Make coding-agent completion inspectable with four proof fields
At a glance
A developer makes an agent show the outcome, evidence, scope, and unchecked areas before calling a coding task done. This workflow is listed under coding agents, verification, software testing, review workflow.
- Agents / tools
- AI coding agent, Git, test runner, CI, process registry
SourceBuilding a tool for these workflows? See weekly rates.
What they did
1) Before a coding run, define the observable outcome, scope, exclusions, and the test or observation that will count as evidence. 2) Keep the agent's authority within reversible work such as a branch, tests, or a prepared pull request, and require a result report with DONE, PROOF, SCOPE, and NOT VERIFIED fields. 3) Independently check that the cited evidence is reachable, current, and covers the stated scope; before acceptance, inspect for leftover processes, temporary infrastructure, uncommitted files, open worktrees, or queued jobs.
How it went
The author says this protocol grew out of a run where the task inventory looked empty but a polling process was still running minutes later.
Worth knowing
The author describes Proof of Done as a bounded, inspectable claim rather than mathematical proof; the amount of checking should match the impact and reversibility of the change.
