P1
Will show up in production for users or callers if merged — crashes, wrong results, security/data harm, public contract breaks
Product
Every PR gets the same six categories. Pacecar does not invent a new rubric per author or per agent. Optional repo review.md items fold under these headings — they do not replace them.
No nits. Taste is the product: catch what will show up in production. Style, test pedagogy, and process chatter stay silent.
Severity
P1
Will show up in production for users or callers if merged — crashes, wrong results, security/data harm, public contract breaks
P2
Credible production risk; non-blocking watch item
Approval is applied from severity by rules: any P1 → request changes; only P2s → comment; nothing to raise → approve.
Categories
01
Access control, attack surface, secrets, and data privacy — OWASP-aligned themes, not a full audit.
02
The change matches what the PR claims — especially important for agent-written PRs.
03
The change stays inside the claimed area and fits the repo’s existing patterns.
04
The code works — success path, errors, edges, and calculation accuracy.
05
Safe in production — timeouts, migrations, and tests that cover the real risk.
06
Shared APIs, schemas, and callers stay compatible — or ship a migration plan.
Install once. Reviews run on draft and ready PRs — without training models on your private code.