# Plan: Add Mistral AI as Alternative Chat Provider ## Context bDS currently routes all AI chat through the OpenCode Zen gateway (`opencode.ai/zen/v1/...`) with two code paths: Anthropic Messages API and OpenAI-compatible. The user wants Mistral AI added as a direct alternative provider with frontier models that support chat completion, tool use, and vision. Mistral's API is OpenAI-compatible (`api.mistral.ai/v1/chat/completions`), making integration straightforward. **Important architecture facts:** - HTTP requests are currently **non-streaming** (full response body collected, text emitted after each complete call) — to be converted to SSE streaming as part of this work - Neither `sendAnthropicMessage()` nor `sendOpenAIMessage()` currently sets `tool_choice` - `sendOpenAIMessage()` does **not** convert `view_image` results to `image_url` format — they are JSON-stringified - `generateConversationTitle()` is hardcoded to `claude-haiku-4-5` via `ZEN_ANTHROPIC_URL` - `analyzeMediaImage()` is hardcoded to `claude-sonnet-4-5` via `ZEN_ANTHROPIC_URL` - `checkReady()` only checks the OpenCode key — blocks `sendMessage()` for keyless users ## Target Models | Model ID | Display Name | Vision | Tools | Context Window | Context Budget | |----------|-------------|--------|-------|----------------|----------------| | `mistral-large-2512` | Mistral Large 3 | yes | yes | 40k | 35,000 | | `mistral-medium-2508` | Mistral Medium 3.1 | yes | yes | 40k | 35,000 | | `mistral-small-2506` | Mistral Small 3.2 | yes | yes | 128k | 120,000 | | `devstral-small-2505` | Devstral Small | no | yes | 128k | 120,000 | | `devstral-large-2506` | Devstral 2 | no | yes | 256k | 240,000 | ## Files to Modify ### 1. `src/main/engine/OpenCodeManager.ts` - Core provider logic **A. Add Mistral constants** (near lines 23-25) - `MISTRAL_API_URL = 'https://api.mistral.ai/v1/chat/completions'` - `MISTRAL_MODELS_URL = 'https://api.mistral.ai/v1/models'` **B. Add Mistral models to `MODEL_DISPLAY_NAMES`** (lines 28-69) ``` 'mistral-large-2512': 'Mistral Large 3' 'mistral-medium-2508': 'Mistral Medium 3.1' 'mistral-small-2506': 'Mistral Small 3.2' 'devstral-small-2505': 'Devstral Small' 'devstral-large-2506': 'Devstral 2' ``` **C. Update `detectProvider()`** (lines 1839-1845) - Add: `if (id.startsWith('mistral') || id.startsWith('ministral') || id.startsWith('devstral') || id.startsWith('codestral') || id.startsWith('pixtral')) return 'mistral';` - This covers all current and foreseeable Mistral model prefixes (Mistral, Ministral, Devstral, Codestral, Pixtral) **C2. Update `formatModelName()` and `UPPERCASE_PREFIXES`** - `formatModelName()` (L1869) first checks `MODEL_DISPLAY_NAMES`, then auto-formats via hyphen splitting + capitalization - All 5 Mistral models are in `MODEL_DISPLAY_NAMES`, so auto-format is a fallback for future unknown models — no changes needed - `UPPERCASE_PREFIXES` (L72) contains `['gpt', 'glm']` — no Mistral prefixes need uppercasing, so no changes needed **D. Add Mistral API key storage** - New field: `private mistralApiKey: string = ''` - New methods: `setMistralApiKey()`, `getMistralApiKey()`, `validateMistralApiKey()` - Load on init from settings key `'mistral_api_key'` **E. Update `checkReady()`** - Return `ready: true` if **either** OpenCode key or Mistral key is set - Extend `ChatReadyStatus` to report per-provider availability, e.g. `providers: { opencode: boolean, mistral: boolean }` - Callers (`Sidebar.tsx`, `sendMessage()`) must gate on the relevant provider, not a single boolean **F. Parameterize `sendOpenAIMessage()` for Mistral (no separate method)** - Mistral uses the identical OpenAI-compatible chat/completions format — creating a separate `sendMistralRequest()` would be a near-duplicate - Instead, parameterize `sendOpenAIMessage()` to accept URL, API key, and provider-specific options: - Add params: `apiUrl: string`, `apiKey: string`, `providerOptions?: { parallelToolCalls?: boolean }` - `sendMessage()` determines provider via `detectProvider()` and calls `sendOpenAIMessage()` with the correct URL/key/options - For OpenCode OpenAI path: URL = `ZEN_OPENAI_URL`, key = `this.apiKey` - For Mistral: URL = `MISTRAL_API_URL`, key = `this.mistralApiKey`, `parallelToolCalls: false` - `tool_choice`: omit entirely for all OpenAI-compatible providers (default `"auto"` is correct) - `parallel_tool_calls: false` — set explicitly for Mistral only; our tool executor runs tools sequentially, so parallel tool calls would break the execution loop **F2. Add `MODEL_CONTEXT_BUDGETS` map** - New constant map `MODEL_CONTEXT_BUDGETS: Record` with per-model token budgets - `truncateToTokenBudget()` (L1654) currently defaults to `maxContextTokens = 150000` - In `sendAnthropicMessage()` and `sendOpenAIMessage()`: pass the model's context budget from the map (defaulting to 150,000 for OpenCode models) - The parameterized `sendOpenAIMessage()` looks up `MODEL_CONTEXT_BUDGETS[modelId]` for Mistral models and passes to truncation - Values from Target Models table (35k, 120k, 240k) **G. Fix tool-call message history in OpenAI-compatible path** - Within a single `sendMessage()` call, the tool loop correctly tracks tool results across rounds - However, `tool` role messages are not persisted to DB-backed conversation history — on conversation resume, the model loses context about prior tool results - Ensure `tool` role messages are included when persisting conversation history so cross-session continuity works - This affects all OpenAI-compatible providers (OpenCode OpenAI path + Mistral) **H. Fix vision in OpenAI-compatible path (affects Mistral too)** - `sendOpenAIMessage()` currently JSON-stringifies `view_image` results — no `image_url` conversion - Add `image_url` format conversion for `__isImageResult` objects in the OpenAI path: `{ type: 'image_url', image_url: { url: 'data:image/webp;base64,...' } }` - This fixes vision for all OpenAI-compatible providers, not just Mistral **I. Update `getAvailableModels()` — merge from both providers** - **Model list merge strategy**: fetch models from each configured provider's API endpoint and merge into a single list. When both keys are configured, return models from both; when only one key is set, return only that provider's models; when no key is set, return an empty list (UI disables the dropdown) - OpenCode models: fetched from existing OpenCode API (as today) - Mistral models: fetched from `GET https://api.mistral.ai/v1/models` when Mistral key is set; cross-reference returned IDs with `MODEL_DISPLAY_NAMES` to use display names + static `vision`/`contextBudget` metadata - Every model entry carries `provider: 'opencode' | 'mistral'` so the UI and engine can resolve the correct API URL + key - Invalidate `cachedModels`/`cachedModelsAt` when any provider key is added or removed **J. Update `generateConversationTitle()` — make configurable in Preferences** - Currently hardcoded to `claude-haiku-4-5` via `ZEN_ANTHROPIC_URL` with OpenCode key - Add a **"Title generation model"** preference in Settings so users can pick the cheapest model for this task - Default: `claude-haiku-4-5` (OpenCode) or `mistral-small-2506` (Mistral) based on available keys - Route to the correct provider URL + API key based on the selected model's provider - Must work with any configured provider, not just OpenCode **K. Update `analyzeMediaImage()` — make configurable in Preferences** (lines 2066-2192) - Currently hardcoded to `claude-sonnet-4-5` via `ZEN_ANTHROPIC_URL` - Add an **"Image analysis model"** preference in Settings so users can select a dedicated vision model independent of their chat model (e.g. use Devstral for chat but Mistral Large 3 for images) - **Only vision-capable models may be offered** — filter model list by a `vision` capability flag (e.g. Devstral models have no vision and must be excluded) - Default: `claude-sonnet-4-5` (OpenCode) or first vision-capable Mistral model based on available keys - Route to the correct provider URL + API key based on the selected model's provider - When routed to Mistral: use `image_url` format with base64 data URI - When routed to OpenCode/Anthropic: keep current Anthropic-native `image` block format **L. Update `analyzeTaxonomy()`** - Currently uses `this.apiKey` (OpenCode) for both Anthropic and OpenAI paths - Has an early-return guard `if (!this.apiKey)` that must become provider-aware — check Mistral key when provider is Mistral - When a Mistral model is selected: use Mistral API key + `MISTRAL_API_URL` - Must branch on provider to select correct key and URL **L2. Update `analyzeMediaImage()` API key guard** - Same issue: has `if (!this.apiKey)` early-return guard - Must become provider-aware — check the relevant provider's key based on the selected image analysis model - When routed to Mistral: check `this.mistralApiKey` instead of `this.apiKey` **M. Convert chat HTTP calls to SSE streaming** Currently `httpRequest()` buffers the entire response body before any text reaches the UI. Users wait 5–30s per API round with only a bouncing-dots indicator. All three providers (Anthropic, OpenAI, Mistral) support `stream: true` with SSE. **M1. Core streaming infrastructure — `httpRequestStream()`** - New method (~100 lines) — uses Node.js `https.request()` but reads `res` as a readable stream - Returns an async iterable of parsed SSE events (or accepts an `onEvent` callback) - SSE line protocol: lines separated by `\n\n`, each line prefixed with `event: ` or `data: ` - Must handle: - Buffering partial lines across `data` chunks (TCP may split mid-line) - Empty `data:` lines (keep-alive pings) - `data: [DONE]` sentinel — terminates the stream for OpenAI/Mistral (do NOT try to JSON.parse this) - Multiple `data:` lines between double-newlines (concatenate per SSE spec) - Supports `AbortSignal` — calls `req.destroy()` to terminate immediately - 120-second timeout matching existing `httpRequest()` - On non-2xx status: collect the error body (not streamed) and throw with parsed error message **M2. SSE parser for OpenAI/Mistral format** (~50 lines) OpenAI and Mistral use identical SSE event structure: ``` data: {"id":"...","choices":[{"delta":{"content":"Hello"}}]} data: {"id":"...","choices":[{"delta":{"tool_calls":[{"index":0,"id":"call_...","function":{"name":"search_posts","arguments":""}}]}}]} data: {"id":"...","choices":[{"delta":{"tool_calls":[{"index":0,"function":{"arguments":"{\"query\""}}]}}]} ... data: {"id":"...","usage":{"prompt_tokens":150,"completion_tokens":42,"total_tokens":192}} data: [DONE] ``` - **Text deltas**: `choices[0].delta.content` — emit via `onDelta(content)` immediately - **Tool call start**: `delta.tool_calls[i]` with `id` + `function.name` — begin accumulating arguments for tool call at index `i` - **Tool call argument fragments**: `delta.tool_calls[i].function.arguments` — append to argument accumulator string for index `i` - **Finish reason**: `choices[0].finish_reason === 'tool_calls'` or `'stop'` — signals end of this chunk - **Token usage**: arrives in the **final chunk before `[DONE]`** only if `stream_options: { include_usage: true }` is set in the request body — parse `usage.prompt_tokens`, `usage.completion_tokens`, `usage.total_tokens` - **`[DONE]` sentinel**: stop iteration, do NOT JSON.parse - After stream ends: if tool calls were accumulated, JSON.parse each tool's assembled arguments string and execute **M3. SSE parser for Anthropic format** (~60 lines) Anthropic uses named event types: ``` event: message_start data: {"type":"message_start","message":{"id":"...","model":"...","usage":{"input_tokens":150}}} event: content_block_start data: {"type":"content_block_start","index":0,"content_block":{"type":"text","text":""}} event: content_block_delta data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"Hello"}} event: content_block_start data: {"type":"content_block_start","index":1,"content_block":{"type":"tool_use","id":"toolu_...","name":"search_posts"}} event: content_block_delta data: {"type":"content_block_delta","index":1,"delta":{"type":"input_json_delta","partial_json":"{\"query\""}} event: content_block_stop data: {"type":"content_block_stop","index":1} event: message_delta data: {"type":"message_delta","delta":{"stop_reason":"tool_use"},"usage":{"output_tokens":42}} event: message_stop data: {"type":"message_stop"} ``` - **`message_start`**: extract `usage.input_tokens` (prompt tokens) + `usage.cache_read_input_tokens` + `usage.cache_creation_input_tokens` - **`content_block_start`** with `type: 'text'`: no-op (empty initial text) - **`content_block_start`** with `type: 'tool_use'`: record tool call `id` and `name` at block index - **`content_block_delta`** with `type: 'text_delta'`: emit via `onDelta(delta.text)` immediately - **`content_block_delta`** with `type: 'input_json_delta'`: append `delta.partial_json` to argument accumulator - **`content_block_stop`**: if tool block, JSON.parse the accumulated arguments for that block - **`message_delta`**: extract `usage.output_tokens` (completion tokens), `delta.stop_reason` - **`message_stop`**: stream complete - **`ping`**: ignore (keep-alive) - **`error`**: throw with `data.error.message` — handles mid-stream server errors (e.g. overloaded) **M4. Request body changes** - `sendAnthropicMessage()`: add `"stream": true` to request body - `sendOpenAIMessage()` (used for both OpenCode OpenAI and Mistral): add `"stream": true` and `"stream_options": { "include_usage": true }` to request body — this is **required** to receive token usage in streaming mode (without it, usage is omitted from streamed responses) **M5. Tool call accumulation during streaming** - Tool call arguments arrive as partial JSON fragments across many SSE events - Maintain a per-stream accumulator: `Map` keyed by tool call index - Append each `arguments` fragment to the accumulator string - On stream completion (finish_reason `tool_calls`/`tool_use`, or `content_block_stop` for Anthropic): JSON.parse the full accumulated arguments string and execute the tool - If JSON.parse fails on accumulated arguments, report a tool error to the model and continue **M6. Error handling during streaming** - **Non-2xx status on connection**: do NOT stream; collect the full error body and throw (same as current `httpRequest()` behavior) - **Mid-stream TCP disconnect / network error**: `res.on('error')` handler — emit whatever text was accumulated so far, then throw so the tool-call loop can surface the error to the user - **Mid-stream API error event**: Anthropic sends `event: error` with error details; OpenAI/Mistral return an error JSON in a `data:` line — detect and throw with parsed error message - **Abort during streaming**: `req.destroy()` triggers `res.on('error')` or `res.on('close')` — handle gracefully without surfacing as an error to the user (it's intentional cancellation) **What does NOT change:** - The renderer pipeline — `onDelta` → IPC `chat-stream-delta` → `appendStreamDelta` → React state → live Markdown rendering already works token-by-token; it just receives one big chunk today - `AbortController` abort support — `req.destroy()` stops the stream immediately instead of wasting a buffered response - The tool-call loop structure — still max 10 rounds, still sequential **What to keep non-streaming:** - `generateConversationTitle()` — small one-shot request, buffering is fine - `analyzeMediaImage()` — one-shot, no UI streaming needed - `analyzeTaxonomy()` — one-shot, no UI streaming needed - `validateApiKey()` / `validateMistralApiKey()` — small validation requests - Note: `validateMistralApiKey()` must call `GET https://api.mistral.ai/v1/models` with `Authorization: Bearer ${key}`. Mistral returns `{ data: [{ id, object, created, owned_by }] }` — check for HTTP 200 + non-empty `data` array. On 401, return invalid. On success, optionally cross-reference returned model IDs with `MODEL_DISPLAY_NAMES` to verify expected models are available **Estimated scope:** ~300 lines of new code in `OpenCodeManager.ts` (streaming adds ~100 lines vs original estimate) ### 2. `src/main/engine/ChatEngine.ts` - Settings persistence **A. Add Mistral key helpers** - Use existing generic `getSetting()`/`setSetting()` with key `'mistral_api_key'` — no dedicated methods needed, avoids unnecessary boilerplate - ChatEngine already exposes generic helpers for reading/writing the settings table **B. Default model is user-driven** - `getSelectedModel()` defaults to `'claude-sonnet-4-5'` - When user configures providers in Preferences, they explicitly select their default model — no automatic fallback logic needed - All surfaces (ChatPanel, AssistantSidebar, ImportAnalysisView) use this preference as default - If selected model's provider key is later removed: - `sendMessage()` returns a clear error string: "The selected model requires a {provider} API key. Configure it in Settings." - `checkReady()` still returns `ready: true` if any other provider is available - ChatPanel shows an inline error banner (not a toast) with a link/button to open Settings - i18n key: `chat.providerKeyMissing` — "The model '{model}' requires a {provider} API key. Go to Settings to configure it." - Add this key to all 5 locale files **C. Add per-purpose model preferences** - `getTitleModel()` / `setTitleModel(modelId)` — settings key `'chat_title_model'` - `getImageAnalysisModel()` / `setImageAnalysisModel(modelId)` — settings key `'chat_image_analysis_model'` - Both default to `null` (= use hardcoded defaults per provider) ### 3. `src/main/ipc/chatHandlers.ts` - IPC bridge **A. Add Mistral-specific handlers** - `chat:setMistralApiKey` - validate + persist Mistral key, invalidate model cache - `chat:getMistralApiKey` - return masked key - `chat:validateMistralApiKey` - test key against Mistral API **B. Update `chat:getAvailableModels`** - Include Mistral models when Mistral key is configured - Return provider info per model **C. Update `chat:checkReady`** - Report readiness for both providers independently **D. Update `getOpenCodeManager()` init** - Load `'mistral_api_key'` from settings on first call (alongside OpenCode key) - Call `manager.setMistralApiKey()` during init **E. Add per-purpose model preference handlers** - `chat:setTitleModel` / `chat:getTitleModel` — persist + load title generation model preference - `chat:setImageAnalysisModel` / `chat:getImageAnalysisModel` — persist + load image analysis model preference ### 4. `src/main/shared/electronApi.ts` - Type definitions **A. Extend `ChatModel` interface** - Add `provider: 'opencode' | 'mistral'` field (already optional, ensure populated) - Add `vision: boolean` field — indicates whether the model supports image inputs (used to filter the image analysis model dropdown) **B. Extend `ChatReadyStatus` interface** - Add `providers?: { opencode: boolean; mistral: boolean }` for per-provider status **C. Add Mistral IPC methods to `ElectronAPI.chat`** - `setMistralApiKey(key: string)` - `getMistralApiKey()` - `validateMistralApiKey(key: string)` **D. Add per-purpose model preference methods to `ElectronAPI.chat`** - `setTitleModel(modelId: string | null)` / `getTitleModel()` - `setImageAnalysisModel(modelId: string | null)` / `getImageAnalysisModel()` ### 5. `src/renderer/components/SettingsView/SettingsView.tsx` - UI settings **A. Add Mistral API key section** - Separate input field for Mistral API key (below OpenCode key) - Same pattern: masked display, change button, validation on save **B. Update model selector** - SettingsView uses a native `` dropdown - When both keys configured, show merged list from both providers; when only one key set, show only that provider's models - **Note**: `availableModels` state is currently typed as `{id: string; name: string}[]` — must be updated to `ChatModel[]` (which includes `provider` and `vision` fields) so provider grouping and vision filtering work **C. Add per-purpose model preferences** - "Title generation model" dropdown — select cheapest/fastest model for auto-titling conversations - "Image analysis model" dropdown — select a dedicated vision model for media metadata (independent of chat model, e.g. use Devstral for chat but Mistral Large 3 for images); **only show vision-capable models** (filter out models without vision support like Devstral) - Both show available models from all configured providers, grouped by provider - Both allow a "Default" option that auto-selects per provider defaults ### 6. `src/renderer/components/ChatPanel/ChatPanel.tsx` - Chat UI **A. Update model selector in chat** - ChatPanel uses a custom dropdown (CSS `model-dropdown` with `