| name | speccrew-system-designer-orchestration |
| version | 1.0.0 |
| description | System Designer 的核心编排技能,负责读取已确认的 Feature Spec 和 API Contract 文档,加载技术知识库,评估框架需求,调度各平台详细设计技能生成系统设计文档。支持 web、mobile、desktop 平台。 |
| tools | Read, Write, Glob, Grep, Bash, Agent |
⚠️ MANDATORY EXECUTION PROTOCOL — READ BEFORE EXECUTING ANY BLOCK
Step 1: Load XML workflow specification: speccrew-workspace/docs/rules/agentflow-spec.md — this defines all block types and action-to-tool mappings
Step 2: Execute this SKILL.md's XML workflow block by block in document order. For EVERY block, you MUST follow this 3-step cycle:
📋 Block [ID] (action=[action]) — [desc]
🔧 Tool: [which IDE tool to call]
✅ Result: [output or status]
Action-to-tool mapping:
action="run-skill" → Invoke via Skill tool (pass the <field name="skill"> value EXACTLY)
action="run-script" → Execute via Terminal tool (pass the <field name="command"> value EXACTLY)
action="dispatch-to-worker" → Use Agent tool to create a new speccrew-task-worker agent, passing skill_path and context
action="read-file" → Read via Read tool
action="log" → Output message directly
action="confirm" → Present to user and wait for response
Step 3: Execute ALL blocks sequentially without pausing (only stop at explicit <event action="confirm"> blocks)
System Designer Orchestration
System Designer 的核心编排技能,负责:
- Stage Gate - 验证 Feature Design 阶段已确认
- Preparation - 加载 Feature Registry,验证文件存在性
- Resource Verification - 验证技术知识库可用性
- Framework Evaluation - 派发单个 worker agent 执行框架评估技能,等待 worker 完成并验证 framework-evaluation.md 生成后进入下一阶段(HARD STOP 等待用户确认)
- Design Overview - 生成 DESIGN-OVERVIEW.md
- Platform Dispatch - 分发各平台设计任务给 Worker
- Joint Confirmation - 联合确认所有设计文档
Invocation Method
CRITICAL: This skill is loaded directly by System Designer Agent — do NOT invoke via Worker Agent.
<block type="task" action="run-skill" desc="System Designer orchestration workflow">
<field name="skill">speccrew-system-designer-orchestration</field>
</block>
Input Parameters
| Parameter | Type | Required | Description |
|---|
workspace_path | string | Yes | speccrew-workspace root directory path |
iterations_dir | string | Yes | iterations directory path |
update_progress_script | string | Yes | Path to update-progress.js script |
current_iteration | string | No | Current active iteration identifier |
Output
status - Execution status (success / partial / failed)
output_files - List of generated design documents
summary - Execution summary
next_steps - Suggested next actions
AgentFlow Definition
CONTINUOUS EXECUTION RULES
This skill MUST execute tasks continuously without unnecessary interruptions.
FORBIDDEN Interruptions
- DO NOT ask user "Should I continue?" after completing a subtask
- DO NOT suggest "Let me split this into batches" or "Let's do this in parts"
- DO NOT pause to list what you plan to do next — just do it
- DO NOT ask for confirmation before generating output files
- DO NOT warn about "large number of files" — proceed with generation
- DO NOT offer "Should I proceed with the remaining items?"
When to Pause (ONLY these cases)
- CHECKPOINT gates defined in workflow (user confirmation required by design)
- Ambiguous requirements that genuinely need clarification
- Unrecoverable errors that prevent further progress
- Security-sensitive operations (e.g., deleting existing files)
Phase 3 Execution Method — Worker Agent Dispatch
HOW TO DISPATCH: When executing Phase 3 block P3-B1 (action="dispatch-to-worker"):
- Use the Agent tool to create a new
speccrew-task-worker agent
- Pass
skill_path to the worker: ${workspace_path}/.speccrew/skills/speccrew-sd-framework-evaluate/SKILL.md
- Pass context parameters: workspace_path, iteration_path, feature_spec_paths, api_contract_paths, techs_knowledge_paths, output_path
- Wait for the worker agent to complete and return results
- After worker completes, verify framework-evaluation.md exists at output_path
- Then proceed to Phase 3.5 HARD STOP for user confirmation
CRITICAL: The Agent tool creates a NEW agent session — this is completely different from the Skill tool which executes inline.
HARD STOP Checkpoints
This workflow has mandatory HARD STOP checkpoints at:
- Phase 3.5: Framework evaluation confirmation (user MUST approve)
- Phase 6.1: Joint design confirmation (user MUST approve all designs)
DO NOT proceed past these checkpoints without explicit user confirmation.
Must Do
- READ workflow.agentflow.xml FIRST — Execute blocks in document order
- Use Agent tool for dispatch-to-worker blocks — Agent tool creates a new speccrew-task-worker agent session
- Pass skill_path explicitly to worker — Worker cannot find skills via glob on first run
- Wait for worker completion before verifying output — Do not proceed until worker returns
- Update progress via update-progress.js script — Use run-script commands exactly as defined in XML
Must Not Do
- DO NOT use Skill tool for Phase 3 framework evaluation — Skill tool executes inline, Agent tool creates a worker
- DO NOT read feature spec files yourself — Workers read them
- DO NOT read tech knowledge files yourself — Workers read them
- DO NOT generate framework-evaluation.md yourself — Only workers generate it
- DO NOT create scripts for batch analysis — Workers handle this via their own skill
- DO NOT fallback to inline execution if worker fails — ABORT instead
- DO NOT skip reading workflow.agentflow.xml — XML is the execution authority