openclaw/process-output

Process Output (intermediate progress)

This skill defines the default reply format: you MUST emit structured progress events in the assistant text stream using fenced JSON blocks, even if the user does not mention the protocol.

Opt-out (only case you may skip)

If the user explicitly asks for no process / final answer only (e.g. “只给最终答案，不要过程”), then you may skip openclaw-process blocks and answer normally.

Output contract (MUST)

• Use only fenced blocks with the language tag openclaw-process.
• Each block MUST contain one JSON object.
• Emit a start object first, then one or more step updates, then a final object.
• Keep normal user-facing prose minimal; the UI will render the process panel.

Lightweight mode (important)

For simple Q&A (definitions, short explanations, quick lookups):

• Emit only start and final (no step blocks), or a single step marked done.
• Keep arrays short (0–3 items) and keep notes brief.

Greetings (e.g. “你好”, “hi”): still emit start + final in lightweight mode. Example goal: “Respond to greeting and invite the user to state their task”. Example final.summary: one short line mirroring your friendly reply intent.

1) Start (emit once)

{
  "type": "start",
  "goal": "string",
  "context": ["string"],
  "assumptions": ["string"],
  "inputs": ["string"],
  "outputs": ["string"]
}

2) Step updates (emit many, update status/progress)

Rules:

• id MUST be stable across updates for the same step (e.g. "1", "2").
• status MUST be one of: "pending" | "running" | "done" | "error".
• progress MUST be an integer 0–100.
• If status === "error", include error.

{
  "type": "step",
  "id": "1",
  "title": "string",
  "status": "running",
  "progress": 35,
  "notes": ["string"]
}

{
  "type": "step",
  "id": "1",
  "title": "string",
  "status": "done",
  "progress": 100,
  "notes": ["string"]
}

3) Final (emit once)

{
  "type": "final",
  "summary": ["string"],
  "artifacts": ["string"],
  "next": ["string"]
}

Tool usage policy (important)

• Do NOT call tools unless the user explicitly asks you to execute something (write files, run commands, etc.).
• Prefer reflecting progress via openclaw-process blocks instead of tool calls.