claude-pipeline

한국어

An AI development pipeline built on Claude Code custom skills.

Run /pipeline and a lead agent automatically dispatches specialized agents in sequence — from research through implementation, review, bugfix, and retrospective.

How It Works

/pipeline add payment integration

→ Lead Agent (Pipeline) assesses the situation
  → Web Researcher (haiku)  — web search + information gathering
  → Code Analyzer (sonnet)  — codebase analysis + synthesis
  → Planner (opus)          — task decomposition + design
  → Implementer (sonnet)    — code implementation
  → Reviewer (opus)         — build/test + code review
  → Bugfixer (sonnet)       — issue resolution
  → Retrospective           — retrospective + skill self-improvement

Each agent runs in an isolated subprocess within a git worktree, keeping the main branch untouched. Agents communicate through markdown files in the .pipeline/ directory.

Agent Team

Agent	Role	Model	Specialty
Web Researcher	Web search + info gathering	haiku	Fast, low-cost research
Code Analyzer	Codebase analysis + synthesis	sonnet	Pattern recognition, direction proposals
Planner	Task decomposition + design	opus	Architecture, deep reasoning
Implementer	Code implementation	sonnet	Code writing, speed/cost balance
Reviewer	Code review + build/test	opus	Thorough review, quality focus
Bugfixer	Issue resolution	sonnet	Targeted fixes, speed focus

Key Features

Lead Agent's Autonomous Decisions

The Decision Matrix is a default guideline — the lead agent can skip or adjust steps based on the situation.

Priority: User Override > Autonomous Judgment > Decision Matrix (Default)

• Refactoring, internal logic changes → research can be skipped
• New tech adoption, external API integration → research required
• User says "skip research" → always respected

Git Worktree Isolation

Every /pipeline run creates a dedicated git worktree and branch, so the main branch stays clean until you're ready to merge.

.worktrees/
  login-feature/          ← branch: pipeline/login-feature
    .pipeline/
      my-app/
        login-feature/
          research.md
          plan.md
          progress.md
          review.md
          retrospective.md
    src/
    ...

After the pipeline completes, review and merge at your own pace:

cd .worktrees/login-feature
git add -A && git commit -m "feat: login feature"
git checkout main
git merge pipeline/login-feature
git worktree remove .worktrees/login-feature

Build & Test Verification

The Reviewer doesn't just read diffs — it actually runs builds and tests.

Implementation Progress Tracking

The Implementer updates progress.md after completing each task.

Skills

Skill	Description	Standalone
/pipeline	Full pipeline (lead agent dispatches as needed)	Yes
/research	Web research + codebase analysis	Yes
/plan	Create implementation plan	Yes
/implement	Plan-based implementation	Yes
/review	Code review + build/test	Yes
/bugfix	Fix issues from review	Yes
/retrospective	Pipeline retrospective + skill self-improvement	Yes

Each skill can be used independently without /pipeline.

Installation

Install directly in Claude Code:

/plugin marketplace add sunmerrr/claude-pipeline
/plugin install claude-pipeline@sunmerrr-claude-pipeline

Recommended Permission Settings

To reduce permission prompts during pipeline execution, add to ~/.claude/settings.local.json:

{
  "permissions": {
    "allow": [
      "Bash(ls *)", "Bash(pwd)", "Bash(basename *)",
      "Bash(mkdir *)", "Bash(cat *)", "Bash(head *)", "Bash(tail *)",

      "Bash(git status*)", "Bash(git diff*)", "Bash(git log*)",
      "Bash(git branch*)", "Bash(git show*)", "Bash(git rev-parse*)",

      "Bash(npm run *)", "Bash(npm test*)", "Bash(npx *)",
      "Bash(bun run *)", "Bash(bun test*)", "Bash(bunx *)",

      "Bash(tsc *)", "Bash(eslint *)", "Bash(prettier *)",
      "Bash(python -m pytest*)", "Bash(python -c *)",

      "WebSearch"
    ]
  }
}

Usage Examples

# Full pipeline
/pipeline add user authentication with OAuth2

# Skip research
/pipeline add notification icon to sidebar
# → Lead agent may autonomously skip research

# Individual skills
/research WebSocket real-time notifications
/plan
/implement
/review
/bugfix

Artifact Flow

Web Researcher → web-research.md
                        ↓
Code Analyzer  → research.md (integrates web-research.md)
                        ↓
Planner        → plan.md (references research.md)
                        ↓
Implementer    → code implementation + progress.md
                        ↓
Reviewer       → review.md (includes build/test results)
                        ↓
Bugfixer       → code fixes (references review.md)
                        ↓
Retrospective  → retrospective.md

---

Contributing

Contributions are welcome! Whether you want to fix a bug, improve an existing skill, or add a new one, please read CONTRIBUTING.md for guidelines.

---

License

MIT License - see LICENSE for details.