DoneCheck: proof-of-done for AI coding agents

Your AI coding agent says it is done. Make it prove it.

Works with Codex, Claude Code, Cursor, OpenCode, local CLIs, and GitHub Actions.

donecheck is a zero-dependency proof-of-done gate for AI-assisted code changes. It scans the changed files, runs the verification command you choose, and writes a small DONECHECK.md receipt before anyone claims the work is finished.

Star it if you want a tiny AI coding agent verification gate you can drop into any repo.

python donecheck.py --cmd "pytest -q"
cat DONECHECK.md

If there are no files and no command, it fails. No evidence, no "done".

Quick Start

Add DoneCheck to a repo using AI coding agents:

pipx install git+https://github.com/AtharvaMaik/donecheck
donecheck --init --cmd "pytest -q"
donecheck --agent-prompt --cmd "pytest -q"

Then tell Codex, Claude Code, Cursor, or any agent:

Before claiming done, run donecheck and fix anything it reports.

What It Catches

• unfinished markers and placeholder phrases in changed files
• narrow verification that does not mention changed code paths
• proof files that only say tests passed without output, exit code, and timestamp
• stale DoneCheck receipts when files, commands, or base inputs change
• skipped verification as SKIPPED, not a quiet pass
• swallowed Python exceptions and empty JavaScript catch blocks
• accidental literal secrets
• unsafe eval / exec
• failed verification commands and missing evidence when nothing was checked

20 Second Demo

git init demo && cd demo
curl -O https://raw.githubusercontent.com/AtharvaMaik/donecheck/main/donecheck.py
printf 'def charge_card():\n    # TODO wire Stripe later\n    return True\n' > app.py
git add app.py
python donecheck.py --all

Output:

DoneCheck: FAIL
- unfinished_marker app.py:2 # TODO wire Stripe later

The full receipt is in DONECHECK.md. Fix the file, then run:

python donecheck.py --all --cmd "python -m py_compile app.py"

Now the receipt says PASS and records the command output.

If verification cannot run, say so explicitly:

python donecheck.py --skip-reason "missing DATABASE_URL"

That writes a SKIPPED receipt and exits non-zero.

Use It Anywhere

To create the GitHub Action workflow in a repo:

donecheck --init --cmd "pytest -q"

GitHub Action

name: donecheck
on: [pull_request]

jobs:
  donecheck:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
        with:
          fetch-depth: 0
      - uses: AtharvaMaik/[email protected]
        with:
          command: pytest -q

On pull requests, the action scans the PR diff against the base branch, emits GitHub error annotations for findings, and adds the full receipt to the Actions step summary. Outside pull requests, pass args: --all to scan the whole repo.

Agent Prompt

Before claiming done, run:
python donecheck.py --cmd "<project test command>"

If it fails, fix the work and rerun it. Include the DONECHECK.md status in your final answer.

There is also a drop-in skill at skills/donecheck/SKILL.md.

Why It Exists

AI agents are good at sounding finished. DoneCheck makes them leave evidence:

• changed files
• findings
• commands run
• exit codes
• recent command output

It is not a full linter, security scanner, or test framework. It is the cheap first gate that catches obvious AI-code misses before a human review, CI system, or hosted review bot spends time on them.