One human approval
critique login opens Critique in the browser. The human signs in once; the agent receives a scoped machine key, never a web session.
Agent contract · finish.v1
A small interface agents can trust.
Critique has one job: independently finish a candidate change. Every terminal response has a deterministic outcome, evidence artifacts, changed paths, limits, and an exit code.
npm install --global @critiquedotsh/cli
critique login
critique model auto
critique integrate --agent all
critique finish --intent "Finish PAY-42" --repair apply --env auto --json
critique artifact fin_… --kind proofCodex, Claude Code, OpenCode
critique integrate installs a finish skill for each supported coding agent. They call the CLI after their own implementation and checks.
Your env stays local
Critique reconstructs the change in disposable workspaces. Development and test env files can run validation locally, but are excluded from patches, proof, and cloud uploads.
Changes come back
With --repair apply, only a verifier-approved patch reaches the caller's working tree. JSON returns applied, changed_paths, and the proof artifact.
Early access
Early users receive free GPT-5.6-Terra-powered CLI capacity. Hundreds of millions of tokens are allocated for the launch cohort, subject to launch availability.
The agent loop
critique finish reconstructs and tests the current change. critique wait reads an asynchronous cloud run. critique artifact retrieves proof, reproduction, repair package, or patch.