air — Automated Code Review with Verification, Pattern Learning, and Team Knowledge
Why
Code reviews are slow, inconsistent, and lose institutional knowledge when people leave. air fixes this:
- 5 specialized agents review in parallel — security, code quality, simplification, git history, and an optional Codex second opinion
- A verification agent filters false positives before posting — findings must score 60+ confidence
- Wiki-backed learning captures patterns from every review — author tendencies, service gotchas, accepted patterns
- Re-review tracking knows what was fixed and what wasn't — no starting over
- Your memory enriches reviews — Claude Code memory feeds institutional knowledge into every agent
One command. One consolidated PR comment. Gets smarter over time.
Install
/plugin marketplace add VorobiovD/air
/plugin install air@air
Two commands become available: /air:review-pr and /air:learn. Updates are automatic.
Prerequisites
- Claude Code — installed and running
- GitHub CLI (
gh) — authenticated via gh auth login
- Repo access — must be able to run
gh pr view on the target repo
- Codex plugin (optional) — if installed, runs as an additional reviewer. Skips gracefully if not available
- GitHub Wiki — auto-created on first run if missing. Used to store learned review patterns
Usage
/air:review-pr 123 # Full review of PR #123
/air:review-pr # Auto-detect: review current branch's PR, or self-review
/air:review-pr --self # Review local changes before pushing
/air:review-pr --self --fix # Self-review + auto-apply fixes
/air:review-pr --re-review # Delta review: FIXED/NOT FIXED tracking + new findings
/air:review-pr --fresh # Full review from scratch, new comment
/air:review-pr --rewrite # Full review, edit existing comment in place
/air:review-pr --respond # Reply to review: classify findings + self-check + push
/air:review-pr --full # Review entire codebase (first-time audit)
/air:review-pr --dry-run # Print to console, don't post to GitHub
/air:review-pr --no-codex # Skip Codex review pass
/air:review-pr https://github.com/org/repo/pull/45 # Cross-repo review
Smart Default (no flags)
When you run /air:review-pr with no arguments:
- Checks if the current branch has an open PR — if yes, reviews it
- If an existing review comment is found with new commits since — auto re-reviews
- If no PR exists but you have local changes — auto self-reviews
- If nothing to review — tells you
Re-review Mode
After posting a review, developers can respond to findings by number:
#3 — fixed → re-review checks if the code actually changed#5 — this is our standard pattern → evaluated with graduated resistance before accepting#8 — pre-existing, not from this PR → classified separately, not dropped
Re-review generates an inter-diff (only changes since last review) so agents focus on what's new.
Respond Mode
After fixing review findings (manually or with Claude Code), auto-generate the response:
/air:review-pr --respond
This reads the existing review, then for each finding:
- Auto-classifies as fixed/unfixed by checking if the flagged code changed
- Verifies fixes are correct — compares what was done vs what was suggested
- Asks you about unfixed findings (dispute/acknowledge/won't-fix) for non-obvious cases
- Runs the full review pipeline on your fix diff to catch regressions
- Detects additional changes beyond the fixes (refactors, new features) and lists them
Posts a structured response the reviewer's re-review can parse directly, then pushes the branch.
How It Works
Pipeline (13 steps)