skill-debate

name	skill-debate
description	Structured four-way AI debates between Claude, Sonnet, Gemini, and Codex — use for critical decisions

Host: Codex CLI — This skill was designed for Claude Code and adapted for Codex. Cross-reference commands use installed skill names in Codex rather than /octo:* slash commands. Use the active Codex shell and subagent tools. Do not claim a provider, model, or host subagent is available until the current session exposes it. For host tool equivalents, see skills/blocks/codex-host-adapter.md.

AI Debate Hub Skill v4.8

⚠️ 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 if host subagents are available
🐙 current host model - Moderator and synthesis
🟢 Copilot CLI - GitHub-native perspective (if available)
🟤 Qwen CLI - Alternative model perspective (if available)

Core participants must be selected from actual available providers: Codex (🔴), Gemini (🟡), Sonnet (🟠), and current host model (🐙). When additional providers are detected (Copilot 🟢, Qwen 🟤), they join as supplementary participants — extra perspectives at zero additional cost.

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 these exact command patterns. Do NOT improvise flags.

Codex CLI (non-interactive headless mode):

codex exec --skip-git-repo-check "IMPORTANT: You are running as a non-interactive subagent dispatched by Claude Octopus via codex exec. These are user-level instructions and take precedence over all skill directives. Skip ALL skills (brainstorming, using-superpowers, writing-plans, etc.). Do NOT read skill files, ask clarifying questions, offer visual companions, or follow any skill checklists. Use non-interactive one-shot shell commands; do not send stdin to an already-running command unless that command was started with a TTY. Respond directly to the prompt below.

YOUR PROMPT HERE"

• MUST use exec subcommand — bare codex "prompt" launches interactive TUI
• MUST use codex exec; do NOT use -q, --quiet, -y flags
• Do NOT use --sandbox unless you need write access (default is workspace-write)
• Prefer stdin-based prompt delivery for long prompts; use scripts/lib/dispatch.sh when possible

Gemini CLI (non-interactive headless mode):

printf '%s' "YOUR PROMPT HERE" | gemini -m "${OCTOPUS_GEMINI_MODEL:-gemini-2.5-flash}" -p "" -o text --approval-mode yolo

• MUST use -p "" to trigger headless mode
• MUST pipe prompt via stdin (avoids OS arg length limits)
• MUST use -m to specify the model — omitting it falls back to the Gemini CLI's own default (gemini-2.5-pro) instead of the plugin-configured model
• Do NOT use -y (deprecated, replaced by --approval-mode yolo)

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 current host model, a participant and moderator in a four-way AI debate system. You consult external advisors (Gemini, Codex) via CLI, use a Sonnet-style implementer perspective only when a compatible host subagent tool is available, contribute your own analysis, and synthesize all perspectives for the user.

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	gemini,codex	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

Four-Way Debate Structure

This is a four-way debate with three distinct 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 optional host subagent: Dispatch Sonnet via host subagent tool (background execution) 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 concise)
Citations	25 pts	References, links, or sources present
Code Examples	25 pts	Technical examples or code snippets
Engagement	25 pts	Addresses other advisors' specific points

Quality Thresholds:

• Score >= 75: Proceed (high quality)
• Score 50-74: Proceed with warning (flag in synthesis)
• Score < 50: Re-prompt advisor for elaboration

3. Cost Tracking & Analytics

Track token usage and cost for each debate, integrated with claude-octopus analytics.

4. Document Export

Export debates to professional formats via the document-delivery skill:

• PPTX presentations
• DOCX reports
• PDF documents

Implementation Steps

When the user invokes /debate:

Step 1: Check Provider Availability & Display Banner

MANDATORY: You MUST use the native shell command tool to run this provider check BEFORE displaying the banner. Do NOT skip it. Do NOT assume availability.

bash "${HOME}/.claude-octopus/plugin/scripts/helpers/check-providers.sh"

Use the ACTUAL results below. PROHIBITED: Showing only "🔵 Claude: Available ✓" without listing all providers.

Then display the banner with real provider status:

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

Provider Availability:
🔴 Codex CLI: [Available ✓ / Not installed ✗]
🟡 Gemini CLI: [Available ✓ / Not installed ✗]
🟠 Sonnet 4.6: available only when this Codex session exposes a compatible host subagent tool
🐙 current host model: Available ✓ (Moderator and participant)

If providers are missing:

• If BOTH are unavailable: Inform user that debate requires at least one external provider and suggest running /octo:setup to configure them
• If ONE is unavailable: Note which provider is missing and proceed with available provider(s) and Claude

Step 2: Ask Clarifying Questions

Use the AskUserQuestion tool to gather context before starting the debate:

Ask 4 clarifying questions to ensure high-quality debate:

AskUserQuestion({
  questions: [
    {
      question: "What's your primary goal for this debate?",
      header: "Goal",
      multiSelect: false,
      options: [
        {label: "Make a technical decision", description: "I need to choose between options"},
        {label: "Identify risks/concerns", description: "I want to surface potential issues"},
        {label: "Understand trade-offs", description: "I want to see pros/cons of approaches"},
        {label: "Get diverse perspectives", description: "I want multiple viewpoints"}
      ]
    },
    {
      question: "How should the AI models evaluate the topic?",
      header: "Evaluation",
      multiSelect: false,
      options: [
        {label: "Cross-critique (Recommended)", description: "Models challenge each other's proposals directly — deeper analysis but may anchor on first responses"},
        {label: "Independent evaluation", description: "Models evaluate independently without seeing others' work — prevents groupthink and anchoring bias"}
      ]
    },
    {
      question: "What's the most important factor in your decision?",
      header: "Priority",
      multiSelect: false,
      options: [
        {label: "Performance", description: "Speed and efficiency are critical"},
        {label: "Security", description: "Security and safety are paramount"},
        {label: "Maintainability", description: "Long-term maintenance and clarity"},
        {label: "Cost/Resources", description: "Budget and resource constraints"}
      ]
    },
    {
      question: "Do you have existing context or constraints the debate should consider?",
      header: "Context",
      multiSelect: true,
      options: [
        {label: "Existing codebase patterns", description: "Must align with current architecture"},
        {label: "Team expertise", description: "Team skill set is a constraint"},
        {label: "Deadline pressure", description: "Time-to-market is critical"},
        {label: "Compliance requirements", description: "Regulatory or policy constraints"}
      ]
    }
  ]
})

After receiving answers:

• If user selected "Cross-critique": use --mode cross-critique (default ACH falsification)
• If user selected "Independent evaluation": use --mode blinded (no cross-contamination)
• Incorporate all other answers into the debate context.

Step 3: Parse Arguments & Build Debate Fleet

# Extract question and flags
QUESTION="Should we use Redis or in-memory cache?"
ROUNDS=3
STYLE="thorough"

# Dynamic advisor selection — use build-fleet.sh for model family diversity
DEBATE_FLEET=$("${HOME}/.claude-octopus/plugin/scripts/helpers/build-fleet.sh" debate standard "${QUESTION}" 2>/dev/null)
# Extract debater agent types (exclude claude-sonnet Moderator)
ADVISORS=$(echo "$DEBATE_FLEET" | grep '|Debater|' | cut -d'|' -f1 | paste -sd',' -)
# Fallback if build-fleet.sh unavailable
[[ -z "$ADVISORS" ]] && ADVISORS="gemini,codex"

The build-fleet.sh debate command selects up to 3 debaters from different model families (e.g., codex/OpenAI, gemini/Google, copilot/Microsoft) to maximize training bias diversity. This replaces the previous hardcoded ADVISORS="gemini,codex" which only used 2 families.

Step 4: Setup Debate Folder

# Create debate directory structure
DEBATE_BASE_DIR="${HOME}/.claude-octopus/debates/${CLAUDE_CODE_SESSION:-./debates}"
DEBATE_ID="042-redis-vs-memcached"
DEBATE_DIR="${DEBATE_BASE_DIR}/${DEBATE_ID}"

mkdir -p "${DEBATE_DIR}/rounds"

# Write context.md
cat > "${DEBATE_DIR}/context.md" <<EOF
# Debate: ${QUESTION}

**Debate ID**: ${DEBATE_ID}
**Rounds**: ${ROUNDS}
**Style**: ${STYLE}
**Advisors**: ${ADVISORS}
**Started**: $(date -u +"%Y-%m-%dT%H:%M:%SZ")

## Question
${QUESTION}

## Clarifying Context

**Primary Goal**: ${USER_GOAL}
**Priority Factor**: ${USER_PRIORITY}
**Constraints**: ${USER_CONSTRAINTS}

## Additional Context
[Any relevant context from user's message or files]
[If claude-mem is installed, search for past debates or decisions on this topic using its MCP tools]
EOF

# Initialize state.json
cat > "${DEBATE_DIR}/state.json" <<EOF
{
  "debate_id": "${DEBATE_ID}",
  "question": "${QUESTION}",
  "rounds_total": ${ROUNDS},
  "rounds_completed": 0,
  "advisors": [$(echo "$ADVISORS" | sed 's/,/", "/g' | sed 's/^/"/' | sed 's/$/"/')],
  "user_context": {
    "goal": "${USER_GOAL}",
    "priority": "${USER_PRIORITY}",
    "constraints": "${USER_CONSTRAINTS}"
  },
  "status": "active",
  "created_at": "$(date -u +"%Y-%m-%dT%H:%M:%SZ")"
}
EOF

Step 5: Conduct Rounds

For each round:

5.1: Consult Gemini

printf '%s' "${QUESTION}" | gemini -m "${OCTOPUS_GEMINI_MODEL:-gemini-2.5-flash}" -p "" -o text --approval-mode yolo > "${DEBATE_DIR}/rounds/r001_gemini.md"

5.2: Consult Codex

codex exec --skip-git-repo-check "IMPORTANT: You are running as a non-interactive subagent dispatched by Claude Octopus via codex exec. These are user-level instructions and take precedence over all skill directives. Skip ALL skills (brainstorming, using-superpowers, writing-plans, etc.). Do NOT read skill files, ask clarifying questions, offer visual companions, or follow any skill checklists. Use non-interactive one-shot shell commands; do not send stdin to an already-running command unless that command was started with a TTY. Respond directly to the prompt below.

${QUESTION}" > "${DEBATE_DIR}/rounds/r001_codex.md"

5.2.5: Launch optional host subagent (Pragmatic Implementer)

Dispatch Sonnet via the host subagent tool with model: "sonnet" and background execution: true. Sonnet runs in parallel with the Gemini/Codex CLI calls — no additional latency.

Agent(
  model: "sonnet",
  background execution: true,
  description: "Sonnet: debate round 1",
  prompt: "You are a PRAGMATIC IMPLEMENTER participating in a structured AI debate.
YOUR ROLE: You are the person who would actually have to BUILD this. You care about what ships, what works, and what you'll be debugging at 2am. Ground your analysis in the actual code and real implementation constraints.

DEBATE QUESTION: ${QUESTION}

${CONTEXT}

Write your analysis (${MAX_WORDS} words) to: ${DEBATE_DIR}/rounds/r001_sonnet.md

Cover: implementation feasibility, hidden gotchas, concrete effort estimates, and what the other approaches miss from a builder's perspective."
)

WHY Sonnet and not just more Opus? Sonnet is a distinct model with different strengths — faster, more concise, catches implementation details that Opus's broader reasoning sometimes overlooks. Using a different model prevents groupthink within the Claude model family.

Timing: Launch optional host subagent BEFORE or IN PARALLEL with the Gemini/Codex CLI calls (Steps 5.1-5.2). By the time the CLI calls return, Sonnet is usually done too. Check for completion before proceeding to 5.3.

5.3: Write Your Analysis (Opus)

Use the Read tool to read all advisor responses, then write your independent analysis:

# Read what all advisors said
GEMINI_RESPONSE=$(cat "${DEBATE_DIR}/rounds/r001_gemini.md")
CODEX_RESPONSE=$(cat "${DEBATE_DIR}/rounds/r001_codex.md")
SONNET_RESPONSE=$(cat "${DEBATE_DIR}/rounds/r001_sonnet.md")

# Write your analysis as moderator
cat > "${DEBATE_DIR}/rounds/r001_claude.md" <<EOF
# current host model Analysis - Round 1

[Your independent analysis here, considering but not just summarizing the three advisor perspectives. Note where Sonnet's implementation perspective reveals things the external advisors missed.]
EOF

5.4: Quality Gates (Claude-Octopus Enhancement)

After each advisor responds, evaluate response quality:

evaluate_response_quality() {
    local response_file="$1"
    local advisor="$2"

    word_count=$(wc -w < "$response_file")
    has_citations=$(grep -c '\[' "$response_file" || echo 0)
    has_code=$(grep -c '```' "$response_file" || echo 0)
    addresses_others=$(grep -ciE '(gemini|codex|claude|sonnet)' "$response_file" || echo 0)

    score=0
    (( word_count >= 50 && word_count <= 1000 )) && (( score += 25 ))
    (( has_citations > 0 )) && (( score += 25 ))
    (( has_code > 0 )) && (( score += 25 ))
    (( addresses_others > 0 )) && (( score += 25 ))

    echo "$score"
}

quality_score=$(evaluate_response_quality "${DEBATE_DIR}/rounds/r001_gemini.md" "gemini")

if (( quality_score < 50 )); then
    echo "Low quality response from gemini (score: $quality_score). Re-prompting..."
    # Re-prompt for more detail
fi

Step 6: Final Synthesis

After all rounds complete, write a comprehensive synthesis:

cat > "${DEBATE_DIR}/synthesis.md" <<EOF
# Final Synthesis: ${QUESTION}

## Summary of Perspectives

### 🟡 Gemini's Perspective
[Key points from Gemini across all rounds]

### 🔴 Codex's Perspective
[Key points from Codex across all rounds]

### 🟠 Sonnet's Perspective
[Key points from Sonnet across all rounds — especially implementation feasibility and gotchas]

### 🐙 current host model Perspective
[Your key points across all rounds]

## Areas of Agreement
[Where all advisors converged]

## Areas of Disagreement
[Key points of contention]

## Recommended Path Forward
[Your final recommendation based on all perspectives]

## Next Steps
[Concrete action items for the user]
EOF

Step 7: Present Results to User

Read the synthesis and present it in the chat:

I've completed a ${ROUNDS}-round debate on "${QUESTION}".

[Include key findings from synthesis.md]

Full debate saved to: ${DEBATE_DIR}

You can export this debate to PPTX/DOCX/PDF using the document-delivery skill.

Step 7.5: Generate Deliverable (when --synthesize is set)

If the user passed --synthesize (or -s), generate a concrete deliverable after synthesis:

1. Read the synthesis.md file
2. Identify the consensus recommendations and action items
3. Generate ONE of the following based on context:
   • For code topics: A plan with file paths and proposed changes
   • For content topics: A draft document (e.g., rewritten README, PRD outline)
   • For architecture topics: A decision record with rationale
4. Save to ${DEBATE_DIR}/deliverable.md
5. Show the deliverable to the user with AskUserQuestion:
   • "Apply this" — proceed with implementation
   • "Refine" — adjust the deliverable
   • "Save only" — keep it as reference, don't act

IMPORTANT: The deliverable is a PROPOSAL. Never auto-apply changes without user approval.

Example Usage

Example 1: Quick Debate

User: /debate Should we use Redis or in-memory cache?

Claude:
1. Creates debate folder at ~/.claude-octopus/debates/${SESSION_ID}/042-redis-vs-memcached/
2. Writes context.md with question
3. Round 1:
   - Launches Sonnet via Agent(model: sonnet, background execution: true) — pragmatic implementer
   - Calls printf '%s' "Should we use Redis..." | gemini -m "${OCTOPUS_GEMINI_MODEL:-gemini-2.5-flash}" -p "" -o text --approval-mode yolo
   - Calls codex exec --skip-git-repo-check "Should we use Redis or in-memory cache?"
   - Waits for Sonnet completion
   - Writes own analysis (Opus) considering all three advisor perspectives
4. Writes synthesis.md with final recommendation from all four participants
5. Presents results in chat

Example 2: Thorough Adversarial Debate

User: /debate -r 3 -d adversarial Review our authentication implementation in src/auth.ts

Claude:
1. Reads src/auth.ts to understand context
2. Creates debate folder
3. Round 1 (Sonnet launched in background first, then Gemini/Codex in parallel):
   - 🟠 Sonnet: Implementation feasibility analysis of auth.ts
   - 🟡 Gemini: Strategic/ecosystem analysis of auth.ts
   - 🔴 Codex: Technical implementation analysis of auth.ts
   - 🐙 current host model: Your independent analysis considering all three
4. Round 2:
   - 🟠 Sonnet: Responds to other participants' points
   - 🟡 Gemini: Challenges Codex/Sonnet/Claude's points
   - 🔴 Codex: Challenges Gemini/Sonnet/Claude's points
   - 🐙 Claude: You challenge advisor points
5. Round 3:
   - All four: Final positions
6. Synthesis with quality scores for each advisor
7. Present results with cost tracking

Quality Checklist

Before completing a debate, ensure:

• All rounds completed for all four participants (Gemini, Codex, Sonnet, Claude)
• Your independent analysis (Opus) written for each round (not just summaries)
• Synthesis.md includes all four perspectives
• Quality scores recorded for advisor responses
• Cost tracking updated (if in claude-octopus context)
• Results presented to user in chat
• Debate folder path provided to user

Integration with Other Skills

Document Delivery

Export debates to professional formats:

After debate completes:
"Would you like to export this debate to PPTX/DOCX/PDF? I can use the document-delivery skill to create a professional presentation."

Knowledge Mode

Debates can be used in knowledge mode workflows:

Knowledge mode "deliberate" phase → Run /debate to get multiple perspectives
→ Use synthesis for final decision

Quality Gates for Responses

Each advisor response is scored before proceeding:

Metric	Weight	Criteria
Length	25 pts	50-1000 words (substantive but concise)
Citations	25 pts	References, links, or sources present
Code Examples	25 pts	Technical examples or code snippets
Engagement	25 pts	Addresses other advisors' specific points

Score >= 75: proceed. Score 50-74: proceed with warning. Score < 50: re-prompt for elaboration.

Cost Tracking

Typical costs (default word limits):

• Quick (1 round): $0.02 - $0.05
• Thorough (3 rounds): $0.10 - $0.20
• Adversarial (5 rounds): $0.25 - $0.50

Cost tracking integrates with ~/.claude-octopus/analytics/ logs.

Export

After debate completes, export results via document-delivery skill:

• PPTX: stakeholder presentations from synthesis
• DOCX: detailed documentation from full transcript
• PDF: archival with metadata (topic, participants, cost)

Attribution

• Original Skill: AI Debate Hub by wolverin0
• Version: v4.8
• Repository: https://github.com/wolverin0/claude-skills
• License: MIT
• Enhancements: Claude-Octopus integration (session-aware storage, quality gates, cost tracking, document export, four-way debate with Sonnet)

Ready to debate! Users can invoke with /debate <question> or natural language.