Menu

Field notes

Practical notes on reviewing agent-written pull requests.

  1. Kiro Code Review: Turn Specs Into a Merge Gate

    Connect Kiro specs, review findings, tests, and the exact current head with a spec-to-verdict receipt.

  2. Augment Code Review: Add a Current-Head Gate

    Use Augment for findings and fixes, then require an independent verdict for the exact current head.

  3. Amazon Mandating AI Code Review? What We Know

    Amazon denied the reported mandate. Use TRACE risk routing and an exact-head approval contract instead.

  4. Devin Code Review: Keep Review and Merge Separate

    A four-authority contract for review findings, branch fixes, repository checks, and the exact-head merge decision.

  5. Windsurf Code Review: Build a Current-Head Gate

    Separate local review, PR findings, autofixes, and the exact-head repository verdict.

  6. AI Code Review Benchmark: What to Measure

    A reproducible scorecard for recall, false blockers, repeatability, failures, cost, and merge-gate safety.

  7. Cursor Code Review: Use Bugbot in a Current-Head Gate

    Separate local review, Bugbot findings, and the current-head repository verdict.

  8. Codex Code Review: Turn Findings Into a Merge Gate

    Set up Codex review, write scoped AGENTS.md rules, and bind the verdict to the current head.

  9. Claude Code Review: Keep Four Roles Separate

    A practical operating model for authors, reviewers, evidence policy, and current-head merge gates.

  10. Pull request review comments: a signal-first guide

    A five-step contract for blockers, evidence, scope, resolution, and current-head review state.

  11. PR review prompts for agent-written pull requests

    Eight copy-ready prompts with a concrete evidence and finding-admission contract.

  12. AI code review on GitHub: a merge-safe workflow

    A practical guide to least-privilege setup, evidence-bound findings, and exact-head merge checks.

  13. Best AI code review tools for agent-written PRs

    A practical guide to score-first, merge-safe review tools and agent repair loops.

  14. How to tell if code is AI-generated: what actually works

    Why style is weak evidence, and the review gate that works regardless of who wrote the diff.