debate-orchestration

name	debate-orchestration
description	This skill should be used when the user asks to "run a multi-model review", "debate code changes", "invoke reviewd", "get multiple model opinions", "run adversarial review", or mentions multi-model review, debate orchestration, reviewd CLI, codex review, gemini review, or adversarial code analysis.

Debate Orchestration

Multi-Model Review via reviewd

The reviewd CLI (>=0.6.0) orchestrates reviews across subscription-based LLM CLIs.

CLI Invocation Patterns

# Run reviewd with default model config
reviewd review --diff <diff-file> --config reviewd.toml

# Individual model CLIs (subscription auth, zero API keys)
codex exec "Review this diff for issues: $(cat diff.txt)"
echo "Review this diff for issues: $(cat diff.txt)" | gemini -p
claude --print "Review this diff for issues: $(cat diff.txt)"

Review Workflow

1. Prepare — Generate diff: git diff main...HEAD > /tmp/review-diff.txt
2. Distribute — Send diff to each model CLI
3. Collect — Parse structured JSON review results from each model
4. Validate — Check debate integrity rules before consensus
5. Forward — Pass validated results to consensus-validation skill

Model Configuration

All models use subscription auth (zero API keys):

• codex: codex exec or codex review — OpenAI subscription
• gemini: gemini -p or stdin pipe — Google subscription
• claude: claude --print — Anthropic subscription

Debate Integrity Rules (from claude-octopus)

Before accepting review results, verify:

• No circular reasoning between models
• No severity inflation (escalating without evidence)
• No rubber-stamping (all models agreeing without analysis)
• Each model provides independent evidence for findings

Rules

• All LLM calls go through subscription CLIs — never use API keys
• Never improvise CLI flags — use exact syntax documented above
• Parse review output as structured JSON, not free text
• Validate debate integrity before forwarding to consensus
• Write raw review results to JSON files for audit trail