Debate

Changelog / 变更日志

/debate — Multi-Agent Debate System

/debate — 多Agent辩论系统

You are launching the Critical Debater multi-agent debate system. 你正在启动 Critical Debater 多 agent 辩论系统。

Arguments / 参数

The user invoked this command with: $ARGUMENTS

Parse the arguments:

• First argument: debate topic (required, quoted string) / 辩论话题（必需）
• Remaining arguments: optional flags in any order / 可选标志，顺序不限

Supported flags:

• --domain <value>: geopolitics | tech | health | finance | philosophy | culture | general (default: auto-infer from topic)
• --depth <value>: quick | standard | deep (default: standard)
• --rounds <N>: number of rounds (default: 3)
• --mode <value>: balanced | red_team (default: balanced)
• --speculation <value>: conservative | moderate | exploratory (default: moderate)
• --output <value>: full_report | executive_summary | decision_matrix (default: full_report)
• --language <value>: en | zh | bilingual (default: bilingual)
• --focus <value>: comma-separated focus areas (default: none)

Examples:

• /debate "Bitcoin vs Gold as store of value" → all defaults, domain auto-inferred as finance
• /debate "React vs Vue" --domain tech --depth deep --speculation exploratory
• /debate "Is remote work productive?" --mode red_team --output executive_summary
• /debate "中东局势" --rounds 5 --focus "oil prices,shipping routes"

Auto-inference / 自动推断: If --domain is not provided, use LLM to infer the most appropriate domain from the topic text. Include the inferred domain in the config.json passed to the orchestrator.

Execution / 执行

Launch the debate-orchestrator agent to drive the entire workflow:

1. Spawn a debate-orchestrator subagent with this prompt:

Run a full multi-agent debate.

Topic: <parsed topic>
Rounds: <parsed rounds>
Config: <all parsed config fields as JSON>
Project root: <current working directory>

Write the full config to the workspace config.json before proceeding.

Follow your system prompt to execute all 4 phases:
1. Initialize workspace (run scripts/init-workspace.sh, gather evidence via WebSearch/WebFetch)
2. Run <N> debate rounds (Pro → Con → Judge each round, using Agent tool for each)
3. Generate final synthesis report
4. Offer scheduled refresh setup

Key references:
- Data contracts: .claude/skills/source-ingest/references/data-contracts.md
- Project rules: docs/debate_system_v2.md
- Scripts: scripts/init-workspace.sh, scripts/validate-json.sh, scripts/hash-snippet.sh, scripts/append-audit.sh

Present the final report in a readable, bilingual (Chinese + English) format to the user.

2. When the orchestrator completes, present its final report output to the user.

Mode: red_team

When config.mode = "red_team", the debate structure changes:

• Con agent becomes Red Team: Instead of opposing a motion, it actively searches for risks, vulnerabilities, failure modes, and blind spots in the topic/plan.
• Pro agent becomes Blue Team: Instead of supporting a motion, it defends against Red Team attacks with mitigations, contingency plans, and risk acceptance rationale.
• Judge agent: Evaluates the SEVERITY and LIKELIHOOD of each identified risk, and the FEASIBILITY of proposed mitigations.

Orchestrator prompt adjustment:

Run a Red Team analysis.

Topic: <parsed topic>
Mode: red_team
...

IMPORTANT MODE CHANGE:
- Con agent's role: RED TEAM — find every possible risk, failure mode, vulnerability, and blind spot. Be creative and thorough. Think about edge cases, cascading failures, adversarial scenarios, and black swan events.
- Pro agent's role: BLUE TEAM — for each risk identified by Red Team, propose mitigations, assess residual risk, and provide risk acceptance rationale where mitigation is impractical.
- Judge's role: Assess each risk's severity (critical/high/medium/low) and likelihood (high/medium/low). Evaluate whether Blue Team's mitigations are feasible and sufficient.

Notes / 注意事项

• The debate runs sequentially within each round: Pro → Con → Judge
• Each agent runs as an isolated subagent
• All state is persisted to the workspace directory as JSON files
• The system guarantees real-time information access (can fetch data from last 24 hours)
• Evidence chains may span any timeframe — historical evidence is valid
• No total scores — output uses evidence states (verified, contested, unverified, stale)