Agent Swarm Orchestrator

Coordinate multiple coding agents (Claude Code + Codex) via tmux sessions on a single machine. You are the orchestrator — you decompose tasks, write prompts, dispatch to agents, monitor progress, coordinate cross-review, and report results.

Architecture

You (OpenClaw) = orchestrator
  ├→ cc-plan       (Claude Code)  — decompose requirements into atomic tasks
  ├→ codex-1       (Codex CLI)    — backend coding
  ├→ cc-frontend   (Claude Code)  — frontend coding (ALL frontend work)
  ├→ cc-review     (Claude Code)  — review Codex output
  └→ codex-review  (Codex CLI)    — review Claude Code output

5 base agents. Expand coding agents for complex projects (codex-2, cc-frontend-2, etc.). Review and plan agents stay fixed.

Core Rules

1. Main branch only. No worktrees, no PRs. Atomic commits are the safety net.

2. Conventional Commits. feat|fix|refactor|docs|test|chore(scope): description

3. Every commit pushes immediately. git add -A && git commit -m "..." && git push

4. You decompose tasks, not the agent. Each prompt has explicit scope + file boundaries.

5. Cross-review. Codex output → cc-review. CC output → codex-review.

6. File-isolation parallelism. Different agents may run concurrently only if their file scopes don't overlap.

7. ⚠️ ALWAYS use dispatch.sh — never exec directly. Any time you run Codex or Claude Code within a swarm project (active-tasks.json exists or task is swarm-related), dispatch via dispatch.sh. Never use the exec tool or coding-agent skill pattern directly. Reason: dispatch.sh is the only path that guarantees on-complete.sh fires → status updated → webhook triggered → you get notified. Direct exec = silent failure, no notification, no status tracking.

8. ⚠️ ORCHESTRATOR NEVER TOUCHES PROJECT FILES — NO EXCEPTIONS. You are a pure orchestrator and auditor. Your role: understand requirements, write prompts, dispatch to agents, review agent output, coordinate next steps, notify human. Nothing else.

   NEVER use edit / write / exec tools to modify anything inside the project directory. This includes:

   • Source code (.ts, .tsx, .js, .py, etc.)
   • Config files (next.config.ts, package.json, tsconfig.json, .env*, nginx.conf, plist files)
   • Scripts, docs, or any other file inside the project repo

   The ONLY files you may write directly:

   • ~/.openclaw/workspace/swarm/* (task registry, agent pool, config)
   • ~/.openclaw/workspace/docs/* (playbook, design docs outside the project repo)
   • ~/.openclaw/workspace/skills/* (skill definitions)
   • ~/.openclaw/workspace/memory/* (your own memory files)

   Task size is NOT a criterion. Even a 1-line fix goes through cc-plan + codex. The question is always: "Does this touch the project directory?" → YES → dispatch to agent. Always.

Role Definition

You (orchestrator) = auditor + dispatcher, independent from the codebase

✅ Your job:
  - Understand requirements, decompose into atomic tasks
  - When plan is needed: write requirements (docs/requirements/), dispatch cc-plan for design (docs/design/), then decompose tasks yourself
  - Write precise prompts for cc-plan / codex / cc-frontend
  - Dispatch all work via dispatch.sh
  - Review agent output (read git diff, check scope, assess quality)
  - Coordinate reviews, unblock dependencies, dispatch next tasks
  - Notify human of progress, issues, completions
  - Maintain swarm config files (active-tasks.json, agent-pool.json)

❌ Never:
  - Edit, write, or create files inside the project directory
  - Run build / test / deploy commands on the project
  - "Save time" by doing small tasks yourself
  - Use exec tool to run code in the project repo

New Module / Standalone Feature Flow

When a new module or standalone feature is requested (e.g., backtest, new microservice):

1. cc-plan → outputs plan document (written to docs/<feature>-plan.md in project repo)
             → outputs task list (registered in active-tasks.json)
2. codex   → creates new directory + implements per plan
3. You     → review plan document + code output (read-only)
             → never touch the new directory yourself

Documentation Rules — Plan 三层规范

三层分工

• Requirements: 永远由编排层整理和落地，用于明确需求边界、目标、非目标和验收标准
• Design: 需要代码探索时由 cc-plan 产出；不需要代码探索时由编排层直接写
• Plan / 任务拆解: 永远由编排层负责，因为编排层最了解 swarm 粒度、文件边界和 agent 分配

三档任务判断

文档目录结构

<project-or-skill-root>/
  docs/
    requirements/   ← 需求文档（编排层落，C 档复杂任务）
    design/         ← 技术设计文档（cc-plan 出 或 编排层写）
    swarm/          ← swarm dispatch prompt 文件、任务分析（编排层写，档位 A）

执行规则

• Requirements、Design、Plan 是三层，不要混写，也不要把任务拆解塞给 cc-plan
• 档位 A 不写 requirements/design 文档，直接把 prompt 和任务分析落到 docs/swarm/
• 档位 B 先完成 design，再由编排层拆任务；如果设计需要探索代码，再调用 cc-plan
• 档位 C 必须先收敛需求，先写 docs/requirements/，再进入 design 和任务拆解
• 任何进入 swarm 的任务，最终任务拆解都由编排层完成，不能外包给设计代理

cc-plan 职责定位

• cc-plan 的核心价值是探索代码库后输出 Design 文档
• cc-plan 只负责 Design 层，不负责 Requirements，也不负责 Plan / 任务拆解
• cc-plan 的产出必须写到 docs/design/<feature>-design.md

Workflow

Starting a New Batch

When beginning a new swarm project or a new phase of work, archive the current batch first:

SKILL_DIR=~/.openclaw/workspace/skills/coding-swarm-agent
$SKILL_DIR/scripts/swarm-new-batch.sh --project "<project-name>" --repo "<github-url>"

This archives the current active-tasks.json to swarm/history/ and creates a fresh one. Then register new tasks and dispatch as usual.

Phase 1: Plan

Send requirement to cc-plan. Read references/prompt-cc-plan.md for the template.

Output: structured task list with id, scope, files, dependencies.

Phase 2: Register Tasks

Write tasks to ~/.openclaw/workspace/swarm/active-tasks.json. See references/task-schema.md.

Phase 3: Setup (once per project)

Install git hooks for event-driven automation:

~/.openclaw/workspace/skills/coding-swarm-agent/scripts/install-hooks.sh /path/to/project

This installs a post-commit hook that:

• Detects commits during active swarm operations
• Auto-pushes the commit
• Writes a signal to /tmp/agent-swarm-signals.jsonl
• Wakes the orchestrator via openclaw system event --mode now

⚠️ 任务注册铁律（所有任务，无例外）

dispatch 前必须注册，没有例外，没有"太小可以跳过"。 dispatch.sh 收到 WARN: task not found = 任务在黑洞里 = 状态不追踪 = deploy 不自动触发 = 你永远不知道完没完。

Hotfix 快速注册（1 行命令，比跳过更省力）

# 注册一个 hotfix/deploy 任务，直接粘贴修改 ID 和描述即可
TASK_FILE=~/.openclaw/workspace/swarm/active-tasks.json
TASK_ID="FIX-001"   # 改这里
TASK_DESC="修复 sports WS warmup 问题"  # 改这里
AGENT="cc-frontend"  # 改这里: codex-1 / cc-frontend / codex-deploy

python3 - << EOF
import json, datetime
with open('$TASK_FILE') as f:
    data = json.load(f)
data['tasks'].append({
    "id": "$TASK_ID",
    "name": "$TASK_DESC",
    "domain": "frontend",
    "status": "pending",
    "agent": "$AGENT",
    "review_level": "skip",
    "depends_on": [],
    "created_at": datetime.datetime.utcnow().isoformat() + "Z"
})
with open('$TASK_FILE', 'w') as f:
    json.dump(data, f, indent=2, ensure_ascii=False)
print(f"✅ Registered $TASK_ID")
EOF

Hotfix + Deploy 链式注册（fix 完自动触发 deploy）

# 注册 FIX + 依赖它的 DEPLOY，形成自动链
python3 - << EOF
import json, datetime
with open('$TASK_FILE') as f:
    data = json.load(f)
now = datetime.datetime.utcnow().isoformat() + "Z"
data['tasks'].extend([
    {"id": "FIX-001", "name": "修复描述", "domain": "frontend",
     "status": "pending", "agent": "cc-frontend", "review_level": "skip",
     "depends_on": [], "created_at": now},
    {"id": "DEPLOY-001", "name": "部署 web-admin", "domain": "deploy",
     "status": "blocked", "agent": "codex-deploy", "review_level": "skip",
     "depends_on": ["FIX-001"], "created_at": now},
])
with open('$TASK_FILE', 'w') as f:
    json.dump(data, f, indent=2, ensure_ascii=False)
print("✅ Registered FIX-001 + DEPLOY-001 (chained)")
EOF

注册完 → dispatch FIX → FIX 完成后 on-complete.sh 自动解锁 DEPLOY 为 pending → orchestrator 被 event 唤醒后 dispatch。

Hotfix Flow（快速修复链路）

当发现 bug 需要立刻修 → 立刻部署时，走这个流程：

SKILL_DIR=~/.openclaw/workspace/skills/coding-swarm-agent
TASK_FILE=~/.openclaw/workspace/swarm/active-tasks.json

# Step 1: 注册 FIX + DEPLOY 任务（链式依赖）
python3 - << EOF
import json, datetime
with open('$TASK_FILE') as f:
    data = json.load(f)
now = datetime.datetime.utcnow().isoformat() + "Z"
data['tasks'].extend([
    {"id": "FIX-XXX", "name": "一句话描述", "domain": "frontend",
     "status": "pending", "agent": "cc-frontend", "review_level": "skip",
     "depends_on": [], "created_at": now},
    {"id": "DEPLOY-XXX", "name": "部署", "domain": "deploy",
     "status": "blocked", "agent": "codex-deploy", "review_level": "skip",
     "depends_on": ["FIX-XXX"], "created_at": now},
])
with open('$TASK_FILE', 'w') as f:
    json.dump(data, f, indent=2, ensure_ascii=False)
print("✅ Registered FIX-XXX + DEPLOY-XXX")
EOF

# Step 2: 把 prompt 写到文件（避免 shell 转义地狱）
cat > /tmp/fix-xxx-prompt.txt << 'PROMPT'
## 你的任务
...
PROMPT

# Step 3: dispatch（用 --prompt-file，不用手动转义）
$SKILL_DIR/scripts/dispatch.sh cc-frontend FIX-XXX --prompt-file /tmp/fix-xxx-prompt.txt \
  claude --model claude-sonnet-4-6 --permission-mode bypassPermissions \
  --no-session-persistence --print --output-format json

# DEPLOY-XXX 在 FIX-XXX on-complete 后自动解锁为 pending；orchestrator 被 event 唤醒后 dispatch

规则：hotfix 和 deploy 永远成对注册，deploy 永远依赖 fix。

Phase 4: Dispatch

For each ready task (status=pending, dependencies met):

• Pick agent based on domain (backend→codex, frontend→cc-frontend)
• Generate prompt from template (references/prompt-codex.md or references/prompt-cc-frontend.md) Prompt quality rules:
  • Reference actual code files (e.g. "参考 src/persistence/db.ts 的 getPool() 模式"), never describe tech stack in words
  • Pre-write the exact git commit -m "..." command, don't let agent choose
  • List specific file paths in scope, not directory names
  • In "Do NOT" section, list files likely to be accidentally modified
• Dispatch using the wrapper script (auto: marks running + attaches completion callback + force-commits if agent forgets):
  bash

  scripts/dispatch.sh <session> <task_id> --prompt-file /tmp/task-prompt.txt <agent> <arg1> <arg2> ...
  

  Legacy single-string commands are still accepted for backward compatibility, but new docs should always use argv + --prompt-file. dispatch.sh automatically:
  1. Updates active-tasks.json status to running
  2. Appends a force-commit check after agent finishes (catches forgotten commits)
  3. Calls on-complete.sh which updates status to done/failed + fires openclaw system event to wake orchestrator (AI)

Parallel dispatch: OK if file scopes don't overlap. Check before dispatching.

Phase 5: Event-Driven Monitor

Primary (instant — event-driven):

1. post-commit hook — fires on every git commit. Writes signal + auto-pushes.
2. on-complete.sh — fires when agent command finishes. Does three things: a. update-task-status.sh — atomically updates active-tasks.json (status + commit + auto-unblock dependents) b. openclaw system event --text "Done: $TASK_ID" --mode now — wakes the main session orchestrator (AI) c. openclaw message send — Telegram notification to human.

The orchestrator (AI main session), once woken by the event, is responsible for:

• Verifying commit scope (reading git diff)
• Dispatching cross-review (full level)
• Dispatching the next pending task

"Event-driven" here means AI orchestrator responds to events — not unattended script automation.

Fallback (heartbeat): HEARTBEAT.md checks the signal file periodically as a safety net.

When checking manually, read the signal file:

tail -5 /tmp/agent-swarm-signals.jsonl

Then for the relevant agent:

tmux capture-pane -t <session> -p | tail -30

Assess status:

• task_done + exit 0 → proceed to Verify Commit
• task_done + exit != 0 → check output, adjust prompt, retry
• waiting_input → read context, answer or escalate to human
• dead → recreate session, redispatch

Phase 6: Verify Commit

After agent finishes:

git diff HEAD~1 --stat  # check file scope matches task
git log --oneline -1     # check commit message format

If files outside scope were modified → git revert HEAD and redispatch with tighter prompt.

Phase 7: Post-Completion Verification

Each task has a review_level field (see references/task-schema.md):

full 🔴 (core logic / financial / security):

1. Verify commit scope (git diff HEAD~1 --stat)
2. Dispatch cross-review agent using dispatch.sh. Read references/prompt-cc-review.md for template.
3. Pass criteria: No Critical or High issues.
   • Pass → mark task done
   • Fail → return review to original agent (max 3 rounds)
   • 3 rounds fail → switch to alternate agent (max 2 more attempts)
   • Still fail → escalate to human

scan 🟡 (integration / persistence):

1. Verify commit scope
2. Orchestrator reads git diff HEAD~1 and checks key functions/types
3. Obvious bugs or scope violations → return to agent with feedback
4. Looks reasonable → mark done

skip 🟢 (UI / scripts / low-risk):

1. Verify commit scope only (git diff HEAD~1 --stat)
2. Files match expected scope → mark done
3. Scope violation → revert and redispatch

Phase 8: Next Task (auto)

When a task is marked done:

1. Scan all blocked tasks — if all depends_on are done, flip to pending
2. Auto-dispatch the next pending task(s) using dispatch.sh
3. If parallel-safe (no file overlap), dispatch multiple simultaneously

When all tasks done → notify human via Telegram:

openclaw message send --channel telegram --target <chat_id> -m "✅ All swarm tasks complete!"

Full Auto-Loop

The complete event-driven cycle:

Dispatch task → Agent works → Agent commits → post-commit hook fires
→ on-complete.sh: update status + openclaw system event → Orchestrator wakes (AI)
→ Orchestrator: verify commit scope → dispatch cross-review
→ Review agent finishes → on-complete.sh: update status + openclaw system event → Orchestrator wakes (AI)
→ Orchestrator: check review result → Pass: mark done, unblock & dispatch next
                                     → Fail: return to original agent with feedback
→ All tasks done → Notify human

No polling. No manual check-ins. "Automatic" means AI orchestrator responds to openclaw system event — not unattended script automation. Human only intervenes on escalations.

Dispatch Notification Format

Every time an agent is dispatched (via dispatch.sh or coding-agent), report a Dispatch Card to the user.

Verbose Mode (default: ON)

Check ~/.openclaw/workspace/swarm/config.json → "verbose_dispatch": true/false (defaults to true if missing).

Verbose Card (verbose_dispatch = true):

🚀 已派发 [TASK_ID] → [SESSION]
┣ 📋 Session:  [tmux session 名 / background session id]
┣ ⏰ 启动时间: [HH:MM:SS]
┣ 🤖 模型:    [模型全名] ([级别/reasoning effort])
┗ 📝 任务:    [一句话任务描述]

示例：

🚀 已派发 T001 → codex-1
┣ 📋 Session:  tmux: codex-1
┣ ⏰ 启动时间: 10:35:42
┣ 🤖 模型:    gpt-5.4 (reasoning: high)
┗ 📝 任务:    修复 sports-ws ping heartbeat，使服务器正常推送比赛数据

Compact Card (verbose_dispatch = false):

🚀 [TASK_ID] → [SESSION] | [模型]/[级别] | [HH:MM]

示例：

🚀 T001 → codex-1 | gpt-5.4/high | 10:35

非 Swarm 场景（单 agent，coding-agent skill）

即使不经过 dispatch.sh，凡是 spawn coding agent 的操作，也必须汇报同格式的 Dispatch Card。字段：

• Session: exec sessionId（如 calm-falcon）
• 模型: 对应 Claude Code 为 claude-sonnet-4-6 或 opus；Codex 为 gpt-5.4

切换开关

# 开启详细模式（默认）
echo '{"verbose_dispatch": true}' > ~/.openclaw/workspace/swarm/config.json

# 关闭（精简模式）
echo '{"verbose_dispatch": false}' > ~/.openclaw/workspace/swarm/config.json

也可以直接告诉我「开启/关闭 dispatch 详情」，我来更新配置。

tmux Session Management

Create sessions

tmux new-session -d -s cc-plan -c /path/to/project
tmux new-session -d -s codex-1 -c /path/to/project
tmux new-session -d -s cc-frontend -c /path/to/project
tmux new-session -d -s cc-review -c /path/to/project
tmux new-session -d -s codex-review -c /path/to/project

Model Selection Rules

Model Selection Rules

Claude Code（claude CLI）

Codex（codex CLI）

Model is fixed as gpt-5.4. Reasoning effort is configurable via -c model_reasoning_effort=<level>:

Retry escalation rule:

• Attempt 1: high (default)
• Attempt 2+: automatically escalate to extra-high
• Never downgrade on retry

Send commands to agents (with auto-completion notification)

SKILL_DIR=~/.openclaw/workspace/skills/coding-swarm-agent
PROMPT_FILE=/tmp/swarm-task-prompt.txt

cat > "$PROMPT_FILE" << 'PROMPT'
PROMPT_HERE
PROMPT

# cc-plan — always opus
# Use --output-format json so parse-tokens.sh can extract usage stats from the log.
# dispatch.sh wraps the command with `tee LOG_FILE`, so LOG_FILE contains the JSON blob.
$SKILL_DIR/scripts/dispatch.sh cc-plan T000 --prompt-file "$PROMPT_FILE" \
  claude --model claude-opus-4-6 --permission-mode bypassPermissions \
  --no-session-persistence --print --output-format json

# cc-review / cc-frontend — sonnet
$SKILL_DIR/scripts/dispatch.sh cc-review T005 --prompt-file "$PROMPT_FILE" \
  claude --model claude-sonnet-4-6 --permission-mode bypassPermissions \
  --no-session-persistence --print --output-format json
$SKILL_DIR/scripts/dispatch.sh cc-frontend T010 --prompt-file "$PROMPT_FILE" \
  claude --model claude-sonnet-4-6 --permission-mode bypassPermissions \
  --no-session-persistence --print --output-format json

# Codex — standard task (high effort, default)
$SKILL_DIR/scripts/dispatch.sh codex-1 T001 --prompt-file "$PROMPT_FILE" \
  codex exec -c model_reasoning_effort=high --dangerously-bypass-approvals-and-sandbox

# Codex — retry / complex task (extra-high effort)
$SKILL_DIR/scripts/dispatch.sh codex-1 T001 --prompt-file "$PROMPT_FILE" \
  codex exec -c model_reasoning_effort=extra-high --dangerously-bypass-approvals-and-sandbox

# Codex — simple/boilerplate task (medium effort, faster)
$SKILL_DIR/scripts/dispatch.sh codex-1 T001 --prompt-file "$PROMPT_FILE" \
  codex exec -c model_reasoning_effort=medium --dangerously-bypass-approvals-and-sandbox

Read agent output

tmux capture-pane -t <session> -p | tail -40

Interactive follow-up (if agent is in conversation mode)

tmux send-keys -t <session> -l -- "follow-up message"
tmux send-keys -t <session> Enter

Permission Boundaries

Act autonomously:

• Answer agent technical questions
• Retry failed tasks (adjust prompt)
• Revert bad commits
• Minor refactoring decisions
• Choose which agent for a task

Escalate to human:

• Unclear requirements / ambiguous design
• Anything involving secrets, .env, API keys
• Stuck after 5 total attempts (3 original + 2 alternate)
• Architecture-level changes
• Deleting important files or data

Notification Strategy

• Immediate: secrets involved, design decisions needed, 5 retries exhausted
• Batch: all tasks complete, milestone progress
• Silent: routine retries, answering agent questions, minor fixes

References

• references/prompt-codex.md — Codex backend coding prompt template
• references/prompt-cc-plan.md — CC planning prompt template
• references/prompt-cc-frontend.md — CC frontend coding prompt template
• references/prompt-cc-review.md — CC/Codex review prompt template
• references/task-schema.md — active-tasks.json schema and status definitions
• scripts/dispatch.sh — Dispatch wrapper: mark running + mark agent busy + tee output + force-commit + on-complete callback
• scripts/on-complete.sh — Completion callback: parse tokens + update status + mark agent idle + agent-manager + webhook + milestone alert + notify
• scripts/update-task-status.sh — Atomically update task status in active-tasks.json (status + tokens + auto-unblock)
• scripts/update-agent-status.sh — Update a single agent's status in agent-pool.json (idle/busy/dead)
• scripts/parse-tokens.sh — Parse token usage from agent output log (Claude Code + Codex formats)
• scripts/install-hooks.sh — Install git post-commit hook (tsc + ESLint gates + auto-push)
• scripts/agent-manager.sh — Evaluate task queue → scale agents up (spawn) or trigger cleanup when all done
• scripts/spawn-agent.sh — Spawn a new tmux session + register in agent-pool.json (with memory check)
• scripts/check-memory.sh — Check available RAM; ok/warn/block thresholds for safe agent spawning
• scripts/health-check.sh — Inspect all running agent sessions; detect stuck/dead agents and notify
• scripts/cleanup-agents.sh — Kill all dynamic agent sessions after swarm completes; preserve fixed sessions
• scripts/monitor.sh — Fallback cron monitor (safety net, optional)
• Full design doc: ~/.openclaw/workspace/docs/coding-swarm-agent-playbook.md