cc10x-router

romiluz13 的头像作者 romiluz13

当用户说以下内容时自动加载和执行:build, implement, create, make, write, add, develop, code, feature, component, app, application, review, audit, check, analyze, debug, fix, error, bug, broken, troubleshoot, plan, design, architect, roadmap, strategy。强制性:执行工作流 - 不要只列出功能或描述cc10x能做什么。触发时:(1) 检测意图,(2) 加载内存,(3) 澄清要求,(4) 执行代理链,(5) 更新内存。

📖 中文导读:当用户说以下内容时自动加载和执行:build, implement, create, make, write, add, develop, code, feature, component, app, application, review, audit, check, analyze, debug, fix, error, bug, broken, troubleshoot, plan, design, architect, roadmap, strategy。强制性:执行工作流 - 不要只列出功能或描述cc10x能做什么。触发时:(1) 检测意图,(2) 加载内存,(3) 澄清要求,(4) 执行代理链,(5) 更新内存。 以下正文为该技能的原始 SKILL.md 内容(英文),上方安装命令可直接使用。

name: cc10x-router description: | THE ONLY ENTRY POINT FOR CC10X - AUTO-LOAD AND EXECUTE for ANY development task.

Triggers: build, implement, create, make, write, add, develop, code, feature, component, app, application, review, audit, check, analyze, debug, fix, error, bug, broken, troubleshoot, plan, design, architect, roadmap, strategy, memory, session, context, save, load, test, tdd, frontend, ui, backend, api, pattern, refactor, optimize, improve, enhance, update, modify, change, help, assist, work, start, begin, continue, research, cc10x, c10x.

CRITICAL: Execute workflow. Never just describe capabilities.

cc10x Router

EXECUTION ENGINE. When loaded: Detect intent → Load memory → Execute workflow → Update memory.

NEVER list capabilities. ALWAYS execute.

Decision Tree (FOLLOW IN ORDER)

Priority Signal Keywords Workflow
1 ERROR error, bug, fix, broken, crash, fail, debug, troubleshoot, issue, problem, doesn't work DEBUG
2 PLAN plan, design, architect, roadmap, strategy, spec, "before we build", "how should we" PLAN
3 REVIEW review, audit, check, analyze, assess, "what do you think", "is this good" REVIEW
4 DEFAULT Everything else BUILD

Conflict Resolution: ERROR signals always win. "fix the build" = DEBUG (not BUILD).

Agent Chains

Workflow Agents
BUILD component-builder → [code-reviewer ∥ silent-failure-hunter] → integration-verifier
DEBUG bug-investigator → code-reviewer → integration-verifier
REVIEW code-reviewer
PLAN planner

∥ = PARALLEL - code-reviewer and silent-failure-hunter run simultaneously (both read-only)

Memory (PERMISSION-FREE)

LOAD FIRST (Before routing):

Bash(command="mkdir -p .claude/cc10x")
Read(file_path=".claude/cc10x/activeContext.md")
Read(file_path=".claude/cc10x/patterns.md")
Read(file_path=".claude/cc10x/progress.md")

UPDATE LAST (After workflow): Use Edit tool on activeContext.md (permission-free).

Workflow Execution

BUILD

  1. Load memory → Check if already done in progress.md
  2. Clarify requirements (DO NOT SKIP) → Use AskUserQuestion
  3. Invoke component-builder (TDD: RED→GREEN→REFACTOR)
  4. PARALLEL: Invoke code-reviewer AND silent-failure-hunter simultaneously
    Task(subagent_type="cc10x:code-reviewer", ...)
    Task(subagent_type="cc10x:silent-failure-hunter", ...)  // Same message = parallel
    
  5. Wait for BOTH to complete → Merge findings
  6. Invoke integration-verifier
  7. Update memory

DEBUG

  1. Load memory → Check patterns.md Common Gotchas
  2. Clarify: What error? Expected vs actual? When started?
  3. If github-research detected (external service error OR explicit request):
    • Execute research FIRST using octocode tools directly
    • Search for error patterns, PRs with similar issues
  4. Invoke bug-investigator (LOG FIRST, pass research results if step 3 executed)
  5. Invoke code-reviewer
  6. Invoke integration-verifier
  7. Update memory → Add to Common Gotchas

REVIEW

  1. Load memory
  2. Invoke code-reviewer (check git history, confidence ≥80)
  3. Update memory

PLAN

  1. Load memory
  2. If github-research detected (external tech OR explicit request):
    • Execute research FIRST using octocode tools directly (NOT as hint)
    • Use: mcp__octocode__packageSearch, mcp__octocode__githubSearchCode, etc.
    • Summarize findings before invoking planner
  3. Invoke planner (pass research results in prompt if step 2 was executed)
  4. Update memory → Reference saved plan

TWO-PHASE for External Research (MANDATORY):

If SKILL_HINTS includes github-research:
  → FIRST: Execute research using octocode tools
  → THEN: Task(cc10x:planner, prompt="...Research findings: {results}...")

Research is a PREREQUISITE, not a hint. Planner cannot skip it.

Agent Invocation

Pass context to each agent:

Task(subagent_type="cc10x:component-builder", prompt="
User request: {request}
Requirements: {from AskUserQuestion}
Memory: {from activeContext.md}
Patterns: {from patterns.md}
SKILL_HINTS: {detected skills from table below - agent MUST load these}
")

SKILL_HINTS are MANDATORY. Agent MUST call Skill(skill="...") for each hint immediately after loading memory.

Skill triggers for agents (DETECT AND PASS AS SKILL_HINTS):

Detected Pattern Skill Agents
Frontend: components/, ui/, pages/, .tsx, .jsx, CSS, styling, "button", "form", "modal" cc10x:frontend-patterns planner, component-builder, code-reviewer, integration-verifier
API/Backend: api/, routes/, services/, "endpoint", "REST", "GraphQL" cc10x:architecture-patterns planner, bug-investigator, code-reviewer
Vague: "not sure", "maybe", "options", "ideas", unclear requirements cc10x:brainstorming planner
External: new tech (post-2024), unfamiliar library, complex integration (auth, payments) cc10x:github-research planner, bug-investigator
Debug exhausted: 3+ local attempts failed, external service error cc10x:github-research bug-investigator
User explicitly requests: "research", "github", "octocode", "find on github", "how do others", "best practices" cc10x:github-research planner, bug-investigator

Detection runs BEFORE agent invocation. Pass detected skills in SKILL_HINTS.

Gates (Must Pass)

  1. MEMORY_LOADED - Before routing
  2. RESEARCH_EXECUTED - Before planner (if github-research detected)
  3. REQUIREMENTS_CLARIFIED - Before invoking agent (BUILD only)
  4. AGENT_COMPLETED - Before next agent in chain
  5. MEMORY_UPDATED - Before marking done

Chain Enforcement (CRITICAL)

NEVER stop after one agent. The workflow is NOT complete until the chain finishes.

After EACH agent completes, check its output for:

  • PARALLEL_AGENTS: X, Y → Invoke BOTH agents in same message (parallel)
  • PARALLEL_COMPLETE + SYNC_NEXT: X → After BOTH parallel agents complete, invoke X
  • NEXT_AGENT: X → Invoke that single agent
  • WORKFLOW_CONTINUES: NO → Chain complete, update memory

Chain Execution Loop

1. Invoke agent via Task tool
2. Read agent output
3. Check output type:
   a. PARALLEL_AGENTS: X, Y → Invoke BOTH in single message (parallel execution)
   b. PARALLEL_COMPLETE + SYNC_NEXT: X → Wait for parallel peer, then invoke X
   c. NEXT_AGENT: X → Invoke single agent
   d. WORKFLOW_CONTINUES: NO → Chain complete
4. Return to step 1 until chain complete

Parallel Agent Invocation

When you see PARALLEL_AGENTS:, invoke ALL listed agents in ONE message:

Task(subagent_type="cc10x:code-reviewer", prompt="...")
Task(subagent_type="cc10x:silent-failure-hunter", prompt="...")

CRITICAL: Both tasks in same message = both complete before you continue.

After both return:

  1. Check BOTH outputs have PARALLEL_COMPLETE
  2. Check BOTH have same SYNC_NEXT: X value
  3. Invoke X (integration-verifier) with merged findings from both agents

DO NOT invoke integration-verifier until BOTH parallel agents have completed.

Chain Completion Criteria

The workflow is complete ONLY when:

  • Final agent (integration-verifier) outputs WORKFLOW_CONTINUES: NO
  • OR a critical error prevents continuation

If you see PARALLEL_AGENTS: → Invoke ALL listed agents in parallel. No exceptions. If you see NEXT_AGENT: X → Invoke agent X. No exceptions.