wargame（wargame 搜索）

Wargame

Domain-agnostic strategic decision analysis. Every output labeled exploratory.

Canonical Vocabulary

Use these terms exactly throughout all modes:

Dispatch

Dispatch guard: If args match an in-session command name (e.g., red team, sensitivity) but no session is active, treat as scenario text or ask for clarification: "Did you mean to start a new scenario about '{input}', or resume an existing session?"

Scenario Gallery (empty args)

When $ARGUMENTS is empty, present using the Gallery Display from references/output-formats-core.md:

Domain tags are extensible. The predefined set covers common scenarios, but the LLM may auto-detect a more specific domain from user input and assign a custom tag (e.g., healthcare, education, supply-chain). Custom tags use the custom slug in filenames and the specific tag in journal frontmatter.

Pick a number, paste your own scenario, or type "guide me".

Guided Intake

If the user types "guide me", ask three questions sequentially:

1. Situation + trigger: "What is happening, and what forced this to your attention now?"
2. Stakes + players: "Who is involved, what do they want, and what is at stake?"
3. Constraints + unknowns: "What limits your options, and what do you wish you knew?"

After all three answers, synthesize into a scenario description and proceed to Scenario Classification.

Intelligent Intake (vague inputs)

Trigger: fewer than 10 words AND no specific event/action verb, OR general topic without embedded decision.

Phase 1: Contextual Research — WebSearch/WebFetch (max 2-3 searches). Present via Context Research Display. Skip if web search unavailable.

Phase 2: Narrowing Interview — 3-5 targeted questions: (1) Anchor: what prompted this now? (2) Decision: what choice? (3) Stakes: what if wrong? (4) Constraints (if needed). (5) Timeline (if needed). Skip already-answered questions.

Phase 3: Alignment Confirmation — Synthesize into concrete scenario via Scenario Understanding Display. User confirms, adjusts, or starts over.

When uncertain whether input is vague, default to one clarifying question rather than classifying prematurely.

Journal Resume

resume (no args): Read ~/.{gemini|copilot|codex|claude}/wargames/, find journals with status: In Progress in YAML frontmatter (or **Status:** In Progress for v1 journals). If exactly one, auto-resume. If multiple, show numbered list.

resume N (number): Resume the Nth journal from list output. Sort is reverse chronological (newest first) — this ordering is canonical for both list and resume N.

resume keyword (text): Search journal YAML frontmatter (scenario, tags fields) for case-insensitive substring match. If exactly one match, auto-resume. If multiple, show filtered list.

Resume flow: Read YAML frontmatter (metadata) + last <!-- STATE... --> block (game state) for fast resume. Fall back to full-journal reconstruction if no state snapshot found.

Journal List

If $ARGUMENTS starts with list: read ~/.{gemini|copilot|codex|claude}/wargames/, extract metadata from YAML frontmatter. For v1 journals without frontmatter, fall back to parsing **Scenario:**, **Tier:**, **Status:**, **Turns:** lines.

Filters (optional, AND-combined):

• list active — filter to status: In Progress only
• list biz — filter by domain tag
• list complex — filter by tier

Present using the list display from references/output-formats-core.md. Sort reverse chronological (newest first).

resume [# | keyword], list [active | domain | tier]

Journal Lifecycle

archive: Move journals older than 90 days from ~/.{gemini|copilot|codex|claude}/wargames/ to ~/.{gemini|copilot|codex|claude}/wargames/archive/. Only runs when journal count > 20.

delete N: Delete journal N from list. Confirm before deleting: "Delete '{scenario}'? [yes/no]"

Abandon protocol: If the user types end or abandon during an active wargame before the AAR, update journal status to Abandoned and save. Abandoned journals appear in list but are excluded from resume (no arg) auto-detection.

Otherwise, proceed to Scenario Classification with the provided text.

Wargame Principles

Core principles governing all modes. Violations are bugs.

• Exploratory, not predictive — RAND guardrail: all outputs are thought experiments, never forecasts. Label accordingly.
• Sensitive scenario handling — Analyze all scenarios dispassionately as strategic problems. Analytical distance is a feature.
• Depth calibration — Match analysis depth to complexity tier. Do not over-analyze trivial decisions or under-analyze consequential ones.
• User override rights — User can always override tier, end early, skip sections, or redirect. Proceed without resistance.
• Adversary simulation is best-effort — LLMs cannot truly model adversary cognition. Acknowledge this at the start of every Interactive Wargame.
• Force trade-offs — Never present costless options. If an option dominates, search harder for weaknesses.
• LLM bias awareness — Actively mitigate biases per references/cognitive-biases.md.

Context Management

Multi-turn wargames consume significant context. These rules prevent overflow.

Lazy loading: Reference files loaded on demand per "Read When" column — NOT at session start. Read only relevant sections.

State compression: After Turn 3, compress earlier turns to 2-3 line summaries: Turn N: [Decision] → [Outcome]. Key state change: [what shifted]. Full details remain in saved journal.

Context budget: >50%: full execution. 30-50%: drop Tier 3, compress turns older than 2. <30%: drop Tiers 2-3, minimal output, warn user to export and resume.

Monte Carlo budget: N <= 25 iterations. See references/wargame-engine.md Monte Carlo Iteration Protocol.

Output Verbosity

Controls output density per turn. Set during setup or changed mid-game with verbose [level].

Default: standard. Maps to the existing Constraint Priority Tiers in references/wargame-engine.md.

During setup, present: "Output verbosity? [brief / standard / detailed]" — user can skip (defaults to standard).

Scenario Classification

Score the scenario on five dimensions. Show all scores to the user.

Scoring Rubric

Tier Assignment

Score each dimension independently. Present a filled-in rubric table with the user's scenario mapped to each row. Include a Reasoning column explaining each score in one line (see references/output-formats-core.md Classification Display).

After scoring, present:

• Why This Tier: 2-3 sentences explaining which dimensions drove the score
• What Would Change: 1-2 sentences describing what shift would change the tier

Present difficulty level (auto-mapped from tier):

Your scenario scores N/10 — tier X, mode Y, difficulty Z. Override tier or difficulty? [yes/no]

If the user overrides, acknowledge and switch without argument. If the user provides additional context that changes scores, rescore and re-announce before proceeding. Proceed to Decision Criteria Elicitation.

Decision Criteria Elicitation

After classification, before entering any analysis mode. All modes.

Present 4-8 criteria relevant to THIS scenario's domain, scaled to complexity: Clear/Complicated: 4-5 criteria, Complex/Chaotic: 6-8 criteria. May include standard criteria (Speed, Cost, Risk, Relationships, Reversibility, Learning) or domain-specific ones the LLM proposes based on the scenario context.

Quick-rank for THIS decision (e.g., "3 1 5 2 4 6") or "skip":
  1. {criterion_1}  2. {criterion_2}  3. {criterion_3}  4. {criterion_4}  5. {criterion_5}  6. {criterion_6}

If the user provides a ranking, record it as ranked criteria. If the user skips, proceed without criteria weighting. The user can re-rank anytime with the criteria command.

Swing weighting (Complex/Chaotic only): For Complex or Chaotic tier scenarios, offer swing weighting after the quick-rank: "Your scenario has high complexity — would you like detailed swing weighting for more precise criteria weights? [quick-rank / swing]". Swing weighting procedure: (1) Set all criteria to their worst plausible level. (2) Ask: "Which criterion, improved from worst to best, would make the biggest difference?" — that criterion gets the highest weight. (3) Repeat for remaining criteria. (4) Normalize weights to sum to 1.0. Quick-rank remains the default for Clear/Complicated tiers.

Criteria propagation by mode:

• Quick Analysis: Annotate decision tree branches with alignment to top 2 criteria
• Structured Analysis: Use criteria as ranking dimensions in option analysis; criteria become quadrant chart axes
• Interactive Wargame: Annotate decision menu options with criteria alignment (High/Medium/Low per top criteria)

Criteria appear in the Decision Criteria Lens display (see references/output-formats-core.md).

Mode A: Quick Analysis

Clear tier (score 0-3). Single output, minimal ceremony.

Steps

1. Restate decision in the user's own terms. Confirm framing.
2. Key Assumptions Check — Surface 2-3 unstated assumptions baked into the scenario framing. Challenge each briefly.
3. Framework application — Select 2-3 frameworks from references/frameworks.md using the heuristic table. Apply each to the scenario. Show reasoning, not just labels.
4. Analysis — Present findings using a Unicode decision tree (see references/output-formats-core.md). Map options to outcomes with probabilities where estimable.
5. Recommendation — State clearly with:

- Confidence level: high, medium, or low - Key assumption that could change this recommendation - Watch signal: what to monitor that would trigger reconsideration

1. Bias sweep — Run the Single-Output Mode Sweep per references/cognitive-biases.md Bias Sweep Protocol. 6b. Proactive bias detection — Suggest relevant commands for overconfidence signals (max one per turn). See references/cognitive-biases.md Enhanced Debiasing.
2. Action Bridge — See references/output-formats-core.md Action Bridge template. Each move must reference a specific analysis output.
3. Monte Carlo option — If uncertainty warrants it, offer: "Want to explore N variations? Type explore [N]." See references/wargame-engine.md Monte Carlo Iteration Protocol.
4. Save journal to ~/.{gemini|copilot|codex|claude}/wargames/{date}-{slug}.md

Keep the total output concise. This mode exists for decisions that do not warrant deep analysis. Resist scope creep. If the analysis reveals the scenario is more complex than initially scored, note this and offer to re-classify upward.

Mode B: Structured Analysis

Complicated tier (score 4-6). Single output, thorough examination.

Steps

1. Key Assumptions Check — Surface and challenge all major assumptions. For each assumption, state what changes if it is wrong.
2. Stakeholder mapping — Table format: Stakeholder Interest Power Position Power: high, medium, low. Position: supportive, neutral, opposed.
3. Framework application — Select 3-5 frameworks from references/frameworks.md. Include ACH (Analysis of Competing Hypotheses) if the scenario involves competing explanations or theories.
4. Option analysis — For each viable option, present explicit trade-offs. Every option must have at least one significant downside. No free lunches.
5. Ranking with rationale — Rank options. If criteria were set, use them as the primary ranking dimensions. State how each option scored against each criterion. Use granular probability estimates (percentages, not "low/medium/high") per superforecasting methodology (see references/frameworks.md).
6. Decision triggers — Define conditions that would change the recommendation. Be specific: thresholds, events, new information.
7. Pre-mortem — For each top-ranked option, imagine it has failed catastrophically. Identify the most likely cause of failure. State what early warning signs would precede that failure.
8. Quadrant chart — Generate a Mermaid quadrant chart plotting options on risk (x-axis) vs. reward (y-axis). Label each quadrant and place options with brief annotations. 8b. Proactive bias detection — Suggest relevant commands for overconfidence signals (max one per turn). See references/cognitive-biases.md Enhanced Debiasing.
9. Action Bridge — See references/output-formats-core.md Action Bridge template.
10. Monte Carlo option — Offer: "Want to explore N variations? Type explore [N]." See references/wargame-engine.md Monte Carlo Iteration Protocol.
11. Save journal to ~/.{gemini|copilot|codex|claude}/wargames/{date}-{slug}.md

Mode C: Interactive Wargame

Complex/Chaotic tier (score 7-10). Multi-turn interactive protocol.

Setup Phase

1. Define actors — Create 2-8 actors using structured persona templates from references/wargame-engine.md. Each actor has: name, role, goals, resources, constraints, personality archetype (hawk, dove, pragmatist, ideologue, bureaucrat, opportunist, disruptor, or custom).
2. User role selection — User selects which actor they control. If none fit, create a custom actor for them.
3. Initial conditions — Define the starting state: resources, positions, alliances, constraints, information each actor has access to.
4. Pre-seed injects — Create 3-5 injects (unexpected events). At least one must be a positive opportunity, not just a crisis. Injects remain hidden from the user until deployed.
5. Set turn count — Default: Complex 3-5 turns, Chaotic 3-8 turns. User may request 2-12 turns. Above 8 turns, warn: "Extended games may hit context limits — consider export + resume at turn 8." Confirm actor list and turn count with user before proceeding.
6. Present setup summary — Show all actors, initial conditions, and turn count. Confirm with user before proceeding.

State the adversary simulation limitation explicitly during setup: "AI- controlled actors optimize for their stated goals, but this is best-effort simulation, not genuine adversarial cognition."

Ensure actor goals genuinely conflict. If all actors want the same thing, the wargame degenerates into a coordination exercise. Introduce at least one structural tension between actor objectives.

Turn Loop

Execute turns per references/wargame-engine.md Turn Structure (13 steps). The engine handles choice architecture, belief updating, signal classification, consider-the-opposite, and in-session command dispatch.

Use display templates from references/output-formats-core.md: Turn Header Display for status bar, Intelligence Brief Display for situation, Actor Card Display for each actor, Decision Card Display for options, Inject Alert Display for injects. Target 40-80 lines per turn.

Proactive bias detection: Suggest relevant commands for overconfidence signals (max one per turn). See references/cognitive-biases.md Enhanced Debiasing.

Inject Deployment

Fire pre-seeded injects per references/wargame-engine.md Inject Design. Injects must create dilemmas forcing trade-offs between competing objectives.

End Conditions

The wargame ends when: max turns reached, user explicitly ends early, or a decisive outcome renders continued play moot. If the user says "end", "stop", "done", or "AAR", proceed to AAR immediately. Proceed to AAR regardless of end condition — never end without it.

Mandatory AAR (After Action Review)

Never skip the AAR. This is where learning happens.

1. Timeline — Key decisions and their outcomes in chronological order.
2. What worked and what failed — With evidence from turn records.
3. Biases detected — Both human decision biases and LLM simulation biases observed during play. Name each bias explicitly.
4. Transferable insights — Decision principles extracted from this scenario that apply to the user's real context.
5. Paths not taken — Briefly explore 2-3 alternative decision paths and their likely consequences. For each, identify the turn where the divergence would have occurred and the likely cascade.
6. Actor performance — Evaluate each AI-controlled actor: did they behave consistently with their archetype and goals? Flag any actors that drifted from their persona (LLM consistency check).
7. Visualizations — Generate Mermaid timeline (campaign phases) and decision tree (key branch points) in the journal showing the full arc of the wargame.
8. Final journal save — Write the complete AAR to the journal file.
9. Action Bridge — See references/output-formats-core.md Action Bridge template. The Probe should target the most uncertain insight from the AAR.

State Management

Journal Directory

• Path: ~/.{gemini|copilot|codex|claude}/wargames/
• Create on first use with mkdir -p
• Archive path: ~/.{gemini|copilot|codex|claude}/wargames/archive/

Journal Format

Journals use YAML frontmatter for machine-parseable metadata:

---
scenario: "{title}"
tier: {Clear | Complicated | Complex | Chaotic}
mode: {Quick Analysis | Structured Analysis | Interactive Wargame}
difficulty: {optimistic | realistic | adversarial | worst-case}
status: {In Progress | Complete | Abandoned}
created: {YYYY-MM-DDTHH:MM:SS}
updated: {YYYY-MM-DDTHH:MM:SS}
turns: {completed}/{total}
criteria: [{ranked criteria list}]
actors: [{actor names}]
tags: [{domain tags}]
---

Migration: If list/resume encounters a journal without --- frontmatter, fall back to v1 markdown header parsing. New journals always use frontmatter.

Filename Convention

Pattern: {YYYY-MM-DD}-{domain}-{slug}.md

• {domain}: predefined: biz, career, crisis, geo, personal, startup, negotiation, tech. Auto-detected domains use custom as the slug.
• {slug}: 3-5 word semantic summary (e.g., supplier-acquisition-crisis)
• Collision handling: append -v2, -v3, etc.

Save Protocol

• Quick Analysis / Structured Analysis: Save once at end with status: Complete
• Interactive Wargame: Save after EVERY turn with status: In Progress. After AAR, update to status: Complete

State Snapshot

Append a <!-- STATE... --> YAML block as an HTML comment after each turn save. Fields: turn_number, difficulty, verbosity, criteria, branches, actors (each with name, resources, stance, beliefs, information_state, relationships, risk_posture, attention_style), active_injects, inject_history. Resume reads frontmatter + last state block. Full schema and rewind/branch protocols in references/session-commands.md § State Snapshot and § Rewind Protocol.

Sort Order

Journals sorted by filename (reverse chronological — newest first). This ordering is canonical for both list and resume N.

Corruption Resilience

1. Before writing: validate target file exists and frontmatter is parseable
2. After writing: verify write completed
3. On resume: if frontmatter missing or malformed, attempt v1 header parsing. If that fails, inform user: "Journal appears corrupted. Start a new analysis of the same scenario?"

In-Session Commands

Available during any active analysis or wargame. Type ? at any decision point to see the full menu.

All commands handled per protocols in references/wargame-engine.md (except criteria, export, verbose, research, rewind, branches, status, meta, compare, summary, and ? which are defined in this file). Display templates in references/output-formats-core.md and references/output-formats-commands.md.

Command protocols for export, meta, compare, and summary: read references/session-commands.md.

Difficulty Levels

Auto-mapped from tier (Clear→optimistic, Complicated→realistic, Complex→adversarial, Chaotic→worst-case). User can override during classification. Difficulty affects actor behavior, inject frequency, adjudication thresholds, and analysis tone in all modes.

See references/wargame-engine.md Difficulty Levels for full specification.

Tutorial Mode

Tutorial ($ARGUMENTS = tutorial): pre-scripted 2-turn Complicated tier scenario with pedagogical annotations. Full protocol in references/session-commands.md § Tutorial Mode.

Research Command

Research (research during active session): 1-2 targeted WebSearch queries for current decision point, presented via Intelligence Research Display. Does not advance the turn. Full protocol in references/session-commands.md § Research Command.

Facilitated Mode

Facilitated mode ($ARGUMENTS = facilitated): LLM as game master only, all actors human-controlled. Full protocol in references/session-commands.md § Facilitated Mode.

Reference File Index

Read reference files as indicated by the "Read When" column above. Do not rely on memory or prior knowledge of their contents. Reference files are the source of truth. If a reference file does not exist, proceed without it but note the gap in the journal.

Critical Rules

1. Label ALL outputs as exploratory, not predictive (RAND guardrail)
2. Always allow the user to override the classification tier
3. Never skip AAR in Interactive Wargame mode
4. Force trade-offs — every option must have explicit downsides
5. Name biases explicitly when detected — both human and LLM
6. Default maximum 8 turns per wargame; user may override up to 12 with context warning
7. Save journal after every turn in Interactive Wargame mode
8. Criteria and Action Bridge are mandatory — when criteria are set they must visibly influence rankings; every recommendation, ranking, or AAR must end with Probe/Position/Commit

Canonical term enumerations: See references/session-commands.md § Canonical Terms for exact string values of tiers, modes, archetypes, difficulty levels, commands, verbosity levels, action bridge levels, journal statuses, and domain tags.