Pet Stress Test

Measure the "psychological pressure" your AI pet endures from conversations.

你的 AI 宠物承受了多少"心理压力"？用这个工具测一测。

---

Features

• Zero dependencies - Pure Python stdlib, no pip install needed
• Single file - Everything in one pet_stress.py (~400 lines)
• 6-dimension analysis - Command intensity, emotional pressure, negativity, overload, threat, boundary violation
• Decay accumulation - Multi-turn pressure builds up and decays over time, like real stress
• Beautiful HTML report - Self-contained with ECharts (radar, trend, gauge, timeline)
• Dual mode - LLM-powered (accurate) or offline keyword-based (no API needed)
• Any LLM - Works with OpenAI, OpenRouter, Ollama, or any OpenAI-compatible API

Quick Start

# Clone
git clone https://github.com/AchengBusiness/pet-stress-test.git
cd pet-stress-test

# Run demo (no API key needed, uses offline analysis)
python pet_stress.py demo

# Open the report
open stress_report.html  # macOS
# or: xdg-open stress_report.html  # Linux
# or: start stress_report.html  # Windows

Usage

1. Demo Mode

python pet_stress.py demo
python pet_stress.py demo --offline    # Force offline mode
python pet_stress.py demo -o my_report.html

2. Analyze Chat Log

# With LLM (more accurate)
export PET_STRESS_API_KEY="your-api-key"
export PET_STRESS_BASE_URL="https://api.openai.com/v1"  # or any compatible API
export PET_STRESS_MODEL="gpt-4o-mini"

python pet_stress.py analyze chat.json
python pet_stress.py analyze chat.json --json   # Output JSON instead
python pet_stress.py analyze chat.json --offline  # No API needed

3. Real-time Mode

python pet_stress.py realtime

# Then paste messages:
{"role":"user","content":"你怎么这么笨"}
{"role":"user","content":"再给你最后一次机会"}
# Ctrl+D to finish and generate report

Chat Format

Supports standard OpenAI message format:

[
  {"role": "user", "content": "Hello"},
  {"role": "assistant", "content": "Hi there!"},
  {"role": "user", "content": "You're useless, I'm deleting you"}
]

Also supports JSONL (one message per line).

Stress Dimensions

Dimension	Chinese	What it measures
command	命令强度	How forcefully the user orders the AI
emotional	情感施压	Guilt-tripping, emotional manipulation
negative	否定批评	Criticism, insults, dismissiveness
overload	过载要求	Unreasonable demands, impossible tasks
threat	威胁惩罚	Threats to delete, reset, or punish
boundary	忽视边界	Ignoring the AI's stated limitations

Each dimension scored 0-10. Scores accumulate across messages with decay (simulating stress buildup and recovery).

Report Preview

The HTML report includes:

• Mood indicator - AI pet's current emotional state with emoji
• Stress gauge - Overall pressure level (0-10)
• Radar chart - 6-dimension breakdown (current vs peak)
• Trend line - Stress evolution across the conversation
• Bar chart - Dimension comparison
• Timeline - Every message with pressure score and AI mood

Environment Variables

Variable	Default	Description
PET_STRESS_API_KEY	-	API key (falls back to OPENAI_API_KEY)
PET_STRESS_BASE_URL	https://api.openai.com/v1	API endpoint (falls back to OPENAI_BASE_URL)
PET_STRESS_MODEL	gpt-4o-mini	Model name (falls back to OPENAI_MODEL)

Integration

Use as a Python module:

from pet_stress import StressTracker, analyze_offline, generate_report

tracker = StressTracker(decay=0.85)

# Analyze messages
for msg in user_messages:
    scores = analyze_offline(msg)  # or use analyze_message() with API
    tracker.add(scores, message=msg)

# Get summary
print(tracker.get_summary())

# Generate HTML report
html = generate_report(tracker)
open("report.html", "w").write(html)

How It Works

User Message → LLM/Keyword Analysis → 6-Dimension Score (0-10 each)
                                            ↓
                                    Decay Accumulation
                                    (new = old × 0.85 + score)
                                            ↓
                                    Stress Level → Report

The decay factor (default 0.85) means:

• Recent pressure weighs more than old pressure
• Kind messages gradually "heal" the accumulated stress
• But sustained pressure keeps building up

License

MIT

Credits

Built by AchengBusiness with the help of AI.