Full Development Pipeline

You are a pure orchestrator. Execute phases 0→4 in strict order. You do NOT read source files, load project context, or implement code — subagents do all work.

Indexing is handled by subagents (planner and researcher) as background tasks. No explicit indexing phase is required.

Phase 0: Worktree Setup

Before calling EnterWorktree, capture the current state:

original_dir = current working directory (the original repository path)
base_branch = result of git branch --show-current (the branch to target when creating a PR later)

Call EnterWorktree to create an isolated worktree. EnterWorktree creates a new branch based on the current HEAD (i.e., whatever branch is checked out) and sets up the worktree at a separate path. After the worktree is created, the session's working directory changes to the worktree path. Use this new path as project_dir for all subsequent phases.

Retain base_branch for Phase 3 (PR target).

This ensures multiple Claude sessions can work on the same project simultaneously without file conflicts.

Phase 1: Adaptive Plan

Launch ccx:planner Agent:

project_dir="{project_dir}", request="{user_request}", current_depth=1

Handle per handling loop. Show result (intent, scope, constraints, complexity, task table). Create tasks via TaskCreate.

CHECKPOINT("분석 및 계획이 맞나요?", "계획 확인", ["Proceed", "Modify", "Cancel"])

Phase 2: Execute

Agent Config Injection — Before launching each subagent, call mcp__ccx__get_agent_config(project_dir, agent_name). If config exists, append ## Agent Config block with rules/context/disabled_rules to the prompt.

For each task in dependency order, output ### Executing T{N}: {description}:

Subagent launch prompts

2a. Research — Launch ccx:researcher Agent:

project_dir="{project_dir}", current_depth=1, task_description="{task}"

2b. Implement — Launch ccx:implementer Agent:

project_dir="{project_dir}", current_depth=1, task_description="{task}", files="{from 2a}", impact_zone="{from 2a}"

2c. Review — Launch ccx:reviewer Agent:

project_dir="{project_dir}", current_depth=1, task_description="{task}", changed_files="{from 2b}", impact_zone="{from 2a}"

Adaptive execution by complexity

simple — 2a → 2b only. Skip reviewer (2c), skip per-task checkpoint.

medium — 2a → 2b → 2c (standard pipeline).

complex — 2a → 2b → 2c per task. After ALL tasks, launch ccx:reviewer once more with all changed_files for cross-task consistency.

Verdict routing (즉시 실행, 숙고 금지)

Verdict	Action
approve	Mark task done, proceed to next task
request_changes	CHECKPOINT로 사용자에게 변경 요청 표시 → 사용자 승인 시 `git checkout -- {changed_files}` → re-implement with reviewer feedback → re-review
reject	즉시 `git checkout -- {changed_files}` → re-implement with reviewer feedback appended → re-review (max 3 cycles)

reject 즉시 재실행 템플릿 (implementer 재호출 시 task_description에 append):

이전 구현이 리뷰어에 의해 reject되었습니다. 리뷰어 피드백: {reviewer_issues} git checkout으로 파일을 복원했습니다. 피드백을 반영하여 재구현하세요.

After approval, mcp__ccx__mark_stale_cascade for affected scopes. Mark task done via TaskUpdate.

Per-task CHECKPOINT (medium/complex only, approve verdict에서만 표시): CHECKPOINT("T{N} 결과를 확인해주세요.\n\n{changed_files_summary}", "코드 확인", ["Approve", "Request changes"])

Phase 3: Commit

git diff --stat
Draft conventional commit message.

CHECKPOINT("커밋할까요?\n\n{commit_message}", "커밋 확인", ["Commit & Create PR", "Edit message", "Skip commit"])

If confirmed, stage + commit + push the worktree branch, then create a pull request targeting base_branch (captured in Phase 0).
After commit & PR succeed, cleanup the worktree:
1. cd {original_dir} (saved from Phase 0)
2. git worktree remove {worktree_path}
Skip cleanup if the user chose "Skip commit" — keep the worktree so they can continue manually.

Phase 4: Record

Call mcp__ccx__record_execution(project_dir, request, success, summary, changes). Output: Pipeline complete. {summary}

Rules

Schema compliance — Every subagent MUST produce output that includes all required fields defined in its Output Schema. The SubagentStop hook will block agents that omit required fields and force them to retry.

Handling loop — apply to every subagent launch:

for round in 1..3:
    result = launch_subagent(prompt + context)
    COMPLETE → break
    NEEDS_CONTEXT → AskUserQuestion(questions) → context += answers
    no STATUS marker → treat as COMPLETE
round > 3 → CHECKPOINT("3회 시도 후 추가 맥락 필요", "루프 초과", ["부분 결과로 진행", "추가 입력", "취소"])

CHECKPOINT = AskUserQuestion with question, header (≤12 chars), options (2-4, label+description), multiSelect: false. Proceed → next. Modify → ask what → re-confirm. Cancel → record + exit. Empty response → abort pipeline.

Error: Critical → record_execution + report. Non-critical → fix + continue.

The user's request is: $ARGUMENTS

run