skill-iter-tune

name	skill-iter-tune
disable-model-invocation	true
description	Iterative skill tuning via execute-evaluate-improve feedback loop. Uses maestro delegate Claude to execute skill, Agy to evaluate quality, and Agent to apply improvements. Iterates until quality threshold or max iterations. Triggers on "skill iter tune", "iterative skill tuning", "tune skill".
allowed-tools	["Bash","Edit","Glob","Grep","Read","Write","followup_task","interrupt_agent","list_agents","request_user_input","send_message","spawn_agent","spawn_agents_on_csv","update_plan","wait_agent"]
session-mode	run
version	0.5.83
contract	{"discovery":"self-described","consumes":[],"produces":[],"gates":{"entry":[],"exit":[]}}

Plan tracking: codex 无 TaskCreate/TaskUpdate/TodoWrite 任务板。进度清单用 update_plan({ explanation?, plan: [{ step, status }] }) 维护（整体提交步骤数组，status: pending | in_progress | completed），权威状态始终在 session 工件中；依赖/认领（addBlockedBy/owner）是工件字段，不是工具参数。

<required_reading> @/.maestro/workflows/run-mode.md @/.maestro/workflows/codex-run-mode.md </required_reading>

Skill Iter Tune

Iterative skill refinement through execute-evaluate-improve feedback loops. Each iteration runs the skill via Claude, evaluates output via Agy, and applies improvements via Agent.

Architecture Overview

┌──────────────────────────────────────────────────────────────────────────┐
│  Skill Iter Tune Orchestrator (SKILL.md)                                 │
│  → Parse input → Setup workspace → Iteration Loop → Final Report         │
└────────────────────────────┬─────────────────────────────────────────────┘
                             │
         ┌───────────────────┼───────────────────────────────────┐
         ↓                   ↓                                   ↓
    ┌──────────┐      ┌─────────────────────────────┐     ┌──────────┐
    │ Phase 1  │      │  Iteration Loop (2→3→4)     │     │ Phase 5  │
    │ Setup    │      │  ┌─────┐  ┌─────┐  ┌─────┐ │     │ Report   │
    │          │─────→│  │ P2  │→ │ P3  │→ │ P4  │ │────→│          │
    │ Backup + │      │  │Exec │  │Eval │  │Impr │ │     │ History  │
    │ Init     │      │  └─────┘  └─────┘  └─────┘ │     │ Summary  │
    └──────────┘      │       ↑               │     │     └──────────┘
                      │       └───────────────┘     │
                      │    (if score < threshold    │
                      │     AND iter < max)         │
                      └─────────────────────────────┘

Chain Mode Extension

Chain Mode (execution_mode === "chain"):

Phase 2 runs per-skill in chain_order:
  Skill A → maestro delegate → artifacts/skill-A/
       ↓ (artifacts as input)
  Skill B → maestro delegate → artifacts/skill-B/
       ↓ (artifacts as input)
  Skill C → maestro delegate → artifacts/skill-C/

Phase 3 evaluates entire chain output + per-skill scores
Phase 4 improves weakest skill(s) in chain

Key Design Principles

1. Iteration Loop: Phases 2-3-4 repeat until quality threshold, max iterations, or convergence
2. Two-Tool Pipeline: Claude (write/execute) + Agy (analyze/evaluate) = complementary perspectives
3. Pure Orchestrator: SKILL.md coordinates only — execution detail lives in phase files
4. Progressive Phase Loading: Phase docs read only when that phase executes
5. Skill Versioning: Each iteration snapshots skill state before execution
6. Convergence Detection: Stop early if score stalls (no improvement in 2 consecutive iterations)

Interactive Preference Collection

// ★ Auto mode detection
const autoYes = /\b(-y|--yes)\b/.test($ARGUMENTS)

if (autoYes) {
  workflowPreferences = {
    autoYes: true,
    maxIterations: 5,
    qualityThreshold: 80,
    executionMode: 'single'
  }
} else {
  const prefResponse = request_user_input({
    questions: [
      {
        question: "选择迭代调优配置：",
        header: "Tune Config",
        multiSelect: false,
        options: [
          { label: "Quick (3 iter, 70)", description: "快速迭代，适合小幅改进" },
          { label: "Standard (5 iter, 80) (Recommended)", description: "平衡方案，适合多数场景" },
          { label: "Thorough (8 iter, 90)", description: "深度优化，适合生产级 skill" }
        ]
      }
    ]
  })

  const configMap = {
    "Quick": { maxIterations: 3, qualityThreshold: 70 },
    "Standard": { maxIterations: 5, qualityThreshold: 80 },
    "Thorough": { maxIterations: 8, :  }
  }
   selected = .(configMap).(
    prefResponse[].(k)
  ) || 
  workflowPreferences = { : , ...configMap[selected] }

  
   modeResponse = ({
    : [{
      : ,
      : ,
      : ,
      : [
        { : , :  },
        { : , :  }
      ]
    }]
  });
  workflowPreferences. = modeResponse[].()
    ?  : ;
}

Input Processing

$ARGUMENTS → Parse:
  ├─ Skill path(s): first arg, comma-separated for multiple
  │   e.g., ".claude/skills/my-skill" or "my-skill" (auto-prefixed)
  │   Chain mode: order preserved as chain_order
  ├─ Test scenario: --scenario "description" or remaining text
  └─ Flags: --max-iterations=N, --threshold=N, -y/--yes

Execution Flow

⚠️ COMPACT DIRECTIVE: Context compression MUST check update_plan phase status. The phase currently marked in_progress is the active execution phase — preserve its FULL content. Only compress phases marked completed or pending.

Phase 1: Setup (one-time)

Read and execute: Ref: phases/01-setup.md

• Parse skill paths, validate existence
• Create workspace at {run_dir}/outputs/skill-iter-tune-{ts}/
• Backup original skill files
• Initialize iteration-state.json

Output: workDir, targetSkills[], testScenario, initialized state

Iteration Loop

// Orchestrator iteration loop
while (true) {
  // Increment iteration
  state.current_iteration++;
  state.iterations.push({
    round: state.current_iteration,
    status: 'pending',
    execution: null,
    evaluation: null,
    improvement: null
  });

  // Update update_plan
  update_plan(iterationTask, {
    subject: `Iteration ${state.current_iteration}/${state.max_iterations}`,
    status: 'in_progress',
    activeForm: `Running iteration ${state.current_iteration}`
  });

  // === Phase 2: Execute ===
  // Read: phases/02-execute.md
  // Single mode: one maestro delegate call for all skills
  // Chain mode: sequential maestro delegate per skill in chain_order, passing artifacts
  // Snapshot skill → construct prompt → maestro delegate --to claude --mode write
  // Collect artifacts

  // === Phase 3: Evaluate ===
  // Read: phases/03-evaluate.md
  // Construct eval prompt → maestro delegate --to agy --mode analysis
  // Parse score → write iteration-N-eval.md → check termination

  // Check termination
  if (shouldTerminate(state)) {
    break;  // → Phase 5
  }

  // === Phase 4: Improve ===
  
  

  
  
}

Phase 2: Execute Skill (per iteration)

Read and execute: Ref: phases/02-execute.md

• Snapshot skill → iteration-{N}/skill-snapshot/
• Build execution prompt from skill content + test scenario
• Execute: maestro delegate "..." --to claude --mode write --cd "${iterDir}/artifacts"
• Collect artifacts

Phase 3: Evaluate Quality (per iteration)

Read and execute: Ref: phases/03-evaluate.md

• Build evaluation prompt with skill + artifacts + criteria + history
• Execute: maestro delegate "..." --to agy --mode analysis
• Parse 5-dimension score (Clarity, Completeness, Correctness, Effectiveness, Efficiency)
• Write iteration-{N}-eval.md
• Check termination: score >= threshold | iter >= max | convergence | error limit

Phase 4: Apply Improvements (per iteration, skipped on termination)

Read and execute: Ref: phases/04-improve.md

• Read evaluation suggestions
• Launch general-purpose Agent to apply changes
• Write iteration-{N}-changes.md
• Update state

Phase 5: Final Report (one-time)

Read and execute: Ref: phases/05-report.md

• Generate comprehensive report with score progression table
• Write final-report.md
• Display summary to user

Phase Reference Documents (read on-demand when phase executes):

Phase	Document	Purpose	Compact
1	phases/01-setup.md	Initialize workspace and state	update_plan 驱动
2	phases/02-execute.md	Execute skill via maestro delegate Claude	update_plan 驱动 + 🔄 sentinel
3	phases/03-evaluate.md	Evaluate via maestro delegate Agy	update_plan 驱动 + 🔄 sentinel
4	phases/04-improve.md	Apply improvements via Agent	update_plan 驱动 + 🔄 sentinel
5	phases/05-report.md	Generate final report	update_plan 驱动

Compact Rules:

1. update_plan in_progress → 保留完整内容，禁止压缩
2. update_plan completed → 可压缩为摘要
3. 🔄 sentinel fallback → 若 compact 后仅存 sentinel 而无完整 Step 协议，立即 Read() 恢复

Core Rules

1. Start Immediately: First action is preference collection → Phase 1 setup
2. Progressive Loading: Read phase doc ONLY when that phase is about to execute
3. Snapshot Before Execute: Always snapshot skill state before each iteration
4. Background CLI: maestro delegate runs in background, wait for hook callback before proceeding
5. Parse Every Output: Extract structured JSON from CLI outputs for state updates
6. DO NOT STOP: Continuous iteration until termination condition met
7. Single State Source: iteration-state.json is the only source of truth

Data Flow

User Input (skill paths + test scenario)
    ↓ (+ execution_mode + chain_order if chain mode)
    ↓
Phase 1: Setup
    ↓ workDir, targetSkills[], testScenario, iteration-state.json
    ↓
┌─→ Phase 2: Execute (maestro delegate claude)
│   ↓ artifacts/ (skill execution output)
│   ↓
│   Phase 3: Evaluate (maestro delegate agy)
│   ↓ score, dimensions[], suggestions[], iteration-N-eval.md
│   ↓
│   [Terminate?]─── YES ──→ Phase 5: Report → final-report.md
│   ↓ NO
│   ↓
│   Phase 4: Improve (Agent)
│   ↓ modified skill files, iteration-N-changes.md
│   ↓
└───┘ next iteration

update_plan Pattern

// Initial state
update_plan({ subject: "Phase 1: Setup workspace", activeForm: "Setting up workspace" })
update_plan({ subject: "Iteration Loop", activeForm: "Running iterations" })
update_plan({ subject: "Phase 5: Final Report", activeForm: "Generating report" })

// Chain mode: create per-skill tracking tasks
if (state.execution_mode === 'chain') {
  for (const skillName of state.chain_order) {
    update_plan({
      subject: `Chain: ${skillName}`,
      activeForm: `Tracking ${skillName}`,
      description: `Skill chain member position ${state.chain_order.indexOf(skillName) + 1}`
    })
  }
}

// During iteration N
// Single mode: one score per iteration (existing behavior)
// Chain mode: per-skill status updates
if (state.execution_mode === 'chain') {
  // After each skill executes in Phase 2:
  update_plan(chainSkillTask, {
    subject: `Chain: ${skillName} — Iter ${N} executed`,
    activeForm: 
  })
  
  (chainSkillTask, {
    : ,
    : 
  })
}  {
  
  ({
    : ,
    : ,
    : 
  })
}


(iterLoop, {
  : ,
  : 
})

Termination Logic

function shouldTerminate(state) {
  // 1. Quality threshold met
  if (state.latest_score >= state.quality_threshold) {
    return { terminate: true, reason: 'quality_threshold_met' };
  }
  // 2. Max iterations reached
  if (state.current_iteration >= state.max_iterations) {
    return { terminate: true, reason: 'max_iterations_reached' };
  }
  // 3. Convergence: ≤2 points improvement over last 2 iterations
  if (state.score_trend.length >= 3) {
    const last3 = state.score_trend.slice(-3);
    if (last3[2] - last3[0] <= 2) {
      state.converged = true;
      return { terminate: true, reason: 'convergence_detected' };
    }
  }
  // 4. Error limit
  if (state.error_count >= state.max_errors) {
    return { terminate: true, reason: 'error_limit_reached' };
  }
   { :  };
}

Error Handling

Phase	Error	Recovery
2: Execute	CLI timeout/crash	Retry once with simplified prompt, then skip
3: Evaluate	CLI fails	Retry once, then use score 50 with warning
3: Evaluate	JSON parse fails	Extract score heuristically, save raw output
4: Improve	Agent fails	Rollback from iteration-{N}/skill-snapshot/
Any	3+ consecutive errors	Terminate with error report

Error Budget: Each phase gets 1 retry. 3 consecutive failed iterations triggers termination.

Coordinator Checklist

Pre-Phase Actions

• Read iteration-state.json for current state
• Verify workspace directory exists
• Check error count hasn't exceeded limit

Per-Iteration Actions

• Increment current_iteration in state
• Create iteration-{N} subdirectory
• Update update_plan with iteration status
• After Phase 3: check termination before Phase 4
• After Phase 4: write state, proceed to next iteration

Post-Workflow Actions

• Execute Phase 5 (Report)
• Display final summary to user
• Update all update_plan tasks to completed