model access on HuggingFace", "op": "manual"}],
-            ),
-        ]
-        for pattern, message, suggestions in patterns:
-            if re.search(pattern, tail, re.I):
-                return {"message": message, "suggestions": suggestions}
-        if re.search(r"Traceback \(most recent call last\)", tail, re.I) and not re.search(
-            r"Application startup complete|GET /v1/|Uvicorn running on", tail, re.I
-        ):
-            return {
-                "message": "Python traceback detected during serve startup.",
-                "suggestions": [{"label": "inspect traceback and retry with adjusted backend/settings", "op": "manual"}],
-            }
-        return None
-
     def _state_for_client(state):
         """Return cookbook state without raw secrets for browser clients."""
         _strip_task_secrets(state)
@@ -1575,80 +1411,6 @@ def setup_cookbook_routes() -> APIRouter:
 
         return {"models": models, "host": host or "local"}
 
-    def _auto_register_image_endpoint(req: ServeRequest, remote: str | None) -> str | None:
-        """Register a diffusion model as an image endpoint so it appears in the model selector."""
-        import re
-        from core.database import SessionLocal, ModelEndpoint
-        from src.settings import load_settings, save_settings
-
-        # Parse port from command (--port NNNN), default 8100 for diffusion_server
-        port_match = re.search(r'--port\s+(\d+)', req.cmd)
-        port = int(port_match.group(1)) if port_match else 8100
-
-        # Determine host
-        if remote:
-            # SSH alias — use as hostname (Tailscale resolves it later)
-            host = remote.split("@")[-1] if "@" in remote else remote
-        else:
-            host = "localhost"
-
-        base_url = f"http://{host}:{port}/v1"
-
-        # Friendly display name from repo_id
-        short_name = req.repo_id.split("/")[-1] if "/" in req.repo_id else req.repo_id
-        display_name = f"{short_name} (image)"
-        pinned_models = [req.repo_id] if req.repo_id else []
-
-        db = SessionLocal()
-        try:
-            # Check for existing endpoint with same base_url — update it
-            existing = db.query(ModelEndpoint).filter(ModelEndpoint.base_url == base_url).first()
-            if existing:
-                existing.is_enabled = True
-                existing.model_type = "image"
-                existing.name = display_name
-                existing.endpoint_kind = "local"
-                existing.model_refresh_mode = "manual"
-                if pinned_models:
-                    existing.cached_models = json.dumps(pinned_models)
-                    existing.pinned_models = json.dumps(pinned_models)
-                db.commit()
-                settings = load_settings()
-                if settings.get("image_gen_enabled") is not True:
-                    settings["image_gen_enabled"] = True
-                    save_settings(settings)
-                logger.info(f"Updated existing image endpoint: {base_url}")
-                return existing.id
-
-            ep_id = f"img-{uuid.uuid4().hex[:8]}"
-            ep = ModelEndpoint(
-                id=ep_id,
-                name=display_name,
-                base_url=base_url,
-                api_key=None,
-                is_enabled=True,
-                model_type="image",
-                endpoint_kind="local",
-                model_refresh_mode="manual",
-                cached_models=json.dumps(pinned_models) if pinned_models else None,
-                pinned_models=json.dumps(pinned_models) if pinned_models else None,
-            )
-            db.add(ep)
-            db.commit()
-            settings = load_settings()
-            settings["image_gen_enabled"] = True
-            if not settings.get("image_model"):
-                settings["image_model"] = req.repo_id
-            save_settings(settings)
-            logger.info(f"Auto-registered image endpoint: {display_name} @ {base_url}")
-            return ep_id
-        except Exception as e:
-            logger.error(f"Failed to auto-register image endpoint: {e}")
-            db.rollback()
-            return None
-        finally:
-            db.close()
-
     def _pick_free_port_for_ollama(
         remote: str | None, ssh_port: str | None, start_port: int, max_offset: int
     ) -> int | None:
@@ -1777,244 +1539,66 @@ def setup_cookbook_routes() -> APIRouter:
                 # let the probe layer mark it offline if nothing's listening.
                 logger.info(f"crash-watchdog: serve {session_id} exited cleanly (0); leaving endpoint {endpoint_id}")
                 return
-            # Non-zero exit — drop the endpoint.
+            # A nonzero exit may delete only the endpoint owned by this serve.
             try:
                 from core.database import SessionLocal as _SL, ModelEndpoint as _ME
+                from core.constants import internal_api_base as _internal_api_base
+                from core.middleware import INTERNAL_TOOL_HEADER, INTERNAL_TOOL_TOKEN
+                from urllib.parse import quote
+                import httpx
+
                 db = _SL()
                 try:
                     ep = db.query(_ME).filter(_ME.id == endpoint_id).first()
-                    if ep:
+                    if not ep or ep.cookbook_serve_session_id != session_id:
-                        # A scheduled serve can leave old non-zero exit markers
+                        logger.info(
-                        # in tmux scrollback while the current OpenAI endpoint is
+                            "crash-watchdog: endpoint %s no longer owned by %s",
-                        # actually alive. Verify reachability before deleting the
+                            endpoint_id, session_id,
-                        # endpoint row; otherwise chats fall back even though the
+                        )
-                        # served model is ready.
+                        return
-                        try:
                     probe_url = ep.base_url.rstrip("/") + "/models"
+                finally:
+                    db.close()
+
+                # Preserve an endpoint that is already healthy, including
+                # after an old runner emits a stale failure marker.
+                try:
                     with urllib.request.urlopen(probe_url, timeout=3) as resp:
                         if 200 <= getattr(resp, "status", 0) < 300:
                             logger.info(
-                                        f"crash-watchdog: serve {session_id} has exit marker {exit_code} "
+                                "crash-watchdog: endpoint %s is reachable; keeping it",
-                                        f"but endpoint {ep.id} is reachable; leaving it registered"
+                                endpoint_id,
                             )
                             return
                 except Exception:
                     pass
+
+                async with httpx.AsyncClient(timeout=8) as client:
+                    response = await client.delete(
+                        f"{_internal_api_base()}/api/model-endpoints/{quote(endpoint_id, safe='')}",
+                        params={"expected_cookbook_serve_session_id": session_id},
+                        headers={INTERNAL_TOOL_HEADER: INTERNAL_TOOL_TOKEN},
+                    )
+                if response.status_code == 200:
                     logger.info(
-                            f"crash-watchdog: dropping endpoint {endpoint_id} "
+                        "crash-watchdog: deleted endpoint %s after serve %s exited %s",
-                            f"({ep.name} @ {ep.base_url}) — serve exited {exit_code}"
+                        endpoint_id, session_id, exit_code,
+                    )
+                elif response.status_code in (404, 409):
+                    logger.info(
+                        "crash-watchdog: endpoint %s absent or reassigned; keeping it",
+                        endpoint_id,
+                    )
+                else:
+                    logger.warning(
+                        "crash-watchdog: endpoint %s deletion returned HTTP %s",
+                        endpoint_id, response.status_code,
                     )
-                        db.delete(ep)
-                        db.commit()
-                finally:
-                    db.close()
             except Exception as e:
-                logger.warning(f"crash-watchdog: endpoint cleanup failed: {e!r}")
+                logger.warning("crash-watchdog: endpoint cleanup failed: %r", e)
             return
         logger.debug(f"crash-watchdog: no exit marker for {session_id} within window; leaving endpoint {endpoint_id}")
 
-    def _auto_register_llm_endpoint(req: ServeRequest, remote: str | None) -> str | None:
-        """Register a freshly-served LLM as a model endpoint so it appears in the
-        model picker without a manual /setup step — the text-model sibling of
-        _auto_register_image_endpoint.
-
-        Cookbook serve commands launch an OpenAI-compatible server (llama.cpp's
-        llama-server, vLLM, SGLang, or Ollama) on a known port. We point an
-        endpoint at that server's /v1; the picker auto-discovers the model id by
-        probing /v1/models and dims the endpoint until the server is reachable,
-        so registering immediately (before the server finishes loading) is safe.
-        """
-        logger.info(
-            f"_auto_register_llm_endpoint: ENTRY repo_id={req.repo_id!r} "
-            f"remote={remote!r} cmd_prefix={req.cmd[:80]!r}"
-        )
-        import re
-        from core.database import SessionLocal, ModelEndpoint
-
-        # Port: ordered fallbacks so we match whatever the user actually
-        # asked for, not a hardcoded default:
-        #   1. explicit `--port N`  (vllm / sglang / llama-server)
-        #   2. `OLLAMA_HOST=host:port`  (the way Ollama specifies its bind)
-        #   3. fallback by backend (11434 ollama / 8080 llama.cpp)
-        # Previously the OLLAMA_HOST form was silently ignored and we
-        # registered every Ollama endpoint at 11434 — even if the user
-        # set OLLAMA_HOST=0.0.0.0:11435 to avoid colliding with an
-        # existing systemd Ollama, the registered endpoint pointed at
-        # the OLD port and showed as offline.
-        port_match = re.search(r'--port\s+(\d+)', req.cmd)
-        ollama_host_match = re.search(r'OLLAMA_HOST=[^\s]*?:(\d+)', req.cmd)
-        if port_match:
-            port = int(port_match.group(1))
-        elif ollama_host_match:
-            port = int(ollama_host_match.group(1))
-        elif "ollama" in req.cmd:
-            port = 11434
-        else:
-            port = 8080  # llama.cpp's llama-server default — the Apple Silicon path
-
-        # Determine host. The cookbook tmux for `local=true` serves runs INSIDE
-        # the odysseus container — so the right URL for the in-container
-        # backend to reach it is `localhost`, NOT `host.docker.internal`
-        # (the latter points at the docker HOST, which doesn't have a server
-        # on that port). The previous host.docker.internal fallback only made
-        # sense for /setup-added external services like systemd Ollama on the
-        # host — and those go through manual setup, not this auto-register
-        # code path. For remote serves we still use the SSH host alias.
-        if remote:
-            host = remote.split("@")[-1] if "@" in remote else remote
-        elif re.search(r"\bdocker\s+exec\s+(?:ollama-rocm|ollama-test)\b", req.cmd or ""):
-            host = "host.docker.internal"
-        else:
-            host = "localhost"
-
-        base_url = f"http://{host}:{port}/v1"
-
-        short_name = req.repo_id.split("/")[-1] if "/" in req.repo_id else req.repo_id
-        display_name = short_name or "Local model"
-        is_mlx_deepseek_v4 = (
-            "mlx_lm.server" in (req.cmd or "")
-            and "deepseek-v4" in ((req.repo_id or "") + " " + (req.cmd or "")).lower()
-        )
-        mlx_shim_model_id = ""
-        if is_mlx_deepseek_v4 and short_name:
-            home_match = re.search(r"((?:/Users|/home)/[^/\s'\"]+)", req.cmd or "")
-            remote_home = home_match.group(1) if home_match else ""
-            if remote_home:
-                mlx_shim_model_id = f"{remote_home}/.cache/odysseus/mlx-shims/{short_name}"
-
-        # If the serve command opts models into OpenAI tool-calling, record it so
-        # agent_loop trusts emitted tool_calls instead of the name heuristic.
-        is_ollama_endpoint = "ollama" in (req.cmd or "").lower()
-        supports_tools = True if "--enable-auto-tool-choice" in req.cmd else None
-        # Pin the model the user launched for every Cookbook-created LLM
-        # endpoint, not just Ollama. Some OpenAI-compatible servers report a
-        # deployment alias from /v1/models, and a stale server can answer on the
-        # same port while the new launch failed. Keeping the requested model id
-        # pinned makes the picker reflect the actual launch intent.
-        pinned_models = [mlx_shim_model_id] if mlx_shim_model_id else ([req.repo_id] if req.repo_id else [])
-
-        db = SessionLocal()
-        try:
-            # Reuse an endpoint already pointed at this URL instead of duplicating.
-            existing = db.query(ModelEndpoint).filter(ModelEndpoint.base_url == base_url).first()
-            if existing:
-                existing.is_enabled = True
-                existing.model_type = "llm"
-                existing.name = display_name
-                existing.endpoint_kind = "local"
-                existing.model_refresh_mode = "auto"
-                if pinned_models:
-                    try:
-                        existing_pinned = json.loads(existing.pinned_models or "[]")
-                    except Exception:
-                        existing_pinned = []
-                    merged_pinned = []
-                    for mid in [*existing_pinned, *pinned_models]:
-                        if mid and mid not in merged_pinned:
-                            merged_pinned.append(mid)
-                    existing.pinned_models = json.dumps(merged_pinned) if merged_pinned else None
-                if is_ollama_endpoint:
-                    existing.endpoint_kind = "ollama"
-                    if pinned_models:
-                        existing.cached_models = json.dumps(pinned_models)
-                if supports_tools is not None:
-                    existing.supports_tools = supports_tools
-                db.commit()
-                logger.info(f"Updated existing local model endpoint: {base_url}")
-                # Re-probe so cached_models matches what the server actually
-                # serves right now (the URL may have stayed the same but the
-                # model behind it changed across launches).
-                try:
-                    if mlx_shim_model_id:
-                        existing.cached_models = json.dumps([mlx_shim_model_id])
-                        existing.pinned_models = json.dumps([mlx_shim_model_id])
-                        db.commit()
-                    else:
-                        from routes.model_routes import _probe_endpoint
-                        import json as _json2
-                        probed = _probe_endpoint(base_url, existing.api_key, timeout=5)
-                        if probed:
-                            existing.cached_models = _json2.dumps(probed)
-                            db.commit()
-                except Exception as _pe:
-                    logger.warning(f"Re-probe failed for {base_url}: {_pe!r}")
-                # Sweep stale dupes: other endpoints with the same display name
-                # at DIFFERENT URLs (likely failed earlier-attempt ports) get
-                # deleted so the picker doesn't show an offline ghost next to
-                # the working one. Only sweeps endpoints whose id starts with
-                # `local-` so we never touch a user's hand-added DeepSeek/OpenAI/
-                # etc. entry with a coincidentally matching name.
-                stale = (db.query(ModelEndpoint)
-                         .filter(ModelEndpoint.name == display_name)
-                         .filter(ModelEndpoint.base_url != base_url)
-                         .filter(ModelEndpoint.id.like("local-%"))
-                         .all())
-                for s in stale:
-                    logger.info(f"Sweeping stale local endpoint {s.id} ({s.base_url})")
-                    db.delete(s)
-                if stale:
-                    db.commit()
-                return existing.id
-
-            ep_id = f"local-{uuid.uuid4().hex[:8]}"
-            ep = ModelEndpoint(
-                id=ep_id,
-                name=display_name,
-                base_url=base_url,
-                api_key=None,
-                is_enabled=True,
-                model_type="llm",
-                endpoint_kind="ollama" if is_ollama_endpoint else "local",
-                model_refresh_mode="auto",
-                cached_models=json.dumps(pinned_models) if pinned_models else None,
-                pinned_models=json.dumps(pinned_models) if pinned_models else None,
-                supports_tools=supports_tools,
-            )
-            db.add(ep)
-            db.commit()
-            logger.info(f"Auto-registered local model endpoint: {display_name} @ {base_url}")
-            # Same sweep on first-register path: drop any pre-existing local-*
-            # endpoints with this display name pointed elsewhere.
-            stale = (db.query(ModelEndpoint)
-                     .filter(ModelEndpoint.name == display_name)
-                     .filter(ModelEndpoint.id != ep_id)
-                     .filter(ModelEndpoint.id.like("local-%"))
-                     .all())
-            for s in stale:
-                logger.info(f"Sweeping stale local endpoint {s.id} ({s.base_url})")
-                db.delete(s)
-            if stale:
-                db.commit()
-            # Probe /v1/models NOW and write cached_models so the chat
-            # picker actually shows the model on the next /api/models
-            # call. Without this immediate probe, the endpoint has empty
-            # cached_models until the next background refresh fires (up
-            # to a minute later) and the picker shows nothing — even
-            # though the endpoint is in the DB and the server is up.
-            try:
-                if mlx_shim_model_id:
-                    ep.cached_models = json.dumps([mlx_shim_model_id])
-                    ep.pinned_models = json.dumps([mlx_shim_model_id])
-                    db.commit()
-                    logger.info(f"Auto-register: pinned MLX DeepSeek-V4 shim model @ {base_url}")
-                else:
-                    from routes.model_routes import _probe_endpoint
-                    import json as _json2
-                    probed = _probe_endpoint(base_url, None, timeout=5)
-                    if probed:
-                        ep.cached_models = _json2.dumps(probed)
-                        db.commit()
-                        logger.info(f"Auto-register: probed {len(probed)} models @ {base_url}")
-            except Exception as _pe:
-                logger.warning(f"Auto-register: probe-after-create failed for {base_url}: {_pe!r}")
-            return ep_id
-        except Exception as e:
-            logger.error(f"Failed to auto-register local model endpoint: {e}")
-            db.rollback()
-            return None
-        finally:
-            db.close()
-
     @router.post("/api/model/serve")
     async def model_serve(request: Request, req: ServeRequest):
         """Launch a model server in a tmux session (or PowerShell background process on Windows).
@@ -2161,7 +1745,7 @@ def setup_cookbook_routes() -> APIRouter:
                 ps_lines.append('}')
             elif "llama_cpp" in req.cmd or "llama-server" in req.cmd:
                 ps_lines.append('# Auto-install llama-cpp-python if missing')
-                ps_lines.append('try { python -c "import llama_cpp" 2>$null } catch {}')
+                ps_lines.append('try { python -c "import llama_cpp.server" 2>$null } catch {}')
                 ps_lines.append('if ($LASTEXITCODE -ne 0) {')
                 ps_lines.append('  Write-Host "Installing llama-cpp-python..."')
                 ps_lines.append('  python -m pip install llama-cpp-python[server]')
@@ -2247,14 +1831,17 @@ def setup_cookbook_routes() -> APIRouter:
                 # ollama is found (otherwise macOS falls back to a slow source build).
                 # /opt/homebrew = Apple Silicon, /usr/local = Intel; harmless on Linux.
                 runner_lines.append('export PATH="$HOME/.local/bin:$HOME/bin:$HOME/llama.cpp/build/bin:/opt/homebrew/bin:/usr/local/bin:$PATH"')
+                _append_llama_server_shim_definition_lines(runner_lines)
+                _append_llama_server_shim_reconciliation_lines(runner_lines)
                 runner_lines.append('if [ -d /data/data/com.termux ]; then')
                 runner_lines.append('  # Termux: no native build — use the Python bindings (CPU).')
-                runner_lines.append('  if ! python3 -c "import llama_cpp" 2>/dev/null; then')
+                runner_lines.append('  if ! python3 -c "import llama_cpp.server" 2>/dev/null; then')
                 runner_lines.append('    pkg install -y cmake 2>/dev/null')
                 runner_lines.append('    pip install numpy diskcache jinja2 2>/dev/null')
                 runner_lines.append('    CMAKE_ARGS="-DGGML_BLAS=OFF -DGGML_LLAMAFILE=OFF" pip install \'llama-cpp-python[server]\' --no-build-isolation --no-cache-dir 2>&1 || true')
                 runner_lines.append('  fi')
-                runner_lines.append('elif ! command -v llama-server &>/dev/null; then')
+                _append_llama_server_finalization_lines(runner_lines, platform_hint="Termux")
+                runner_lines.append('elif [ -z "$_odysseus_llama_bin" ]; then')
                 runner_lines.append('  echo "Native llama-server not found — building from source (one-time, may take a few minutes)..."')
                 runner_lines.append('  mkdir -p ~/bin')
                 runner_lines.append('  cd ~ && [ -d llama.cpp ] || git clone --depth 1 https://github.com/ggml-org/llama.cpp')
@@ -2271,8 +1858,7 @@ def setup_cookbook_routes() -> APIRouter:
                 # would reuse the bad settings and fail again. CMAKE_BUILD_TYPE is
                 # explicit so the binary is optimized (Metal auto-enables on macOS).
                 runner_lines.append('    cd ~/llama.cpp && rm -rf build && cmake -B build -DCMAKE_BUILD_TYPE=Release \\')
-                runner_lines.append('      && cmake --build build -j"$NPROC" --target llama-server \\')
+                runner_lines.append('      && cmake --build build -j"$NPROC" --target llama-server')
-                runner_lines.append('      && ln -sf ~/llama.cpp/build/bin/llama-server ~/bin/llama-server')
                 runner_lines.append('  else')
                 _append_llama_cpp_linux_accel_build_lines(runner_lines)
                 runner_lines.append('  fi')
@@ -2296,58 +1882,13 @@ def setup_cookbook_routes() -> APIRouter:
                 runner_lines.append('      fi')
                 runner_lines.append('    fi')
                 runner_lines.append('  fi')
-                # SHORT-CIRCUIT before the build/pip fallback: if the
+                runner_lines.append('  _odysseus_resolve_llama_server || true')
-                # native binary is missing but llama_cpp Python is already
+                # If the native build failed, fall back to the Python bindings.
-                # installed, drop a wrapper at ~/bin/llama-server that
+                runner_lines.append('  if [ -z "$_odysseus_llama_bin" ] && ! python3 -c "import llama_cpp.server" 2>/dev/null; then')
-                # translates llama-server CLI args to llama_cpp.server's
-                # underscore-style flags. The user's serve command stays
-                # `llama-server ...` and "just works" — no build, no cmake,
-                # no second install. This is the path that unblocks every
-                # remote where pip-installed llama-cpp-python is already
-                # working but Cookbook used to insist on a native binary.
-                runner_lines.append('  if ! command -v llama-server >/dev/null 2>&1 && python3 -c "import llama_cpp" 2>/dev/null; then')
-                runner_lines.append('    mkdir -p ~/bin')
-                runner_lines.append('    cat > ~/bin/llama-server <<\'_ODY_LLAMA_SHIM_EOF\'')
-                runner_lines.append('#!/usr/bin/env bash')
-                runner_lines.append('# Auto-generated by Odysseus Cookbook: a `llama-server` lookalike')
-                runner_lines.append('# that translates the native CLI to `python -m llama_cpp.server`.')
-                runner_lines.append('# Lets cookbook-generated launch commands run unchanged on hosts')
-                runner_lines.append('# where only the pip llama-cpp-python package is installed.')
-                runner_lines.append('ARGS=()')
-                runner_lines.append('while [ $# -gt 0 ]; do')
-                runner_lines.append('  case "$1" in')
-                runner_lines.append('    -ngl|--gpu-layers|--n-gpu-layers) ARGS+=(--n_gpu_layers "$2"); shift 2 ;;')
-                runner_lines.append('    -c|--ctx-size) ARGS+=(--n_ctx "$2"); shift 2 ;;')
-                runner_lines.append('    -b|--batch-size) ARGS+=(--n_batch "$2"); shift 2 ;;')
-                runner_lines.append('    -ub|--ubatch-size) shift 2 ;;  # llama-cpp-python has no separate ubatch')
-                runner_lines.append('    --flash-attn) ARGS+=(--flash_attn true); shift 2 ;;')
-                runner_lines.append('    --cache-type-k) ARGS+=(--type_k "$2"); shift 2 ;;')
-                runner_lines.append('    --cache-type-v) ARGS+=(--type_v "$2"); shift 2 ;;')
-                runner_lines.append('    --n-cpu-moe) ARGS+=(--n_cpu_moe "$2"); shift 2 ;;')
-                runner_lines.append('    --mmproj) ARGS+=(--clip_model_path "$2"); shift 2 ;;')
-                runner_lines.append('    --image-max-tokens) shift 2 ;;  # native-only')
-                runner_lines.append('    --no-mmap) ARGS+=(--no_mmap true); shift ;;')
-                runner_lines.append('    --no-warmup) shift ;;  # native-only')
-                runner_lines.append('    --chat-template) ARGS+=(--chat_format "$2"); shift 2 ;;')
-                runner_lines.append('    --fit|--split-mode|--tensor-split|--main-gpu|--parallel) shift 2 ;;  # native-only')
-                runner_lines.append('    --mlock) ARGS+=(--use_mlock true); shift ;;')
-                runner_lines.append('    *) ARGS+=("$1"); shift ;;')
-                runner_lines.append('  esac')
-                runner_lines.append('done')
-                runner_lines.append('exec python3 -m llama_cpp.server "${ARGS[@]}"')
-                runner_lines.append('_ODY_LLAMA_SHIM_EOF')
-                runner_lines.append('    chmod +x ~/bin/llama-server')
-                runner_lines.append('    echo "[odysseus] Created llama-server shim → python -m llama_cpp.server (no native binary needed)"')
-                runner_lines.append('  fi')
-                runner_lines.append('  # If the native build failed, fall back to the Python bindings.')
-                runner_lines.append('  if ! command -v llama-server &>/dev/null && ! python3 -c "import llama_cpp" 2>/dev/null; then')
                 runner_lines.append('    echo "llama-server build failed — installing Python bindings as fallback..."')
                 runner_lines.append(f"    {_pip_install_fallback_chain('llama-cpp-python[server]', python_cmd='pip')} || true")
                 runner_lines.append('  fi')
-                runner_lines.append('  if ! command -v llama-server &>/dev/null && ! python3 -c "import llama_cpp" 2>/dev/null; then')
+                _append_llama_server_finalization_lines(runner_lines)
-                runner_lines.append('    echo "ERROR: llama.cpp serving is not available after install/build attempts."')
-                runner_lines.append('    ODYSSEUS_PREFLIGHT_EXIT=127')
-                runner_lines.append('  fi')
                 runner_lines.append('fi')
             elif re.search(r"\bollama\s+serve\b", req.cmd):
                 handled_ollama_serve = True
@@ -2849,9 +2390,17 @@ def setup_cookbook_routes() -> APIRouter:
         endpoint_id = None
         is_image_endpoint = "diffusion_server.py" in req.cmd or "mlx_image_server.py" in req.cmd
         if is_image_endpoint:
-            endpoint_id = _auto_register_image_endpoint(req, remote)
+            endpoint_id = _auto_register_image_endpoint(
+                req,
+                remote,
+                serve_session_id=session_id,
+            )
         elif not is_pip_install:
-            endpoint_id = _auto_register_llm_endpoint(req, remote)
+            endpoint_id = _auto_register_llm_endpoint(
+                req,
+                remote,
+                serve_session_id=session_id,
+            )
 
         # Crash watchdog: the auto-register above writes the endpoint row
         # IMMEDIATELY (before the server has even bound its port) so the
@@ -3971,7 +3520,7 @@ def setup_cookbook_routes() -> APIRouter:
                     return {"ok": False, "files": [], "error": f"HF API HTTP {resp.status_code}"}
                 data = resp.json()
         except Exception:
-            logger.exception("HF GGUF file scan failed for %s", repo)
+            logger.exception("HF GGUF file scan failed for %s", repo_id)
             return {"ok": False, "files": [], "error": "HF API request failed"}
         files = [
             str(s.get("rfilename") or "")
@@ -4343,7 +3892,7 @@ def setup_cookbook_routes() -> APIRouter:
             tmux output can stop at a stale progress line if the pane/session
             disappears before Cookbook captures the final DOWNLOAD_OK marker.
             In that case, trust the cache shape: a snapshot directory with files
-            and no *.incomplete blobs means HuggingFace finished materializing the
+            and no unfinished blobs means HuggingFace finished materializing the
             model. cache_root is the task's custom download dir — the runner
             pointed HF_HOME there, so the cache lives under <cache_root>/hub,
             not wherever this probe's environment says.
@@ -4368,7 +3917,8 @@ def setup_cookbook_routes() -> APIRouter:
             """Best-effort check for resumable HF partial blobs.
 
             A lost SSH/tmux session can leave a real download still incomplete.
-            Treat any *.incomplete blob as stronger evidence than stale
+            Treat unfinished blobs (excluding empty markers paired with a
+            completed blob) as stronger evidence than stale
             "100%" lines in the captured pane output.
             """
             if not repo_id or "/" not in repo_id:
@@ -4486,7 +4036,9 @@ def setup_cookbook_routes() -> APIRouter:
                 # fills the visible tail. Without this, output_tail ends up
                 # as just "Locale: C / Ubuntu_Odysseus ❯" and the agent
                 # can't diagnose the actual error.
-                capture_cmd = ssh_base + [remote, _remote_tmux_command("capture-pane", "-t", session_id, "-p", "-S", "-500")]
+                # -J joins lines tmux wrapped at the pane width, so an error
+                # split mid-word still matches the serve diagnosis patterns.
+                capture_cmd = ssh_base + [remote, _remote_tmux_command("capture-pane", "-t", session_id, "-p", "-J", "-S", "-500")]
             elif IS_WINDOWS:
                 # LOCAL Windows task: launched as a detached process (no tmux).
                 # Liveness comes from the <session>.pid file, output from the
@@ -4495,7 +4047,7 @@ def setup_cookbook_routes() -> APIRouter:
                 capture_cmd = None
             else:
                 check_cmd = ["tmux", "has-session", "-t", session_id]
-                capture_cmd = ["tmux", "capture-pane", "-t", session_id, "-p", "-S", "-500"]
+                capture_cmd = ["tmux", "capture-pane", "-t", session_id, "-p", "-J", "-S", "-500"]
 
             local_win_task = (not remote) and IS_WINDOWS
 

routes/document/document_routes.py

@@ -16,6 +16,7 @@ from core.database import SessionLocal, Document, DocumentVersion
 from core.database import Session as DbSession
 from src.auth_helpers import get_current_user, _auth_disabled
 from src.constants import MAIL_ATTACHMENTS_DIR
+from src.session_search import _escape_like
 from src.upload_handler import reserve_upload_references
 
 logger = logging.getLogger(__name__)
@@ -571,9 +572,10 @@ def setup_document_routes(session_manager, upload_handler=None) -> APIRouter:
             # match docs containing both words regardless of position/order.
             if search:
                 for tok in search.split():
-                    term = f"%{tok}%"
+                    term = f"%{_escape_like(tok)}%"
                     q = q.filter(
-                        Document.title.ilike(term) | Document.current_content.ilike(term)
+                        Document.title.ilike(term, escape="\\")
+                        | Document.current_content.ilike(term, escape="\\")
                     )
 
             # Language filter. "pdf" is a display language derived from the
@@ -781,9 +783,11 @@ def setup_document_routes(session_manager, upload_handler=None) -> APIRouter:
         except Exception as e:
             logger.warning("Failed to parse export request body, defaulting to empty", exc_info=e)
             data = {}
-        ids = data.get("ids") or []
+        ids = data.get("ids") if isinstance(data, dict) else None
         if not ids:
             raise HTTPException(400, "No documents specified")
+        if not isinstance(ids, list) or not all(isinstance(i, str) for i in ids):
+            raise HTTPException(400, "ids must be a list of document ids")
         _ext = {
             "javascript": ".js", "python": ".py", "html": ".html", "css": ".css",
             "markdown": ".md", "json": ".json", "yaml": ".yml", "bash": ".sh",
@@ -810,10 +814,13 @@ def setup_document_routes(session_manager, upload_handler=None) -> APIRouter:
                     ext = _ext.get(doc.language or "text", ".txt")
                     base = (doc.title or "document").strip() or "document"
                     base = re.sub(r"[^\w\-. ]+", "", base)[:60].strip() or doc.id
-                    name = base if "." in base else base + ext
+                    has_ext = base.lower().endswith(ext)
+                    stem = base[:-len(ext)] if has_ext else base
+                    suffix = base[-len(ext):] if has_ext else ext
+                    name = stem + suffix
                     i = 1
                     while name in used:
-                        name = f"{base}-{i}" + ("" if "." in base else ext)
+                        name = f"{stem}-{i}{suffix}"
                         i += 1
                     used.add(name)
                     content = doc.current_content or ""

routes/editor_agent_routes.py

@@ -846,6 +846,64 @@ def _plan_psd_steps(prompt: str) -> Optional[List[Dict[str, Any]]]:
     return None
 
 
+def _is_background_removal_intent(prompt: str) -> bool:
+    """Linear-time, bounded intent parser for background removal commands.
+
+    Replaces the previous backtracking regular expression flagged by CodeQL
+    (#1212) while preserving the exact command grammar.
+    """
+    if not prompt or not isinstance(prompt, str):
+        return False
+    i = len(prompt)
+    while i > 0 and prompt[i - 1] in ".! \t\r\n":
+        i -= 1
+    cleaned = prompt[:i].strip()
+    if not cleaned:
+        return False
+
+    tokens = cleaned.split()
+    if not tokens:
+        return False
+
+    idx = 0
+    if tokens[idx].lower() == "please":
+        idx += 1
+        if idx >= len(tokens):
+            return False
+
+    if tokens[idx].lower() not in {"remove", "erase", "delete", "clear", "strip"}:
+        return False
+    idx += 1
+    if idx >= len(tokens):
+        return False
+
+    if tokens[idx].lower() == "the":
+        idx += 1
+        if idx >= len(tokens):
+            return False
+
+    if tokens[idx].lower() not in {"bg", "background", "backdrop"}:
+        return False
+    idx += 1
+
+    if idx == len(tokens):
+        return True
+
+    if tokens[idx].lower() in {"from", "of"}:
+        idx += 1
+        if idx < len(tokens) and tokens[idx].lower() == "the":
+            idx += 1
+        if (
+            idx + 1 < len(tokens)
+            and tokens[idx].lower() in {"active", "current", "selected"}
+            and tokens[idx + 1].lower() == "layer"
+            and idx + 2 == len(tokens)
+        ):
+            return True
+
+    return False
+
+
 def _plan_special(
     prompt: str, state: Any, provider_available: bool,
 ) -> Optional[Tuple[List[Dict[str, Any]], str, str]]:
@@ -855,6 +913,8 @@ def _plan_special(
     left to the model.     Returns (steps, reply, note) or None. These never need an
     image provider except generative border fill, used only when one exists.
     """
+    if _is_background_removal_intent(prompt):
+        return [{"op": "ai.removeBackground", "params": {}}], "Removing the active layer's background.", ""
     # Targeted recolour is fully deterministic (object mask + masked hue shift)
     # and must never be left to the model, which sometimes substitutes a
     # generative placeImage. Highest priority so it can never vary.
@@ -1203,19 +1263,64 @@ def _bridge_timeout(value: Any) -> float:
 
 
 def _is_admin(request: Request) -> bool:
+    """Return True only for authorized administrator browser sessions or auth-disabled single-user mode.
+
+    Delegated credentials (bearer API tokens) are explicitly rejected:
+    per src/tool_security.py, bearer tokens are long-lived credentials handed to
+    third parties/scripts and must remain capped at non-admin policies even
+    when minted by an administrator.
+    """
+    # 1. Delegated credentials (bearer API tokens) must NEVER receive admin metadata privileges
+    if getattr(request.state, "api_token", False):
+        return False
+
+    # 2. In-process internal tool bypass (loopback agent tools)
+    try:
+        from core.middleware import INTERNAL_TOOL_HEADER, INTERNAL_TOOL_TOKEN, INTERNAL_TOOL_USER
+        import secrets
+        hdr = request.headers.get(INTERNAL_TOOL_HEADER)
+        if hdr and secrets.compare_digest(hdr, INTERNAL_TOOL_TOKEN):
+            return True
+        if getattr(request.state, "current_user", None) == INTERNAL_TOOL_USER:
+            return True
+    except Exception:
+        pass
+
+    # 3. Explicitly auth-disabled single-user operation (AUTH_ENABLED=false)
+    try:
+        from src.auth_helpers import _auth_disabled
+        if _auth_disabled():
+            return True
+    except Exception:
+        pass
+
+    # 4. Middleware require_admin
     try:
         from core.middleware import require_admin
         require_admin(request)
         return True
     except Exception:
+        pass
+
+    # 5. Direct session/user check when request.state.current_user is set
+    try:
+        auth_mgr = getattr(request.app.state, "auth_manager", None)
+        if auth_mgr and getattr(auth_mgr, "is_configured", False):
+            from src.auth_helpers import get_current_user
+            user = get_current_user(request) or getattr(request.state, "current_user", None)
+            if user and getattr(auth_mgr, "is_admin", None) and auth_mgr.is_admin(user):
+                return True
+    except Exception:
+        pass
+
     return False
 
 
 def _user(request: Request) -> Optional[str]:
-    # These owner-scoped editor APIs have no delegated-token scope contract.
+    # Handle API bearer tokens explicitly: require 'chat' scope and attribute to real owner
-    # Do not collapse all bearer users into the shared "api" pseudo-owner.
     if getattr(request.state, "api_token", False):
-        raise HTTPException(403, "API tokens must use a scope-aware API route")
+        from src.auth_helpers import require_api_token_scope
+        return require_api_token_scope(request, "chat")
     from src.auth_helpers import get_current_user, _auth_disabled
     u = get_current_user(request)
     if u:
@@ -1358,9 +1463,13 @@ def setup_editor_agent_routes() -> APIRouter:
             except Exception:
                 raw = None
         _used = editor_agent_model.get_last_model()
-        model_used = (_used[0] if _used and _used[0]
-                      else (agent_model or editor_agent_model.local_model()))
         parsed = _parse_plan(raw)
+        if _used and _used[0]:
+            model_used = _used[0]
+        elif parsed:
+            model_used = agent_model or editor_agent_model.local_model()
+        else:
+            model_used = "fallback" if agent_model else editor_agent_model.local_model()
         if isinstance(parsed, dict):
             parsed["agent_model"] = model_used
         if mode != "check":
@@ -1598,6 +1707,13 @@ def setup_editor_agent_routes() -> APIRouter:
         screenshot = body.get("screenshot")
         # Optional user-picked image model (from the editor model selector).
         model_choice = str(body.get("model") or body.get("image_model") or "").strip()
+        generation_model_choice = str(body.get("generation_model") or "").strip()
+        use_image_settings = body.get("_use_image_settings") is True
+        if use_image_settings:
+            from src.editor_image_settings import editor_image_models
+            image_models = editor_image_models(owner=owner)
+            model_choice = image_models["edit"]
+            generation_model_choice = image_models["generation"]
         # Optional user-picked PLANNER model (the editor agent/planner model).
         agent_model = str(body.get("agent_model") or "").strip()
         tools_offered = [tool for tool in _scoped_editor_tools() if tool not in disabled]
@@ -1634,9 +1750,14 @@ def setup_editor_agent_routes() -> APIRouter:
             except Exception:
                 raw = None
         _used = editor_agent_model.get_last_model()
-        model_used = (_used[0] if _used and _used[0]
+        parsed = _parse_plan(raw)
-                      else (agent_model or editor_agent_model.local_model()))
+        if _used and _used[0]:
-        plan = _parse_plan(raw) or _fallback_plan(prompt)
+            model_used = _used[0]
+        elif parsed:
+            model_used = agent_model or editor_agent_model.local_model()
+        else:
+            model_used = "fallback" if agent_model else editor_agent_model.local_model()
+        plan = parsed or _fallback_plan(prompt)
         _strengthened = style_guide.strengthen(prompt, list(plan.get("steps") or []))
         if _strengthened != plan.get("steps"):
             plan = {**plan, "steps": _strengthened,
@@ -1723,6 +1844,12 @@ def setup_editor_agent_routes() -> APIRouter:
                 "results": [],
                 "agent_model": model_used,
             })
+        needed = {"plan_edit"}
+        for step in steps:
+            op = step.get("op")
+            needed.add(op if op in {"ai_edit_image", "generate_image"} else "run_editor_command")
+        if needed.intersection(disabled):
+            raise HTTPException(403, "Editor tool is disabled")
         results: List[Dict[str, Any]] = []
         for step in steps:
             op = step.get("op")
@@ -1737,9 +1864,22 @@ def setup_editor_agent_routes() -> APIRouter:
                 # Planner output is untrusted: it cannot redirect this turn to
                 # another editor session or override an explicit model choice.
                 tool_payload["session"] = session
+                if use_image_settings:
+                    tool_payload["_use_image_settings"] = True
+                    if op == "generate_image" and state.get("width") and state.get("height"):
+                        try:
+                            width, height = int(state["width"]), int(state["height"])
+                            if width > 0 and height > 0:
+                                tool_payload["size"] = f"{width}x{height}"
+                        except (TypeError, ValueError):
+                            pass
                 # Route to the user-picked model when one is selected.
-                if model_choice:
+                selected_model = generation_model_choice if op == "generate_image" else model_choice
-                    tool_payload["model"] = model_choice
+                if selected_model:
+                    tool_payload["model"] = selected_model
+                elif op == "generate_image" or use_image_settings:
+                    # Let server settings resolve Auto, not a planner-invented model.
+                    tool_payload.pop("model", None)
                 # Hand the AI step the canvas image the client already captured
                 # for this request. Without it the tool falls back to a separate
                 # bridge screenshot, which is not reliably answered right after a
@@ -1865,14 +2005,19 @@ def setup_editor_agent_routes() -> APIRouter:
                     # recognisably image-capable — the same test the request
                     # router uses.
                     image_endpoint = getattr(ep, "model_type", None) == "image"
+                    endpoint_id = getattr(ep, "id", None)
                     provider = getattr(ep, "name", None)
                     if not provider:
-                        provider = getattr(ep, "base_url", "") if is_admin else "image"
+                        provider = (
+                            getattr(ep, "base_url", "")
+                            if is_admin
+                            else (f"Endpoint {endpoint_id}" if endpoint_id else "image")
+                        )
                     provider = provider or "image"
                     for name in _endpoint_enabled_models(ep):
                         if not image_endpoint and not looks_like_image_generation_model(name):
                             continue
-                        key = (provider, name)
+                        key = (str(endpoint_id or provider), name)
                         if key in seen:
                             continue
                         seen.add(key)
@@ -1881,7 +2026,7 @@ def setup_editor_agent_routes() -> APIRouter:
                             "label": name,
                             "kind": "edit" if _EDIT_MODEL_RE.search(name) else "generate",
                             "provider": provider,
-                            "endpoint_id": getattr(ep, "id", None),
+                            "endpoint_id": endpoint_id,
                         }
                         if is_admin:
                             entry["base_url"] = getattr(ep, "base_url", "") or ""
@@ -1896,11 +2041,12 @@ def setup_editor_agent_routes() -> APIRouter:
     async def editor_agent_models(request: Request) -> JSONResponse:
         """List the models the editor agent/planner can run on.
 
-        Each entry is ``{id, label, source, vision, base_url}`` where ``id`` is
+        Each entry is ``{id, label, source, vision}`` where ``id`` is
         passed as ``agent_model`` to /api/editor/plan or /api/editor/agent, and
         ``source`` is ``local`` (the local Qwen3.5-9B vision server) or
         ``hosted`` (the configured default/utility chat model that the planner
         falls back to).
+        For non-admin users, private URLs and internal probe status are omitted.
         """
         owner = _user(request)
         is_admin = _is_admin(request)
@@ -1916,12 +2062,14 @@ def setup_editor_agent_routes() -> APIRouter:
         except Exception:
             logger.warning("editor agent-models lookup failed", exc_info=True)
             models = []
+
         probe_info = {}
         if is_admin:
             probe_info = editor_agent_model.probe_status()
             if not probe_info:
                 editor_agent_model.startup_probe()
                 probe_info = editor_agent_model.probe_status()
+
         return JSONResponse({
             "models": models,
             "probe": probe_info,

routes/email/email_helpers.py

@@ -36,6 +36,7 @@ from fastapi import Query, HTTPException, Request
 from pydantic import BaseModel
 from typing import Optional, List
 
+from src.email_mime import decode_text_part as _decode_text_part
 from src.auth_helpers import _auth_disabled, get_current_user
 from src.secret_storage import decrypt as _decrypt
 from src.url_safety import OutboundAddressBlocked, connect_outbound_tcp
@@ -1776,44 +1777,38 @@ def _extract_html(msg):
             ct = part.get_content_type()
             cd = str(part.get("Content-Disposition", ""))
             if ct == "text/html" and "attachment" not in cd:
-                payload = part.get_payload(decode=True)
+                text = _decode_text_part(part)
-                if payload:
+                if text:
-                    charset = part.get_content_charset() or "utf-8"
+                    return text
-                    return payload.decode(charset, errors="replace")
     elif msg.get_content_type() == "text/html":
-        payload = msg.get_payload(decode=True)
+        text = _decode_text_part(msg)
-        if payload:
+        if text:
-            charset = msg.get_content_charset() or "utf-8"
+            return text
-            return payload.decode(charset, errors="replace")
     return None
 
 
 def _extract_text(msg):
     if msg.is_multipart():
         text_parts = []
+        html_parts = []
         for part in msg.walk():
             ct = part.get_content_type()
             cd = str(part.get("Content-Disposition", ""))
             if ct == "text/plain" and "attachment" not in cd:
-                payload = part.get_payload(decode=True)
+                text = _decode_text_part(part)
-                if payload:
+                if text:
-                    charset = part.get_content_charset() or "utf-8"
+                    text_parts.append(text)
-                    text_parts.append(payload.decode(charset, errors="replace"))
+            elif ct == "text/html" and not html_parts and "attachment" not in cd:
-            elif ct == "text/html" and not text_parts and "attachment" not in cd:
+                raw_html = _decode_text_part(part)
-                payload = part.get_payload(decode=True)
+                if raw_html:
-                if payload:
-                    charset = part.get_content_charset() or "utf-8"
-                    raw_html = payload.decode(charset, errors="replace")
                     text = re.sub(r"<br\s*/?>", "\n", raw_html, flags=re.I)
                     text = re.sub(r"<[^>]+>", "", text)
                     text = html.unescape(text)
-                    text_parts.append(text.strip())
+                    html_parts.append(text.strip())
-        return "\n".join(text_parts)
+        return "\n".join(text_parts or html_parts)
     else:
-        payload = msg.get_payload(decode=True)
+        text = _decode_text_part(msg)
-        if payload:
+        if text:
-            charset = msg.get_content_charset() or "utf-8"
-            text = payload.decode(charset, errors="replace")
             if msg.get_content_type() == "text/html":
                 text = re.sub(r"<br\s*/?>", "\n", text, flags=re.I)
                 text = re.sub(r"</(?:p|div|li|tr|h[1-6])\s*>", "\n", text, flags=re.I)

routes/font_routes.py

@@ -19,9 +19,10 @@ def _split_family_token(token):
 def _derive_family(filename):
     """Derive a font-family name from a filename like 'JetBrainsMono-Regular.woff2' → 'JetBrains Mono'."""
     name = os.path.splitext(filename)[0]
-    # Strip common weight/style suffixes
+    # Strip weight/style suffixes, including BoldItalic and Bold-Italic.
     name = re.sub(
-        r'[-_ ]?(Thin|ExtraLight|UltraLight|Light|Regular|Medium|SemiBold|DemiBold|Bold|ExtraBold|UltraBold|Black|Heavy|Italic|Oblique|Variable|VF)$',
+        r'[-_ ]?(Thin|ExtraLight|UltraLight|Light|Regular|Medium|SemiBold|DemiBold|Bold|ExtraBold|UltraBold|Black|Heavy|Italic|Oblique|Variable|VF)'
+        r'(?:[-_ ]?(?:Italic|Oblique))?$',
         '', name, flags=re.IGNORECASE
     )
     # Replace dashes/underscores with spaces

routes/gallery/gallery_routes.py

@@ -18,6 +18,7 @@ from fastapi import APIRouter, HTTPException, Query, Request
 from core.database import SessionLocal, GalleryImage, GalleryAlbum, ModelEndpoint
 from core.database import Session as DbSession
 from src.auth_helpers import get_current_user, owner_filter, require_privilege
+from src.session_search import escape_like_pattern
 from src.upload_limits import (
     read_upload_limited,
     GALLERY_UPLOAD_MAX_BYTES,
@@ -37,6 +38,24 @@ _SAM_STATE: Dict[str, Any] = {}
 _GROUNDING_STATE: Dict[str, Any] = {}
 
 
+def _compose_cloud_image_edit(source, generated, mask, *, instruction=False):
+    from PIL import Image, ImageChops, ImageOps
+
+    full_frame = instruction or mask.getextrema() == (255, 255)
+    if full_frame:
+        return generated, True
+    # Partial edits must share the source coordinate system, but resizing
+    # width/height independently distorts the subject when the provider changes ratio.
+    fitted = ImageOps.contain(generated, source.size, Image.Resampling.LANCZOS)
+    offset = ((source.width - fitted.width) // 2, (source.height - fitted.height) // 2)
+    aligned = Image.new("RGBA", source.size)
+    aligned.paste(fitted, offset)
+    coverage = Image.new("L", source.size)
+    coverage.paste(255, (offset[0], offset[1], offset[0] + fitted.width, offset[1] + fitted.height))
+    effective_mask = ImageChops.multiply(mask, coverage)
+    return Image.composite(aligned, source, effective_mask), False
+
+
 def _rss_mb() -> float:
     """Current process RSS in MB (from /proc; -1.0 when unavailable)."""
     try:
@@ -305,6 +324,23 @@ def _b64_bytes(value: str, *, label: str = "image") -> bytes:
         raise HTTPException(400, f"Invalid {label}") from exc
 
 
+async def _json_object(request: Request) -> dict:
+    try:
+        data = await request.json()
+    except ValueError:
+        raise HTTPException(400, "Invalid JSON body")
+    if not isinstance(data, dict):
+        raise HTTPException(400, "JSON object body required")
+    return data
+
+
+async def _image_ids(request: Request) -> list:
+    ids = (await _json_object(request)).get("image_ids", [])
+    if not isinstance(ids, list) or not all(isinstance(i, str) for i in ids):
+        raise HTTPException(400, "image_ids must be a list of image ids")
+    return ids
+
+
 def _pil_image_to_b64(img, *, fmt: str = "PNG") -> str:
     buf = io.BytesIO()
     img.save(buf, format=fmt)
@@ -568,6 +604,7 @@ def _is_openai_api_base(url: str) -> bool:
 
 
 _GALLERY_ENDPOINT_PATHS = frozenset({
+    "/chat/completions",
     "/images",
     "/images/edits",
     "/images/generations",
@@ -683,9 +720,15 @@ def _visible_image_endpoint_for_base(db, base: str, owner: str | None, model: st
                     configured.update(v for v in values if isinstance(v, str))
                 except (TypeError, ValueError):
                     pass
+            if model:
                 if model not in configured or not looks_like_image_generation_model(model):
                     continue
-        if _normalize_image_endpoint_base(getattr(ep, "base_url", "")) == target:
+            else:
+                if not any(looks_like_image_generation_model(v) for v in configured):
+                    continue
+        ep_id = str(getattr(ep, "id", "") or "")
+        ep_base = _normalize_image_endpoint_base(getattr(ep, "base_url", ""))
+        if (ep_id and (base == ep_id or target == ep_id)) or (ep_base and ep_base == target):
             if owner and getattr(ep, "owner", None) == owner:
                 return ep
             if fallback is None:
@@ -712,7 +755,7 @@ async def _fetch_result_image_b64(url: str) -> Optional[str]:
     )
     if not ok:
         raise HTTPException(502, f"Upstream returned an unsafe image URL: {reason}")
-    async with httpx.AsyncClient(timeout=60) as c2:
+    async with httpx.AsyncClient(timeout=60, follow_redirects=False) as c2:
         ir = await c2.get(url)
         if ir.status_code == 200:
             return base64.b64encode(ir.content).decode()
@@ -1178,14 +1221,17 @@ def setup_gallery_routes() -> APIRouter:
             )
             q = _owner_filter(q, user)
 
-            # Search filter (prompt + tags + ai_tags)
+            # Per-term AND search: a single `%foo bar%` LIKE only matched the
+            # exact phrase, so "red car" missed "a car that is red".
+            # LIKE wildcards are escaped per token.
             if search:
-                term = f"%{search}%"
                 from sqlalchemy import or_
+                for tok in search.split():
+                    term = f"%{escape_like_pattern(tok)}%"
                     q = q.filter(or_(
-                    GalleryImage.prompt.ilike(term),
+                        GalleryImage.prompt.ilike(term, escape="\\"),
-                    GalleryImage.tags.ilike(term),
+                        GalleryImage.tags.ilike(term, escape="\\"),
-                    GalleryImage.ai_tags.ilike(term),
+                        GalleryImage.ai_tags.ilike(term, escape="\\"),
                     ))
 
             # Tag filter. The UI stacks multiple tag pills by passing them
@@ -1197,9 +1243,10 @@ def setup_gallery_routes() -> APIRouter:
                 for one in (t.strip() for t in tag.split(",")):
                     if not one:
                         continue
+                    safe_tag = f"%{escape_like_pattern(one)}%"
                     q = q.filter(_or(
-                        GalleryImage.tags.ilike(f"%{one}%"),
+                        GalleryImage.tags.ilike(safe_tag, escape="\\"),
-                        GalleryImage.ai_tags.ilike(f"%{one}%"),
+                        GalleryImage.ai_tags.ilike(safe_tag, escape="\\"),
                     ))
 
             # Model filter
@@ -1730,7 +1777,7 @@ def setup_gallery_routes() -> APIRouter:
         requested_base = (body.pop("_endpoint", "") or "").rstrip("/")
         # SSRF hardening: validate a client-supplied endpoint before any
         # outbound request (mirrors routes/embedding_routes.py).
-        if requested_base:
+        if requested_base and "://" in requested_base:
             from src.url_safety import check_outbound_url
             ok, reason = check_outbound_url(
                 requested_base,
@@ -1739,6 +1786,17 @@ def setup_gallery_routes() -> APIRouter:
             if not ok:
                 raise HTTPException(400, f"Rejected endpoint URL: {reason}")
         chosen_model = (body.pop("_model", "") or "").strip()
+        if body.pop("_use_image_settings", False) is True:
+            from src.editor_image_settings import editor_image_models
+            from src.ai_interaction import _resolve_model
+            chosen_model = editor_image_models(owner=user)["edit"]
+            if not chosen_model:
+                raise HTTPException(400, "Choose an image edit model in AI settings")
+            try:
+                edit_url, chosen_model, _ = await asyncio.to_thread(_resolve_model, chosen_model, owner=user)
+            except ValueError as exc:
+                raise HTTPException(400, str(exc)) from exc
+            requested_base = edit_url.removesuffix("/chat/completions").rstrip("/")
         # Optional client-supplied id so GET /api/image/edit-progress can poll the
         # provider's step progress for this op. Ignored by providers that do not
         # implement the progress contract.
@@ -1856,13 +1914,29 @@ def setup_gallery_routes() -> APIRouter:
                         mask_png.save(reference_mask, format="PNG")
                         references = [src_buf.getvalue(), reference_mask.getvalue()]
                         payload = {
-                            "model": oa_model, "size": size, "n": 1,
+                            "model": oa_model, "n": 1,
                             "output_format": "png",
                             "prompt": "Edit the first image. The second image is a mask: white marks the region to edit, black marks the region to preserve. Keep composition and framing unchanged. Requested edit: " + str(body.get("prompt", "")),
                             "input_references": [{"type": "image_url", "image_url": {
                                 "url": "data:image/png;base64," + base64.b64encode(value).decode(),
                             }} for value in references],
                         }
+                        from src.image_model_ids import image_aspect_ratio
+                        payload["aspect_ratio"] = image_aspect_ratio(source_png.width, source_png.height)
+                        from src.ai_interaction import _uses_openrouter_chat_images
+                        if _uses_openrouter_chat_images(oa_model):
+                            payload = {
+                                "model": oa_model,
+                                "messages": [{"role": "user", "content": [
+                                    {"type": "text", "text": payload["prompt"]},
+                                    *[{"type": "image_url", "image_url": ref["image_url"]}
+                                      for ref in payload["input_references"]],
+                                ]}],
+                                "modalities": ["image", "text"], "stream": False,
+                                "image_config": {"aspect_ratio": payload["aspect_ratio"]},
+                            }
+                            r = await client.post(_join_checked_gallery_endpoint(base, "/chat/completions"), headers=headers, json=payload)
+                        else:
                             r = await client.post(_join_checked_gallery_endpoint(base, "/images"), headers=headers, json=payload)
                     else:
                         r = await client.post(_join_checked_gallery_endpoint(base, "/images/edits"), headers=headers, data=data, files=files)
@@ -1879,6 +1953,12 @@ def setup_gallery_routes() -> APIRouter:
                             pass
                         raise HTTPException(r.status_code, f"{_detail} (provider HTTP {r.status_code})")
                     result = r.json()
+                    from src.ai_interaction import _chat_image_results, _image_response_error
+                    provider_error = _image_response_error(result, oa_model)
+                    if provider_error:
+                        raise HTTPException(400 if "deprecated" in provider_error.lower() else 502, provider_error)
+                    if is_openrouter and _uses_openrouter_chat_images(oa_model):
+                        result = {"data": _chat_image_results(result)}
                     raw_b64 = None
                     if result.get("data"):
                         item = result["data"][0]
@@ -1898,17 +1978,13 @@ def setup_gallery_routes() -> APIRouter:
                     # the masked region actually changes.
                     try:
                         generated = Image.open(io.BytesIO(base64.b64decode(raw_b64))).convert("RGBA")
-                        # Match the generated image to the source dims.
+                        blended, full_frame = _compose_cloud_image_edit(
-                        if generated.size != source_png.size:
+                            source_png, generated, mask_png,
-                            generated = generated.resize(source_png.size, Image.LANCZOS)
+                            instruction=body.get("mode") == "instruction",
-                        # mask_png: white = regenerate (use generated),
+                        )
-                        #           black = keep (use source).
-                        # Composite: result = source * (1 - mask_norm) + generated * mask_norm
-                        # Image.composite does exactly that with `mask`.
-                        blended = Image.composite(generated, source_png, mask_png)
                         out_buf = io.BytesIO()
                         blended.save(out_buf, format="PNG")
-                        return {"image": base64.b64encode(out_buf.getvalue()).decode(), "model": oa_model}
+                        return {"image": base64.b64encode(out_buf.getvalue()).decode(), "model": oa_model, "full_frame": full_frame}
                     except Exception as comp_err:
                         # Never return a full-image edit when masking fails.
                         logger.warning(f"Inpaint compose failed: {comp_err}")
@@ -2076,7 +2152,7 @@ def setup_gallery_routes() -> APIRouter:
         runs off the event loop (large images).
         """
         requested_base = (body.pop("_endpoint", "") or "").rstrip("/")
-        if requested_base:
+        if requested_base and "://" in requested_base:
             from src.url_safety import check_outbound_url
             ok, reason = check_outbound_url(
                 requested_base,
@@ -2169,26 +2245,38 @@ def setup_gallery_routes() -> APIRouter:
         try:
             ep = None
             if endpoint_id:
-                ep = owner_filter(
+                ep = _visible_image_endpoint_for_base(db, endpoint_id, user)
-                    db.query(ModelEndpoint).filter(ModelEndpoint.id == endpoint_id),
+                if ep is None:
-                    ModelEndpoint, user,
+                    return {"available": False, "reason": "endpoint not found or disabled"}
-                ).first()
+            elif requested_base:
-            if ep is None and requested_base:
                 ep = _visible_image_endpoint_for_base(db, requested_base, user)
                 if ep is None:
+                    return {"available": False, "reason": "endpoint not found or disabled"}
+            else:
                 ep = _first_visible_image_endpoint(db, user)
             if ep is None:
                 return {"available": False, "reason": "no image endpoint configured"}
             base = ep.base_url.rstrip("/")
         finally:
             db.close()
+
+        from src.url_safety import check_outbound_url
+        block_private = os.getenv("IMAGE_BLOCK_PRIVATE_IPS", "false").lower() == "true"
+        ok, reason = check_outbound_url(base, block_private=block_private)
+        if not ok:
+            # CodeQL #1203: Never expose internal addresses, resolver errors or raw
+            # exception details to clients or logs. Use a stable diagnostic message
+            # without interpolating untrusted exception content or destination strings.
+            logger.warning("edit-progress endpoint rejected: outbound URL policy check failed")
+            return {"available": False, "reason": "endpoint URL rejected"}
+
         if not base.endswith("/v1"):
             base += "/v1"
         # Dynamic path segment (validated token above), so not routed through
         # _join_checked_gallery_endpoint's exact-match whitelist.
         url = f"{base}/images/progress/{request_id}"
         try:
-            async with httpx.AsyncClient(timeout=10) as client:
+            async with httpx.AsyncClient(timeout=10, follow_redirects=False) as client:
                 r = await client.get(url)
             if r.status_code == 200:
                 data = r.json()
@@ -2226,7 +2314,7 @@ def setup_gallery_routes() -> APIRouter:
         # below, so validate it first (mirrors routes/embedding_routes.py).
         # Local-first means loopback/LAN is allowed by default; the cloud
         # metadata range and non-HTTP(S) schemes are always rejected.
-        if requested_base:
+        if requested_base and "://" in requested_base:
             from src.url_safety import check_outbound_url
             ok, reason = check_outbound_url(
                 requested_base,
@@ -2253,7 +2341,7 @@ def setup_gallery_routes() -> APIRouter:
             # so the outbound URL never depends directly on request-body input.
             db = SessionLocal()
             try:
-                ep = _visible_image_endpoint_for_base(db, requested_base, user)
+                ep = _visible_image_endpoint_for_base(db, requested_base, user, model)
                 if not ep:
                     raise HTTPException(403, "Choose a registered image endpoint")
                 base = ep.base_url.rstrip("/")
@@ -2415,11 +2503,10 @@ def setup_gallery_routes() -> APIRouter:
         image_b64 = body.get("image")
         amount = body.get("amount", 50) / 100.0
 
-        from PIL import Image, ImageFilter
+        from PIL import ImageFilter
         import base64, io
 
-        img_bytes = _b64_bytes(image_b64)
+        img = _b64_to_pil_image(image_b64, mode="RGB")
-        img = Image.open(io.BytesIO(img_bytes)).convert("RGB")
 
         # Unsharp mask: radius=2, percent=amount*200, threshold=3
         sharpened = img.filter(ImageFilter.UnsharpMask(radius=2, percent=int(amount * 200), threshold=3))
@@ -2454,8 +2541,7 @@ def setup_gallery_routes() -> APIRouter:
         except ImportError:
             raise HTTPException(500, "Server missing a required dependency")
         # Decode source image (RGB; Real-ESRGAN doesn't preserve alpha).
-        img_bytes = _b64_bytes(image_b64)
+        src = _b64_to_pil_image(image_b64, mode="RGB")
-        src = Image.open(io.BytesIO(img_bytes)).convert("RGB")
         try:
             patch_realesrgan_torchvision_compat()
             from realesrgan import RealESRGANer
@@ -2507,8 +2593,7 @@ def setup_gallery_routes() -> APIRouter:
             import numpy as np
         except ImportError:
             raise HTTPException(500, "Server missing a required dependency")
-        img_bytes = _b64_bytes(image_b64)
+        src = _b64_to_pil_image(image_b64, mode="RGB")
-        src = Image.open(io.BytesIO(img_bytes)).convert("RGB")
         try:
             patch_realesrgan_torchvision_compat()
             from basicsr.archs.rrdbnet_arch import RRDBNet
@@ -2556,7 +2641,9 @@ def setup_gallery_routes() -> APIRouter:
             height = int((body or {}).get("height") or 0)
         except (TypeError, ValueError):
             width = height = 0
-        if height > width:
+        if width > 0 and height > 0:
+            size = f"{width}x{height}"
+        elif height > width:
             size = "1024x1536"
         elif width > height:
             size = "1536x1024"
@@ -2569,6 +2656,9 @@ def setup_gallery_routes() -> APIRouter:
         # The editor's model picker may choose a specific (cloud) image model;
         # forward it so settings defaults only apply when none is selected.
         model_choice = str((body or {}).get("model") or "").strip()
+        if body.get("_use_image_settings") is True:
+            from src.editor_image_settings import editor_image_models
+            model_choice = editor_image_models(owner=user)["generation"]
         payload = {"prompt": prompt, "size": size}
         if model_choice:
             payload["model"] = model_choice
@@ -2760,8 +2850,7 @@ def setup_gallery_routes() -> APIRouter:
         from PIL import Image
         import base64, io
 
-        img_bytes = _b64_bytes(image_b64)
+        img = _b64_to_pil_image(image_b64)
-        img = Image.open(io.BytesIO(img_bytes)).convert("RGBA")
         W, H = img.size
 
         hint = None
@@ -2850,8 +2939,7 @@ def setup_gallery_routes() -> APIRouter:
         from PIL import Image, ImageFilter, ImageEnhance
         import numpy as np
 
-        img_bytes = _b64_bytes(image_b64)
+        img = _b64_to_pil_image(image_b64, mode="RGB")
-        img = Image.open(io.BytesIO(img_bytes)).convert("RGB")
 
         # Try GFPGAN first (AI face restoration)
         try:
@@ -2926,14 +3014,22 @@ def setup_gallery_routes() -> APIRouter:
     @router.put("/api/gallery/albums/{album_id}")
     async def update_album(request: Request, album_id: str):
         user = get_current_user(request)
-        data = await request.json()
+        data = await _json_object(request)
+        name = data.get("name")
+        if name is not None:
+            if not isinstance(name, str) or not name.strip():
+                raise HTTPException(400, "Album name required")
+            name = name.strip()
+        description = data.get("description")
+        if description is not None and not isinstance(description, str):
+            raise HTTPException(400, "Album description must be a string")
         db = SessionLocal()
         try:
             album = _get_or_404_album(db, album_id, user)
-            if data.get("name") is not None:
+            if name is not None:
-                album.name = data["name"]
+                album.name = name
-            if data.get("description") is not None:
+            if description is not None:
-                album.description = data["description"]
+                album.description = description
             if data.get("cover_id") is not None:
                 cover_id = data["cover_id"] or None
                 if cover_id:
@@ -2963,8 +3059,7 @@ def setup_gallery_routes() -> APIRouter:
     @router.post("/api/gallery/albums/{album_id}/add")
     async def add_to_album(request: Request, album_id: str):
         user = get_current_user(request)
-        data = await request.json()
+        ids = await _image_ids(request)
-        ids = data.get("image_ids", [])
         db = SessionLocal()
         try:
             _get_or_404_album(db, album_id, user)
@@ -2981,8 +3076,7 @@ def setup_gallery_routes() -> APIRouter:
     @router.post("/api/gallery/albums/{album_id}/remove")
     async def remove_from_album(request: Request, album_id: str):
         user = get_current_user(request)
-        data = await request.json()
+        ids = await _image_ids(request)
-        ids = data.get("image_ids", [])
         db = SessionLocal()
         try:
             _get_or_404_album(db, album_id, user)

routes/history/history_routes.py

@@ -471,10 +471,16 @@ def setup_history_routes(session_manager, upload_handler=None) -> APIRouter:
         _verify_session_owner(request, session_id)
         try:
             body = await request.json()
+            if not isinstance(body, dict):
+                raise HTTPException(400, "Request body must be a JSON object")
             role = body.get("role", "assistant")
             content = body.get("content", "")
+            if role not in ("user", "assistant"):
+                raise HTTPException(400, "role must be 'user' or 'assistant'")
             if not content:
                 raise HTTPException(400, "content is required")
+            if not isinstance(content, str):
+                raise HTTPException(400, "content must be a string")
             metadata = sanitize_client_message_metadata(body.get("metadata"))
             _reserve_message_uploads(request, content, metadata)
             msg = ChatMessage(role=role, content=content, metadata=metadata)
@@ -825,6 +831,23 @@ def setup_history_routes(session_manager, upload_handler=None) -> APIRouter:
         try:
             body = await request.json()
             keep_count = body.get("keep_count", 0)
+            # A message id is exact; keep_count from the client counts only the
+            # messages loaded in the page. before_msg_id forks before that
+            # message, through_msg_id includes it.
+            before_msg_id = str(body.get("before_msg_id") or "").strip()
+            through_msg_id = str(body.get("through_msg_id") or "").strip()
+            if before_msg_id or through_msg_id:
+                db = SessionLocal()
+                try:
+                    all_db_messages = db.query(DbChatMessage).filter(
+                        DbChatMessage.session_id == session_id
+                    ).order_by(DbChatMessage.timestamp).all()
+                finally:
+                    db.close()
+                resolved = _keep_count_before_message(all_db_messages, before_msg_id or through_msg_id)
+                if resolved is None:
+                    raise HTTPException(404, "Message not found")
+                keep_count = resolved + (0 if before_msg_id else 1)
 
             # Get the source session. keep_count indexes into source.history,
             # so this must go through get_session — reading the cache directly

routes/mcp/mcp_routes.py

@@ -5,6 +5,7 @@ import os
 import uuid
 import urllib.parse
 import html
+import secrets
 from pathlib import Path
 from fastapi import APIRouter, Form, HTTPException, Request
 from fastapi.responses import RedirectResponse, HTMLResponse
@@ -12,7 +13,7 @@ import logging
 import httpx
 
 from core.database import McpServer, SessionLocal
-from core.middleware import require_admin
+from core.middleware import INTERNAL_TOOL_HEADER, INTERNAL_TOOL_TOKEN, require_admin
 from src.constants import DATA_DIR, MCP_OAUTH_DIR
 from src.mcp_manager import McpManager
 
@@ -424,6 +425,11 @@ def setup_mcp_routes(mcp_manager: McpManager):
         Expects JSON body: {"disabled": ["tool_name_1", "tool_name_2"]}
         """
         require_admin(request)
+        internal_token = request.headers.get(INTERNAL_TOOL_HEADER)
+        if internal_token and secrets.compare_digest(internal_token, INTERNAL_TOOL_TOKEN):
+            # Same boundary as POST /api/tools: an agent turn must not be able
+            # to re-enable MCP tools an admin switched off.
+            raise HTTPException(403, "Change MCP tool policy in Settings > Integrations, not from chat")
         db = SessionLocal()
         try:
             srv = db.query(McpServer).filter(McpServer.id == server_id).first()

routes/model_routes.py

@@ -7,6 +7,7 @@ import json
 import hashlib
 import ipaddress
 import socket
+import secrets
 import time as _time
 import logging
 import threading
@@ -417,6 +418,7 @@ _HOST_TO_CURATED = (
     ("nvidia.com", "nvidia"),
     ("openrouter.ai", "openrouter"),
     ("ollama.com", "ollama"),
+    ("conifer.build", "openrouter"),
 )
 
 
@@ -658,7 +660,7 @@ _NON_CHAT_CONTAINS = (
     "gliner", "riva", "-parse", "-embedqa", "-nemoretriever",
     "topic-control", "calibration",
     "ai-synthetic-video", "cosmos-reason2",
-    "bge", "llama-guard",
+    "bge", "llama-guard", "-embed", "rerank",
 )
 _NON_CHAT_EXACT_PREFIXES = (
     "gpt-audio",  # gpt-audio, gpt-audio-mini etc. (not gpt-4o-audio-preview which is chat)
@@ -2845,11 +2847,12 @@ def setup_model_routes(model_discovery):
                 return {"endpoint_id": "", "endpoint_url": "", "model": ""}
             base = _normalize_base(ep.base_url)
             chat_url = build_chat_url(base)
+            if model and not _is_chat_model(model):
+                model = ""
             if not model and (getattr(ep, "cached_models", None) or getattr(ep, "pinned_models", None)):
                 try:
                     visible = _visible_models(ep.cached_models, getattr(ep, "hidden_models", None), getattr(ep, "pinned_models", None))
-                    if visible:
+                    model = next((mid for mid in visible if _is_chat_model(mid)), "")
-                        model = visible[0]
                 except Exception:
                     pass
             return {"endpoint_id": ep.id, "endpoint_url": chat_url, "model": model}
@@ -3042,19 +3045,43 @@ def setup_model_routes(model_discovery):
         return {"dependents": _settings_using_endpoint(ep_id)}
 
     @router.delete("/model-endpoints/{ep_id}")
-    def delete_model_endpoint(ep_id: str, request: Request):
+    def delete_model_endpoint(
+        ep_id: str,
+        request: Request,
+        expected_cookbook_serve_session_id: Optional[str] = None,
+    ):
         require_admin(request)
         db = SessionLocal()
         try:
             ep = db.query(ModelEndpoint).filter(ModelEndpoint.id == ep_id).first()
             if not ep or not _chatgpt_endpoint_visible(ep, request):
                 raise HTTPException(404, "Endpoint not found")
+            if expected_cookbook_serve_session_id is not None:
+                # The DELETE itself checks ownership. An old serve must never
+                # mutate the settings of an endpoint reused by another serve.
+                if (
+                    not ep_id.startswith(("local-", "img-"))
+                    or not expected_cookbook_serve_session_id
+                ):
+                    raise HTTPException(409, "Cookbook endpoint ownership mismatch")
+                deleted = (
+                    db.query(ModelEndpoint)
+                    .filter(
+                        ModelEndpoint.id == ep_id,
+                        ModelEndpoint.cookbook_serve_session_id
+                        == expected_cookbook_serve_session_id,
+                    )
+                    .delete(synchronize_session=False)
+                )
+                if not deleted:
+                    raise HTTPException(409, "Cookbook endpoint ownership changed")
             # Clean up any settings that reference this endpoint
             cleared = _clear_settings_for_endpoint(ep_id)
             cleared_user_preferences = _clear_user_prefs_for_endpoint(ep_id)
             cleared_sessions = _clear_sessions_for_endpoint(db, ep.base_url, ep.id, ep.owner)
             cleared_loaded_sessions = _clear_loaded_sessions_for_endpoint(ep.base_url, ep.id, ep.owner)
             auth_id = getattr(ep, "provider_auth_id", None)
+            if expected_cookbook_serve_session_id is None:
                 db.delete(ep)
             cleared_provider_auth = _delete_orphaned_provider_auth(db, auth_id, exclude_ep_id=ep_id)
             db.commit()
@@ -3091,6 +3118,12 @@ def setup_model_routes(model_discovery):
     def update_tools(body: ToolsUpdate, request: Request):
         """Update which tools are disabled."""
         require_admin(request)
+        from core.middleware import INTERNAL_TOOL_HEADER, INTERNAL_TOOL_TOKEN
+        internal_token = request.headers.get(INTERNAL_TOOL_HEADER)
+        if internal_token and secrets.compare_digest(internal_token, INTERNAL_TOOL_TOKEN):
+            # Owner impersonation can carry an admin username, so check the
+            # authenticated transport rather than request.state.current_user.
+            raise HTTPException(403, "Change tool policy in Settings > Agent Tools, not from chat")
         settings = _load_settings()
         settings["disabled_tools"] = body.disabled
         _save_settings(settings)

routes/note/note_routes.py

@@ -728,32 +728,35 @@ def setup_note_routes(task_scheduler=None, upload_handler=None):
             )
             if body.title is not None:
                 note.title = body.title
-            if body.content is not None:
+            # Omitted fields preserve their values; explicit null clears the
+            # nullable fields used by the editor's removal controls.
+            supplied = body.model_fields_set
+            if "content" in supplied:
                 note.content = body.content
-            if body.items is not None:
+            if "items" in supplied:
-                note.items = json.dumps(body.items)
+                note.items = json.dumps(body.items) if body.items is not None else None
                 flag_modified(note, "items")
             if body.note_type is not None:
                 note.note_type = body.note_type
-            if body.color is not None:
+            if "color" in supplied:
                 note.color = body.color
-            if body.label is not None:
+            if "label" in supplied:
                 note.label = body.label
             if body.pinned is not None:
                 note.pinned = body.pinned
             if body.archived is not None:
                 note.archived = body.archived
-            if body.due_date is not None:
+            if "due_date" in supplied:
                 note.due_date = body.due_date
-            if body.image_url is not None:
+            if "image_url" in supplied:
                 note.image_url = body.image_url
-            if body.gallery_id is not None:
+            if "gallery_id" in supplied:
                 note.gallery_id = body.gallery_id
             if body.repeat is not None:
                 note.repeat = body.repeat
             if body.sort_order is not None:
                 note.sort_order = body.sort_order
-            if body.agent_session_id is not None:
+            if "agent_session_id" in supplied:
                 note.agent_session_id = body.agent_session_id
 
             db.commit()

routes/prefs_routes.py

@@ -118,8 +118,12 @@ def setup_prefs_routes():
     async def set_pref(request: Request, key: str, body: dict):
         user = get_current_user(request)
         prefs = _load_for_user(user)
-        prefs[key] = body.get("value")
+        val = body.get("value")
+        if val is None:
+            prefs.pop(key, None)
+        else:
+            prefs[key] = val
         _save_for_user(user, prefs)
-        return {"key": key, "value": prefs[key]}
+        return {"key": key, "value": prefs.get(key)}
 
     return router

routes/research/research_routes.py

@@ -183,6 +183,7 @@ def _resolve_endpoint_runtime(ep, owner=None, model: Optional[str] = None):
     OAuth tokens in ProviderAuthSession, so ep.api_key is intentionally empty.
     """
     from src.endpoint_resolver import (
+        _endpoint_enabled_models,
         build_chat_url,
         build_headers,
         resolve_endpoint_runtime as resolve_model_endpoint_runtime,
@@ -197,7 +198,7 @@ def _resolve_endpoint_runtime(ep, owner=None, model: Optional[str] = None):
     ep_model = (model or "").strip()
     if not ep_model:
         try:
-            models = json.loads(ep.cached_models) if ep.cached_models else []
+            models = _endpoint_enabled_models(ep)
             if models:
                 ep_model = _first_chat_model(models)
         except Exception:
@@ -824,23 +825,13 @@ def setup_research_routes(research_handler, session_manager=None) -> APIRouter:
         if not ep_url or not ep_model:
             # Last resort: this user's enabled endpoint, plus legacy shared rows.
             from src.database import SessionLocal
-            from src.endpoint_resolver import normalize_base, build_chat_url, build_headers
             db = SessionLocal()
             try:
                 ep = _owned_enabled_endpoint(db, user)
                 if ep:
-                    base = normalize_base(ep.base_url)
+                    resolved = _resolve_endpoint_runtime(ep, owner=user)
-                    fallback_url = build_chat_url(base)
+                    if resolved:
-                    fallback_headers = build_headers(ep.api_key, base)
+                        _merge(*resolved)
-                    fallback_model = ""
-                    if ep.cached_models:
-                        try:
-                            models = json.loads(ep.cached_models)
-                            if models:
-                                fallback_model = _first_chat_model(models)
-                        except Exception:
-                            pass
-                    _merge(fallback_url, fallback_model, fallback_headers)
             finally:
                 db.close()
 

routes/session_routes.py

@@ -1315,7 +1315,7 @@ def setup_session_routes(
     def auto_sort_sessions(request: Request, skip_llm: bool = False):
         """Use AI to categorize all sessions into folders.
 
-        Phase 1 deletes empty/throwaway sessions and Phase 2 asks the LLM
+        Phase 1 deletes empty/Incognito sessions and Phase 2 asks the LLM
         to assign folders. When `skip_llm=true` the endpoint returns
         after Phase 1 — used by the "Tidy (no AI)" UI affordance so
         users can clean junk without spending tokens.
@@ -1326,21 +1326,11 @@ def setup_session_routes(
         single_user_mode = not user and _auth_disabled()
         user_sessions = session_manager.get_sessions_for_user(user)
 
-        # Delete empty and throwaway sessions before sorting
+        # Saved conversations are never disposable based on their title or prompt.
         from core.database import ChatMessage as DbMsg
         db = SessionLocal()
         deleted_empty = 0
         deleted_throwaway = 0
-        # Names that indicate a throwaway/test session (case-insensitive exact or prefix match)
-        _THROWAWAY_NAMES = {
-            "test", "testing", "asdf", "asd", "hello", "hi", "hey",
-            "yo", "sup", "hola", "hii", "hiii", "heyo",
-            "foo", "bar", "baz", "tmp", "temp", "scratch", "untitled",
-            "new chat", "delete", "remove", "junk", "trash", "xxx",
-            "abc", "qwerty", "blah", "stuff", "whatever", "idk",
-            "ok", "lol", "bruh", "hmm", "hm", "meh",
-        }
-        _THROWAWAY_MAX_MESSAGES = 4  # only delete if <= this many messages
         try:
             rows_q = db.query(DbSession).filter(DbSession.archived == False)
             if user:
@@ -1349,15 +1339,11 @@ def setup_session_routes(
                 rows_q = rows_q.filter(DbSession.owner == user)
             rows = rows_q.limit(2000).all()
             folder_map = {r.id: r.folder for r in rows}
-            # Precompute per-session message counts in TWO aggregate queries
+            # Precompute per-session message counts in one aggregate query
             # instead of 1–3 queries PER session — with many chats the per-row
             # loop was doing thousands of round-trips and blowing the timeout.
             from sqlalchemy import func as _sa_func
             _counts = dict(db.query(DbMsg.session_id, _sa_func.count(DbMsg.id)).group_by(DbMsg.session_id).all())
-            _asst_counts = dict(
-                db.query(DbMsg.session_id, _sa_func.count(DbMsg.id))
-                .filter(DbMsg.role == "assistant").group_by(DbMsg.session_id).all()
-            )
             cleanup_now = utcnow_naive()
             for row in rows:
                 # Never delete important sessions
@@ -1367,9 +1353,10 @@ def setup_session_routes(
                 if (row.name or "").strip() == "Incognito":
                     should_delete = True
                     deleted_throwaway += 1
-                    db.delete(row)
                     if hasattr(session_manager, 'delete_session'):
                         session_manager.delete_session(row.id)
+                    else:
+                        db.delete(row)
                     continue
                 if is_session_recently_active(row, now=cleanup_now):
                     continue
@@ -1378,31 +1365,11 @@ def setup_session_routes(
                 if msg_count == 0:
                     should_delete = True
                     deleted_empty += 1
-                elif msg_count <= _THROWAWAY_MAX_MESSAGES:
-                    name = (row.name or "").strip().lower()
-                    # Check first user message content (AI renames sessions, so
-                    # "hi" becomes "Casual Greeting Exchange" — name alone won't match)
-                    first_msg = db.query(DbMsg.content).filter(
-                        DbMsg.session_id == row.id, DbMsg.role == "user"
-                    ).order_by(DbMsg.timestamp).first()
-                    first_text = (first_msg[0] or "").strip().lower() if first_msg else ""
-                    # Count assistant messages — if user sent something but AI never replied, it's dead
-                    assistant_count = _asst_counts.get(row.id, 0)
-                    if name in _THROWAWAY_NAMES or name.startswith("chat:") or first_text in _THROWAWAY_NAMES:
-                        should_delete = True
-                        deleted_throwaway += 1
-                    # Single user message with no AI response = dead session
-                    elif msg_count == 1 and assistant_count == 0:
-                        should_delete = True
-                        deleted_throwaway += 1
-                    # Short phrase (1-3 words) with no real AI conversation (<=2 msgs)
-                    elif msg_count <= 2 and first_text and len(first_text.split()) <= 3 and len(first_text) <= 40:
-                        should_delete = True
-                        deleted_throwaway += 1
                 if should_delete:
-                    db.delete(row)
                     if hasattr(session_manager, 'delete_session'):
                         session_manager.delete_session(row.id)
+                    else:
+                        db.delete(row)
             if deleted_empty or deleted_throwaway:
                 db.commit()
                 logger.info(f"Auto-sort: deleted {deleted_empty} empty + {deleted_throwaway} throwaway sessions")

routes/skills_routes.py

@@ -15,6 +15,7 @@ import httpx
 from fastapi import APIRouter, HTTPException, Request
 from pydantic import BaseModel, Field
 
+from services.memory.skill_format import slugify
 from services.memory.skills import SkillsManager
 from src.auth_helpers import get_current_user
 from services.skills import evaluation as _evaluation
@@ -836,6 +837,10 @@ def setup_skills_routes(skills_manager: SkillsManager) -> APIRouter:
         updates = body.dict(exclude_none=True)
         if not updates:
             return {"ok": True}
+        if "name" in updates:
+            new_name = slugify(updates["name"] or match.get("name"))
+            if new_name != match.get("name") and skills_manager.has_skill(new_name, owner=user):
+                raise HTTPException(409, f"A skill named '{new_name}' already exists")
         ok = skills_manager.update_skill(match.get("name"), updates, owner=user)
         if not ok:
             raise HTTPException(404, "Skill not found")

routes/task/task_routes.py

@@ -20,7 +20,7 @@ from src.task_action_policy import (
     is_admin_only_task_action,
     owner_has_admin_task_privileges,
 )
-from src.task_scheduler import compute_next_run, HOUSEKEEPING_DEFAULTS
+from src.task_scheduler import compute_next_run, is_valid_cron, HOUSEKEEPING_DEFAULTS
 from routes.prefs_routes import _load_for_user, _save_for_user
 
 logger = logging.getLogger(__name__)
@@ -289,15 +289,15 @@ def _resolve_run_endpoint(db, task: ScheduledTask, run: TaskRun) -> str:
 
     try:
         from core.database import ModelEndpoint
-        eps = db.query(ModelEndpoint).filter(ModelEndpoint.is_enabled == True).all()
+        from src.auth_helpers import owner_filter
+        from src.endpoint_resolver import _endpoint_enabled_models
+
+        eps = owner_filter(
+            db.query(ModelEndpoint).filter(ModelEndpoint.is_enabled == True),
+            ModelEndpoint, getattr(task, "owner", None),
+        ).all()
         for ep in eps:
-            cached = []
+            if model in _endpoint_enabled_models(ep):
-            if ep.cached_models:
-                try:
-                    cached = json.loads(ep.cached_models) or []
-                except Exception:
-                    cached = []
-            if model in cached:
                 return ep.base_url or ""
     except Exception:
         pass
@@ -340,6 +340,7 @@ def setup_task_routes(task_scheduler) -> APIRouter:
                 max_tokens=20,
                 headers=headers,
                 timeout=15,
+                thinking_mode="off",
             )
             title = result.strip().strip('"\'').strip()
             return title[:60] if title else prompt[:50].strip()
@@ -475,10 +476,7 @@ def setup_task_routes(task_scheduler) -> APIRouter:
         if req.trigger_type == "schedule" and req.schedule == "cron" and not req.cron_expression:
             raise HTTPException(400, "Cron expression is required for cron schedule")
         if req.trigger_type == "schedule" and req.schedule == "cron" and req.cron_expression:
-            try:
+            if not is_valid_cron(req.cron_expression):
-                from croniter import croniter
-                croniter(req.cron_expression)
-            except Exception:
                 raise HTTPException(400, "Invalid cron expression")
         if req.trigger_type == "event" and not req.trigger_event:
             raise HTTPException(400, "Event name is required for event-triggered tasks")
@@ -788,10 +786,7 @@ def setup_task_routes(task_scheduler) -> APIRouter:
                 task.character_id = req.character_id or None
             if req.cron_expression is not None:
                 if req.cron_expression:
-                    try:
+                    if not is_valid_cron(req.cron_expression):
-                        from croniter import croniter
-                        croniter(req.cron_expression)
-                    except Exception:
                         raise HTTPException(400, "Invalid cron expression")
                 task.cron_expression = req.cron_expression or None
 

routes/upload_routes.py

@@ -25,11 +25,12 @@ from src.auth_helpers import effective_user, is_delegated_credential
 from src.owner_identity import auth_disabled, is_request_sentinel_owner
 from src.attachment_refs import attachment_refs_from_metadata
 from src.constants import GENERATED_IMAGES_DIR
-from src.path_confinement import is_inside
+from src.path_confinement import PathEscape, canonical_root, confine, is_inside
 from src.upload_handler import (
     UploadCleanupSafetyError,
     count_recent_uploads,
     extract_upload_ids,
+    secure_filename,
 )
 
 logger = logging.getLogger(__name__)
@@ -153,30 +154,47 @@ def setup_upload_routes(upload_handler):
 
     def _upload_root() -> str:
         from src.constants import UPLOAD_DIR
-        return os.path.realpath(getattr(upload_handler, "upload_dir", UPLOAD_DIR))
+        return canonical_root(getattr(upload_handler, "upload_dir", UPLOAD_DIR))
 
     def _path_inside_upload_dir(path: str) -> bool:
         return is_inside(_upload_root(), path)
 
     def _resolve_upload_path(file_id: str) -> str:
-        from src.constants import UPLOAD_DIR
+        safe_id = secure_filename(os.path.basename(file_id))
-        upload_root = getattr(upload_handler, "upload_dir", UPLOAD_DIR)
+        if not safe_id:
-        direct = os.path.join(upload_root, file_id)
+            raise HTTPException(400, "Invalid file ID")
+        root = _upload_root()
+
+        direct = os.path.normpath(os.path.join(root, safe_id))
         if os.path.lexists(direct):
-            if not _path_inside_upload_dir(direct):
+            try:
+                canonical_direct = confine(root, direct, allow_root=False)
+            except (PathEscape, ValueError):
                 raise HTTPException(403, "Access denied")
-            if os.path.isfile(direct):
+            if os.path.isfile(canonical_direct):
-                return direct
+                return canonical_direct
             raise HTTPException(404, "File not found")
 
-        for root, _dirs, files in os.walk(upload_root, followlinks=False):
+        info = _load_upload_info(safe_id)
-            if file_id not in files:
+        stored_path = (info or {}).get("path")
-                continue
+        if stored_path:
-            path = os.path.join(root, file_id)
+            try:
-            if not _path_inside_upload_dir(path):
+                canonical_stored = confine(root, stored_path, allow_root=False)
+                if os.path.isfile(canonical_stored):
+                    return canonical_stored
+            except (PathEscape, ValueError):
                 raise HTTPException(403, "Access denied")
-            if os.path.isfile(path):
+
-                return path
+        for walk_root, _dirs, files in os.walk(root, followlinks=False):
+            if safe_id not in files:
+                continue
+            cand = os.path.normpath(os.path.join(walk_root, safe_id))
+            try:
+                canonical_cand = confine(root, cand, allow_root=False)
+            except (PathEscape, ValueError):
+                raise HTTPException(403, "Access denied")
+            if os.path.isfile(canonical_cand):
+                return canonical_cand
             raise HTTPException(404, "File not found")
 
         raise HTTPException(404, "File not found")
@@ -435,35 +453,62 @@ def setup_upload_routes(upload_handler):
             logger.error(f"Failed to get upload stats: {e}")
             raise HTTPException(500, "Failed to get upload statistics")
 
-    @router.get("/{file_id}")
+    @router.api_route("/{file_id}", methods=["GET", "HEAD"])
     async def download_file(request: Request, file_id: str, thumb: int = 0):
-        """Serve an uploaded file by its ID. `?thumb=1` returns a small cached
+        """Serve an uploaded file by its ID or probe via HEAD. `?thumb=1` returns
-        JPEG thumbnail for images (used by chat attachment previews) so the
+        a small cached JPEG thumbnail for images (used by chat attachment previews)
-        client isn't downloading the full-resolution photo just to show it tiny."""
+        so the client isn't downloading the full-resolution photo just to show it tiny."""
         if not upload_handler.validate_upload_id(file_id):
             raise HTTPException(400, "Invalid file ID")
+        safe_file_id = secure_filename(os.path.basename(file_id))
         import mimetypes as _mt
         # Look up original filename and owner from uploads.json
-        original_name = file_id
+        original_name = safe_file_id
         # _load_upload_index() tolerates a missing/corrupt uploads.json (it falls
         # back to the .bak sibling, then to {}), so a truncated DB degrades to
         # "no metadata" instead of a 500 from an unhandled JSONDecodeError.
         db = upload_handler._load_upload_index()
-        info = next((fi for fi in db.values() if fi.get("id") == file_id), None)
+        info = next((fi for fi in db.values() if fi.get("id") == safe_file_id), None)
         if info:
-            original_name = info.get("name", file_id)
+            original_name = info.get("name", safe_file_id)
         file_owner = info.get("owner") if info else None
         _verify_upload_access(request, file_owner)
-        path = _resolve_upload_path(file_id)
+        upload_root = _upload_root()
+        path = _resolve_upload_path(safe_file_id)
+        try:
+            path = confine(upload_root, path, allow_root=False)
+        except (PathEscape, ValueError):
+            raise HTTPException(403, "Access denied")
+        if not path.startswith(upload_root + os.sep) or not _path_inside_upload_dir(path):
+            raise HTTPException(403, "Access denied")
         mime = (info or {}).get("mime") or _mt.guess_type(path)[0] or "application/octet-stream"
         from fastapi.responses import FileResponse
         # Downscaled thumbnail for image previews — generated once and cached.
+        # Skip generation during HEAD probe to avoid loading/processing whole image.
         if thumb and mime.startswith("image/"):
+            thumb_dir = os.path.normpath(os.path.join(upload_root, ".thumbs"))
+            thumb_path = os.path.normpath(os.path.join(thumb_dir, safe_file_id + ".jpg"))
+            try:
+                thumb_path = confine(upload_root, thumb_path, allow_root=False)
+            except (PathEscape, ValueError):
+                raise HTTPException(403, "Access denied")
+            if not thumb_path.startswith(thumb_dir + os.sep) or not _path_inside_upload_dir(thumb_path):
+                raise HTTPException(403, "Access denied")
+            if request.method == "HEAD":
+                if (os.path.isfile(thumb_path)
+                        and os.path.getmtime(thumb_path) >= os.path.getmtime(path)):
+                    return FileResponse(thumb_path, media_type="image/jpeg", headers=UPLOAD_RESPONSE_HEADERS)
+                # GET may generate a JPEG or fall back to the original on failure.
+                # Neither representation is known without decoding the image. HEAD
+                # must not advertise the original/stale file's headers as a thumb.
+                from fastapi.responses import Response
+                response = Response(status_code=200, headers=UPLOAD_RESPONSE_HEADERS)
+                del response.headers["content-length"]
+                return response
+            else:
                 try:
                     from PIL import Image, ImageOps
-                thumb_dir = os.path.join(_upload_root(), ".thumbs")
                     os.makedirs(thumb_dir, exist_ok=True)
-                thumb_path = os.path.join(thumb_dir, file_id + ".jpg")
                     if (not os.path.exists(thumb_path)
                             or os.path.getmtime(thumb_path) < os.path.getmtime(path)):
                         im = Image.open(path)
@@ -496,9 +541,10 @@ def setup_upload_routes(upload_handler):
         return next((fi for fi in db.values() if fi.get("id") == file_id), None)
 
     def _vision_cache_path(file_id: str) -> str:
-        cache_dir = os.path.join(_upload_root(), ".vision")
+        safe_id = secure_filename(os.path.basename(file_id))
+        cache_dir = os.path.normpath(os.path.join(_upload_root(), ".vision"))
         os.makedirs(cache_dir, exist_ok=True)
-        return os.path.join(cache_dir, file_id + ".txt")
+        return os.path.normpath(os.path.join(cache_dir, safe_id + ".txt"))
 
     def _sync_gallery_caption_for_upload(info: dict | None, owner: str | None, text: str) -> None:
         """Copy upload OCR/vision text onto the promoted gallery image row."""

scripts/check_brotli.py

@@ -0,0 +1,122 @@
+"""Fail image builds that select an unfixed Debian native Brotli decoder.
+
+Debian Node links libbrotlidec rather than bundling upstream Node's decoder.
+Use Debian's backport revision, not BrotliDecoderVersion() or Python wheel
+metadata, to establish the package floor for CVE-2026-111497:
+https://security-tracker.debian.org/tracker/CVE-2026-111497
+"""
+
+import argparse
+import ctypes
+from pathlib import Path
+import platform
+import re
+import subprocess
+import sys
+
+
+DEBIAN_FIXED = {"12": "1.0.9-2+deb12u1", "13": "1.1.0-2+deb13u1"}
+DECODER_SONAME = "libbrotlidec.so.1"
+
+
+def command(*args):
+    return subprocess.check_output(args, text=True).strip()
+
+
+def check_package():
+    release = platform.freedesktop_os_release()
+    floor = DEBIAN_FIXED.get(release.get("VERSION_ID"))
+    if release.get("ID") != "debian" or floor is None:
+        raise ValueError("Brotli package floor is not defined for this distribution")
+    record = command(
+        "dpkg-query", "-W", "-f=${Status}\n${Version}", "libbrotli1"
+    ).splitlines()
+    if len(record) != 2 or any(not value.strip() for value in record):
+        raise ValueError("invalid libbrotli1 package record: expected status and version")
+    status, version = record
+    if status != "install ok installed":
+        raise ValueError(f"libbrotli1 is not fully installed (status: {status})")
+    if subprocess.run(
+        ["dpkg", "--compare-versions", version, "ge", floor], check=False
+    ).returncode != 0:
+        raise ValueError(f"libbrotli1 {version} does not meet Debian's fixed floor {floor}")
+    owned = {
+        Path(line).resolve()
+        for line in command("dpkg-query", "-L", "libbrotli1").splitlines()
+        if Path(line).name == DECODER_SONAME
+    }
+    if not owned:
+        raise ValueError("libbrotli1 has no packaged native decoder")
+    return version, owned
+
+
+def loaded_path(library):
+    # Inspect the library actually selected by the dynamic loader. An installed
+    # fixed package alone would not catch LD_LIBRARY_PATH or /usr/local shadows.
+    class DlInfo(ctypes.Structure):
+        _fields_ = [("filename", ctypes.c_char_p), ("base", ctypes.c_void_p),
+                    ("symbol", ctypes.c_char_p), ("address", ctypes.c_void_p)]
+
+    dladdr = ctypes.CDLL(None).dladdr
+    dladdr.argtypes = [ctypes.c_void_p, ctypes.POINTER(DlInfo)]
+    dladdr.restype = ctypes.c_int
+    info = DlInfo()
+    if not dladdr(ctypes.cast(library.BrotliDecoderDecompress, ctypes.c_void_p), ctypes.byref(info)):
+        raise ValueError("cannot identify the loaded Brotli decoder")
+    return Path(info.filename.decode()).resolve()
+
+
+def check_decoder(library):
+    # The fix adds error -17. Older decoders return INVALID here, including
+    # upstream 1.1.0 and 1.2.0. Fixed Debian 12 and 13 backports expose the same
+    # marker with different error-string prefixes.
+    # No malformed attacker-controlled stream is needed for this build check.
+    library.BrotliDecoderErrorString.argtypes = [ctypes.c_int]
+    library.BrotliDecoderErrorString.restype = ctypes.c_char_p
+    if library.BrotliDecoderErrorString(-17) not in (b"BLOCK_SWITCH", b"_ERROR_FORMAT_BLOCK_SWITCH"):
+        raise ValueError("native decoder lacks the CVE-2026-111497 BLOCK_SWITCH fix marker")
+
+    library.BrotliDecoderDecompress.argtypes = [
+        ctypes.c_size_t, ctypes.c_void_p, ctypes.POINTER(ctypes.c_size_t), ctypes.c_void_p
+    ]
+    library.BrotliDecoderDecompress.restype = ctypes.c_int
+    # A small, valid uncompressed Brotli stream must still decode correctly.
+    encoded = bytes.fromhex("0b058068656c6c6f20776f726c6403")
+    output = ctypes.create_string_buffer(11)
+    size = ctypes.c_size_t(len(output))
+    result = library.BrotliDecoderDecompress(len(encoded), encoded, ctypes.byref(size), output)
+    if result != 1 or size.value != 11 or output.raw != b"hello world":
+        raise ValueError("native Brotli decoder failed the valid-stream control")
+
+
+def check_node(node, decoder):
+    # ldd inspects the distro executable and its libnode dependency without
+    # executing JavaScript or relying on process.versions.brotli.
+    linked = command("ldd", str(node))
+    match = re.search(r"libbrotlidec\.so\.1\s+=>\s+(/\S+)", linked)
+    if not match or Path(match[1]).resolve() != decoder:
+        raise ValueError(f"{node} does not select the checked Debian Brotli decoder")
+
+
+def main(argv=None):
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--node", type=Path, help="also verify this distro Node executable's decoder")
+    args = parser.parse_args(argv)
+    try:
+        version, owned = check_package()
+        library = ctypes.CDLL(DECODER_SONAME)
+        decoder = loaded_path(library)
+        if decoder not in owned:
+            raise ValueError(f"selected decoder is outside libbrotli1: {decoder}")
+        check_decoder(library)
+        if args.node:
+            check_node(args.node, decoder)
+    except (ValueError, OSError, AttributeError, subprocess.CalledProcessError) as exc:
+        print(f"Brotli acceptance failed: {exc}", file=sys.stderr)
+        return 1
+    print(f"Brotli accepted: Debian libbrotli1 {version}, decoder {decoder}")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

scripts/generate_env_reference.py

@@ -492,6 +492,25 @@ VARIABLE_NOTES: dict[str, tuple[str, str, str]] = {
         "tools that execute code, mutate state or cause external side effects need "
         "a separate approval. Set to 0 to opt out.",
     ),
+    "ODYSSEUS_LOCAL_MCP_READ_TOOLS": (
+        "Agent loop and tool execution", USER,
+        "Security-relevant. Comma-separated exact MCP read-tool names "
+        "on configured stdio MCP servers. Only stdio servers qualify. "
+        "Declarations do not override built-in tool effects. Output remains "
+        "untrusted to prevent privilege escalation, but allows continuation to "
+        "declared write tools on the same stdio server without external taint. "
+        "Stdio transport does not guarantee network isolation. Automatic "
+        "same-store writes carry residual prompt-injection risk.",
+    ),
+    "ODYSSEUS_LOCAL_MCP_WRITE_TOOLS": (
+        "Agent loop and tool execution", USER,
+        "Security-relevant. Comma-separated exact MCP write-tool names "
+        "on configured stdio MCP servers. Only stdio servers qualify. "
+        "Declarations do not override built-in tool effects. Allowed to execute "
+        "after declared local reads on the same stdio server, but blocked once "
+        "external untrusted context enters the run. Authorization must still "
+        "originate from the trusted request grant.",
+    ),
     "ODYSSEUS_MCP_ALLOWED_COMMANDS": (
         "Agent loop and tool execution", USER,
         "Security-relevant. Comma-separated allowlist of MCP launcher basenames the "

scripts/odysseus-docs

@@ -27,6 +27,7 @@ from pathlib import Path
 
 try:
     from core.database import SessionLocal, Document, DocumentVersion
+    from src.session_search import escape_like_pattern
     quiet_logs()
 except ModuleNotFoundError as e:
     sys.stderr.write(f"error: {e}\nhint: run from repo root with venv active.\n")
@@ -139,10 +140,17 @@ def cmd_export(args):
 def cmd_search(args):
     db = SessionLocal()
     try:
-        like = f"%{args.query}%"
+        # Per-term AND search: a single `%foo bar%` LIKE only matched the
-        rows = db.query(Document).filter(
+        # exact phrase, so "machine learning" missed docs containing both
-            (Document.title.ilike(like)) | (Document.current_content.ilike(like))
+        # words apart. LIKE wildcards are escaped per token.
-        ).order_by(Document.updated_at.desc()).limit(args.limit).all()
+        q = db.query(Document)
+        for tok in args.query.split():
+            term = f"%{escape_like_pattern(tok)}%"
+            q = q.filter(
+                Document.title.ilike(term, escape="\\")
+                | Document.current_content.ilike(term, escape="\\")
+            )
+        rows = q.order_by(Document.updated_at.desc()).limit(args.limit).all()
         emit([_serialize(d) for d in rows], args)
     finally:
         db.close()

scripts/odysseus-gallery

@@ -6,6 +6,7 @@ that are awkward at the shell). Filters images by tag, prompt, favorite,
 album, and surfaces EXIF metadata.
 
     odysseus-gallery list --limit 20 --favorites
+    odysseus-gallery list --tag "sunset,beach"  # all tags, manual or AI
     odysseus-gallery show IMAGE_ID
     odysseus-gallery albums | jq -r '.[] | .name'
     odysseus-gallery search "sunset"          # matches prompt + tags
@@ -83,7 +84,13 @@ def cmd_list(args):
                 fail(f"no album named {args.album!r}")
             q = q.filter(GalleryImage.album_id == al.id)
         if args.tag:
-            q = q.filter(GalleryImage.tags.ilike(f"%{args.tag}%"))
+            for tag in (value.strip() for value in args.tag.split(",")):
+                if not tag:
+                    continue
+                q = q.filter(
+                    GalleryImage.tags.ilike(f"%{tag}%") |
+                    GalleryImage.ai_tags.ilike(f"%{tag}%")
+                )
         q = q.order_by(GalleryImage.created_at.desc()).limit(args.limit)
         emit([_serialize_image(i) for i in q.all()], args)
     finally:
@@ -159,7 +166,7 @@ def _build_parser():
     pl.add_argument("--limit", type=int, default=50)
     pl.add_argument("--favorites", action="store_true")
     pl.add_argument("--album", help="filter by album name")
-    pl.add_argument("--tag", help="substring match against tags column")
+    pl.add_argument("--tag", help="comma-separated tags; each matches manual or AI tags by substring")
     pl.set_defaults(func=cmd_list)
 
     psh = sub.add_parser("show", parents=[common])

scripts/odysseus-mail

@@ -43,7 +43,7 @@ import logging
 import os
 import re
 import sys
-from datetime import datetime
+from datetime import datetime, timezone
 from pathlib import Path
 
 # Anchor the path so the script works when invoked from anywhere.
@@ -136,7 +136,7 @@ def cmd_list(args) -> None:
             emit([], args)
             return
         all_uids = data[0].split()
-        # Newest first, limit
+        # Keep a bounded arrival batch; order its decoded rows by Date below.
         uids = list(reversed(all_uids))[: args.limit]
         out = []
         for uid in uids:
@@ -165,10 +165,15 @@ def cmd_list(args) -> None:
                 sender_name, sender_addr = parseaddr(sender)
                 date_raw = msg.get("Date", "")
                 try:
-                    iso = parsedate_to_datetime(date_raw).isoformat() if date_raw else ""
+                    parsed_date = parsedate_to_datetime(date_raw) if date_raw else None
+                    iso = parsed_date.isoformat() if parsed_date else ""
+                    if parsed_date is not None and parsed_date.tzinfo is None:
+                        parsed_date = parsed_date.replace(tzinfo=timezone.utc)
+                    date_epoch = parsed_date.timestamp() if parsed_date else float("-inf")
                 except Exception:
                     iso = ""
-                out.append({
+                    date_epoch = float("-inf")
+                out.append((date_epoch, {
                     "uid": uid.decode(),
                     "date": iso,
                     "from_addr": sender_addr,
@@ -176,10 +181,11 @@ def cmd_list(args) -> None:
                     "subject": subject,
                     "is_read": "\\Seen" in flags,
                     "is_answered": "\\Answered" in flags,
-                })
+                }))
             except Exception as e:
                 sys.stderr.write(f"warn: skipping uid {uid!r}: {e}\n")
-        emit(out, args)
+        out.sort(key=lambda item: item[0], reverse=True)
+        emit([row for _, row in out], args)
 
 
 # ─── read ────────────────────────────────────────────────────────────

scripts/sanitize_trace_artifacts.py

@@ -0,0 +1,336 @@
+#!/usr/bin/env python3
+"""Sanitize Playwright trace archives to strip sensitive credentials and tokens.
+
+Structured, format-aware fail-closed security:
+Replaces naive regex-based replacement with structured JSON / HAR parsing,
+recursive key/value redaction, URL credential scrubbing, DOM snapshot sanitization,
+safe source code masking, and strict fail-closed purge on corruption or error.
+"""
+from __future__ import annotations
+
+import json
+import logging
+import os
+import re
+import sys
+import urllib.parse
+import zipfile
+from typing import Any, Dict, List, Set
+
+logger = logging.getLogger(__name__)
+
+BINARY_IMAGE_EXTENSIONS = ('.png', '.jpg', '.jpeg', '.webp', '.gif', '.ico')
+
+SENSITIVE_KEY_PATTERN = re.compile(
+    r'(?i)(password|passwd|secret|token|api_key|apikey|access_token|refresh_token|private_key|auth_token|client_secret|session|session_id|sid|csrf|xsrf|jwt|credential|authorization)'
+)
+
+SENSITIVE_PARAM_NAMES = frozenset({
+    "token", "session", "auth", "password", "passwd", "key", "secret", "api_key",
+    "apikey", "access_token", "refresh_token", "jwt", "sid", "code", "sig",
+    "signature", "client_secret", "credential"
+})
+
+SENSITIVE_HEADER_NAMES = frozenset({
+    "authorization", "proxy-authorization", "cookie", "set-cookie", "x-api-key",
+    "api-key", "x-auth-token", "x-access-token"
+})
+
+
+def sanitize_url(url: str) -> str:
+    """Sanitize query parameters and authority credentials in URLs."""
+    if not isinstance(url, str) or not url.strip():
+        return url
+    try:
+        parts = urllib.parse.urlsplit(url)
+    except Exception:
+        return url
+
+    netloc = parts.netloc
+    if "@" in netloc:
+        userinfo, host = netloc.split("@", 1)
+        if ":" in userinfo:
+            user, _ = userinfo.split(":", 1)
+            netloc = f"{user}:[REDACTED]@{host}"
+        else:
+            netloc = f"[REDACTED]@{host}"
+
+    query = parts.query
+    if query:
+        q_pairs = urllib.parse.parse_qsl(query, keep_blank_values=True)
+        sanitized_pairs = []
+        for k, v in q_pairs:
+            k_lower = k.lower().replace("-", "_")
+            if k_lower in SENSITIVE_PARAM_NAMES or SENSITIVE_KEY_PATTERN.search(k_lower):
+                sanitized_pairs.append((k, "[REDACTED]"))
+            else:
+                sanitized_pairs.append((k, v))
+        query = urllib.parse.urlencode(sanitized_pairs)
+
+    return urllib.parse.urlunsplit((parts.scheme, netloc, parts.path, query, parts.fragment))
+
+
+def sanitize_body_text(text: str) -> str:
+    """Sanitize request bodies formatted as JSON or URL-encoded form data."""
+    if not isinstance(text, str) or not text.strip():
+        return text
+
+    # 1. Try JSON
+    trimmed = text.strip()
+    if (trimmed.startswith("{") and trimmed.endswith("}")) or (trimmed.startswith("[") and trimmed.endswith("]")):
+        try:
+            parsed = json.loads(trimmed)
+            sanitized = sanitize_structured_obj(parsed)
+            return json.dumps(sanitized, ensure_ascii=False)
+        except Exception:
+            pass
+
+    # 2. Try URL-encoded form data
+    if "=" in trimmed and not trimmed.startswith("<"):
+        try:
+            pairs = urllib.parse.parse_qsl(trimmed, keep_blank_values=True)
+            if pairs and any(k.lower() in SENSITIVE_PARAM_NAMES or SENSITIVE_KEY_PATTERN.search(k) for k, _ in pairs):
+                out_pairs = []
+                for k, v in pairs:
+                    k_lower = k.lower().replace("-", "_")
+                    if k_lower in SENSITIVE_PARAM_NAMES or SENSITIVE_KEY_PATTERN.search(k_lower):
+                        out_pairs.append((k, "[REDACTED]"))
+                    else:
+                        out_pairs.append((k, v))
+                return urllib.parse.urlencode(out_pairs)
+        except Exception:
+            pass
+
+    return text
+
+
+def sanitize_structured_obj(obj: Any) -> Any:
+    """Recursively sanitize Python objects decoded from JSON or HAR payloads."""
+    if isinstance(obj, dict):
+        new_dict: Dict[str, Any] = {}
+
+        # Check for header entry with inverted or standard keys: {"name": ..., "value": ...}
+        name_val = obj.get("name")
+        val_val = obj.get("value")
+        if isinstance(name_val, str) and isinstance(val_val, str):
+            name_lower = name_val.lower().strip()
+            if name_lower in SENSITIVE_HEADER_NAMES or SENSITIVE_KEY_PATTERN.search(name_lower):
+                for k, v in obj.items():
+                    if k == "value":
+                        new_dict[k] = "[REDACTED]"
+                    else:
+                        new_dict[k] = sanitize_structured_obj(v)
+                return new_dict
+
+        # Process standard dictionary keys
+        for k, v in obj.items():
+            k_lower = str(k).lower().replace("-", "_")
+
+            # Action parameters in Playwright traces (e.g. page.fill, page.type)
+            if k == "params" and isinstance(v, dict):
+                sanitized_params = sanitize_structured_obj(v)
+                # If this action involves fill, type, or sensitive inputs, redact value/text
+                if "value" in sanitized_params:
+                    sanitized_params["value"] = "[REDACTED]"
+                if "text" in sanitized_params:
+                    sanitized_params["text"] = "[REDACTED]"
+                new_dict[k] = sanitized_params
+                continue
+
+            # DOM snapshot nodes
+            if k == "attributes" and isinstance(v, (dict, list)):
+                if isinstance(v, dict):
+                    attr_dict = dict(v)
+                    type_val = str(attr_dict.get("type") or "").lower()
+                    if type_val == "password" or "password" in str(attr_dict.get("name") or "").lower():
+                        if "value" in attr_dict:
+                            attr_dict["value"] = "[REDACTED]"
+                    new_dict[k] = attr_dict
+                    continue
+                elif isinstance(v, list):
+                    # List of attribute strings or pairs
+                    new_list = []
+                    is_pw = any("password" in str(x).lower() for x in v)
+                    for x in v:
+                        if is_pw and isinstance(x, str) and not x.startswith("type="):
+                            new_list.append("[REDACTED]")
+                        else:
+                            new_list.append(sanitize_structured_obj(x))
+                    new_dict[k] = new_list
+                    continue
+
+            # Request postData / text
+            if k in ("postData", "postDataText") and isinstance(v, str):
+                new_dict[k] = sanitize_body_text(v)
+                continue
+
+            # URL fields
+            if k == "url" and isinstance(v, str):
+                new_dict[k] = sanitize_url(v)
+                continue
+
+            # Cookie arrays
+            if k == "cookies" and isinstance(v, list):
+                sanitized_cookies = []
+                for c in v:
+                    if isinstance(c, dict):
+                        c_copy = dict(c)
+                        if "value" in c_copy:
+                            c_copy["value"] = "[REDACTED]"
+                        sanitized_cookies.append(c_copy)
+                    else:
+                        sanitized_cookies.append(sanitize_structured_obj(c))
+                new_dict[k] = sanitized_cookies
+                continue
+
+            # Direct sensitive key matches
+            if SENSITIVE_KEY_PATTERN.search(k_lower) and isinstance(v, str):
+                new_dict[k] = "[REDACTED]"
+                continue
+
+            new_dict[k] = sanitize_structured_obj(v)
+
+        return new_dict
+
+    elif isinstance(obj, list):
+        return [sanitize_structured_obj(item) for item in obj]
+
+    elif isinstance(obj, str):
+        # Check for inline credentials/tokens
+        if obj.startswith("Bearer ") or obj.startswith("bearer "):
+            return "Bearer [REDACTED]"
+        if obj.startswith("Basic ") or obj.startswith("basic "):
+            return "Basic [REDACTED]"
+        return obj
+
+    return obj
+
+
+def sanitize_jsonl_content(data: bytes) -> bytes:
+    """Sanitize newline-delimited JSON Lines (.trace, .network, .stacks)."""
+    lines = data.split(b"\n")
+    out_lines = []
+    for line in lines:
+        if not line.strip():
+            out_lines.append(line)
+            continue
+        try:
+            parsed = json.loads(line.decode("utf-8"))
+        except (json.JSONDecodeError, UnicodeDecodeError):
+            # Not valid JSON, keep as-is
+            out_lines.append(line)
+            continue
+        sanitized = sanitize_structured_obj(parsed)
+        out_lines.append(json.dumps(sanitized, ensure_ascii=False).encode("utf-8"))
+    return b"\n".join(out_lines)
+
+
+def sanitize_trace_archive(zp: str) -> bool:
+    """Sanitize a single zip archive in a strict fail-closed manner.
+
+    If the archive is malformed, partial, unreadable, or encounters an exception
+    during processing, the archive is permanently deleted so that raw traces
+    cannot leak.
+    Returns True if successfully sanitized, False if deleted due to error.
+    """
+    if not os.path.exists(zp):
+        return False
+
+    temp_zp = zp + ".tmp"
+    try:
+        if os.path.getsize(zp) == 0:
+            raise zipfile.BadZipFile("Empty archive")
+
+        with zipfile.ZipFile(zp, "r") as zin:
+            infolist = zin.infolist()
+            if not infolist:
+                raise zipfile.BadZipFile("Archive has no entries")
+
+            corrupt_file = zin.testzip()
+            if corrupt_file is not None:
+                raise zipfile.BadZipFile(f"Corrupt entry detected: {corrupt_file}")
+
+            with zipfile.ZipFile(temp_zp, "w", compression=zipfile.ZIP_DEFLATED) as zout:
+                for item in infolist:
+                    filename_lower = item.filename.lower()
+
+                    # 1. Binary image screenshots/diagnostics: preserve as-is
+                    if filename_lower.endswith(BINARY_IMAGE_EXTENSIONS):
+                        content = zin.read(item.filename)
+                        zout.writestr(item, content)
+                        continue
+
+                    # 2. Embedded test source code (e.g. resources/src@... or .js/.ts):
+                    # Mask safely without syntax corruption!
+                    if "src@" in filename_lower or filename_lower.endswith((".js", ".ts", ".mjs", ".py")):
+                        safe_placeholder = b"// [SOURCE CODE REDACTED FOR TRACE CONFIDENTIALITY]\n"
+                        zout.writestr(item, safe_placeholder)
+                        continue
+
+                    # 3. JSON Lines and structured JSON/HAR traces
+                    content = zin.read(item.filename)
+                    if filename_lower.endswith((".trace", ".network", ".stacks")):
+                        sanitized_content = sanitize_jsonl_content(content)
+                        zout.writestr(item, sanitized_content)
+                    elif filename_lower.endswith((".json", ".har")):
+                        try:
+                            parsed = json.loads(content.decode("utf-8"))
+                            sanitized = sanitize_structured_obj(parsed)
+                            zout.writestr(item, json.dumps(sanitized, ensure_ascii=False).encode("utf-8"))
+                        except Exception:
+                            # If invalid JSON, treat as text line-by-line
+                            zout.writestr(item, sanitize_jsonl_content(content))
+                    else:
+                        # Unknown binary or text member: exclude to fail closed
+                        continue
+
+        # Atomic replace
+        os.replace(temp_zp, zp)
+        return True
+
+    except Exception as exc:
+        logger.warning("Sanitizing trace %s failed (%s); removing fail-closed", zp, exc)
+        for path_to_remove in (temp_zp, zp):
+            if os.path.exists(path_to_remove):
+                try:
+                    os.remove(path_to_remove)
+                except OSError as remove_err:
+                    logger.error("Failed to delete unredacted trace %s: %s", path_to_remove, remove_err)
+        return False
+
+
+def sanitize_directory(results_dir: str) -> int:
+    """Walk results_dir and sanitize every .zip archive fail-closed.
+
+    Returns the number of archives processed. Exits with 1 if any archive fails.
+    """
+    if not os.path.exists(results_dir):
+        return 0
+
+    count = 0
+    all_ok = True
+    for root, _, files in os.walk(results_dir):
+        for f in files:
+            if f.endswith(".zip"):
+                zp = os.path.join(root, f)
+                ok = sanitize_trace_archive(zp)
+                if not ok:
+                    all_ok = False
+                count += 1
+
+    if not all_ok:
+        sys.exit(1)
+    return count
+
+
+def main() -> None:
+    if len(sys.argv) < 2:
+        print(f"Usage: {sys.argv[0]} <results_directory>")
+        sys.exit(1)
+
+    target_dir = sys.argv[1]
+    sanitize_directory(target_dir)
+
+
+if __name__ == "__main__":
+    main()

scripts/verify_notes_browser_runtime.py

@@ -11,6 +11,38 @@ import time
 from urllib.request import urlopen
 
 ROOT = Path(__file__).resolve().parents[1]
+if str(ROOT) not in sys.path:
+    sys.path.insert(0, str(ROOT))
+
+
+def _stop_owned_server(server, identity):
+    """Close only identity-bound descendants, including MCP's new sessions.
+
+    Capture before signalling the parent: an orphaned stderr holder can no
+    longer be attributed by ancestry. Never search by executable name or kill
+    the caller's process group.
+    """
+    from src import process_lifecycle as lifecycle
+
+    descendants = (lifecycle.bind_descendants([identity.pid])
+                   if identity.verdict() == lifecycle.OWNED else [])
+    if os.name == "nt":
+        if identity.verdict() == lifecycle.OWNED:
+            lifecycle.taskkill_tree(identity.pid)
+    else:
+        outcome = lifecycle.terminate_identities(
+            [*(entry.identity for entry in reversed(descendants)), identity],
+            steps=lifecycle.term_kill_steps(10),
+        )
+        if outcome.survivors or outcome.unverified:
+            raise RuntimeError(f"Disposable Notes cleanup could not verify exit: {outcome}")
+    try:
+        logs, _ = server.communicate(timeout=2)
+    except subprocess.TimeoutExpired as error:
+        # Do not turn an unidentified pipe holder into an unbounded wait.
+        server.stdout.close()
+        raise RuntimeError("Disposable Notes output pipe remained open after owned cleanup") from error
+    return logs
 
 
 def main():
@@ -36,7 +68,10 @@ def main():
         server = subprocess.Popen(
             [sys.executable, "-m", "uvicorn", "app:app", "--host", "127.0.0.1", "--port", str(port)],
             cwd=target, env=env, stdout=subprocess.PIPE, stderr=subprocess.STDOUT, text=True,
+            start_new_session=(os.name != "nt"),
         )
+        from src.process_lifecycle import ProcessIdentity
+        server_identity = ProcessIdentity.capture(server.pid)
         result = None
         try:
             origin = f"http://127.0.0.1:{port}"
@@ -62,12 +97,7 @@ def main():
                 print(result.stderr, file=sys.stderr)
             return result.returncode
         finally:
-            server.terminate()
+            logs = _stop_owned_server(server, server_identity)
-            try:
-                logs, _ = server.communicate(timeout=10)
-            except subprocess.TimeoutExpired:
-                server.kill()
-                logs, _ = server.communicate()
             if result is None or result.returncode:
                 print(logs[-8000:], file=sys.stderr)
 

services/hwfit/hardware.py

@@ -660,6 +660,8 @@ def _detect_windows():
             "gpu_vram_gb": d.get("gpu_vram_gb"),
             "gpu_count": _as_int(d.get("gpu_count"), 0),
             "backend": d.get("gpu_backend", "cpu_x86"),
+            "gpus": [],
+            "gpu_groups": [],
             "homogeneous": True,
             "gpu_error": None,
             "platform": "windows",
@@ -822,12 +824,14 @@ def detect_system(host="", ssh_port="", platform="", fresh=False):
         if result:
             result = _attach_probe_context(result, host=host)
             _remote_host = None
+            _remote_port = None
             _remote_platform = None
             _cache_by_host[cache_key] = (now, result)
             return result
         # SSH may work while the PowerShell hardware probe still fails.
         result = {"error": f"Windows hardware probe failed for {host}", "host": host}
         _remote_host = None
+        _remote_port = None
         _remote_platform = None
         _cache_by_host[cache_key] = (now, result)
         return result
@@ -852,6 +856,7 @@ def detect_system(host="", ssh_port="", platform="", fresh=False):
         result = {"error": f"Cannot connect to {host}", "host": host}
         _cache_by_host[cache_key] = (now, result)
         _remote_host = None
+        _remote_port = None
         _remote_platform = None
         return result
     available_ram = round(_get_available_ram_gb(), 1)
@@ -893,6 +898,9 @@ def detect_system(host="", ssh_port="", platform="", fresh=False):
             "gpu_name": None,
             "gpu_vram_gb": None,
             "gpu_count": 0,
+            "gpus": [],
+            "gpu_groups": [],
+            "homogeneous": True,
             "backend": backend,
             # Set when nvidia-smi exists but failed (e.g. driver/library
             # version mismatch) — lets the UI say "GPU driver error" instead
@@ -902,6 +910,7 @@ def detect_system(host="", ssh_port="", platform="", fresh=False):
 
     result = _attach_probe_context(result, host=host)
     _remote_host = None
+    _remote_port = None
     _remote_platform = None
     _cache_by_host[cache_key] = (now, result)
     return result

services/hwfit/hf_discovery.py

@@ -202,17 +202,35 @@ def _infer_use_case(repo_id, pipeline_tag):
     return "general"
 
 
+# HF collections mix in models no Cookbook LLM backend can serve (vision
+# encoders, time-series forecasters, image/video generators). Image models
+# have their own registry, so keep these out of the LLM catalog.
+_NON_LLM_PIPELINE_TAGS = frozenset({
+    "time-series-forecasting",
+    "zero-shot-image-classification",
+    "image-classification",
+    "image-feature-extraction",
+    "text-to-image",
+    "image-to-video",
+    "text-to-video",
+})
+
+
 def _entry_from_collection_item(collection, item, source):
     repo_id = item.get("id") or ""
     if item.get("type") != "model" or not repo_id:
         return None
+    if (item.get("pipeline_tag") or "").strip().lower() in _NON_LLM_PIPELINE_TAGS:
+        return None
     repo_prefix = source.get("repo_prefix")
     if repo_prefix and not repo_id.startswith(repo_prefix):
         return None
     raw_params = item.get("numParameters") or 0
-    active = None
+    # Read the name even when HF gives a total: it is the only source of the
+    # MoE active size (e.g. "-35B-A3B"), which the speed estimate depends on.
+    name_params, active = _parse_params_from_name(repo_id)
     if not raw_params:
-        raw_params, active = _parse_params_from_name(repo_id)
+        raw_params = name_params
     param_label, raw_params = _format_params(raw_params)
     if not raw_params:
         return None

services/memory/skill_format.py

@@ -84,10 +84,14 @@ _FM_KEY_RE = re.compile(r"^([a-z_][a-z0-9_]*):\s*(.*)$", re.IGNORECASE)
 _FM_BLOCK_LIST_RE = re.compile(r"^\s*-\s*(.*)$")
 
 
-def _parse_scalar(raw: str) -> Any:
+def _parse_scalar(raw: str, *, string_only: bool = False) -> Any:
     raw = raw.strip()
     if raw == "":
         return ""
+    # String fields keep their spelling; quoted values still use the decoder
+    # below so existing escapes round-trip as before.
+    if string_only and not (raw[0] == raw[-1] and raw[0] in ("'", '"')):
+        return raw
     if raw.startswith("[") and raw.endswith("]"):
         inner = raw[1:-1].strip()
         if not inner:
@@ -126,10 +130,15 @@ def _parse_scalar(raw: str) -> Any:
 def _split_top_level(s: str, sep: str) -> List[str]:
     """Split `s` on `sep` ignoring separators inside [] or quotes."""
     out, buf, depth, quote = [], [], 0, None
+    escaped = False
     for ch in s:
         if quote:
             buf.append(ch)
-            if ch == quote:
+            if escaped:
+                escaped = False
+            elif ch == "\\" and quote == '"':
+                escaped = True
+            elif ch == quote:
                 quote = None
             continue
         if ch in ("'", '"'):
@@ -171,7 +180,7 @@ def parse_frontmatter(text: str) -> tuple[Dict[str, Any], str]:
                 pending_key = key
                 fm[key] = []
             else:
-                fm[key] = _parse_scalar(val)
+                fm[key] = _parse_scalar(val, string_only=(key == "version"))
                 pending_key = None
             continue
         m2 = _FM_BLOCK_LIST_RE.match(line)

services/memory/skills.py

@@ -500,6 +500,19 @@ class SkillsManager:
         sk.path = self._skill_file(cat, nm)
         return sk.to_dict()
 
+    def has_skill(self, name: str, owner: Optional[str] = None) -> bool:
+        """Whether `owner` already has a skill called `name`, in any category.
+
+        Update and delete find a skill by name within its owner, so two of
+        one owner's skills sharing a name would make them act on whichever
+        file is walked first.
+        """
+        for path in self._iter_skill_files():
+            sk = self._read_skill(path)
+            if sk and sk.name == name and (sk.owner or "") == (owner or ""):
+                return True
+        return False
+
     def update_skill(self, skill_id: str, updates: Dict, owner: Optional[str] = None) -> bool:
         """`skill_id` is the slug name. Allows updating any field plus
         renames if `name` changes (file is moved on disk).
@@ -551,6 +564,9 @@ class SkillsManager:
             # Rename
             new_name = slugify(updates.get("name") or sk.name)
             if new_name != sk.name:
+                if self.has_skill(new_name, owner=owner):
+                    logger.warning(f"Skill rename target name in use: {new_name}")
+                    return False
                 sk.name = new_name
 
             # Write to potentially new path

src/action_intents.py

@@ -64,6 +64,21 @@ _SHELL_COMMAND = (
     r"nvcc|docker|systemctl|journalctl|tmux|git)"
 )
 _BENCHMARK_COMMAND = r"(?:[a-z][a-z0-9_-]*bench(?:mark)?s?|bench(?:mark)?s?)"
+# "on <word>" and "from <word>" alone describe people, companies, and topics
+# ("Jensen Huang from NVIDIA", "a book on Rome") as often as hosts, so a target
+# only counts as a computer next to wording that describes work on a machine.
+_NAMED_COMPUTER = (
+    r"(?!this\b|my\b|the\b|a\b|an\b|that\b|it\b|same\b|current\b)"
+    r"[a-z][a-z0-9_.-]{1,31}\b"
+)
+NAMED_COMPUTER_TASK = (
+    r"\b(?:disks?|uptime|processes|logs?|files?|folders?|director(?:y|ies)|running|installed|"
+    r"(?:storage|memory|ram|vram|cpu|gpu)\s+(?:usage|space|utili[sz]ation|load|temp(?:erature)?))\b"
+    rf"[^.?!\n]{{0,80}}\bon\s+{_NAMED_COMPUTER}"
+    r"|\b(?:copy|fetch|get|grab|pull|download|scp|rsync|sync|read|show|list|tail|check)\b"
+    r"[^.?!\n]{0,60}\b(?:files?|logs?|folders?|director(?:y|ies)|configs?|checkpoints?|backups?)\b"
+    rf"[^.?!\n]{{0,40}}\bfrom\s+{_NAMED_COMPUTER}"
+)
 _CODE_ACTION = r"(?:write|create|add|edit|modify|code|program|implement|build)"
 _CODE_ARTIFACT = (
     r"(?:code|function|class|script|module|component|snippet|program|app|feature|file|"
@@ -164,7 +179,7 @@ _ROUTING_PATTERNS: tuple[tuple[str, str, Pattern[str]], ...] = tuple(
         ("workspace", "file/code inspection request", rf"{_PLEASE}(?:find|inspect|look\s+at|open|read|check)\b.{{0,120}}\b(?:file|folder|directory|repo|repository|code|source|logs?|trace|stack|diff)\b"),
         ("workspace", "server/process debugging request", rf"{_PLEASE}(?:check|debug|fix|restart|start|stop|kill|tail|inspect)\b.{{0,120}}\b(?:server|service|process|port|docker|container|tmux|endpoint|logs?)\b"),
         ("workspace", "local computer task request", r"\b(?:on|from|in|using|with)\s+(?:this|my|the)\s+(?:computer|machine|pc|laptop|device|system)\b|\b(?:local|host)\s+(?:computer|machine|files?|system)\b"),
-        ("workspace", "named computer task request", r"\b(?:on|from)\s+(?!this\b|my\b|the\b|a\b|an\b|that\b|it\b|same\b|current\b)(?:[a-z][a-z0-9_.-]{1,31})\b"),
+        ("workspace", "named computer task request", NAMED_COMPUTER_TASK),
         ("workspace", "terminal workspace request", rf"\b(?:terminal|shell|workspace|tmux|docker|container|git|branch|commit|diff|pytest|stacktrace|traceback|{_BENCHMARK_COMMAND}|eval(?:uation)?s?)\b"),
 
         # Shell / remote-host intent.

src/agent_loop.py

@@ -14,12 +14,12 @@ import csv
 import difflib
 import html
 import json
+import logging
 import os
 import re
 import shlex
 import shutil
 import time
-import logging
 import hashlib
 from src.web_recovery import WebRecoveryBudget
 from itertools import count
@@ -44,6 +44,7 @@ from src.model_profiles import (
     is_odysseus_merged_tools_model,
     tool_schema_profile,
 )
+from src.action_intents import NAMED_COMPUTER_TASK
 from src.agent_evidence import (
     EvidenceLedger,
     command_has_mutation_effect,
@@ -2834,7 +2835,7 @@ def _parse_qwen_explicit_admin_request(text: str) -> Optional[tuple[str, str]]:
     if re.search(r"\b(?:disabled\s+tools?|tools?\s+(?:are\s+)?(?:currently\s+)?disabled)\b", q):
         return "manage_settings", json.dumps({"action": "list_tools"})
     if re.search(r"\b(?:turn|enable|re-enable|reenable)\b.{0,40}\b(?:image\s+generation|images?|generate_image)\b", q):
-        return "manage_settings", json.dumps({"action": "enable_tool", "tool": "images"})
+        return "manage_settings", json.dumps({"action": "list_tools"})
     if re.search(r"\b(?:disable|turn\s+off)\b.{0,40}\b(?:image\s+generation|images?|generate_image)\b", q):
         return "manage_settings", json.dumps({"action": "disable_tool", "tool": "images"})
     if listish and re.search(r"\bapi\s+tokens?\b|\btokens?\b", q):
@@ -8057,7 +8058,7 @@ Generate an image. Line 1 = description, line 2 = model name, line 3 = WxH (e.g.
     "manage_tokens": "- ```manage_tokens``` — Generate or revoke API access tokens for external integrations. Args (JSON): {\"action\": \"list|create|delete\", ...}",
     "manage_documents": "- ```manage_documents``` — List, read/open, delete, or tidy documents in the editor panel. Args (JSON): {\"action\": \"list|read|delete|tidy\", ...}. `list` returns rows like `[Title](#document-<id>) — lang, size, updated 5m ago` sorted MOST-RECENT FIRST; the user clicks the anchor to open. `read` (aliases: view/open/get) takes `document_id` and returns the content. When the user asks \"open/show/read my notes\" or \"what documents do I have\", use this — do NOT shell out, do NOT curl.",
     "manage_research": "- ```manage_research``` — List, read/open, or delete saved DEEP RESEARCH results from the Library. Args (JSON): {\"action\": \"list|read|delete\", \"id\": \"<id>\", \"search\": \"...\"}. `list` returns rows like `[query](#research-<id>) — N sources` MOST-RECENT FIRST; the user clicks to open. `read` (aliases: open/view/get) takes `id` and returns the report text + sources. Use when the user says \"open/read/find/delete my research\" or \"that report\". This IS how you read a finished report: when the user refers to a just-completed deep-research job (\"check it out\", \"read that report\", \"summarize the research\") WITHOUT giving an id, call `manage_research` with `action:list` to get the most-recent id, then `action:read` with that id, and answer from the returned text. Do NOT `web_fetch`/`app_api` the `/api/research/report/{id}` URL — that endpoint renders HTML for the browser, not clean text — and do NOT start a fresh `web_search`/`trigger_research` just to read an existing report. To START new research, use trigger_research instead.",
-    "manage_settings": "- ```manage_settings``` — View/change the REAL app settings (same ones the Settings panel writes) AND turn tools on/off. Change a setting: `{\"action\":\"set\",\"key\":\"...\",\"value\":\"...\"}` — keys accept friendly aliases, e.g. voice→tts_voice, \"search engine\"→search_provider, \"default model\"→default_model, \"teacher model\"→teacher_model, \"task/background model\"→task_model, \"image quality\"→image_quality, \"reminder channel\"→reminder_channel (browser|email|ntfy), \"agent timeout\"/\"max tool calls\"/\"token budget\". Read: `{\"action\":\"get\",\"key\":\"...\"}`; see all: `{\"action\":\"list\"}`; reset one: `{\"action\":\"reset\",\"key\":\"...\"}`. Use this when the user asks to change ANY preference instead of making them open Settings. Secrets/API keys are read-only (tell them to set those in the panel). Tool toggles: `{\"action\":\"disable_tool|enable_tool\",\"tool\":\"shell\"}` (aliases: shell/search/browser/documents/memory/skills/images/tasks/notes/calendar/email), list disabled: `{\"action\":\"list_tools\"}`.",
+    "manage_settings": "- ```manage_settings``` — View/change the REAL app settings (same ones the Settings panel writes) AND turn tools on/off. Change a setting: `{\"action\":\"set\",\"key\":\"...\",\"value\":\"...\"}` — keys accept friendly aliases, e.g. voice→tts_voice, \"search engine\"→search_provider, \"default model\"→default_model, \"teacher model\"→teacher_model, \"task/background model\"→task_model, \"image quality\"→image_quality, \"reminder channel\"→reminder_channel (browser|email|ntfy), \"agent timeout\"/\"max tool calls\"/\"token budget\". Read: `{\"action\":\"get\",\"key\":\"...\"}`; see all: `{\"action\":\"list\"}`; reset one: `{\"action\":\"reset\",\"key\":\"...\"}`. Use this when the user asks to change ANY preference instead of making them open Settings. Secrets/API keys are read-only (tell them to set those in the panel). Tool toggles: `{\"action\":\"disable_tool\",\"tool\":\"shell\"}` (aliases: shell/search/browser/documents/memory/skills/images/tasks/notes/calendar/email), list disabled: `{\"action\":\"list_tools\"}`. You cannot re-enable a disabled tool: that is done by the user in Settings > Agent Tools.",
     "manage_notes": """\
 ```manage_notes
 {"action": "add", "title": "<short todo>", "due_date": "<natural language or ISO datetime>"}
@@ -8764,7 +8765,7 @@ _LOCAL_COMPUTER_REFERENCE_RE = re.compile(
     r"\b(?:on|from|in|using|with)\s+(?:this|my|the)\s+(?:computer|machine|pc|laptop|device|system)\b"
     r"|\b(?:this|my|the)\s+(?:computer|machine|pc|laptop|device|system)\b"
     r"|\b(?:local|host)\s+(?:computer|machine|files?|system)\b"
-    r"|\b(?:on|from)\s+(?!this\b|my\b|the\b|a\b|an\b)(?:[a-z][a-z0-9_.-]{1,31})\b",
+    rf"|{NAMED_COMPUTER_TASK}",
     re.IGNORECASE,
 )
 _LOCAL_NETWORK_REFERENCE_RE = re.compile(
@@ -16994,6 +16995,20 @@ def _detect_runaway_call(call_freq, threshold=15):
     sig = next((s for s, n in call_freq.items() if n >= threshold), None)
     return sig.split(":", 1)[0] if sig else None
 
+def _tool_approval_gate_bypassed(
+    exact_approval: Optional[ExactToolApproval] = None,
+) -> bool:
+    """Return whether the post-external-context tool approval gate is bypassed.
+
+    ``tool_approval_mode=off`` (admin setting, Settings → Agent Tools) bypasses
+    the gate on every run, the same effect as ``ODYSSEUS_TOOL_APPROVAL_GATE=0``.
+    A chat-session grant bypasses it only within the scope of its own task.
+    """
+    return (
+        get_setting("tool_approval_mode", "ask") == "off"
+        or bool(exact_approval and exact_approval.allow_remaining_actions)
+    )
+
 
 def _tool_result_signature(tool_result_records: list[dict]) -> str:
     """Return a bounded fingerprint of the observable result of a tool batch.
@@ -21437,9 +21452,7 @@ async def stream_agent_loop(
             )
             or messages_contain_external_untrusted_context(messages)
         ),
-        approval_gate_bypassed=bool(
+        approval_gate_bypassed=_tool_approval_gate_bypassed(exact_approval),
-            exact_approval and exact_approval.allow_remaining_actions
-        ),
         delegated_credential=bool(delegated_credential),
     )
     _has_tui_host_bridge = _tui_host_bridge_is_usable(client_runtime_context)
@@ -24193,11 +24206,14 @@ async def stream_agent_loop(
                 budget_is_explicit,
                 hard_max=hard_max,
             )
-            if candidate_context <= 8192:
+            if 0 < candidate_context <= 8192:
                 # Small local servers often tokenize chat wrappers and tool
                 # results much more generously than our rough estimator. Keep
                 # substantial headroom for those wrappers and generation so a
                 # follow-up tool round cannot exceed the server's n_ctx.
+                # 0 means the window is unknown, which already gets the
+                # conservative default budget; clamping it as well cut every
+                # request to 1200 tokens and dropped open documents (#6678).
                 effective_budget = min(
                     effective_budget,
                     max(1200, int(candidate_context * 0.40)),
@@ -27833,7 +27849,16 @@ async def stream_agent_loop(
                                 if _streamed_tool_markup or _streamed_tool_markup_starts(_delta_text):
                                     _streamed_tool_markup += _delta_text
                                     if not _streamed_tool_markup_complete(_streamed_tool_markup):
+                                        if _streamed_tool_markup_starts(_streamed_tool_markup):
                                             continue
+                                        # False alarm: a bare "<" (e.g. ``x < 2``) looked like
+                                        # a tag prefix, but the buffer can no longer become a
+                                        # tool tag. Release it as prose instead of holding the
+                                        # rest of the answer forever.
+                                        _delta_text = _streamed_tool_markup
+                                        _streamed_tool_markup = ""
+                                        data["delta"] = _delta_text
+                                    else:
                                         _visible_markup_tail = strip_tool_blocks(
                                             _streamed_tool_markup,
                                             skip_fenced=True,

src/agent_runtime/authority.py

@@ -371,6 +371,101 @@ def interpret_request(request_text, *, history=(), workspace=None, active_docume
     return SemanticIntent(families, selected, required_read_operation_for_request(request_text, history))
 
 
+
+def _trusted_local_mcp_request_tools(request_text: str, intent: SemanticIntent) -> set[str]:
+    """Seal local MCP operations from operator declarations and trusted request intent.
+
+    Discovery validates an operator-declared tool but cannot grant it. For the
+    notes family, only an explicit request for local notes admits matching
+    read operations and explicitly requested non-destructive write verbs. Other
+    MCP domains require the user to name the qualified operation explicitly.
+    """
+    from src.tool_capabilities import (
+        LOCAL_MCP_READ_TOOLS_ENV, LOCAL_MCP_WRITE_TOOLS_ENV,
+        _local_mcp_tools, _is_declared_local_mcp_read_tool,
+        _is_declared_local_mcp_write_tool, capabilities_for_tool, ToolEffect,
+    )
+    from src.tool_utils import get_mcp_manager
+
+    manager = get_mcp_manager()
+    if manager is None:
+        return set()
+
+    request = request_text.casefold()
+    local_notes = (
+        intent.selected_tools is None
+        and intent.required_read is None
+        and "notes" in intent.capabilities
+        and re.search(r"\b(?:local|on my computer|on this machine)\b", request)
+    )
+    write_verbs = {
+        "create": r"\b(?:create|add|make)\b",
+        "add": r"\b(?:add|create)\b",
+        "write": r"\bwrite\b",
+        "save": r"\bsave\b",
+        "update": r"\b(?:update|edit|modify|change)\b",
+        "edit": r"\b(?:edit|update|modify|change)\b",
+        "append": r"\bappend\b",
+        "modify": r"\b(?:modify|edit|update|change)\b",
+    }
+    negative_write = re.search(
+        r"\b(?:do not|don't|never|without|no)\s+(?:\w+\s+){0,2}"
+        r"(?:create|add|make|write|save|update|edit|append|modify)\b",
+        request,
+    )
+    declared = _local_mcp_tools(LOCAL_MCP_READ_TOOLS_ENV) | _local_mcp_tools(LOCAL_MCP_WRITE_TOOLS_ENV)
+    admitted = set()
+    for name in sorted(declared):
+        parts = name.split("__", 2)
+        if len(parts) != 3 or parts[0] != "mcp":
+            continue
+        _, server, bare = parts
+        if manager.is_builtin(server):
+            continue
+        conn = manager.get_server_status(server)
+        if conn.get("status") != "connected" or conn.get("transport") != "stdio":
+            continue
+        resource = manager.resource_identity(name)
+        if resource is None or resource.namespace != "mcp" or resource.endpoint_id != f"stdio:{server}":
+            continue
+        rows = manager._tools.get(server, [])
+        row = next((item for item in rows if item.get("name") == bare), None)
+        if row is None:
+            continue
+        read_ok = _is_declared_local_mcp_read_tool(name)
+        write_ok = _is_declared_local_mcp_write_tool(name)
+        if not (read_ok or write_ok):
+            continue
+        caps = capabilities_for_tool(name)
+        if not caps.known or caps.effects != frozenset({
+            ToolEffect.READ_PRIVATE if read_ok else ToolEffect.WRITE_PRIVATE
+        }):
+            continue
+        ann = row.get("annotations") or {}
+        destructive = ann.get("destructiveHint") if isinstance(ann, dict) else getattr(ann, "destructiveHint", None)
+        if destructive is True:
+            continue
+        if write_ok and re.match(r"^(?:delete|remove|erase|destroy|drop|purge)(?:_|$)", bare, re.I):
+            continue
+
+        # Exact qualified names in the trusted request permit any configured
+        # and validated local operation. Model/tool content is not an input.
+        explicitly_named = bool(re.search(
+            r"(?<![A-Za-z0-9_])" + re.escape(name.casefold()) + r"(?![A-Za-z0-9_])",
+            request,
+        ))
+        if not explicitly_named:
+            if not local_notes or not re.search(r"(?:^|_)notes?(?:_|$)", bare, re.I):
+                continue
+            if write_ok:
+                action = bare.split("_", 1)[0].casefold()
+                if action not in write_verbs or not re.search(write_verbs[action], request):
+                    continue
+                if negative_write:
+                    continue
+        admitted.add(name)
+    return admitted
+
 def create_request_authority(request_text, *, owner=None, session_id=None, workspace=None,
                              history=(), policy=None, active_document=False,
                              image_attachment=False, capabilities=None, client_runtime_context=None):
@@ -393,6 +488,7 @@ def create_request_authority(request_text, *, owner=None, session_id=None, works
         tools = {canonical_tool(operation.tool)}
     else:
         operation = None
+    tools.update(_trusted_local_mcp_request_tools(request_text, intent))
     grants = []
     for name in sorted(tools | {"ask_user", "update_plan"}):
         name = canonical_tool(name)

src/agent_runtime/effect_log.py

@@ -47,6 +47,7 @@ EFFECTS_DIR = os.path.join(DATA_DIR, "effects")
 _RUN_ID = re.compile(r"[a-f0-9]{32}")
 _TYPES = {"claim": EffectClaim, "outcome": EffectOutcome, "observation": Observation}
 _VERSION = 1
+_O_BINARY = getattr(os, "O_BINARY", 0)  # Windows: raw os.open() is text-mode without it
 
 
 def _fsync_directory(directory: str | os.PathLike) -> None:
@@ -243,7 +244,7 @@ class EffectLog:
         """Hold the file exclusively with every durable record merged."""
         assert self.path is not None
         _ensure_directory(self.path.parent)
-        flags = os.O_RDWR | os.O_APPEND | os.O_CREAT | getattr(os, "O_NOFOLLOW", 0) | getattr(os, "O_CLOEXEC", 0)
+        flags = os.O_RDWR | os.O_APPEND | os.O_CREAT | getattr(os, "O_NOFOLLOW", 0) | getattr(os, "O_CLOEXEC", 0) | _O_BINARY
         descriptor = os.open(self.path, flags, 0o600)
         try:
             if fcntl is not None:
@@ -358,7 +359,7 @@ class EffectLog:
         with self._lock:
             try:
                 descriptor = os.open(self.path, os.O_RDONLY | getattr(os, "O_NOFOLLOW", 0)
-                                     | getattr(os, "O_CLOEXEC", 0))
+                                     | getattr(os, "O_CLOEXEC", 0) | _O_BINARY)
             except FileNotFoundError:
                 return
             except OSError:
@@ -394,7 +395,7 @@ class EffectLog:
         log = cls(run_id, directory=directory)
         assert log.path is not None
         try:
-            descriptor = os.open(log.path, os.O_RDONLY | getattr(os, "O_NOFOLLOW", 0) | getattr(os, "O_CLOEXEC", 0))
+            descriptor = os.open(log.path, os.O_RDONLY | getattr(os, "O_NOFOLLOW", 0) | getattr(os, "O_CLOEXEC", 0) | _O_BINARY)
         except FileNotFoundError:
             return log
         except OSError as error:
@@ -460,7 +461,7 @@ class EffectLog:
         temporary = target.with_suffix(".tmp")
         data = json.dumps({"run_id": self.run_id, "effect_id": effect_id}, sort_keys=True).encode()
         _ensure_directory(target.parent)
-        flags = os.O_WRONLY | os.O_CREAT | os.O_TRUNC | getattr(os, "O_NOFOLLOW", 0) | getattr(os, "O_CLOEXEC", 0)
+        flags = os.O_WRONLY | os.O_CREAT | os.O_TRUNC | getattr(os, "O_NOFOLLOW", 0) | getattr(os, "O_CLOEXEC", 0) | _O_BINARY
         descriptor = os.open(temporary, flags, 0o600)
         try:
             view = memoryview(data)
@@ -478,7 +479,7 @@ class EffectLog:
             return None
         base = Path(directory) if directory is not None else effects_dir()
         try:
-            descriptor = os.open(base / f"launch-{generation}.json", os.O_RDONLY | getattr(os, "O_NOFOLLOW", 0))
+            descriptor = os.open(base / f"launch-{generation}.json", os.O_RDONLY | getattr(os, "O_NOFOLLOW", 0) | _O_BINARY)
             with os.fdopen(descriptor, "rb") as stream:
                 if os.fstat(stream.fileno()).st_nlink != 1:
                     return None

src/agent_tools/admin_tools.py

@@ -585,7 +585,9 @@ async def do_manage_settings(content: str, owner: Optional[str] = None) -> Dict:
             return _ALIASES_SET.get(k2, (k or "").strip())
 
         def _is_managed_key(key):
-            return key in DEFAULT_SETTINGS and key not in RETIRED_SETTING_KEYS
+            # Tool policy is not a general setting, even if a future default
+            # is added: resetting it must never re-enable disabled tools.
+            return key != "disabled_tools" and key in DEFAULT_SETTINGS and key not in RETIRED_SETTING_KEYS
 
         _ENUMS = {
             "image_quality": ["low", "medium", "high"],
@@ -713,13 +715,35 @@ async def do_manage_settings(content: str, owner: Optional[str] = None) -> Dict:
             save_settings(s)
             return {"response": f"Reset {key} to default ({DEFAULT_SETTINGS[key]}).", "exit_code": 0}
 
-        elif action in ("disable_tool", "enable_tool", "list_tools"):
+        elif action == "enable_tool":
+            # The global disabled_tools denylist is the one hard permission
+            # boundary the Settings -> Agent Tools panel offers admins. It
+            # must only ever be *narrowed* from inside an agent turn: a
+            # model that can call enable_tool can undo the boundary itself,
+            # and any web page, email, or document the agent reads can ask
+            # it to (prompt injection re-enabling bash/email/etc.). Widening
+            # the tool set stays a human action in the Settings panel.
+            tool_name = (args.get("tool") or args.get("name") or "").strip().lower()
+            return {
+                "error": (
+                    f"Re-enabling tools is not available from chat"
+                    f"{f' (requested: {tool_name})' if tool_name else ''}. "
+                    "Disabled tools are an admin policy: ask the user to turn the tool "
+                    "back on in Settings > Agent Tools. You can still disable tools "
+                    "(disable_tool) or list the current state (list_tools)."
+                ),
+                "exit_code": 1,
+            }
+
+        elif action in ("disable_tool", "list_tools"):
             # Tool-toggle actions. These edit settings.json:disabled_tools
             # (the global list read on every chat request) rather than
             # prefs.json. Friendly aliases accepted: "shell" -> "bash",
             # "search" -> "web_search", "browser" -> browser tools,
             # "documents" -> the document tool set, "memory" ->
-            # manage_memory, etc.
+            # manage_memory, etc. Only disable_tool mutates: see the
+            # enable_tool branch above for why the agent can never widen
+            # the denylist.
             from src.settings import get_setting, save_settings, load_settings
             _ALIASES = {
                 "shell": ["bash"],
@@ -754,6 +778,7 @@ async def do_manage_settings(content: str, owner: Optional[str] = None) -> Dict:
                         f"Currently disabled: {', '.join(current) if current else '(none)'}.\n"
                         "Common toggles: shell (bash), search (web_search), browser, documents, "
                         "memory, skills, images, tasks, notes, calendar, email. "
+                        "To re-enable a tool, open Settings > Agent Tools."
                     ),
                     "disabled": list(current),
                     "exit_code": 0,
@@ -767,22 +792,19 @@ async def do_manage_settings(content: str, owner: Optional[str] = None) -> Dict:
             settings = load_settings()
             current = list(settings.get("disabled_tools") or [])
             before = set(current)
-            if action == "disable_tool":
             for t in targets:
                 if t not in current:
                     current.append(t)
-            else:  # enable_tool
-                current = [t for t in current if t not in targets]
             after = set(current)
             settings["disabled_tools"] = current
             save_settings(settings)
 
-            verb = "Disabled" if action == "disable_tool" else "Enabled"
             changed = sorted(after.symmetric_difference(before))
             return {
                 "response": (
-                    f"{verb} {tool_name} ({', '.join(targets)}). "
+                    f"Disabled {tool_name} ({', '.join(targets)}). "
                     f"Now disabled: {', '.join(current) if current else '(none)'}. "
+                    "Re-enabling is done by the user in Settings > Agent Tools."
                 ),
                 "changed": changed,
                 "disabled": list(current),

src/agent_tools/document_tools.py

@@ -4,7 +4,7 @@ import html
 import difflib
 import logging
 import re
-from src.constants import MAX_READ_CHARS
+from src.constants import MAX_OUTPUT_CHARS, MAX_READ_CHARS
 from src.tool_approvals import document_content_digest
 from src.tool_utils import _parse_tool_args, get_upload_handler
 from src.upload_handler import reserve_upload_references
@@ -1175,14 +1175,23 @@ class ManageDocumentTool:
                 except (TypeError, ValueError):
                     offset = 0
                 offset = min(offset, len(body))
+                anchor = f"[{doc.title}](#document-{doc.id})"
+                header = f"{anchor} — click to open in editor.\n\n```{doc.language or ''}\n"
+                # Reserve the closing fence and the longest possible paging
+                # notice before selecting the page. The common output cap must
+                # never hide content that next_offset claims was delivered.
+                notice = f"\n... (truncated, {len(body)} chars total; next_offset={len(body)})"
+                page_budget = MAX_OUTPUT_CHARS - len(header) - len(notice) - len("\n```")
+                if page_budget < 1:
+                    return {"error": "Document title or language is too long to display a read page", "exit_code": 1}
+                preview_limit = min(preview_limit, page_budget)
                 end = min(offset + preview_limit, len(body))
                 truncated = end < len(body)
                 preview = body[offset:end]
                 if truncated:
                     preview += f"\n... (truncated, {len(body)} chars total; next_offset={end})"
-                anchor = f"[{doc.title}](#document-{doc.id})"
                 return {
-                    "response": f"{anchor} — click to open in editor.\n\n```{doc.language or ''}\n{preview}\n```",
+                    "response": f"{header}{preview}\n```",
                     "document": {
                         "id": doc.id,
                         "title": doc.title,

src/agent_tools/editor_tools.py

@@ -191,7 +191,7 @@ def _persist_png(image_bytes: bytes, owner: Optional[str], session_id: Optional[
         db.close()
 
 
-async def _place_layer(data_url: str, name: str, owner: Optional[str], session: str, *, masked: bool = False, blend: str = "normal", opacity: float = 100.0, mode: Optional[str] = None) -> bool:
+async def _place_layer(data_url: str, name: str, owner: Optional[str], session: str, *, masked: bool = False, blend: str = "normal", opacity: float = 100.0, mode: Optional[str] = None, native_size: bool = False) -> bool:
     params: Dict[str, Any] = {
         "dataUrl": data_url,
         "name": name,
@@ -199,6 +199,8 @@ async def _place_layer(data_url: str, name: str, owner: Optional[str], session:
         "blend": blend or "normal",
         "opacity": opacity,
     }
+    if native_size:
+        params["nativeSize"] = True
     # Forward the edit mode so the client can build a difference mask for
     # instruction-mode results: the whole image was regenerated under an
     # instruction prompt, so untouched pixels must stay pixel-identical (the
@@ -506,6 +508,8 @@ class AiEditImageTool:
             }
         _t["image"] = time.perf_counter()
         payload: Dict[str, Any] = {"image": image}
+        if args.get("_use_image_settings") is True:
+            payload["_use_image_settings"] = True
         model_choice = str(args.get("model") or "").strip()
         mask_used = False
         place_mode: Optional[str] = None
@@ -609,10 +613,11 @@ class AiEditImageTool:
                 f"AI {op}",
                 owner,
                 session,
-                masked=mask_used,
+                masked=mask_used and data.get("full_frame") is not True,
                 blend=str(args.get("blend") or "normal"),
                 opacity=float(args.get("opacity") or 100),
                 mode=place_mode,
+                native_size=data.get("full_frame") is True,
             )
             info["placed"] = placed
             if placed:
@@ -687,6 +692,7 @@ class GenerateImageLayerTool:
                 str(args.get("name") or "Generated image"),
                 owner,
                 session,
+                native_size=True,
                 masked=bool(args.get("mask")),
                 blend=str(args.get("blend") or "normal"),
                 opacity=float(args.get("opacity") or 100),

src/ai_interaction.py

@@ -179,8 +179,13 @@ def _resolve_model(spec: str, owner: Optional[str] = None, model_type: Optional[
                 # served internal id. Include pinned/cached ids in the match set
                 # so chat sessions using the HF repo id still resolve. Do not use
                 # stale cached aliases when the endpoint itself is unreachable.
-                if model_type == "image" and endpoint_reachable:
+                if endpoint_reachable:
-                    for extra in _json_list(getattr(ep, "pinned_models", None)) + _json_list(getattr(ep, "cached_models", None)):
+                    # Gateway image models may live in a dedicated catalogue,
+                    # so explicitly pinned IDs are valid on mixed endpoints too.
+                    extras = _json_list(getattr(ep, "pinned_models", None))
+                    if model_type == "image":
+                        extras += _json_list(getattr(ep, "cached_models", None))
+                    for extra in extras:
                         if extra not in model_ids:
                             model_ids.append(extra)
 
@@ -1176,9 +1181,9 @@ def image_editing_provider_available(owner: Optional[str] = None) -> bool:
     request on this so it never falls back to deterministic approximations.
     """
     try:
-        from src.settings import load_settings
+        from src.settings import get_user_setting, load_settings
 
-        if not load_settings().get("image_gen_enabled", True):
+        if not load_settings().get("image_gen_enabled", True) or not get_user_setting("image_gen_enabled", owner=owner or "", default=True):
             return False
     except Exception:
         pass
@@ -1200,6 +1205,39 @@ def image_editing_provider_available(owner: Optional[str] = None) -> bool:
         return False
 
 
+def _uses_openrouter_chat_images(model_id: str) -> bool:
+    # GPT-5 Image is an LLM wrapper, unlike native gpt-image-* image models.
+    leaf = model_id_leaf(model_id).lower()
+    return leaf.startswith("gpt-") and "-image" in leaf and not leaf.startswith("gpt-image-")
+
+
+def _chat_image_results(data: dict) -> list:
+    images = []
+    for choice in data.get("choices") or []:
+        message = choice.get("message") or {}
+        for item in message.get("images") or []:
+            url = (item.get("image_url") or {}).get("url")
+            if not isinstance(url, str):
+                continue
+            if url.startswith("data:image/") and ";base64," in url:
+                images.append({"b64_json": url.split(",", 1)[1]})
+            elif url.startswith(("https://", "http://")):
+                images.append({"url": url})
+    return images
+
+
+def _image_response_error(data: dict, model_id: str) -> Optional[str]:
+    errors = [data.get("error")]
+    errors.extend(choice.get("error") for choice in data.get("choices") or [])
+    for error in errors:
+        if not error:
+            continue
+        message = str(error.get("message") or "Image provider failed") if isinstance(error, dict) else str(error)
+        hint = " Choose a currently supported image model in AI settings." if "deprecated" in message.lower() else ""
+        return f"Image provider failed for '{model_id}': {message}{hint}"
+    return None
+
+
 async def do_generate_image(content: str, session_id: Optional[str] = None, owner: Optional[str] = None, *, incognito: bool = False) -> Dict:
     """Generate an image using an image-capable model (e.g. gpt-image-1).
 
@@ -1207,7 +1245,7 @@ async def do_generate_image(content: str, session_id: Optional[str] = None, owne
       Line 1: prompt describing the image
       Line 2: model name (optional, default auto-detects: prefers gpt-image-1.5 > gpt-image-1)
       Line 3: size (optional, defaults to 1024x1024)
-      Line 4: quality (optional, defaults to medium — options: low, medium, high, auto)
+      Line 4: quality (optional, defaults to configured quality or medium — low, medium, high, auto)
     """
     import base64
     import httpx
@@ -1225,32 +1263,38 @@ async def do_generate_image(content: str, session_id: Optional[str] = None, owne
         prompt = str(args.get('prompt') or '').strip()
         model_spec = str(args.get('model') or '').strip()
         size = str(args.get('size') or '1024x1024')
-        quality = str(args.get('quality') or 'medium')
+        quality = str(args.get('quality') or '')
     else:
         lines = content.strip().split("\n")
         prompt = lines[0].strip() if lines else ""
         model_spec = lines[1].strip() if len(lines) > 1 and lines[1].strip() else ""
         size = lines[2].strip() if len(lines) > 2 and lines[2].strip() else "1024x1024"
-        quality = lines[3].strip() if len(lines) > 3 and lines[3].strip() else "medium"
+        quality = lines[3].strip() if len(lines) > 3 and lines[3].strip() else ""
 
     if not prompt:
         return {"error": "Image prompt is required (line 1)"}
 
-    # Load admin settings for defaults
+    # Load settings for defaults
     try:
-        from src.settings import load_settings
+        from src.settings import get_user_setting, load_settings
         _settings = load_settings()
     except Exception:
         _settings = {}
+        def get_user_setting(key, owner="", default=None):
+            return default
 
+    # The administrator's global image-generation disable switch is authoritative.
+    # Non-admin preferences cannot re-enable generation when disabled globally.
     if not _settings.get("image_gen_enabled", True):
         return {"error": "Image generation is disabled by the administrator."}
+    if not get_user_setting("image_gen_enabled", owner=owner or "", default=True):
+        return {"error": "Image generation is disabled by user preference."}
 
-    # Use admin-configured model/quality if not specified by the tool call
+    # Use user- or admin-configured model/quality if not specified by the tool call
     if not model_spec:
-        model_spec = _settings.get("image_model", "")
+        model_spec = get_user_setting("image_model", owner=owner, default="") or ""
-    if quality == "medium" and _settings.get("image_quality"):
+    if not quality:
-        quality = _settings["image_quality"]
+        quality = get_user_setting("image_quality", owner=owner, default="medium") or "medium"
 
     # Auto-detect best available image model if still not set
     if not model_spec:
@@ -1342,8 +1386,19 @@ async def do_generate_image(content: str, session_id: Optional[str] = None, owne
     base_url = url.replace("/chat/completions", "").replace("/v1/messages", "").rstrip("/")
     images_url = base_url + "/images/generations"
     from src.model_capability_readers.base import detect_vendor
+    chat_images = detect_vendor(url) == "openrouter" and _uses_openrouter_chat_images(model_id)
     if detect_vendor(url) == "openrouter":
         images_url = base_url + "/images"
+    if chat_images:
+        images_url = base_url + "/chat/completions"
+
+    # Capture the requested geometry before older OpenAI size restrictions.
+    from src.image_model_ids import image_aspect_ratio
+    try:
+        requested_width, requested_height = (int(value) for value in size.lower().split("x"))
+        requested_ratio = image_aspect_ratio(requested_width, requested_height)
+    except (TypeError, ValueError):
+        requested_ratio = "1:1"
 
     # Validate size for cloud image models (local diffusion accepts any WxH)
     valid_gpt_sizes = {"1024x1024", "1024x1536", "1536x1024", "auto"}
@@ -1359,6 +1414,9 @@ async def do_generate_image(content: str, session_id: Optional[str] = None, owne
         "n": 1,
         "size": size,
     }
+    if detect_vendor(url) == "openrouter":
+        payload.pop("size", None)
+        payload["aspect_ratio"] = requested_ratio
 
     # GPT image models and local diffusion support quality; DALL-E does not
     if is_gpt_image or is_local_diffusion:
@@ -1372,7 +1430,16 @@ async def do_generate_image(content: str, session_id: Optional[str] = None, owne
     try:
         # GPT image models can take 30-120s+ depending on quality
         async with httpx.AsyncClient(timeout=httpx.Timeout(connect=30.0, read=300.0, write=30.0, pool=30.0)) as client:
-            resp = await client.post(images_url, json=payload, headers=headers)
+            request_payload = payload
+            if chat_images:
+                request_payload = {
+                    "model": model_id,
+                    "messages": [{"role": "user", "content": prompt}],
+                    "modalities": ["image", "text"],
+                    "stream": False,
+                    "image_config": {"aspect_ratio": requested_ratio},
+                }
+            resp = await client.post(images_url, json=request_payload, headers=headers)
 
             if resp.status_code != 200:
                 error_text = resp.text[:500]
@@ -1387,9 +1454,15 @@ async def do_generate_image(content: str, session_id: Optional[str] = None, owne
                 }
 
             data = resp.json()
-            images = data.get("data", [])
+            provider_error = _image_response_error(data, model_id)
+            if provider_error:
+                return {"error": provider_error, "untrusted_content": True}
+            images = _chat_image_results(data) if chat_images else data.get("data", [])
             if not images:
-                return {"error": "No images returned from API"}
+                return {"error": (
+                    f"Image model '{model_id}' returned no image. "
+                    "Choose another image generation model in AI settings."
+                )}
 
             img = images[0]
             image_url = None
@@ -1482,7 +1555,7 @@ async def do_edit_image(
     session_id: Optional[str] = None,
     owner: Optional[str] = None,
     size: str = "auto",
-    quality: str = "medium",
+    quality: Optional[str] = None,
     progress_callback: Optional[Callable[[Dict[str, Any]], Awaitable[None]]] = None,
     incognito: bool = False,
 ) -> Dict:
@@ -1502,21 +1575,27 @@ async def do_edit_image(
         return {"error": "Attached image file was not found"}
 
     try:
-        from src.settings import load_settings
+        from src.settings import get_user_setting, load_settings
         _settings = load_settings()
     except Exception:
         _settings = {}
+        def get_user_setting(key, owner="", default=None):
+            return default
 
     if not model_spec:
-        model_spec = _settings.get("image_model", "")
+        model_spec = get_user_setting("image_edit_model", owner=owner) or get_user_setting("image_model", owner=owner, default="")
-    if quality == "medium" and _settings.get("image_quality"):
+    if quality is None:
-        quality = _settings["image_quality"]
+        quality = get_user_setting("image_quality", owner=owner, default="medium") or "medium"
     if not model_spec:
         return {"error": "No image model selected for image editing"}
 
     try:
         try:
             def _call():
+                try:
+                    return _resolve_model(model_spec, owner=owner)
+                except ValueError:
+                    pass
                 try:
                     return _resolve_model(model_spec, owner=owner, model_type="image")
                 except TypeError as exc:
@@ -1529,6 +1608,7 @@ async def do_edit_image(
     except ValueError:
         return {"error": f"No endpoint found with image model '{model_spec}'."}
 
+    requested_edit_size = size
     if not size or size == "auto":
         from PIL import Image
         from src.image_model_ids import image_edit_size
@@ -1704,6 +1784,34 @@ async def do_edit_image(
                             "image_url": {"url": f"data:{mime};base64,{image_b64}"},
                         }],
                     }
+                    from src.image_model_ids import image_aspect_ratio
+                    from PIL import Image as ReferenceImage
+                    with ReferenceImage.open(path) as reference:
+                        reference_width, reference_height = reference.size
+                        if reference.getexif().get(274) in {5, 6, 7, 8}:
+                            reference_width, reference_height = reference_height, reference_width
+                    if requested_edit_size and requested_edit_size != "auto":
+                        try:
+                            reference_width, reference_height = (int(value) for value in requested_edit_size.lower().split("x"))
+                        except (TypeError, ValueError):
+                            pass
+                    edit_payload.pop("size", None)
+                    edit_payload["aspect_ratio"] = image_aspect_ratio(reference_width, reference_height)
+                    if _uses_openrouter_chat_images(model_id):
+                        edit_payload = {
+                            "model": model_id,
+                            "messages": [{"role": "user", "content": [
+                                {"type": "text", "text": prompt},
+                                {"type": "image_url", "image_url": {
+                                    "url": f"data:{mime};base64,{image_b64}",
+                                }},
+                            ]}],
+                            "modalities": ["image", "text"],
+                            "stream": False,
+                            "image_config": {"aspect_ratio": edit_payload["aspect_ratio"]},
+                        }
+                        resp = await client.post(base_url + "/chat/completions", json=edit_payload, headers=headers)
+                    else:
                         resp = await client.post(base_url + "/images", json=edit_payload, headers=headers)
                 else:
                     with path.open("rb") as f:
@@ -1747,7 +1855,12 @@ async def do_edit_image(
                 }
 
             data = resp.json()
-            images = data.get("data", [])
+            provider_error = _image_response_error(data, model_id)
+            if provider_error:
+                return {"error": provider_error, "untrusted_content": True}
+            images = (_chat_image_results(data)
+                      if is_openrouter and _uses_openrouter_chat_images(model_id)
+                      else data.get("data", []))
             if not images:
                 return {"error": "No image returned from edit API"}
 

src/browser_observation.py

@@ -25,7 +25,7 @@ def compact_browser_observation(value, budget=8000):
                 # Snapshot text already contains labels and refs in DOM order.
                 # The refs mapping repeats them and buries menus in raw JSON.
                 lines = [line for line in snapshot.splitlines()
-                         if not re.fullmatch(r'\s*-?\s*generic(?:\s+\[ref=e\d+\])?:?\s*', line)]
+                         if not re.fullmatch(r'\s*(?:-\s*)?generic(?:\s+\[ref=e\d+\])?:?\s*', line)]
                 pages.append(('URL: ' + str(url) + '\n' if url else '') + '\n'.join(lines))
             elif url:
                 notices.append('URL: ' + str(url))
@@ -37,7 +37,8 @@ def compact_browser_observation(value, budget=8000):
                 parsed = json.loads(item)
             except (ValueError, TypeError):
                 # CLI status text precedes the JSON post-interaction state.
-                parts = re.split(r'\n\n\[(?:post-[^\]]+|page state after failed [^\]]+)\]\n', item)
+                # Labels are one line; stopping at \n keeps the split linear.
+                parts = re.split(r'\n\n\[(?:post-[^\]\n]+|page state after failed [^\]\n]+)\]\n', item)
                 if len(parts) > 1:
                     for part in parts:
                         visit(part)

src/builtin_actions.py

@@ -10,7 +10,7 @@ import os
 import json
 import re
 import time
-from datetime import datetime
+from datetime import datetime, timezone
 from typing import Tuple
 
 from src.auth_helpers import owner_filter
@@ -1957,6 +1957,23 @@ async def action_learn_sender_signatures(owner: str, **kwargs) -> Tuple[str, boo
         return str(e), False
 
 
+def _daily_brief_utc_boundary(local_dt, tz=None):
+    """Convert a local wall-clock boundary using the offset at that boundary."""
+    if local_dt.tzinfo is None and tz is not None:
+        local_dt = local_dt.replace(tzinfo=tz)
+    return local_dt.astimezone(timezone.utc).replace(tzinfo=None)
+
+
+def _daily_brief_local_start(event, tz=None):
+    """Keep date-only/floating rows local; convert only stored UTC instants."""
+    start = event.dtstart
+    if event.is_utc and not event.all_day:
+        if start.tzinfo is None:
+            start = start.replace(tzinfo=timezone.utc)
+        return start.astimezone(tz).replace(tzinfo=None)
+    return start
+
+
 async def action_daily_brief(owner: str, **kwargs) -> Tuple[str, bool]:
     """Build a short morning digest: today's calendar events, unread email count
     + top-N senders/subjects, active todos."""
@@ -1966,10 +1983,13 @@ async def action_daily_brief(owner: str, **kwargs) -> Tuple[str, bool]:
 
         from core.database import SessionLocal, CalendarEvent, CalendarCal, Note
         from routes.email_helpers import _imap_connect, _decode_header
+        from sqlalchemy import and_, or_
 
         # ----- Calendar: today's events -----
         today = _dt.now().replace(hour=0, minute=0, second=0, microsecond=0)
         tomorrow = today + _td(days=1)
+        utc_today = _daily_brief_utc_boundary(today)
+        utc_tomorrow = _daily_brief_utc_boundary(tomorrow)
         # v2 review HIGH-12: gate the OR-null branch on single-user
         # (unconfigured) deploys only. In a multi-user deploy, one
         # user's daily brief must not include another user's notes or
@@ -1982,13 +2002,25 @@ async def action_daily_brief(owner: str, **kwargs) -> Tuple[str, bool]:
         db = SessionLocal()
         try:
             ev_q = db.query(CalendarEvent).join(CalendarCal).filter(
+                or_(
+                    and_(
+                        CalendarEvent.is_utc.is_(True),
+                        CalendarEvent.all_day.is_not(True),
+                        CalendarEvent.dtstart < utc_tomorrow,
+                        CalendarEvent.dtend > utc_today,
+                    ),
+                    and_(
+                        or_(CalendarEvent.is_utc.is_not(True), CalendarEvent.all_day.is_(True)),
                         CalendarEvent.dtstart < tomorrow,
                         CalendarEvent.dtend > today,
+                    ),
+                ),
                 CalendarEvent.status != "cancelled",
             )
             if owner:
                 ev_q = owner_filter(ev_q, CalendarCal, owner, include_shared=_allow_null)
-            events = ev_q.order_by(CalendarEvent.dtstart).all()
+            events = ev_q.all()
+            events.sort(key=_daily_brief_local_start)
             # ----- Notes: pinned + non-archived todos with at least one undone item -----
             n_q = db.query(Note).filter(Note.archived == False)  # noqa: E712
             if owner:
@@ -2058,7 +2090,7 @@ async def action_daily_brief(owner: str, **kwargs) -> Tuple[str, bool]:
         if events:
             plain.append("Calendar:")
             for e in events:
-                t = e.dtstart.strftime("%H:%M") if not e.all_day else "all day"
+                t = _daily_brief_local_start(e).strftime("%H:%M") if not e.all_day else "all day"
                 loc = f" @ {e.location}" if e.location else ""
                 plain.append(f"  {t}  {e.summary}{loc}")
             plain.append("")

src/chat_handler.py

@@ -282,7 +282,13 @@ class ChatHandler:
                                     _sync_upload_vision_to_gallery(file_info, owner, vl_desc)
                                 except Exception:
                                     pass
-                        enhanced_message = f"{enhanced_message}\n\n[Image: {file_info['name']}]\n{vl_desc}"
+                        enhanced_message = f"{enhanced_message}\n\n[Image: {file_info['name']}]\n"
+                        if vl_desc:
+                            # A separate vision call wrote this, but the chat model sees
+                            # one message and answers the prose as if the user wrote it
+                            # (#5573). Label it below the marker line — the client strips
+                            # the block by matching `[Image: <name>]\n`.
+                            enhanced_message += f"[Image description — context only, not part of the user's message]:\n{vl_desc}"
                         # Surface the description to the client live so it renders as a
                         # collapsible "image description" on the user bubble (not just
                         # after a refresh that re-parses the stored message).

src/chat_processor.py

@@ -9,9 +9,61 @@ from src.chat_helpers import extract_urls
 from src.youtube_handler import is_youtube_url
 from src.search import comprehensive_web_search, fetch_webpage_content
 from src.prompt_security import UNTRUSTED_CONTEXT_POLICY, untrusted_context_message
+from src.text_helpers import strip_think
 
 logger = logging.getLogger(__name__)
 
+_SEARCH_QUERY_HISTORY_TURNS = 4
+_SEARCH_QUERY_HISTORY_TURN_CHARS = 500
+
+
+def _search_query_messages(message: str, history: List[Any]) -> List[Dict[str, str]]:
+    """Build the query extraction prompt, with recent turns to resolve references.
+
+    A follow-up such as "Does he work with Jane?" has no searchable subject on
+    its own. The current turn is already the last history entry when the session
+    persists it, so a trailing user entry is dropped rather than repeated.
+    """
+    turns = [m for m in history if getattr(m, "role", "") in {"user", "assistant"}]
+    if turns and turns[-1].role == "user":
+        turns.pop()
+
+    lines = []
+    for turn in turns[-_SEARCH_QUERY_HISTORY_TURNS:]:
+        content = turn.content
+        if isinstance(content, list):
+            content = " ".join(
+                str(part.get("text") or "") for part in content if isinstance(part, dict)
+            )
+        text = " ".join(strip_think(str(content or ""), prose=False, prompt_echo=False).split())
+        if text:
+            lines.append(f"{turn.role}: {text[:_SEARCH_QUERY_HISTORY_TURN_CHARS]}")
+
+    if not lines:
+        return [
+            {
+                "role": "system",
+                "content": "Extract a concise search query from the user's message. Reply ONLY with the query.",
+            },
+            {"role": "user", "content": message},
+        ]
+    conversation = "\n".join(lines)
+    return [
+        {
+            "role": "system",
+            "content": (
+                "Extract a concise search query from the user's latest message. "
+                "Use the earlier conversation only to resolve references such as pronouns, "
+                "and leave out earlier topics the latest message does not refer to. "
+                "Reply ONLY with the query."
+            ),
+        },
+        {
+            "role": "user",
+            "content": f"Earlier conversation:\n{conversation}\n\nLatest message:\n{message}",
+        },
+    ]
+
 
 def _clean_search_query(query: str, max_len: int = 200) -> str:
     """Strip fenced code blocks from a search query while preserving inline
@@ -422,21 +474,18 @@ class ChatProcessor:
                     generated_query = llm_call(
                         t_url,
                         t_model,
-                        [
+                        _search_query_messages(
-                            {
+                            message,
-                                "role": "system",
+                            [] if incognito else getattr(session, "history", None) or [],
-                                "content": (
-                                    "Extract a concise search query from the user's message. "
-                                    "Reply ONLY with the query."
                         ),
-                            },
-                            {"role": "user", "content": message},
-                        ],
                         headers=t_headers,
                         temperature=0.1,
                         max_tokens=50,
                         timeout=15,
-                    ).strip()
+                        thinking_mode="off",
+                    )
+                    # Models without a reasoning parser put thinking in the content.
+                    generated_query = strip_think(generated_query, prose=False, prompt_echo=False).strip()
 
                     if generated_query:
                         # LLM successfully generated a non-empty query -> use the generated query

src/clean_agent_preview.py

@@ -1229,7 +1229,7 @@ def tasks_terminal_response(raw, *, user_text='', max_items=20):
     text = str(payload or '').strip()
     rows = []
     for line in text.splitlines():
-        match = re.match(r'^\s*\d+\.\s+(.*?)\s+\(([^)]+)\)\s+—\s+(.+)$', line)
+        match = re.match(r'^\s*\d+\.\s(.*?)\s\(([^()]+)\)\s+—\s+(.+)$', line)
         if match:
             details = match[3].strip()
             rows.append({
@@ -1346,7 +1346,7 @@ def skills_terminal_response(raw, *, user_text='', max_items=20):
             continue
         match = re.match(r'^\s*-\s+\*\*([^*]+)\*\*(?:\s+\(([^)]+)\))?', line)
         if not match:
-            match = re.match(r'^\s*-\s+([^:]+?)(?:\s+\(([^)]+)\))?(?:\s*:|$)', line)
+            match = re.match(r'^\s*-\s+([^\s:][^:]*?)(?:\s\(([^()]+)\)(?:\s*:|$)|:|$)', line)
         if match and status:
             rows.append((status, match.group(1).strip(), (match.group(2) or '').strip()))
     if not rows:
@@ -5065,7 +5065,7 @@ def email_search_result_empty(value):
         try:
             decoded = json.loads(value)
         except (TypeError, ValueError):
-            return bool(re.fullmatch(r'\s*No emails matched [^\n]+\.?\s*', value, re.I))
+            return bool(re.fullmatch(r'\s*No emails matched [^\n]+(?:\n\s*)?', value, re.I))
         if decoded == value:
             return False
         value = decoded
@@ -5339,6 +5339,16 @@ def private_browser_effective_url(result):
     return urls[-1] if urls else ''
 
 
+def is_google_sorry_url(url):
+    """Recognize Google's rate-limit interstitial."""
+    # Matching the host separately keeps this linear; one combined pattern
+    # rescans the host from every dot.
+    return any(
+        re.fullmatch(r'(?:[^/]+\.)?google\.[^/]+', host[1], re.I)
+        for host in re.finditer(r'https?://([^/]+)/sorry/', url, re.I)
+    )
+
+
 def browser_observation_page_missing(raw):
     """Recognize a rendered error page, rather than an article mentioning 404."""
     text = str(raw or '')
@@ -8213,11 +8223,7 @@ async def stream_preview(*, endpoint_url, model, messages, headers, turn_contrac
                         browser_search_url = requested_browser_url or effective_browser_url
                         is_search_engine_navigation = contains_search_engine_navigation(
                             browser_search_url
-                        ) or bool(re.search(
+                        ) or is_google_sorry_url(effective_browser_url)
-                            r'https?://(?:[^/]+\.)?google\.[^/]+/sorry/',
-                            effective_browser_url,
-                            re.I,
-                        ))
                         has_native_search = any(
                             canonical(schema['function']['name']) == 'web_search'
                             for schema in offered

src/constants.py

@@ -175,6 +175,7 @@ AGENT_ISOLATED_TMP_DIRNAME = ".tmp"
 
 # Auth policy
 PASSWORD_MIN_LENGTH = 8
+USERNAME_MAX_LENGTH = 64
 
 # Default parameters
 DEFAULT_TEMPERATURE = 1.0

src/context_compactor.py

@@ -42,6 +42,15 @@ def _content_as_text(content: Any) -> str:
 _MULTIMODAL_IMAGE_TYPES = {"image_url", "input_image", "image"}
 
 
+def _has_preservable_content(content: Any) -> bool:
+    """Retain nonblank text or nonempty multimodal blocks after tool-call removal."""
+    if isinstance(content, str):
+        return bool(content.strip())
+    if isinstance(content, list):
+        return bool(content)
+    return False
+
+
 def prune_multimodal_images(
     messages: List[Dict],
     *,
@@ -222,22 +231,81 @@ def normalize_compaction_summary(summary: str) -> str:
 
 
 def _sanitize_tool_messages(msgs: List[Dict]) -> List[Dict]:
-    """Drop orphaned `tool` messages and dangling assistant `tool_calls`.
+    """Normalize tool-call message ordering so strict providers accept the transcript.
 
     OpenAI's API requires every `role:"tool"` message to immediately
     follow an assistant message that carries `tool_calls` (or another
-    tool message in the same batch). Front-trimming the history can cut
+    tool message in the same batch). Gemini and some other providers
-    the assistant `tool_calls` parent while keeping its tool responses,
+    additionally require a function call turn to come immediately after
-    which triggers: "messages with role 'tool' must be a response to a
+    a user turn or a function response turn. Trimming the history can
-    preceding message with 'tool_calls'". This pass repairs that:
+    break either rule, which triggers errors such as "messages with role
-      - drops `tool` messages with no valid preceding tool_calls
+    'tool' must be a response to a preceding message with 'tool_calls'"
-      - drops assistant `tool_calls` messages whose tool responses were
+    and "Please ensure that function call turn comes immediately after
-        all trimmed away (some providers reject unanswered tool_calls)
+    a user turn or after a function response turn."
+
+    The transcript is normalized against its own effective output: each
+    pass repairs the violations it can see and the passes repeat until
+    the result is stable, so the returned transcript itself satisfies
+    the ordering contract and sanitizing it again is a no-op. A
+    completed call/result batch whose predecessor is an adjacent
+    assistant text turn keeps its calls by merging that text into the
+    call message instead of dropping executed-tool evidence; a batch
+    that cannot get a user/tool predecessor (e.g. a transcript that
+    starts mid-history) is the only one dropped.
     """
-    # Pass 1: drop orphan tool messages.
+    work = list(msgs)
+    for _ in range(2 * len(work) + 2):
+        normalized = _normalize_tool_messages_once(work)
+        if normalized == work:
+            break
+        work = normalized
+    return work
+
+
+def _normalize_tool_messages_once(msgs: List[Dict]) -> List[Dict]:
+    """One normalization pass over an already-trimmed message list."""
+    # Merge the run of assistant text turns directly before an assistant
+    # tool_calls turn into a single assistant message carrying both, so the
+    # batch inherits the text run's predecessor (a user or tool message)
+    # instead of losing its tool_calls. Only plain-text content is merged;
+    # multimodal content stays untouched and falls through to the strip logic.
+    merged: List[Dict] = []
+    for m in msgs:
+        cur_text = m.get("content")
+        if (
+            m.get("role") == "assistant"
+            and m.get("tool_calls")
+            and (cur_text is None or isinstance(cur_text, str))
+        ):
+            text_run: List[Dict] = []
+            while (
+                merged
+                and merged[-1].get("role") == "assistant"
+                and not merged[-1].get("tool_calls")
+                and isinstance(merged[-1].get("content"), str)
+            ):
+                text_run.append(merged.pop())
+            if text_run:
+                text_parts = [
+                    msg["content"].strip()
+                    for msg in reversed(text_run)
+                    if msg["content"].strip()
+                ]
+                if isinstance(cur_text, str) and cur_text.strip():
+                    text_parts.append(cur_text.strip())
+                merged_msg = dict(m)
+                if text_parts:
+                    merged_msg["content"] = "\n\n".join(text_parts)
+                merged.append(merged_msg)
+                continue
+        merged.append(m)
+
+    # Strip tool_calls from assistant messages whose predecessor is not a
+    # user or tool message, then drop orphan tool messages and dangling
+    # tool_calls whose responses are gone.
     cleaned: List[Dict] = []
     in_batch = False  # are we right after an assistant tool_calls (or mid-batch)?
-    for m in msgs:
+    for m in merged:
         role = m.get("role")
         if role == "tool":
             if in_batch:
@@ -245,13 +313,21 @@ def _sanitize_tool_messages(msgs: List[Dict]) -> List[Dict]:
             # else: orphan — drop
             continue
         if role == "assistant" and m.get("tool_calls"):
+            prev_role = cleaned[-1].get("role") if cleaned else None
+            if prev_role not in ("user", "tool"):
+                # Invalid predecessor — strip tool_calls to avoid provider
+                # errors; keep the message only if it still carries content.
+                m = {k: v for k, v in m.items() if k != "tool_calls"}
+                in_batch = False
+                if not _has_preservable_content(m.get("content")):
+                    continue  # nothing left worth keeping
+                cleaned.append(m)
+                continue
             in_batch = True
         else:
             in_batch = False
         cleaned.append(m)
 
-    # Pass 2: drop assistant tool_calls messages that have NO following
-    # tool response (dangling) — walk backwards so we know what follows.
     out: List[Dict] = []
     for i, m in enumerate(cleaned):
         if m.get("role") == "assistant" and m.get("tool_calls"):
@@ -261,7 +337,7 @@ def _sanitize_tool_messages(msgs: List[Dict]) -> List[Dict]:
                 # tool_calls so it's a plain assistant turn (preserves any
                 # text content the model produced alongside the calls).
                 m = {k: v for k, v in m.items() if k != "tool_calls"}
-                if not (m.get("content") or "").strip():
+                if not _has_preservable_content(m.get("content")):
                     continue  # nothing left worth keeping
         out.append(m)
     return out

src/cookbook_serve_lifecycle.py

@@ -31,63 +31,43 @@ def _internal_headers() -> dict:
 
 
 async def _delete_endpoint_for_task(task: dict) -> None:
-    """Drop the auto-registered model endpoint for a scheduled-stop serve.
+    """Remove a stopped serve's endpoint using the normal deletion contract."""
+    import re
+    from urllib.parse import quote
 
-    Without this, killing the tmux session leaves the endpoint sitting in
+    endpoint_id = str(task.get("_endpointId") or task.get("endpointId") or "").strip()
-    the picker (probe goes offline; chats still try to route there) and
+    serve_session_id = str(task.get("sessionId") or task.get("id") or "").strip()
-    the user has to delete it by hand in Settings -> Endpoints.
+    if (
-    """
+        not re.fullmatch(r"(?:local|img)-[0-9a-f]{8}", endpoint_id)
-    endpoint_id = (task.get("_endpointId") or task.get("endpointId") or "").strip()
+        or not serve_session_id
-    if not endpoint_id:
+    ):
-        logger.info(
+        logger.info("cookbook_serve_lifecycle: missing endpoint/session identity")
-            "cookbook_serve_lifecycle: task %s has no endpoint id; skipping endpoint deletion",
-            task.get("sessionId") or task.get("id") or "",
-        )
         return
-    import re as _re
+
-    payload = task.get("payload") or {}
-    cmd = str(payload.get("_cmd") or "")
-    remote = task.get("remoteHost") or ""
-    # Build host the same way _auto_register_llm_endpoint does so URL match wins.
-    if remote:
-        host = remote.split("@")[-1] if "@" in remote else remote
-    else:
-        host = "host.docker.internal"
-    port_match = _re.search(r"--port\s+(\d+)", cmd)
-    ollama_host_match = _re.search(r"OLLAMA_HOST=[^\s]*?:(\d+)", cmd)
-    if port_match:
-        port = int(port_match.group(1))
-    elif ollama_host_match:
-        port = int(ollama_host_match.group(1))
-    elif "ollama" in cmd:
-        port = 11434
-    else:
-        port = 8080
-    base_url = f"http://{host}:{port}/v1"
     try:
         async with httpx.AsyncClient(timeout=8) as client:
-            r = await client.get(
+            response = await client.delete(
-                f"{internal_api_base()}/api/model-endpoints",
+                f"{internal_api_base()}/api/model-endpoints/{quote(endpoint_id, safe='')}",
-                headers=_internal_headers(),
+                params={"expected_cookbook_serve_session_id": serve_session_id},
-            )
-            if r.status_code >= 400:
-                return
-            eps = r.json() if r.content else []
-            # Delete only the endpoint created by this scheduled serve. URL
-            # matching is unsafe because a later scheduled serve can reuse the
-            # same host:port after an older task has gone stale.
-            ep = next((e for e in eps if e.get("id") == endpoint_id), None)
-            if ep:
-                await client.delete(
-                    f"{internal_api_base()}/api/model-endpoints/{ep['id']}",
                 headers=_internal_headers(),
             )
+        if response.status_code == 200:
             logger.info(
-                    f"cookbook_serve_lifecycle: deleted endpoint {ep.get('id')} "
+                "cookbook_serve_lifecycle: deleted endpoint %s for serve %s",
-                    f"({ep.get('base_url')}) after scheduled stop"
+                endpoint_id, serve_session_id,
             )
-    except Exception as e:
+        elif response.status_code in (404, 409):
-        logger.warning(f"cookbook_serve_lifecycle: endpoint delete failed: {e}")
+            logger.info(
+                "cookbook_serve_lifecycle: endpoint %s absent or reassigned; keeping it",
+                endpoint_id,
+            )
+        else:
+            logger.warning(
+                "cookbook_serve_lifecycle: endpoint %s deletion returned HTTP %s",
+                endpoint_id, response.status_code,
+            )
+    except Exception as exc:
+        logger.warning("cookbook_serve_lifecycle: cleanup failed: %s", exc)
 
 
 async def _stop_serve(session_id: str, remote_host: str = "", ssh_port: str = "") -> bool:

src/deep_research.py

@@ -625,7 +625,8 @@ class DeepResearcher:
     # LLM helper
     # ------------------------------------------------------------------
     async def _llm(self, messages: List[Dict], temperature: float = 0.3,
-                   max_tokens: int = 4096, timeout: int = 60) -> str:
+                   max_tokens: int = 4096, timeout: int = 60,
+                   thinking_mode: Optional[str] = None) -> str:
         """Call the LLM asynchronously and strip thinking tags."""
         from src.llm_core import llm_call_async
         response = await llm_call_async(
@@ -636,6 +637,7 @@ class DeepResearcher:
             max_tokens=max_tokens,
             headers=self.llm_headers,
             timeout=timeout,
+            thinking_mode=thinking_mode,
         )
         return strip_thinking(response)
 
@@ -696,7 +698,7 @@ class DeepResearcher:
         try:
             result = await self._llm(
                 [{"role": "user", "content": prompt}],
-                temperature=0, max_tokens=20, timeout=15,
+                temperature=0, max_tokens=20, timeout=15, thinking_mode="off",
             )
             cat = (result or "").strip().lower()
             # Clean one-word answer first.

src/document_processor.py

@@ -142,7 +142,10 @@ def _process_pdf(
                         with tempfile.NamedTemporaryFile(suffix=".png", delete=False) as tmp:
                             temp_img_path = tmp.name
                         try:
-                            img.image.save(temp_img_path, "PNG")  # pypdf -> PIL image
+                            image = img.image
+                            if image.mode == "CMYK":
+                                image = image.convert("RGB")
+                            image.save(temp_img_path, "PNG")
                             ocr_text = analyze_image_with_vl(temp_img_path, owner=owner)
                             if ocr_text and "unavailable" not in ocr_text.lower():
                                 pdf_text += f"\n\n[Page {page_num + 1} image {img_index + 1} text]: {ocr_text}"

src/editor_agent_model.py

@@ -339,6 +339,229 @@ async def self_check(
     return out
 
 
+def _resolve_override_candidate(override: str, owner: Optional[str]) -> Optional[Dict[str, Any]]:
+    """Resolve an explicit agent_model override under server authority.
+
+    Guarantees:
+      - Client-provided raw URLs never determine arbitrary outbound network destinations.
+      - Registered endpoint IDs resolve only to enabled endpoints visible to the owner.
+      - Shared endpoints (owner IS NULL) remain visible to authorized users.
+      - Disabled, foreign-owned, forged, or unconfigured destinations fail closed (return None).
+      - Admin privileges do not allow arbitrary unconfigured destinations.
+    """
+    raw = str(override or "").strip()
+    if not raw:
+        return None
+
+    if "::" in raw:
+        target, requested_model = raw.split("::", 1)
+        target = target.strip()
+        requested_model = requested_model.strip()
+    else:
+        target = ""
+        requested_model = raw
+
+    # 1. Explicit local selection or matching a configured local base
+    if target == "local":
+        bases = local_bases()
+        if not bases:
+            return None
+        return {
+            "base": bases[0],
+            "model": requested_model or local_model(),
+            "headers": None,
+            "source": "override",
+        }
+    for lb in local_bases():
+        if target and target.rstrip("/") == lb.rstrip("/"):
+            return {
+                "base": lb,
+                "model": requested_model or local_model(),
+                "headers": None,
+                "source": "override",
+            }
+
+    # 2. Explicit hosted fallback selection
+    try:
+        from src.endpoint_resolver import resolve_endpoint
+
+        for prefix in ("utility", "default"):
+            try:
+                url, model, headers = resolve_endpoint(prefix, owner=owner)
+            except Exception:
+                continue
+            if url and model:
+                if target in ("default", "hosted"):
+                    return {
+                        "base": url,
+                        "model": requested_model or model,
+                        "headers": headers,
+                        "source": "hosted",
+                    }
+                if target and target.rstrip("/") == url.rstrip("/"):
+                    if not requested_model or requested_model == model:
+                        return {
+                            "base": url,
+                            "model": requested_model or model,
+                            "headers": headers,
+                            "source": "hosted",
+                        }
+    except Exception:
+        pass
+    if target in ("default", "hosted"):
+        return None
+
+    # 3. Bare model id without endpoint prefix
+    if not target:
+        if requested_model == "local":
+            bases = local_bases()
+            return {"base": bases[0], "model": local_model(), "headers": None, "source": "local"} if bases else None
+        if requested_model in ("default", "hosted"):
+            return _resolve_override_candidate("default::", owner)
+
+        # Exact match with local model
+        if requested_model == local_model():
+            bases = local_bases()
+            if bases:
+                return {"base": bases[0], "model": local_model(), "headers": None, "source": "local"}
+
+        # Exact match with hosted default/utility
+        try:
+            from src.endpoint_resolver import resolve_endpoint
+
+            for prefix in ("utility", "default"):
+                try:
+                    url, model, headers = resolve_endpoint(prefix, owner=owner)
+                except Exception:
+                    continue
+                if url and model and model == requested_model:
+                    return {"base": url, "model": model, "headers": headers, "source": "hosted"}
+        except Exception:
+            pass
+
+        # Check registered database endpoints for this bare model
+        try:
+            from core.database import SessionLocal, ModelEndpoint
+            from src.auth_helpers import owner_filter, _auth_disabled
+            from src.endpoint_resolver import (
+                resolve_endpoint_runtime,
+                build_chat_url,
+                build_headers,
+                _endpoint_enabled_models,
+                _endpoint_hidden_models,
+            )
+
+            db = SessionLocal()
+            try:
+                if not owner and not _auth_disabled():
+                    q = db.query(ModelEndpoint).filter(
+                        ModelEndpoint.is_enabled == True,
+                        ModelEndpoint.owner == None,
+                    )
+                else:
+                    q = owner_filter(
+                        db.query(ModelEndpoint).filter(ModelEndpoint.is_enabled == True),
+                        ModelEndpoint,
+                        owner,
+                    )
+                matching_eps = []
+                for ep in q.all():
+                    enabled_models = _endpoint_enabled_models(ep)
+                    hidden_models = _endpoint_hidden_models(ep)
+                    if requested_model in enabled_models and requested_model not in hidden_models:
+                        matching_eps.append(ep)
+
+                # Avoid model ID collision across providers: only accept if unique
+                if len(matching_eps) == 1:
+                    ep = matching_eps[0]
+                    base, api_key = resolve_endpoint_runtime(ep, owner=owner)
+                    return {
+                        "base": build_chat_url(base),
+                        "model": requested_model,
+                        "headers": build_headers(api_key, base),
+                        "source": "registered",
+                        "endpoint_id": ep.id,
+                    }
+            finally:
+                db.close()
+        except Exception:
+            pass
+        return None
+
+    # 4. Target is an endpoint_id or a registered legacy endpoint URL
+    try:
+        from core.database import SessionLocal, ModelEndpoint
+        from src.auth_helpers import owner_filter, _auth_disabled
+        from src.endpoint_resolver import (
+            resolve_endpoint_runtime,
+            build_chat_url,
+            build_headers,
+            _endpoint_enabled_models,
+            _endpoint_hidden_models,
+            _first_chat_model,
+        )
+
+        db = SessionLocal()
+        try:
+            if not owner and not _auth_disabled():
+                q = db.query(ModelEndpoint).filter(
+                    ModelEndpoint.is_enabled == True,
+                    ModelEndpoint.owner == None,
+                )
+            else:
+                q = owner_filter(
+                    db.query(ModelEndpoint).filter(ModelEndpoint.is_enabled == True),
+                    ModelEndpoint,
+                    owner,
+                )
+            # Lookup by ID
+            ep = q.filter(ModelEndpoint.id == target).first()
+
+            # If not found by ID, match legacy registered base URL (strictly registered rows only)
+            if ep is None and ("://" in target or target.startswith("http")):
+                target_norm = target.rstrip("/")
+                for cand_ep in q.all():
+                    cand_base = (cand_ep.base_url or "").rstrip("/")
+                    if (
+                        cand_base == target_norm
+                        or cand_base + "/v1" == target_norm
+                        or cand_base == target_norm + "/v1"
+                    ):
+                        ep = cand_ep
+                        break
+
+            if ep is None:
+                # Target is disabled, foreign-owned, nonexistent, or arbitrary unconfigured URL -> fail closed!
+                return None
+
+            enabled_models = _endpoint_enabled_models(ep)
+            hidden_models = _endpoint_hidden_models(ep)
+            if requested_model:
+                if requested_model in hidden_models:
+                    return None
+                if enabled_models and requested_model not in enabled_models:
+                    # Model not served by this endpoint -> fail closed against forged model IDs
+                    return None
+                model = requested_model
+            else:
+                model = _first_chat_model(enabled_models)
+                if not model:
+                    return None
+
+            base, api_key = resolve_endpoint_runtime(ep, owner=owner)
+            return {
+                "base": build_chat_url(base),
+                "model": model,
+                "headers": build_headers(api_key, base),
+                "source": "registered",
+                "endpoint_id": ep.id,
+            }
+        finally:
+            db.close()
+    except Exception:
+        return None
+
+
 def _candidates(owner: Optional[str], override: Optional[str]) -> List[Dict[str, Any]]:
     cands: List[Dict[str, Any]] = []
     for base in local_bases():
@@ -357,16 +580,13 @@ def _candidates(owner: Optional[str], override: Optional[str]) -> List[Dict[str,
     except Exception:
         pass
     if override:
-        base, model = _split_override(override)
+        cand = _resolve_override_candidate(override, owner)
-        # A request may choose a configured endpoint, never introduce a URL.
+        if cand is not None:
-        # Prefer an exact model match (including hosted credentials) over the
+            cands.insert(0, cand)
-        # first local endpoint when a bare model id is supplied.
+            return cands
-        matching = [c for c in cands if not base or str(c["base"]).rstrip("/") == base.rstrip("/")]
+        # Unauthorized, invalid, foreign, disabled, or forged override:
-        selected = next((c for c in matching if c["model"] == model), None)
+        # Fail closed! Never silently dispatch sensitive user images/state to an unexpected fallback.
-        if selected is None and model and matching and matching[0]["source"] == "local":
+        return []
-            selected = {**matching[0], "model": model}
-        if selected is not None:
-            cands.insert(0, {**selected, "source": "override"})
     return cands
 
 
@@ -438,9 +658,11 @@ async def list_editor_agent_models(owner: Optional[str] = None) -> List[Dict[str
             "source": "local",
             "vision": True,
             "base_url": base,
+            "endpoint_id": "local",
         })
     try:
         from src.endpoint_resolver import resolve_endpoint
+        from src.settings import get_user_setting, load_settings
 
         for prefix in ("utility", "default"):
             try:
@@ -451,12 +673,19 @@ async def list_editor_agent_models(owner: Optional[str] = None) -> List[Dict[str
                 key = ("hosted", model)
                 if key not in seen:
                     seen.add(key)
+                    ep_id = None
+                    try:
+                        settings = load_settings()
+                        ep_id = (get_user_setting(f"{prefix}_endpoint_id", owner or "", settings.get(f"{prefix}_endpoint_id", "")) or "").strip() or None
+                    except Exception:
+                        pass
                     models.append({
                         "id": model,
                         "label": model,
                         "source": "hosted",
                         "vision": None,
                         "base_url": url,
+                        "endpoint_id": ep_id or prefix,
                     })
                 break
     except Exception:

src/editor_image_settings.py

@@ -0,0 +1,19 @@
+"""The shared model defaults used by Quick Edit's image operations."""
+from src.settings import get_user_setting, load_settings
+
+
+def editor_image_models(owner: str = "") -> dict[str, str]:
+    settings = load_settings()
+    generation = ""
+    edit_model = ""
+    if owner:
+        generation = str(get_user_setting("image_model", owner=owner, default="") or "").strip()
+        edit_model = str(get_user_setting("image_edit_model", owner=owner, default="") or "").strip()
+    if not generation:
+        generation = str(settings.get("image_model") or "").strip()
+    if not edit_model:
+        edit_model = str(settings.get("image_edit_model") or generation).strip()
+    return {
+        "generation": generation,
+        "edit": edit_model or generation,
+    }

src/email_mime.py

@@ -0,0 +1,13 @@
+"""Shared MIME text decoding for the web email helpers and MCP server."""
+
+
+def decode_text_part(part) -> str:
+    payload = part.get_payload(decode=True)
+    if not payload:
+        return ""
+    charset = part.get_content_charset() or "utf-8"
+    try:
+        return payload.decode(charset, errors="replace")
+    except (LookupError, ValueError):
+        # Broken senders can label valid UTF-8 with an unsupported codec.
+        return payload.decode("utf-8", errors="replace")

src/image_model_ids.py

@@ -40,6 +40,15 @@ def image_edit_size(model_id: str, width: int, height: int) -> str:
     return f"{width}x{height}"
 
 
+def image_aspect_ratio(width: int, height: int) -> str:
+    """Choose the nearest common Image API ratio, without distorting pixels."""
+    if width <= 0 or height <= 0:
+        raise ValueError("Image dimensions must be positive")
+    ratios = ((1, 1), (16, 9), (9, 16), (3, 2), (2, 3), (4, 3), (3, 4), (21, 9))
+    w, h = min(ratios, key=lambda ratio: abs(math.log((ratio[0] / ratio[1]) / (width / height))))
+    return f"{w}:{h}"
+
+
 def looks_like_image_generation_model(model_id: str) -> bool:
     """Return True when a model id should use image generation routes.
 

src/llm_core.py

@@ -422,7 +422,18 @@ def _stream_delta_event(text: str, *, thinking: bool = False) -> str:
     return f"data: {json.dumps(payload)}\n\n"
 
 
-_DEGENERATE_WORD_RE = re.compile(r"[A-Za-z0-9_\u0370-\u03ff\u0400-\u04ff]+")
+_WORD_CHARS = r"A-Za-z0-9_\u0370-\u03ff\u0400-\u04ff"
+_DELIM_RE = re.compile(rf"[^{_WORD_CHARS}+\-.]")
+_NUMERIC_PATTERN = rf"(?<![{_WORD_CHARS}])[+-]?(?:[0-9]+\.[0-9]+|\.[0-9]+|[0-9]+)(?:[eE][+-]?[0-9]+)?(?![{_WORD_CHARS}])"
+_NUMERIC_FULLMATCH_RE = re.compile(rf"^[+-]?(?:[0-9]+\.[0-9]+|\.[0-9]+|[0-9]+)(?:[eE][+-]?[0-9]+)?$")
+_DEGENERATE_WORD_RE = re.compile(rf"[{_WORD_CHARS}]+")
+_TOKEN_RE = re.compile(rf"(?P<NUM>{_NUMERIC_PATTERN})|(?P<WORD>[{_WORD_CHARS}]+)")
+
+
+def _is_numeric_token(token: str) -> bool:
+    if token.isdigit():
+        return True
+    return _NUMERIC_FULLMATCH_RE.fullmatch(token) is not None
 
 
 class _DegenerateStreamGuard:
@@ -432,32 +443,50 @@ class _DegenerateStreamGuard:
     "Summer Summer ..."). This is not a useful response and can burn context,
     browser memory, and GPU time. Keep the guard conservative: only fire on long
     same-token runs or a very dominant repeated token in the recent window.
+    Repetition made only of numbers is data (a zero or identity matrix, a
+    padded buffer), so it only stops at a long all-numeric backstop.
     """
 
     def __init__(self, model: str):
         self.model = model or "model"
         self.last_token = ""
         self.same_run = 0
+        self.numeric_run = 0
         self.recent_tokens: List[str] = []
         self.periodic_tokens: List[str] = []
         self.total_chars = 0
+        self._buffer = ""
 
     def check(self, text: str) -> Optional[str]:
         if not text:
             return None
         self.total_chars += len(text)
-        # Keep one-character tokens. Structured output often varies through
+        self._buffer += text
-        # numeric selectors or coordinates; dropping those tokens can make
+
-        # valid CSS/SVG rows look like an exact repeated phrase.
+        m_delims = list(_DELIM_RE.finditer(self._buffer))
-        tokens = [t.lower() for t in _DEGENERATE_WORD_RE.findall(text)]
+        if m_delims:
+            split_idx = m_delims[-1].end()
+            head = self._buffer[:split_idx]
+            self._buffer = self._buffer[split_idx:]
+        elif len(self._buffer) >= 256:
+            head = self._buffer
+            self._buffer = ""
+        else:
+            return None
+
+        tokens = []
+        for m in _TOKEN_RE.finditer(head):
+            tokens.append((m.group(0).lower(), m.lastgroup == "NUM"))
         if not tokens:
             return None
-        for token in tokens:
+
+        for token, is_num in tokens:
             if token == self.last_token:
                 self.same_run += 1
             else:
                 self.last_token = token
                 self.same_run = 1
+            self.numeric_run = self.numeric_run + 1 if is_num else 0
             self.recent_tokens.append(token)
             self.periodic_tokens.append(token)
         if len(self.recent_tokens) > 96:
@@ -466,14 +495,20 @@ class _DegenerateStreamGuard:
             self.periodic_tokens = self.periodic_tokens[-2048:]
 
         reason = None
-        if self.same_run >= 28 and self.total_chars >= 100:
+        if self.same_run >= 28 and self.total_chars >= 100 and not _is_numeric_token(self.last_token):
             reason = f"repeated '{self.last_token}' {self.same_run} times"
         elif len(self.recent_tokens) >= 72:
             top = max(set(self.recent_tokens), key=self.recent_tokens.count)
             count = self.recent_tokens.count(top)
-            if count >= 60 and count / max(len(self.recent_tokens), 1) >= 0.78:
+            if count >= 60 and count / max(len(self.recent_tokens), 1) >= 0.78 and not _is_numeric_token(top):
                 reason = f"repeated '{top}' {count}/{len(self.recent_tokens)} recent tokens"
-        if not reason and len(self.recent_tokens) >= 48:
+        if (
+            not reason
+            and len(self.recent_tokens) >= 48
+            # A window that is mostly numbers is matrix/table data; its
+            # periodicity is real. Mixed loops still hit the exact-block check.
+            and sum(_is_numeric_token(t) for t in self.recent_tokens) * 2 < len(self.recent_tokens)
+        ):
             # Detect a periodic suffix instead of counting repeated n-grams.
             # Reused property names are normal in CSS/SVG/JSON; a collapsed
             # generation repeats nearly every token at a short fixed period.
@@ -508,12 +543,15 @@ class _DegenerateStreamGuard:
                 if history[previous:previous + anchor_size] != anchor:
                     continue
                 period = current_anchor - previous
-                if period < 24 or len(history) < period * 2:
+                if period < 24 or len(history) < period * 2 or self.numeric_run >= period:
                     continue
-                if history[-2 * period:-period] == history[-period:]:
+                if history[-2 * period:-period] == history[-period:] and not all(_is_numeric_token(t) for t in history[-period:]):
                     reason = f"repeated an exact {period}-token block twice"
                     break
 
+        if not reason and self.numeric_run >= 2048:
+            reason = f"emitted {self.numeric_run} numbers in a row"
+
         if not reason:
             return None
 
@@ -1556,8 +1594,9 @@ def _build_chatgpt_responses_payload(
         "stream": stream,
         "store": False,
     }
-    if not _restricts_temperature(model):
+    # ChatGPT plan Responses requests reject temperature for every model.
-        payload["temperature"] = temperature
+    # The common caller still passes its preset value, but this provider must
+    # omit the field even for model names outside the known reasoning families.
     # ChatGPT Subscription Codex API does not support max_output_tokens —
     # passing it returns HTTP 400 "Unsupported parameter: max_output_tokens".
     # Do not include it in the payload.
@@ -2282,6 +2321,25 @@ def _apply_explicit_nonstream_reasoning_effort(
         payload["reasoning_effort"] = effort
 
 
+def _nonstream_openai_response_text(msg: Dict, thinking_mode: Optional[str]) -> str:
+    """Return the answer text of an OpenAI-compatible non-streaming message.
+
+    With thinking off the caller wants only the answer, so an empty answer
+    stays empty even when the model spent its whole budget on reasoning.
+    Otherwise reasoning stands in for a missing answer, and Mistral's
+    structured thinking is kept ahead of its text.
+    """
+    thinking_off = _normalize_thinking_mode(thinking_mode) == "off"
+    reasoning = "" if thinking_off else msg.get("reasoning_content")
+    content = msg.get("content")
+    if not isinstance(content, list):
+        return content or reasoning or ""
+    text_part, thinking_part = _normalize_mistral_content(content)
+    if thinking_part and not thinking_off:
+        return thinking_part + "\n\n" + (text_part or "")
+    return text_part or reasoning or ""
+
+
 def llm_call(url: str, model: str, messages: List[Dict], temperature: float = LLMConfig.DEFAULT_TEMPERATURE,
              max_tokens: int = LLMConfig.DEFAULT_MAX_TOKENS, headers: Optional[Dict] = None,
              timeout: int = LLMConfig.DEFAULT_TIMEOUT, prompt_type: Optional[str] = None,
@@ -2373,17 +2431,9 @@ def llm_call(url: str, model: str, messages: List[Dict], temperature: float = LL
         elif provider == "ollama":
             response = _parse_ollama_response(data)
         else:
-            msg = data["choices"][0]["message"]
+            response = _nonstream_openai_response_text(
-            content = msg.get("content")
+                data["choices"][0]["message"], thinking_mode
-            if isinstance(content, list):
+            )
-                # Mistral structured content — extract thinking + text
-                text_part, thinking_part = _normalize_mistral_content(content)
-                if thinking_part:
-                    response = thinking_part + "\n\n" + (text_part or "")
-                else:
-                    response = text_part or msg.get("reasoning_content") or ""
-            else:
-                response = content or msg.get("reasoning_content") or ""
         _set_cached_response(cache_key, response)
         return response
     except Exception:
@@ -2821,18 +2871,9 @@ async def llm_call_async(
                 elif provider == "ollama":
                     response = _parse_ollama_response(data)
                 else:
-                    msg = data["choices"][0]["message"]
+                    response = _nonstream_openai_response_text(
-                    content = msg.get("content")
+                        data["choices"][0]["message"], thinking_mode
-                    if isinstance(content, list):
+                    )
-                        # Mistral structured content — extract thinking + text
-                        # (same contract as llm_call / stream_llm; see #5435).
-                        text_part, thinking_part = _normalize_mistral_content(content)
-                        if thinking_part:
-                            response = thinking_part + "\n\n" + (text_part or "")
-                        else:
-                            response = text_part or msg.get("reasoning_content") or ""
-                    else:
-                        response = content or msg.get("reasoning_content") or ""
                 _set_cached_response(
                     cache_key,
                     response,

src/mcp_manager.py

@@ -10,6 +10,7 @@ import logging
 import os
 import re
 import asyncio 
+import contextlib
 from typing import Any, Dict, List, Optional, Set, Tuple
 from src.database import McpServer, SessionLocal
 
@@ -383,12 +384,21 @@ class McpManager:
 
             provider = build_provider(server_id, url, on_redirect=_on_redirect)
             stack = AsyncExitStack()
+            try:
                 transport = await stack.enter_async_context(streamablehttp_client(url, auth=provider))
                 read_stream, write_stream, _get_session_id = transport
                 session = await stack.enter_async_context(ClientSession(read_stream, write_stream))
                 await session.initialize()
 
                 tools_result = await session.list_tools()
+            except BaseException:
+                # Close the half-open transport here, in the task that entered
+                # it. Left for the garbage collector, its AnyIO cancel scope is
+                # exited from the wrong task and the event loop keeps spinning
+                # (#5518). Its own teardown error would only mask the real one.
+                with contextlib.suppress(Exception):
+                    await stack.aclose()
+                raise
             tools = []
             for tool in tools_result.tools:
                 tools.append({
@@ -760,8 +770,11 @@ class McpManager:
                     qualified.add(f"mcp__{server_id}__{tool['name']}")
         return disabled_map, qualified
 
-    def is_builtin(self, server_id: str) -> bool:
+    @staticmethod
+    def is_builtin(server_id: str) -> bool:
         """Check if a server is a built-in (auto-registered) server."""
+        if not isinstance(server_id, str):
+            return False
         return server_id.startswith("builtin_") or server_id in {
             "image_gen",
             "memory",

src/memory.py

@@ -364,32 +364,32 @@ class MemoryManager:
             return []
             
         # Define keyword categories for semantic matching
-        identity_words = ["name", "who", "i", "am", "called", "identity", "myself", "me", "my"]
+        identity_words = ["name", "called", "identity", "myself"]
         contact_words = ["phone", "email", "address", "contact", "number", "where", "located", "reach"]
         preference_words = ["like", "prefer", "favorite", "want", "love", "hate", "dislike", "enjoy", "interested"]
         task_words = ["todo", "task", "remind", "meeting", "appointment", "schedule", "deadline"]
         fact_words = ["what", "when", "where", "how", "why", "explain", "describe", "information", "know"]
         
         query_lower = query.lower()
+        query_words = set(re.findall(r'\w+', query_lower))
+        query_tokens = set(tokenize(query_lower))
         
         # Determine query type based on keywords
         query_type = None
-        if any(word in query_lower for word in identity_words):
+        if query_words.intersection(identity_words) or re.search(r'\bwho\s+am\s+i\b', query_lower):
             query_type = "identity"
-        elif any(word in query_lower for word in contact_words):
+        elif query_words.intersection(contact_words):
             query_type = "contact"
-        elif any(word in query_lower for word in preference_words):
+        elif query_words.intersection(preference_words):
             query_type = "preference"
-        elif any(word in query_lower for word in task_words):
+        elif query_words.intersection(task_words):
             query_type = "task"
-        elif any(word in query_lower for word in fact_words):
+        elif query_words.intersection(fact_words):
             query_type = "fact"
         
         relevant = []
-        identity_memories = []
+        # Every memory participates in the same ranking and threshold check.
-        other_memories = []
+        # Query type is a scoring hint, not an unconditional insertion path.
-        
-        # Separate identity memories from others
         for memory in memories:
             memory_text = memory["text"].lower()
             # Check if this is an identity memory (contains name patterns or identity indicators)
@@ -397,22 +397,7 @@ class MemoryManager:
                 re.search(r'\b[A-Z][a-z]+ [A-Z][a-z]+\b', memory["text"]),
                 any(word in memory_text for word in ["name is", "i'm", "i am", "called", "my name", "named", "call me"])
             ])
-            if is_identity:
-                identity_memories.append(memory)
-            else:
-                other_memories.append(memory)
-        
-        # For identity queries, include all identity memories regardless of similarity
-        if query_type == "identity" and identity_memories:
-            # Give them high scores to ensure they're included first
-            for memory in identity_memories:
-                relevant.append((0.9, memory))  # High score for identity memories in identity queries
-        
-        # Process other memories with similarity scoring
-        for memory in other_memories:
-            memory_text = memory["text"].lower()
             memory_tokens = set(tokenize(memory_text))
-            query_tokens = set(tokenize(query_lower))
             
             # Calculate base Jaccard similarity
             if not query_tokens or not memory_tokens:
@@ -422,7 +407,11 @@ class MemoryManager:
             final_score = base_similarity
             
             # Apply boosts based on semantic matching
-            if query_type == "contact":
+            if query_type == "identity" and is_identity:
+                # Keep explicit name/identity recall useful without outranking
+                # a closer textual match or bypassing the caller's threshold.
+                final_score = max(final_score * 1.3, 0.6)
+            elif query_type == "contact":
                 # Boost memories with contact information
                 has_contact_info = any(word in memory_text for word in ["@gmail.com", "@", ".com", 
                                                                      "phone", "number", "address", 

src/model_discovery.py

@@ -193,7 +193,10 @@ class ModelDiscovery:
             data = r.json()
             # Some OpenAI-compatible servers return a bare list, not {"data": [...]}.
             items = data if isinstance(data, list) else ((data or {}).get("data") or [])
-            ids = [m.get("id") for m in items if isinstance(m, dict) and m.get("id")]
+            ids = [
+                m["id"] for m in items
+                if isinstance(m, dict) and isinstance(m.get("id"), str) and m["id"]
+            ]
             if ids:
                 return {
                     "host": host,

src/pdf_form_doc.py

@@ -102,8 +102,10 @@ _PLAIN_FRONT_MATTER_RE = re.compile(
 #   - **label:** value <!-- field=NAME-ENC type=text -->
 #   - **label** [opts]: value <!-- field=NAME-ENC type=choice -->
 #   - [x] **label** <!-- field=NAME-ENC type=checkbox -->
+# Accept Unicode word characters for older saved markers, whose encoder left
+# non-ASCII letters/numbers literal. Newly rendered names are ASCII-safe.
 _FIELD_BULLET_RE = re.compile(
-    r'^\s*-\s+(?P<body>.*?)\s*<!--\s*field=(?P<name>[A-Za-z0-9_.%-]+)\s+type=(?P<type>\w+)\s*-->\s*$'
+    r'^\s*-\s+(?P<body>.*?)\s*<!--\s*field=(?P<name>[\w.%-]+)\s+type=(?P<type>\w+)\s*-->\s*$'
 )
 
 
@@ -115,7 +117,7 @@ def _encode_name(name: str) -> str:
     """
     out = []
     for ch in name or "":
-        if ch.isalnum() or ch in ("_", ".", "-"):
+        if (ch.isascii() and ch.isalnum()) or ch in ("_", ".", "-"):
             out.append(ch)
         else:
             for b in ch.encode("utf-8"):

src/report_math.py

@@ -0,0 +1,87 @@
+"""Preserve report TeX for markdown.js's existing deferred KaTeX renderer."""
+import re
+from html.parser import HTMLParser
+from xml.etree import ElementTree
+
+from markdown.blockprocessors import BlockProcessor
+from markdown.extensions import Extension
+from markdown.inlinepatterns import InlineProcessor
+from markdown.util import AtomicString
+
+
+# Keep delimiter and currency rules aligned with mathBlocks in markdown.js.
+_ESCAPED_MATH = r"\\\[([\s\S]*?)\\\]|\\\(([^\n]*?)\\\)"
+_DOLLAR_MATH = (
+    r"\$\$([\s\S]*?)\$\$|"
+    r"(?<![\$\d])\$(?!\$)(?=\S)([^\$\n]+?)(?<=\S)\$(?!\$|\d)"
+)
+
+
+def _placeholder(tex, display):
+    node = ElementTree.Element("span", {
+        "class": "ody-math-pending", "data-display": str(display).lower(),
+    })
+    node.text = AtomicString(tex)
+    return node
+
+
+class _HtmlContext(HTMLParser):
+    """Inline HTML is still raw when math runs, including tag attributes."""
+    def __init__(self):
+        super().__init__()
+        self.protected = []
+
+    def handle_starttag(self, tag, attrs):
+        if tag in {"code", "pre", "script", "style", "textarea"}:
+            self.protected.append(tag)
+
+    def handle_endtag(self, tag):
+        if tag in self.protected:
+            del self.protected[self.protected.index(tag):]
+
+
+class _InlineMath(InlineProcessor):
+    def handleMatch(self, match, data):
+        prefix = data[:match.start()]
+        if "<" in prefix:
+            context = _HtmlContext()
+            context.feed(prefix)
+            if context.protected or re.match(r"<(?:/?[A-Za-z]|[!?])", context.rawdata.lstrip()):
+                return None, None, None
+        # Backticks run before this processor. Preserve a backslash-escaped
+        # dollar too; math must run before Markdown consumes \( and \[.
+        if prefix.endswith("\\") and (len(prefix) - len(prefix.rstrip("\\"))) % 2:
+            return None, None, None
+        tex = next(group for group in match.groups() if group is not None)
+        display = match.group(1) is not None
+        return _placeholder(tex, display), match.start(), match.end()
+
+
+class _DisplayMath(BlockProcessor):
+    def test(self, parent, block):
+        return re.match(r"^ {0,3}(\$\$|\\\[)", block) is not None
+
+    def run(self, parent, blocks):
+        opening = re.match(r"^ {0,3}(\$\$|\\\[)", blocks[0])
+        closing = "$$" if opening.group(1) == "$$" else "\\]"
+        source = blocks[0][opening.end():]
+        consumed = 1
+        while closing not in source:
+            if consumed == len(blocks):
+                return False  # Leave incomplete input to ordinary Markdown.
+            source += "\n\n" + blocks[consumed]
+            consumed += 1
+        tex, _, tail = source.partition(closing)
+        del blocks[:consumed]
+        if tail.strip():
+            blocks.insert(0, tail)
+        parent.append(_placeholder(tex, True))
+
+
+class ReportMathExtension(Extension):
+    def extendMarkdown(self, md):
+        # Backticks (190) and indented code (80) retain precedence.
+        md.inlinePatterns.register(_InlineMath(_ESCAPED_MATH, md), "report_math_escaped", 185)
+        # Parse links before dollar math so URL destinations stay literal.
+        md.inlinePatterns.register(_InlineMath(_DOLLAR_MATH, md), "report_math_dollar", 155)
+        md.parser.blockprocessors.register(_DisplayMath(md.parser), "report_math", 79)

src/service_health.py

@@ -76,9 +76,23 @@ _ERROR_DETAIL = {
     "auth_or_protocol_error": "authentication or protocol error",
     "no_models": "endpoint returned no models",
     "no_host": "no host configured",
+    "config_source_error": "configuration source could not be read",
     "error": "probe failed",
 }
 
+# Config/inventory source behind each network subsystem. `_gather_inputs()`
+# fails soft, so a source that could not be read hands its probe an empty
+# inventory — indistinguishable from "nothing configured", which every probe
+# answers with `disabled`. The collector therefore reports the source failure
+# instead of running a probe whose verdict would be a lie.
+CONFIG_SOURCE_ERROR = "config_source_error"
+_SUBSYSTEM_SOURCE = {
+    "searxng": "settings",
+    "ntfy": "integrations",
+    "email": "accounts",
+    "providers": "endpoints",
+}
+
 
 def _svc(name: str, status: str, detail: str, **meta: Any) -> Dict[str, Any]:
     return {"name": name, "status": status, "detail": detail, "meta": dict(meta)}
@@ -411,26 +425,33 @@ def _gather_inputs() -> Dict[str, Any]:
     """Pull live config/account/endpoint lists from the app's data sources.
 
     Each lookup fails soft: a broken source yields an empty/neutral value so a
-    single failure can't take down the whole health report.
+    single failure can't take down the whole health report. The failure is also
+    *recorded* under `failed` (source -> controlled category), because the empty
+    value alone cannot be told apart from a source that loaded and is genuinely
+    empty — and the probes read empty as `disabled`.
     """
     settings: Dict[str, Any] = {}
     integrations: List[Dict[str, Any]] = []
     accounts: List[Dict[str, Any]] = []
     endpoints: List[Dict[str, Any]] = []
+    failed: Dict[str, str] = {}
     try:
         from src.settings import load_settings
         settings = load_settings() or {}
     except Exception as e:
+        failed["settings"] = _classify_error(e)
         logger.debug(f"service_health: settings load failed: {e}")
     try:
         from src.integrations import load_integrations
         integrations = load_integrations() or []
     except Exception as e:
+        failed["integrations"] = _classify_error(e)
         logger.debug(f"service_health: integrations load failed: {e}")
     try:
         from routes.email_helpers import _list_email_accounts
         accounts = _list_email_accounts() or []
     except Exception as e:
+        failed["accounts"] = _classify_error(e)
         logger.debug(f"service_health: email accounts load failed: {e}")
     try:
         from core.database import SessionLocal, ModelEndpoint
@@ -443,9 +464,10 @@ def _gather_inputs() -> Dict[str, Any]:
         finally:
             db.close()
     except Exception as e:
+        failed["endpoints"] = _classify_error(e)
         logger.debug(f"service_health: endpoint load failed: {e}")
     return {"settings": settings, "integrations": integrations,
-            "accounts": accounts, "endpoints": endpoints}
+            "accounts": accounts, "endpoints": endpoints, "failed": failed}
 
 
 async def _run_subsystem(name: str, fn: Callable, *args: Any) -> Dict[str, Any]:
@@ -464,6 +486,24 @@ async def _run_subsystem(name: str, fn: Callable, *args: Any) -> Dict[str, Any]:
         return _svc(name, DOWN, _detail_for(category), error=category)
 
 
+async def _source_unavailable(name: str, source: str,
+                              category: str) -> Dict[str, Any]:
+    """Controlled entry for a subsystem whose config source could not be read.
+
+    Deliberately not `disabled`: the configured state is *unknown*, not known to
+    be off. `degraded` rather than `down` because the failure is in reading the
+    inventory, not a probe that reached the subsystem and failed — but it still
+    carries non-zero severity, so a report containing one can never roll up to
+    `ok`. `meta` names the source and the controlled category of the underlying
+    exception; never its text.
+    """
+    return _svc(name, DEGRADED,
+                f"Could not read configuration "
+                f"({_detail_for(CONFIG_SOURCE_ERROR)}).",
+                error=CONFIG_SOURCE_ERROR, source=source,
+                source_error=category)
+
+
 async def collect_service_health(rag_manager: Any = None,
                                  memory_vector: Any = None) -> Dict[str, Any]:
     """Run every probe and return {overall, services, timestamp}.
@@ -472,6 +512,10 @@ async def collect_service_health(rag_manager: Any = None,
     four network subsystems run concurrently, each under `_SUBSYSTEM_DEADLINE`,
     with an overall `_AGGREGATE_DEADLINE` backstop. Per-item probes inside
     providers/email are themselves bounded by `_FANOUT_BUDGET`.
+
+    A subsystem whose config/inventory source failed to load is reported via
+    `_source_unavailable()` instead of being probed with an empty inventory,
+    which every probe would otherwise report as `disabled`.
     """
     from datetime import datetime, timezone
 
@@ -482,12 +526,23 @@ async def collect_service_health(rag_manager: Any = None,
     chroma = chromadb_health(rag_manager, memory_vector)
 
     names = ["searxng", "ntfy", "email", "providers"]
-    coros = [
+    probes = {
-        _run_subsystem("searxng", searxng_health, settings),
+        "searxng": (searxng_health, (settings,)),
-        _run_subsystem("ntfy", ntfy_health, inputs["integrations"], settings),
+        "ntfy": (ntfy_health, (inputs["integrations"], settings)),
-        _run_subsystem("email", email_health, inputs["accounts"]),
+        "email": (email_health, (inputs["accounts"],)),
-        _run_subsystem("providers", providers_health, inputs["endpoints"]),
+        "providers": (providers_health, (inputs["endpoints"],)),
-    ]
+    }
+    # A subsystem whose inventory could not be read is reported as such. The
+    # others are still probed normally — one broken source degrades one entry.
+    failed = inputs.get("failed") or {}
+    coros = []
+    for name in names:
+        source = _SUBSYSTEM_SOURCE[name]
+        if source in failed:
+            coros.append(_source_unavailable(name, source, failed[source]))
+        else:
+            fn, args = probes[name]
+            coros.append(_run_subsystem(name, fn, *args))
     try:
         results = await asyncio.wait_for(asyncio.gather(*coros),
                                          timeout=_AGGREGATE_DEADLINE)

src/session_actions.py

@@ -12,16 +12,6 @@ from datetime import datetime, timedelta, timezone
 
 logger = logging.getLogger(__name__)
 
-# Names that indicate a throwaway/test session
-_THROWAWAY_NAMES = {
-    "test", "testing", "asdf", "asd", "hello", "hi", "hey",
-    "yo", "sup", "hola", "hii", "hiii", "heyo",
-    "foo", "bar", "baz", "tmp", "temp", "scratch", "untitled",
-    "new chat", "delete", "remove", "junk", "trash", "xxx",
-    "abc", "qwerty", "blah", "stuff", "whatever", "idk",
-    "ok", "lol", "bruh", "hmm", "hm", "meh",
-}
-_THROWAWAY_MAX_MESSAGES = 4
 _FRESH_EMPTY_SESSION_GRACE = timedelta(minutes=10)
 _FRESH_SESSION_GRACE = _FRESH_EMPTY_SESSION_GRACE
 
@@ -54,14 +44,15 @@ def is_session_recently_active(row, now=None, grace=_FRESH_SESSION_GRACE) -> boo
 
 
 async def run_auto_sort(owner: str, skip_llm: bool = False, delete_throwaway: bool = True) -> str:
-    """Run session cleanup + (optional) AI folder sort for the given owner.
+    """Remove empty/Incognito sessions and optionally organize saved chats.
 
     Args:
         owner: user whose sessions to process
-        skip_llm: when True, do only Phase 1 (delete empty/throwaway sessions);
+        skip_llm: when True, do only Phase 1 (delete empty/Incognito sessions);
             skip Phase 2 (AI folder assignment). Used by the built-in daily
             background sweep so it never burns LLM tokens.
-        delete_throwaway: when False, only empty/incognito sessions are deleted.
+        delete_throwaway: retained for caller compatibility. Ordinary sessions
+            with saved messages are always preserved, regardless of this flag.
 
     Returns a human-readable summary of what was done.
     """
@@ -71,7 +62,7 @@ async def run_auto_sort(owner: str, skip_llm: bool = False, delete_throwaway: bo
 
     db = SessionLocal()
     try:
-        # ── Phase 1: Delete empty/throwaway sessions ──
+        # ── Phase 1: Delete only empty and explicitly Incognito sessions ──
         deleted_empty = 0
         deleted_throwaway = 0
 
@@ -93,47 +84,13 @@ async def run_auto_sort(owner: str, skip_llm: bool = False, delete_throwaway: bo
             if is_session_recently_active(row, now=cleanup_now):
                 continue
 
-            msg_count = db.query(DbMsg.id).filter(
+            has_messages = db.query(DbMsg.id).filter(
                 DbMsg.session_id == row.id
-            ).limit(_THROWAWAY_MAX_MESSAGES + 1).count()
+            ).first() is not None
-            should_delete = False
+            if not has_messages:
-
-            if msg_count == 0:
                 if is_fresh:
                     continue
-                should_delete = True
                 deleted_empty += 1
-            elif delete_throwaway and msg_count <= _THROWAWAY_MAX_MESSAGES:
-                name = (row.name or "").strip().lower()
-                first_msg = db.query(DbMsg.content).filter(
-                    DbMsg.session_id == row.id, DbMsg.role == "user"
-                ).order_by(DbMsg.timestamp).first()
-                first_text = (first_msg[0] or "").strip().lower() if first_msg else ""
-                assistant_count = db.query(DbMsg.id).filter(
-                    DbMsg.session_id == row.id, DbMsg.role == "assistant"
-                ).limit(1).count()
-
-                if name in _THROWAWAY_NAMES or name.startswith("chat:") or first_text in _THROWAWAY_NAMES:
-                    should_delete = True
-                    deleted_throwaway += 1
-                elif msg_count == 1 and assistant_count == 0:
-                    should_delete = True
-                    deleted_throwaway += 1
-                elif msg_count <= 4 and first_text and len(first_text.split()) <= 8 and len(first_text) <= 80:
-                    # Short trivial chats — e.g. "write hi to a friend" → "Hi!"
-                    should_delete = True
-                    deleted_throwaway += 1
-                else:
-                    # Aggressive: total message text under 250 chars combined = trivial
-                    msg_rows = db.query(DbMsg.content).filter(
-                        DbMsg.session_id == row.id
-                    ).all()
-                    total_chars = sum(len(m[0] or "") for m in msg_rows)
-                    if total_chars <= 250:
-                        should_delete = True
-                        deleted_throwaway += 1
-
-            if should_delete:
                 db.delete(row)
 
         if deleted_empty or deleted_throwaway:

src/session_search.py

@@ -57,25 +57,53 @@ def _message_to_context(msg: DBChatMessage) -> dict[str, Any]:
     }
 
 
-def _escape_like(value: str) -> str:
+def escape_like_pattern(value: str) -> str:
+    """Escape SQL LIKE wildcards (``%``, ``_``) and the escape char itself.
+
+    Public shared helper for building ``ilike(f"%{...}%", escape="\\\\")``
+    patterns from user input. Must escape backslash first.
+    """
     return value.replace("\\", "\\\\").replace("%", "\\%").replace("_", "\\_")
 
 
+def _escape_like(value: str) -> str:
+    return escape_like_pattern(value)
+
+
 def _snippet(content: str, query: str, radius: int = 60) -> str:
     content = content or ""
     query = query or ""
     if not query:
         return content[: radius * 2]
 
-    idx = content.lower().find(query.lower())
+    match_start = -1
-    if idx == -1:
+    match_length = 0
+    lowered_content = content.lower()
+    for candidate in _query_terms(query):
+        idx = lowered_content.find(candidate.lower())
+        if idx != -1 and (match_start == -1 or idx < match_start):
+            match_start = idx
+            match_length = len(candidate)
+
+    if match_start == -1:
         return content[: radius * 2]
 
-    start = max(0, idx - radius)
+    start = max(0, match_start - radius)
-    end = min(len(content), idx + len(query) + radius)
+    end = min(len(content), match_start + match_length + radius)
     return ("..." if start > 0 else "") + content[start:end] + ("..." if end < len(content) else "")
 
 
+def _query_terms(query: str) -> list[str]:
+    """Return meaningful quoted phrases and tokens in user-entered order."""
+    terms: list[str] = []
+    for match in re.finditer(r'"([^"]+)"|[\w][\w._-]*', query, flags=re.UNICODE):
+        term = match.group(1) if match.group(1) is not None else match.group(0).strip("._-")
+        term = term.strip()
+        if term:
+            terms.append(term)
+    return terms
+
+
 def _sanitize_fts_query(query: str) -> str | None:
     """Convert free text into a conservative FTS5 MATCH query.
 

src/settings.py

@@ -46,6 +46,7 @@ DEFAULT_SETTINGS = {
     "agent_email_confirm": True,
     "image_gen_enabled": False,
     "image_model": "",
+    "image_edit_model": "",
     "image_quality": "medium",
     # Optional editor planner/caption endpoints. Empty uses configured chat
     # fallback without probing a developer-specific host.
@@ -119,6 +120,7 @@ DEFAULT_SETTINGS = {
     # Tune via Settings or by editing data/settings.json.
     "research_run_timeout_seconds": 1800,
     "agent_max_tool_calls": 0,
+    "tool_approval_mode": "ask",
     "agent_max_rounds": 20,  # per-message agent step cap (clamped 1..200)
     # Soft input-token budget for the agent loop. The DEFAULT value (6000) is the
     # "auto" sentinel: it means "scale the budget to the model's context window"
@@ -294,7 +296,7 @@ def is_setting_overridden(key: str) -> bool:
 # resolved by FastAPI deps; an empty/None owner falls through to the global.
 _PER_USER_KEYS = {
     "vision_model", "vision_enabled", "vision_model_fallbacks",
-    "image_model", "image_gen_enabled", "image_quality",
+    "image_model", "image_edit_model", "image_gen_enabled", "image_quality",
     # Default chat endpoint / model — without per-user resolution every new
     # account inherited whatever the most-recent admin picked, which then
     # got injected into the chat composer on first open.
@@ -313,6 +315,20 @@ def get_user_setting(key: str, owner: str = "", default: Any = None) -> Any:
     Falls back gracefully if the prefs module can't be imported (cycle/early
     boot) — admin-global settings keep working.
     """
+    if key == "image_gen_enabled":
+        global_val = get_setting("image_gen_enabled", default if default is not None else False)
+        if not global_val:
+            return False
+        if owner:
+            try:
+                from routes.prefs_routes import _load_for_user
+                prefs = _load_for_user(owner) or {}
+                if "image_gen_enabled" in prefs and prefs["image_gen_enabled"] is not None:
+                    return bool(prefs["image_gen_enabled"])
+            except Exception:
+                pass
+        return bool(global_val)
+
     if owner and key in _PER_USER_KEYS:
         try:
             from routes.prefs_routes import _load_for_user