workflow-orchestrator（工作流程协调器）

Workflow Orchestrator

Design, configure, and execute AI-driven development workflows (Ralph loops) that iterate until tasks complete successfully using configurable patterns, multi-agent coordination, and intelligent stop conditions.

When to use me

Use this skill when:

• Building complex features requiring multiple AI agent passes
• Implementing automated development workflows with verification steps
• Coordinating parallel AI agents for different aspects of a task
• Creating self-correcting loops that iterate until success criteria met
• Designing workflow patterns for team consistency
• Need to orchestrate multiple AI agents in sequence or parallel
• Want to implement "Ralph loops" (AI workflows that feed failures back into input)

What I do

1. Workflow Pattern Selection

• Analyze task requirements to recommend optimal workflow patterns
• Match complexity levels with appropriate workflow types
• Compare trade-offs between speed, cost, control, and quality
• Select from proven patterns (simple, verified, planned, parallel, custom)

2. Configuration Generation

• Generate YAML configurations for workflow execution
• Define agent roles and responsibilities (builder, verifier, planner, merger)
• Set success criteria and stop conditions for automated completion
• Configure iteration limits and escape hatches for safety
• Specify tool permissions and model selections per agent

3. Multi-Agent Coordination

• Sequence agents for linear workflows (build → verify → plan)
• Parallelize agents for concurrent execution
• Merge results from multiple agent perspectives
• Handle failures and retries with intelligent routing
• Implement steering patterns (linear, conditional, parallel, unblocking)

4. Ralph Loop Implementation

• Create feedback loops where failures inform subsequent attempts
• Implement stop hooks to detect completion promises (<promise>...</promise>)
• Build contextual pressure cookers that iterate toward success
• Manage iteration limits to prevent infinite loops
• Track progress and telemetry across workflow executions

5. Integration & Tooling

• OpenCode server integration for HTTP-based workflows
• SSE event streaming for real-time monitoring
• TUI/UI dashboards for workflow visualization
• Git hook integration for automated quality gates
• Cost tracking and budgeting across agent executions

Workflow Patterns

Pattern 1: Builder Only (Simple)

Single Agent → Build → [Success?] → Done
                     ↓ No
                     Retry (max iterations)

Best for: Simple tasks, bug fixes, small features Agents: 1 (builder) Control: Low, Cost: Low, Speed: High

Pattern 2: Build + Verify

Builder → [Success?] → Verifier → [Pass?] → Done
           ↓ No                 ↓ No
           Retry                Feedback → Builder

Best for: Quality-critical work, production code Agents: 2 (builder, verifier) Control: Medium, Cost: Medium, Speed: Medium

Pattern 3: Build + Verify + Plan

Planner → Builder → Verifier → [Pass?] → Done
                            ↓ No
                            Feedback → Planner

Best for: Complex features, architectural changes Agents: 3+ (planner, builder, verifier) Control: High, Cost: High, Speed: Low

Pattern 4: Multi-Agent Pipeline

[Agent A] → [Merge]
[Agent B] → [Results] → Done
[Agent C] →

Best for: Multiple perspectives, comprehensive solutions Agents: 3+ parallel agents Control: High, Cost: High, Speed: Medium

Pattern 5: Custom Workflow

User-defined sequence of agents with custom routing logic.

Configuration Examples

Simple Builder Loop (builder-only.yaml):

workflow:
  version: "1.0"
  settings:
    max_iterations: 20
    stop_on_success: true

  loop:
    agent: builder
    task: "Build the requested feature"
    stop_condition: "<promise>BUILD_COMPLETE</promise>"
    tools:
      read: true
      write: true
      edit: true
      bash: true

Build + Verify Loop:

workflow:
  version: "1.0"

  agents:
    builder:
      role: "Build the feature"
      stop_condition: "<promise>BUILT</promise>"

    verifier:
      role: "Verify quality and correctness"
      stop_condition: "<promise>VERIFIED</promise>"

  sequence:
    - agent: builder
    - agent: verifier
      condition: "builder.stop_condition met"

  feedback:
    verifier_failed: "builder"

Multi-Agent Parallel Workflow:

workflow:
  version: "1.0"

  agents:
    frontend_specialist:
      role: "Implement UI components"

    backend_specialist:
      role: "Implement API and business logic"

    qa_specialist:
      role: "Test implementation"

  parallel:
    - frontend_specialist
    - backend_specialist

  sequence:
    - parallel: [frontend_specialist, backend_specialist]
    - agent: qa_specialist

  merger:
    agent: lead_developer
    task: "Merge parallel work into cohesive solution"

Examples

# Analyze project and suggest workflow pattern
npm run workflow:analyze -- --task "Add user authentication"

# Generate workflow configuration
npm run workflow:generate -- --pattern build-verify --output auth-workflow.yaml

# Execute workflow
npm run workflow:execute -- --config auth-workflow.yaml --max-iterations 20

# Monitor running workflow
npm run workflow:monitor -- --workflow-id auth-123

# List available patterns
npm run workflow:patterns -- --list

# Create custom workflow from template
npm run workflow:custom -- --agents "planner,builder,verifier" --sequence "sequential"

Output format

Workflow Analysis Report:

Workflow Orchestrator Analysis
──────────────────────────────
Task: Add user authentication to React/Node.js app
Date: 2026-02-26

Recommended Pattern: Build + Verify + Plan
Rationale: Authentication requires security review, testing, and planning

Pattern Comparison:
┌─────────────────┬─────────┬─────────┬─────────┬──────────┐
│ Pattern         │ Control │ Cost   │ Speed   │ Quality  │
├─────────────────┼─────────┼─────────┼─────────┼──────────┤
│ Builder Only    │ Low     │ $      │ Fast    │ Medium   │
│ Build + Verify  │ Medium  │ $$     │ Medium  │ High     │
│ Build+Verify+Plan│ High    │ $$$    │ Slow    │ Highest  │
│ Multi-Agent     │ Highest │ $$$$   │ Medium  │ Highest  │
└─────────────────┴─────────┴─────────┴─────────┴──────────┘

Agents Required:
1. Planner (architect authentication flow)
2. Builder (implement components and API)
3. Verifier (security and testing specialist)

Estimated:
- Iterations: 3-5
- Cost: $15-25
- Time: 45-60 minutes
- Success Probability: 85%

Configuration Generated: auth-workflow.yaml

Workflow Configuration:

workflow:
  name: "user-authentication"
  version: "1.1"
  created: "2026-02-26T18:00:00Z"

  settings:
    max_iterations: 20
    stop_on_success: true
    escape_hatch: true
    cost_limit: 50.00
    time_limit_minutes: 120

  agents:
    planner:
      model: "anthropic/claude-sonnet-4-20250514"
      temperature: 0.3
      role: "Plan authentication architecture"
      tools: ["read", "write", "question"]
      stop_condition: "<promise>PLAN_COMPLETE</promise>"

    builder:
      model: "anthropic/claude-sonnet-4-20250514"
      temperature: 0.7
      role: "Implement authentication components"
      tools: ["read", "write", "edit", "bash", "webfetch"]
      stop_condition: "<promise>BUILD_COMPLETE</promise>"

    verifier:
      model: "anthropic/claude-sonnet-4-20250514"
      temperature: 0.2
      role: "Verify security and functionality"
      tools: ["read", "bash", "question"]
      stop_condition: "<promise>VERIFIED</promise>"

  sequence:
    - agent: planner
    - agent: builder
      condition: "planner.stop_condition met"
    - agent: verifier
      condition: "builder.stop_condition met"

  feedback:
    verifier_failed:
      target: planner
      message: "Verification failed, replan needed"

    builder_stuck:
      target: planner
      message: "Builder stuck, needs guidance"

  success_criteria:
    - "All tests pass"
    - "Security review passed"
    - "API endpoints documented"
    - "User stories implemented"

  telemetry:
    track_costs: true
    track_iterations: true
    track_duration: true
    dashboard_url: "http://localhost:3000/dashboard"

Workflow Execution Log:

Workflow Execution: user-authentication
─────────────────────────────────────
Status: Running
Start Time: 2026-02-26T18:00:00Z
Current Iteration: 3/20
Cost: $8.75
Duration: 25m 30s

Agent Timeline:
00:00:00 - Planner started
00:05:15 - Planner completed: <promise>PLAN_COMPLETE</promise>
00:05:30 - Builder started
00:15:45 - Builder: Created 12 files, 457 lines
00:20:30 - Builder completed: <promise>BUILD_COMPLETE</promise>
00:20:45 - Verifier started
00:25:30 - Verifier: Running security checks

Current Agent: Verifier
Progress: 75%
Estimated Completion: 00:35:00

Success Criteria Status:
✅ Tests passing: 42/42 tests
🔍 Security review: In progress
📚 Documentation: 80% complete
👤 User stories: 3/4 implemented

Next: Verifier completion → Workflow success

Notes

• Ralph loops work best with clear success criteria - vague goals lead to infinite loops
• Start simple - use Builder Only pattern for small tasks before scaling up
• Monitor costs - parallel agents and multiple iterations increase expense
• Implement escape hatches - always have manual stop conditions
• Use stop promises (<promise>...</promise>) for reliable completion detection
• Consider feedback mechanisms - failed verification should inform subsequent attempts
• Balance control vs. autonomy - more agents = more control but higher cost/complexity
• Test workflows with simple tasks before applying to critical work
• Document patterns for team consistency and knowledge sharing
• Review telemetry to optimize workflow configurations over time