Dependabot dependency-update PRs do not benefit from BLOCK/FIX/NIT/QUESTION
code notes. Route them to a dependency-risk assessment instead: the semver
update type, a safe/low_risk/risky/breaking call, the packages bumped, and
reasons, grounded in the Sea Haven Dependabot merge policy (patch/minor
generally safe; majors need changelog review; grouped PRs assessed at the
riskiest package). Feedback focuses on PR title/description quality.
review() dispatches on the author to a dependabot or code path, each
stamping a "_kind" so consumers can tell the shapes apart (missing "_kind"
reads as code, keeping older cached reviews valid). Enum fields are clamped
to allowlists with cautious defaults so a hallucinated or injected value
cannot reach the posted event. The handbook distillation also captures the
dependency policy, though the prompt carries it regardless.
Group the review queue into a collapsible section per repo (collapse
state persisted in localStorage), with PRs ordered oldest to newest by
creation date, so a large multi-repo queue is easier to scan.
Add an optional per-PR auto-merge control: a method choice (squash by
default per handbook, merge, or rebase) enables GitHub auto-merge via a
GraphQL mutation, so the PR merges once required checks pass. It only
fires when clicked; nothing merges automatically.
Back this with created_at and node_id from the PR search, two new
nullable store columns added via an idempotent PRAGMA-guarded migration,
and a cheap-gate metadata backfill so already-cached PRs gain node_id
without being re-reviewed. New endpoint POST /api/automerge.
Some Fireworks models emit chain-of-thought before the JSON review, and
that preamble can contain stray "{". The old first-"{"-to-last-"}" span
then grabbed reasoning braces and failed to parse (seen live on a real
PR). Scan each "{" and return the first substring that actually decodes
to an object instead, so a preamble or trailing prose no longer breaks
the review.
Feed the reviewer a distilled digest of the Sea Haven engineering-handbook
so findings reflect our naming, commit, PR, secrets, and IaC conventions
instead of generic code-review judgment.
A new handbook module keeps an app-managed shallow clone of the (private)
handbook, distills the review-relevant pages into a compact conventions
checklist via the Fireworks model once a day, caches it under ~/.cache,
and hands it to the reviewer to inject into every review's system prompt.
The refresh runs in-process at the start of each worker cycle; failures
keep the last good digest and back off, so a handbook outage never blocks
reviews. Set HANDBOOK_ENABLED=false to disable.
Extract a shared fireworks_complete helper used by both the reviewer and
the distiller, so the handbook provider needs no reviewer reference and
the guidance callable is set once at construction. Clone auth uses a
Basic http.extraHeader (GitHub git-over-HTTPS rejects Bearer), and the
distiller wraps its answer in delimiters to strip a reasoning model's
chain-of-thought preamble. Adds GET /api/handbook and a header status
line. Stdlib-only, no new dependencies.
Pre-compute PR reviews so they are ready the moment a PR is opened in
the queue, instead of waiting on an on-demand Fireworks call each time.
A daemon worker polls the queue every POLL_INTERVAL and reviews new or
changed non-draft PRs into a local SQLite cache (gitignored). Change
detection is two-level: skip when the PR's updated_at is unchanged, and
even when it moved, skip the model call when the diff's SHA-256 matches,
so comment-only bumps do not burn tokens. Failed reviews retry up to
MAX_REVIEW_ATTEMPTS with 429 backoff; departed PRs are closed with a
grace window before purge.
Singletons are initialized eagerly in the FastAPI lifespan before the
worker thread starts to avoid an init race; all cache writes are
serialized. New endpoints GET /api/reviews and POST /api/refresh back a
dashboard that polls for status (reviewing/ready/error) and opens a
ready review instantly. Nothing is posted automatically; the human
still decides. Stdlib-only, so no new runtime or test dependencies.
A single-user local dashboard that pulls open PRs from the
Sea-Haven-Industries org, reviews each with a Fireworks model in the
BLOCK/FIX/NIT/QUESTION format, and posts the review to GitHub as the
token owner. Runs only on localhost; secrets stay in a gitignored .env
and never reach the browser.
Structured as an app/ package plus a static/ frontend so the module
imports and static mount resolve. HTTP uses httpx2 (the runtime lib
starlette's TestClient now prefers), pinned in requirements.txt.
Includes a stdlib-only pytest suite (network mocked, no extra test
deps so CI needs only pytest) with a skip-by-default live Fireworks
test, and CI wired to the org ci-python-app reusable workflow to lint
app and tests and run the mocked suite fully offline. Dependabot covers
pip and github-actions.
