Tools/ci/ai_review reviews a PR with Claude Code on Amazon Bedrock and
writes a pr-review artifact for pr-review-poster.py.
- Pipeline: gather (metadata, linked issues, commits, CI status, per-file
patches, earlier comments), a reviewer pass returning line findings,
PR-level findings and a contribution checklist, an independent
validator pass per finding, and code-owned routing: inline only at high
confidence, otherwise a before-merge list or a collapsed section. No PR
is skipped for size; generated files are listed but not diffed.
- Trust: the PR is untrusted data. The agent runs --bare --restricted
with read-only tools. Code it writes to check its math runs only in a
px4-dev container (no network, no credentials, read-only checkout,
gVisor in CI), and `harden` blocks the runner's metadata service first.
Output containing a credential is never written.
- A model refusal produces no review and passes the provider's exact
reason, category and request ID through.
- Models and per-role defaults live in models.json; models-check verifies
access before switching.
- Process checks (tests, description, upgrade notes, docs) live in a
CI-only prompt so the interactive skill stays lean.
Assisted-by: Claude:claude-opus-5-5
Signed-off-by: Ramon Roche <mrpollo@gmail.com>