Skip to content
Critique/docs
Platform

Checkpoint

Pre-review trust gates that block or flag low-signal pull requests before automated AI review runs.

Checkpoint runs before the main review pipeline on pull requests you enable it for. It applies fast, deterministic checks — contributor signals, PR shape, boilerplate patterns — so you do not spend credits reviewing obvious spam, empty drive-by PRs, or AI-slop templates.

Marketing overview: critique.sh/checkpoint. Configuration lives in Dashboard → Checkpoint and per-repo at /dashboard/[owner]/[repo]/checkpoint.

When it runs

Checkpoint evaluates a PR at gate time (typically when the PR becomes eligible for Critique automation). Outcomes are recorded as gate events you can inspect in the dashboard, including which rules passed or failed.

Checkpoint does not replace the security or architecture specialists — it filters who and what deserves a full review.

Example rule categories

Rules are toggled per repository. Common categories include:

CategoryWhat it checks (conceptually)
Identity & accountProfile signals, account age, public activity
PR shapeFile count, description quality, daily PR velocity
Content qualityBoilerplate phrases, weak commit messages, incoherent large diffs without context
HygieneRepeat offenders, optional language or test requirements

Exact defaults evolve; the dashboard rule catalog is authoritative for your installation.

Outcomes

Each event resolves to an outcome such as allow, warn, or block (labels may vary in the UI). A block stops or defers expensive downstream review until a human overrides or the PR is fixed.

Operators can override a specific gate event from the event detail page when a false positive should proceed.

Configure Checkpoint

  1. Open Dashboard → Checkpoint for a cross-repo view, or a repo’s checkpoint page for rules.
  2. Enable the rules that match your risk tolerance (stricter for public OSS, looser for trusted internal repos).
  3. Watch gate events for the first week and tune — start with account age, file count, and AI-slop patterns if noise is high.

Relationship to review policy

LayerQuestion it answers
CheckpointShould we spend review effort on this PR at all?
Review policyHow hard should we judge findings (strictness, specialists, models)?

See Policy fields for review-side configuration.