Goal

Let users plug in a Mistral API key as an LLM provider (alongside Anthropic, OpenAI, OpenRouter, Google, and OpenAI-compatible).

Background

The AI watcher calls an LLM adapter to decide on next prompts / summaries / HITL asks. Adapters live in src/server/services/ai/llm/. Each one implements the LlmAdapter interface (types.ts). Mistral speaks a simple JSON API over HTTPS, similar to OpenAI chat completions.

Where to start

• Read src/server/services/ai/llm/anthropic.ts as a reference (also openai-compatible.ts which covers the OpenAI chat shape).
• Create src/server/services/ai/llm/mistral.ts exporting createMistralAdapter({ apiKey, baseUrl? }).
• Wire it in src/server/services/ai/llm/registry.ts — add "mistral" to ProviderKind in types.ts, return the new adapter from getAdapter().
• Add pricing in src/server/services/ai/llm/pricing.ts (mistral-small, mistral-large, mistral-nemo input/output cents/1M).
• Add a PROVIDER_KINDS entry in src/server/routes/ai.ts and the UI dropdown in src/web/components/settings/AiSettingsPanel.tsx.

Acceptance criteria

• POST /api/v1/ai/providers with kind: "mistral" persists cleanly.
• The existing watcher test (src/server/services/ai/llm/llm.test.ts) is extended with a Mistral mock covering happy-path + 429 + 5xx.
• "Load available models" in the provider form works (Mistral has GET /v1/models).
• Typecheck + Biome + tests all green.