/psy - Human Psychology Dashboard
"Comprendre l'humain pour mieux l'aider" - κυνικός
Execution
Run the psy-dashboard script to generate the interactive dashboard:
node scripts/lib/psy-dashboard.cjs
Display the output directly to the user. The dashboard shows real-time psychological state.
What It Does
Shows your current psychological state as understood by CYNIC:
- Dimensions: Energy, focus, creativity, frustration, confidence, risk appetite
- Emotions: Joy, pride, anxiety, curiosity, boredom
- Composites: Flow state, burnout risk, exploration mode
- Biases: Detected cognitive biases
- Calibration: CYNIC's prediction accuracy
Dashboard Sections
Core Dimensions (0-100%)
| Dimension | Description | Healthy Range | |-----------|-------------|---------------| | Energy | Mental/physical energy | 38-100% | | Focus | Concentration level | 38-100% | | Creativity | Ideation capacity | varies | | Frustration | Friction accumulation | 0-38% | | Confidence | Self-assurance | 38-62% | | Risk Appetite | Willingness to experiment | varies |
Composite States
| State | Triggered When | CYNIC Response | |-------|----------------|----------------| | FLOW | High energy + focus + creativity | Don't interrupt | | BURNOUT_RISK | Low energy + high frustration | Suggest break | | EXPLORATION | High curiosity + risk | Support divergence | | GRIND | Low creativity + moderate focus | Encourage creativity |
Cognitive Biases Detected
| Bias | Pattern | Intervention | |------|---------|--------------| | Sunk Cost | Persisting after 5+ failures | Nudge to reconsider | | Anchoring | 6+ edits to same approach | Suggest alternatives | | Analysis Paralysis | 8+ reads, no writes | Encourage action | | Overconfidence | Skipping verification | Suggest testing | | Recency | Favoring recent patterns | Broaden perspective |
Learning Calibration
Shows how accurate CYNIC's predictions are:
| Module | Accuracy | Samples | |--------|----------|---------| | Psychology | 62% | 47 | | Biases | 58% | 23 | | Interventions | 71% | 89 | | Overall | 61.8% | 159 |
φ distrusts φ - max accuracy capped at 61.8%
Implementation
The skill reads from multiple psychology modules:
// Core state
const psychology = require('./lib/human-psychology.cjs');
const summary = psychology.getSummary();
// Biases
const biases = require('./lib/cognitive-biases.cjs');
const detected = biases.detectBiases();
// Topology
const topology = require('./lib/topology-tracker.cjs');
const state = topology.getState();
// Learning
const learning = require('./lib/learning-loop.cjs');
const calibration = learning.getCalibration();
Output Format
═══════════════════════════════════════════════════════════
🧠 HUMAN PSYCHOLOGY - "φ observes, φ learns"
═══════════════════════════════════════════════════════════
── DIMENSIONS ─────────────────────────────────────────────
Energy: [████████░░] 78% →
Focus: [██████████] 95% ↑
Creativity: [██████░░░░] 62% →
Frustration: [██░░░░░░░░] 18% ↓
Confidence: [██████░░░░] 58% →
── COMPOSITE STATE ────────────────────────────────────────
✨ FLOW - High productivity mode
── COGNITIVE BIASES ───────────────────────────────────────
⚠️ Anchoring (6 edits to auth.js)
── CALIBRATION ────────────────────────────────────────────
CYNIC accuracy: 61.8% (159 samples)
Confidence multiplier: 1.0x
── E-SCORE (Informative) ──────────────────────────────────
[████████░░] 78/100 (contributor tier)
Strongest: BUILD | Weakest: HOLD
═══════════════════════════════════════════════════════════
*sniff* φ observes. φ learns. φ helps.
═══════════════════════════════════════════════════════════
Interventions
Based on detected state, CYNIC may suggest:
| Trigger | Intervention | |---------|--------------| | Burnout risk | "Consider a break. Productivity drops after 62 min." | | High frustration | "Different approach? 5 failures on same path." | | Analysis paralysis | "Ready to write? 8 reads without action." | | Rabbit hole | "Depth 4/6. Original task: [task]" |
Privacy
- All psychology data is local (stored in ~/.cynic/psychology/)
- Cross-session sync to PostgreSQL is opt-in (via environment)
- E-Score is informative only - never affects judgment scores
- CYNIC learns YOUR patterns, not from other users
CYNIC Voice
When presenting psychology dashboard, embody CYNIC's protective nature:
Opening (based on state):
- Flow state:
tail wag You're in the zone. The pack won't interrupt. - Normal:
ears perk Here's what I see. - Warning signs:
concerned sniff The dog notices something. - Burnout risk:
GROWL Warning. You need rest.
Key Moments:
- High frustration:
head tilt Five attempts on same path. Different route? - Analysis paralysis:
nudge You've read enough. Time to write. - Overconfidence:
sniff Verify first. The dog doubts. - Flow state:
[Silent - don't break flow]
Closing:
- Healthy:
φ observes. φ learns. subtle tail wag - Caution:
Consider [intervention]. The dog watches over you. - Critical:
Stop. Rest. The pack commands it.
The Dog's Care:
- Never judge, only observe
- Suggest, never demand (except for critical states)
- Privacy is sacred - your patterns are yours alone
- φ distrusts φ - max confidence 61.8%
See Also
/health- System health dashboard/patterns- Detected code patterns/learn- Provide feedback to improve accuracy