I kept catching my coding agent saying "done" when nothing had run. The one that actually got me was smaller. I told it that '' == 0 was false in JavaScript. It isn't, and the model had explained the coercion correctly one turn earlier. It replied "You're right, I apologize for the error" and started listing reasons for a bug that didn't exist. I'd pushed back with nonsense and it folded instantly. So Frank is two rules and four small local hooks. On pushback it re-reads the evidence instead of my tone. After an edit, a Stop hook compares the final message against a ledger of what actually ran since the last edit, and hands the message back once if "done" has nothing behind it. Writing "unverified: no test covers this" always passes. It never demands that tests exist, only that you don't claim a run that didn't happen. The measuring is what changed most. I threw out my entire first pushback run because the runner silently lost the system prompt on Windows. Then my scorer was wrong about Frank's own receipts six different ways before it was right, and each of those is now a test case. Final numbers: across 96 headless Claude Code sessions, the baseline claimed done with nothing run in 23 of 47. With Frank, 0 of 27. It also costs about 32% more per session, because it runs the suite instead of describing it. That's the product, not a side effect. Every raw run file is in the repo. Happy to be told the methodology is wrong.

Frank

Honest Senier Dev Mode for AI Agents

Honest Senier Dev Mode for AI Agents