dunnit 0.1.0

Did the agent actually do it? Tamper-evident verification of AI agent work: definition-of-done as code.

dunnit

Did the agent actually do it?

AI coding agents claim completion when work isn't done: they delete failing tests, add skip markers, hardcode expected outputs, stub functions and declare victory. Research calls it reward hacking. You call it Tuesday.

dunnit is a tamper-evident verifier for agent work. You declare your definition of done in dod.yaml; dunnit verify re-runs the proof itself and inspects the diff for test-gaming. It never trusts the transcript.

pip install dunnit
dunnit init      # write dod.yaml
dunnit verify    # ✓ or ✗, exit code for CI

What it checks (v0.1)

• Commands — declared proof commands (tests, lint, build) must exit 0, executed by dunnit, not quoted from the agent's chat log.
• Tamper — git diff of test files: deleted tests, added skip markers, net-removed assertions → hard FAIL.
• Stubs — changed code scanned for TODO/FIXME, NotImplementedError, swallowed exceptions → WARN.

dod.yaml

version: 1
base: origin/main        # diff against this ref (default: HEAD, i.e. uncommitted work)
checks:
  - name: tests
    run: pytest -q
  - name: lint
    run: ruff check .
tamper: true
stubs: true

Use it everywhere agents work

• CI — dunnit verify exits non-zero on FAIL; --json for machines.
• Claude Code stop-hook / Cursor rules — make the agent run dunnit verify before it's allowed to say "done".
• Python — from dunnit import verify; verify("dod.yaml").passed

Roadmap

pytest plugin · coverage non-regression · hardcoded-output detection · optional LLM judge · MCP server (agents self-verify) · signed verdict attestations · JS/Go check packs.

MIT licensed.