War Room — Adversarial Decision Engine

Run adversarial multi-agent war-room evaluations for any strategic decision. Spawns 5 parallel subagents (Analyst, Guardian, Treasurer, Builder, Strategist) to challenge a proposal from different angles, then synthesizes a GO/NO-GO/REWORK ruling. Use when: (1) evaluating proposals that need multi-perspective stress-testing, (2) making go/no-go decisions on investments, products, hires, or architecture, (3) any decision where adversarial challenge improves quality. Supports finance, product, engineering, and hiring domains. NOT for: simple questions, routine tasks, or decisions that do not need formal evaluation.

War Room

Structured adversarial evaluation of any strategic proposal using 5 parallel subagents.

Language Rule

Subagents and the final report MUST use the same language as the user's request. If the user writes in Chinese, all agents respond in Chinese and the report is in Chinese. If English, all English. Match the user's language — do not default to English.

Roles

Role	Focus	Must Answer
Analyst	Data, math, quantitative modeling	"Show the numbers and formulas."
Guardian	Risk, failure modes, worst cases	"If [X] fails, what is the maximum loss?"
Treasurer	Resource efficiency, ROI, costs	"Per $1 invested, what is the expected return?"
Builder	Execution feasibility, timeline, tooling	"What is the time/cost/risk to implement?"
Strategist	Strategic fit, alternatives, long-term vision	"How does this fit the long-term strategy?"

4-Phase Flow

Phase 1: Stance

State the proposal, key assumptions, and GO/NO-GO criteria.

Phase 2: Spawn Subagents

Spawn all 5 in parallel with sessions_spawn, mode: "run".

Language instruction: Add to each agent's task prompt: "Respond in {user's language}."

Token optimization (recommended for large proposals):

1. Write proposal data to a temp file: /tmp/rt_{topic}.md
2. Keep task prompts small (~500 tokens): role definition + deliverables + "Read /tmp/rt_{topic}.md for full context"
3. Subagents use the read tool to load the file themselves

This cuts input tokens by ~95% vs inlining all data in each prompt.

If the proposal is short (under 1500 words), inline it directly in the task prompt.

Label pattern: {role}_{topic}_{YYYYMMDD}

Phase 3: Collect and Critique

Wait for all 5 (auto-announced). Then apply the Critic lens:

• Consensus (4/5+ agree)
• Disputes and contradictions
• Stress-test: "If [X] fails, the entire logic collapses."
• Blind spots no agent raised

Phase 4: Ruling and Report

Generate the ruling report and save to file:

1. Write the full report to ~/roundtable/RT{N}_{TOPIC}_{YYYYMMDD}.md
   • Create ~/roundtable/ directory if it doesn't exist
2. Reply to the user with the report content (not just a file path)

Report must include ALL sections:

1. Participants table (role, label, runtime, key contribution)
2. Per-agent summaries with key numbers and arguments
3. Process highlights: Notable quotes, strongest challenges, turning points
4. Consensus points (4/5+ agree)
5. Disputes and contradictions with explicit rulings and rationale
6. Final plan with concrete numbers
7. Scenario projections (bull/base/bear with probabilities)
8. Retained doubts (mandatory: intellectual honesty)
9. Ruling: GO / NO-GO / REWORK + conditions
10. Suggested action items: P0/P1/P2 with owners and deadlines

Audit ID format: RT{N}-{TOPIC}-{YYYYMMDD}

Domain Adaptation

The 5 roles adapt to any domain. See references/domains.md for domain-specific mandatory questions and prompt guidance for: Finance, Product, Engineering, Hiring.

Role Details and Prompt Templates

See references/roles.md for full role definitions. See references/prompts.md for spawn patterns and ruling templates.

Post-Ruling Checklist

1. Report file saved to ~/roundtable/ and report content replied to user
2. Store key decisions to long-term memory with audit ID
3. Git commit the report if in a managed repo
4. Update daily log