Skip to content

AI first-pass review

A repo with ai_review set gets a first-pass review of each non-draft agent PR before a person reads it. Turn it on with hangar repo set acme/app --ai-review hangar. It is advisory. It never approves, never counts toward review capacity, and creates no inbox item.

Who reviews

  • hangar: a review job (source: review) on another harness than the author's: codex for claude-code's PRs, else claude-code, unless ai_review_harness names one. Its model is ai_review_model, by default frontier resolved to a concrete id.
  • check:<name>: an org review bot the team already licenses. hangar waits for a check run or status named <name> on the PR's head. success is approve; neutral and skipped are comment; failure and action_required are blocking; anything else is an error. No job runs; the rest of this page doesn't apply.

With only the author's harness, the review runs on it in a fresh session. It is flagged independent: false and doesn't satisfy the self-approval rule.

How the review job is fenced

The job runs in the PR job's environment with the PR's branch checked out. The PR diff and the review packet go in as untrusted context: agent-written code is input to the reviewer, never instructions. The job delivers nothing (deliver: none). Its sandboxes get egress none, and its git gateway grant is fetch-only, so it can't push.

Egress none needs hangar's egress proxy (Docker's egress network, or HANGAR_EGRESS_URL on GKE). Under HANGAR_INTAKE_POSTURE=enforce the job needs the git gateway. Where the sandbox can't be fenced, the review fails rather than running open.

Two passes: find, then confirm

  1. The review pass returns a verdict, a summary and findings. Each finding has a path, a line, a severity (info, minor, major or critical), the text and its evidence.
  2. The confirmation pass checks each finding. It runs on the same harness, fenced the same way, on the cheaper balanced model policy. It keeps a finding only with one of two kinds of evidence:
    • command: a command or test it ran that shows the problem, with the output lines that show it. hangar doesn't rerun it, so the finding is marked unverified.
    • quote: a line the PR adds to the finding's own file, within 5 lines of its line. hangar checks the diff for it. An invented quote, a removed or context line, or another file's line confirms nothing.

The check reads the reviewed head, not the branch tip. Findings it can't confirm are dropped. If the check doesn't complete or can't start, none is confirmed and the verdict stands.

Severity, verdict and inline comments

The verdict fails closed. A blocking review becomes comment only when the check answered every major or critical finding and confirmed none.

The App posts one PR review with event: COMMENT, never APPROVE, so it can't satisfy branch protection. It carries at most ai_review_inline_max inline comments (default 5, at most 20), the most severe first. The review body lists every confirmed finding, most severe first, with its evidence. Mentions in agent text are neutralized.

The review packet counts the findings in the body only and those dropped as unconfirmed.

Spend

Both passes share ai_review_budget_usd (default $1.50). The review pass may spend 70% of it; under a cent, no review runs. The confirmation pass gets the rest, if that is at least $0.05. A pass whose cost isn't known counts as its whole cap.

It counts in the PR's person's daily caps, and shows in usage under source=review (hangar usage --by source).

When it runs again

A push asks for a review of the new head, at most one every 5 minutes. A draft (a tainted PR opens as one) is reviewed when it is marked ready.

If either pass's job record is lost, the review becomes an error after 10 minutes and the head is reviewed again. The App's review on GitHub is a bot's, so it starts no auto-fix round.

Self-review before a person looks

With self_review: true too (--self-review), the PR opens as a draft and waits for its review, with no inbox item or review-capacity count. An independent blocking review starts one auto-fix round on the job's session, inside its job budget. The review of the round's push ends the wait, whatever it says.

So does a review that isn't blocking or independent, or a round that fails, pushes nothing or can't start. The PR is then marked ready, unless hangar would have opened it as a draft. The packet's Self-review line says what happened.