Todo-list

Execute the operator's standing directive without making them re-type it:

Make a numbered todo list for yourself with the list you gave me above, remember the list, and execute on it in the recommended order above. Also set a timer for yourself for every 10 x 10 minutes in case one of the running tasks doesn't wake you to review and continue on next steps.

1. Extract and remember the numbered list

1. Read the preceding assistant turn (and any explicit user reordering or filtering in the current turn) and extract the actionable items in the recommended execution order. Exclude items the user explicitly cancelled, deferred, or marked out of scope. If no clear prior list exists in the conversation, ask the user which items to include before arming a timer or mutating state.
2. Render the numbered todo list (1..N) at the top of your response, with each item's concise title, target/acceptance check, and status ([ ] pending, [-] in flight, [x] verified complete, [!] blocked).
3. Persist the list in two places so context compaction and 10-minute wakes never lose track:
   • In-conversation state: update the session's native todo/task tool when available and echo checklist transitions in chat on material progress.
   • Workspace resume note: write one unique, non-overwriting temp/todo-list-YYYY-MM-DD-<session-token>.md file in the current repo (temp/ is gitignored). Include:
     • Start timestamp, hard deadline (start + 100 minutes), and 10 x 10m wake counter (wake 0 / 10).
     • Scheduler/job identity and exact cancellation method (from handsfree below).
     • The full numbered [ ] / [-] / [x] / [!] checklist in recommended order, the currently active item number, awaited task/run/command identity, and latest observed evidence.
     • Never store credentials in the note.

2. Arm the 10 x 10 minutes watchdog via handsfree

Delegate all scheduler capability checks, job creation, active-job readback, and teardown rules to handsfree (skills/1-hourly/handsfree/SKILL.md in this repo) with three subroutine overrides:

1. Window & iteration cap override (10 x 10 minutes): Set the hard deadline to 100 minutes from start and cap recurring 10-minute wakes at 10 iterations (whichever is reached first), instead of handsfree's standalone 3-hour / 18-wake default.

2. Shared state-note override: Point the scheduled wake prompt at temp/todo-list-YYYY-MM-DD-<session-token>.md (one shared note and one job identity; never create a second parallel note or duplicate timer).

3. Checklist wake-action override: Use this scheduled wake prompt (substituting the exact note path and job ID):

   Resume the authorized numbered todo list in temp/todo-list-...md in this conversation. Check the hard deadline (100 minutes / 10 wakes max) first. Inspect the active numbered item ([-]) and its live task/command/CI output. If the active item finished and its check passed, mark it [x] in the note and chat, then immediately start the next [ ] item in recommended order. If all items are [x], or a hard blocker / user decision / user stop / 10 x 10m cap is reached, cancel this scheduled job, verify it is inactive, and report the final checklist status. Never mark an item [x] on pending, empty, or unverified output.

If the current harness lacks a verified native same-conversation scheduler (per handsfree's create/readback/cancel check), leave the wake unarmed, record scheduler: unarmed (<reason>) in temp/todo-list-...md, tell the user plainly, and continue executing the numbered list now.

3. Execute immediately in recommended order

1. Do not wait for the first 10-minute wake to start work. Immediately mark item 1 as [-] in temp/todo-list-YYYY-MM-DD-<session-token>.md and begin executing item 1 in the current turn.
2. Whenever an item completes synchronously in the current turn (or wakes the agent via normal reactive background-task completion before the 10-minute timer fires), verify its outcome, mark it [x] in the note and native task list, and advance immediately to item k + 1.
3. When an item launches an asynchronous command, review, or CI run that leaves nothing else actionable in the current turn, record the exact task/run ID in temp/todo-list-...md and end the turn so either the task's reactive completion wakeup or the next 10-minute watchdog wake resumes execution.
4. As soon as the final item N is verified [x] (or execution halts on a hard blocker, required user decision, or the earlier of 100 minutes / 10 wakes), cancel the scheduled job in the same conversation, verify readback confirms it is inactive/absent per handsfree, update temp/todo-list-...md, and report the completed numbered checklist.