enhance-cross-file（增强交叉文件）

enhance-cross-file

Analyze cross-file semantic consistency across agents, skills, and workflows.

Parse Arguments

const args = '$ARGUMENTS'.split(' ').filter(Boolean);
const targetPath = args.find(a => !a.startsWith('--')) || '.';

Purpose

Detects issues that span multiple files - things single-file analysis misses:

• Tools used in prompt body but not declared in frontmatter
• Agent references that don't exist
• Duplicate instructions across files (maintenance burden)
• Contradictory rules (ALWAYS vs NEVER conflicts)
• Orphaned agents not referenced by any workflow
• Skill tool mismatches (allowed-tools vs actual usage)

Workflow

1. Run Analyzer - Execute the JavaScript analyzer to get findings: node -e "const a = require('./lib/enhance/cross-file-analyzer.js'); console.log(JSON.stringify(a.analyze('.'), null, 2));" For a specific path: a.analyze('./plugins/enhance')
2. Parse Results - The analyzer returns JSON with summary and findings
3. Report - Return findings grouped by category

The JavaScript analyzer (lib/enhance/cross-file-analyzer.js) implements all cross-file detection. The patterns below are reference documentation.

Detection Patterns

1. Tool Consistency (MEDIUM Certainty)

tool_not_in_allowed_list: Tool used in prompt body but not in frontmatter tools: list

# Frontmatter declares:
tools: Read, Grep

# But body uses:
Use Write({ file_path: "/out" })  # <- Not declared!

skill_tool_mismatch: Skill's allowed-tools doesn't match actual tool usage in skill body

2. Workflow Consistency (MEDIUM Certainty)

missing_workflow_agent: subagent_type: "plugin:agent-name" references non-existent agent

orphaned_prompt: Agent file exists but no workflow references it (may be entry point - check manually)

incomplete_phase_transition: Workflow phase mentions "Phase N" but no corresponding section

3. Instruction Consistency (MEDIUM Certainty)

duplicate_instructions: Same MUST/NEVER instruction in 3+ files (extract to shared location)

contradictory_rules: One file says "ALWAYS X" while another says "NEVER X"

Output Format

## Cross-File Analysis

**Files Analyzed**: {agents} agents, {skills} skills, {commands} commands

### Tool Consistency ({n})
| Agent | Issue | Fix |
|-------|-------|-----|
| exploration-agent | Uses Write but not in tools list | Add Write to frontmatter |

### Workflow Issues ({n})
| Source | Issue | Fix |
|--------|-------|-----|
| workflow.md | References nonexistent agent | Check spelling or create agent |

### Instruction Consistency ({n})
| Instruction | Files | Fix |
|-------------|-------|-----|
| "NEVER push --force" | 4 files | Extract to CLAUDE.md |

Constraints

• All patterns are MEDIUM certainty (require context)
• No auto-fix (cross-file changes need human review)
• Skip content inside <bad-example>, <bad_example>, <badexample> tags
• Skip content inside code blocks with "bad" in info string
• Entry point agents (orchestrator, validator, discoverer) are not orphaned

Pattern Statistics