resume

Crew Resume

Resume an interrupted crew run from where it left off.

RUN_ID="$ARGUMENTS"

If $ARGUMENTS is empty, list available runs from .crew/runs/ (filter to ones with state.json.status != "completed") and ask the user to pick one.

Steps

1. Verify the run exists:

test -d ".crew/runs/$RUN_ID" || { echo "No such run: $RUN_ID"; exit 1; }

1. Read .crew/runs/$RUN_ID/state.json. Inspect status and phase.

2. Tell the user what state the run is in:

Run <RUN_ID>:
  status: <status>
  phase: <phase>
  intent: <intent>
  last activity: <last journal entry timestamp>

1. Branch on phase:

   • router — restart from PHASE 1 of /crew:do. Treat as new run with same intent.
   • scout — re-run scout (PHASE 2). Brief may exist partial; scout will overwrite or extend.
   • decide_checkpoint / plan_approval — re-render plan.md if it exists and re-prompt the user (PHASE 4 from /crew:do).
   • worker_execute — re-spawn worker in last known mode with existing brief/plan/reflections/journal as inputs.
   • dual_verification — re-spawn critic + fresh-eyes (PHASE 7 from /crew:do).
   • debrief — read all verdicts and render debrief (PHASE 8 from /crew:do).
   • completed — tell user the run is already done, point them at .crew/runs/<RUN_ID>/.
   • cancelled_at_plan — tell user the run was cancelled, offer to restart with /crew:do.

2. Update state.json with resumed_at: "<ISO timestamp>" before proceeding.

3. Read the relevant phase logic from commands/do.md and follow it from that phase forward. You do not have to re-execute earlier phases — trust the on-disk state.

Rules

• Don't lose work. If existing files (brief.md, plan.md, journal.md) are present, treat them as authoritative. Append, don't overwrite, unless re-running the producing phase.
• Confirm with user before re-running expensive phases. Scout takes minutes; worker can take much longer. If you're about to re-run something the user might prefer to skip, ask via AskUserQuestion.
• State.json is the source of truth. Trust it over your own inference.