feat: 添加令牌消耗诊断功能，优化输出和状态管理

2026-02-05 01:50:27 +08:00 · 2026-01-14 21:40:00 +08:00
parent 266f6f11ec
commit f9c1216eec
7 changed files with 540 additions and 60 deletions
--- a/.claude/skills/skill-tuning/SKILL.md
+++ b/.claude/skills/skill-tuning/SKILL.md
@@ -70,6 +70,7 @@ Based on comprehensive analysis, skill-tuning addresses **core skill issues** an
 | **P2** | Agent Coordination | Fragile call chains, merge complexity | error_wrapping, result_validation |
 | **P3** | Context Explosion | Token accumulation, multi-turn bloat | sliding_window, context_summarization |
 | **P4** | Long-tail Forgetting | Early constraint loss | constraint_injection, checkpoint_restore |
+| **P5** | Token Consumption | Verbose prompts, excessive state, redundant I/O | prompt_compression, lazy_loading, output_minimization |

 ### General Optimization Areas (按需分析 via Gemini CLI)

@@ -202,47 +203,27 @@ RULES: $(cat ~/.claude/workflows/cli-templates/protocols/analysis-protocol.md) |
 │  → Initialize state.json with target skill info                             │
 │  → Create backup of target skill files                                       │
 ├─────────────────────────────────────────────────────────────────────────────┤
-│  action-analyze-requirements: Requirement Analysis (NEW)                     │
+│  action-analyze-requirements: Requirement Analysis                           │
 │  → Phase 1: 维度拆解 (Gemini CLI) - 单一描述 → 多个关注维度                   │
 │  → Phase 2: Spec 匹配 - 每个维度 → taxonomy + strategy                       │
 │  → Phase 3: 覆盖度评估 - 以"有修复策略"为满足标准                             │
 │  → Phase 4: 歧义检测 - 识别多义性描述，必要时请求澄清                         │
-│  → Output: requirement-analysis.json, 自动优化 focus_areas                   │
+│  → Output: state.json (requirement_analysis field)                           │
 ├─────────────────────────────────────────────────────────────────────────────┤
-│  action-diagnose-context: Context Explosion Analysis                         │
-│  → Scan for token accumulation patterns                                      │
-│  → Detect multi-turn dialogue growth                                         │
-│  → Output: context-diagnosis.json                                            │
-├─────────────────────────────────────────────────────────────────────────────┤
-│  action-diagnose-memory: Long-tail Forgetting Analysis                       │
-│  → Trace constraint propagation through phases                               │
-│  → Detect early instruction loss                                             │
-│  → Output: memory-diagnosis.json                                             │
-├─────────────────────────────────────────────────────────────────────────────┤
-│  action-diagnose-dataflow: Data Flow Analysis                                │
-│  → Map state transitions between phases                                      │
-│  → Detect format inconsistencies                                             │
-│  → Output: dataflow-diagnosis.json                                           │
-├─────────────────────────────────────────────────────────────────────────────┤
-│  action-diagnose-agent: Agent Coordination Analysis                          │
-│  → Analyze agent call patterns                                               │
-│  → Detect result passing issues                                              │
-│  → Output: agent-diagnosis.json                                              │
-├─────────────────────────────────────────────────────────────────────────────┤
-│  action-diagnose-docs: Documentation Structure Analysis (Optional)           │
-│  → Detect definition duplicates across files                                 │
-│  → Detect conflicting definitions                                            │
-│  → Output: docs-diagnosis.json                                               │
+│  action-diagnose-*: Diagnosis Actions (context/memory/dataflow/agent/docs/   │
+│                      token_consumption)                                      │
+│  → Execute pattern-based detection for each category                         │
+│  → Output: state.json (diagnosis.{category} field)                           │
 ├─────────────────────────────────────────────────────────────────────────────┤
 │  action-generate-report: Consolidated Report                                 │
-│  → Merge all diagnosis results                                               │
+│  → Generate markdown summary from state.diagnosis                            │
 │  → Prioritize issues by severity                                             │
-│  → Output: tuning-report.md                                                  │
+│  → Output: state.json (final_report field)                                   │
 ├─────────────────────────────────────────────────────────────────────────────┤
 │  action-propose-fixes: Fix Proposal Generation                               │
 │  → Generate fix strategies for each issue                                    │
 │  → Create implementation plan                                                │
-│  → Output: fix-proposals.json                                                │
+│  → Output: state.json (proposed_fixes field)                                 │
 ├─────────────────────────────────────────────────────────────────────────────┤
 │  action-apply-fix: Apply Selected Fix                                        │
 │  → User selects fix to apply                                                 │
@@ -255,9 +236,9 @@ RULES: $(cat ~/.claude/workflows/cli-templates/protocols/analysis-protocol.md) |
 │  → Update iteration count                                                    │
 ├─────────────────────────────────────────────────────────────────────────────┤
 │  action-complete: Finalization                                               │
-│  → Generate final report                                                     │
-│  → Cleanup temporary files                                                   │
-│  → Output: tuning-summary.md                                                 │
+│  → Set status='completed'                                                    │
+│  → Final report already in state.json (final_report field)                   │
+│  → Output: state.json (final)                                                │
 └─────────────────────────────────────────────────────────────────────────────┘
 ```

@@ -267,31 +248,25 @@ RULES: $(cat ~/.claude/workflows/cli-templates/protocols/analysis-protocol.md) |
 const timestamp = new Date().toISOString().slice(0,19).replace(/[-:T]/g, '');
 const workDir = `.workflow/.scratchpad/skill-tuning-${timestamp}`;

-Bash(`mkdir -p "${workDir}/diagnosis"`);
+// Simplified: Only backups dir needed, diagnosis results go into state.json
 Bash(`mkdir -p "${workDir}/backups"`);
-Bash(`mkdir -p "${workDir}/fixes"`);
 ```

 ## Output Structure

 ```
 .workflow/.scratchpad/skill-tuning-{timestamp}/
-├── state.json                      # Session state (orchestrator-managed)
-├── diagnosis/
-│   ├── context-diagnosis.json      # Context explosion analysis
-│   ├── memory-diagnosis.json       # Long-tail forgetting analysis
-│   ├── dataflow-diagnosis.json     # Data flow analysis
-│   ├── agent-diagnosis.json        # Agent coordination analysis
-│   └── docs-diagnosis.json         # Documentation structure analysis (optional)
-├── backups/
-│   └── {skill-name}-backup/        # Original skill files backup
-├── fixes/
-│   ├── fix-proposals.json          # Proposed fixes
-│   └── applied-fixes.json          # Applied fix history
-├── tuning-report.md                # Consolidated diagnosis report
-└── tuning-summary.md               # Final summary
+├── state.json                      # Single source of truth (all results consolidated)
+│   ├── diagnosis.*                 # All diagnosis results embedded
+│   ├── issues[]                    # Found issues
+│   ├── proposed_fixes[]            # Fix proposals
+│   └── final_report                # Markdown summary (on completion)
+└── backups/
+    └── {skill-name}-backup/        # Original skill files backup
 ```

+> **Token Optimization**: All outputs consolidated into state.json. No separate diagnosis files or report files.
+
 ## State Schema

 详细状态结构定义请参阅 [phases/state-schema.md](phases/state-schema.md)。
@@ -316,6 +291,7 @@ Bash(`mkdir -p "${workDir}/fixes"`);
 | [phases/actions/action-diagnose-dataflow.md](phases/actions/action-diagnose-dataflow.md) | Data flow diagnosis |
 | [phases/actions/action-diagnose-agent.md](phases/actions/action-diagnose-agent.md) | Agent coordination diagnosis |
 | [phases/actions/action-diagnose-docs.md](phases/actions/action-diagnose-docs.md) | Documentation structure diagnosis |
+| [phases/actions/action-diagnose-token-consumption.md](phases/actions/action-diagnose-token-consumption.md) | Token consumption diagnosis |
 | [phases/actions/action-generate-report.md](phases/actions/action-generate-report.md) | Report generation |
 | [phases/actions/action-propose-fixes.md](phases/actions/action-propose-fixes.md) | Fix proposal |
 | [phases/actions/action-apply-fix.md](phases/actions/action-apply-fix.md) | Fix application |
--- a/.claude/skills/skill-tuning/phases/actions/action-diagnose-token-consumption.md
+++ b/.claude/skills/skill-tuning/phases/actions/action-diagnose-token-consumption.md
@@ -0,0 +1,200 @@
+# Action: Diagnose Token Consumption
+
+Analyze target skill for token consumption inefficiencies and output optimization opportunities.
+
+## Purpose
+
+Detect patterns that cause excessive token usage:
+- Verbose prompts without compression
+- Large state objects with unnecessary fields
+- Full content passing instead of references
+- Unbounded arrays without sliding windows
+- Redundant file I/O (write-then-read patterns)
+
+## Detection Patterns
+
+| Pattern ID | Name | Detection Logic | Severity |
+|------------|------|-----------------|----------|
+| TKN-001 | Verbose Prompts | Prompt files > 4KB or high static/variable ratio | medium |
+| TKN-002 | Excessive State Fields | State schema > 15 top-level keys | medium |
+| TKN-003 | Full Content Passing | `Read()` result embedded directly in prompt | high |
+| TKN-004 | Unbounded Arrays | `.push`/`concat` without `.slice(-N)` | high |
+| TKN-005 | Redundant Write→Read | `Write(file)` followed by `Read(file)` | medium |
+
+## Execution Steps
+
+```javascript
+async function diagnoseTokenConsumption(state, workDir) {
+  const issues = [];
+  const evidence = [];
+  const skillPath = state.target_skill.path;
+
+  // 1. Scan for verbose prompts (TKN-001)
+  const mdFiles = Glob(`${skillPath}/**/*.md`);
+  for (const file of mdFiles) {
+    const content = Read(file);
+    if (content.length > 4000) {
+      evidence.push({
+        file: file,
+        pattern: 'TKN-001',
+        severity: 'medium',
+        context: `File size: ${content.length} chars (threshold: 4000)`
+      });
+    }
+  }
+
+  // 2. Check state schema field count (TKN-002)
+  const stateSchema = Glob(`${skillPath}/**/state-schema.md`)[0];
+  if (stateSchema) {
+    const schemaContent = Read(stateSchema);
+    const fieldMatches = schemaContent.match(/^\s*\w+:/gm) || [];
+    if (fieldMatches.length > 15) {
+      evidence.push({
+        file: stateSchema,
+        pattern: 'TKN-002',
+        severity: 'medium',
+        context: `State has ${fieldMatches.length} fields (threshold: 15)`
+      });
+    }
+  }
+
+  // 3. Detect full content passing (TKN-003)
+  const fullContentPattern = /Read\([^)]+\)\s*[\+,]|`\$\{.*Read\(/g;
+  for (const file of mdFiles) {
+    const content = Read(file);
+    const matches = content.match(fullContentPattern);
+    if (matches) {
+      evidence.push({
+        file: file,
+        pattern: 'TKN-003',
+        severity: 'high',
+        context: `Full content passing detected: ${matches[0]}`
+      });
+    }
+  }
+
+  // 4. Detect unbounded arrays (TKN-004)
+  const unboundedPattern = /\.(push|concat)\([^)]+\)(?!.*\.slice)/g;
+  for (const file of mdFiles) {
+    const content = Read(file);
+    const matches = content.match(unboundedPattern);
+    if (matches) {
+      evidence.push({
+        file: file,
+        pattern: 'TKN-004',
+        severity: 'high',
+        context: `Unbounded array growth: ${matches[0]}`
+      });
+    }
+  }
+
+  // 5. Detect write-then-read patterns (TKN-005)
+  const writeReadPattern = /Write\([^)]+\)[\s\S]{0,100}Read\([^)]+\)/g;
+  for (const file of mdFiles) {
+    const content = Read(file);
+    const matches = content.match(writeReadPattern);
+    if (matches) {
+      evidence.push({
+        file: file,
+        pattern: 'TKN-005',
+        severity: 'medium',
+        context: `Write-then-read pattern detected`
+      });
+    }
+  }
+
+  // Calculate severity
+  const highCount = evidence.filter(e => e.severity === 'high').length;
+  const mediumCount = evidence.filter(e => e.severity === 'medium').length;
+
+  let severity = 'none';
+  if (highCount > 0) severity = 'high';
+  else if (mediumCount > 2) severity = 'medium';
+  else if (mediumCount > 0) severity = 'low';
+
+  return {
+    status: 'completed',
+    issues_found: evidence.length,
+    severity: severity,
+    execution_time_ms: Date.now() - startTime,
+    details: {
+      patterns_checked: ['TKN-001', 'TKN-002', 'TKN-003', 'TKN-004', 'TKN-005'],
+      patterns_matched: [...new Set(evidence.map(e => e.pattern))],
+      evidence: evidence,
+      recommendations: generateRecommendations(evidence)
+    }
+  };
+}
+
+function generateRecommendations(evidence) {
+  const recs = [];
+  const patterns = [...new Set(evidence.map(e => e.pattern))];
+
+  if (patterns.includes('TKN-001')) {
+    recs.push('Apply prompt_compression: Extract static instructions to templates, use placeholders');
+  }
+  if (patterns.includes('TKN-002')) {
+    recs.push('Apply state_field_reduction: Remove debug/cache fields, consolidate related fields');
+  }
+  if (patterns.includes('TKN-003')) {
+    recs.push('Apply lazy_loading: Pass file paths instead of content, let agents read if needed');
+  }
+  if (patterns.includes('TKN-004')) {
+    recs.push('Apply sliding_window: Add .slice(-N) to array operations to bound growth');
+  }
+  if (patterns.includes('TKN-005')) {
+    recs.push('Apply output_minimization: Use in-memory data passing, eliminate temporary files');
+  }
+
+  return recs;
+}
+```
+
+## Output
+
+Write diagnosis result to `${workDir}/diagnosis/token-consumption-diagnosis.json`:
+
+```json
+{
+  "status": "completed",
+  "issues_found": 3,
+  "severity": "medium",
+  "execution_time_ms": 1500,
+  "details": {
+    "patterns_checked": ["TKN-001", "TKN-002", "TKN-003", "TKN-004", "TKN-005"],
+    "patterns_matched": ["TKN-001", "TKN-003"],
+    "evidence": [
+      {
+        "file": "phases/orchestrator.md",
+        "pattern": "TKN-001",
+        "severity": "medium",
+        "context": "File size: 5200 chars (threshold: 4000)"
+      }
+    ],
+    "recommendations": [
+      "Apply prompt_compression: Extract static instructions to templates"
+    ]
+  }
+}
+```
+
+## State Update
+
+```javascript
+updateState({
+  diagnosis: {
+    ...state.diagnosis,
+    token_consumption: diagnosisResult
+  }
+});
+```
+
+## Fix Strategies Mapping
+
+| Pattern | Strategy | Implementation |
+|---------|----------|----------------|
+| TKN-001 | prompt_compression | Extract static text to variables, use template inheritance |
+| TKN-002 | state_field_reduction | Audit and consolidate fields, remove non-essential data |
+| TKN-003 | lazy_loading | Pass paths instead of content, agents load when needed |
+| TKN-004 | sliding_window | Add `.slice(-N)` after push/concat operations |
+| TKN-005 | output_minimization | Use return values instead of file relay |
--- a/.claude/skills/skill-tuning/phases/orchestrator.md
+++ b/.claude/skills/skill-tuning/phases/orchestrator.md
@@ -93,7 +93,7 @@ function selectNextAction(state) {
  }

  // 4. Run diagnosis in order (only if not completed)
-  const diagnosisOrder = ['context', 'memory', 'dataflow', 'agent', 'docs'];
+  const diagnosisOrder = ['context', 'memory', 'dataflow', 'agent', 'docs', 'token_consumption'];

  for (const diagType of diagnosisOrder) {
    if (state.diagnosis[diagType] === null) {
@@ -221,6 +221,7 @@ async function runOrchestrator(workDir) {
    console.log(`[Loop ${iteration}] Executing: ${actionId}`);

    // 3. Update state: current action
+    // FIX CTX-001: sliding window for action_history (keep last 10)
    updateState({
      current_action: actionId,
      action_history: [...state.action_history, {
@@ -229,13 +230,24 @@ async function runOrchestrator(workDir) {
        completed_at: null,
        result: null,
        output_files: []
-      }]
+      }].slice(-10)  // Sliding window: prevent unbounded growth
    });

    // 4. Execute action
    try {
      const actionPrompt = Read(`phases/actions/${actionId}.md`);
-      const stateJson = JSON.stringify(state, null, 2);
+      // FIX CTX-003: Pass state path + key fields only instead of full state
+      const stateKeyInfo = {
+        status: state.status,
+        iteration_count: state.iteration_count,
+        issues_by_severity: state.issues_by_severity,
+        quality_gate: state.quality_gate,
+        current_action: state.current_action,
+        completed_actions: state.completed_actions,
+        user_issue_description: state.user_issue_description,
+        target_skill: { name: state.target_skill.name, path: state.target_skill.path }
+      };
+      const stateKeyJson = JSON.stringify(stateKeyInfo, null, 2);

      const result = await Task({
        subagent_type: 'universal-executor',
@@ -245,8 +257,12 @@ async function runOrchestrator(workDir) {
 You are executing action "${actionId}" for skill-tuning workflow.
 Work directory: ${workDir}

-[STATE]
-${stateJson}
+[STATE KEY INFO]
+${stateKeyJson}
+
+[FULL STATE PATH]
+${workDir}/state.json
+(Read full state from this file if you need additional fields)

 [ACTION INSTRUCTIONS]
 ${actionPrompt}
@@ -295,6 +311,7 @@ After completing the action:
      console.log(`[Loop ${iteration}] Error in ${actionId}: ${error.message}`);

      // Error handling
+      // FIX CTX-002: sliding window for errors (keep last 5)
      updateState({
        current_action: null,
        errors: [...state.errors, {
@@ -302,7 +319,7 @@ After completing the action:
          message: error.message,
          timestamp: new Date().toISOString(),
          recoverable: true
-        }],
+        }].slice(-5),  // Sliding window: prevent unbounded growth
        error_count: state.error_count + 1
      });
    }
--- a/.claude/skills/skill-tuning/phases/state-schema.md
+++ b/.claude/skills/skill-tuning/phases/state-schema.md
@@ -31,6 +31,7 @@ interface TuningState {
    dataflow: DiagnosisResult | null;
    agent: DiagnosisResult | null;
    docs: DocsDiagnosisResult | null;  // 文档结构诊断
+    token_consumption: DiagnosisResult | null;  // Token消耗诊断
  };

  // === Issues Found ===
@@ -69,6 +70,9 @@ interface TuningState {
  work_dir: string;
  backup_dir: string;

+  // === Final Report (consolidated output) ===
+  final_report: string | null;  // Markdown summary generated on completion
+
  // === Requirement Analysis (新增) ===
  requirement_analysis: RequirementAnalysis | null;
 }
@@ -176,7 +180,7 @@ interface Evidence {

 interface Issue {
  id: string;                   // e.g., "ISS-001"
-  type: 'context_explosion' | 'memory_loss' | 'dataflow_break' | 'agent_failure';
+  type: 'context_explosion' | 'memory_loss' | 'dataflow_break' | 'agent_failure' | 'token_consumption';
  severity: 'critical' | 'high' | 'medium' | 'low';
  priority: number;             // 1 = highest
  location: {
@@ -214,6 +218,10 @@ type FixStrategy =
  | 'schema_enforcement'        // Add data contract validation
  | 'orchestrator_refactor'     // Refactor agent coordination
  | 'state_centralization'      // Centralize state management
+  | 'prompt_compression'        // Extract static text, use templates
+  | 'lazy_loading'              // Pass paths instead of content
+  | 'output_minimization'       // Return minimal structured JSON
+  | 'state_field_reduction'     // Audit and consolidate state fields
  | 'custom';                   // Custom fix

 interface FileChange {
@@ -270,7 +278,8 @@ interface ErrorEntry {
    "memory": null,
    "dataflow": null,
    "agent": null,
-    "docs": null
+    "docs": null,
+    "token_consumption": null
  },
  "issues": [],
  "issues_by_severity": {
@@ -294,6 +303,7 @@ interface ErrorEntry {
  "max_errors": 3,
  "work_dir": null,
  "backup_dir": null,
+  "final_report": null,
  "requirement_analysis": null
 }
 ```
--- a/.claude/skills/skill-tuning/specs/category-mappings.json
+++ b/.claude/skills/skill-tuning/specs/category-mappings.json
@@ -93,6 +93,14 @@
      "detection_focus": null,
      "needs_gemini_analysis": true,
      "priority_order": [1, 2, 3, 4]
+    },
+    "token_consumption": {
+      "pattern_ids": ["TKN-001", "TKN-002", "TKN-003", "TKN-004", "TKN-005"],
+      "severity_hint": "medium",
+      "strategies": ["prompt_compression", "lazy_loading", "output_minimization", "state_field_reduction", "sliding_window"],
+      "risk_levels": ["low", "low", "low", "low", "low"],
+      "detection_focus": "Verbose prompts, excessive state fields, full content passing, unbounded arrays, redundant I/O",
+      "priority_order": [1, 2, 3, 4, 5]
    }
  },
  "keywords": {
@@ -149,7 +157,12 @@
      "中间文件": "authoring_principles_violation",
      "临时文件": "authoring_principles_violation",
      "文件中转": "authoring_principles_violation",
-      "state膨胀": "authoring_principles_violation"
+      "state膨胀": "authoring_principles_violation",
+      "token消耗": "token_consumption",
+      "token优化": "token_consumption",
+      "产出简化": "token_consumption",
+      "冗长": "token_consumption",
+      "精简": "token_consumption"
    },
    "english": {
      "token": "context_explosion",
@@ -198,7 +211,12 @@
      "feedback": "user_experience",
      "intermediate": "authoring_principles_violation",
      "temp": "authoring_principles_violation",
-      "relay": "authoring_principles_violation"
+      "relay": "authoring_principles_violation",
+      "verbose": "token_consumption",
+      "minimize": "token_consumption",
+      "compress": "token_consumption",
+      "simplify": "token_consumption",
+      "reduction": "token_consumption"
    }
  },
  "category_labels": {
@@ -213,6 +231,7 @@
    "output_quality": "Output Quality",
    "user_experience": "User Experience",
    "authoring_principles_violation": "Authoring Principles Violation",
+    "token_consumption": "Token Consumption",
    "custom": "Custom"
  },
  "category_labels_chinese": {
@@ -227,6 +246,7 @@
    "output_quality": "Output Quality",
    "user_experience": "User Experience",
    "authoring_principles_violation": "Authoring Principles Violation",
+    "token_consumption": "Token Consumption Optimization",
    "custom": "Other Issues"
  },
  "category_descriptions": {
@@ -241,12 +261,13 @@
    "output_quality": "Output validation or completeness issues",
    "user_experience": "Interaction or feedback clarity issues",
    "authoring_principles_violation": "Violation of skill authoring principles",
+    "token_consumption": "Excessive token usage from verbose prompts, large state objects, or redundant I/O patterns",
    "custom": "Requires custom analysis"
  },
  "fix_priority_order": {
    "P0": ["dataflow_break", "authoring_principles_violation"],
    "P1": ["agent_failure"],
-    "P2": ["context_explosion"],
+    "P2": ["context_explosion", "token_consumption"],
    "P3": ["memory_loss"]
  },
  "cross_category_dependencies": {
--- a/.claude/skills/skill-tuning/specs/problem-taxonomy.md
+++ b/.claude/skills/skill-tuning/specs/problem-taxonomy.md
@@ -180,7 +180,35 @@ Classification of skill execution issues with detection patterns and severity cr

 ---

-### 6. Documentation Conflict (P6)
+### 6. Token Consumption (P6)
+
+**Definition**: Excessive token usage from verbose prompts, large state objects, or inefficient I/O patterns.
+
+**Root Causes**:
+- Long static prompts without compression
+- State schema with too many fields
+- Full content embedding instead of path references
+- Arrays growing unbounded without sliding windows
+- Write-then-read file relay patterns
+
+**Detection Patterns**:
+
+| Pattern ID | Regex/Check | Description |
+|------------|-------------|-------------|
+| TKN-001 | File size > 4KB | Verbose prompt files |
+| TKN-002 | State fields > 15 | Excessive state schema |
+| TKN-003 | `/Read\([^)]+\)\s*[\+,]/` | Full content passing |
+| TKN-004 | `/.push\|concat(?!.*\.slice)/` | Unbounded array growth |
+| TKN-005 | `/Write\([^)]+\)[\s\S]{0,100}Read\([^)]+\)/` | Write-then-read pattern |
+
+**Impact Levels**:
+- **High**: Multiple TKN-003/TKN-004 issues causing significant token waste
+- **Medium**: Several verbose files or state bloat
+- **Low**: Minor optimization opportunities
+
+---
+
+### 7. Documentation Conflict (P7)

 **Definition**: 同一概念在不同文件中定义不一致，导致行为不可预测和文档误导。

@@ -262,6 +290,7 @@ function calculateIssueSeverity(issue) {
 | Long-tail Forgetting | constraint_injection, state_constraints_field, checkpoint | 1, 2, 3 |
 | Data Flow Disruption | state_centralization, schema_enforcement, field_normalization | 1, 2, 3 |
 | Agent Coordination | error_wrapping, result_validation, flatten_nesting | 1, 2, 3 |
+| **Token Consumption** | prompt_compression, lazy_loading, output_minimization, state_field_reduction | 1, 2, 3, 4 |
 | **Documentation Redundancy** | consolidate_to_ssot, centralize_mapping_config | 1, 2 |
 | **Documentation Conflict** | reconcile_conflicting_definitions | 1 |

--- a/.claude/skills/skill-tuning/specs/tuning-strategies.md
+++ b/.claude/skills/skill-tuning/specs/tuning-strategies.md
@@ -1308,3 +1308,230 @@ async function checkpoint(name, summary, options) {
  return response;
 }
 ```
+
+---
+
+## Token Consumption Strategies
+
+Strategies for reducing token usage and simplifying skill outputs.
+
+---
+
+### Strategy: prompt_compression
+
+**Purpose**: Reduce verbose prompts by extracting static text and using templates.
+
+**Implementation**:
+```javascript
+// Before: Long inline prompt
+const prompt = `
+You are an expert code analyzer specializing in identifying patterns.
+Your role is to examine the provided code and identify any issues.
+Please follow these detailed instructions carefully:
+1. Read the code thoroughly
+2. Identify any anti-patterns
+3. Check for security vulnerabilities
+... (continues for many lines)
+
+Code to analyze:
+${fullCodeContent}
+`;
+
+// After: Compressed with template reference
+const PROMPT_TEMPLATE_PATH = 'templates/analyzer-prompt.md';
+
+const prompt = `
+[TEMPLATE: ${PROMPT_TEMPLATE_PATH}]
+[CODE_PATH: ${codePath}]
+[FOCUS: patterns, security]
+`;
+
+// Or use key instructions only
+const prompt = `
+Analyze ${codePath} for: patterns, security, performance.
+Return JSON: { issues: [], severity: string }
+`;
+```
+
+**Risk**: Low
+**Verification**: Compare token count before/after compression
+
+---
+
+### Strategy: lazy_loading
+
+**Purpose**: Pass file paths instead of full content, let consumers load if needed.
+
+**Implementation**:
+```javascript
+// Before: Full content in prompt
+const content = Read(filePath);
+const prompt = `Analyze this content:\n${content}`;
+
+// After: Path reference with lazy loading instruction
+const prompt = `
+Analyze file at: ${filePath}
+(Read the file content if you need to examine it)
+Return: { summary: string, issues: [] }
+`;
+
+// For agent calls
+const result = await Task({
+  subagent_type: 'universal-executor',
+  prompt: `
+    [FILE_PATH]: ${dataPath}
+    [TASK]: Analyze the file and extract key metrics.
+    [NOTE]: Read the file only if needed for your analysis.
+  `
+});
+```
+
+**Risk**: Low
+**Verification**: Verify agents can still access required data
+
+---
+
+### Strategy: output_minimization
+
+**Purpose**: Configure agents to return minimal, structured output instead of verbose text.
+
+**Implementation**:
+```javascript
+// Before: Verbose output expectation
+const prompt = `
+Analyze the code and provide a detailed report including:
+- Executive summary
+- Detailed findings with explanations
+- Code examples for each issue
+- Recommendations with rationale
+...
+`;
+
+// After: Minimal structured output
+const prompt = `
+Analyze the code. Return ONLY this JSON:
+{
+  "status": "pass|review|fail",
+  "issues": [{ "id": string, "severity": string, "file": string, "line": number }],
+  "summary": "one sentence"
+}
+Do not include explanations or code examples.
+`;
+
+// Validation
+const result = JSON.parse(agentOutput);
+if (!result.status || !Array.isArray(result.issues)) {
+  throw new Error('Invalid output format');
+}
+```
+
+**Risk**: Low
+**Verification**: JSON.parse succeeds, output size reduced
+
+---
+
+### Strategy: state_field_reduction
+
+**Purpose**: Audit and consolidate state fields to minimize serialization overhead.
+
+**Implementation**:
+```typescript
+// Before: Bloated state
+interface State {
+  status: string;
+  target: TargetInfo;
+  user_input: string;
+  parsed_input: ParsedInput;      // Remove - temporary
+  intermediate_result: any;       // Remove - not persisted
+  debug_info: DebugInfo;          // Remove - debugging only
+  analysis_cache: any;            // Remove - session cache
+  full_history: HistoryEntry[];   // Remove - unbounded
+  step1_output: any;              // Remove - intermediate
+  step2_output: any;              // Remove - intermediate
+  step3_output: any;              // Remove - intermediate
+  final_result: FinalResult;
+  error_log: string[];            // Remove - debugging
+  metrics: Metrics;               // Remove - optional
+}
+
+// After: Minimal state (≤15 fields)
+interface State {
+  status: 'pending' | 'running' | 'completed' | 'failed';
+  target: { name: string; path: string };
+  input_summary: string;          // Summarized user input
+  result_path: string;            // Path to final result
+  quality_gate: 'pass' | 'fail';
+  error?: string;                 // Only if failed
+}
+```
+
+**Audit Checklist**:
+```javascript
+function auditStateFields(stateSchema) {
+  const removeCandidates = [];
+
+  for (const [key, type] of Object.entries(stateSchema)) {
+    // Identify removal candidates
+    if (key.startsWith('debug_')) removeCandidates.push(key);
+    if (key.endsWith('_cache')) removeCandidates.push(key);
+    if (key.endsWith('_temp')) removeCandidates.push(key);
+    if (key.includes('intermediate')) removeCandidates.push(key);
+    if (key.includes('step') && key.includes('output')) removeCandidates.push(key);
+  }
+
+  return {
+    total_fields: Object.keys(stateSchema).length,
+    remove_candidates: removeCandidates,
+    estimated_reduction: removeCandidates.length
+  };
+}
+```
+
+**Risk**: Medium (ensure no essential data removed)
+**Verification**: State field count ≤ 15, all essential data preserved
+
+---
+
+### Strategy: in_memory_consolidation
+
+**Purpose**: Consolidate outputs into single file, eliminate redundant report files.
+
+**Implementation**:
+```javascript
+// Before: Multiple output files
+Write(`${workDir}/diagnosis-report.md`, reportMarkdown);
+Write(`${workDir}/diagnosis-summary.json`, summaryJson);
+Write(`${workDir}/state.json`, JSON.stringify(state));
+Write(`${workDir}/tuning-report.md`, tuningReport);
+Write(`${workDir}/tuning-summary.md`, finalSummary);
+
+// After: Single source of truth
+const consolidatedState = {
+  ...state,
+  final_report: {
+    summary: summaryJson,
+    details_available_in_state: true,
+    generated_at: new Date().toISOString()
+  }
+};
+Write(`${workDir}/state.json`, JSON.stringify(consolidatedState, null, 2));
+
+// Report can be rendered from state on-demand
+function renderReport(state) {
+  return `
+# Tuning Report: ${state.target_skill.name}
+Status: ${state.status}
+Quality: ${state.quality_gate}
+Issues: ${state.issues.length}
+...
+`;
+}
+```
+
+**Benefits**:
+- Single file to read/write
+- No data duplication
+- On-demand rendering
+
+**Risk**: Low
+**Verification**: Only state.json exists as output, rendering works correctly