[HINT] Download the complete skill directory including SKILL.md and all related files
| name | init-plan |
| description | Create a detailed implementation plan based on actual codebase |
| argument-hint | <task description> [--qa quick|light|standard|thorough|adversarial] |
Caller note: planning benefits from
highorxhigheffort; lower effort may miss call sites in cross-file analysis.
Check if a similar plan already exists in .claude_reports/plans/. Behavior depends on plan status:
active: Always ask the user — "기존에 진행 중인 plan이 있습니다. 이어서 진행할까요, 새로 만들까요?" Do NOT proceed until confirmed.done/failed: Note it for reference and auto-proceed with new plan creation (no prompt).partial: Auto-create a new plan covering only the failed steps (read failed_steps from plan frontmatter). No prompt.Record any user-facing pause for the pipeline_summary.md Decision Points table.
Invoke the plan-team (기획팀) agent as a subagent with the following prompt:
Plan mode. Create a new implementation plan.
Task: {$ARGUMENTS}
Save English plan to: .claude_reports/plans/{YYYY-MM-DD}_{short-task-name}/plan/plan.md
Date: {YYYY-MM-DD}
{If a done/failed/partial plan exists: "Reference previous plan: [path], status: [status]"}
{If partial: "Failed steps from previous execution: [list from plan frontmatter failed_steps]"}
Read all relevant source files, analyze the current state, and create the plan.
Write the plan files directly. Return ONLY the file paths and a 3-5 line Korean summary. Do NOT return the plan content itself.
The agent writes the plan file directly; the orchestrator only receives paths and a summary.
If $ARGUMENTS contains --qa quick|light|standard|thorough|adversarial, use that level and strip the flag from the task description. Otherwise, auto-detect from the plan's scope. When qa_level is set in plan frontmatter, it overrides auto-detect.
| Level | Auto-detect condition | Action |
|---|---|---|
| Quick | (manual only — never auto-selected) | 1× 품질관리팀 (model: "sonnet"), single pass, max 1 review round (no iteration even if 🔴 found — 🔴 are recorded as 미해결 이슈 and loop exits) |
| Light | ≤3 steps, mechanical, single-variant | 1× 품질관리팀 (model: "sonnet") |
| Standard | 4-10 steps, logic changes, single module | 1× 품질관리팀 (default opus) |
| Thorough | >10 steps, cross-module/variant, architectural | 2-3× 품질관리팀 in parallel: Agent A correctness (opus), B completeness (sonnet), C risk (opus, optional, >15 steps); each writes round_{N}_{focus}.md; all 🔴 issues must be resolved |
| Adversarial | Cross-variant (SE+SS+CSS), shared modules (utils/, network.py), or >20 steps with architectural impact — AND Codex available | Thorough-level 품질관리팀 + 1× codex-review-team (adversarial-review) in parallel; Codex writes round_{N}_codex.md; all 🔴 from ANY agent (including Codex) must be resolved |
Codex availability check: Before selecting Adversarial, run codex --version (suppress stderr). If the command fails or Codex is not authenticated, fall back to Thorough silently. This check is skipped if --qa adversarial is explicitly specified (fail loudly instead).
The log directory is the task root folder (parent of plan/). Example: .claude_reports/plans/2026-03-18_task/plan/plan.md → log dir is .claude_reports/plans/2026-03-18_task/. Run mkdir -p {log_dir}/_internal/plan_reviews before invoking QA.
Round counting: Initialize round = 0. A round = one plan-team fix → QA review cycle; all parallel Thorough agents count as one round. Increment round only when QA is re-invoked after a revision. "max 3 rounds" means 기획팀 is invoked at most 3 times to fix issues. quick mode: max rounds = 1 — after the single review pass, exit regardless of 🔴 (record residuals as 미해결 이슈 and skip the fix-round).
QA level lock: QA level is determined once at loop start; only upward escalation allowed (no downgrade). If --qa was NOT specified, the orchestrator MAY upgrade once (starting round 2) when 🔴 count ≥3 in the one-line verdict (no review file reading needed); round counter does NOT reset. If --qa was manually specified, no change allowed. quick is never auto-upgraded (user opted in for fastest path).
After the 기획팀 agent returns:
model: 'sonnet'. Thorough: 2-3 parallel agents with focus suffix and separate output files; pass model: 'sonnet' for the B (completeness) agent, default opus for A (correctness) and C (risk). Do NOT read the review file unless relaying verdict to user.round, re-invoke 품질관리팀. Repeat until no 🔴 or round >= 3.round >= 3: Auto-proceed — invoke 기획팀: "Refine mode. Add remaining 🔴 issues to the plan's 리스크 section under ## 미해결 이슈. Return brief Korean summary." Then report to user: plan path, resolved issues, and unresolved issues with reasons.Record any user-facing pause (e.g., active-plan ambiguity) so the pipeline skill can surface it in pipeline_summary.md.
After the review loop completes, invoke 기획팀 one final time:
Translate mode. English plan file: {plan_path}. Save Korean version to: {same directory}/plan_ko.md.
Full Korean translation (NOT a summary). Same detail level. Section titles: 목표, 현황 분석, 변경 계획, 리스크, 검증 방법. Code identifiers stay in English. Return ONLY the file path.
Then report to the user: English plan path, Korean plan path, plan summary, and QA verdict.
$ARGUMENTS