phase: Skill for Claude Code

Adding a Deck-Aware AI Feature + Policy

This is the authoritative checklist for adding a new feature to the three-layer AI architecture (Features → Plan → Policies). Nine features exist as reference: landfall, mana_ramp, tribal, control, aristocrats, aggro_pressure, tokens_wide, plus_one_counters, spellslinger_prowess. Each follows the same pattern; mirror it.

Before you start: Read landfall.rs (simplest), then aristocrats.rs (geometric-mean commitment with identity lookup), then one feature similar in shape to yours. Trace from features/<name>.rs::detect() through DeckFeatures::analyze and AiSession::from_game through policies/<name>.rs::verdict so the data flow is concrete.

CR Verification Rule. Every CR number in annotations MUST be verified by grepping docs/MagicCompRules.txt BEFORE you write it. Do NOT rely on memory or prior prompts. The 701.x and 702.x sections are arbitrary sequential assignments that hallucinate consistently. Run grep -n "^701.34" docs/MagicCompRules.txt for every number. CR 122.1d is Stun (NOT +1/+1 — that's CR 122.1a). CR 701.27 is Transform (NOT Proliferate — that's CR 701.34). Two real catches from the existing batch — you will have similar misses unless you grep.

1. Architecture (memorize this)

AiSession (per-game cache; built once in AiSession::from_game)
  ├─ features: HashMap<PlayerId, DeckFeatures>    ← Layer 1: structural detection
  ├─ plan:     HashMap<PlayerId, PlanSnapshot>    ← Layer 2: bottoming + curve schedule
  ├─ deck_profile / strategy: per-player strategic profiles
  ├─ synergy:  HashMap<PlayerId, SynergyGraph>
  └─ memory:   PolicyMemory

DeckFeatures
  ├─ landfall:       LandfallFeature
  ├─ mana_ramp:      ManaRampFeature
  ├─ tribal:         TribalFeature
  ├─ control:        ControlFeature
  ├─ aristocrats:    AristocratsFeature
  ├─ aggro_pressure: AggroPressureFeature
  ├─ tokens_wide:    TokensWideFeature
  ├─ plus_one_counters: PlusOneCountersFeature
  ├─ spellslinger_prowess: SpellslingerProwessFeature
  └─ <your new feature>

Policies (Layer 3)
  ├─ TacticalPolicy — in-game decisions (CastSpell, ActivateAbility, DeclareAttackers, ...)
  └─ MulliganPolicy — pre-game hand evaluation (sibling trait)

Both run inside their respective registries (`PolicyRegistry::default`,
`MulliganRegistry::default`). Activation is the single multiplicative knob.