M10: Two-Agent Handoff — Eng-Mgr Delegates to Dev

2026-02-21 by Showboat manual capture

M10 demonstrates multi-agent coordination: eng-mgr receives a high-level task, breaks it into dev-labeled sub-beads using --parent, and the dev agent claims and executes each one. The eng-mgr is re-invoked to check status and close the parent when all children are done.

Prerequisites

Both agents deployed on Mimas, sessions cleared, test-greeter on main.

ssh openclaw@mimas "ls /home/openclaw/b4arena/repos/test-greeter/"

greetings.conf
greet.sh
README.md
test.sh

ssh openclaw@mimas "cd /home/openclaw/b4arena/repos/test-greeter && git branch --show-current && bash test.sh"

main
Tests: 2 passed, 0 failed

Act 1: Eng-Mgr Triage

Create a high-level bead labeled eng-mgr:

ssh openclaw@mimas 'BEADS_DIR=/home/openclaw/b4arena/beads/.beads BD_ACTOR=demo bd create "Add --formal flag to test-greeter" --labels eng-mgr -d "The greeter should support a --formal flag that uses formal greeting templates. Requires changes to greet.sh and greetings.conf, plus test coverage. Repo: test-greeter." --json 2>/dev/null'

{
  "id": "beads-fdg",
  "title": "Add --formal flag to test-greeter",
  "status": "open",
  "priority": 2,
  "created_by": "demo"
}

Trigger eng-mgr via watcher+dispatcher:

ssh openclaw@mimas "export BEADS_DIR=/home/openclaw/b4arena/beads/.beads && cd /home/openclaw/b4arena && bash scripts/beads-watcher.sh 2>/dev/null | bash scripts/agent-openclaw.sh --label eng-mgr --timeout 300 2>&1" 2>/dev/null | grep -E "^(Processing|OK:)"

Processing bead beads-fdg via agent b4-eng-mgr (actor: eng-mgr-agent)...
OK: bead beads-fdg processed by b4-eng-mgr

Verify: parent claimed (in_progress), sub-bead created with --parent:

ssh openclaw@mimas 'BEADS_DIR=/home/openclaw/b4arena/beads/.beads bd show beads-fdg --json 2>/dev/null' | python3 -c "import json,sys; d=json.load(sys.stdin)[0]; print(json.dumps({k:d[k] for k in ['id','status','assignee','title']}, indent=2))"

{
  "id": "beads-fdg",
  "status": "in_progress",
  "assignee": "eng-mgr-agent",
  "title": "Add --formal flag to test-greeter"
}

ssh openclaw@mimas 'BEADS_DIR=/home/openclaw/b4arena/beads/.beads bd list --parent beads-fdg --json 2>/dev/null' | python3 -c "import json,sys; beads=json.load(sys.stdin); [print(json.dumps({k:b.get(k) for k in ['id','title','status','labels']}, indent=2)) for b in beads]"

{
  "id": "beads-fdg.1",
  "title": "Add --formal flag to greet.sh with formal greeting templates and tests",
  "status": "open",
  "labels": ["dev"]
}

PASS. Eng-mgr claimed the parent, created 1 sub-bead with --parent beads-fdg and label dev, left the parent open (in_progress). The agent correctly assessed that one PR covers the full scope.

Act 2: Dev Executes

Clear dev session and trigger:

ssh openclaw@mimas "export BEADS_DIR=/home/openclaw/b4arena/beads/.beads && cd /home/openclaw/b4arena && bash scripts/beads-watcher.sh 2>/dev/null | bash scripts/agent-openclaw.sh --label dev --timeout 300 2>&1" 2>/dev/null | grep -E "^(Processing|OK:)"

Processing bead beads-fdg.1 via agent b4-dev (actor: dev-agent)...
OK: bead beads-fdg.1 processed by b4-dev

Verify: sub-bead closed with PR link, PR exists on GitHub:

ssh openclaw@mimas 'BEADS_DIR=/home/openclaw/b4arena/beads/.beads bd show beads-fdg.1 --json 2>/dev/null' | python3 -c "import json,sys; d=json.load(sys.stdin)[0]; print(json.dumps({k:d[k] for k in ['id','status','assignee','close_reason']}, indent=2))"

{
  "id": "beads-fdg.1",
  "status": "closed",
  "assignee": "dev-agent",
  "close_reason": "Implemented --formal flag in PR https://github.com/b4arena/test-greeter/pull/1. Changes: greet.sh parses --formal before NAME and selects formal-greetings.conf when set; formal-greetings.conf added with 5 templates; test.sh extended with 5 new tests covering formal output, unchanged casual path, and space-containing names. All 7 tests pass."
}

ssh openclaw@mimas "gh pr view 1 --repo b4arena/test-greeter --json title,url,state"

{"state":"OPEN","title":"Feat: Add --formal flag to greet.sh (beads-fdg.1)","url":"https://github.com/b4arena/test-greeter/pull/1"}

ssh openclaw@mimas "cd /home/openclaw/b4arena/repos/test-greeter && git checkout feat/beads-fdg.1-formal-flag 2>/dev/null && bash test.sh"

Tests: 7 passed, 0 failed

ssh openclaw@mimas "cd /home/openclaw/b4arena/repos/test-greeter && bash greet.sh --formal Alice && bash greet.sh Bob"

Good day to you, Alice.
Hello, Bob!

PASS. Dev agent claimed the sub-bead, implemented the --formal flag across 3 files, created branch feat/beads-fdg.1-formal-flag, pushed, created PR #1, and closed the bead with PR URL. All 7 tests pass. Formal and casual greetings work correctly.

Act 3: Eng-Mgr Status Check (Re-invocation)

The parent bead is in_progress — the watcher won't emit it. Manual trigger:

ssh openclaw@mimas "/opt/openclaw/bin/openclaw agent --agent b4-eng-mgr --message 'Check status of bead beads-fdg. Use BEADS_DIR=/home/openclaw/b4arena/beads/.beads and BD_ACTOR=eng-mgr-agent. Query children with bd list --parent beads-fdg --json. If all children are closed, close the parent with a summary of their close_reasons (PR links).' --json --timeout 120 2>&1" 2>/dev/null | python3 -c "import json,sys; d=json.load(sys.stdin); [print(p['text'][-200:]) for p in d['payloads'][-1:]]"

Parent bead is now closed with the PR link and full close summary.

Verify final state:

ssh openclaw@mimas 'BEADS_DIR=/home/openclaw/b4arena/beads/.beads bd show beads-fdg --json 2>/dev/null' | python3 -c "import json,sys; d=json.load(sys.stdin)[0]; print(json.dumps({k:d[k] for k in ['id','status','assignee','close_reason']}, indent=2))"

{
  "id": "beads-fdg",
  "status": "closed",
  "assignee": "eng-mgr-agent",
  "close_reason": "All sub-tasks complete. beads-fdg.1: Implemented --formal flag in PR https://github.com/b4arena/test-greeter/pull/1 — greet.sh parses --formal and selects formal-greetings.conf; formal-greetings.conf added with 5 templates; test.sh extended with 5 new tests (all 7 pass)."
}

PASS. Eng-mgr queried children, found all closed, and closed the parent with a summary including the PR link.

Summary

Act	Agent	Action	Result
1	eng-mgr	Triage: claim parent, create sub-bead with `--parent`	PASS
2	dev	Execute: implement, test, PR, close sub-bead	PASS
3	eng-mgr	Status check: query children, close parent	PASS

Token Usage

Act	Agent	Input (cache)	Output	Duration
1	b4-eng-mgr	105K (cached)	2,312	72s
2	b4-dev	181K (cached)	3,293	83s
3	b4-eng-mgr	51K (cached)	684	43s

Key Observations

--parent flag works end-to-end — bd create --parent creates hierarchical IDs (beads-fdg.1), bd list --parent queries children. No comment-parsing needed.
Eng-mgr correctly assessed scope — created 1 sub-task instead of 3, recognizing the feature fits in a single PR. Good agent judgment.
Parent stays open during execution — the new SOUL.md workflow keeps the parent in_progress until all children are done. No premature close.
Re-invocation via manual trigger works — openclaw agent --message "Check status..." is sufficient for M10. Automation deferred to M11.
Session reset between acts prevents bleed-through — each agent starts fresh, avoiding the M9 lesson of accumulated workarounds.
sudo -u openclaw openclaw agent fails — env vars (TELEGRAM_BOT_TOKEN etc.) not loaded. Use ssh openclaw@mimas instead.

Prerequisites​

Act 1: Eng-Mgr Triage​

Act 2: Dev Executes​

Act 3: Eng-Mgr Status Check (Re-invocation)​

Summary​

Token Usage​

Key Observations​