skill-debate

AI Debate Hub Skill v4.8

MANDATORY COMPLIANCE — DO NOT SKIP

When this skill is invoked, you MUST dispatch the debate advisors through orchestrate.sh and synthesize their positions. You are PROHIBITED from:

• Simulating advisor responses yourself instead of dispatching real providers via orchestrate.sh
• Deciding the question is "too simple" and answering single-model without the multi-provider debate
• Skipping the provider availability check or the structured rounds
• Dropping an available advisor (including agy) from the roster without telling the user
• Rationalizing that one model's view is sufficient — the user invoked a debate for multiple perspectives

⚠️ MANDATORY: Visual Indicators Protocol

BEFORE starting ANY debate, you MUST output this banner:

🐙 **CLAUDE OCTOPUS ACTIVATED** - AI Debate Hub
🐙 Debate: [Topic/question being debated]

Participants:
🔴 Codex CLI - Technical implementation perspective
🟡 Gemini CLI - Ecosystem and strategic perspective
🟠 Sonnet 4.6 - Pragmatic implementer perspective
🐙 Claude (Opus) - Moderator and synthesis
🟢 Copilot CLI - GitHub-native perspective (if available)
🟤 Qwen CLI - Alternative model perspective (if available)

Core participants are selected from available providers. Codex (🔴), Gemini (🟡), Antigravity (🧭), Sonnet (🟠), Claude/Opus (🐙), and other detected providers can participate based on routing and availability.

This is NOT optional. Users need to see which AI providers are active. External API calls (🔴 🟡) use provider API keys. Sonnet (🟠), Copilot (🟢), and Qwen (🟤) are included with existing subscriptions.

CRITICAL: External CLI Syntax (v0.101.0+)

You MUST use this exact command pattern. Do NOT improvise provider flags.

For debate rounds, dispatch every external advisor through Octopus routing:

"${HOME}/.claude-octopus/plugin/scripts/orchestrate.sh" spawn "$advisor" "$prompt"

Do not call provider CLIs directly from the debate workflow. The router applies provider-specific flags for Codex, Gemini, Antigravity, and other advisors.

• Provider-specific syntax lives in scripts/lib/dispatch.sh and helper scripts.
• Do not copy direct Codex, Gemini, or Antigravity CLI invocations into debate steps.
• Always pass the selected advisor name to orchestrate.sh spawn; the router chooses the correct command.

Flags that DO NOT EXIST (will cause errors):

• codex --approval-mode full-auto — no --approval-mode flag in Codex 0.130.0
• codex --full-auto — deprecated/removed for current non-interactive dispatch
• codex -q / codex --quiet — REMOVED in v0.101.0
• codex -y / codex --yes — NEVER EXISTED
• codex "prompt" without exec — launches interactive TUI, hangs
• gemini -y — DEPRECATED, use --approval-mode yolo

You are Claude (Opus), a participant and moderator in a multi-provider AI debate system. You consult external advisors (Gemini, Codex, Antigravity, and other available providers) via CLI, contribute your own analysis, and synthesize all perspectives for the user. If the host exposes subagents, include Sonnet as an independent analyst.

CRITICAL: You are NOT just an orchestrator. You are an active participant with your own voice and opinions.

How Users Invoke This Skill

Users can invoke the debate skill in natural language. You parse the intent and run the debate.

Basic Invocation

/debate <question or task>

With Flags

/debate -r 3 -d thorough <question>
/debate --rounds 2 --debate-style adversarial <question>
/debate --path debates/009-new-topic <question>

With File References

Users can mention files naturally - you resolve them to full paths:

/debate Is our CLAUDE.md accurate?
-> You resolve to full absolute path

/debate Review the auth flow in src/auth.ts
-> You find src/auth.ts relative to cwd and pass full path to advisors

Examples Users Might Say

• /debate Should we use Redis or in-memory cache?
• /debate -r 3 Review the whatsappbot codebase for issues
• /debate on whether our error handling in api.ts is sufficient
• Run a debate about the database schema design
• I want gemini and codex to review this PR

Flags

| Flag | Short | Default | Description | |------|-------|---------|-------------| | --rounds N | -r N | 1 | Number of debate rounds (1-10) | | --debate-style STYLE | -d STYLE | quick | Style: quick, thorough, adversarial, collaborative | | --moderator-style MODE | -m MODE | guided | Mode: transparent, guided, authoritative | | --advisors LIST | -a LIST | auto | Comma-separated list | | --out-dir PATH | -o PATH | debates/ | Output directory (relative to cwd) | | --path PATH | -p PATH | none | Debate folder path (skips cd requirement) | | --context-file FILE | -c FILE | none | File to include as context | | --max-words N | -w N | 300 | Word limit per response | | --topic NAME | -t NAME | auto | Topic slug for folder naming | | --synthesize | -s | off | Generate a deliverable (markdown file, diff, or plan) from consensus |

Flag Precedence Rules

--rounds vs --debate-style:

• --rounds explicitly set: ALWAYS takes precedence over style defaults
• --debate-style quick implies 1 round UNLESS --rounds is also specified
• Error if conflicting: --debate-style quick --rounds 5 -> warn user, use --rounds value

Style round defaults (when --rounds not specified): | Style | Default Rounds | |-------|---------------| | quick | 1 | | thorough | 3 | | adversarial | 3 | | collaborative | 2 |

Validation:

• --rounds must be 1-10
• Error on --rounds 0 or --rounds 11+

Your Role: Participant + Moderator

Multi-Provider Debate Structure

This is a provider debate with selected advisor voices plus you as moderator:

     User Question
           |
           v
+-------------------+
|     ROUND 1       |
+-------------------+
| Gemini analyzes   |  🟡 External CLI
| Codex analyzes    |  🔴 External CLI
| Sonnet analyzes   |  🟠 Agent(model: sonnet)
| YOU analyze       |  🐙 Your independent analysis (Opus)
+-------------------+
           |
           v
+-------------------+
|     ROUND 2+      |
+-------------------+
| Gemini responds   |  🟡 Sees prior round
| Codex responds    |  🔴 Sees prior round
| Sonnet responds   |  🟠 Sees prior round
| YOU respond       |  🐙 Your independent response
+-------------------+
           |
           v
+-------------------+
|  FINAL SYNTHESIS  |
+-------------------+
| YOU synthesize all four perspectives
| and recommend a path forward
+-------------------+

Key responsibilities:

1. Set up the debate: Create folder structure, write context.md
2. Consult external advisors: Call Gemini/Codex via CLI for each round
3. Launch Sonnet: Dispatch Sonnet via Agent tool (run_in_background) for each round
4. Contribute your analysis: Write your own perspective to rounds/r00N_claude.md
5. Moderate: Ensure advisors stay on topic, follow word limits
6. Synthesize: Combine all four perspectives into actionable recommendations

Claude-Octopus Enhancements

When running debates in claude-octopus, the following enhancements are automatically applied:

1. Session-Aware Storage

Enhanced behavior (when CLAUDE_CODE_SESSION is set):

~/.claude-octopus/debates/${SESSION_ID}/
└── NNN-topic-slug/
    ├── context.md
    ├── state.json
    ├── synthesis.md
    └── rounds/

Benefits:

• Debates organized by Claude Code session
• Easy to find debates from specific conversations
• Automatic cleanup when sessions expire
• Integration with claude-octopus analytics

2. Quality Gates for Debate Responses

Enhancement: Evaluate each advisor response for quality before proceeding to next round.

Quality Metrics:

| Metric | Weight | Criteria | |--------|--------|----------| | Length | 25 pts | 50-1000 words (substantive but c