[lawrence3699]

Fixes for the package findings of the 2026-11-05 review (pkg-code-*, pkg-pypi-*, pkg-docs-*, mirrors-crosslinks-1/7/8). Model revisions and weights are unchanged; the version is bumped to 0.6.4 (tag and PyPI release are done separately, after merge).

Code

• Error bodies of both shapes (pkg-code-1, pkg-code-v1): the client (parse_error_body), the guard and the MCP server read the hosted API's flat {"error": "<code>", "message": ...} as well as the nested contract body; no more http_error: unauthorized or AttributeError. JevStyleError.input_too_long covers max_tokens_exceeded (local, 400), input_too_large (hosted, 413) and the pre-0.5.0 input_budget_exceeded. The server keeps 400 max_tokens_exceeded; docs note the hosted 413.
• Guard fail-closed (pkg-code-3, pkg-code-v2): a 200 without a usable value for every question (no answers, wrong type, no number) is now an error → ask, even with emit_allow. tests/test_guard.py covers thresholds → ask/deny, emit_allow, 4xx flat/nested bodies (real local stub server), malformed 200s.
• Key routing (pkg-code-2): JEV_STYLE_API_KEY goes only to loopback servers or the server in JEV_STYLE_URL; any other base_url needs an explicit key.
• MCP (pkg-code-v3, pkg-code-v2): instructions are built per backend (budgets from GET /v1/models; remote --allow-remote --url says the state leaves the machine); HTTP error path tested for both body shapes.
• Local server (pkg-code-5): Host / Origin check (403 forbidden_host / forbidden_origin), loopback + --host only; --allow-any-host / create_app(allowed_hosts=None) to disable; --host 0.0.0.0 accepts any Host but same-host/loopback Origin only. Tests address the TestClient as http://127.0.0.1.
• Downloads (pkg-pypi-1, pkg-pypi-v1, pkg-pypi-v2): backend checked before fetching (--no-backend-check to override); GGUF refuses to download while no scorer can be found (--scorer, env, build/ in a cached snapshot, PATH) and says how to get one; download / offline failures are one error: line (DownloadError); missing --model-dir is error: file not found; JEV_STYLE_RELEASE validated before parsing; --backend mlx off Apple silicon points at [torch].
• Integrity (pkg-code-7): jev-style download verifies against manifest.json by default (--no-verify), serve / decide --verify pass verify=True to the runtime, models.verify_folder().
• Custom repos (pkg-code-4): trust_remote_code repos get generic per-backend file patterns instead of the 0.8B v3 file names.
• CLI help (pkg-pypi-5): top-level --help lists all eight commands; guard / guard-replay usage lines carry their own name.

Release, metadata, docs

• release.yml (pkg-code-6): manual runs publish to TestPyPI only; PyPI only from a version-checked tag. server.json (mirrors-crosslinks-7): description names the current models (2B v4 + 0.8B, text + images), version 0.6.4.
• pyproject (pkg-pypi-4, pkg-pypi-2, pkg-docs-7): Windows classifier + a windows-latest CI job (no torch); [mlx] pin explained in a comment; Models URL → the Hub profile (the v4 collection has no 0.8B); Discussions URL. NOTICE (pkg-pypi-6) lists 2B v3 / v4 sources.
• README (mirrors-crosslinks-1, pkg-docs-*): Decision Index 0.3 board entry (2B v4 listed 2026-11-04: Public 44.37 / Full 34.53 / Breadth 41.75; Vision 57.55 / 51.94; 0.8B v3 17.0 / 12.5) replaces "self-run, v3 34.05 on the board" (the outdated banner image is removed); hosted API context 8,192 / input 8,000, no fixed per-account quota, 429 + Retry-After; bare install = client only, [mlx] pin note; key rule; both error shapes; noul / choice examples replace the urgency score ones; Discussions + SECURITY links.
• Skills (pkg-docs-4, pkg-docs-5): serve / mcp / jev-style cover --release 2b, per-release GGUF scorers, model starting with jev-style-; stale "mcp extra" advice dropped; guard numbers marked as 0.8B v3; reference.md error table gains the hosted rows.
• SECURITY.md, CODE_OF_CONDUCT.md, PR template, issue config → Discussions (pkg-docs-12).

Verification

pytest -q: 290 passed (233 before). uv build + twine check --strict pass. CLI smoke: --help, decide --fake, guard-replay --fake, eval --fake, JEV_STYLE_RELEASE=9b → one-line error.

🤖 Generated with Claude Code

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has insufficient remaining quota. Try requesting a review at a lower effort level.

[lawrence3699]

Verification pass (fresh clone, Python 3.12 venv, .[dev] only, no torch/mlx):

• pytest -q: 292 passed (290 on the branch as opened, +2 tests added here). python -m build + twine check --strict pass; the wheel's METADATA carries License-Expression: Apache-2.0, License-File: LICENSE / NOTICE, the Windows classifier and the new Models / Discussions URLs. The wheel installs in a clean venv and runs the release workflow's smoke.
• Error bodies: the live api.jevstyle.com 401 ({"error":"unauthorized","message":...}) is parsed as unauthorized by the client, guard.error_fields and the MCP HttpBackend; nested, bare-message and non-JSON bodies covered; 413 input_too_large / 400 max_tokens_exceeded / 422 input_budget_exceeded all set input_too_long.
• Guard: a 200 with no / partial / wrong-type / non-numeric answers is ask (source error) even with emit_allow; thresholds still produce allow/ask/deny.
• Downloads: download --backend torch|mlx on a bare install, serve|decide --backend gguf without a scorer, HF_HUB_OFFLINE=1, and a missing --model-dir all stop with one error: line and the HF cache stays empty. All 9 pinned repos (and their bf16/8bit, F16/Q8_0/Q4_K_M variants) have a manifest.json whose sizes and sha256 match the Hub's file metadata, so the default download --verify passes on real repos.
• Serve: with a real serve --fake, loopback Host / [::1] / localhost:3000 Origin / same-origin Playground POST pass; spoofed Host, cross-site text/plain POST and Origin: null get 403; --host 0.0.0.0 accepts any Host but refuses foreign Origins; --allow-any-host accepts everything. The Python client, the MCP HttpBackend and jev-style guard --check all work against it unchanged.
• release.yml: pypi runs only on ref_type == 'tag', testpypi only on dispatch; version is 0.6.4 in __init__.py, server.json (both fields) and CHANGELOG. All 68 README links answer 200 (SECURITY.md 404s on main only until this merges). The Decision Index figures match data/index.json (public 44.37, balanced 34.53, breadth 41.75; 0.8B 17.0 / 12.5).

Two fixes pushed on top (763d733, 9df2666): jev-style guard is exempt from the new JEV_STYLE_RELEASE gate (a hook must exit 0), the reference.md example answer no longer shows an urgency score, and decide --url gets --api-key-env so a non-loopback server can be given a key explicitly (eval and cascades already had their own). CI is green on 9df2666.

[lawrence3699]

Added ec8873b from the 2026-11-05 SEO/GEO audit. It changes metadata only; no code paths change.

• jev-style-serve skill: the curl … | sh uv install line is replaced with brew install uv / pipx install uv. Skill directories graded this skill F for that line.
• jev-style-guard skill: the description now says "recursive deletes, force-pushes or key exfiltration" instead of literal commands.
• Plugin manifests: both now use homepage: https://jevstyle.com/, so plugin listings link to the site instead of only GitHub.
• Directory files: context7.json (Context7) and glama.json (Glama MCP directory claim) are added at the repo root. Neither file is in the sdist or wheel; I checked both build outputs.
• CHANGELOG: one new 0.6.4 bullet.

pytest: 292 passed. No real models were loaded.

🤖 Generated with Claude Code