resume

Resume a paused, interrupted, or crashed RePPITS workflow by reconstructing state from Claude Code Tasks.

The full /run flow keeps its state in the conversation context, which is lost on a crash, a stop/pause, or a new session. The durable record is the Claude Code Tasks created during Phase 3 (Plan) and advanced during Implement/Test/Secure. This command rebuilds from that record and continues from the last incomplete step.

Arguments

• topic (optional): The original topic or Linear issue ID. If omitted, infer from the open Tasks and the local plans/ directory.

Steps

1. Read the durable state:
   • Run TaskList to get all tasks, their status, owners, and blockedBy edges.
   • For the parent/plan context, read the local plan (e.g. plans/<feature>/00-parent-design-doc.md) and any matching Linear parent issue if Linear MCP is available.
2. Reconstruct progress:
   • completed tasks = work already done — do not redo them. Trust the Task record over conversation memory.
   • in_progress tasks = work that was interrupted mid-flight. Re-read the corresponding files/diff to see how far it got before continuing. (A goal-loop stamps metadata.heartbeatAt; an in_progress task with a stale or absent heartbeat was interrupted, not running — reclaim and continue it, never redo.)
   • pending tasks with an empty blockedBy = the next available work, lowest ID first.
   • pending tasks still blockedBy open tasks = not yet startable.
3. Confirm the reconstructed state with the user before resuming: list what's done, what was in-flight, and what's next. Play the gate sound (afplay /System/Library/Sounds/Glass.aiff &) and ask: "Resume from here?"
4. Continue the workflow from the first available task, following .agents/skills/run/SKILL.md from the matching phase. Re-apply the Phase 0 tier and gate policy that was recorded in the plan.

Notes

• Goal-loop resume: if a metadata.kind=goal task or a plans/loop/<slug>.md with status: active exists, this is a goal-loop, not a one-shot /run. Resume by following .agents/skills/loop/SKILL.md → Self-healing — do this on every (re)entry (reclaim stale in_progress, idempotent-continue), then run cycles continuously to a real exit condition. Confirm once at resume, then do not pause between cycles — that's the loop's contract, and the Stop hook (bin/loop-rearm.py) keeps it alive across turns.
• This command never re-implements completed work; the Task record is the source of truth.
• If TaskList is empty and no plans/ artifact exists, there is nothing to resume — say so and suggest starting a fresh /run.
• Pairs with the Phase 6 convergence loop, which externalizes each iteration's outcome so a mid-loop crash is resumable.