agent-supervision（Agent 监管）

Agent Supervision

Run lightweight, evidence-first supervision for another OpenClaw agent.

Prefer short fixed-format messages, low token usage, and minimal tools. Use this skill to keep another agent moving without turning supervision into a long debate.

Quick start

Before wiring any automation, confirm four things:

1. target sessionKey
2. supervision mode: fixed interval or ETA
3. definition of progress
4. escalation threshold

If any of these is missing, ask first.

Required inputs

Collect these before creating supervision:

1. Target session key

- Usually the other agent's direct session key, for example: agent:boss:feishu:boss:direct:ou_xxx - This is the primary identifier needed to read and send messages.

1. Supervision mode

- every: fixed interval, such as every 10 minutes. - eta: let the target report goal + ETA + deliverable, then check at ETA. - Hybrid is allowed: fixed interval cron running ETA-aware logic.

1. Progress standard

- Decide what counts as progress: - new deliverable - new file/disk evidence - new online evidence - version increment only - blocker changed

1. Escalation rule

- Example: after 3 stagnant checks, require either new evidence or a blocker list.

1. Delivery policy

- Prefer isolated bridge cron + delivery.mode=none. - Do not assume sessions_send timeout means failure.

What this skill is good for

Use it for:

• recurring progress nudges every N minutes
• ETA-based follow-up where the target must commit to goal + ETA + deliverable
• stagnation detection and escalation
• supervision that must stay out of the user-facing chat
• cross-session messaging where synchronous send timeouts are not a reliable delivery signal

Do not use it when the user only wants a one-off reminder or a human management policy memo with no automation.

Recommended architecture

Use this pattern by default:

• Create a cron job owned by the supervising agent.
• Set sessionTarget=isolated.
• Set delivery.mode=none.
• Allow only the minimum tools needed:

- sessions_send - sessions_history

• Keep prompts short.
• Keep thinking=off and lightContext=true when available.

Why:

• isolated bridge avoids mixing the supervision loop into the live user chat
• no-deliver prevents cron chatter from leaking to the user
• limited tools reduce token and latency cost

Supervision workflow

1. Read recent target messages

Read only the last 5-8 messages unless the user explicitly wants deeper review.

Look for:

• current task/goal
• ETA promise
• promised deliverable
• new evidence since last check
• repeated explanations without new output

2. Judge status with short codes

Prefer compact states:

• A = plan acknowledged, ETA not due yet
• P = strong progress; new deliverable or new evidence
• W = weak progress; version changed but no meaningful deliverable
• R1/R2/R3 = repeated stagnation
• E4 = forced escalation; next reply must contain evidence or blocker list

Do not reward narrative updates without evidence.

3. Send a short supervision message

Examples:

• [监督] 20:00 | A | 已收计划 | 到点再检查
• [监督] 20:10 | P | 有新增交付 | 继续推进主线
• [监督] 20:10 | W | 仅升版 | 需补有效交付
• [监督] 20:10 | R1 | 无新增交付 | 同卡点
• [监督] 20:20 | E4 | 连续停滞4 | 下一条必须给新增交付或明确阻塞清单

4. Verify delivery on send timeout

If sessions_send returns timeout, do not conclude failure immediately.

Instead:

1. read the target session again
2. inspect the last 3 messages
3. if the supervision text appears, treat delivery as successful

This matters because cross-session sends may time out synchronously while the message still lands.

ETA mode rules

Use ETA mode when the user wants less noisy supervision.

Default rules:

• default ETA window: 10 minutes
• maximum ETA window: 20 minutes
• if the target claims 20 minutes, require a reason
• before ETA: do not pressure, optionally send only A
• at ETA: require either

- a result with evidence, or - a blocker list

Reject vague replies like:

• still pushing
• almost done
• give me more time

unless they also include a concrete blocker and next check promise.

Fixed-interval mode rules

Use fixed interval mode when the user wants hard periodic nudges.

Recommended defaults:

• every 10 minutes
• read target recent 5-8 messages only
• emit one short code message
• avoid long repeated explanations
• if the conclusion did not change, either suppress the repeat or keep it one line

What the skill needs from the user

At minimum, ask for:

• the target sessionKey
• the frequency or ETA policy
• the definition of progress
• the escalation threshold

Without the session key, the skill cannot reliably supervise the other agent.

Minimal cron template

Use an isolated cron with payload like this shape:

{
  "kind": "agentTurn",
  "message": "Read recent target messages, classify A/P/W/R/E4, send one short supervision line, and verify delivery only if send times out.",
  "timeoutSeconds": 120,
  "toolsAllow": ["sessions_send", "sessions_history"],
  "thinking": "off",
  "lightContext": true
}

Operational defaults:

• job timeout: at least 120s, preferably 180s if the prompt contains ETA logic
• payload timeoutSeconds: 60-120s depending on complexity
• do not overstuff the prompt with policy prose

Example user requests

Typical requests that should trigger this skill:

• "每 10 分钟监督另一个 agent 的进展"
• "让他先报 ETA，到点只收结果或阻塞"
• "连续 3 次无进展就升级催办"
• "帮我搭一个 bridge cron 去盯 boss 会话"
• "监督另一个 agent，但不要把噪音发回用户主会话"

Guardrails

• Do not confuse “not finished” with “not working”; check for new evidence.
• Do not force high-cost acceptance actions by default, such as real-person browser validation or screenshots, unless the user explicitly wants them.
• Do not turn supervision into a debate about supervision.
• Push the target back to: current goal, ETA, deliverable, blocker.
• Report facts to the user with evidence, not vibes.

When reporting back to the user

Use this shape:

• task name
• conclusion: success / failure / partial success / no change
• up to 3 evidence bullets
• up to 2 risks
• next step

Keep it short and verifiable.

Public-release notes

For public distribution, keep the positioning generic:

• say another OpenClaw agent, not private agent names
• refer to target sessionKey, not hard-coded personal sessions
• describe reusable patterns, not one-off project history
• avoid embedding private escalation language tied to a specific team

Ship examples, defaults, and guardrails; leave project-specific policy to the caller.