Article: Nexus Lab has open-sourced ack-is-not-done-guard, a tiny utility that stops Claude Code file-based agents from claiming a task is finished until they reach a verified “proven done” state. Developers tired of phantom completions now have a concrete guard to drop into their pipelines.
Why the tool matters
Claude Code agents read and write files, a workflow that easily confuses the start of a process with its end.
How it works
The utility defines four explicit states:
- Acknowledged – the task is recognized.
- Working – the agent is actively processing.
- Artifact delivered – required outputs have been written.
- Proven done – all completion conditions are satisfied.
Only when an agent reaches “proven done” does the guard let a completion claim pass. The check is lightweight; it doesn’t monitor unrelated processes or act as a full-scale orchestration framework. It plugs directly into the file-based workflow Claude Code expects.
Who wins, who is left out
Because the guard focuses narrowly on file-based states, it won’t help projects that use other agent architectures or need broader workflow supervision. Those teams must still rely on external monitoring or custom logic.
Takeaway
ack-is-not-done-guard proves that a modest, well-defined state model can fill a persistent reliability hole in AI-driven file workflows, giving developers a clear, low-overhead way to trust that an agent’s “done” really means done.