PR review
Hyrax reviews every pull request in your repo automatically, the moment it's opened, and again on every push.
Once the GitHub App is installed, no further setup is required. Reviews appear on every PR — from your team, from contractors, from Dependabot, from Hyrax itself. The review posts on the PR within a couple of minutes of the push.
Two modes
You pick one mode per repo, in Settings → PR review.
Scanner mode
Pattern-based, fast, cheap. Around $1–2 per PR.
The scanner runs over the diff, identifies suspicious lines using the same pattern library that powers a fast audit, and posts inline comments on the relevant lines. Each comment names the issue and the suggested fix. No reasoning, no context — just the patterns Hyrax knows.
Right for high-velocity repos where most PRs are small and you want a fast safety net without an LLM bill on every push.
LLM mode
Full agent reasoning. Around $1–2.50 per PR.
A single Claude agent reads the diff in context with the rest of the repo and any related repos, then posts a single rolling checklist comment on the PR. As the author pushes more commits, the comment updates in place — you never have to scroll through stale review threads to find the latest state.
The agent is capped at 7 findings per review, ranked by impact. Each is tiered:
- Must fix — block merge until addressed.
- Consider — advisory; reviewer's call.
Right for repos where the cost of merging a bad change is higher than the cost of an LLM review — services in production, libraries that other repos depend on, anything user-facing.
How tiers behave
You can wire must fix items to your branch protection rules so they actually block merge. The Hyrax review posts a check status that GitHub recognizes; configure it in your repo's branch protection settings the same way you'd configure any required check.
Consider items post as part of the same checklist comment but never block. They're suggestions for the reviewer to evaluate.
What the review sees
Both modes have access to:
- The PR diff (full file context, not just the patch).
- The base branch's existing code.
- Your repo's discovery bundle, if you've published one.
- Summaries of any related repos you've linked.
The discovery context is the single biggest quality lift. A repo with discovery published gets reviews that respect your conventions; a repo without one gets generic best-practice advice. If you're going to wire up PR review, run discovery first.
When the review runs
- On PR open — first review.
- On every subsequent push — review re-runs. The previous review's items are tracked in hidden JSON inside the comment, so the new review sees what was already raised and folds in only what changed.
- On PRs Hyrax itself opens (fix, task) — yes, Hyrax reviews its own PRs, with the same gauntlet you'd review them with.
Cost notes
- Scanner mode is roughly fixed cost per PR — pattern matching is cheap.
- LLM mode scales with diff size more than repo size, because the diff is what the agent reads first. A 5,000-line PR review will cost more than a 50-line PR review.
- Bot PRs (Dependabot, Renovate) get the same treatment as human PRs. If you want to skip them, exclude the bot users in Settings → PR review → Skip authors.