Phase 2: providers + chat + tasks + IPC rewire

src/main/engine/ai/chat.ts

@@ -0,0 +1,513 @@
+/**
+ * ChatService — streaming chat using AI SDK's streamText().
+ *
+ * Replaces OpenCodeManager's sendAnthropicMessage/sendOpenAIMessage/
+ * streaming.ts with a single, provider-agnostic code path.
+ *
+ * AI SDK handles:
+ *   - SSE parsing, reconnection, abort
+ *   - Provider-specific request/response format (Anthropic Messages, OpenAI Chat Completions)
+ *   - Tool call/result loop (maxSteps)
+ *   - Token usage extraction
+ */
+
+import { streamText, generateText, stepCountIs } from 'ai';
+import type { ModelMessage, LanguageModelUsage } from 'ai';
+import type { BrowserWindow } from 'electron';
+import type { ChatEngine, ChatMessageData } from '../ChatEngine';
+import { isRenderTool, generateFromToolCall } from '../../a2ui/generator';
+import type { A2UIServerMessage } from '../../a2ui/types';
+import { ProviderRegistry, detectProvider } from './providers';
+import { createBlogTools, type BlogToolDeps } from './blog-tools';
+import { createA2UITools } from './a2ui-tools';
+
+// ---------------------------------------------------------------------------
+// Types
+// ---------------------------------------------------------------------------
+
+export interface ChatCallbacks {
+  onDelta?: (delta: string) => void;
+  onToolCall?: (toolCall: { name: string; args: unknown }) => void;
+  onToolResult?: (result: { name: string; result: unknown }) => void;
+  onA2UIMessage?: (message: A2UIServerMessage) => void;
+  onTokenUsage?: (usage: TokenUsageReport) => void;
+}
+
+export interface TokenUsageReport {
+  inputTokens: number;
+  outputTokens: number;
+  cacheReadTokens: number;
+  cacheWriteTokens: number;
+  totalTokens: number;
+  cumulativeInputTokens: number;
+  cumulativeOutputTokens: number;
+  cumulativeCacheReadTokens: number;
+  cumulativeCacheWriteTokens: number;
+  cumulativeTotalTokens: number;
+}
+
+export interface SendResult {
+  success: boolean;
+  message?: string;
+  error?: string;
+  toolCalls?: Array<{ name: string; args: unknown }>;
+}
+
+// Maximum tool-call rounds per request
+const MAX_TOOL_ROUNDS = 10;
+
+// ---------------------------------------------------------------------------
+// Message serialization — DB flat rows ↔ AI SDK messages
+// ---------------------------------------------------------------------------
+
+/**
+ * Convert DB message rows into AI SDK Message[] for `streamText({ messages })`.
+ * DB stores flat rows: role, content, toolCallId, toolCalls (JSON).
+ * AI SDK expects structured messages with content parts.
+ *
+ * Per Open Questions #3: only user/assistant messages are sent, tool call
+ * details from previous turns are appended as text annotations.
+ */
+function dbMessagesToAIMessages(
+  dbMessages: Pick<ChatMessageData, 'role' | 'content' | 'toolCalls'>[],
+): ModelMessage[] {
+  const messages: ModelMessage[] = [];
+
+  for (const msg of dbMessages) {
+    if (msg.role === 'user') {
+      messages.push({ role: 'user', content: msg.content || '' });
+    } else if (msg.role === 'assistant') {
+      let content = msg.content || '';
+      // Append tool-call annotation from previous turns (same as OpenCodeManager)
+      if (msg.toolCalls) {
+        try {
+          const calls = JSON.parse(msg.toolCalls) as Array<{ name: string; args: unknown }>;
+          if (calls.length > 0) {
+            const summary = calls
+              .map(tc => `- ${tc.name}(${JSON.stringify(tc.args)})`)
+              .join('\n');
+            content += `\n\n[Tools used in this turn:\n${summary}\n]`;
+          }
+        } catch {
+          // Ignore malformed tool call JSON
+        }
+      }
+      messages.push({ role: 'assistant', content });
+    }
+    // System and tool messages from DB are not sent — system is passed separately,
+    // tool results are only used within the same request via maxSteps.
+  }
+
+  return messages;
+}
+
+// ---------------------------------------------------------------------------
+// System prompt augmentation
+// ---------------------------------------------------------------------------
+
+/** Append live blog stats to the system prompt for data-volume awareness. */
+async function appendBlogStats(
+  basePrompt: string,
+  blogToolDeps: BlogToolDeps,
+): Promise<string> {
+  try {
+    const stats = await blogToolDeps.postEngine.getBlogStats();
+    const mediaList = await blogToolDeps.mediaEngine.getAllMedia();
+
+    if (stats.totalPosts === 0) return basePrompt;
+
+    const dateRange = stats.oldestPostDate && stats.newestPostDate
+      ? `from ${stats.oldestPostDate.toISOString().split('T')[0]} to ${stats.newestPostDate.toISOString().split('T')[0]}`
+      : 'unknown';
+
+    const yearBreakdown = Object.entries(stats.postsPerYear)
+      .sort(([a], [b]) => Number(a) - Number(b))
+      .map(([year, count]) => `${year}: ${count}`)
+      .join(', ');
+
+    return basePrompt + `
+
+--- CURRENT BLOG DATA SUMMARY ---
+Total posts: ${stats.totalPosts} (${stats.publishedCount} published, ${stats.draftCount} drafts, ${stats.archivedCount} archived)
+Date range: ${dateRange}
+Posts per year: ${yearBreakdown}
+Unique tags: ${stats.tagCount}, Unique categories: ${stats.categoryCount}
+Total media files: ${mediaList.length}
+NOTE: Use pagination (offset/limit) in list_posts and search_posts to access all data. Default page size is 20.`;
+  } catch {
+    return basePrompt;
+  }
+}
+
+// ---------------------------------------------------------------------------
+// Token estimation (for context truncation)
+// ---------------------------------------------------------------------------
+
+function estimateTokens(text: string): number {
+  return Math.ceil(text.length / 3.5);
+}
+
+/**
+ * Drop oldest user+assistant pairs to fit within context budget.
+ * Preserves the most recent messages for continuity.
+ */
+function truncateMessages(
+  messages: ModelMessage[],
+  systemPrompt: string,
+  toolsJson: string,
+  maxContextTokens: number,
+): ModelMessage[] {
+  const systemTokens = estimateTokens(systemPrompt);
+  const toolsTokens = estimateTokens(toolsJson);
+  const responseReserve = 4096;
+  const availableBudget = maxContextTokens - systemTokens - toolsTokens - responseReserve;
+
+  if (availableBudget <= 0) return messages.slice(-1);
+
+  const messageTokens = () =>
+    messages.reduce((sum, m) => sum + estimateTokens(typeof m.content === 'string' ? m.content : JSON.stringify(m.content)), 0);
+
+  if (messageTokens() <= availableBudget) return messages;
+
+  let truncated = [...messages];
+  while (truncated.length > 2 && messageTokens.call(null) > availableBudget) {
+    if (truncated[0].role === 'user') {
+      truncated = truncated.slice(2); // Drop user + assistant pair
+    } else {
+      truncated = truncated.slice(1);
+    }
+  }
+
+  return truncated;
+}
+
+// ---------------------------------------------------------------------------
+// ChatService
+// ---------------------------------------------------------------------------
+
+export class ChatService {
+  private chatEngine: ChatEngine;
+  private providers: ProviderRegistry;
+  private blogToolDeps: BlogToolDeps;
+  private getMainWindow: () => BrowserWindow | null;
+
+  // Abort controllers per conversation
+  private abortControllers = new Map<string, AbortController>();
+
+  // Cumulative token usage per conversation
+  private conversationUsage = new Map<string, {
+    inputTokens: number;
+    outputTokens: number;
+    cacheReadTokens: number;
+    cacheWriteTokens: number;
+  }>();
+
+  constructor(
+    chatEngine: ChatEngine,
+    providers: ProviderRegistry,
+    blogToolDeps: BlogToolDeps,
+    getMainWindow: () => BrowserWindow | null,
+  ) {
+    this.chatEngine = chatEngine;
+    this.providers = providers;
+    this.blogToolDeps = blogToolDeps;
+    this.getMainWindow = getMainWindow;
+  }
+
+  /**
+   * Send a user message, stream the AI response with tool use.
+   * This is the main entry point — replaces OpenCodeManager.sendMessage().
+   */
+  async sendMessage(
+    conversationId: string,
+    userMessage: string,
+    callbacks: ChatCallbacks = {},
+  ): Promise<SendResult> {
+    try {
+      // Readiness check
+      if (!this.providers.isReady()) {
+        return { success: false, error: 'API key not configured' };
+      }
+
+      // Load conversation
+      const conversation = await this.chatEngine.getConversation(conversationId);
+      if (!conversation) {
+        return { success: false, error: 'Conversation not found' };
+      }
+
+      // Add user message to DB
+      await this.chatEngine.addMessage({
+        conversationId,
+        role: 'user',
+        content: userMessage,
+        createdAt: new Date(),
+      });
+
+      // Abort controller
+      const abortController = new AbortController();
+      this.abortControllers.set(conversationId, abortController);
+
+      const modelId = conversation.model || 'claude-sonnet-4';
+      const provider = detectProvider(modelId);
+
+      // Verify provider key is available
+      if (!this.providers.isProviderKeySet(provider)) {
+        const providerLabel = provider === 'mistral' ? 'Mistral' : 'OpenCode';
+        return { success: false, error: `The model '${modelId}' requires a ${providerLabel} API key. Configure it in Settings.` };
+      }
+
+      // Build system prompt with live blog stats
+      const systemMessage = conversation.messages.find(m => m.role === 'system');
+      const basePrompt = systemMessage?.content || await this.chatEngine.getDefaultSystemPrompt();
+      const systemPrompt = await appendBlogStats(basePrompt, this.blogToolDeps);
+
+      // Convert DB messages to AI SDK format
+      const dbMessages = conversation.messages.filter(m => m.role !== 'system');
+      dbMessages.push({
+        conversationId,
+        role: 'user',
+        content: userMessage,
+        createdAt: new Date(),
+      });
+
+      const aiMessages = dbMessagesToAIMessages(dbMessages);
+
+      // Build tools
+      const blogTools = createBlogTools(this.blogToolDeps);
+      const a2uiToolsRaw = createA2UITools();
+      const allTools = { ...blogTools, ...a2uiToolsRaw };
+
+      // Get context window for truncation
+      const contextWindow = await this.providers.getModelCatalogEngine().getContextWindow(modelId) ?? 150_000;
+      const truncatedMessages = truncateMessages(
+        aiMessages,
+        systemPrompt,
+        JSON.stringify(Object.keys(allTools)),
+        contextWindow,
+      );
+
+      // Resolve model
+      const model = this.providers.resolveModel(modelId);
+
+      // Compute turn index for A2UI messages
+      const turnIndex = dbMessages.filter(m => m.role === 'user').length - 1;
+
+      // Track tool calls for response
+      const allToolCalls: Array<{ name: string; args: unknown }> = [];
+
+      // Build Anthropic-specific provider options for cache control
+      const providerOptions = modelId.startsWith('claude')
+        ? { anthropic: { cacheControl: { type: 'ephemeral' as const } } }
+        : undefined;
+
+      try {
+        // --- streamText: the AI SDK replaces our entire SSE/accumulator/tool-loop ---
+        const result = streamText({
+          model,
+          system: systemPrompt,
+          messages: truncatedMessages,
+          tools: allTools,
+          stopWhen: stepCountIs(MAX_TOOL_ROUNDS),
+          abortSignal: abortController.signal,
+          maxRetries: 3,
+          providerOptions,
+          onChunk: ({ chunk }) => {
+            if (chunk.type === 'text-delta' && callbacks.onDelta) {
+              callbacks.onDelta(chunk.text);
+            }
+          },
+          onStepFinish: ({ staticToolCalls: stepToolCalls, staticToolResults: stepToolResults }) => {
+            // Emit tool call/result events for each step
+            if (stepToolCalls) {
+              for (const tc of stepToolCalls) {
+                allToolCalls.push({ name: tc.toolName, args: tc.input });
+                callbacks.onToolCall?.({ name: tc.toolName, args: tc.input });
+              }
+            }
+            if (stepToolResults) {
+              for (const tr of stepToolResults) {
+                const toolName = tr.toolName;
+                const toolResult = tr.output;
+
+                // Handle A2UI render tools
+                if (isRenderTool(toolName)) {
+                  // Find the matching tool call args
+                  const matchingCall = stepToolCalls?.find(tc => tc.toolName === toolName);
+                  if (matchingCall) {
+                    const a2uiMessages = generateFromToolCall(
+                      conversationId,
+                      toolName,
+                      matchingCall.input as Record<string, unknown>,
+                    );
+                    if (a2uiMessages && callbacks.onA2UIMessage) {
+                      for (const msg of a2uiMessages) {
+                        if (msg.type === 'createSurface') {
+                          msg.metadata = { ...msg.metadata, turnIndex };
+                        }
+                        callbacks.onA2UIMessage(msg);
+                      }
+                    }
+                  }
+                }
+
+                callbacks.onToolResult?.({ name: toolName, result: toolResult });
+              }
+            }
+          },
+        });
+
+        // Consume the stream to completion
+        const finalResult = await result.response;
+
+        // Extract usage from the response
+        const usage = await result.usage;
+        this.emitUsage(conversationId, usage, callbacks);
+
+        // Get the final text
+        const fullResponse = await result.text;
+
+        // Save assistant response to DB
+        if (fullResponse) {
+          await this.chatEngine.addMessage({
+            conversationId,
+            role: 'assistant',
+            content: fullResponse,
+            toolCalls: allToolCalls.length > 0 ? JSON.stringify(allToolCalls) : undefined,
+            createdAt: new Date(),
+          });
+        }
+
+        // Generate title after first user message
+        const userMsgCount = conversation.messages.filter(m => m.role === 'user').length;
+        if (userMsgCount === 0 && fullResponse) {
+          this.generateConversationTitle(conversationId, userMessage).catch(err =>
+            console.error('[ChatService] Error generating title:', err),
+          );
+        }
+
+        return {
+          success: true,
+          message: fullResponse,
+          toolCalls: allToolCalls.length > 0 ? allToolCalls : undefined,
+        };
+      } catch (error) {
+        const isAborted = abortController.signal.aborted || (error as Error).message === 'Request cancelled';
+        if (!isAborted) throw error;
+        return { success: true, message: '' };
+      } finally {
+        this.abortControllers.delete(conversationId);
+      }
+    } catch (error) {
+      console.error('[ChatService] Error sending message:', error);
+      return { success: false, error: (error as Error).message };
+    }
+  }
+
+  /** Abort an in-flight request for a conversation. */
+  async abortMessage(conversationId: string): Promise<{ success: boolean; error?: string }> {
+    const controller = this.abortControllers.get(conversationId);
+    if (!controller) {
+      return { success: false, error: 'No active request for this conversation' };
+    }
+    controller.abort();
+    this.abortControllers.delete(conversationId);
+    return { success: true };
+  }
+
+  /** Abort all in-flight requests. */
+  async stop(): Promise<void> {
+    for (const [, controller] of this.abortControllers) {
+      controller.abort();
+    }
+    this.abortControllers.clear();
+  }
+
+  // ---- Private helpers ----
+
+  /**
+   * Generate a short conversation title from the first user message.
+   * Non-streaming one-shot call using the configured title model.
+   */
+  private async generateConversationTitle(
+    conversationId: string,
+    userMessage: string,
+  ): Promise<void> {
+    try {
+      let titleModel = await this.chatEngine.getSetting('chat_title_model');
+
+      // Fallback chain: setting → haiku → mistral-small
+      if (!titleModel || !this.providers.isProviderKeySet(detectProvider(titleModel))) {
+        titleModel = this.providers.getOpencodeKey()
+          ? 'claude-haiku-4-5'
+          : this.providers.getMistralKey()
+            ? 'mistral-small-latest'
+            : null;
+      }
+      if (!titleModel) return;
+
+      const model = this.providers.resolveModel(titleModel);
+
+      const { text } = await generateText({
+        model,
+        system: 'Generate an ultra-short title (2-3 words, max 25 characters) for this conversation. Focus ONLY on the topic. Ignore any capability disclaimers. Output ONLY the title text.',
+        prompt: `Topic: ${userMessage.substring(0, 100)}`,
+        maxOutputTokens: 20,
+        maxRetries: 2,
+      });
+
+      let title = text.trim().replace(/^["']|["']$/g, '').replace(/[.!?]+$/, '');
+      const MAX_TITLE_LENGTH = 30;
+      if (title.length > MAX_TITLE_LENGTH) {
+        title = title.substring(0, MAX_TITLE_LENGTH - 1) + '…';
+      }
+
+      if (title) {
+        await this.chatEngine.updateConversation(conversationId, { title });
+        const mainWindow = this.getMainWindow();
+        if (mainWindow) {
+          mainWindow.webContents.send('chat-title-updated', { conversationId, title });
+        }
+      }
+    } catch (error) {
+      console.error('[ChatService] Error generating title:', error);
+    }
+  }
+
+  /** Emit per-turn + cumulative token usage. */
+  private emitUsage(
+    conversationId: string,
+    usage: LanguageModelUsage | undefined,
+    callbacks: ChatCallbacks,
+  ): void {
+    if (!usage || !callbacks.onTokenUsage) return;
+
+    // AI SDK v6 normalizes usage into inputTokens/outputTokens
+    // Cache tokens are in inputTokenDetails
+    const inputTokens = usage.inputTokens ?? 0;
+    const outputTokens = usage.outputTokens ?? 0;
+    const cacheReadTokens = usage.inputTokenDetails?.cacheReadTokens ?? 0;
+    const cacheWriteTokens = usage.inputTokenDetails?.cacheWriteTokens ?? 0;
+    const adjustedInputTokens = inputTokens - cacheReadTokens - cacheWriteTokens;
+    const totalTokens = inputTokens + outputTokens;
+
+    const prev = this.conversationUsage.get(conversationId) || {
+      inputTokens: 0, outputTokens: 0, cacheReadTokens: 0, cacheWriteTokens: 0,
+    };
+    const cumulative = {
+      inputTokens: prev.inputTokens + adjustedInputTokens,
+      outputTokens: prev.outputTokens + outputTokens,
+      cacheReadTokens: prev.cacheReadTokens + cacheReadTokens,
+      cacheWriteTokens: prev.cacheWriteTokens + cacheWriteTokens,
+    };
+    this.conversationUsage.set(conversationId, cumulative);
+
+    callbacks.onTokenUsage({
+      inputTokens: adjustedInputTokens, outputTokens, cacheReadTokens, cacheWriteTokens, totalTokens,
+      cumulativeInputTokens: cumulative.inputTokens,
+      cumulativeOutputTokens: cumulative.outputTokens,
+      cumulativeCacheReadTokens: cumulative.cacheReadTokens,
+      cumulativeCacheWriteTokens: cumulative.cacheWriteTokens,
+      cumulativeTotalTokens: cumulative.inputTokens + cumulative.outputTokens + cumulative.cacheReadTokens + cumulative.cacheWriteTokens,
+    });
+  }
+}

src/main/engine/ai/providers.ts

@@ -0,0 +1,347 @@
+/**
+ * Provider registry — single source of truth for AI provider routing.
+ *
+ * Two provider sources:
+ *   1. OpenCode Zen gateway — routes claude* → Anthropic Messages API,
+ *      everything else → OpenAI Chat Completions API
+ *   2. Mistral direct — uses Mistral's native API
+ *
+ * Model listing uses raw HTTP (AI SDK has no listing API).
+ *
+ * IMPORTANT: OpenAI SDK v6 defaults to Responses API (/responses).
+ * OpenCode Zen only supports Chat Completions. Use provider.chat(modelId).
+ */
+
+import { customProvider } from 'ai';
+import { createAnthropic } from '@ai-sdk/anthropic';
+import { createOpenAI } from '@ai-sdk/openai';
+import { createMistral } from '@ai-sdk/mistral';
+import type { LanguageModel, Provider } from 'ai';
+import { ModelCatalogEngine } from '../ModelCatalogEngine';
+import type { ChatModel } from '../../shared/electronApi';
+
+// ---------------------------------------------------------------------------
+// Constants
+// ---------------------------------------------------------------------------
+
+export const ZEN_BASE_URL = 'https://opencode.ai/zen/v1';
+export const ZEN_MODELS_URL = 'https://opencode.ai/zen/v1/models';
+export const MISTRAL_MODELS_URL = 'https://api.mistral.ai/v1/models';
+
+const MODEL_CACHE_TTL = 5 * 60 * 1000; // 5 minutes
+
+// ---------------------------------------------------------------------------
+// Gateway factory
+// ---------------------------------------------------------------------------
+
+/**
+ * Creates the OpenCode Zen gateway custom provider.
+ * Routes claude* → Anthropic Messages API, everything else → OpenAI Chat Completions.
+ */
+export function createOpenCodeGateway(apiKey: string): Provider {
+  const anthropicProvider = createAnthropic({
+    baseURL: ZEN_BASE_URL,
+    apiKey,
+  });
+  const openaiProvider = createOpenAI({
+    baseURL: ZEN_BASE_URL,
+    apiKey,
+  });
+
+  // Build a ProviderV3 that routes claude* → Anthropic, else → OpenAI Chat Completions
+  const gatewayRouter: import('@ai-sdk/provider').ProviderV3 = {
+    specificationVersion: 'v3',
+    languageModel: (modelId: string) => {
+      if (modelId.startsWith('claude')) {
+        return anthropicProvider(modelId);
+      }
+      // Use .chat() for Chat Completions — Zen doesn't support Responses API
+      return openaiProvider.chat(modelId);
+    },
+    embeddingModel: () => {
+      throw new Error('Embeddings not supported via OpenCode gateway');
+    },
+    imageModel: () => {
+      throw new Error('Image models not supported via OpenCode gateway');
+    },
+  };
+
+  return customProvider({
+    languageModels: {},
+    fallbackProvider: gatewayRouter,
+  });
+}
+
+// ---------------------------------------------------------------------------
+// Provider detection — shared utility
+// ---------------------------------------------------------------------------
+
+/** Determine which provider backend a model ID belongs to. */
+export function detectProvider(modelId: string): string {
+  const id = modelId.toLowerCase();
+  if (id.startsWith('claude')) return 'anthropic';
+  if (id.startsWith('gpt') || id.startsWith('o3') || id.startsWith('o4')) return 'openai';
+  if (id.startsWith('gemini')) return 'google';
+  if (
+    id.startsWith('mistral') ||
+    id.startsWith('ministral') ||
+    id.startsWith('devstral') ||
+    id.startsWith('codestral') ||
+    id.startsWith('pixtral')
+  ) return 'mistral';
+  return 'other';
+}
+
+// ---------------------------------------------------------------------------
+// ProviderRegistry — manages keys, providers, model resolution
+// ---------------------------------------------------------------------------
+
+export class ProviderRegistry {
+  private opencodeKey = '';
+  private mistralKey = '';
+  private opencodeGateway: Provider | null = null;
+  private mistralProvider: ReturnType<typeof createMistral> | null = null;
+  private modelCatalogEngine = new ModelCatalogEngine();
+
+  // Model cache
+  private cachedModels: ChatModel[] | null = null;
+  private cachedModelsAt = 0;
+
+  // ---- Key management ----
+
+  setOpencodeKey(key: string): void {
+    this.opencodeKey = key;
+    this.opencodeGateway = null; // rebuild on next use
+    this.invalidateModelCache();
+  }
+
+  getOpencodeKey(): string {
+    return this.opencodeKey;
+  }
+
+  setMistralKey(key: string): void {
+    this.mistralKey = key;
+    this.mistralProvider = null; // rebuild on next use
+    this.invalidateModelCache();
+  }
+
+  getMistralKey(): string {
+    return this.mistralKey;
+  }
+
+  /** Check whether at least one provider key is configured. */
+  isReady(): boolean {
+    return !!(this.opencodeKey || this.mistralKey);
+  }
+
+  /** Check whether the key for a specific provider is set. */
+  isProviderKeySet(provider: string): boolean {
+    if (provider === 'mistral') return !!this.mistralKey;
+    return !!this.opencodeKey;
+  }
+
+  /** Returns status of all configured providers. */
+  getProviderStatus(): { opencode: boolean; mistral: boolean } {
+    return {
+      opencode: !!this.opencodeKey,
+      mistral: !!this.mistralKey,
+    };
+  }
+
+  // ---- Provider resolution ----
+
+  /** Resolve a model ID to an AI SDK LanguageModel. */
+  resolveModel(modelId: string): LanguageModel {
+    const provider = detectProvider(modelId);
+
+    if (provider === 'mistral') {
+      if (!this.mistralKey) {
+        throw new Error(`Mistral API key not configured for model '${modelId}'`);
+      }
+      if (!this.mistralProvider) {
+        this.mistralProvider = createMistral({ apiKey: this.mistralKey });
+      }
+      return this.mistralProvider(modelId);
+    }
+
+    // Everything else goes through the OpenCode gateway
+    if (!this.opencodeKey) {
+      throw new Error(`OpenCode API key not configured for model '${modelId}'`);
+    }
+    if (!this.opencodeGateway) {
+      this.opencodeGateway = createOpenCodeGateway(this.opencodeKey);
+    }
+    return this.opencodeGateway.languageModel(modelId);
+  }
+
+  // ---- Model listing (raw HTTP — AI SDK has no listing API) ----
+
+  invalidateModelCache(): void {
+    this.cachedModels = null;
+    this.cachedModelsAt = 0;
+  }
+
+  /** Get the model catalog engine for context window lookups. */
+  getModelCatalogEngine(): ModelCatalogEngine {
+    return this.modelCatalogEngine;
+  }
+
+  /** Get available models across all configured providers (cached 5 min). */
+  async getAvailableModels(): Promise<ChatModel[]> {
+    if (this.cachedModels && Date.now() - this.cachedModelsAt < MODEL_CACHE_TTL) {
+      return this.cachedModels;
+    }
+
+    const allModels: ChatModel[] = [];
+    let fetched = false;
+    const { vision: catalogVision, names: catalogNames } = await this.getCatalogLookups();
+
+    // Fetch OpenCode models
+    if (this.opencodeKey) {
+      try {
+        const models = await this.fetchModelsFromEndpoint(
+          ZEN_MODELS_URL,
+          { Authorization: `Bearer ${this.opencodeKey}`, 'x-api-key': this.opencodeKey },
+          catalogVision,
+          catalogNames,
+        );
+        allModels.push(...models);
+        fetched = true;
+      } catch {
+        // Fall through
+      }
+    }
+
+    // Fetch Mistral models
+    if (this.mistralKey) {
+      try {
+        const models = await this.fetchModelsFromEndpoint(
+          MISTRAL_MODELS_URL,
+          { Authorization: `Bearer ${this.mistralKey}` },
+          catalogVision,
+          catalogNames,
+          'mistral', // only keep mistral-family models
+        );
+        allModels.push(...models);
+        fetched = true;
+      } catch {
+        // Fall through
+      }
+    }
+
+    if (fetched && allModels.length > 0) {
+      this.cachedModels = allModels;
+      this.cachedModelsAt = Date.now();
+      return allModels;
+    }
+
+    // Fallback: model catalog DB, filtered by available provider keys
+    return this.getModelsFromCatalog();
+  }
+
+  /** Validate an OpenCode API key against the models endpoint. */
+  async validateOpencodeKey(apiKey: string): Promise<{ isValid: boolean; models: ChatModel[] }> {
+    if (!apiKey || apiKey.length < 3) return { isValid: false, models: [] };
+
+    const { vision: catalogVision, names: catalogNames } = await this.getCatalogLookups();
+
+    const headerSets: Record<string, string>[] = [
+      { Authorization: `Bearer ${apiKey}` },
+      { 'x-api-key': apiKey },
+    ];
+
+    for (const headers of headerSets) {
+      try {
+        const models = await this.fetchModelsFromEndpoint(
+          ZEN_MODELS_URL, headers, catalogVision, catalogNames,
+        );
+        return { isValid: true, models };
+      } catch {
+        // Try next
+      }
+    }
+    return { isValid: false, models: [] };
+  }
+
+  /** Validate a Mistral API key against the Mistral models endpoint. */
+  async validateMistralKey(apiKey: string): Promise<{ isValid: boolean; models: ChatModel[] }> {
+    if (!apiKey || apiKey.length < 3) return { isValid: false, models: [] };
+
+    const { vision: catalogVision, names: catalogNames } = await this.getCatalogLookups();
+
+    try {
+      const models = await this.fetchModelsFromEndpoint(
+        MISTRAL_MODELS_URL,
+        { Authorization: `Bearer ${apiKey}` },
+        catalogVision,
+        catalogNames,
+        'mistral',
+      );
+      return { isValid: true, models };
+    } catch {
+      return { isValid: false, models: [] };
+    }
+  }
+
+  // ---- Private helpers ----
+
+  private async fetchModelsFromEndpoint(
+    url: string,
+    headers: Record<string, string>,
+    catalogVision: Map<string, boolean>,
+    catalogNames: Map<string, string>,
+    filterProvider?: string,
+  ): Promise<ChatModel[]> {
+    const response = await fetch(url, { method: 'GET', headers });
+    if (!response.ok) throw new Error(`HTTP ${response.status}`);
+
+    const data = await response.json() as { data?: Array<{ id: string }> };
+    if (!data.data || !Array.isArray(data.data)) return [];
+
+    let models = data.data;
+    if (filterProvider) {
+      models = models.filter(m => detectProvider(m.id) === filterProvider);
+    }
+
+    return models.map(m => ({
+      id: m.id,
+      name: catalogNames.get(m.id) ?? m.id,
+      provider: detectProvider(m.id),
+      vision: catalogVision.get(m.id) ?? false,
+    }));
+  }
+
+  private async getCatalogLookups(): Promise<{ vision: Map<string, boolean>; names: Map<string, string> }> {
+    const vision = new Map<string, boolean>();
+    const names = new Map<string, string>();
+    try {
+      const catalog = await this.modelCatalogEngine.getAll();
+      for (const m of catalog) {
+        vision.set(m.id, m.inputModalities.includes('image'));
+        names.set(m.id, m.name);
+      }
+    } catch {
+      // Catalog unavailable
+    }
+    return { vision, names };
+  }
+
+  private async getModelsFromCatalog(): Promise<ChatModel[]> {
+    try {
+      const catalog = await this.modelCatalogEngine.getAll();
+      if (catalog.length > 0) {
+        return catalog
+          .map(m => ({
+            id: m.id,
+            name: m.name,
+            provider: detectProvider(m.id),
+            vision: m.inputModalities.includes('image'),
+          }))
+          .filter(m => this.isProviderKeySet(m.provider));
+      }
+    } catch {
+      // Fall through
+    }
+    return [];
+  }
+}

src/main/engine/ai/tasks.ts

@@ -0,0 +1,258 @@
+/**
+ * OneShotTasks — non-streaming AI tasks using generateText().
+ *
+ * Replaces OpenCodeManager.analyzeTaxonomy() and analyzeMediaImage()
+ * with provider-agnostic AI SDK calls.
+ */
+
+import { generateText } from 'ai';
+import type { ChatEngine } from '../ChatEngine';
+import type { MediaEngine } from '../MediaEngine';
+import { ProviderRegistry, detectProvider } from './providers';
+
+// ---------------------------------------------------------------------------
+// Types
+// ---------------------------------------------------------------------------
+
+export interface TaxonomyAnalysisResult {
+  success: boolean;
+  categoryMappings?: Record<string, string>;
+  tagMappings?: Record<string, string>;
+  error?: string;
+}
+
+export interface ImageAnalysisResult {
+  success: boolean;
+  title?: string;
+  alt?: string;
+  caption?: string;
+  error?: string;
+}
+
+// ---------------------------------------------------------------------------
+// Language map for image analysis prompts
+// ---------------------------------------------------------------------------
+
+const LANGUAGE_NAMES: Record<string, string> = {
+  en: 'English', de: 'German', es: 'Spanish', fr: 'French', it: 'Italian',
+  pt: 'Portuguese', nl: 'Dutch', pl: 'Polish', ru: 'Russian', ja: 'Japanese',
+  zh: 'Chinese', ko: 'Korean', ar: 'Arabic', hi: 'Hindi', tr: 'Turkish',
+  sv: 'Swedish', da: 'Danish', no: 'Norwegian', fi: 'Finnish', cs: 'Czech',
+};
+
+// ---------------------------------------------------------------------------
+// OneShotTasks
+// ---------------------------------------------------------------------------
+
+export class OneShotTasks {
+  private providers: ProviderRegistry;
+  private chatEngine: ChatEngine;
+  private mediaEngine: MediaEngine;
+
+  constructor(
+    providers: ProviderRegistry,
+    chatEngine: ChatEngine,
+    mediaEngine: MediaEngine,
+  ) {
+    this.providers = providers;
+    this.chatEngine = chatEngine;
+    this.mediaEngine = mediaEngine;
+  }
+
+  /**
+   * Analyze taxonomy items from a WordPress import and suggest mappings
+   * from NEW items to EXISTING items to avoid duplicates.
+   */
+  async analyzeTaxonomy(
+    categories: Array<{ name: string; slug: string; existsInProject: boolean }>,
+    tags: Array<{ name: string; slug: string; existsInProject: boolean }>,
+    modelId: string,
+  ): Promise<TaxonomyAnalysisResult> {
+    const provider = detectProvider(modelId);
+    if (!this.providers.isProviderKeySet(provider)) {
+      const providerLabel = provider === 'mistral' ? 'Mistral' : 'OpenCode';
+      return { success: false, error: `${providerLabel} API key not set` };
+    }
+
+    const existingCategories = categories.filter(c => c.existsInProject).map(c => c.name);
+    const newCategories = categories.filter(c => !c.existsInProject).map(c => c.name);
+    const existingTags = tags.filter(t => t.existsInProject).map(t => t.name);
+    const newTags = tags.filter(t => !t.existsInProject).map(t => t.name);
+
+    const systemPrompt = `You are an expert at analyzing taxonomy terms (tags and categories) for a blog import system.
+
+Your task is to identify NEW tags/categories from an import that should be mapped to EXISTING tags/categories in the project to avoid creating duplicates.
+
+CRITICAL RULES:
+1. ONLY map NEW items to EXISTING items - never map new to new
+2. The goal is to prevent duplicate creation, NOT to reduce the number of new items
+3. A new item should only map to an existing item if they represent the same concept
+4. Consider language differences: a new tag can match an existing tag in a different language (e.g., "Photography" should map to "Fotografie" if that exists)
+5. Consider variations like: different casing, singular/plural, abbreviations, hyphenation, synonyms
+6. Only suggest mappings where there is a clear semantic match - not every new item needs a mapping
+
+EXAMPLES OF VALID MAPPINGS (new → existing):
+- "Photos" → "Photography" (if Photography exists)
+- "Fotografie" → "Photography" (language variation, if Photography exists)  
+- "tech" → "Technology" (abbreviation, if Technology exists)
+- "Web Dev" → "Web Development" (abbreviation, if Web Development exists)
+
+DO NOT:
+- Map a new item to another new item
+- Suggest mappings just because items are in the same topic area
+- Create mappings for items that are distinct concepts
+
+RESPONSE FORMAT:
+You MUST respond with valid JSON only, no other text. Use this exact structure:
+{
+  "categoryMappings": { "New Category": "Existing Category", ... },
+  "tagMappings": { "New Tag": "Existing Tag", ... }
+}
+
+The source (key) MUST be from the NEW items list, and the target (value) MUST be from the EXISTING items list.
+If there are no sensible mappings to suggest, return empty objects.`;
+
+    const userPrompt = `Analyze these taxonomy items from a WordPress import. Identify NEW items that should be mapped to EXISTING items to avoid duplicates.
+
+EXISTING CATEGORIES IN PROJECT (map TO these):
+${existingCategories.length > 0 ? existingCategories.join(', ') : '(none)'}
+
+NEW CATEGORIES FROM IMPORT (map FROM these):
+${newCategories.length > 0 ? newCategories.join(', ') : '(none)'}
+
+EXISTING TAGS IN PROJECT (map TO these):
+${existingTags.length > 0 ? existingTags.join(', ') : '(none)'}
+
+NEW TAGS FROM IMPORT (map FROM these):
+${newTags.length > 0 ? newTags.join(', ') : '(none)'}
+
+Remember: Only suggest mappings from NEW items to EXISTING items. Consider language differences (e.g., German/English equivalents). Response must be valid JSON only.`;
+
+    try {
+      const model = this.providers.resolveModel(modelId);
+
+      const { text } = await generateText({
+        model,
+        system: systemPrompt,
+        prompt: userPrompt,
+        maxOutputTokens: 4096,
+        maxRetries: 2,
+      });
+
+      // Extract and parse JSON from response
+      const jsonMatch = text.match(/\{[\s\S]*\}/);
+      if (!jsonMatch) {
+        return { success: false, error: 'Invalid response format from AI' };
+      }
+
+      const result = JSON.parse(jsonMatch[0]);
+
+      // Validate mappings: only new→existing allowed
+      const validatedCategoryMappings: Record<string, string> = {};
+      const validatedTagMappings: Record<string, string> = {};
+
+      const newCatSet = new Set(newCategories);
+      const existingCatSet = new Set(existingCategories);
+      for (const [source, target] of Object.entries(result.categoryMappings || {})) {
+        if (newCatSet.has(source) && existingCatSet.has(target as string)) {
+          validatedCategoryMappings[source] = target as string;
+        }
+      }
+
+      const newTagSet = new Set(newTags);
+      const existingTagSet = new Set(existingTags);
+      for (const [source, target] of Object.entries(result.tagMappings || {})) {
+        if (newTagSet.has(source) && existingTagSet.has(target as string)) {
+          validatedTagMappings[source] = target as string;
+        }
+      }
+
+      return {
+        success: true,
+        categoryMappings: validatedCategoryMappings,
+        tagMappings: validatedTagMappings,
+      };
+    } catch (error) {
+      return { success: false, error: (error as Error).message };
+    }
+  }
+
+  /**
+   * Analyze an image and generate title, alt text, and caption.
+   * Uses multimodal input — AI SDK handles the provider-specific format.
+   */
+  async analyzeMediaImage(
+    mediaId: string,
+    language: string = 'en',
+  ): Promise<ImageAnalysisResult> {
+    // Determine model with smart fallback
+    let modelId = await this.chatEngine.getSetting('chat_image_analysis_model');
+    if (!modelId || !this.providers.isProviderKeySet(detectProvider(modelId))) {
+      modelId = this.providers.getOpencodeKey()
+        ? 'claude-sonnet-4-5'
+        : this.providers.getMistralKey()
+          ? 'mistral-large-latest'
+          : null;
+    }
+    if (!modelId) {
+      return { success: false, error: 'API key not configured. Please set an API key in Settings.' };
+    }
+
+    // Get media metadata
+    const mediaItem = await this.mediaEngine.getMedia(mediaId);
+    if (!mediaItem) return { success: false, error: 'Media item not found' };
+    if (!mediaItem.mimeType.startsWith('image/')) {
+      return { success: false, error: `Cannot analyze this file type: ${mediaItem.mimeType}. Only images are supported.` };
+    }
+
+    // Get thumbnail
+    let dataUrl = await this.mediaEngine.getThumbnailDataUrl(mediaId, 'large');
+    if (!dataUrl) dataUrl = await this.mediaEngine.getThumbnailDataUrl(mediaId, 'medium');
+    if (!dataUrl) {
+      return { success: false, error: 'Image thumbnail not available. Try regenerating thumbnails from Settings.' };
+    }
+
+    const base64Data = dataUrl.replace(/^data:image\/\w+;base64,/, '');
+    const languageName = LANGUAGE_NAMES[language] || language;
+
+    const systemPrompt = `Generate title, alt text, and caption for this image in ${languageName}.
+
+TITLE: A short, descriptive title for display in lists and search results (3-8 words). Should identify the main subject.
+ALT: Describe ONLY what is visually present in the image. Be factual, neutral, and concise (5-12 words max). No interpretations, emotions, or "Image of" prefix. Example: "Red bicycle leaning against white brick wall"
+CAPTION: Short, engaging blog caption (5-20 words).
+
+Respond with JSON only: {"title": "...", "alt": "...", "caption": "..."}`;
+
+    try {
+      const model = this.providers.resolveModel(modelId);
+
+      // AI SDK handles provider-specific multimodal format automatically
+      const { text } = await generateText({
+        model,
+        system: systemPrompt,
+        messages: [{
+          role: 'user',
+          content: [
+            { type: 'image', image: `data:image/webp;base64,${base64Data}` },
+            { type: 'text', text: 'Analyze and respond with JSON.' },
+          ],
+        }],
+        maxOutputTokens: 200,
+        maxRetries: 2,
+      });
+
+      const jsonMatch = text.match(/\{[\s\S]*\}/);
+      if (!jsonMatch) return { success: false, error: 'Invalid response format from AI' };
+
+      const result = JSON.parse(jsonMatch[0]);
+      return {
+        success: true,
+        title: result.title || undefined,
+        alt: result.alt || undefined,
+        caption: result.caption || undefined,
+      };
+    } catch (error) {
+      return { success: false, error: (error as Error).message };
+    }
+  }
+}