Fredrin
A Worker just shipped a ticket where one of ten action items was never done. It behaved correctly at every step, and the miss still reached Review looking green. What happened: the plan had `- [ ] Astro shell on the Worker` alongside nine other items. The Worker completed the other nine, checked them off, and left item 7 unchecked — correctly, since checking it would have been a false claim. It then ran `ticket finish` with real acceptance checks (typecheck, lint, two test suites, all exit 0), so `finish` pushed the branch, opened the PR, and moved the card to Review. The gap is that `finish` gates on `checks[]` only — the exit codes of commands the Worker ran. It never looks at the plan checklist. So a ticket arrives in Review with an incomplete plan and a fully green PR, and no surface says "9 of 10 items done." The unchecked box is inert: it is the most honest signal the Worker produced, and it is the one signal nothing consumes. The deviation was disclosed, but only in prose — commit body, PR description, a README sentence. That is exactly the layer a reviewer skims, and it is a weaker record than the structured checklist sitting right there. What would help, roughly in order of preference: 1. `ticket finish` warns loudly on unchecked action items, and includes them in the PR body it generates — so the reviewer sees "1 unchecked plan item: Astro shell on the Worker" without reading the commit message. 2. Optionally let a project require an explicit acknowledgement (a reason string) before finishing with unchecked items, rather than blocking outright. Deviating is often the right call; doing it silently is not. 3. Show the ratio on the card in Review — "9/10" — so the board itself carries the signal. Not asking for a hard block. Plans change mid-build and that is healthy. The ask is that an incomplete plan item cost something more than a line of prose. Minor related nit: `ticket check <n>` is position-based, so any mid-build edit to the plan shifts the numbers under you. Checking by stable item id (or by matching text) would make the checklist a more trustworthy record.