session-analyzer（会话分析器）

Session Analyzer Skill

Post-hoc analysis tool for validating Claude Code session behavior against SKILL.md specifications.

Purpose

Analyze completed sessions to verify:

1. Expected vs Actual Behavior - Did the skill follow SKILL.md workflow?
2. Component Invocations - Were SubAgents, Hooks, and Tools called correctly?
3. Artifacts - Were expected files created/deleted?
4. Bug Detection - Any unexpected errors or deviations?

Input Requirements

Phase 1: Locate Session Files

Step 1.1: Find Session Files

Session files are located in ~/.claude/:

# Main session log
~/.claude/projects/-{encoded-cwd}/{sessionId}.jsonl

# Debug log (detailed)
~/.claude/debug/{sessionId}.txt

# Agent transcripts (if subagents were used)
~/.claude/projects/-{encoded-cwd}/agent-{agentId}.jsonl

Use script to locate files:

${baseDir}/scripts/find-session-files.sh {sessionId}

Step 1.2: Verify Files Exist

Check all required files exist before proceeding. If debug log is missing, analysis will be limited.

Phase 2: Parse Target SKILL.md

Step 2.1: Extract Expected Components

Read the target SKILL.md and identify:

From YAML Frontmatter:

• hooks.PreToolUse - Expected PreToolUse hooks and matchers
• hooks.PostToolUse - Expected PostToolUse hooks
• hooks.Stop - Expected Stop hooks
• hooks.SubagentStop - Expected SubagentStop hooks
• allowed-tools - Tools the skill is allowed to use

From Markdown Body:

• SubAgents mentioned (Task(subagent_type="..."))
• Skills called (Skill("..."))
• Artifacts created (.dev-flow/drafts/, .dev-flow/plans/, etc.)
• Workflow steps and conditions

Step 2.2: Build Expected Behavior Checklist

Create checklist from SKILL.md analysis:

## Expected Behavior

### SubAgents
- [ ] Explore agent called (parallel, run_in_background)
- [ ] gap-analyzer called before plan generation
- [ ] reviewer called after plan creation

### Hooks
- [ ] PreToolUse[Edit|Write] triggers plan-guard.sh
- [ ] Stop hook validates reviewer approval

### Artifacts
- [ ] Draft file created at .dev-flow/drafts/{name}.md
- [ ] Plan file created at .dev-flow/plans/{name}.md
- [ ] Draft file deleted after OKAY

### Workflow
- [ ] Interview Mode before Plan Generation
- [ ] User explicit request triggers plan generation
- [ ] Reviewer REJECT causes revision loop

Phase 3: Analyze Debug Log

The debug log (~/.claude/debug/{sessionId}.txt) contains detailed execution traces.

Step 3.1: Extract SubAgent Calls

Search patterns:

SubagentStart with query: {agent-name}
SubagentStop with query: {agent-id}

Use script:

${baseDir}/scripts/extract-subagent-calls.sh {debug-log-path}

Step 3.2: Extract Hook Events

Search patterns:

Getting matching hook commands for {HookEvent} with query: {tool-name}
Matched {N} unique hooks for query "{query}"
Hooks: Processing prompt hook with prompt: {prompt}
Hooks: Prompt hook condition was met/not met
permissionDecision: allow/deny

Use script:

${baseDir}/scripts/extract-hook-events.sh {debug-log-path}

Step 3.3: Extract Tool Calls

Search patterns:

executePreToolHooks called for tool: {tool-name}
File {path} written atomically

Step 3.4: Extract Hook Results

For prompt-based hooks, find the model response:

Hooks: Model response: {
  "ok": true/false,
  "reason": "..."
}

Phase 4: Verify Artifacts

Step 4.1: Check File Creation

For each expected artifact:

1. Search debug log for FileHistory: Tracked file modification for {path}
2. Search for File {path} written atomically
3. Verify current filesystem state

Step 4.2: Check File Deletion

For files that should be deleted:

1. Search for rm commands in Bash calls
2. Verify file no longer exists on filesystem

Phase 5: Compare Expected vs Actual

Step 5.1: Build Comparison Table

| Component | Expected | Actual | Status |
|-----------|----------|--------|--------|
| Explore agent | 2 parallel calls | 2 calls at 09:39:26 | ✅ |
| gap-analyzer | Called before plan | Called at 09:43:08 | ✅ |
| reviewer | Called after plan | 2 calls (REJECT→OKAY) | ✅ |
| PreToolUse hook | Edit\|Write matcher | Triggered for Write | ✅ |
| Stop hook | Validates approval | Returned ok:true | ✅ |
| Draft file | Created then deleted | Created→Deleted | ✅ |
| Plan file | Created | Exists (10KB) | ✅ |

Step 5.2: Identify Deviations

Flag any mismatches:

• Missing component calls
• Wrong order of operations
• Hook failures
• Missing artifacts
• Unexpected errors

Phase 6: Generate Report

Report Template

# Session Analysis Report

## Session Info
- **Session ID**: {sessionId}
- **Target Skill**: {skillPath}
- **Analysis Date**: {date}

---

## 1. Expected Behavior (from SKILL.md)

[Summary of expected workflow]

---

## 2. Skill/SubAgent/Hook Verification

### SubAgents
| SubAgent | Expected | Actual | Time | Result |
|----------|----------|--------|------|--------|
| ... | ... | ... | ... | ✅/❌ |

### Hooks
| Hook | Matcher | Triggered | Result |
|------|---------|-----------|--------|
| ... | ... | ... | ✅/❌ |

---

## 3. Artifacts Verification

| Artifact | Path | Expected State | Actual State |
|----------|------|----------------|--------------|
| ... | ... | ... | ✅/❌ |

---

## 4. Issues/Bugs

| Severity | Description | Location |
|----------|-------------|----------|
| ... | ... | ... |

---

## 5. Overall Result

**Verdict**: ✅ PASS / ❌ FAIL

**Summary**: [1-2 sentence summary]

Scripts Reference

Usage Example

User: "Analyze session 3cc71c9f-d27a-4233-9dbc-c4f07ea6ec5b against .claude/skills/specify/SKILL.md"

1. Find session files
2. Parse SKILL.md → Expected: Explore, gap-analyzer, reviewer, hooks
3. Analyze debug log → Extract actual calls
4. Verify artifacts → Check .dev-flow/
5. Compare → Build verification table
6. Generate report → PASS/FAIL with details

Additional Resources

Reference Files

• references/analysis-patterns.md - Detailed grep patterns for log analysis
• references/common-issues.md - Known issues and troubleshooting

Scripts

• scripts/find-session-files.sh - Session file locator
• scripts/extract-subagent-calls.sh - SubAgent call extractor
• scripts/extract-hook-events.sh - Hook event extractor