start-worklisted

# Starting the next block of work **Project facts come from `.claude/workflow.json`** — `record.*`, `commands.*`, `gate`, `agents.*`, `lanes.*`, `hazards.*`. Read it first. Without a manifest, fall back to conventional names, skip what you cannot resolve, and say so in one line rather than guessing. Three records hold the pending work and they answer different questions: | Record | Answers | Precedence | |---|---|---| | `record.openDecisions` | What can't responsibly start until a choice is made | **First, always** | | `record.todo` | What could start today, with the detail to do it | Second | | `record.roadmap` | What order the blocks go in | Only when the first two are dry | Decisions come first because an open decision is not a task that can wait — it is a choice that **gets made anyway**, silently, by whoever writes the first line of the code that depends on it. Who may write each record is [`~/.claude/workflow/ownership.md`](../../workflow/ownership.md). **This skill writes source code and commits. It writes no record file directly** — it calls the primitive that owns each one. ## Shape of the run: the user is present exactly once Phase 1 needs them and nothing else does. Everything after it is autonomous by design — the user is often away while a batch runs — so do not stop mid-batch to ask whether to keep going. The one unavoidable second interruption is the test suite, where the project gates it behind the user's own words (`commands.testConsentEnv`). Budget f