Isomorphic tool system: toolDefinition() with Zod schemas, .server() and .client() implementations, passing tools to both chat() on server and useChat/clientTools on client, tool approval flows with needsApproval and bound interrupts (resolveInterrupt), generic middleware interrupts with defineInterrupt(), lazy tool discovery with lazy:true, rendering ToolCallPart and ToolResultPart in UI.

Type-safe JSON schema responses from LLMs using outputSchema on chat() and useChat(). Supports Zod, ArkType, and Valibot schemas. The adapter handles provider-specific strategies transparently — never configure structured output at the provider level. Pass stream:true alongside outputSchema for incremental JSON deltas + a completed typed object via the `structured-output.complete` event. Each successfully completed structured-output run adds a typed `StructuredOutputPart` to message history. partial/final derive from the most recent structured-output part after the latest user message. convertSchemaToJsonSchema() for manual schema conversion.

Chat lifecycle middleware hooks: onConfig, onStart, onChunk, onBeforeToolCall, onAfterToolCall, onUsage, onFinish, onAbort, onError. Use for analytics, event firing, tool caching (toolCacheMiddleware), logging, and tracing. Middleware array in chat() config, left-to-right execution order. NOT onEnd/onFinish callbacks on chat() — use middleware.

Image, audio, video, speech (TTS), and transcription generation using activity-specific adapters: generateImage() with openaiImage/geminiImage/byteplusImage, generateAudio() with geminiAudio/falAudio, generateVideo() with async polling (openaiVideo/geminiVideo/grokVideo/falVideo/byteplusVideo/openRouterVideo, per-model typed durations), generateSpeech() with openaiSpeech/byteplusSpeech, generateTranscription() with openaiTranscription/byteplusTranscription. React hooks: useGenerateImage, useGenerateAudio, useGenerateSpeech, useTranscription, useGenerateVideo. TanStack Start server function integration with toServerSentEventsResponse.

LockStore, InMemoryLockStore, LocksCapability and withLocks for multi-instance coordination in TanStack AI. Ships in @tanstack/ai — NOT in @tanstack/ai-persistence. Separate from AIPersistence state stores — not a stores key, not composable. InMemoryLockStore vs a distributed (e.g. Cloudflare Durable Object) lock, lease recovery, AbortSignal in critical sections. Use when sandbox or other middleware needs cross-worker mutual exclusion — NOT for storing messages/runs (use withPersistence).

Pluggable, category-toggleable debug logging for TanStack AI activities. Toggle with `debug: true | false | DebugConfig` on chat(), summarize(), generateImage(), generateSpeech(), generateTranscription(), generateVideo(). Categories: request, provider, output, middleware, tools, agentLoop, config, errors. Pipe into pino/winston/etc via `debug: { logger }`. Errors log by default even when `debug` is omitted; silence with `debug: false`.

Connect useChat to a non-TanStack-AI backend through custom connection adapters. ConnectConnectionAdapter (single async iterable) vs SubscribeConnectionAdapter (separate subscribe/send). Customize fetchServerSentEvents() and fetchHttpStream() with auth headers, custom URLs, and request options. Import from framework package, not @tanstack/ai-client.

Browser chat persistence on useChat / ChatClient: localStoragePersistence, sessionStoragePersistence, indexedDBPersistence. Client-authoritative (adapter, full transcript) vs server-authoritative (persistence: true, no client cache). Reload restore, pending interrupts, mid-stream rejoin with delivery durability. Use for SPA reload durability — NOT server history alone. Also covers generation hooks (useGenerateImage etc.), which take only the server-driven mode: persistence: true hydrates the last generation for the (REQUIRED) threadId from the server on mount and repaints status/result/error, nothing is cached in the browser. No extra package: the adapters ship in the framework packages.

Server-side AG-UI streaming protocol implementation: StreamChunk event types (RUN_STARTED, TEXT_MESSAGE_START/CONTENT/END, TOOL_CALL_START/ARGS/END, RUN_FINISHED, RUN_ERROR, STEP_STARTED/STEP_FINISHED, STATE_SNAPSHOT/DELTA, CUSTOM), toServerSentEventsStream() for SSE format, toHttpStream() for NDJSON format. For backends serving AG-UI events without client packages.

Provider adapter selection and configuration: openaiText, anthropicText, geminiText, ollamaText, grokText, groqText, openRouterText, bedrockText, byteplusText, openaiCompatible. Per-model type safety with modelOptions, reasoning/thinking configuration, runtime adapter switching, extendAdapter() for custom models, createModel(). Generic OpenAI-compatible providers (DeepSeek, Together, Fireworks, etc.) via openaiCompatible({ baseURL, apiKey, models }) from @tanstack/ai-openai/compatible. API key env vars: OPENAI_API_KEY, ANTHROPIC_API_KEY, GOOGLE_API_KEY/GEMINI_API_KEY, XAI_API_KEY, GROQ_API_KEY, OPENROUTER_API_KEY, OLLAMA_HOST, BEDROCK_API_KEY (or AWS_BEARER_TOKEN_BEDROCK). BytePlus needs TWO keys: ARK_API_KEY (ModelArk — chat/video/image) and BYTEPLUS_VOICE_API_KEY (Seed Speech — TTS/transcription); neither is a fallback for the other.

Run harness adapters (Claude Code, Codex, OpenCode) INSIDE isolated sandboxes via defineSandbox + withSandbox + a provider (localProcessSandbox / dockerSandbox). Covers declarative provisioning: createSecrets + secret/bearer, skills (agentSkill/gitSkill/mcpSkill/ fileSkill), plugins, instructions → canonical AGENTS.md + symlinks projected per harness; shallow-clone default with depth opt-out; serial/parallel setup callback over a persistent shell; snapshot-after-setup default with snapshotMaxAge TTL. It also covers portable snapshots after a successful terminal run with withPersistence before withSandbox and memorySandboxSnapshots for local examples. It covers named saves with snapshots.save, selected-checkpoint forks with snapshots.fork, and authorized artifact reads with snapshots.readArtifact. See docs/sandbox/portable-snapshots.md. It covers defineWorkspace (git/setup/scripts/skills/secrets/ instructions/plugins), defineSandboxPolicy (allow/ask/deny), lifecycle/resume, the SandboxHandle (fs/git/process/ports), capability tokens, defineSandbox hooks (onFile/onFileCreate/onFileChange/onFileDelete/onReady/onError/ onDestroy) + fileEvents flag, chat middleware sandbox group (defineChatMiddleware sandbox hooks), the sandbox debug category, watchWorkspace as a low-level building block, the file.changed / sandbox.file / claude-code.session-id events, and the run journal (spawnNdjson journal option, runId uniqueness, follow vs bounded-poll reading, alignToStoredLog replay alignment, chunkFingerprint, createRunScopedIdGen), and takeover of detached runs (withSandbox runs+durability as one opt-in, detach vs cancel via requestRunCancel / RUN_CANCEL_REASON, sandboxRunDriver on the resume path, single-writer fencing of BOTH the event log and the run record, replay-from-zero with JournalReplayDivergedError, the distributed LockStore requirement). Use whenever a harness adapter needs a sandbox or when building sandbox providers.

Implement the MessageStore, RunStore, InterruptStore, MetadataStore contracts for @tanstack/ai-persistence against any database. defineAIPersistence, composePersistence overrides, critical invariants (full-replace saveThread, insert-if-absent createOrResume and interrupt create), authorize thread access, runPersistenceConformance testkit. Use whenever you need server persistence — the package ships contracts, not a backend for your database.

Server chat state with withPersistence from @tanstack/ai-persistence. Authoritative transcript, run lifecycle, durable interrupts/approvals, chatParamsFromRequest, reconstructChat, snapshotStreaming. Use when the server owns history, multi-device, or durable tool approvals. NOT client localStorage (see ai-core/client-persistence in @tanstack/ai) and NOT stream reconnect alone.

Use when an app already runs Prisma and needs TanStack AI chat persistence — writes a chat-persistence.ts into the app against its existing PrismaClient and schema.prisma. Covers the four models, BigInt timestamps, JSON-as-string columns, upsert-with-empty-update idempotency, and model renaming.

Use when an app already runs Drizzle ORM and needs TanStack AI chat persistence — writes a chat-persistence.ts into the app against its existing db handle, schema file, and drizzle-kit journal. Covers the four tables (SQLite/Postgres/MySQL), the onConflict idempotency rules, JSON columns, and per-request bindings like D1.

Use when an app needs TanStack AI chat persistence on a database with no dedicated recipe — raw Postgres (pg/postgres.js), Kysely, node:sqlite, MongoDB, Supabase, Redis. Writes a chat-persistence.ts against the app's existing client, covering the four stores, the idempotency invariants, and the conformance gate. Route to the Drizzle, Prisma, or Cloudflare skills instead when one of those matches.

Use when a Cloudflare Worker needs TanStack AI chat persistence — writes a chat-persistence.ts into the app against its D1 binding (raw or via Drizzle), plus a Durable Object LockStore. Covers per-request bindings, wrangler config, D1 migrations, and lease-based locks.

AI becomes provocative and mocks the opponent during gameplay

WeWrite 写作模块：在公众号选题明确后完成文章任务书、主张与证据、素材和初稿。由主流程调用， 或响应“就这个选题写正文”。通用写作、博客和短视频文案不触发。

WeWrite 视觉模块：为公众号文章生成封面和必要的内文配图，或只交付提示词。触发词： 封面图、公众号配图、给文章配图、换封面。通用绘图和 logo 设计不触发。