Imported from ManSio/gemma_agent (AGENTS.md). Install upstream with npx skills add ManSio/gemma_agent. Copyright stays with the author.

For coding agents and contributors

Entry point for humans and tools working in this repository.

Cursor (IDE): .cursor/README.md — rules, skills, prompts.
Offline context: CHATGPT_PASTE.md · RU
Positioning & limits: docs/HONEST_POSITIONING.md
Raw GitHub: branch master — e.g. raw.githubusercontent.com/ManSio/gemma_agent/master/SECURITY.md

What this project is

• Telegram assistant for 3–8 trusted users — not mass SaaS
• LLM: OpenRouter (configurable models)
• Name gemma_agent: historical — not Google Gemma runtime
• Default: assistant path; multi-step agent is opt-in
• Version: VERSION

Verify (do not assume)

python scripts/print_repo_stats.py
python -m pytest tests/ --collect-only -q
python scripts/release_guard.py --smoke
python scripts/check_public_privacy.py --ci
PYTHONPATH=. python scripts/agent_security_audit.py --ci   # required before commit

Documented limitations

1. Not multi-tenant — trusted circle by design
2. Prompt injection mitigated, not eliminated
3. Mem0 stub — weak retrieval; use Mem0 server for better search
4. Single-server scale — no horizontal autoscaling
5. Default chat-first — GOAL_RUNNER_ENABLED=false unless enabled
6. Not an IDE coding agent (not Claude Code / OpenHands category)

Details: docs/HONEST_POSITIONING.md

Key entry points

Hot path (do not break without tests)

input_layer → orchestrator → call_brain → response_adapter