M7 Capstone: pre-review

The pre-review agent that scores PRs and ranks what blocks production: path-scoped rules, a read-only reviewer with structured findings, a verify gate, blockers from repository signals, metrics against labels, and the case for running it.

Lessons

• 01 Review as a measurable pipeline
• 02 Conventions of a Go, TypeScript, proto and migrations repository
• 03 Reading a PR with a read-only agent
• 04 Blocker ranking from repository signals
• 05 From pre-review to what blocks prod
• 06 Interview angle

Exercise

M7 capstone: build capstone, a pre-review gate over the go-ts-mini fixture and its twelve pull requests, with tests and hints.

1. 01-rule-loader: load .claude/rules with their paths frontmatter, match globs with **, *, ? and braces, and select the rules for a PR’s changed files.
2. 02-review-runner: a read-only Agent SDK reviewer with a structured findings schema, a prompt built from the matched rules and the patch, and every failure raised as a review error.
3. 03-filter-and-dedupe: drop findings on untouched files and below the confidence threshold, merge duplicates, and let the verify gate override a summary that claims the tests pass.
4. 04-verify-gate: apply the patch to a copy and run go test and node --test, with no test run when the patch does not apply (needs Go and Node).
5. 05-blocker-ranking: conflict, failed verify, stale migration, missing second approval and unreviewed age, weighted and ranked per PR.
6. 06-review-metrics: match findings to labels by file, line window and category, one to one, and compute precision, recall and blocker recall.
7. 07-review-policy: decide merge, hold or block for all twelve PRs and name the categories that block without a human (judgment stage; needs Go and Node).
8. 08-live-review: review all twelve PRs with a real model and write the report with cost, verdicts and metrics (needs a Claude login, Go and Node).