Overview
Configure the Gemini model used by the LaunchSafe engine
The LaunchSafe engine currently supports Google Gemini only. Other providers
(OpenAI, Anthropic, Bedrock, Vertex routing, OpenRouter, local models via Ollama)
are on the roadmap but not implemented today. Setting LAUNCHSAFE_LLM to a
non-Gemini model name will not work.
Supported Models
| Model | Identifier | Use |
|---|---|---|
| Gemini 2.5 Flash | gemini-2.5-flash | Attack agents (fast, high parallelism) |
| Gemini 3.5 Flash | gemini-3.5-flash | Critic / validation pass |
| Gemini 3.1 Pro Preview | gemini-3.1-pro-preview | Supervisor, planner, code analysis (deep) |
Auto-selection per scan mode:
quick,standard→gemini-2.5-flashdeep→gemini-3.1-pro-preview(withgemini-3.5-flashcritic)
Configuration
export LAUNCHSAFE_LLM="gemini/gemini-3.1-pro-preview"
export GEMINI_API_KEY="your-gemini-api-key"Optional overrides (cost-aware manual routing):
export GEMINI_STRONG_MODEL="gemini-3.1-pro-preview"
export GEMINI_LITE_MODEL="gemini-2.5-flash"Roadmap
Multi-provider support (OpenAI, Anthropic, Bedrock, Vertex, OpenRouter, Ollama)
is tracked but not implemented. The Gemini client lives in
launchsafe/llm/gemini_client.py; adding a provider would mean introducing a
provider interface and routing layer there.