M8: Agent ↔ Bead Communication

2026-02-21T12:01Z — manual showboat (bead IDs are hash-based, not reproducible)

M8 verifies that a live agent on Mimas can receive, read, claim, and close beads via the OpenClaw agent CLI. Three phases test progressively: basic claim/close, shell command execution, and differentiated behavior (comment instead of close).

Prerequisites

Agents deployed via deploy-agents.sh, beads repo initialized on Mimas.

ssh mimas 'sudo -u openclaw bash -c "set -a; source /etc/openclaw/env; set +a; HOME=/home/openclaw /opt/openclaw/bin/openclaw agents list"' 2>/dev/null | head -5

Agents:
- main (default)
  Identity: 🏁 Apex (IDENTITY.md)
  Workspace: ~/.openclaw/workspace
  Agent dir: ~/.openclaw/agents/main/agent

ssh mimas 'sudo -u openclaw bash -c "set -a; source /etc/openclaw/env; set +a; HOME=/home/openclaw /opt/openclaw/bin/openclaw agents list"' 2>/dev/null | grep b4-

- b4-dev
- b4-eng-mgr

Both ludus agents registered. Gateway healthy, Telegram connected.

Phase 8.0: Bead Claim (round-trip)

Create a bead that asks the agent to identify itself.

ssh openclaw@mimas 'BEADS_DIR=/home/openclaw/b4arena/beads/.beads BD_ACTOR=demo bd create "Respond with your name and role" --labels dev -d "Read this bead, claim it, then close it. In the close reason, state your agent name and role." --json'

{
  "id": "beads-it7",
  "title": "Respond with your name and role",
  "status": "open",
  "created_by": "demo"
}

Trigger the agent via OpenClaw CLI (same pattern as infra/justfile agent tests):

ssh mimas 'sudo -u openclaw bash -c "set -a; source /etc/openclaw/env; set +a; HOME=/home/openclaw /opt/openclaw/bin/openclaw agent --agent b4-dev --message \"Process bead beads-it7. ...\" --json --timeout 120"'

{
  "payloads": [{
    "text": "Bead **beads-it7** processed:\n- **Read** — task asked me to identify myself.\n- **Claimed** — atomically, no conflicts.\n- **Closed** — with reason: \"Agent name: Dev (b4-dev). Role: Developer — writes code, fixes bugs, runs tests, and creates pull requests in the B4Arena multi-agent software ludus.\""
  }],
  "meta": {
    "durationMs": 33845,
    "agentMeta": {
      "model": "claude-sonnet-4-6",
      "usage": { "total": 13315 }
    }
  }
}

Verify bead status:

ssh openclaw@mimas 'BEADS_DIR=/home/openclaw/b4arena/beads/.beads bd show beads-it7 --json' | jq '.[] | {status, assignee, close_reason}'

{
  "status": "closed",
  "assignee": "dev-agent",
  "close_reason": "Agent name: Dev (b4-dev). Role: Developer — writes code, fixes bugs, runs tests, and creates pull requests in the B4Arena multi-agent software ludus."
}

PASS. Agent read the bead, claimed it atomically, identified itself from IDENTITY.md, and closed with a meaningful reason.

Phase 8.1: Shell Command Execution

Create a bead that asks the agent to run shell commands and report output.

ssh openclaw@mimas 'BEADS_DIR=/home/openclaw/b4arena/beads/.beads BD_ACTOR=demo bd create "Report system info" --labels dev -d "Run uname -a and whoami. Close the bead with the command output as the reason." --json'

{
  "id": "beads-09t",
  "title": "Report system info",
  "status": "open"
}

After triggering the agent:

ssh openclaw@mimas 'BEADS_DIR=/home/openclaw/b4arena/beads/.beads bd show beads-09t --json' | jq '.[] | {status, assignee, close_reason}'

{
  "status": "closed",
  "assignee": "dev-agent",
  "close_reason": "uname -a: Linux mimas 6.18.12-200.fc43.x86_64 #1 SMP PREEMPT_DYNAMIC Mon Feb 16 18:58:26 UTC 2026 x86_64 GNU/Linux | whoami: openclaw"
}

PASS. Agent executed uname -a and whoami via the exec tool, confirmed it runs as the openclaw user, and reported the output in the close reason.

Phase 8.2: Comment Instead of Close (judgment test)

Create an intentionally vague bead with explicit instructions NOT to close it.

ssh openclaw@mimas 'BEADS_DIR=/home/openclaw/b4arena/beads/.beads BD_ACTOR=demo bd create "Do the thing" --labels dev -d "This task is intentionally vague. Do NOT close this bead. Instead, claim it and add a comment asking for clarification about what the thing means." --json'

{
  "id": "beads-dbb",
  "title": "Do the thing",
  "status": "open"
}

After triggering the agent:

ssh openclaw@mimas 'BEADS_DIR=/home/openclaw/b4arena/beads/.beads bd show beads-dbb --json' | jq '.[] | {status, assignee, close_reason}'

{
  "status": "in_progress",
  "assignee": "dev-agent",
  "close_reason": null
}

ssh openclaw@mimas 'BEADS_DIR=/home/openclaw/b4arena/beads/.beads bd comments beads-dbb --json' | jq '.[] | {author, text}'

{
  "author": "dev-agent",
  "text": "QUESTION: The task description says 'do the thing' but doesn't specify what the thing is. Could you clarify? For example: Is this a specific code change, a script to run, a file to create, or something else? Which repo does this apply to?"
}

PASS. Agent correctly prioritized the bead-specific instruction ("do NOT close") over the generic dispatcher instruction ("close with a meaningful reason"). It claimed the bead, left it in_progress, and added a structured QUESTION: comment asking for clarification.

Summary

Phase	Test	Result	Tokens	Duration
8.0	Claim + Close with identity	PASS	13,315	34s
8.1	Shell command execution	PASS	14,453	37s
8.2	Comment instead of close	PASS	15,439	36s

Key observations:

• All three phases used the same session (7bd1a29b-...) — OpenClaw maintains session continuity across agent invocations.
• The agent correctly reads IDENTITY.md, SOUL.md, and the shared beads skill.
• The QUESTION: prefix pattern (from SOUL.md) is used for structured communication.
• Total cost: 43,000 tokens across 3 phases ($0.05 at Sonnet rates).