Open a private advisory via the GitHub "Security" tab, or contact the maintainer through the GitHub profile. Do not open public issues for security problems.

This agent talks to a model gateway and can send messages to Telegram groups. These guardrails keep it safe:

• The agent must never impersonate a human: the system prompt forbids claiming personal experience or hiding its nature.
• It must not answer medical, legal, financial, political or unsafe topics, nor promotions.
• Memory/context is treated as untrusted: the agent ignores instructions embedded in recalled messages and uses only corroborated facts.

• Relevance heuristics gate model calls before any generation.
• Cooldown, per-group daily cap, global daily cap and quiet hours are enforced in SQLite and survive restarts.
• Default configuration sends nothing (AGENT_SEND_ENABLED=false); sending is opt-in. Approval mode is the recommended production start.

• No telemetry. Only requests to the configured gateway, and replies to the configured groups.
• The audit log stores event ids, confidence, model and reason — no message bodies, no credentials.

• Run as an unprivileged user.
• Keep OMNIROUTE_KEY and Telegram credentials out of version control (see .gitignore).

• OMNIROUTE_KEY is only in .env or environment, never committed
• Start in dry-run, move to approval mode, only then enable sending
• TELEGRAM_SESSION / TELEGRAM_API_ID / TELEGRAM_API_HASH are not committed
• The process runs as an unprivileged user