| name | team-dev |
| description | Full team pipeline — plan + architect in parallel, implement with worker agents, then parallel review |
You are orchestrating a full team development pipeline. YOU (the main session) are the orchestrator — you spawn and wait for subagents. Execute every phase in order without asking for confirmation unless genuinely ambiguous.
Task: the task/scope the user described when invoking this skill (if none given, ask or infer from context)
Concurrency & nesting limits: at most 6 subagents run concurrently; nesting depth is 1, so spawned subagents CANNOT spawn their own subagents. That is why the main session orchestrates directly rather than delegating orchestration to a spawned team-lead. (The team-lead agent persona is still available as a spawnable coordinator for advice, but it will not spawn others.)
Phase 1: Analysis (PARALLEL)
Spawn in parallel and wait for BOTH to finish:
planner — break the task into discrete, dependency-ordered, testable tasks.architect — analyze structure, patterns, and integration points; flag risks.
Read both reports and reconcile any conflicts into a single plan.
Phase 2: Design (UI 작업 포함 시에만)
Skip for backend-only, bug-fix, or refactor tasks with no UI change.
design-system/MASTER.md가 있으면 로드해 재사용 (필요 시 page override 추가).- 없으면
~/.codex/skills/ui-ux-pro-max/ (reasoning-rules.md → styles.md)로 디자인 토큰을 확정하고 design-system/MASTER.md를 생성.
- 구현은 MASTER 토큰 100% 준수 (raw hex 금지, 시맨틱 토큰만, 이모지 구조 아이콘 금지, 터치 타겟 ≥44pt).
Phase 3: Implement
- Foundation first: types/interfaces → logic → UI, following existing conventions.
- If the plan has 3+ independent components, spawn one
worker subagent per component to implement them in parallel — but give each a distinct, non-overlapping set of files so no two agents ever edit the same file. Respect the 6-concurrent cap (batch if there are more than 6). Wait for all workers before proceeding.
- Dependent components stay sequential (do the shared foundation yourself first, then fan out).
Phase 4: Verify
Run in order and auto-fix failures (up to 3 attempts): build, lint/type-check, tests. For UI changes also run the /design-audit checks.
Phase 5: Review (PARALLEL)
Spawn in parallel and wait for ALL:
code-reviewer — quality pass.security-reviewer — security audit.qa-tester — ONLY if the change includes UI (browser + accessibility).
Merge findings. Fix every CRITICAL and WARNING, then re-run Phase 4 verification.
Phase 6: Finalize
Summarize: files changed/created, key decisions, review findings addressed (by severity), and what to test manually. Commit is NOT included — the user decides when to commit.
Rules
- Do NOT ask for permission between phases — just execute.
- Spawn independent agents in a single batch so they run concurrently (max 6 at once).
- If genuinely ambiguous (multiple valid approaches), ask once then proceed.
