---

name: restart-task-recovery description: Preserve and resume in-progress multi-agent work across OpenClaw config patch/apply restarts. Use when a restart is required during active tasks, when users ask to minimize interruption, or when agent runs/tool calls were interrupted by gateway restart/timeouts.

Restart Task Recovery

Use this workflow to maximize successful recovery after OpenClaw restart.

1) Pre-restart checkpoint (required)

Before any gateway.config.patch, gateway.config.apply, gateway.update.run, or gateway.restart:

1. List active sessions that may be impacted (sessions_list).
2. For each active work session, capture the latest context (sessions_history, limit 20-50).
3. Write a compact checkpoint file at:
   • memory/restart-checkpoints/<YYYY-MM-DD>/<HHmmss>.md
4. Include per session:
   • sessionKey / label / agent
   • goal
   • last completed step
   • next exact step
   • blocked dependencies (if any)
   • a ready-to-send resume message (1-2 lines)

Keep checkpoint concise and executable.

2) Restart with explicit recovery intent

When calling gateway restart/config change, set note to include recovery intent, e.g.:

• “配置已更新并重启；将按 checkpoint 恢复中断任务。”

3) Post-restart recovery sweep

After restart:

1. Re-list sessions (sessions_list) and compare against checkpoint.
2. For each interrupted/idle target session, send resume message via sessions_send:
   • “Continue where you left off. Last completed: . Next: . If previous tool call failed, retry from .”
3. Do not poll in tight loops. Check on-demand only.
4. Summarize recovery status to user:
   • recovered sessions
   • still blocked sessions
   • manual follow-up needed

4) Idempotent task design rules

When resuming tasks, enforce:

1. Re-run-safe steps (idempotency key / upsert / duplicate-safe writes).
2. Small step boundaries with explicit “done markers”.
3. External writes batched, not one-by-one loops.
4. On uncertainty, verify state first then continue.

5) V2 automation helper

Use script: scripts/build_checkpoint.py to generate checkpoint markdown from structured JSON.

Example:

cat session-snapshot.json | python3 scripts/build_checkpoint.py memory/restart-checkpoints/$(date +%F)/$(date +%H%M%S).md

Expected stdin JSON shape:

{
  "sessions": [
    {
      "sessionKey": "agent:engineer:main",
      "agentId": "engineer",
      "goal": "Finish regression verification",
      "lastDone": "401/幂等/时区/retention case passed",
      "nextStep": "Publish final acceptance summary",
      "blockers": "none"
    }
  ]
}

6) V3 resume-plan automation

Use script: scripts/generate_resume_plan.py to parse the latest checkpoint and produce a structured resume plan.

Example:

python3 scripts/generate_resume_plan.py memory/restart-checkpoints/2026-03-09/162200.md /tmp/resume-plan.json

Then send each items[].resumeMessage to items[].sessionKey via sessions_send.

Rules:

• Send once per session (no loop polling).
• If a session is already active and progressing, skip resend.
• After sends, post one concise recovery summary to user.

7) V4 one-click recovery payload generator

Use script: scripts/recover_from_latest_checkpoint.py.

It auto-selects the latest checkpoint file and emits a ready JSON payload list for sessions_send calls.

Examples:

# Use latest checkpoint automatically
python3 scripts/recover_from_latest_checkpoint.py > /tmp/recover-actions.json

# Use a specific checkpoint
python3 scripts/recover_from_latest_checkpoint.py memory/restart-checkpoints/2026-03-09/162200.md > /tmp/recover-actions.json

Execution guidance:

• Read /tmp/recover-actions.json
• Execute each actions[] item with sessions_send
• Post one concise summary to user

8) V5 pre-resume verifier + manual confirmation gate

Use script: scripts/pre_resume_verify.py to score resume actions before sending.

Examples:

python3 scripts/pre_resume_verify.py /tmp/recover-actions.json /tmp/recover-verified.json

Behavior:

• Marks each action as risk=normal|high
• high risk actions are set to decision=hold and requiresManualConfirm=true
• Only send decision=send automatically
• Ask user confirmation before executing held actions

Recommended execution flow:

1. Generate actions with V4
2. Verify with V5
3. Send all decision=send
4. Present decision=hold list to user for explicit confirmation

9) V6 execution-plan generator (auto-send safe items)

Use script: scripts/execute_verified_recovery.py with V5 output.

Example:

python3 scripts/execute_verified_recovery.py /tmp/recover-verified.json > /tmp/recover-exec.json

Behavior:

• Emits sendActions[] for auto-safe resumes (decision=send)
• Emits holdForManualConfirm[] for risky resumes (decision=hold)

Execution:

1. Execute all sendActions[] with sessions_send
2. Ask user to confirm holdForManualConfirm[]
3. Execute confirmed held items
4. Post concise summary

10) Message templates

Read and use: references/templates.md