DoneAudit — your AI says “done”; verify the evidence before you trust it

[kafidog]

I kept running into the same problem with coding agents:

“Done. All tests pass.”

But sometimes a test actually failed, a required check was missing, or the evidence no longer matched the source.

So I released DoneAudit, a small MIT-licensed tool that verifies the actual evidence before accepting an AI completion claim.

Example

AI says:

Done ✅

DoneAudit checks the evidence:

DoneAudit
Trust score: 70 / 100 🔴

❌ test: FAILED
✅ build: PASSED
✅ required checks: PASSED
✅ Git state: PASSED
✅ completion evidence: PASSED

Conclusion: Cannot confirm completion

Install

Run this inside an existing Git repository:

npx doneaudit@0.1.0 init

Then use Codex normally.

DoneAudit checks real test, build and required-check results, Git/source state, and completion evidence.

It runs locally, needs no account or server, and GitHub Actions can independently rerun the checks.

The score is deterministic evidence coverage — not another AI judging the AI.

Current validation

DoneAudit v0.1.0 has been tested with a real Codex workflow on Windows:

• Success case: 100 / VERIFIED
• Intentional failing case: 70 / FAILED
• Fresh npm installation on Windows: PASS
• GitHub Actions validation: PASS

Links

GitHub
https://github.com/kafidog/DoneAudit

npm
https://www.npmjs.com/package/doneaudit

v0.1.0 release
https://github.com/kafidog/DoneAudit/releases/tag/v0.1.0

I’m looking for a few Codex users to try DoneAudit on real repositories.

If it catches a false “done”, or if DoneAudit itself gives an incorrect result, please open an issue with a reproducible example and remove any sensitive information first.

Real-world reports are especially useful for improving the project.

If you find it useful, starring the repo also helps other coding-agent users discover it.