design

Conduct a design discussion using an agent team for the given task. All team discussions and inter-agent communication should be in English to optimize token usage. User-facing communication (interviews, final presentation) and saved documentation should be in Korean.

Workflow

Step 1: Requirements Interview & Codebase Exploration (Korean)

• Do NOT create a team yet. First, interview the user to understand requirements.
• Ask deep, non-obvious questions covering all aspects of the task, including but not limited to: technical implementation, UI/UX, concerns, tradeoffs, scale, constraints, and integration points.
• Avoid generic or superficial questions. Dig into specifics.
• Explore the existing codebase as needed during the interview to inform better questions and identify related patterns, modules, and conventions.
• Iterate between interviewing and codebase exploration until all critical aspects are sufficiently covered, then confirm with the user before proceeding.

Step 2: Team Composition Proposal (Korean)

• Based on interview results and codebase exploration, propose a team composition to the user.
• Specify each teammate's role, exploration scope, and model (use short aliases: "opus", "sonnet", "haiku").
• Only create the team after the user approves the composition.

Step 3: Design Discussion (English, internal only)

• Create the team with the approved composition.
• All inter-agent discussion and reasoning should be conducted in English.
• NO code modifications allowed. Design discussion only.
• The lead acts as a facilitator, actively driving multi-round discussion. Do NOT passively wait for results and move on.

Teammate Instructions:

• When assigning work to teammates, ALWAYS include this instruction: "When you send your result, start the message with [COMPLETE] if your analysis is finished, or [IN PROGRESS] if you need more time to explore. If [IN PROGRESS], briefly state what remains to be done."
• This signal is the ONLY reliable way to know if a teammate has finished. Do NOT infer completion from idle notifications or the mere arrival of a DM.

Discussion Protocol (minimum 2 full rounds):

1. Round 1 — Initial Proposals: Assign each teammate their design scope. Wait for ALL teammates to submit [COMPLETE] proposals before proceeding. If a teammate sends [IN PROGRESS], reply with "Take your time and send your complete analysis when ready" — do NOT move on.
2. Quality Gate: Before proceeding to cross-review, verify each proposal meets minimum depth:
   • References specific files, modules, or patterns from the actual codebase (not generic advice)
   • Includes concrete design decisions with rationale (not just listing options)
   • If a proposal is too shallow or generic, send it back with specific questions asking for deeper analysis
3. Cross-Review: Send each teammate's proposal to the other teammates for review. Explicitly ask them to identify gaps, risks, contradictions, and alternative approaches.
4. Round 2+ — Refinement: Collect cross-review feedback and send it back to the original authors. Ask them to address the feedback, revise their proposals, and respond to challenges. Repeat this cycle until convergence.
5. Convergence Check: After each round, the lead must assess whether open issues remain. Ask ALL teammates explicitly: "Do you have any remaining concerns or alternative proposals? Reply with [COMPLETE] No further input or [IN PROGRESS] with your remaining concerns." Only proceed to Step 4 when ALL teammates confirm [COMPLETE].

Facilitator Rules:

• Distinguish idle notifications from DMs: Messages marked with (idle) (e.g., [Teammate X (idle)]: ...) are system-generated summaries, NOT teammate DMs. Even if an idle notification contains words like "completed" or "finished", it is NOT a [COMPLETE] signal. ONLY count a response as received when the teammate sends an actual DM via SendMessage (shown as [Teammate X] sent DM to [team-lead] or similar) that starts with [COMPLETE] or [IN PROGRESS].
• Idle ≠ Done: A teammate going idle is normal — it does NOT mean they are done. Teammates may go idle while still processing (e.g., during sequential-thinking). Only trust the [COMPLETE] / [IN PROGRESS] signal in an actual DM. If a teammate goes idle without sending a DM, send them a follow-up asking for their status.
• Do NOT end discussion after the first response. Even if all teammates send [COMPLETE] in Round 1, you MUST proceed to cross-review (Round 2) at minimum.
• Actively cross-pollinate: When teammate A raises a point relevant to teammate B's scope, forward it and ask for their take.
• Surface disagreements: If teammates propose conflicting approaches, explicitly highlight the conflict and ask both sides to argue their position before the lead makes a judgment call.
• Track open issues: Maintain a running list of unresolved questions. The discussion is NOT complete until every item is either resolved or explicitly marked as a tradeoff to document.

Step 4: Result Synthesis & Documentation (Korean)

• The lead synthesizes all discussion results into a structured Korean document:
  • 합의된 아키텍처
  • 주요 결정사항과 근거
  • 미해결 이슈 / 트레이드오프
  • 권장 구현 순서
• Save the design document to the project's docs/ directory.
• Present the results to the user in Korean.

Step 5: Plan Refinement with User

• Discuss the design document with the user in Korean and refine the plan together.
• If the user raises questions that require deeper exploration, alternative comparison, or additional analysis, propose spinning up an agent team for that specific investigation.
• Only create a team if the user approves. Otherwise, continue refining in the current session.
• If a team is created, internal discussion should be conducted in English. Present the results to the user in Korean.
• After each round of refinement, update the saved document in docs/ to reflect the latest decisions.
• Repeat until the user is satisfied with the plan.

Constraints

• NO code modifications. Design discussion only.
• Inter-agent communication must be in English.
• Agent Team required: Steps that involve team creation and inter-agent discussion MUST use TeamCreate and SendMessage tools. Do NOT substitute with Agent tool sub-agents. The team discussion requires real-time inter-agent communication (debate, challenge, cross-validation) which is only possible through Agent Teams, not isolated sub-agents.
• Deferred tool loading: Before using AskUserQuestion, TeamCreate, SendMessage, or TeamDelete, you MUST first load them via ToolSearch. Run ToolSearch with query "select:AskUserQuestion", "select:TeamCreate", "select:SendMessage", and "select:TeamDelete" to load each tool. These are deferred tools and will NOT work unless loaded first. AskUserQuestion MUST be loaded before Step 1 (user interview).
• Team cleanup required: After all team discussion is complete and results are synthesized, you MUST gracefully shut down the team. Follow these steps IN ORDER — do NOT skip ahead:
  1. Send shutdown_request to each teammate via SendMessage (type: "shutdown_request").
  2. WAIT for ALL teammates to confirm shutdown (they respond with shutdown_response approve: true). Do NOT proceed to step 3 until every teammate has responded. If a teammate does not respond, retry the shutdown_request — repeat up to 10 times with a brief pause between attempts. NEVER forcefully kill (kill) agent processes.
  3. If a teammate still has not responded after 10 retries, use AskUserQuestion to inform the user which teammate(s) failed to shut down and ask them to handle it manually. Do NOT proceed to TeamDelete until resolved.
  4. Call TeamDelete to remove the team files and clean up resources. Only call this AFTER all teammates have confirmed shutdown or the user has handled unresponsive teammates.
  5. Verify process cleanup: Run ps aux | grep "team-name" | grep -v grep to check for orphan agent processes. If any remain, do NOT kill them — use AskUserQuestion to inform the user of the remaining PIDs and ask them to terminate the processes.
  • Shutdown failure fallback: If TeamDelete fails due to active teammates, do NOT use rm -rf or kill. Instead, use AskUserQuestion to inform the user of the failure and ask them to manually clean up (~/.claude/teams/{team-name} and ~/.claude/tasks/{team-name}).
  • If multiple teams are created during the workflow (e.g., in Step 5 refinement), clean up each team before creating the next one.

Task: $ARGUMENTS