DoneAudit
Trust score: 70 / 100 🔴

❌ test: FAILED — the actual test failed
✅ build: PASSED
✅ required: PASSED
✅ Git state: PASSED
✅ Completion evidence: PASSED

Conclusion: Cannot confirm completion

Saying done doesn't make it done. Verify the evidence before you ship.

Install inside your Git project (Node.js 18+, Git; PowerShell on Windows, Bash elsewhere):

npx --yes --package=https://github.com/kafidog/DoneAudit/archive/refs/tags/v0.1.0.tar.gz doneaudit init

Then use Codex as usual:

codex "Fix the failing test. Follow the DoneAudit completion rule."

DoneAudit installs the rule that tells Codex to write its completion claim and run node .doneaudit/tool/bin/doneaudit.js run. The tool executes your checks, captures real receipts and displays the result. Commit the generated files and claim; the installed GitHub Actions workflow reruns the same checks and posts the same score format in the job summary. There is no account, server, or AI judge.

One-time setup: review doneaudit.config.json. npm test, build and lint (or check) scripts are detected automatically. Replace any null with a real shell command for your project. Missing checks never silently count as passes. Run from a Git repository root with at least one commit. Install project dependencies before checking. Review the generated workflow if your project uses a toolchain other than Node.js.

Real Windows terminal result · GitHub checks · Releases

What gets verified?

Evidence	Points
Tests	30
Build	20
Required checks	25
Git/source binding	10
Completion evidence	15

Within each command category, points are the category weight multiplied by the fraction of passing checks, rounded down. A failed required check always means FAILED, even with a high score. Missing, inconsistent, or older-than-one-hour evidence means INSUFFICIENT EVIDENCE. Only all required items passing produces VERIFIED and exit code 0. Failure exits 1; insufficient evidence exits 2. A build for an interpreted project can be its real syntax/type/package check.

The score is evidence coverage, not a probability that the software is correct. DoneAudit checks command exit codes, receipt/log integrity and source state. It cannot tell whether your chosen tests are good or cover every requirement.

{
  "version": 1,
  "checks": [
    { "label": "test", "group": "test", "command": "npm test" },
    { "label": "build", "group": "build", "command": "npm run build" },
    { "label": "lint", "group": "required", "command": "npm run lint" }
  ]
}

Completion claim (doneaudit.claim.json, written by Codex):

{ "completed": true, "summary": "Fixed addition and verified the configured checks." }

The claim is an assertion to check, not proof by itself. Commands generate the proof. Reports and upstream receipts live in .doneaudit/evidence/ and are uploaded as workflow artifacts. To inspect existing evidence without rerunning:

node .doneaudit/tool/bin/doneaudit.js report

Supported scope and trust boundary

Windows PowerShell and Unix Bash execution, with zero npm runtime dependencies.
Codex through project AGENTS.md. This is an instruction-based integration, not a hard Codex stop hook. GitHub independently reruns the checks.
Dirty working trees are allowed; changes to tracked or nonignored untracked file contents after execution invalidate the receipts. Ignored files (such as dependencies and build output) are outside the source fingerprint.
Only use configured commands you trust. On GitHub, use pull_request, never pull_request_target for executing contribution code. The generated workflow has read-only repository permissions.
Local receipts are not signed or an adversarial sandbox. Someone able to edit both tooling and evidence can forge local results. Review changes to the configuration, tests and vendored tool. CI reruns instead of trusting receipts.
Upstream Claude Code hooks and agent-done-or-not CLI are retained; see the original integration instructions. They retain upstream behavior; DoneAudit scoring is provided by the doneaudit command.

Origin and license

DoneAudit is built on mohamedzhioua/agent-done-or-not. Its receipt engines do the execution capture. We add deterministic scoring, stronger working-tree binding, Codex setup and concise terminal/GitHub results. The original code is not presented as our own.

Contributing

Found an incorrect result? Open an issue with your OS, Node version, config and redacted terminal output. Never attach secrets from check logs. Run npm test for DoneAudit regression tests; original upstream tests remain in tests/. Keep v0.1 focused on completion verification. If this helped you catch a false “done”, a star and a reproducible report help other people find and improve the project.