DeepSeek And LLM Configuration

DeepSecLLM routes Shield reviews to the chat model and Spear reasoning to the reasoning model. Providers are independent clients with a deterministic fallback chain, so a failed DeepSeek request cannot reuse the wrong endpoint for OpenAI, Claude, or a local server.

[llm]
provider = "deepseek"
fallbacks = ["claude", "openai", "local"]

[llm.providers.deepseek]
api_key = ""
base_url = ""

[llm.providers.openai]
api_key = ""
base_url = ""

[llm.shield]
chat_model = "deepseek-chat"

[llm.spear]
reason_model = "deepseek-reasoner"
explore_model = "deepseek-chat"

Use DEEPSEC_DEEPSEEK_API_KEY for a provider-specific secret or DEEPSEC_LLM_API_KEY for the primary provider. DeepSeek requests enable prompt caching through extra_body.prompt_cache; usage-only metadata containing provider, model, token counts, and estimated cost is appended to ~/.deepsec/runs/llm-usage.jsonl. Prompts and credentials are never written there.