resume-after-gateway-restart（网关重启后恢复）

Gateway Resume

Gateway restarts kill the current agent process. If you need to restart mid-task, you must create a durable wakeup *before* restarting.

This skill uses cron + isolated agentTurn + explicit delivery route (not systemEvent) so it:

• survives restarts (cron is persisted to disk)
• replies back to the same place (Discord channel, Telegram DM/topic, etc)
• can pin the cron run to the same model/thinking that was active in the requesting session

Procedure (recommended)

0) Tell the user you’re about to go down (own message)

Send a short message in the current chat before you schedule the restart:

• "Heads up — restarting the gateway for ~10–20s. I’ll be back momentarily and will resume automatically."

Keep this as its own message (don’t bundle it with other info).

Common failure mode: writing a plan but forgetting to actually schedule the cron + restart. If you don’t run Steps 3–4, the gateway won’t go down and you won’t come back.

Do not require a user 'go' message. Instead, schedule the restart to occur ~15s in the future (see Step 4) so this message has time to deliver before the gateway is killed.

1) Save resume context

Write memory/post-restart-task.md:

# Post-restart: <brief description>
- Channel: <where to reply>
- Was doing: <what was in progress>
- Next step: <what to do next on resume>
- Status: pending

2) Resolve the return route + session model (Option B)

From the current chat/session, capture:

• sessionKey (e.g. agent:main:discord:channel:<id> or agent:main:telegram:direct:<id>)
• deliveryChannel (e.g. discord, telegram)
• deliveryTo:

- Discord channel: channel:<channelId> - Discord DM: user:<userId> - Telegram DM/topic: use the chat id or -100…:topic:<id>

• modelProvider + model for the session (e.g. openai-codex + gpt-5.2)
• thinkingLevel for the session

Tip: openclaw sessions --active 240 --json includes these fields.

3) Schedule two one-shot cron jobs (back message + resume)

Use the bundled script (Option B: it infers route + model from the session store):

skills/gateway-resume/scripts/schedule-resume-cron.sh \
  --back-delay 75s \
  --delay 90s \
  --session-key "<sessionKey>"

The script will:

• infer the delivery route from the session key (Discord channel/DM, Telegram DM/group)
• read openclaw sessions --json to infer modelProvider/model and thinking level
• schedule two isolated agentTurn cron jobs with explicit delivery:

1) a short, personable "I’m back" message (own message) 2) a resume job that reads memory/post-restart-task.md and completes the next step

Notes:

• Use --delay 90s (or more) so the job fires well after the gateway finishes restarting.
• The cron job runs isolated, but delivers back to the captured route.
• The “back” job is a *separate* message to make the restart feel responsive, even if the resume step takes longer.

4) Restart (delayed)

Do not restart immediately (it can kill the process before the pre-restart message delivers). Instead schedule a restart a few seconds in the future using systemd-run:

skills/gateway-resume/scripts/delayed-gateway-restart.sh 15

This schedules a systemctl --user restart openclaw-gateway.service in a transient user unit.

5) On the cron wake

The scheduled jobs should:

1) Send a short, personable "I’m back" message (by itself). 2) Then resume: - read memory/post-restart-task.md - do Next step - append a handled marker (avoid needing deletes) - reply back to the captured destination with the result

Do not include raw markers like BACK_FROM_RESTART in user-facing text.

Troubleshooting

• Nothing happens after restart: confirm the cron jobs exist *before* restarting:

- openclaw cron list

• Cron fires but no message is delivered: route inference may be wrong for this sessionKey pattern.

- Re-run the schedule script with the correct --session-key.

• Model inference fails: the session may have fallen outside the --active window.

- Edit schedule-resume-cron.sh to use a larger --active range or no filter.

Why not systemEvent?

systemEvent targets the main session, which may be a different model/provider and may fail (or respond in the wrong place). Isolated agentTurn + explicit delivery is deterministic.