buff（buff 搜索）

Buff

*"All effects are buffs. Some are just shitty."*

Buffs modify stats, abilities, or behavior. They have durations, can stack, and come from various sources. Curses are just negative buffs — no separate system.

Characters Only

Buffs only target characters. This is a design constraint, not a limitation.

• Single closure signature: (world, subject, verb, object)
• subject is always a character — no type checking needed
• Rooms that need buffs get a "room spirit" character

# Room needs to be "haunted"? Create its spirit.
character:
  id: dark-cave-spirit
  name: "Spirit of the Dark Cave"
  location: room/dark-cave
  buffs:
    - ref: buff/haunted

Structure

buff:
  name: "Caffeinated"
  source: "Espresso"
  effect: { energy: +2, focus: +1 }
  duration: 5  # simulation turns
  stacks: false

Buff Types

Numeric

Traditional stat modifiers:

buff:
  name: "Caffeinated"
  effect: { energy: +2, focus: +1 }
  duration: 5

Semantic

Arbitrary effect prompts interpreted by the LLM — not predefined stats, just vibes:

• "feeling lucky"
• "cats seem to like you today"
• "slightly cursed"
• "radiating calm energy"
• "shadows feel watchful"

How it works:

Buff: "cats seem to like you today"
Action: PAT TERPIE
LLM: Gives bonus, narrates extra warmth

Mixed

Combine numeric and semantic:

buff:
  name: "Terpie's Blessing"
  effect:
    calm: +2
    vibe: "cats trust you more"
  duration: "a while"

Standard Properties Buffs Affect

Player/NPC Stats (Sims-Style Needs)

# Numeric needs — decay over time, restored by actions
needs:
  hunger: 80      # 0=starving, 100=full
  energy: 65      # 0=exhausted, 100=rested
  social: 45      # 0=lonely, 100=connected
  hygiene: 90     # 0=filthy, 100=clean
  bladder: 30     # 0=desperate, 100=empty
  fun: 55         # 0=bored, 100=entertained
  comfort: 70     # 0=miserable, 100=cozy

Mind-Mirror Stats (Cognitive/Emotional)

# Mental state — affects decision-making and narration
mind:
  focus: 75       # Concentration (0-100)
  mood: 20        # Emotional valence (-100 to +100)
  stress: 35      # Anxiety level (0-100)
  creativity: 60  # Creative capacity (0-100)
  confidence: 50  # Self-assurance (0-100)
  patience: 40    # Frustration tolerance (0-100)
  curiosity: 80   # Exploration drive (0-100)

Room Spirit Stats

Room spirits are characters whose stats affect the room they haunt:

character:
  id: forge-spirit
  name: "Spirit of the Forge"
  location: room/blacksmith-forge

  # These stats affect everyone in the room
  production_speed: 120   # +20% crafting speed
  error_rate: 8           # 8% chance of mistakes
  mood_influence: +5      # Slight pride boost
  comfort_bonus: -10      # Hot and uncomfortable
  discovery_chance: 15    # Sometimes find rare materials
  danger_level: 25        # Burns, sparks, accidents

  buffs:
    - id: master-craftsman-blessing
      source: "Pleased the forge spirit"
      effect: { production_speed: +30, error_rate: -5 }
      duration: "until you leave"

Sources

Lifecycle Hooks

Three hooks control buff behavior, written as natural language and compiled to JS:

Example: Poison Buff

buff:
  id: poison
  name: "Poisoned"
  tags: [curse, damage-over-time, dispellable]

  # Natural language prompts (author writes these)
  start: "Mark character as poisoned, turn them slightly green"
  simulate: "Reduce HP by 1, chance of groaning sound"
  is_finished: "Return true after 5 ticks OR if HP drops below 10"

  # Compiled by buff compiler (generated)
  start_js: |
    subject.poisoned = true;
    subject.tint = 'green';
  simulate_js: |
    subject.hp -= 1;
    if (Math.random() < 0.3) world.emit('*groan*');
  is_finished_js: |
    return subject.poisonTicks >= 5 || subject.hp < 10;

Closure Signature

All compiled hooks use the same signature:

(world, subject, verb, object) => { ... }

• world — shared game state (never null)
• subject — the character with the buff (never null for buffs)
• verb — context-dependent (may be null)
• object — context-dependent (may be null)

Body-only in YAML: Write just the code body, engine wraps it.

Buff Interactions

Buffs can look up and modify other buffs by tag:

Cancel Example

buff:
  id: cleanse
  name: "Cleanse"
  tags: [holy, dispel]
  cancels: [curse, poison, disease]  # Remove all matching
  start: "Holy light purges dark afflictions"

Boost Example

buff:
  id: fire-attunement
  name: "Fire Attunement"
  boosts:
    tags: [fire]
    multiplier: 2.0
    extend_duration: 5
  start: "Fire spells burn twice as hot"

Merge Example

buff:
  id: rage
  name: "Rage"
  tags: [combat, aggression]
  merges_with:
    tags: [focus, discipline]
    result: battle-trance  # Creates new combined buff

buff:
  id: battle-trance
  name: "Battle Trance"
  tags: [combat, legendary]
  effect: { damage: +50%, focus: +30, pain_immunity: true }
  start: "Fury and focus unite — you become a weapon"

Blocked By Example

buff:
  id: poison
  name: "Poisoned"
  tags: [poison, damage-over-time]
  blocked_by: [immunity-poison, divine-protection]
  # Won't apply if target has these tags

Weight Trees (ML-Style Mixtures)

Buffs can form hierarchical weighted mixtures, like neural network layers:

Blend → Strains → Terpenes → Effects
  ↓        ↓          ↓          ↓
weights  weights    weights    final
  ↑        ↑          ↑          ↑
 BUFFS   BUFFS      BUFFS     BUFFS   ← Each stage can be modified by buffs!

Meta-buffs can modify the weight tree itself:

Tolerance Relationships

Tolerances use the character relationship map — same system as NPC friendships:

character:
  id: player
  name: "Don"

  # Relationships include people AND substances
  # Terpenes are unidirectional — they don't have feelings back
  relationships:
    # NPCs (bidirectional)
    bob: { trust: 45, friendship: 60 }
    alice: { trust: 80, friendship: 75 }

    # Terpene tolerances (unidirectional — no reciprocal)
    terpene/myrcene: { tolerance: 45 }      # Couch-lock less effective
    terpene/limonene: { tolerance: 12 }     # Citrus hits hard
    terpene/pinene: { tolerance: 30 }
    terpene/linalool: { tolerance: 5 }      # Lavender knocks you out
    terpene/caryophyllene: { tolerance: 60 } # Need more for pain relief
    terpene/humulene: { tolerance: 20 }
    terpene/terpinolene: { tolerance: 8 }   # Full creative boost
    terpene/ocimene: { tolerance: 3 }       # Maximum effect

Key difference from NPC relationships:

Tolerance mechanics:

Tolerance changes:

# Each use increases tolerance
on_use:
  tolerance_gain: 2-5 points per use

# Tolerance decays over time (T-break!)
on_rest:
  tolerance_decay: 1 point per day of abstinence

# Full reset after extended break
t_break:
  duration: 2 weeks
  effect: "Reset to 50% of current tolerance"

Effective weight calculation:

def get_effective_terpene_weight(character, terpene, base_weight):
    tolerance = character.tolerances.get(terpene, 0)

    # Convert tolerance to multiplier
    if tolerance < 25:
        multiplier = 1.5 - (tolerance / 50)  # 1.5x → 1.0x
    elif tolerance < 75:
        multiplier = 1.0 - ((tolerance - 50) / 100)  # 1.0x → 0.75x
    else:
        multiplier = 0.75 - ((tolerance - 75) / 100)  # 0.75x → 0.5x

    return base_weight * multiplier

Layer 1: Terpenes → Effects

Each terpene has weighted effects:

myrcene-blessing:
  effects_weighted:
    relaxation: { value: +30, weight: 1.0 }   # Full effect
    pain_relief: { value: +20, weight: 0.8 }  # 80%
    sedation: { value: +25, weight: 0.9 }     # 90%

Layer 2: Strains → Terpenes

Each strain is a weighted mixture of terpenes:

strain-og-kush:
  terpene_profile:
    myrcene: 0.35      # 35% of profile
    limonene: 0.25     # 25%
    caryophyllene: 0.20
    linalool: 0.10
    humulene: 0.10

Layer 3: Blends → Strains

Blends mix multiple strains:

blend-wake-and-bake:
  strain_mixture:
    sour-diesel: 0.50     # Half the blend
    jack-herer: 0.30      # 30%
    pineapple-express: 0.20

Computing Final Effects

# Blend → Strain → Terpene → Effect propagation
def compute_blend_effects(blend):
    final_terpenes = {}

    # Layer 3→2: Blend weights × Strain terpene profiles
    for strain_id, strain_weight in blend.strain_mixture.items():
        strain = get_strain(strain_id)
        for terpene, terpene_weight in strain.terpene_profile.items():
            final_terpenes[terpene] += strain_weight * terpene_weight

    # Layer 2→1: Terpene amounts × Effect weights
    final_effects = {}
    for terpene, amount in final_terpenes.items():
        terpene_buff = get_terpene_buff(terpene)
        for effect, config in terpene_buff.effects_weighted.items():
            final_effects[effect] += amount * config.weight * config.value

    return final_effects

Example Calculation

Wake & Bake Blend:
├── Sour Diesel (50%)
│   ├── limonene: 0.30 × 0.50 = 0.15
│   └── pinene: 0.15 × 0.50 = 0.075
├── Jack Herer (30%)
│   ├── limonene: 0.20 × 0.30 = 0.06
│   └── pinene: 0.25 × 0.30 = 0.075
└── Pineapple Express (20%)
    ├── limonene: 0.30 × 0.20 = 0.06
    └── pinene: 0.25 × 0.20 = 0.05

Final limonene: 0.15 + 0.06 + 0.06 = 0.27
Final pinene: 0.075 + 0.075 + 0.05 = 0.20

Then: limonene × mood_boost, pinene × focus → final character effects

This is essentially a mini neural network where:

• Weights are terpene profiles and strain mixtures
• Activations are effect values
• Forward pass computes final buff effects

Buff Orchestration (Simulation Loop)

The orchestrator runs buff rounds during simulation ticks:

1. Scan Phase

Orchestrator collects all active buffs across all characters:

# Orchestrator builds active-buff manifest
active_buffs:
  - character: player
    buff_ref: "skills/buff/buffs/INDEX.yml#caffeinated"
    remaining: 6
    stacks: 2

  - character: player
    buff_ref: "skills/buff/buffs/INDEX.yml#high"
    remaining: 8
    stacks: 1

  - character: bob-npc
    buff_ref: "skills/buff/buffs/INDEX.yml#drunk"
    remaining: 4
    stacks: 1

2. Event Generation

Create buff-tick events with pointers:

buff_round:
  tick: 42
  events:
    - type: buff-simulate
      character: player
      buff: caffeinated
      simulate_js: "subject.energy_effective += 20; subject.focus_effective += 15;"

    - type: buff-simulate
      character: player
      buff: high
      simulate: "Deep thoughts about random topics, food cravings"
      simulate_js: "if (Math.random() < 0.3) world.emit('*ponders existence*');"

    - type: buff-simulate
      character: bob-npc
      buff: drunk
      simulate: "Occasional slurred speech, may say embarrassing things"

3. LLM Simulation Prompt

Orchestrator instructs LLM to enumerate and simulate:

prompt: |
  BUFF ROUND — Tick 42

  Enumerate and simulate each active buff:

  1. PLAYER — Caffeinated (6 ticks remaining, 2 stacks)
     Effect: +20 energy, +15 focus per stack
     Simulate: Apply effects, note jitteriness if 2+ stacks

  2. PLAYER — High (8 ticks remaining)
     Effect: -25 stress, +20 creativity, +30 hunger
     Simulate: "Deep thoughts about random topics, food cravings"
     → Narrate any random musings or munchie urges

  3. BOB — Drunk (4 ticks remaining)
     Effect: +30 confidence, -25 focus, -30 judgement
     Simulate: "Occasional slurred speech, may say embarrassing things"
     → Decide if Bob says something regrettable this tick

  For each buff:
  - Apply stat modifications to _effective values
  - Run simulate behavior (chance-based events)
  - Check is_finished conditions
  - Decrement remaining duration
  - Remove expired buffs, trigger spawns_after

  Return updated character states and any narration.

4. Buff Lifecycle Per Tick

┌─────────────────────────────────────────────────────────────────┐
│                        BUFF TICK                                 │
├─────────────────────────────────────────────────────────────────┤
│                                                                  │
│  For each character:                                             │
│    For each active buff:                                         │
│                                                                  │
│      1. APPLY EFFECTS                                            │
│         stat_effective += buff.effect × buff.stacks             │
│                                                                  │
│      2. RUN SIMULATE                                             │
│         Execute simulate_js OR let LLM interpret simulate       │
│         (chance-based events, narration, random behaviors)      │
│                                                                  │
│      3. CHECK IS_FINISHED                                        │
│         If is_finished_js returns true → mark for removal       │
│         If remaining <= 0 → mark for removal                    │
│                                                                  │
│      4. DECREMENT DURATION                                       │
│         remaining -= 1                                           │
│                                                                  │
│      5. HANDLE EXPIRATION                                        │
│         If marked for removal:                                   │
│           - Remove buff from character                           │
│           - Trigger spawns_after buffs (with delay/chance)      │
│           - Emit buff-expired event                              │
│                                                                  │
│      6. HANDLE INTERACTIONS                                      │
│         Check for cancels, boosts, replaces, merges             │
│         Apply buff-on-buff effects                               │
│                                                                  │
└─────────────────────────────────────────────────────────────────┘

5. Compiled vs Interpreted

buff:
  id: drunk
  # LLM interprets this for narration
  simulate: "Occasional slurred speech, may say embarrassing things"
  # Engine runs this for mechanics
  simulate_js: |
    if (Math.random() < 0.2) {
      world.emit(subject.name + " slurs something incomprehensible");
    }

6. Attention Concentration (Time-Slicing)

The event-based design concentrates LLM attention on specific tasks:

┌──────────────────────────────────────────────────────────────────┐
│              LLM ATTENTION TIME-SLICING                          │
├──────────────────────────────────────────────────────────────────┤
│                                                                   │
│  Instead of: "Simulate everything at once" (diffuse attention)   │
│                                                                   │
│  We do: Series of focused micro-tasks                            │
│                                                                   │
│  ┌─────────┐   ┌─────────┐   ┌─────────┐   ┌─────────┐          │
│  │ Buff 1  │ → │ Buff 2  │ → │ Buff 3  │ → │ Buff 4  │          │
│  │ PLAYER  │   │ PLAYER  │   │ BOB     │   │ ROOM    │          │
│  │ caffein │   │ high    │   │ drunk   │   │ haunted │          │
│  └─────────┘   └─────────┘   └─────────┘   └─────────┘          │
│       ↓             ↓             ↓             ↓                │
│   [focused]    [focused]    [focused]    [focused]               │
│   attention    attention    attention    attention               │
│                                                                   │
└──────────────────────────────────────────────────────────────────┘

Why this works:

Iteration Pattern:

# Orchestrator feeds LLM one task at a time
iteration_1:
  focus: "PLAYER's Caffeinated buff"
  context: [player_state, buff_definition, tick_number]
  task: "Apply effects, check finish condition, narrate if needed"
  output: [updated_state, narration, events]

iteration_2:
  focus: "PLAYER's High buff"
  context: [player_state, buff_definition, tick_number]
  task: "Apply effects, chance of munchies event, narrate thoughts"
  output: [updated_state, narration, events]

# ... and so on

Benefits:

1. Focused attention — LLM only thinks about one buff
2. Predictable structure — Same input/output format each time
3. Debuggable — Can trace exactly which buff caused what
4. Parallelizable — Independent buffs can run in parallel
5. Interruptible — Can pause/resume between iterations
6. Cacheable — Compiled _js buffs skip LLM entirely

Speed-of-Light Compatible:

This fits the speed-of-light pattern — many focused micro-operations in a single LLM call, or batched across calls:

# Single call, multiple focused tasks
prompt: |
  Process these buff events in sequence:

  [1/4] PLAYER — Caffeinated
  → Apply: energy +40, focus +30 (2 stacks)
  → Check: is_finished? No (6 remaining)
  → Output: state changes only

  [2/4] PLAYER — High
  → Apply: stress -25, creativity +20, hunger +30
  → Simulate: "Deep thoughts" — roll for musing
  → Output: state + optional narration

  [3/4] BOB — Drunk
  → Apply: confidence +30, focus -25
  → Simulate: "Slurred speech" — roll for embarrassment
  → Output: state + optional narration

  [4/4] LIBRARY-SPIRIT — Haunted
  → Apply: error_rate +15, mood_influence -20
  → Simulate: "Poltergeist activity" — roll for book fall
  → Output: room effects + optional event

Stacking

• Same source: Doesn't stack — refresh duration instead
• Different sources: Stack additively up to category limit

Category Limits

terpene_effects: 3
charm_effects: 5
consumable_effects: 4
negative_effects: 3  # 3+ same negative = LEGENDARY

Synergies

Some buffs COMBINE into stronger effects:

• Myr + Lily = "Sedation Stack"
• Lemon + Pine = "Focus Boost"
• All 8 kittens = "ENTOURAGE EFFECT" (legendary)

Negative Buffs (Curses)

Curses are just shitty buffs. Same structure, negative effects.

buff:
  name: "Scratched"
  source: "Failed BELLY RUB"
  effect: { hp: -1, visible_marks: true }
  duration: "Until healed"

Persistent Curses

Long-term negative buffs with lift conditions:

buff:
  name: "Curse of Darkness"
  effect: { lamp_efficiency: -25% }
  duration: conditional
  lift_condition: "Light 3 dark places"
  reward_on_lift: "LIGHT-BEARER title"

Duration Types

Natural Language Durations

We're not tracking real time — the LLM interprets and makes its best guess:

• "forever"
• "5 minutes"
• "a day"
• "until sunset"
• "randomly 50%"
• "a while"
• "briefly"
• "until you forget"

See time/ for full natural duration examples.

Decay

When LLM judges turn(s) have passed:

1. Decrement duration on timed buffs
2. Remove buffs that hit 0
3. Apply new buffs from current turn

Effective Derived Values: Flags Edition

This is the effective derived values protocol for booleans.

Same pattern:

• Numeric: base + sum(modifiers) = effective
• Boolean: base OR any(conditions) = effective flag

Push / Pull / Latch

The LLM can handle any combination:

# PULL — derived on demand, not stored
in_darkness: (room.lit == false) AND (has_lamp == false)

# PUSH — buff explicitly sets
buff:
  sets_flags: [urgent_situation]

# LATCH — persists in state until cleared
player:
  visited_rooms: [room-a, room-b]  # grows, never shrinks

Traditional reactive systems pick one mode. The LLM does all three simultaneously — it sees the whole context and figures out which pattern applies.

Tweening and Animation

Values don't have to snap — they can interpolate over time:

buff:
  name: "Warming Up"
  effect: { warmth: +3 }
  tween: ease-in    # Gradual increase
  duration: 5

animation:
  entering_room:
    from: hallway
    to: pub
    frames: [approaching, at_door, stepping_in, arrived]

The LLM narrates intermediate states. "You feel yourself warming up..." not just "You are warm now."

Velocity

Any reactive variable can have a rate of change:

energy:
  value: 5
  velocity: -1      # Draining 1 per turn

trust:
  value: 45
  velocity: +3      # Building rapport

mood:
  value: "content"
  velocity: "improving"  # Semantic velocity works too

The LLM reads velocity to predict and narrate: *"You're running low on energy and fading fast..."* vs *"Low energy but recovering."*

Physics Simulation

Extend to full 2D/3D cartoon physics:

thrown_ball:
  position: [5, 3]
  velocity: [2, 4]       # Moving up-right
  acceleration: [0, -1]  # Gravity pulling down

bouncing:
  elasticity: 0.8        # Loses 20% on bounce

cartoon_physics:
  hang_time: true        # Pause at apex
  squash_stretch: true   # Deform on impact
  delayed_fall: true     # Look down first, then fall

The LLM narrates physics with cartoon timing:

*The ball arcs gracefully upward... hangs for a moment at the peak... then plummets, SQUASHING flat against the floor before bouncing back slightly less enthusiastically.*

Works for:

• Thrown objects (ball, inventory items)
• Character movement (jumping, falling, knockback)
• Environmental effects (swinging doors, rolling boulders)
• Looney Tunes logic (run off cliff, pause, look down, THEN fall)
• Temperature cooling or warming (ice cream melting, water freezing)

Commands