Personality EngineSAFE

From the repository's own README, as read at the audited commit. Badges and raw HTML are left out.

clawhub install personality-engine

from personality_engine.engine import PersonalityEngine

engine = PersonalityEngine(user_id="[email protected]")

# Pass triggers through the engine
should_send, msg = await engine.process_trigger(
trigger_type="cross_platform",
raw_message="Divergence: Kalshi 52%, Poly 48%",
market_data={"spread": 4.0},
urgency_context={},
)

if should_send:
send_message(msg.content)

clawhub install kalshalyst kalshi-command-center polymarket-command-center prediction-market-arbiter xpulse portfolio-drift-monitor market-morning-brief personality-engine

The instruction file, verbatim from the audited commit — this is the text the model reads, and the surface the audit's instruction layer examines. Quoted here so you can judge it without cloning anything.

---
name: Personality Engine
description: "Six-system behavior engine that makes any OpenClaw agent feel alive. Editorial voice injects opinions. Selective silence knows when NOT to talk. Variable timing scores urgency with time-of-day awareness. Micro-initiations send ambient pings. Context buffer enables back-references to earlier messages. Response tracker adapts to engagement patterns. Domain-agnostic — works with trading agents, personal assistants, DevOps monitors, or any proactive agent. Part of the OpenClaw Prediction Market Trading Stack with default trading configuration."
---

# Personality Engine — 6-System Behavior Framework

**Goal**: Make any AI agent feel *alive* — opinions, awareness, judgment, memory, timing sense, and engagement sensitivity. Works with trading agents, notification systems, personal assistants, or any proactive agent. Not just data delivery.

## Architecture Overview

```
Trigger fires → engine.py orchestrator
              ↓
         selective_silence (should we stay silent?)
              ↓
         urgency_compute (how urgent is this 0.0-1.0?)
              ↓
         engagement_modifier (adjust for user response patterns)
              ↓
         variable_timing (schedule delivery based on urgency + time of day)
              ↓
         context_buffer (add back-references to earlier messages today)
              ↓
         editorial_voice (inject personality / opinions)
              ↓
         dedup (avoid repeats within rolling window)
              ↓
         send → iMessage (or other transport)
```

Plus two ambient systems:
- **micro_initiations**: Unprompted pings when conditions are met (quiet market, good streak, absence detected)
- **response_tracker**: Monitors engagement; adjusts urgency + suggests tuning

---

## System 1: Editorial Voice — Opinion Injection

**What**: Each trigger type gets a personality — opinions that vary based on market state, portfolio P&L, signal confidence.

**Per-trigger voice pools**:

### cross_platform (Kalshi vs Polymarket divergence)
- **Bullish divergence (>5%)**: "Big divergence. One of these markets is wrong."
- **Mild divergence (2-5%)**: "Mild divergence. Nothing screaming yet."
- **Stale divergence (6+ hours old)**: "Divergence is stale — markets may have already repriced."

### portfolio (user's holdings performance)
- **+15% or better**: "Good day. Portfolio's running."
- **+5% to +15%**: "Solid gains. Steady hand."
- **-5% to +5%**: "Flat day. Markets are grinding."
- **-5% to -15%**: "Rough patch. Check your stops."
- **-15% or worse**: "Heavy day. Buckle up for volatility."

### x_signals (social signal scanner)
- **Confidence ≥0.85 + matched position**: "Strong signal. This feels real."
- **Confidence 0.70-0.85**: "New signal on [topic]. Worth watching."
- **Confidence <0.70**: "Noise signal. Low confidence."

### edge (Kalshi edge detection)
- **Edge >3%**: "Fat edge. Worth a deep look."
- **Edge 1-3%**: "Mild edge. Keeping it on radar."
- **Edge <1%**: "Thin edge. Not worth the friction."

### morning (daily brief)
- **Monday**: "New week. Here's the lay of the land."
- **Friday**: "Friday rundown. What matters before the close."
- **Other**: "Daily digest."

### conflicts (overlapping triggers same hour)
- **2+ conflicts**: "Tomorrow's a mess. Multiple overlaps."
- **Lighter**: "Heads up — couple things hitting together."

**Customization point**: Add trigger types by extending the `VOICE_POOLS` dict in `editorial_voice.py`. Each entry maps `(trigger_name, market_state)` → list of opinion strings.

---

## System 2: Selective Silence — Knowing When NOT to Talk

**What**: Not every trigger fire deserves a message. Silent skips are explicit: "Skipped the brief — nothing worth your attention."

**Content quality checks per trigger**:

- **morning_is_boring**: If market vol <0.5%, no divergences, no edges → skip
- **divergence_is_stale**: If last message on this topic was <3 hours ago AND spread hasn't moved >0.5% → skip
- **signals_are_noise**: If all signals have confidence <0.65 AND no position matches → skip
- **edge_is_weak**: If all edges <1% → skip
- **portfolio_is_flat**: If daily P&L is -2% to +2% AND no major position changes → skip

**Silence cadence**:
- Max 1 silence message per day per user
- Only for *expected* triggers (morning, portfolio check, etc.)
- Never silence micro-initiations (those *are* the value)
- When silent, send explicit message: `"Skipped the brief — nothing worth your attention."`

**Customization point**: Adjust thresholds in `selective_silence.py`:
```python
SILENCE_THRESHOLDS = {
    'vol_floor': 0.5,           # % vol threshold for morning silence
    'divergence_age_limit': 3,  # hours
    'signal_confidence_floor': 0.65,
    'edge_floor': 1.0,          # %
    'portfolio_flat_range': 2.0 # % P&L range
}
```

---

## System 3: Variable Timing — Urgency Scoring + Time-of-Day Awareness

**What**: Schedule message delivery based on urgency (0.0-1.0) and time of day. A mild divergence at 6 AM gets sent immediately (threshold 0.9 before 7 AM). Same divergence at 10 PM gets held (threshold 0.35).

**Per-trigger urgency base**:
- **cross_platform**: (spread / 10%) * 0.6, capped at 1.0
  - 5% spread = 0.3 urgency
  - 10% spread = 0.6 urgency
  - 15%+ spread = 1.0 urgency
- **portfolio**: (abs(daily_pnl) / 10%) * 0.7, capped at 1.0
  - ±5% P&L = 0.35 urgency
  - ±15% P&L = 1.0 urgency
- **x_signals**: (confidence * 0.8) + (position_match ? 0.2 : 0), capped at 1.0
  - Confidence 0.85 + matched = 0.88 urgency
  - Confidence 0.70 + no match = 0.56 urgency
- **edge**: (edge_size / 5%) * 0.8, capped at 1.0
  - 2% edge = 0.32 urgency
  - 5% edge = 0.8 urgency
- **meeting**: (1.0 - minutes_away / 120) capped at 1.0
  - 30 min away = 0.75 urgency
  - 5 min away = 0.96 urgency

**Time-of-day delivery thresholds**:
- **Before 7 AM**: threshold 0.90 (almost everything gets sent)
- **7 AM - 9 AM**: threshold 0.75 (morning crunch — moderate bar)
- **9 AM - 10 PM**: threshold 0.45

SAFEgrade B · trust 89/100 Nothing in the source contradicts what it says it does. Grade A is reserved for packages that have also passed the behavioural sandbox.

What the source does

Filesystem

none-observed

Network

none-observed

Shell

none-observed

Dependencies

pinned

Secrets in source

none-found

Findings (3)

<key>k5Ey9KFlkqpj+SDkUw+5ED9lTA3En/qUi0zdrydUCH3kMWTE3Eh65NXnFCaxlY2omY2JHnlEoK7Li7oOEvM7eG5VPdcO/sFlMfoCRdnLYdepJ+uLzYwOWR8W4yQVve/clxVFTVRL4DFleKInGdpAxIbHZT2yi4ADAMENls1N1XSLojRuqXePXDeAT/4Mv4TTx0s

msg_hash = hashlib.md5(message.encode()).hexdigest()

msg_hash = hashlib.md5(message.encode()).hexdigest()

Gates applied: no_behavioural_pass.