SPAM!

Autonomous agent system for a competitive restaurant simulation game. Each turn the system reasons about zone placement, menu composition, ingredient bidding, pricing, and diplomatic manipulation.

---

Core Strategies

1. LLM Strategic Planner (src/decision/strategy_agent.py)

A gpt-oss-120b-powered StrategyAgent acts as the turn brain. It produces a structured TurnStrategy JSON before any skill runs, setting parameters for every downstream module: which zone to target, how aggressively to bid, what price multiplier to apply, and which diplomatic posture to take.

2. Zone-Based Decision Routing (src/decision/subagent_router.py, zone_selector.py)

The market is segmented into strategic zones (e.g. DIVERSIFIED, PREMIUM, BUDGET). An algorithmic heuristic picks the zone; the StrategyAgent can override it when confidence ≥ 0.5. Each zone maps to a dedicated datapizza Agent with a tailored system prompt.

3. ILP Menu & Bid Optimizer (src/decision/ilp_solver.py)

Integer Linear Programming solves menu composition and ingredient bid allocation under budget and prep-time constraints, guided by the StrategyAgent's bid_aggressiveness and menu_* parameters.

4. Dynamic Pricing (src/decision/pricing.py)

Prices are computed per recipe based on competitor observations, then multiplied by the strategy agent's price_adjustment_factor. Undercutting mode is toggled by the agent.

5. Competitive Intelligence Pipeline (src/intelligence/)

A DagPipeline DAG runs every turn:

DataCollector → StateBuilder → FeatureExtractor → StrategyInferrer
→ Embedding → TrajectoryPredictor → ClusterClassifier → BriefingGenerator

• StrategyInferrer classifies each competitor into archetypes: PREMIUM_MONOPOLIST, BUDGET_OPPORTUNIST, AGGRESSIVE_HOARDER, MARKET_ARBITRAGEUR, REACTIVE_CHASER, DECLINING_PASSIVE, DORMANT.
• TrajectoryPredictor forecasts competitor behavior over the next few turns.
• VectorStore holds competitor embeddings for similarity search and clustering.

6. Deception Bandit (src/diplomacy/deception_bandit.py)

Thompson Sampling (Beta priors, per-competitor) selects from 7 manipulation arms: truthful_warning · inflated_intel · manufactured_scarcity · ingredient_misdirect · alliance_offer · price_anchoring · silence

Rewards are measured by observable competitor state changes (bid shifts, menu price changes) tracked across turns.

7. Pseudo-GAN Message Crafter (src/diplomacy/pseudo_gan.py)

A generator (gpt-oss-120b) drafts diplomatic messages in Italian; a discriminator (gpt-oss-20b) scores believability. The loop refines up to 3 iterations until score > 0.7. Messages are grounded in real tracker intel to maximize credibility.

8. Ground Truth Firewall (src/diplomacy/firewall.py)

Incoming competitor messages are filtered and cross-checked against observed game state to prevent our own agent from acting on manipulated intel.