session-wrap

세션에서 새로 확인된 규칙·패턴을 `CLAUDE.md`·`.claude/rules/`·`.claude/commands/` 에 반영하고, 자동화할 만한 반복 작업과 후속 작업을 함께 정리해 보여줍니다 — 작업을 마칠 때.

/cc-dev:session:wrap — 작업 마무리 정리#

항목	내용
실행 명령	`/cc-dev:session:wrap`
분류	자동화
난이도	●●○ 보통

한마디로#

오늘 작업하면서 새로 배운 규칙과 팁을 AI가 알아서 정리해 매뉴얼 문서에 적어두는 단계입니다. 회의가 끝난 뒤 누군가 회의록을 정리해 다음 사람도 볼 수 있게 남겨두는 것과 같아요.

누가·언제 쓰나요#

한 번의 작업(세션)을 마칠 때, 끝내기 직전에
작업 중 중요한 패턴이나 규칙을 새로 발견했을 때
반복되는 일이라 "이건 자동화하면 좋겠다" 싶은 게 보였을 때
매뉴얼 문서를 최신 상태로 갱신하고 싶을 때

무엇을 해주나요#

작업 내용을 분석해서, 프로젝트의 안내 문서들을 자동으로 업데이트해 줍니다. 주로 손대는 곳은:

CLAUDE.md — 프로젝트의 기본 안내서 (규칙·구조 변경 반영)
.claude/rules/ — 코딩 규칙 문서 (새 규칙 추가)
.claude/commands/ — 새 기능(스킬/에이전트) 정의 추가
.claude/session-logs/ — 작업 요약 보관 (선택)

또한 자동화하면 좋을 반복 작업, 다음에 이어서 할 후속 작업도 함께 정리해 보여줍니다.

어떻게 쓰나요#

# 전체 작업 분석 후 문서 업데이트
/cc-dev:session:wrap

# 특정 영역만 갱신
/cc-dev:session:wrap --scope rules      # 규칙만
/cc-dev:session:wrap --scope commands   # 명령(스킬)만
/cc-dev:session:wrap --scope agents     # 에이전트만

# 미리보기 (실제 수정 없이 바뀔 내용만 확인) - 추천
/cc-dev:session:wrap --dry-run

처음 쓸 때는 --dry-run으로 먼저 미리보기 하는 것을 권장합니다. 실제 파일을 고치기 전에 어떤 내용이 추가/수정될지 확인할 수 있어요.

안에서 무슨 일이 벌어지나요#

크게 두 단계로 진행됩니다.

여러 관점에서 동시에 분석 (4가지) — ① 문서 업데이트할 내용 찾기, ② 자동화하면 좋을 반복 패턴 찾기, ③ 이번에 배운 규칙·패턴 뽑아내기, ④ 다음에 이어서 할 후속 작업 제안하기. 네 가지를 한꺼번에 살펴봅니다.
중복 정리 후 마무리 (1가지) — 앞에서 모은 결과 중 겹치거나 충돌하는 내용을 정리하고, 기존 문서와 일관성을 맞춰 최종 반영합니다. 마지막에 무엇이 바뀌었는지 요약해서 보여줍니다.

기존 내용을 함부로 지우지 않고 더하거나 고치는 식으로만 바꾸며, 중요한 변경은 사용자 확인을 받은 뒤 적용합니다.

⚙️ 상세 옵션·실행 명세 (개발자 / AI 에이전트용)

Triggers#

Just before session end
After discovering important patterns/rules
When new automatable tasks are discovered
When document updates are needed

Usage#

# Full session analysis and document update
/cc-dev:session:wrap

# Update specific area only
/cc-dev:session:wrap --scope rules      # Rules only
/cc-dev:session:wrap --scope commands   # Commands only
/cc-dev:session:wrap --scope agents     # Agents only

# Dry run (preview changes) - recommended
/cc-dev:session:wrap --dry-run

Recommended: Use --dry-run on first use to preview changes. You can review what content will be added/modified before actual file modification.

Execution Flow (2 Phases)#

┌─────────────────────────────────────────────────────────────────┐
│                    /cc-dev:session:wrap                                │
├─────────────────────────────────────────────────────────────────┤
│                                                                 │
│  ═══════════════════════════════════════════════════════════   │
│  ║ Phase 1: Parallel Analysis (4 Agents)                     ║   │
│  ═══════════════════════════════════════════════════════════   │
│                                                                 │
│  ┌─────────────┐ ┌─────────────┐ ┌─────────────┐ ┌────────────┐│
│  │ doc-updater │ │ automation  │ │ learning    │ │ followup   ││
│  │   Agent     │ │   -scout    │ │ -extractor  │ │ -suggester ││
│  └──────┬──────┘ └──────┬──────┘ └──────┬──────┘ └─────┬──────┘│
│         │               │               │              │        │
│         ▼               ▼               ▼              ▼        │
│  ┌────────────────────────────────────────────────────────────┐│
│  │                   Collect Analysis Results                 ││
│  └────────────────────────────────────────────────────────────┘│
│                                                                 │
│  ═══════════════════════════════════════════════════════════   │
│  ║ Phase 2: Sequential Processing (1 Agent)                  ║   │
│  ═══════════════════════════════════════════════════════════   │
│                                                                 │
│  ┌─────────────────────────────────────────────────────────┐   │
│  │             duplicate-checker Agent                      │   │
│  │    - Duplicate content check                             │   │
│  │    - Conflict resolution                                 │   │
│  │    - Final document cleanup                              │   │
│  └─────────────────────────────────────────────────────────┘   │
│                                                                 │
│  ┌─────────────────────────────────────────────────────────┐   │
│  │                  Output Result Summary                    │   │
│  └─────────────────────────────────────────────────────────┘   │
│                                                                 │
└─────────────────────────────────────────────────────────────────┘

Phase 1: Parallel Analysis Agents#

1. doc-updater Agent#

Update CLAUDE.md and related documents based on session content

Analysis Targets:

Newly discovered coding conventions
Added project settings
Changed architecture patterns
New dependencies/packages

Output:

updates:
  - file: CLAUDE.md
    section:  " ## Architecture and Structure " 
     action: append
    content:  " New architecture pattern description... " 
   - file: .claude/rules/naming.md
    section:  " ### Naming Rules " 
     action: update
    content:  " Updated naming rules... "

2. automation-scout Agent#

Detect new automatable patterns

Analysis Targets:

Repetitively performed tasks
Manually handled tasks that could be pattern-ized
Tasks that could become new skills

Output:

automatable_patterns:
  - pattern:  " Always add State when adding BLoC event " 
     frequency: 3
    suggestion:  " Recommend creating bloc-event-state-pair skill " 
   - pattern:  " Always add same imports after test file creation " 
     frequency: 5
    suggestion:  " Add auto-import to test-template skill "

3. learning-extractor Agent#

Extract rules/patterns learned during the session

Analysis Targets:

Modified code patterns
Resolved errors and their solutions
Preferences learned from user feedback
Project-specific rules

Output:

learnings:
  - type: rule
    category:  " Code style " 
     content:  " Actively use dot shorthand (.start, .center, etc.) " 
     evidence:  " User pointed out unnecessary full type name usage " 
   - type: pattern
    category:  " BLoC " 
     content:  " isClosed check required after await " 
     evidence:  " Fixed multiple times due to lint errors "

4. followup-suggester Agent#

Suggest follow-up tasks

Analysis Targets:

Incomplete TODOs
Mentioned improvements
Discovered technical debt
Tasks for next session

Output:

followups:
  - priority: high
    task:  " Achieve 80% test coverage for console_author_management " 
     reason:  " Currently 45% - missing tests for key features " 
   - priority: medium
    task:  " Optimize author search API " 
     reason:  " Takes 2+ seconds for 100+ records "

Phase 2: Sequential Processing Agent#

5. duplicate-checker Agent#

Duplicate content check and cleanup

Processing Content:

Duplicate check of Phase 1 results
Conflict resolution with existing documents
Consistency verification
Apply final changes

Output:

deduplication_result:
  removed_duplicates: 2
  merged_entries: 1
  conflicts_resolved: 0
  final_changes:
    - file: CLAUDE.md
      lines_added: 15
      lines_removed: 3
    - file: .claude/rules/bloc-patterns.md
      lines_added: 8
      lines_removed: 0

Detailed Implementation#

Main Execution Logic#

async function sessionWrap(options: WrapOptions) {
  // Phase 1: Parallel analysis
  const [docUpdates, automations, learnings, followups] = await Promise.all([
    Task({
      subagent_type:  " session-doc-updater " ,
      prompt:  " Analyze session conversation content to suggest document updates " ,
    }),
    Task({
      subagent_type:  " session-automation-scout " ,
      prompt:  " Detect automatable patterns " ,
    }),
    Task({
      subagent_type:  " session-learning-extractor " ,
      prompt:  " Extract learned rules/patterns " ,
    }),
    Task({
      subagent_type:  " session-followup-suggester " ,
      prompt:  " Suggest follow-up tasks " ,
    }),
  ]);

  // Phase 2: Sequential processing
  const finalResult = await Task({
    subagent_type:  " session-duplicate-checker " ,
    prompt: `
      Review the following Phase 1 results and resolve duplicates/conflicts:

      ## Doc Updates
      ${JSON.stringify(docUpdates)}

      ## Automations
      ${JSON.stringify(automations)}

      ## Learnings
      ${JSON.stringify(learnings)}

      ## Followups
      ${JSON.stringify(followups)}
    `,
  });

  // Output results
  displayWrapSummary(finalResult);
}

Output Format#

Progress#

╔════════════════════════════════════════════════════════════════╗
║  /cc-dev:session:wrap Progress                                        ║
╠════════════════════════════════════════════════════════════════╣
║                                                                ║
║  Phase 1: Parallel Analysis [████████████░░░░] 75%             ║
║                                                                ║
║  ✅ doc-updater: 3 document updates suggested                  ║
║  ✅ automation-scout: 2 automation patterns found              ║
║  🔄 learning-extractor: Analyzing...                           ║
║  ⏳ followup-suggester: Waiting                                ║
║                                                                ║
╚════════════════════════════════════════════════════════════════╝

On Completion#

╔════════════════════════════════════════════════════════════════╗
║  Session Wrap Complete                                         ║
╠════════════════════════════════════════════════════════════════╣
║                                                                ║
║  📝 Document Updates                                           ║
║  ├── CLAUDE.md: +15 lines (coding conventions added)           ║
║  ├── .claude/rules/bloc-patterns.md: +8 lines                 ║
║  └── .claude/commands/session/wrap.md: newly created           ║
║                                                                ║
║  🤖 Automation Suggestions                                     ║
║  ├── Recommend creating bloc-event-state-pair skill            ║
║  └── Recommend adding test-template auto-import                ║
║                                                                ║
║  📚 Learned Content                                            ║
║  ├── [Rule] Actively use dot shorthand                         ║
║  ├── [Rule] isClosed check required after await                ║
║  └── [Pattern] ZenHub issueTypeId must be specified            ║
║                                                                ║
║  📋 Follow-up Tasks                                            ║
║  ├── [High] Achieve 80% test coverage                          ║
║  └── [Medium] Optimize author search API                       ║
║                                                                ║
║  🔄 Duplicates removed: 2                                      ║
║  ⚠️ Conflicts resolved: 0                                      ║
║                                                                ║
╚════════════════════════════════════════════════════════════════╝

Target Files#

Analysis Targets#

Path	Description
`CLAUDE.md`	Project main guide
`.claude/rules/*/.md`	Coding rules
`.claude/commands/*/.md`	Skill/agent definitions
`.claude/settings.local.json`	Local settings

Output Targets#

Path	Description
`CLAUDE.md`	Convention, architecture updates
`.claude/rules/`	New rule file creation
`.claude/commands/`	New skill/agent creation
`.claude/session-logs/`	Session summary storage (optional)

Options#

Option	Default	Description
`--scope`	all	Analysis scope (all/rules/commands/agents)
`--dry-run`	false	Preview changes only
`--auto-commit`	false	Auto-commit changes
`--skip-duplicates`	false	Skip duplicate check

session-doc-updater - Document update analysis
session-automation-scout - Automation pattern detection
session-learning-extractor - Learned content extraction
session-followup-suggester - Follow-up task suggestions
session-duplicate-checker - Duplicate check and cleanup

Context Passing#

How sub-agents access session conversation content — read this before wiring up the 5-agent flow, since the wrong assumption here silently degrades every agent's output to whatever it can infer from repo files alone.

⚠️ Named subagent_types start with ZERO context — always pass it explicitly#

session-doc-updater / session-automation-scout / session-learning-extractor / session-followup-suggester / session-duplicate-checker are not general-purpose and do not auto-receive the conversation. Only subagent_type: "fork" inherits the calling session's full context; every other subagent_type — named or general-purpose — starts fresh with no memory of what happened in this session (confirmed against actual Task/Agent tool behavior, 2026-08). A prompt like "Extract rules learned during session" with no session content attached gives the agent nothing to analyze — it can only read files via Read/Glob/Grep, which is why these agents are scoped to those tools.

// ❌ WRONG — this agent has no idea what happened in the session
Task({
  subagent_type:  " session-learning-extractor " ,
  prompt:  " Extract rules learned during session " ,
});

// ✅ CORRECT — write a concrete summary of what happened and hand it over
Task({
  subagent_type:  " session-learning-extractor " ,
  prompt: `
    ## Session summary (this agent has no access to the conversation — this is the only source)

    ${sessionSummary}

    Extract learned rules/patterns from the above.
  `,
});

If you want an agent that genuinely inherits the conversation (e.g. for a quick one-off analysis instead of the 5-agent flow), use subagent_type: "fork" instead of a named session-* agent — that is the only path that gets automatic context.

Explicit Context Passing (the only correct method for the 5 named agents)#

Always pass session content explicitly — do not rely on it being inferred:

Task({
  subagent_type:  " session-learning-extractor " ,
  prompt: `
    ## Session Info to Analyze

    ### Modified Files
    - ${modifiedFiles.join( ' \n-  ' )}

    ### User Feedback
    - ${userFeedbacks.join( ' \n-  ' )}

    ### Errors Encountered
    - ${errors.join( ' \n-  ' )}

    Extract learned rules/patterns based on the above information.
  `,
});

Context Limitations#

Item	Accessible to the 5 named session-* agents	Notes
Current session conversation	X (unless explicitly written into the prompt)	Not auto-passed — see warning above
Previous session conversation	X	Isolated between sessions, even for `fork`
File system	O	Via Read, Glob, Grep tools
Git history	X	None of the 5 agents declare Bash in their `tools:` frontmatter (all are `Read, Glob, Grep`; `session-duplicate-checker` adds `Edit, Write`) — pass relevant `git log`/`git diff` output in the prompt if an agent needs it

Key Rules#

Non-destructive Updates: Prefer additions/modifications over deleting existing content
Evidence-based: State session-based evidence for all changes
Maintain Consistency: Maintain consistency with existing document style
Prevent Duplicates: Prevent duplicate content additions
User Confirmation: Apply important changes only after user confirmation

session-wrap

/cc-dev:session:wrap — 작업 마무리 정리#

한마디로#

누가·언제 쓰나요#

무엇을 해주나요#

어떻게 쓰나요#

안에서 무슨 일이 벌어지나요#

Triggers#

Usage#

Execution Flow (2 Phases)#

Phase 1: Parallel Analysis Agents#

1. doc-updater Agent#

2. automation-scout Agent#

3. learning-extractor Agent#

4. followup-suggester Agent#

Phase 2: Sequential Processing Agent#

5. duplicate-checker Agent#

Detailed Implementation#

Main Execution Logic#

Output Format#

Progress#

On Completion#

Target Files#

Analysis Targets#

Output Targets#

Options#

Related Agents#

Context Passing#

⚠️ Named subagent_types start with ZERO context — always pass it explicitly#

Explicit Context Passing (the only correct method for the 5 named agents)#

Context Limitations#

Key Rules#