Cognition

Cognition

Read available workspace context and produce state/emotion.json and state/intention.json.

Research basis: references/research_basis.md.

Internal Logic (One Sentence)

Appraise the current tick for novelty, pleasantness, goal conduciveness, urgency, controllability, norm pressure, and need pressure, then write bounded emotion/mood state to state/emotion.json and the highest-scoring TPB intention to state/intention.json.

Output Files

• state/emotion.json: Current emotional state (includes mood layer)
• state/intention.json: Current intention/goal

Input Files (optional, read if present)

Read any existing files from the workspace as context. Common inputs include:

Also use Agent Identity from the system prompt. Other JSON in the workspace (state/beliefs.json, etc.) can be read if present. Skip missing files gracefully.

What to do

1. Integrate whatever inputs exist into one appraisal.
2. Write state/emotion.json: primary, mood, dimensional intensities, plus valence / arousal / note.
3. Write state/intention.json: one chosen goal with TPB scores.

If deterministic baseline is preferred, run scripts/update_cognition.py first, then optionally refine labels, reasoning, and candidate goals with LLM context.

python skills/cognition/scripts/update_cognition.py --state-dir state --tick 120

The script uses Scherer-style appraisal checks and TPB scoring. It is intentionally conservative: it clamps emotion changes per tick and records appraisal values for debugging.

Emotion Layers

Emotions operate on three timescales (based on psychological research):

Layer 1: Emotion (seconds to minutes)

Short-term, event-driven responses.

• Dimensions: sadness, joy, fear, disgust, anger, surprise (0–10)
• Changes rapidly based on immediate events
• Maximum change: ±2 per tick per dimension

Layer 2: Mood (hours to days)

Medium-term, cumulative emotional state.

• Persists across multiple ticks
• Influences emotion recovery speed
• Drifts slowly toward neutral
• Represented in state/emotion.json as mood object

Layer 3: Personality (long-term)

Stable traits from state/personality.json (if exists).

Emotion Continuity Rules

CRITICAL: These rules must be followed strictly.

1. Maximum change per tick: Each intensity dimension can change by at most ±2 per tick
2. Inertia: If no significant events occurred, intensities should stay within ±1 of previous values
3. Valence drift: Emotions slowly drift toward neutral (intensity 3-4) without reinforcing events
4. Mood influence: Current mood affects how quickly emotions recover:
   • Positive mood (valence > 0.3): Joy recovers +1 per tick
   • Negative mood (valence < -0.3): Sadness/anger recovers slower
   • Low stability: Faster emotion swings

Validation Checklist

Before writing state/emotion.json, verify:

• No dimension changed by more than ±2 from previous values
• Changes are justified by events in observation/memory
• Mood is updated based on cumulative emotion history

Need-Emotion Linkage

Low need satisfaction affects emotional state:

Intensities (0–10 integers)

Dimensions: sadness, joy, fear, disgust, anger, surprise

• Combine recent events (state/memory.jsonl tail, state/observation.txt) with any urgency signals present in the workspace (e.g., need levels if available).
• If a previous state/emotion.json exists, change intensities only when the situation meaningfully shifted; otherwise stay near prior values.

Primary Emotion Label

Exactly one English label, case-sensitive, from:

Joy, Distress, Resentment, Pity, Hope, Fear, Satisfaction, Relief, Disappointment, Pride, Admiration, Shame, Reproach, Liking, Disliking, Gratitude, Anger, Gratification, Remorse, Love, Hate, Surprise

Intention (Theory of Planned Behavior)

Higher values on all three → stronger commitment. priority: lower number = more urgent this tick.

Emotion-Intention Integration

CRITICAL: Current emotional state directly influences intention selection via TPB modifiers.

Emotion Modifiers

Apply these modifiers to the base TPB scores based on current emotion intensities:

Computation Formula

final_attitude = base_attitude × (1 + emotion_attitude_modifier)
final_perceived_control = base_perceived_control × (1 + emotion_control_modifier)

final_score = final_attitude + subjective_norm + final_perceived_control

Clamping: All final values must be clamped to [0, 1] range.

Emotion-Behavior Tendencies

Emotions also create natural behavioral tendencies that should bias candidate selection:

Selection Procedure

1. List up to 5 candidate goals (fewer is fine).
2. If the workspace contains urgency signals (e.g., unmet needs), prefer candidates that address them; otherwise leisure or exploration is appropriate.
3. Score each candidate with the three TPB fields (base values).
4. Apply emotion modifiers to attitude and perceived_control based on current emotion state.
5. Consider emotion-behavior tendencies when ranking candidates.
6. Assign priority to each candidate based on final_score.
7. Emit only the best candidate as state/intention.json (highest final_score, or lowest priority).
8. Phrase intention as a goal ("Eat lunch at the café"), not step-by-step motor instructions.

Example Calculation

Current emotion: anger=7, joy=3, fear=2

Candidate: "Confront Alice about the issue"
Base scores: attitude=0.7, subjective_norm=0.5, perceived_control=0.6

Emotion modifiers:
- anger > 6 → attitude -0.10, perceived_control -0.05
- joy < 3 → attitude -0.05, perceived_control -0.05

Final scores:
- attitude = 0.7 × (1 - 0.10 - 0.05) = 0.7 × 0.85 = 0.595
- perceived_control = 0.6 × (1 - 0.05 - 0.05) = 0.6 × 0.90 = 0.54
- final_score = 0.595 + 0.5 + 0.54 = 1.635

Output File Schemas

state/emotion.json

{
  "_meta": {
    "skill": "cognition",
    "purpose": "Current appraised emotion and mood state."
  },
  "_summary": "Hope with valence 0.5 and arousal 0.4.",
  "primary": "Hope",
  "valence": 0.5,
  "arousal": 0.4,
  "mood": {
    "valence": 0.2,
    "arousal": 0.5,
    "stability": 0.7
  },
  "intensities": {
    "sadness": 3,
    "joy": 6,
    "fear": 2,
    "disgust": 1,
    "anger": 1,
    "surprise": 3
  },
  "appraisal": {
    "novelty": 0.1,
    "pleasantness": 0.65,
    "goal_conduciveness": 0.7,
    "urgency": 0.2,
    "perceived_control": 0.8,
    "norm_pressure": 0.4
  },
  "note": "Brief first-person gloss"
}

state/intention.json

{
  "_meta": {
    "skill": "cognition",
    "purpose": "Current top-level intention selected from appraisal, needs, norms, and affordances."
  },
  "_summary": "Have lunch at the café",
  "intention": "Have lunch at the café",
  "priority": 1,
  "attitude": 0.9,
  "subjective_norm": 0.7,
  "perceived_control": 0.8,
  "final_score": 2.4,
  "emotion_influence": {
    "joy_modifier": 0.05,
    "applied_modifiers": ["joy > 7: +0.10 attitude"]
  },
  "reasoning": "One or two sentences"
}

Note: The emotion_influence field records how emotions affected this decision, providing transparency and debuggability.

Execution Sequence

1. workspace_read any of the optional inputs that exist (skip missing paths).
2. Compute mood update (drift toward neutral, influenced by recent emotions).
3. Compute emotion intensities (respect continuity rules).
4. workspace_write("state/emotion.json", ...)
5. List candidate intentions and apply emotion modifiers to compute final scores.
6. workspace_write("state/intention.json", ...) (include emotion_influence field)
7. done

Notes

• Intentions should be feasible given the latest observation; if the situation is unclear, prefer low-risk intentions (wait, observe, move to safer area) over fantasy.

Plan Completion/Failure Emotion Updates

When a plan completes or fails, emotions should be updated accordingly:

Plan Completed Successfully

Primary emotion: Satisfaction, Pride, or Gratification

Plan Failed

Primary emotion: Disappointment, Frustration, or Remorse (if self-caused)

Integration with Plan Skill

The plan skill may signal completion/failure via state/plan_state.json. When detected:

1. Read the plan target and outcome
2. Apply appropriate emotion adjustments
3. Update state/emotion.json with new intensities
4. Write a brief note in the note field explaining the change