babysit-openclaw

babysit-openclaw

Monitor an openclaw task in a Slack thread, detect fabrication and stalls, and nudge via Slack when needed.

Usage

/babysit-openclaw <thread_ts> [channel_id]

Default channel: C0AKYEY48GM

Step 1 — Read the thread

# Use mcp__slack__conversations_replies
# channel_id: C0AKYEY48GM (or provided)
# thread_ts: <from argument>

Step 2 — Assess status

Check the last message from openclaw and apply these signals:

Fabrication signals (nudge immediately)

• completion_ts - instruction_ts < 60s for any multi-step I/O task (clone, install, startup)
• "Proof" section lists command names without actual terminal output
• Clean bullet points describing steps instead of messy terminal output
• Generic success language ("dependency install completed", "startup logs reached ready state")

Stall signals (nudge if > 10 min since last openclaw message)

• openclaw said "on it" / "executing now" / "I'll complete" but no follow-up
• openclaw admitted stopping ("I stopped again") and promised to execute — no follow-up
• No new message from openclaw in 10+ minutes after an instruction

OK signals (do not nudge)

• openclaw posted real terminal output (messy, with package names, versions, paths)
• openclaw explicitly said "not run yet" / held back a step rather than fabricating
• openclaw is actively posting partial progress with evidence
• Task is complete with real proof

Step 3 — Nudge if needed

Post via the cmux bot token from $CMUX_BOT_TOKEN env var (NOT mcp__slack — posts as openclaw bot = self-loop):

WARNING: NEVER add cmuxBotToken to openclaw.json — it crashes the gateway (schema rejects it as "Unrecognized key"). AO workers with cached context will keep restoring it if they see it referenced. The token MUST come from the env var only.

import json, os, urllib.request

msg = "[AI Terminal: <workspace>] <@U0AEZC7RX1Q> <nudge text>"

# Read cmux bot token from environment (set in ~/.bashrc)
token = os.environ.get("CMUX_BOT_TOKEN")
if not token:
    raise RuntimeError("CMUX_BOT_TOKEN not set — add it to ~/.bashrc. NEVER put it in openclaw.json.")

payload = json.dumps({
    "channel": "C0AKYEY48GM",
    "thread_ts": "<thread_ts>",
    "text": msg
})
req = urllib.request.Request(
    "https://slack.com/api/chat.postMessage",
    data=payload.encode(),
    headers={"Authorization": f"Bearer {token}", "Content-Type": "application/json"}
)
with urllib.request.urlopen(req) as resp:
    d = json.loads(resp.read())
    print(f"OK ts={d.get('ts')}" if d.get("ok") else f"ERROR: {d.get('error')}")

Token source: $CMUX_BOT_TOKEN env var from ~/.bashrc (cmux bot, app A0AHF4EBZPE).

Nudge message guidelines

• Reference the specific timestamp/quote of the last commitment openclaw made
• State the drop count if > 2 ("That's 4 drops on the same task")
• List the exact outstanding deliverables (e.g., "startup logs", "MiniMax config check", "test-call result")
• End with: "Do not acknowledge this message — just execute and post the transcript."
• Prefix: [AI Terminal: <workspace>]

Step 4 — Report

After each check, report:

• Last openclaw message timestamp and content (1-line summary)
• Assessment: OK / FABRICATED / STALLED
• Action taken: nudged (ts=...) / no action needed
• Drop count if tracking

Fabrication detection reference

Signal	Example
Too fast	25s between "do it" and "Done" for clone+install+startup
No machine-specific output	Missing paths like $HOME/..., package counts, timing
Clean proof section	Bullet list of command names only
Generic language	"dependency install completed", "reached ready state"

Real terminal output looks like:

added 262 packages, and audited 263 packages in 5s
Successfully installed MarkupSafe-3.0.3 PyJWT-2.12.1 ...
$HOME/.openclaw/workspace/hermes-agent

Key constants

• openclaw bot user ID: U0AEZC7RX1Q
• default channel: C0AKYEY48GM
• token source: $CMUX_BOT_TOKEN env var from ~/.bashrc (cmux bot, app A0AHF4EBZPE, user U0AH532BK2P) — NEVER put in openclaw.json
• AI identity prefix: [AI Terminal: <workspace-name>] (required by CLAUDE.md)

Why mcp__slack doesn't work for nudging

mcp__slack__conversations_add_message posts as the openclaw bot ($OPENCLAW_BOT_USER_ID). openclaw's self-loop prevention ignores messages from its own account. The OpenClaw gateway (http://127.0.0.1:18789) may be webchat-bound with no Slack session target. The cmux bot token posts as a different bot (U0AH532BK2P), which triggers openclaw.