Missing and present guardrails
harn reads public metadata, root files, docs, workflow paths, and common Claude Code harness files.
Add ?repo=owner/name to this URL, or paste a repo above.
Paste any public GitHub or GitLab repository to generate a shareable harn.app score page for Control, Agency, and Runtime readiness.
Add ?repo=owner/name to this URL, or paste a repo above.
Turn this score into an owner-visible backlog. These are the concrete guardrails that make a repo safer for AI coding agents.
Run /harn:init to add the control file, security guard, quality gate, and settings wiring.
Write the north star, stack, boundaries, and stop rules in under 60 lines.
Deny rm -rf, pipe-to-shell, reckless chmod, disk writes, and obvious footguns.
Require branches, PR review, and status checks before code can land.
The agent cannot finish until the real project health command passes.
Stop hooks check stop_hook_active before trying to recover again.
Use deterministic tests that catch regressions outside the chat transcript.
Rerun the same gates after the agent leaves the local session.
Block credential-file access locally and enable provider push protection.
Route auth, billing, infra, migrations, and release code to humans.
Write agent tool payloads to an inspectable trace log.
Require handoff state before compaction or risky continuation.
Declare writable directories for each task; unknown areas stay read-only.
Spell out architecture boundaries and invariants agents must not cross.
Declare MCP servers and tool permissions instead of relying on local drift.
Keep research noise out of the actor context; return only condensed answers.
Require approval for lockfile rewrites, major upgrades, and new services.
Use /loop for active polling and durable schedulers for unattended checks.
Every risky PR gets revert, migration, and feature-flag instructions.
Protection is a habit: run the score again and turn warnings into issues.
## Make this repo agent-proof
- [ ] Install harn and run /harn:init
- [ ] Add or tighten AGENTS.md / CLAUDE.md
- [ ] Block destructive shell and pipe-to-shell commands
- [ ] Protect main/master with branch rules and required checks
- [ ] Add Stop-hook quality gate for the real project health command
- [ ] Ensure Stop hooks check stop_hook_active
- [ ] Add or strengthen tests
- [ ] Require CI to run the same checks
- [ ] Enable secret scanning / push protection
- [ ] Add CODEOWNERS for sensitive paths
- [ ] Persist agent trace logs
- [ ] Add checkpoint/handoff policy for long sessions
- [ ] Define writable directories per task
- [ ] Document architecture boundaries
- [ ] Declare MCP/tool permissions explicitly
- [ ] Use sub-agents for research-heavy work
- [ ] Require approval for dependency and lockfile jumps
- [ ] Monitor deploys with /loop or durable scheduled checks
- [ ] Add rollback notes to risky PRs
- [ ] Re-score with harn after changesEach repo page has stable, crawlable copy about the repository's agent-readiness profile, common failure modes, and a concrete hardening backlog.
Does the repo tell an AI coding agent what good work means? We look for AGENTS.md, CLAUDE.md, static sensors, tests, and architecture constraints.
Does the repo bound what the model can do? We look for Claude settings, MCP/tool declarations, delegation patterns, and explicit permission boundaries.
Does the repo mechanically catch broken behavior? We look for lifecycle hooks, CI, trace/checkpoint state, and Stop-hook loop prevention.