team-shinchan-debate

name	team-shinchan:debate
description	Use when you want multiple agents to debate and find optimal solutions.
user-invocable	false

EXECUTE IMMEDIATELY

Step 0: Validate Input

If args empty: ask user for topic and STOP. If args > 2000 chars: truncate + warn.

All debates invoke Midori via Task tool.

Auto-Trigger Conditions

Trigger YES	Trigger NO
2+ approaches, architecture change, pattern break, perf vs readability, security, tech stack	Simple CRUD, clear bug fix, user already decided

On detection: announce "Design decision needed: [situation]", proceed Steps 1-3. Record decision in REQUESTS.md (Stage 1) or PROGRESS.md (Stage 2+).

Escalation: Tier 1 (this skill) vs Tier 2 (fierce-debate)

This skill (Midori, consensus-seeking, cheap, auto-triggerable) is Tier 1 and the default. For irreversible / high-stakes decisions, escalate to Tier 2 — team-shinchan:fierce-debate (a deterministic main-loop Workflow with a non-skippable refutation round + scored judge).

Stay on Tier 1 (this skill)	Escalate to fierce-debate
Reversible, low-stakes, simple 2-option	Irreversible: schema migration, public API contract, security boundary, data-loss
Quick consensus acceptable	A guaranteed refutation round is worth the cost
Auto-triggered, silent	Explicit user opt-in only (Workflow can't fire silently)

Auto-detect never silently jumps to Tier 2 — on a high-stakes detection, announce the design decision and offer the opt-in; the user launches /team-shinchan:fierce-debate. Both tiers write to the same .shinchan-docs/debate-decisions.md ledger (see agents/_shared/debate-decisions.md).

Competitive Code Mode Detection

Triggers: args contains any of: "경쟁 구현", "competitive", "best of N", "best of", "parallel implementation", "competitive-code"

If competitive-code trigger detected:

1. Parse: N = number of implementations (default 2, max 3). Extract from "best of N" pattern or use default.
2. Extract implementation request = args after trigger keyword removal.
3. Proceed to Step 1-CC (Competitive Code routing).

Tier 2 escalation: this Midori/worktree path is the cheap default. For a high-value contested implementation where you want a deterministic, schema-validated tournament (read-only builders → scored judge → winner picked by score, applied + tested), escalate to team-shinchan:fierce-compete (opt-in main-loop Workflow). Never silently jump to it; offer the opt-in.

Step 1-CC: Invoke Midori for Competitive Code

Task(
  subagent_type="team-shinchan:midori",
  model="sonnet",
  prompt=`Competitive Code mode invoked.

## Request
${implementation_request}

## Parameters
- N implementations: ${N} (max 3)
- Mode: competitive-code

## Instructions
Run the competitive-code workflow:
1. ToolSearch for EnterWorktree schema before starting (HR-3)
2. Set .current-agent to midori before parallel Task calls (HR-1)
3. Dispatch N Bo subagents each in an isolated worktree via EnterWorktree
4. Collect implementation reports from all Bo agents
5. Dispatch Action Kamen to judge all implementations using rubric scoring (FR-2.1)
6. Announce winner and score comparison table
7. Merge winner worktree to main, ExitWorktree all others
8. Guarantee cleanup on error (NFR-3)
9. Record decision in .shinchan-docs/debate-decisions.md

See agents/midori.md ## Competitive Code Workflow for full procedure.`
)

Step 2-CC: Deliver Competitive Code Results

After Midori returns:

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
🏆 Competitive Code Results
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
📋 Request: {implementation_request}
👥 Implementations: {N}
🥇 Winner: Bo-{N} | Score: {score}/15
📊 Score Comparison: [all Bo scores]
✅ Merged: winner changes applied
🧹 Cleanup: all worktrees removed
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Step 3-CC: Confirm

Ask user if they accept the winner's implementation. If yes: proceed to completion. If no: offer to run competitive-code again or accept a specific implementation.

STOP. Steps 1-CC through 3-CC handle the full competitive-code workflow. Do NOT proceed to Step 1 (standard debate) when competitive-code mode was detected.

---

Step 1: Invoke Midori

Task(subagent_type="team-shinchan:midori", model="sonnet",
  prompt="Debate topic: {topic}\nPanel: {panel list}\nProcedure: Announce, collect opinions (parallel), ALWAYS run a separate refutation round (even on apparent agreement), then a separate scored judge picks the winner with surviving dissent, report decision.")

Step 2: Deliver Results

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
💭 Debate Results
📋 Topic: {topic}
🎤 Opinions: [Agent]: {summary} ...
✅ Decision: {conclusion} | 📝 Rationale: {reasoning}
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Step 3: Confirm + Finalize

Ask user if they agree. If yes: document and proceed. If no: revise reflecting concerns. Never proceed without confirmation.

• After decision recording, Midori archives the full debate transcript to .shinchan-docs/debates/DECISION-{NNN}.json for traceability

Panel Selection

See agents/midori.md for full criteria. Quick reference:

Topic	Panel
Frontend/UI	Aichan, Hiroshi
Backend/API	Buriburi, Hiroshi
DevOps/Infra	Masao, Hiroshi
Architecture	Hiroshi, Nene, Misae
Security	Hiroshi, Buriburi, Masao

Auto-Detection Signals

Keywords: "A or B", "vs", "method1/method2", "schema change", "layer", "structure", "differs from existing pattern", "trade-off", "at the cost of", "auth", "encryption", "permission"