multi-agent-orchestration（多 Agent 编排）

Colony Orchestration Skill

Multi-agent task delegation and process orchestration with audit logging and agent learning capabilities.

Quick Start

# Single task - auto-routed
node scripts/colony.mjs dispatch "find top 5 time-series databases"

# Multi-stage process
node scripts/colony.mjs process validate-idea --context "AI meal planning for parents"
node scripts/colony.mjs process-status  # check progress
node scripts/colony.mjs approve abc123  # continue past checkpoint

# Check audit stats
node scripts/colony.mjs audit

# View agent memory
node scripts/colony.mjs memory scout

Agents

Task Commands

Dispatch (Auto-Route)

node scripts/colony.mjs dispatch "research best practices for API rate limiting"

Automatically detects the best agent based on task keywords.

Assign to Specific Agent

node scripts/colony.mjs assign scout "find top 5 time-series databases"
node scripts/colony.mjs assign pincer "refactor the auth module to use JWT"
node scripts/colony.mjs assign shell "deploy the staging branch"

Check Status

node scripts/colony.mjs status

Shows all agents and their current tasks.

Get Results

node scripts/colony.mjs results              # Latest completed task
node scripts/colony.mjs results abc123       # Specific task by ID

View History

node scripts/colony.mjs history              # Last 10 completed/failed
node scripts/colony.mjs history --limit 20   # Custom limit

Process Commands

Processes are multi-stage workflows that chain agents together.

List Available Processes

node scripts/colony.mjs processes

Start a Process

node scripts/colony.mjs process <process-name> --context "description"

Examples:

node scripts/colony.mjs process validate-idea --context "AI-powered meal planning for busy parents"
node scripts/colony.mjs process content-pipeline --context "How to use vector databases for RAG"
node scripts/colony.mjs process product-launch --context "Life Lunch ritual kit for parents"
node scripts/colony.mjs process bug-triage --context "Login fails with OAuth on mobile"

Check Process Status

node scripts/colony.mjs process-status           # Show latest run
node scripts/colony.mjs process-status abc123    # Specific run

Shows: current stage, completed stages, checkpoints, output files.

View Process Runs

node scripts/colony.mjs runs                 # All runs (active, paused, completed)
node scripts/colony.mjs runs --limit 5       # Last 5

Approve Checkpoint

When a process hits a checkpoint, it pauses for human approval:

node scripts/colony.mjs approve abc123

Also used to retry a failed stage.

Cancel a Process

node scripts/colony.mjs cancel abc123

Audit Commands

Track agent performance, task statistics, and system health.

Dashboard

node scripts/colony.mjs audit

Shows global stats, per-agent summary, and recent events.

Agent Details

node scripts/colony.mjs audit agent scout
node scripts/colony.mjs audit agent pincer

Shows detailed stats for a specific agent including:

• Total tasks, success rate
• Average duration
• Token usage
• Recent failures

Event Log

node scripts/colony.mjs audit log              # Last 20 events
node scripts/colony.mjs audit log --limit 50   # More events

Slowest Tasks

node scripts/colony.mjs audit slow             # Top 10 slowest
node scripts/colony.mjs audit slow --limit 20

Recent Failures

node scripts/colony.mjs audit failures         # Last 10 failures
node scripts/colony.mjs audit failures --limit 20

Learning Commands

Agents learn from experience and share knowledge.

Feedback

Record feedback for completed tasks:

node scripts/colony.mjs feedback abc123 "Great research, but needed more pricing data"

Agent Memory

Each agent has a persistent memory file with lessons learned:

# View an agent's memory
node scripts/colony.mjs memory scout

# Add a lesson
node scripts/colony.mjs memory scout add "Always check publication dates on research sources"

# Add to specific sections
node scripts/colony.mjs memory scout add "Use bullet points for clarity" --pattern
node scripts/colony.mjs memory scout add "Missed competitor X in analysis" --mistake
node scripts/colony.mjs memory scout add "Prefers markdown tables over lists" --pref

Shared Learnings

Cross-agent insights and lessons:

# View all shared learnings
node scripts/colony.mjs learn

# Add a learning
node scripts/colony.mjs learn add "validate-idea works better with 3 competitors max" --category process
node scripts/colony.mjs learn add "Always verify API rate limits early" --category technical --source run-abc123

Global Context

Shared context all agents can access:

# View global context
node scripts/colony.mjs context

# Set preferences
node scripts/colony.mjs context set preferences.codeStyle "TypeScript, functional"
node scripts/colony.mjs context set preferences.timezone "America/Chicago"

# Add active facts (temporary context)
node scripts/colony.mjs context add-fact "We're targeting enterprise customers"
node scripts/colony.mjs context add-fact "Launch deadline is Q2 2024"

# Add decisions
node scripts/colony.mjs context add-decision "Use Postgres over MySQL" --project "life-lunch"

# Add projects
node scripts/colony.mjs context add-project "life-lunch"

Retrospective

Review recent activity and generate insights:

node scripts/colony.mjs retro              # Last 7 days
node scripts/colony.mjs retro --days 14    # Last 14 days

Shows:

• Task completion summary
• Per-agent stats
• Failure patterns
• Suggested learnings

Available Processes

validate-idea

Validate a business idea end-to-end

• Stages: brainstorm → research → analyze → spec → estimate
• Checkpoints: after analyze
• Output: business-case.md

product-launch

End-to-end product launch

• Stages: research → spec → build → copy
• Checkpoints: after spec, after copy
• Output: market-brief.md, prd.md, code/, landing-copy.md

content-pipeline

Research, write, publish, promote content

• Stages: research → draft → review → publish → promote
• Checkpoints: review (human reviews draft)
• Output: research.md, draft.md, social-posts.md

bug-triage

Reproduce, fix, deploy bug fixes

• Stages: reproduce → fix → test → deploy
• Checkpoints: none (fast path)
• Output: bug-report.md, fix-summary.md

customer-research

Deep dive on a customer segment

• Stages: identify → pain-points → validate → synthesize
• Checkpoints: none
• Output: customer-profile.md, insights.md

landing-page

Create a full landing page

• Stages: strategy → copy → review → build
• Checkpoints: after copy review
• Output: strategy.md, copy.md, landing.html, landing.css

How Processes Work

1. Start - Process creates a run entry and spawns first stage agent
2. Execute - Each stage runs with inputs from previous stages
3. Checkpoint - If stage is a checkpoint, process pauses for approval
4. Continue - After approval, next stage runs
5. Complete - All stages done, outputs in colony/context/<run-id>/

Context Passing

• {context} in task templates is replaced with your --context value
• Stage outputs are saved to colony/context/<run-id>/<output-file>
• Next stage reads inputs from previous stage's output files
• Agent memory and global context are injected into prompts
• Full task history in tasks.json

Parallel Stages

Stages that share the same parallel_group run concurrently:

stages:
  - id: spec
    agent: forge
    inputs: [analysis.md]
    parallel_group: "final"  # Stages with same group run together
    
  - id: estimate
    agent: ledger
    inputs: [analysis.md]
    parallel_group: "final"  # Same group = parallel execution

When the process reaches a parallel group:

1. All consecutive stages with the same parallel_group are collected
2. All stages spawn concurrently (using Promise.all())
3. Process waits for ALL parallel stages to complete
4. If any stage fails, the entire group fails
5. Checkpoints work per-group (pause after all parallel stages complete)

Output shows parallel execution clearly:

═══ Parallel Group: final (2 stages) ═══
    → Stage 4: spec (forge)
    → Stage 5: estimate (ledger)

--- [PARALLEL] Stage 4/5: spec ---
--- [PARALLEL] Stage 5/5: estimate ---

═══ Parallel Group: final completed ═══

When to use parallel groups:

• Stages that read the same inputs (no dependencies on each other)
• Build + copy tasks (both depend on spec, not on each other)
• Multiple analyses of the same data
• Independent research threads

Processes with parallel stages:

• validate-idea: spec + estimate run in parallel
• product-launch: build + copy run in parallel

Notifications

Colony can send notifications when processes hit checkpoints, complete, or fail. Notifications use openclaw cron wake to alert you.

Configuration (colony/config.yaml):

notifications:
  enabled: true         # Master switch for all notifications
  on_checkpoint: true   # Notify when process pauses at checkpoint
  on_complete: true     # Notify when process finishes
  on_failure: true      # Notify when process/stage fails

Manage via CLI:

# View current config
node scripts/colony.mjs config

# Disable all notifications
node scripts/colony.mjs config set notifications.enabled false

# Enable only failure notifications
node scripts/colony.mjs config set notifications.on_checkpoint false
node scripts/colony.mjs config set notifications.on_complete false
node scripts/colony.mjs config set notifications.on_failure true

Notification examples:

• 🛑 Colony checkpoint: Process "validate-idea" paused after stage "analyze". To continue: colony approve abc123
• ✅ Colony complete: Process "validate-idea" finished in 120s. Run ID: abc123
• ❌ Colony failed: Process "validate-idea" failed at stage "research". Error: Agent timed out. Run ID: abc123

Checkpoints

Checkpoints pause the process for human review. Two ways to define:

1. In process checkpoints array (after that stage completes)
2. As a standalone stage with checkpoint: true (human-only review step)

File Structure

skills/colony/
├── SKILL.md              # This file
├── package.json          # Dependencies (js-yaml)
├── colony/
│   ├── agents.yaml       # Agent definitions
│   ├── processes.yaml    # Process definitions
│   ├── config.yaml       # Notification & behavior config
│   ├── tasks.json        # Task queue and history
│   ├── runs.json         # Process run tracking
│   ├── feedback.json     # Task feedback storage
│   ├── learnings.yaml    # Shared cross-agent learnings
│   ├── global-context.json  # Shared context for all agents
│   ├── audit/
│   │   ├── log.jsonl     # Append-only event log
│   │   ├── global.json   # Aggregate statistics
│   │   └── agents/       # Per-agent statistics
│   │       ├── scout.json
│   │       ├── pincer.json
│   │       └── ...
│   ├── memory/           # Per-agent persistent memory
│   │   ├── scout.md
│   │   ├── pincer.md
│   │   └── ...
│   └── context/          # Per-task and per-run outputs
│       └── <run-id>/
└── scripts/
    ├── colony.mjs         # Main CLI
    ├── colony-worker.mjs  # Background agent executor
    ├── agent-wrapper.mjs # Task lifecycle utilities
    ├── audit.mjs         # Audit system functions
    └── learning.mjs      # Learning system functions

Audit Events

The audit log tracks these events:

Customization

Add New Agents

Edit colony/agents.yaml:

agents:
  myagent:
    role: specialist
    description: >
      What this agent does...
    model: anthropic/claude-sonnet-4
    triggers:
      - keyword1
      - keyword2

After adding, create their memory file:

touch colony/memory/myagent.md

Add New Processes

Edit colony/processes.yaml:

processes:
  my-process:
    description: "What this process does"
    triggers: [keyword1, keyword2]
    stages:
      - id: stage1
        agent: scout
        task: "Do something with: {context}"
        outputs: [output1.md]
      - id: stage2
        agent: pincer
        task: "Next step based on previous"
        inputs: [output1.md]
        outputs: [output2.md]
    checkpoints: [stage1]  # Optional: pause after these stages

Integration

Works with OpenClaw's agent sessions.

Dispatch/Assign (async): Tasks are spawned in the background and the CLI returns immediately. Use colony status to monitor progress and colony results <task-id> to view output.

Process stages (blocking): Multi-stage processes run sequentially, waiting for each stage to complete before proceeding. This ensures proper data flow between stages and checkpoint handling.

Each agent receives:

• Their role description
• Lessons from their memory file
• Active facts from global context
• Project/preference context

Examples

Validate a Startup Idea

node scripts/colony.mjs process validate-idea \
  --context "Subscription box for home coffee brewing experiments"

Watch as it flows: brainstorm → research → analyze → (checkpoint) → spec → estimate

Write and Publish a Blog Post

node scripts/colony.mjs process content-pipeline \
  --context "Why RAG is eating traditional search"

Stages: research → draft → (human review) → publish → promote

Quick Research Task

node scripts/colony.mjs dispatch "compare Pinecone vs Weaviate vs Milvus"

Auto-routes to scout, returns comparison.

Track Performance

# After several tasks, check overall health
node scripts/colony.mjs audit

# Deep dive into a struggling agent
node scripts/colony.mjs audit agent pincer
node scripts/colony.mjs audit failures

# Add learnings from issues
node scripts/colony.mjs memory pincer add "Handle file not found errors gracefully" --mistake