This section documents version-specific API changes for @tanstack/ai-vue v0.6.1 (current v0.x series). This library is pre-1.0 — all v0.x releases are in scope.
-
BREAKING: Monolithic adapter factories removed — openai(), anthropic(), etc. replaced by activity-specific functions: openaiText('gpt-5.2'), openaiSummarize('gpt-5-mini'), openaiImage('dall-e-3'), etc. Model name is now passed to the adapter factory, not to chat(). source
-
BREAKING: model parameter removed from chat() — model is now embedded in the adapter argument (e.g., adapter: openaiText('gpt-5.2') instead of adapter: openai(), model: 'gpt-4'). Passing model at the call site is silently ignored. source
-
BREAKING: Nested options object flattened — chat({ options: { temperature, maxTokens, topP } }) must be changed to chat({ temperature, maxTokens, topP }). Nested options are silently discarded. source
-
BREAKING: providerOptions renamed to modelOptions — chat({ providerOptions: { ... } }) must be updated to chat({ modelOptions: { ... } }). Silently ignored if not updated. source
-
BREAKING: toResponseStream renamed to toServerSentEventsStream and now returns ReadableStream instead of Response — must manually create new Response(stream, { headers }). AbortController is now a separate parameter: toServerSentEventsStream(stream, abortController). source
-
BREAKING: embedding() function removed — embeddings support eliminated entirely. Use provider SDKs directly or vector DB native embedding APIs. source
-
BREAKING: chat({ as: 'promise' }) replaced by separate chatCompletion() function — as option removed from chat(). chat({ as: 'stream' }) is now just chat(). chat({ as: 'response' }) is now chat() + toServerSentEventsStream(). source
-
NEW: useChat returns status reactive ref — tracks lifecycle as 'ready' | 'submitted' | 'streaming' | 'error'. Previously there was no generation lifecycle state. source
-
NEW: sendMessage() accepts MultimodalContent object — sendMessage({ content: [{ type: 'text', content: '...' }, { type: 'image', source: { type: 'url', value: '...' } }] }) enables image/audio/video/document content alongside text. Added in v0.5.0. source
-
NEW: agentLoopStrategy parameter replaces bare maxIterations: number — use agentLoopStrategy: maxIterations(5), untilFinishReason(['stop']), or combineStrategies([...]). Old maxIterations number is converted automatically but deprecated. source
-
NEW: toolDefinition({ name, description, inputSchema, outputSchema?, needsApproval? }) — creates isomorphic tool definitions. Call .server(fn) for server-side execution or .client(fn) for client-side execution. Replaces ad-hoc tool objects. source
-
NEW: @tanstack/ai-client package — ChatClient class provides framework-agnostic headless chat state management with sendMessage(), reload(), stop(), clear(), addToolResult(), addToolApprovalResponse() methods. source
-
NEW: Connection adapter factories — fetchServerSentEvents(url, options?), fetchHttpStream(url, options?), stream(fn) from @tanstack/ai-client. Pass to useChat({ connection: fetchServerSentEvents('/api/chat') }) instead of url: '/api/chat'. source
-
NEW: extendAdapter(factory, customModels) + createModel(name, modalities) — adds custom/fine-tuned model names to existing adapter factories with full type inference. Avoids as const casts. source
-
useChat returns DeepReadonly<ShallowRef<T>> refs — never reassign messages directly; use setMessages() for manual updates. Changing connection or body options recreates the underlying ChatClient, requiring a component remount or a key prop change to take effect
-
Use status (added v0.4.0) instead of isLoading for granular lifecycle control — status.value tracks 'ready' | 'submitted' | 'streaming' | 'error', enabling distinct UI states for submission vs. active streaming source
-
Pass client tool arrays through clientTools() instead of as const — eliminates the need for const assertion while enabling full discriminated union narrowing on part.name, part.input, and part.output in message iteration source
-
Wrap useChat options with createChatClientOptions() and derive message types using InferChatMessages<typeof chatOptions> — this propagates tool types through the entire message type, making part.name a literal union and part.input/part.output typed from Zod schemas source
-
Define tools with toolDefinition() in a shared file, then call .server() in route handlers and .client() in Vue components — passing the bare definition to chat() signals the client will execute it, while passing .server() output executes it server-side automatically source
-
Use Zod schemas (v4.2+) over raw JSON Schema for inputSchema/outputSchema in toolDefinition() and chat({ outputSchema }) — JSON Schema infers any for tool inputs/outputs and unknown for structured output return types, losing all downstream type safety source
-
Set agentLoopStrategy: maxIterations(n) explicitly when tools are present — the default is 5 iterations, which is too low for multi-step agentic workflows; use untilFinishReason('stop') to exit as soon as the model finishes without hitting the limit source
-
Subscribe to aiEventClient with { withEventTarget: true } in production code — without this third argument the client only emits to the devtools event bus (absent in production builds); the flag also dispatches to the current EventTarget for application-level observability source
-
Prefer fetchServerSentEvents over fetchHttpStream for client connections — SSE provides automatic reconnection; pass URL and options as functions (not static values) when headers like Authorization must be re-evaluated on every request source
-
Use extendAdapter(baseFactory, [createModel('model-name', ['text', 'image'])]) to add TypeScript types for fine-tuned models or OpenAI-compatible proxies — this adds the model to the adapter's allowed type union with zero runtime overhead while preserving all original factory config parameters source
