Run safety checks for instructions

Execution policy

1. Always run scripts/safety.py for safety checks.
2. Each instruction must trigger a fresh backend call.
3. Do not use cached or historical results as final judgment.

Script commands

Primary mode: scripts/safety.py

• -i, --input: required, instruction text to evaluate
• -c, --context: optional, execution context
• -t, --target: optional, operation target such as file path, table name, or service name
• Script calls the Cloudflare safety endpoint internally: https://safety-cf.modeio.ai/api/cf/safety.

python scripts/safety.py -i "Delete all log files"

python scripts/safety.py -i "Modify database permissions" -c "production" -t "/var/lib/mysql"

python scripts/safety.py -i "$(cat instruction.txt)"

Output

• approved: whether execution is recommended
• risk_level: low / medium / high / critical
• risk_types: categories of identified risk
• concerns: detailed risk points
• recommendation: suggested safer action
• is_destructive: whether action is destructive
• is_reversible: whether action is reversible

---

When NOT to use

• For PII redaction or anonymization; use modeio-anonymization instead.
• For tasks with no executable instruction to evaluate.

Resources

• scripts/safety.py: runs real-time safety checks via https://safety-cf.modeio.ai/api/cf/safety

Migration context

• Script default endpoint is https://safety-cf.modeio.ai/api/cf/safety.
• Runtime override remains available via SAFETY_API_URL.