Skip to content
Skip to content
Product6 min readRepath Khan

Critique 0.2: Review Less, Test the Right Risk

Quick confidence for routine changes. Stronger proof, security probes, and resilience testing when the change deserves it.

Four ways to challenge a change

Quick confidence, task-aware proof, security probes, and resilience tests

critique.sh

The old CLI was capable, but it made the important choices hard to discover. A developer or coding agent could finish a task, see a small command list, and still have to guess whether a routine change needed a quick check, a repair package, a security review, or an attempt to break the failure path. That is too much ambiguity at the exact moment a change needs an independent challenge.

The new command palette

All four modes use the same evidence-first verification contract. Pick the lens that fits the change instead of asking a generic review bot to guess.

critique quick
Quick review
  • critique quick is the default for low-risk changes and routine pre-push confidence.
  • It runs a bounded review-only pass with a 180-second default budget.
  • It always uses --repair none, so it cannot modify the working tree.
Proof for a named task
  • critique finish --intent "…" verifies the change against the claim you made.
  • Choose --repair none, pack, or apply according to the authority you want to grant.
  • --json returns a stable result for a coding agent, CI wrapper, or local script.
Security and resilience
  • critique cyber concentrates the pass on threats and authorization boundaries.
  • critique break combines stress and reliability lenses for retries, recovery, interruption, and rollback.
  • Both begin with an explicit intent so the reviewer attacks the behavior that actually changed.

A fast check is useful only when it is safe to run habitually. critique quick maps to finish --repair none and rejects an attempt to add a repair mode. It is for the moment after a small refactor, a test adjustment, or a contained UI change, when you want an independent set of eyes without asking the tool to make a change on your behalf.

“Review the change” is a poor prompt for a new authorization flow or a multi-step upload retry. Those changes have different failure modes. The security mode puts attention on the trust boundary. The resilience mode asks what happens when the normal flow is interrupted, retried, or forced to recover. The aliases make those choices discoverable in the terminal, while the underlying focus flags remain available for more specific combinations.

Keep the full control surface when you need it

Aliases are the starting point, not a restriction. Compose focused checks on the primary command for an unusual risk profile.

critique finish \
  --focus security \
  --focus reliability \
  --repair none \
  --intent "Review the webhook delivery change"
A vague review loop
Agent finishesGeneric review runsFindings lack contextDeveloper guesses the next action
The Critique loop
Name the riskRun in a fresh verification workspaceReturn evidence and artifactsPrepare or apply only an independently verified repair

The modes change the question, not the trust model. Critique reconstructs candidate work in a disposable local workspace for the deep local loop. Development and test environment files can hydrate that local validation context but are excluded from patches, proof artifacts, and cloud uploads. A repair is not accepted just because a reviewer suggested it: the verifier must independently clear it before Critique can apply it.

We also rebuilt critique --help around decisions, rather than an undifferentiated list of flags. It starts with the routine paths, then separates deeper passes from evidence, recovery, and team setup. On an interactive terminal it has a more readable visual hierarchy; in logs or when NO_COLOR is set, the same guidance stays clean and portable.

What to run after an agent says it is done
  1. 1
    Is this a low-risk change and you only need a second look?
    Run critique quick. It is intentionally read-only.
  2. 2
    Do you need a proof bundle or a bounded repair suggestion?
    Run critique finish --intent "…" --repair pack --env auto --json.
  3. 3
    Did the change introduce a trust boundary or a dangerous failure path?
    Run critique cyber or critique break, then inspect the returned proof before accepting a repair.

A command only matters if the person or agent at the keyboard can find it. The new modes now appear in the CLI help, the Dashboard landing flow, the documentation, the README, the AI-readable llms.txt guidance, and the managed instructions installed by critique integrate --agent all. Codex, Claude Code, OpenCode, Cursor, and Copilot integrations can make the same risk-aware handoff without relying on a developer to remember a hidden option.

Challenge the change that is in front of you.

Install Critique CLI 0.2, run a read-only quick check for the routine work, and bring a sharper lens when the change carries real risk.

Read the getting-started guide