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):
- Write proposal data to a temp file:
/tmp/rt_{topic}.md - Keep task prompts small (~500 tokens): role definition + deliverables + "Read /tmp/rt_{topic}.md for full context"
- Subagents use the
readtool 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:
- Write the full report to
~/roundtable/RT{N}_{TOPIC}_{YYYYMMDD}.md- Create
~/roundtable/directory if it doesn't exist
- Create
- Reply to the user with the report content (not just a file path)
Report must include ALL sections:
- Participants table (role, label, runtime, key contribution)
- Per-agent summaries with key numbers and arguments
- Process highlights: Notable quotes, strongest challenges, turning points
- Consensus points (4/5+ agree)
- Disputes and contradictions with explicit rulings and rationale
- Final plan with concrete numbers
- Scenario projections (bull/base/bear with probabilities)
- Retained doubts (mandatory: intellectual honesty)
- Ruling: GO / NO-GO / REWORK + conditions
- 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
- Report file saved to
~/roundtable/and report content replied to user - Store key decisions to long-term memory with audit ID
- Git commit the report if in a managed repo
- Update daily log