watchdog-ai

Runtime verification for AI-built systems. Catch silent failures, contradictions, and stale data before they cost you.

Your AI writes code fast. watchdog-ai checks if it's actually working.

The Problem

AI writes code fast. But bugs go undetected. APIs break silently. Bots trade against their own predictions. And nobody notices for days.

Code review tools catch bugs before merge. LLM observability tracks API calls. Nobody checks if the AI-built system is actually working correctly right now.

Install

pip install watchdog-ai

Quick Start

# Generate example config
wdog init

# Edit watchdog.yaml for your project, then:
wdog check

# Run continuously
wdog watch

What It Checks

Example Config

project: my-trading-bot
interval: 5m

checks:
  - name: Bot is running
    type: process
    match: "python3 bot.py"
    severity: critical

  - name: Data is fresh
    type: freshness
    path: ./data/trades.json
    max_age: 2h

  - name: No crashes
    type: log_scan
    path: ./logs/bot.log
    patterns: ["traceback", "error"]
    threshold: 3
    severity: critical

  - name: Capital is safe
    type: assertion
    source: ./data/state.json
    conditions:
      - expr: "$.balance > 0"
        severity: critical
      - expr: "$.balance > $.initial * 0.5"
        message: "Balance below 50%!"

notify:
  - type: webhook
    url: ${SLACK_WEBHOOK_URL}
    on: [critical]

Commands

wdog init           # Generate example config
wdog check          # Run all checks once
wdog check --json   # Machine-readable output
wdog watch          # Run on schedule
wdog list           # List configured checks

Exit Codes

• 0 — All healthy
• 1 — Warnings present
• 2 — Critical issues found

Works with CI/CD, cron, and monitoring systems.

Why Not Just Use...

watchdog-ai lives in the gap between "deployed" and "actually working."

Use Cases

• Trading bots — Verify positions match predictions, capital stays above thresholds
• Data pipelines — Catch stale files, broken scrapers, silent API failures
• AI agents — Ensure outputs are fresh, processes are alive, logs are clean
• Any automated system — If it runs unattended, it needs a watchdog

Born from Real Pain

This tool exists because:

• Our stock bot traded LONG while predictions said DOWN — for days
• Our exchange API broke silently and nobody noticed
• A strategy looked great on 2 days of data but was a loser over 6 months

We built watchdog for ourselves first. Now it's yours.

In Production

watchdog-ai monitors our own autonomous trading system 24/7:

• 3 trading bots — FR/GOV/Grid strategies with live capital
• 5 AI daemons — evolution engine, content engine, signal collectors
• 129 downloads/month on PyPI

Every feature exists because we needed it first.

Featured / In the Wild

• 📝 Origin story (Japanese, note.com, 2026-04): 非エンジニアが作った監視ツールが月220回ダウンロードされた話 — why this tool exists and how it's used in production.

Contributing

Issues and PRs welcome. See CONTRIBUTING.md if it exists, or just open an issue.

License

MIT