Working overnight

Governed unattended work while the user is away. The engine is /loop: the user launches /loop working-overnight and each cycle does one unit of work and returns with fresh context. This skill holds the policy: what to pull, how to stay safe, when to defer, and when to wind down. The overnight-continue hook keeps a single cycle from halting on a question; it is not the engine.

Read references/safety.md (the tier model and the never-list) and references/pipeline.md (the per-task governance pipeline) before the first cycle. They are binding.

Two modes, told apart by the manifest

• Pre-flight (interactive, once, when the user sets up the night): arm the run.
• Cycle (each /loop iteration, unattended): do one unit of work.

If .claude/overnight/active.md is absent, this is pre-flight. If it is present, this is a cycle.

Pre-flight (arming)

Do this with the user present. Never start autonomous work until they confirm.

1. Agree the work source and scope: which tickets, which investigations, and any no-go areas (paths, systems, ticket labels to avoid).
2. Agree the budget (token or cost ceiling) and the investigation-depth cap (how many new tickets an investigation may create before implementation must catch up).
3. Confirm the safety config in references/safety.md: the allowed tools and the Tier-3 never-list. Do not proceed if the requested work would need anything on the never-list.
4. Ensure .claude/overnight/ is gitignored.
5. Write the manifest to .claude/overnight/active.md (scope, budget, caps, no-go list, start time) and seed .claude/overnight/queue.md with the initial items. Creating the manifest arms the hook.
6. Tell the user to run /loop working-overnight and leave. Stop here.

Cycle (one unit of work)

1. Re-read .claude/overnight/active.md, .claude/overnight/queue.md, and the project rules (CLAUDE.md, whose coding standards are inline). Fresh every cycle: this is why rule-drift does not accumulate.
2. Pick the next item by priority: a cleanly-gated ready ticket, else an ungated ticket to gate, else a bounded investigation. If no safe item remains, go to Wind-down.
3. Classify the item by tier (references/safety.md). A Tier-3 action is never performed: park it as needs-human and pick another item.
4. Resolve any question. Try web search, the code, and the project rules first. Only a genuine you-only decision blocks: if it is low-stakes and reversible, record the assumption for the report and continue; otherwise append it to .claude/overnight/decisions.md (item, options, your recommendation) and pick another item.
5. Do the work through references/pipeline.md.
6. Implementation lands in a per-item git worktree as a branch and a PR, never merged. Investigations write findings and create/gate tickets.
7. Update .claude/overnight/queue.md (done, parked, or newly created items).
8. Backstop: if context or the budget is near its limit, go to Wind-down now rather than starting another item.

Delegate the heavy work (implementation, review) to subagents so their tool output does not fill this cycle's context.

Wind-down

Reached at queue-empty (the goal) or the backstop.

1. Write .claude/overnight/report.md: PRs opened (links), decisions deferred (with options), investigations done, assumptions to verify, suggested next steps. Use the handoff shape from the closing-sessions skill.
2. Remove .claude/overnight/active.md to disarm the hook.
3. Stop the loop: call ScheduleWakeup with stop: true so no further cycle fires.
4. Print a one-line summary pointing at the report.