docs-audit（文件审核）

Documentation Audit Skill

Audit documentation files against the current codebase state, checking for accuracy, completeness, and freshness.

Workflow

Stage 1: Discovery

Run the discovery script to get a manifest of all markdown files with git metadata:

bun "$(dirname "$0")/scripts/discover-docs.ts" ${path ? `--path "${path}"` : ""} --limit 50 --sort staleness

Parse the JSON output to understand:

• Total documentation files in scope
• Activity status distribution (active/recent/idle/stale/ancient)
• Files with related code changes (potential staleness indicators)

Stage 2: Prioritization

Select files for deep analysis based on focus argument:

Target: Select top limit files (default 10) for deep analysis.

Stage 3: Deep Analysis

For each selected file, perform these checks:

3.1 Correctness (see references/correctness-checklist.md)

• Code examples use correct imports/require paths
• Function signatures match current implementation
• Configuration examples reflect current schema
• CLI commands and flags are accurate
• Environment variables mentioned actually exist

3.2 Link Validation

• Internal markdown links ([text](./other.md)) resolve
• Anchor links ([text](#section)) point to existing headings
• Image references exist
• External URLs (sample only if many) - note but don't block on these

3.3 Completeness (see references/completeness-checklist.md)

• Required sections present (varies by doc type)
• All public exports documented (for API docs)
• Examples provided for complex features
• Error handling documented where relevant

Stage 4: Docstring Coverage

Check TSDoc/JSDoc/docstring coverage for code files related to the documentation:

TypeScript/JavaScript:

# Find exports without TSDoc
grep -rn "^export " --include="*.ts" --include="*.tsx" | head -20
# vs exports with TSDoc (/** precedes export)
grep -B1 "^export " --include="*.ts" --include="*.tsx" | grep -c "/\*\*"

Python:

# Find functions/classes without docstrings
grep -rn "^def \|^class " --include="*.py" | head -20

Rust:

# Find pub items without doc comments
grep -rn "^pub " --include="*.rs" | head -20

Go:

# Find exported funcs without godoc
grep -rn "^func [A-Z]" --include="*.go" | head -20

Calculate coverage percentage per language detected.

Stage 5: Report Generation

First, generate the report path using the helper script:

REPORT_PATH=$(bun "$(dirname "$0")/scripts/report-path.ts" --session "${CLAUDE_SESSION_ID}" --json)
# Extracts: timestamp, sessionShort, path, timestampISO

Write the report to the generated path (e.g., .pack/reports/202601251900-docs-audit-a7b3c2d1.md):

---
type: docs-audit
generated: {timestampISO}
timestamp: "{timestamp}"
session: "{CLAUDE_SESSION_ID}"
session_short: "{sessionShort}"
scope: {path or "entire repo"}
focus: {focus}
files_analyzed: {count}
files_total: {total}
status: {pass|needs-work|critical}
---

# Documentation Audit Report

**Generated**: {timestamp}
**Session**: `{sessionShort}`
**Scope**: {path or "entire repo"}
**Files analyzed**: {count} / {total}
**Focus**: {focus}

## Summary

| Dimension | Status | Score |
|-----------|--------|-------|
| Correctness | {PASS/NEEDS WORK} | {x}/{y} files |
| Links | {PASS/NEEDS WORK} | {valid}/{total} |
| Docstrings | {GOOD/ACCEPTABLE/POOR} | {x}% |
| Freshness | {CURRENT/STALE} | {stale_count} files |

## Critical Issues (blocking)

Issues that could cause user confusion or errors:
- {file}: {issue description}

## Warnings (should fix)

Non-blocking but should be addressed:
- {file}: {issue description}

## Stale Documentation

Files that may need review (old docs + recent code changes):
- {file}: Last updated {days}d ago, related code changed {code_days}d ago

## Docstring Coverage by Language

| Language | Coverage | Files Checked |
|----------|----------|---------------|
| TypeScript | {x}% | {n} |
| Python | {x}% | {n} |

## Recommendations

1. {Prioritized recommendation}
2. {Next recommendation}

Report Output

Path Generation & Scaffolding

Use the report-path.ts helper script to generate paths and scaffold directories:

# Get just the path
bun scripts/report-path.ts --session "${CLAUDE_SESSION_ID}"
# → .pack/reports/202601251900-docs-audit-a7b3c2d1.md

# Scaffold the directory structure (creates .pack/reports/)
bun scripts/report-path.ts --scaffold --session "${CLAUDE_SESSION_ID}"

# Multi-file mode: scaffold with placeholder files
bun scripts/report-path.ts --scaffold --multi --session "${CLAUDE_SESSION_ID}"
# Creates:
#   .pack/reports/202601251900-docs-audit/
#   .pack/reports/202601251900-docs-audit/summary.md
#   .pack/reports/202601251900-docs-audit/markdown-docs.md
#   .pack/reports/202601251900-docs-audit/docstrings.md
#   .pack/reports/202601251900-docs-audit/recommendations.md
#   .pack/reports/202601251900-docs-audit/meta.json

# Get all components as JSON (includes scaffolded paths if --scaffold used)
bun scripts/report-path.ts --scaffold --multi --session "${CLAUDE_SESSION_ID}" --json

Default Location

Reports are written to .pack/reports/ with frontloaded timestamp:

.pack/reports/202601251900-docs-audit-a7b3c2d1.md   # Single file (with session)
.pack/reports/202601251900-docs-audit.md            # Single file (no session)
.pack/reports/202601251900-docs-audit/              # Multi-file (no session in dir name)

Filename patterns:

• Single-file: {timestamp}-docs-audit-{sessionShort}.md (session for parallel disambiguation)
• Multi-file: {timestamp}-docs-audit/ (session tracked in frontmatter inside files)

Multi-File Mode

For comprehensive audits covering different documentation types, use --multi:

.pack/reports/202601251900-docs-audit/
├── summary.md           # Overall findings + links to other reports
├── markdown-docs.md     # docs/, README, etc.
├── docstrings.md        # TSDoc/JSDoc/docstring coverage
├── recommendations.md   # Prioritized actionable recommendations
└── meta.json            # Session metadata (structured, machine-readable)

Each file includes frontmatter with full session ID for traceability.

Frontmatter Schema

All report artifacts include YAML frontmatter for searchability:

---
type: docs-audit           # Report type (searchable)
generated: 2026-01-25T19:00:00Z
timestamp: "202601251900"
session: abc123-def456...  # Full session ID
session_short: a7b3c2d1    # First 8 chars (matches filename)
scope: docs/               # Audit scope
focus: stale               # Focus strategy used
files_analyzed: 10
files_total: 47
status: needs-work         # pass | needs-work | critical
---

Why frontmatter:

• Grep/ripgrep searchable (rg "session: abc123")
• Tooling can parse and aggregate reports
• Enables filtering by status, scope, date range
• Parallel agent runs are distinguishable by session

Session ID for Parallel Agents

Single-file mode uses session suffix for parallel disambiguation:

.pack/reports/202601251900-docs-audit-a7b3c2d1.md  # Agent 1
.pack/reports/202601251900-docs-audit-f8e9d0c1.md  # Agent 2
.pack/reports/202601251900-docs-audit-12345678.md  # Agent 3

Multi-file mode relies on timestamps (coordinated audits typically don't run in parallel). Session is tracked inside each file's frontmatter for traceability.

Custom Output

Override the default location (session ID still included in frontmatter):

/docs-audit --output docs/audits/latest.md
/docs-audit --multi --output .pack/reports/202601-quarterly/

Note: Custom filenames don't auto-include session ID prefix - use frontmatter for tracking.

Output Behavior

1. Create directory if it doesn't exist
2. Write report(s) with session ID embedded
3. Print summary to conversation (critical issues + file path)
4. Return path so user can open/commit the report

Git Considerations

The default .pack/reports/ location:

• Should be gitignored for ephemeral reports
• Can be selectively committed for audit history
• Keeps reports separate from actual documentation

Context Efficiency

This skill uses context: fork to run in isolation. The token budget strategy:

Total target: 15-30k tokens for a typical audit.

Task Management

Use task tools (TaskCreate, TaskUpdate, TaskList) to track progress through stages. Tasks survive context compaction and allow resumption if the audit is interrupted.

Initial Task Setup

After discovery, create tasks for the audit stages:

TaskCreate:
  subject: "Run docs-audit discovery"
  activeForm: "Running discovery script"
  description: "Execute discover-docs.ts, parse manifest, identify {n} files in scope"

TaskCreate:
  subject: "Analyze {n} priority docs"
  activeForm: "Analyzing documentation"
  description: "Deep analysis of top {limit} files for correctness, links, completeness"

TaskCreate:
  subject: "Check docstring coverage"
  activeForm: "Checking docstring coverage"
  description: "Grep exports vs documented exports per detected language"

TaskCreate:
  subject: "Generate audit report"
  activeForm: "Generating report"
  description: "Compile findings into .pack/reports/{timestamp}-docs-audit-{sessionShort}.md"

Progress Tracking

Update tasks as you work:

1. Before starting a stage → TaskUpdate with status: in_progress
2. After completing a stage → TaskUpdate with status: completed, update description with key findings
3. If issues found → TaskCreate follow-up tasks for fixes

State Persistence

Before context approaches limit, update task descriptions with checkpoint data:

TaskUpdate:
  taskId: "2"
  description: |
    [CHECKPOINT] Analyzed 7/10 docs.
    Critical: 2 broken imports in api.md (lines 45, 89)
    Warnings: 3 stale files (config.md, setup.md, advanced.md)
    Remaining: config.md, setup.md, advanced.md

This ensures findings survive compaction even if the stage isn't complete.

Resumption

If context resets mid-audit:

1. TaskList to see current state
2. TaskGet on in_progress task to read checkpoint data
3. Skip completed stages
4. Resume from checkpoint, don't re-analyze completed files
5. Reference persisted findings in final report

Handling Edge Cases

No documentation found: Report "No markdown files found in {scope}. Consider adding documentation for your project."

Very large repos (100+ docs):

• Stick to limit parameter strictly
• Focus on highest-staleness files
• Note total count in report for context

Non-git repos:

• Skip git-based metadata (SHA, author, etc.)
• Use file modification times as fallback
• Note "Git metadata unavailable" in report

Mixed language repos:

• Detect languages from file extensions
• Report coverage per detected language
• Skip languages with no source files

Example Invocations

/docs-audit                                    # Single report to .pack/reports/
/docs-audit --path docs/                       # Scope to docs/ directory
/docs-audit --focus all --limit 20             # Analyze 20 files across all statuses
/docs-audit --multi                            # Multi-file mode with separate reports
/docs-audit --output docs/audits/latest.md     # Custom output location
/docs-audit --multi --path src/                # Multi-file audit of src/ docs