Critique Documentation

Critique is an AI Change Control Platform: a GitHub-native control plane that sits at the merge boundary. It records a Change Passport for every pull request, scores risk, requires evidence for blocking decisions, enforces merge policy, verifies repairs, and feeds incident learnings back into the rules — without replacing the coding tools your team already uses.

What you can do

How Critique fits in your stack

flowchart LR
  Dev[Developer opens PR] --> GH[GitHub]
  GH --> CP[Checkpoint optional]
  CP --> RV[Automated review]
  RV --> Out[Checks and comments]
  Out --> Dash[Dashboard]
  Dash --> Chat[Chat and Remedy]

Critique focuses on whether a change should ship — semantic risk, tests, architecture, and security — not just whether CI is green. You stay in GitHub for day-to-day work; the dashboard is where you inspect runs, tune policy, and follow up.

Product surfaces

• Change Passports — The PR-level system of record: provenance, risk, gate, evidence, merge policy, proof, and memory. See Change control.
• Critique Intake — Agentic bug intake for turning user reports into investigated, agent-ready engineering work. See Critique Intake.
• Automated PR review — Triggered when a PR is opened or updated (and on demand via @critique /review). Each run is an evidence run linked to a passport.
• Checkpoint (Agent Firewall) — Optional pre-review trust gates so expensive review does not run on obvious low-signal PRs.
• Remedy — Managed fix execution in an isolated sandbox when you want Critique to apply a bounded patch.
• Chat & workspace — Repo-grounded Q&A and optional cloud jobs for exploration and builder workflows. See Introducing Critique Workspace for the unified /workspace surface.
• BYOK — Optional LLM Gateway, OpenRouter, or CrofAI key so review and chat models bill your provider account ($8/month harness).
• BYOA — Cursor Cloud Agents on Composer 2.5 (Agent SDK), Claude Managed Agents, and OpenAI Codex from review runs, plus Fix prompt and Claude Code chat handoff — execution stays on your existing agent subscriptions. Configure keys under Settings → Agents.
• Coding Agent API — POST /api/v1/coding-agent/runs for OpenCode + E2B jobs with optional draft PR publish (managed credits or per-run OpenRouter key).
• Platform API (v6.0) — POST /api/v1/reviews queues PR reviews; structured findings[] on webhooks and GET /api/v1/review-runs/:id?findings=all; MCP with scoped tools; OpenAPI at GET /api/v1/openapi; installation and per-run webhooks. See Merge Gate API.
• Inference API (v5.2) — GET /api/v1/models and POST /api/v1/chat/completions for token-in/token-out chat on Critique credits — DeepSeek V4 Flash (default), Tencent Hy3 Preview, NVIDIA Nemotron 3 Ultra on Western-hosted capacity. Usage dashboard at /inference-dashboard.
• Agent Skill Marketplace — Browse, publish, and install critique-review lenses with a performance leaderboard driven by finding feedback.
• Insights — Velocity, risk, cost, retrospectives, and compliance exports at /dashboard/insights.
• Connections & Platform API — Linear, Slack, MCP, and REST v1 on scoped crt_ keys — Merge Gate, Inference and Coding Agent from v5, review queueing and webhooks from v6.

Deeper reference

• Policy fields — Strictness, specialists, path escalations, model overrides.
• Specialists — What each review lane looks for.
• Chat tools — What Critique Chat can do on your behalf.
• Environment variables — change-control flags, secrets key, and webhook/API scopes for operators.
• Remedy — How bounded fix runs are scoped.

Architecture (for operators)

If you run Critique for a team or need to reason about reliability:

• Event-driven pipeline — Webhooks, async workers, retries.
• Model engine — Plans, roles, and credit floors.
• Memory & context — Repository awareness and recall.