Harness Score is a deterministic scanner that measures the harness around an AI coding agent in a repository: the context files, rules, skills, hooks, sensors, and guardrails that steer the agent, check its work, and stop damaging actions. It reports a maturity level from L0 to L4, a 108-point breakdown across six dimensions, and a ranked list of what to fix next. Each failed check names the file, explains what is missing, and links to a remediation recipe.
Every one of its 36 checks is a filesystem fact (a file exists, parses, or matches a pattern), with no LLM calls and no network access, so the same commit always gets the same score. That makes it usable as a CI gate: a minimum-level flag fails the build below a chosen level, and a diff mode compares runs. It scores harnesses for Cursor, Claude Code, Devin, Windsurf, Cline, Continue, Codex, and other tools the same way.
Levels gate on the shape of the harness, not only on points, so a repository with extensive docs and no tests stays at L1:
| Level | What it means |
|---|---|
| L0 Unharnessed | Agents rediscover the project every session |
| L1 Documented | A substantive AGENTS.md orients every session |
| L2 Guided | Scoped rules, at least one skill or command, basic hygiene |
| L3 Sensing | Tests, a linter, type checking, and CI verify every push |
| L4 Self-correcting | Gate hooks block destructive actions and feedback hooks lint and format inline |
The six dimensions are Context and Guides (20 points), Skills and Commands (17), Hooks and Guardrails (14), Sensors and Feedback (20), CI Feedback (14), and Hygiene and Safety (23). The scanner does not judge whether tests are good, whether rules are current, functional correctness, or team practice such as branch protection. Alongside the CLI, the project ships a guide to harness engineering, a GitHub Action, and editor plugins with an audit command that lets the agent apply the fixes.
Features
- Maturity level: L0 to L4, with the exact requirement blocking the next level
- Six-dimension breakdown: 108 points across context, skills, hooks, sensors, CI, and hygiene
- Deterministic checks: 36 filesystem checks with zero LLM calls, zero network access, and zero runtime dependencies
- Remediation recipes: each failed check names the file and links to a fix in the guide
- CI gate: fails a pipeline below a minimum level, with a diff mode that flags maturity model changes
- Output formats: terminal, JSON, Markdown, and branded SVG badges and share cards
- Multi-harness detection: scores Cursor, Claude Code, Windsurf, Cline, Continue, Codex, and other tools with one model
- Team customization: an optional config file turns off checks that do not apply, and exclusions show in every output
- Leak checks: three checks for leaked credentials that can never be excluded
- Editor plugins: an audit command and harness-engineering skill that let the agent fix gaps, with Cursor as the flagship
- GitHub Action: scans on every push, gates on a level, and regenerates the badge
- Harness engineering guide: chapters on guides, sensors, guardrails, and the maturity model
