Critique CLI 0.3.0

Critique is an independent code reviewer and persistent engineering second brain for coding agents. Claude Code, Codex, Cursor, Copilot, or another agent remains the author; Critique collaborates during implementation and starts a fresh judgment context for review.

npm install --global @critiquedotsh/cli
critique doctor
critique integrate --agent all

The reviewer runtime is pinned and installed with Critique. No separate review harness, engine selection, or setup install is required.

The current release is 0.3.0:

npm install --global @critiquedotsh/[email protected]
critique --version

Work with an external coding agent

Create durable task context at the start:

critique task start "Implement bounded payment retries" \
  --agent claude-code \
  --accept "Retries cannot charge twice" \
  --json

Consult Critique while the implementation is still fluid:

critique ask "Trace the retry path and challenge the current design" --session payment-retries --json
critique chat --session payment-retries

The second brain retains its runtime session, reads the complete repository, runs bounded diagnostics, and can delegate focused research. It does not edit the author's working tree.

After meaningful work, record an explicit checkpoint—not the author agent's transcript or hidden reasoning:

critique checkpoint "Added retry state and idempotency key handling" \
  --agent claude-code \
  --worklog "Updated the charge worker" \
  --claim "The retry tests pass" \
  --idempotency-key claude-turn-12 \
  --json

Checkpoints inform collaboration. The fresh reviewer is given only the task contract and current repository state. Author transcript, plan, worklog, claims, conclusions, confidence, claimed checks, and second-brain commentary are withheld so they cannot anchor findings.

Request independent review

critique review "Implement bounded payment retries" --json

Review runs in an isolated reconstruction of the complete repository. The diff locates change; it does not limit inspection. The reviewer traces callers, tests, schemas, configuration, deployment, and operational paths before adjudicating findings.

The JSON response distinguishes:

• changed_paths: paths in the working-tree candidate;
• review_methodologies: bundled methods selected as reviewer instructions, each with status: "selected";
• findings and their evidence;
• the durable task_id when a task is active.

A selected methodology is never presented as a completed check. Use findings, command evidence, and proof artifacts to understand what actually ran.

critique findings <run-id> --json
critique proof <run-id> --json
critique recheck <run-id> --json

Bundled review intelligence

critique skills

Critique ships 38 substantive review methodologies across correctness, testing, security, performance, reliability, frontend systems, languages, data, operations, dependencies, AI systems, and documentation. Every method contains:

• an investigation procedure;
• required evidence;
• disconfirmation and false-positive standards.

Relevant methods are physically bundled into the published CLI and injected in full into the reviewer prompt. Repository-specific customization lives in .critique/profile.json:

{
  "schema_version": "critique.review-profile.v1",
  "methodologies": {
    "include": ["concurrency"],
    "exclude": ["docs-contracts"]
  },
  "reviewer": {
    "repository_recon": "deep",
    "max_findings": 8,
    "require_reproduction": true,
    "verifier_model": "qwen/qwen3.7-flash"
  },
  "collaboration": {
    "style": "challenging",
    "interrupt_on": ["test_failure", "public_contract", "migration", "completion_intent"]
  }
}

This portable profile is suitable for hand editing or export from a configuration UI while execution remains in the CLI. It controls repository recon, methodology selection, finding limits, reproduction requirements, independent verification, and collaboration style.

Runtime and evidence

Critique installs exact matching versions of its reviewer distribution, typed SDK, and plugin. The runtime:

• resolves only the package-local binary, never an ambient executable;
• disables auto-update and uncontrolled project runtime configuration;
• uses isolated config, data, cache, and home directories;
• communicates through the typed SDK rather than raw routes or JSONL parsing;
• loads the bundled Critique plugin and refuses startup if its handshake is absent;
• shuts down cleanly without orphan processes.

The plugin provides critique_check, an explicit-argv evidence tool. It records exit status, duration, stdout/stderr digests, working directory, and before/after tree digests. Strict repair verification accepts only fresh-session, reset-workspace, different-model-family evidence captured through this Critique runner.

Commands

critique task start "..."     Start durable collaboration context
critique task status          Inspect the active task
critique checkpoint "..."    Record explicit author progress
critique ask "..."           Consult the persistent second brain
critique chat                 Open an interactive collaboration
critique review "..."        Start a fresh independent review
critique findings <run-id>    Read complete findings
critique proof <run-id>       Inspect evidence
critique recheck <run-id>     Re-evaluate after edits
critique skills               List bundled methodologies
critique doctor               Verify repository and bundled runtime
critique integrate --agent all Install the compact agent workflow
critique init                 Create repository policy and profile

Use --json for structured agent output and --events for NDJSON progress on stderr.

Documentation: critique.sh/docs

Package page: npmjs.com/package/@critiquedotsh/cli