AI code review

developerz.ai includes a native AI code reviewer — the same loop that triages issues also watches PRs. It runs on your BYOK model, on your own box, with no per-review token markup. If you already have CodeRabbit, migrating is a one-liner.

CodeRabbit compatibility

The reviewer reads your existing .coderabbit.yaml and honours the fields below. You do not need to change anything. To migrate off CodeRabbit:

1. Uninstall the CodeRabbit GitHub App.
2. Keep your .coderabbit.yaml as-is (or delete it and switch to review: below).

That's it. The developer experience is identical; the inference bill goes to your key.

.coderabbit.yaml field support

Every key below is recognized. Honored keys are translated and applied; deferred keys are recognized but not yet acted on (the reviewer posts a warning comment so the gap stays visible); ignored keys are cosmetic or platform-specific and are silently no-op'd. Any key not listed is unknown and surfaces an unknown key warning.

{/* BEGIN coderabbit-generated: disposition-table */}

Honored — translated and applied

Deferred — recognized, not yet applied (the reviewer posts a warning comment)

Ignored — recognized platform/cosmetic settings, silently no-op

Generated from the KNOWN disposition registry (@developerz/review). Run bun run coderabbit:doc to regenerate. {/* END coderabbit-generated: disposition-table */}

Precedence. If a repo has both .coderabbit.yaml and a review: block in .maintainer.yml, the review: block wins.

Native review: block

Add a review: section to .maintainer.yml for the full feature set. The native block is a superset of .coderabbit.yaml:

# yaml-language-server: $schema=https://developerz.ai/schemas/maintainer.v1.json
version: 1

review:
  profile: chill                # chill | assertive | custom
  request_changes: true         # block merge on findings
  learnings: true               # write review learnings back to KB

  anti_noise:
    severity_floor: minor       # drop findings below this level
    max_comments: 30            # cap per-review comment volume

  auto_review:
    base_branches: [main, release/*]
    ignore_title_keywords: [WIP, "DO NOT MERGE"]

  path_filters:
    - "!**/*.lock"
    - "!dist/**"

  path_instructions:
    - path: "packages/**/*.ts"
      instructions: "Enforce the repo principles in docs/idea/principles.md."

anti_noise — quiet by default

The #1 complaint about CodeRabbit is noise. developerz.ai is quiet by default:

• Findings below severity_floor are dropped before they reach GitHub.
• Volume is capped at max_comments per push (blocking findings always bypass the cap).
• Identical findings on re-push are de-duplicated.

Knowledge base write-back (learnings: true)

When learnings is enabled, the reviewer writes what it learned from each review back to the repo knowledge base (kind: review-learning). On the next PR in the same repo, the reviewer reads those learnings before commenting — so the bot gets smarter over time without any manual maintenance.

Bot etiquette

The reviewer always defers when another review bot (CodeRabbit, Copilot, Dependabot) is active on the repo. There are no bot-on-bot loops, no overrides. To activate the native reviewer: uninstall the other bot.

Triggering a review

Reviews trigger automatically on every push to a PR. To re-run on the same commit:

# MCP (from Claude Code / Claude Desktop)
box: ci review rerun --repo owner/repo --pr 42

Or use the review MCP verb directly — see the MCP verb reference.

Review output

The reviewer posts:

1. High-level summary — one paragraph, what changed and why.
2. Inline comments — per-hunk findings at the diff line.
3. Status check — developerz/review passes (no blockers) or requests changes.

Every comment includes the disclosure footer and is logged to the audit trail.