crimes

Local, deterministic change-risk and agent-risk analysis for TypeScript, JavaScript and Python. Built for agents, readable by humans.

npx crimes scan .
npx crimes context src/file.ts --root . --format json
npx crimes scan --changed --format json

The human report prioritizes five files. JSON contains all visible findings with evidence, scores, stable fingerprints and suggested next steps. Review coverage warnings and context analysis status; a clean report does not prove safety. Run behavior tests separately.

Analysis reuses source and parsed data within each command; no persistent cache is required. Context and scoped scans still analyze the repository. crimes migrate-pins --format json previews legacy decisions before reviewed --apply; --recover restores originals after an interrupted replacement. Schema remains 0.8.0. Migration guide.

Install agent integration once per repository:

crimes init --agents

Normal terminal use refreshes unchanged generated skills after CLI upgrades. Agent/JSON calls show the safe update action on stderr and never auto-write skills. Customized instructions and executable hook settings are preserved; CI skips maintenance. The optional Claude hook uses your installed CLI and supplies advisory context without downloading tools. Skill setup and updates.

• Getting started
• Agent workflow
• Command and detector reference
• Configuration
• 0.29.0 release evidence
• Performance methods
• Source and contributions

MIT. Node.js >=18. No cloud or model required.