Code-Quality Review

Follow AGENTS.md first. Run the mechanical pass first (make ci-local, and /code-review if available); this skill is the judgment layer on top.

Scan order (security-first)

Review in this order — don't rush past a step to get to style nits:

1. Secrets — hardcoded tokens/keys/URLs that belong in runtime env.
2. Input validation — untrusted args validated against Pydantic schemas/ranges before use.
3. Trust boundary — no caller-token passthrough to upstreams; fixed upstream host (no SSRF); Origin validation intact. (For the full boundary review, use security-review.)
4. Dependencies — new deps pinned in uv.lock, mainstream, no post-install hooks.
5. Correctness — logic, error handling, edge cases, response-envelope shape (success + flat error_code/message/_meta).
6. Efficiency — N+1 upstream calls, unbounded loops/inputs, missing size caps.
7. Style / fleet rules — modern typing (X | None), the per-file line budget (scripts/check_file_size.py), thin handlers, no PII in logs.

Report each finding

Group by file. For each: severity (critical = bug/security · warning = fix before merge · suggestion = optional), the issue, and a concrete fix. Anchor to file:line.

Common mistakes

• Reporting style nits while missing a token-passthrough or SSRF regression.
• Vague findings ("improve error handling") with no location or fix.
• Approving a diff that grows a module past the LOC budget with no split.