openclaw - 💡(How to fix) Fix /steer command for reliable mid-run subagent steering [1 participants]

Problem Summary

The “steer” operation is currently done via sessions_send, which often times‑out and gives no delivery confirmation. Users need a reliable, user‑facing way to inject a message into the active sub‑agent(s) while a session is running.

Root Cause

• sessions_send is a fire‑and‑forget RPC that blocks on the agent’s HTTP client; when the agent is busy (e.g., in a tool call) the request hangs and eventually times out.
• No explicit “steering” endpoint → no ACK/queueing → the UI can’t tell whether the message arrived.

Fix Plan

1. Add a new slash‑command /steer <message>

1. Register the command (example for a FastAPI‑based bot).

   from fastapi import APIRouter, Request
   router = APIRouter()
   
   @router.post("/commands/steer")
   async def steer_cmd(req: Request):
       payload = await req.json()
       user_id = payload["user_id"]
       session_id = payload["session_id"]
       text = payload["text"]          # whole command after /steer
       return await handle_steer(user_id, session_id, text)

2. Expose it to the chat platform (Slack, Discord, etc.) – add /steer to the app manifest.

2. Resolve the active sub‑agent(s)

async def get_active_subagents(session_id: str) -> List[SubAgent]:
    # Example: a Redis sorted‑set `session:{id}:subagents` stores (timestamp, subagent_id)
    ids = await redis.zrevrange(f"session:{session_id}:subagents", 0, -1)
    return [await SubAgent.load(id_) for id_ in ids]

If you only want the most recent, take the first element.

3. Deliver the steering message reliably

Create a steering queue (Redis list or RabbitMQ) per sub‑agent.

STEER_QUEUE = "steer:{sub_id}"          #