@happyvertical/ai 0.110.1 → 0.111.1

package/dist/chunks/{webllm-DVDsgVT7.js.map → webllm-DrwR5V-y.js.map}

@@ -1 +1 @@
-{"version":3,"file":"webllm-DVDsgVT7.js","names":[],"sources":["../../src/shared/providers/webllm.ts"],"sourcesContent":["// biome-ignore-all lint/style/useNamingConvention: WebLLM's OpenAI-shaped wire format and model records are snake_case.\n/**\n * WebLLM provider: runs a chat model entirely in the browser on WebGPU.\n *\n * Backed by `@mlc-ai/web-llm`, an optional peer dependency that is imported\n * lazily on first use, so neither the root entry nor `getAI()` for other\n * provider types ever loads it. Reachable through `getAI({ type: 'webllm' })`\n * and the `@happyvertical/ai/local` subpath.\n *\n * WebLLM's engine API is OpenAI-shaped, so requests and responses map almost\n * one to one. Structured output (`responseSchema`) becomes WebLLM's\n * grammar-constrained `response_format`, which keeps a small model from\n * emitting a value outside the schema.\n */\n\nimport {\n  normalizeBaseAIOptions,\n  normalizeChatOptions,\n  type PreparedRequestControls,\n  prepareRequestControls,\n} from '../safety';\nimport type {\n  AICapabilities,\n  AIInterface,\n  AIMessage,\n  AIModel,\n  AIResponse,\n  AIToolCall,\n  ChatOptions,\n  CompletionOptions,\n  EmbeddingOptions,\n  EmbeddingResponse,\n  ImageDescriptionOptions,\n  ImageEmbeddingOptions,\n  ImageGenerationOptions,\n  ImageGenerationResponse,\n  MessageOptions,\n  TokenUsage,\n  TTSOptions,\n  TTSResponse,\n  VideoGenerationJob,\n  VideoGenerationOptions,\n  VideoGenerationResult,\n  VideoGenerationStatusResult,\n  Voice,\n  VoiceCloneOptions,\n  VoiceDesignOptions,\n  VoiceListOptions,\n  WebLLMEngineLike,\n  WebLLMLoadProgress,\n  WebLLMOptions,\n} from '../types';\nimport {\n  AIError,\n  ContextLengthError,\n  extractTextContent,\n  ModelNotFoundError,\n  WebGPUUnavailableError,\n  WebLLMPeerMissingError,\n} from '../types';\nimport { emitUsage } from './usage';\n\nconst PROVIDER = 'webllm';\n\n/** Small enough for most GPUs (about 0.9 GB of VRAM); override with `defaultModel`. */\nexport const DEFAULT_WEBLLM_MODEL = 'Llama-3.2-1B-Instruct-q4f16_1-MLC';\n\n/** Context window assumed when a model record does not declare one. */\nconst DEFAULT_CONTEXT_LENGTH = 4096;\n\nconst WEBLLM_CAPABILITIES: AICapabilities = {\n  chat: true,\n  completion: true,\n  embeddings: false,\n  streaming: true,\n  functions: true,\n  vision: false,\n  fineTuning: false,\n  imageEmbeddings: false,\n  imageGeneration: false,\n  videoGeneration: false,\n  tts: false,\n  voiceCloning: false,\n  voiceDesign: false,\n  maxContextLength: DEFAULT_CONTEXT_LENGTH,\n  supportedOperations: ['chat', 'completion', 'streaming', 'functions'],\n};\n\n/** Structural subset of the `@mlc-ai/web-llm` module this provider uses. */\ninterface WebLLMModule {\n  CreateMLCEngine(\n    model: string,\n    config?: {\n      appConfig?: unknown;\n      initProgressCallback?: (report: WebLLMLoadProgress) => void;\n    },\n  ): Promise<WebLLMEngineLike>;\n  prebuiltAppConfig: { model_list: WebLLMModelRecord[] };\n  functionCallingModelIds?: string[];\n}\n\ninterface WebLLMModelRecord {\n  model_id: string;\n  vram_required_MB?: number;\n  low_resource_required?: boolean;\n  /** WebLLM `ModelType`: 0 or absent = LLM, 1 = embedding, 2 = VLM. */\n  model_type?: number;\n  overrides?: { context_window_size?: number };\n}\n\ninterface EngineEntry {\n  promise: Promise<WebLLMEngineLike>;\n  listeners: Set<(report: WebLLMLoadProgress) => void>;\n}\n\n/**\n * Per-engine request queues. WebLLM runs one generation at a time per engine\n * and `interruptGenerate()` stops whichever one is running, so requests are\n * serialized here and only the running request may interrupt.\n */\nconst engineQueues = new WeakMap<object, Promise<void>>();\n\n/** Engines shared by every provider instance that uses the prebuilt model list. */\nconst sharedEngines = new Map<string, EngineEntry>();\n\n/**\n * Imports the optional peer. Isolated so a missing install becomes a typed\n * {@link WebLLMPeerMissingError} instead of a bare module-resolution failure.\n */\nexport async function importWebLLM(): Promise<WebLLMModule> {\n  try {\n    const loaded = (await import('@mlc-ai/web-llm')) as unknown as\n      | (WebLLMModule & { default?: WebLLMModule })\n      | undefined;\n    const module =\n      typeof loaded?.CreateMLCEngine === 'function' ? loaded : loaded?.default;\n    if (typeof module?.CreateMLCEngine !== 'function') {\n      throw new Error('@mlc-ai/web-llm did not export CreateMLCEngine');\n    }\n    return module;\n  } catch (error) {\n    throw new WebLLMPeerMissingError(error);\n  }\n}\n\n/** True when this runtime exposes WebGPU (`navigator.gpu`). */\nexport function isWebGPUAvailable(): boolean {\n  const nav = (globalThis as { navigator?: { gpu?: unknown } }).navigator;\n  return Boolean(nav?.gpu);\n}\n\n/**\n * Unloads every engine this module created and forgets them. Caller-supplied\n * engines are never touched.\n */\nexport async function disposeWebLLMEngines(): Promise<void> {\n  const entries = [...sharedEngines.values()];\n  sharedEngines.clear();\n  await Promise.allSettled(\n    entries.map(async (entry) => (await entry.promise).unload?.()),\n  );\n}\n\nfunction unsupported(feature: string): never {\n  throw new AIError(\n    `${feature} is not supported by the webllm provider.`,\n    'NOT_SUPPORTED',\n    PROVIDER,\n  );\n}\n\nfunction stripUndefined<T extends Record<string, unknown>>(value: T): T {\n  return Object.fromEntries(\n    Object.entries(value).filter(([, entry]) => entry !== undefined),\n  ) as T;\n}\n\n/**\n * In-browser chat provider on WebLLM (WebGPU). See the module header.\n */\nexport class WebLLMProvider implements AIInterface {\n  private options: WebLLMOptions;\n  /** Model the caller named, if any; a supplied engine is loaded with it already. */\n  private readonly explicitModel?: string;\n  private factoryEngine?: Promise<WebLLMEngineLike>;\n  private readonly localEngines = new Map<string, EngineEntry>();\n\n  constructor(options: WebLLMOptions) {\n    this.explicitModel = options.defaultModel ?? options.model;\n    this.options = normalizeBaseAIOptions({\n      ...options,\n      defaultModel: this.explicitModel ?? DEFAULT_WEBLLM_MODEL,\n    });\n  }\n\n  // --------------------------------------------------------------------------\n  // Engine lifecycle\n  // --------------------------------------------------------------------------\n\n  private async acquireEngine(\n    model: string,\n    signal?: AbortSignal,\n  ): Promise<WebLLMEngineLike> {\n    const supplied = this.options.engine;\n    if (supplied) {\n      if (typeof supplied !== 'function') return supplied;\n      this.factoryEngine ??= supplied().catch((error) => {\n        this.factoryEngine = undefined;\n        throw error;\n      });\n      return abortable(this.factoryEngine, signal);\n    }\n\n    if (!isWebGPUAvailable()) {\n      throw new WebGPUUnavailableError(PROVIDER);\n    }\n\n    // A custom appConfig changes what a model id means, so it must not share\n    // an engine with other callers.\n    const cache = this.options.appConfig ? this.localEngines : sharedEngines;\n    let entry = cache.get(model);\n    if (!entry) {\n      const listeners = new Set<(report: WebLLMLoadProgress) => void>();\n      const created: EngineEntry = {\n        listeners,\n        promise: this.createEngine(model, listeners),\n      };\n      created.promise.catch(() => {\n        if (cache.get(model) === created) cache.delete(model);\n      });\n      cache.set(model, created);\n      entry = created;\n    }\n\n    const onProgress = this.options.onLoadProgress;\n    if (onProgress) entry.listeners.add(onProgress);\n    try {\n      return await abortable(entry.promise, signal);\n    } finally {\n      if (onProgress) entry.listeners.delete(onProgress);\n    }\n  }\n\n  private async createEngine(\n    model: string,\n    listeners: Set<(report: WebLLMLoadProgress) => void>,\n  ): Promise<WebLLMEngineLike> {\n    const webllm = await importWebLLM();\n    return webllm.CreateMLCEngine(model, {\n      ...(this.options.appConfig ? { appConfig: this.options.appConfig } : {}),\n      initProgressCallback: (report) => {\n        for (const listener of listeners) {\n          try {\n            listener(report);\n          } catch {\n            // A progress listener must never break model loading.\n          }\n        }\n      },\n    });\n  }\n\n  // --------------------------------------------------------------------------\n  // Request / response mapping\n  // --------------------------------------------------------------------------\n\n  private mapMessages(messages: AIMessage[]): Record<string, unknown>[] {\n    return messages.map((message) => {\n      if (message.role === 'user' && Array.isArray(message.content)) {\n        return {\n          role: 'user',\n          content: message.content.map((part) =>\n            part.type === 'text'\n              ? { type: 'text', text: part.text }\n              : { type: 'image_url', image_url: { url: part.image_url.url } },\n          ),\n          ...(message.name ? { name: message.name } : {}),\n        };\n      }\n      const content = extractTextContent(message.content);\n      if (message.role === 'tool' || message.role === 'function') {\n        return {\n          role: 'tool',\n          content,\n          tool_call_id: message.tool_call_id ?? message.name ?? '',\n        };\n      }\n      if (message.role === 'assistant') {\n        return stripUndefined({\n          role: 'assistant',\n          content: content || (message.tool_calls?.length ? null : ''),\n          tool_calls: message.tool_calls?.map((call) => ({\n            id: call.id,\n            type: call.type,\n            function: {\n              name: call.function.name,\n              arguments: call.function.arguments,\n            },\n          })),\n        });\n      }\n      return {\n        role: message.role,\n        content,\n        ...(message.name ? { name: message.name } : {}),\n      };\n    });\n  }\n\n  private mapResponseFormat(\n    options: Pick<ChatOptions, 'responseFormat' | 'responseSchema'>,\n  ): Record<string, unknown> | undefined {\n    if (options.responseSchema !== undefined) {\n      const schema =\n        typeof options.responseSchema === 'string'\n          ? options.responseSchema\n          : JSON.stringify(options.responseSchema);\n      // WebLLM compiles the schema to a grammar and masks logits, so the\n      // output cannot leave it.\n      return { type: 'json_object', schema };\n    }\n    return options.responseFormat?.type === 'json_object'\n      ? { type: 'json_object' }\n      : undefined;\n  }\n\n  private buildRequest(\n    messages: AIMessage[],\n    options: ChatOptions & { maxTokens: number },\n    model: string | undefined,\n    stream: boolean,\n    includeTools: boolean,\n  ): Record<string, unknown> {\n    const tools =\n      includeTools && options.tools?.length\n        ? options.tools.map((tool) => ({\n            type: 'function' as const,\n            function: {\n              name: tool.function.name,\n              description: tool.function.description,\n              parameters: tool.function.parameters,\n            },\n          }))\n        : undefined;\n    return stripUndefined({\n      model,\n      messages: this.mapMessages(messages),\n      max_tokens: options.maxTokens,\n      temperature: options.temperature,\n      top_p: options.topP,\n      n: options.n,\n      stop: options.stop,\n      frequency_penalty: options.frequencyPenalty,\n      presence_penalty: options.presencePenalty,\n      seed: options.seed,\n      tools,\n      tool_choice: tools ? options.toolChoice : undefined,\n      response_format: this.mapResponseFormat(options),\n      stream,\n      stream_options: stream ? { include_usage: true } : undefined,\n    });\n  }\n\n  /** Model id to send. A caller-supplied engine already has its model loaded. */\n  private requestModel(\n    model: string,\n    options: ChatOptions,\n  ): string | undefined {\n    if (!this.options.engine) return model;\n    return options.model ?? this.explicitModel;\n  }\n\n  private mapUsage(usage: unknown): TokenUsage | undefined {\n    const raw = usage as\n      | {\n          prompt_tokens?: number;\n          completion_tokens?: number;\n          total_tokens?: number;\n        }\n      | undefined\n      | null;\n    if (!raw) return undefined;\n    return {\n      promptTokens: raw.prompt_tokens ?? 0,\n      completionTokens: raw.completion_tokens ?? 0,\n      totalTokens: raw.total_tokens ?? 0,\n    };\n  }\n\n  private mapFinishReason(reason: string | null | undefined) {\n    switch (reason) {\n      case 'length':\n        return 'length' as const;\n      case 'tool_calls':\n        return 'tool_calls' as const;\n      default:\n        return 'stop' as const;\n    }\n  }\n\n  /**\n   * WebLLM reports `abort` when generation was interrupted. The provider's own\n   * aborts are raised before this point, so reaching it means someone else\n   * interrupted the engine (for example a caller sharing a supplied engine);\n   * a truncated reply must not look like a normal stop.\n   */\n  private assertNotInterrupted(\n    reason: string | null | undefined,\n    model: string,\n  ): void {\n    if (reason === 'abort') {\n      throw new AIError(\n        'Generation was interrupted before it finished',\n        'AI_INTERRUPTED',\n        PROVIDER,\n        model,\n        true,\n      );\n    }\n  }\n\n  private mapError(error: unknown, model?: string): AIError {\n    if (error instanceof AIError) return error;\n    const name = (error as { name?: string } | undefined)?.name ?? '';\n    const message = error instanceof Error ? error.message : String(error);\n    if (name === 'WebGPUNotAvailableError' || name === 'WebGPUNotFoundError') {\n      return new WebGPUUnavailableError(PROVIDER, message);\n    }\n    if (\n      name === 'ContextWindowSizeExceededError' ||\n      name === 'PrefillChunkSizeSmallerThanImageError'\n    ) {\n      return new ContextLengthError(PROVIDER, model);\n    }\n    if (\n      name === 'ModelNotFoundError' ||\n      name === 'SpecifiedModelNotFoundError' ||\n      name === 'UnsupportedModelIdError' ||\n      name === 'ModelNotLoadedError'\n    ) {\n      return new ModelNotFoundError(model ?? message, PROVIDER);\n    }\n    const mapped = new AIError(\n      message,\n      name === 'DeviceLostError' ? 'WEBGPU_DEVICE_LOST' : 'API_ERROR',\n      PROVIDER,\n      model,\n    );\n    mapped.cause = error;\n    return mapped;\n  }\n\n  /** Throws the right AI error when the request ended by timeout or caller abort. */\n  private abortError(\n    controls: PreparedRequestControls | undefined,\n    options: ChatOptions,\n    model?: string,\n  ): AIError | undefined {\n    if (controls?.didTimeout()) {\n      return new AIError(\n        `AI request timed out after ${controls.timeout}ms`,\n        'AI_TIMEOUT',\n        PROVIDER,\n        model,\n      );\n    }\n    if (options.signal?.aborted) {\n      return new AIError(\n        'AI request aborted by caller',\n        'AI_ABORTED',\n        PROVIDER,\n        model,\n      );\n    }\n    return undefined;\n  }\n\n  /**\n   * Waits for this request's turn on the engine; resolves to a release fn. A\n   * queued request that is aborted or times out leaves the queue at once and\n   * never interrupts the generation ahead of it.\n   */\n  private async lockEngine(\n    engine: WebLLMEngineLike,\n    signal: AbortSignal,\n  ): Promise<() => void> {\n    const previous = engineQueues.get(engine) ?? Promise.resolve();\n    let release: () => void = () => {};\n    const mine = new Promise<void>((resolve) => {\n      release = resolve;\n    });\n    engineQueues.set(\n      engine,\n      previous.then(() => mine),\n    );\n    try {\n      await abortable(previous, signal);\n    } catch (error) {\n      release();\n      throw error;\n    }\n    return release;\n  }\n\n  /** Interrupts generation when the composed signal fires; returns a detach fn. */\n  private bindInterrupt(\n    engine: WebLLMEngineLike,\n    signal: AbortSignal,\n  ): () => void {\n    const interrupt = () => {\n      try {\n        void Promise.resolve(engine.interruptGenerate()).catch(() => {});\n      } catch {\n        // Best effort: the request is already being abandoned.\n      }\n    };\n    if (signal.aborted) interrupt();\n    signal.addEventListener('abort', interrupt, { once: true });\n    return () => signal.removeEventListener('abort', interrupt);\n  }\n\n  // --------------------------------------------------------------------------\n  // Chat\n  // --------------------------------------------------------------------------\n\n  async chat(\n    messages: AIMessage[],\n    options: ChatOptions = {},\n  ): Promise<AIResponse> {\n    const startTime = Date.now();\n    let controls: PreparedRequestControls | undefined;\n    let detach: (() => void) | undefined;\n    let release: (() => void) | undefined;\n    let pending: Promise<any> | undefined;\n    let holdRelease: (() => Promise<unknown> | undefined) | undefined;\n    const model =\n      options.model || this.options.defaultModel || DEFAULT_WEBLLM_MODEL;\n    try {\n      options = normalizeChatOptions(this.options, options, PROVIDER, model);\n      // Load before arming the request timeout: a multi-GB download must not\n      // consume the generation budget.\n      const engine = await this.acquireEngine(model, options.signal);\n      controls = prepareRequestControls(this.options, options);\n      release = await this.lockEngine(engine, controls.signal);\n      detach = this.bindInterrupt(engine, controls.signal);\n\n      pending = engine.chat.completions.create(\n        this.buildRequest(\n          messages,\n          options as ChatOptions & { maxTokens: number },\n          this.requestModel(model, options),\n          false,\n          true,\n        ),\n      );\n      const started: Promise<unknown> = pending;\n      let settled = false;\n      const markSettled = () => {\n        settled = true;\n      };\n      started.then(markSettled, markSettled);\n      holdRelease = () => (settled ? undefined : started);\n      const response = await abortable(pending, controls.signal);\n      if (controls.signal.aborted) {\n        throw (\n          this.abortError(controls, options, model) ?? controls.signal.reason\n        );\n      }\n\n      const choice = response?.choices?.[0];\n      this.assertNotInterrupted(choice?.finish_reason, model);\n      if (!choice) {\n        throw new AIError(\n          'No choices returned from webllm',\n          'NO_CHOICES',\n          PROVIDER,\n        );\n      }\n      const usage = this.mapUsage(response.usage);\n      emitUsage(\n        this.options,\n        PROVIDER,\n        'chat',\n        response.model || model,\n        usage,\n        startTime,\n        options.usageTags,\n      );\n      const toolCalls: AIToolCall[] | undefined =\n        choice.message?.tool_calls?.map(\n          (call: {\n            id: string;\n            function: { name: string; arguments: string };\n          }) => ({\n            id: call.id,\n            type: 'function' as const,\n            function: {\n              name: call.function.name,\n              arguments: call.function.arguments,\n            },\n          }),\n        );\n      return {\n        content: choice.message?.content || '',\n        usage,\n        model: response.model || model,\n        finishReason: this.mapFinishReason(choice.finish_reason),\n        ...(toolCalls?.length ? { toolCalls } : {}),\n      };\n    } catch (error) {\n      throw (\n        this.abortError(controls, options, model) ?? this.mapError(error, model)\n      );\n    } finally {\n      detach?.();\n      releaseWhenIdle(release, holdRelease?.());\n      controls?.cleanup();\n    }\n  }\n\n  async complete(\n    prompt: string,\n    options: CompletionOptions = {},\n  ): Promise<AIResponse> {\n    return this.chat([{ role: 'user', content: prompt }], {\n      model: options.model,\n      maxTokens: options.maxTokens,\n      temperature: options.temperature,\n      topP: options.topP,\n      n: options.n,\n      stop: options.stop,\n      stream: options.stream,\n      onProgress: options.onProgress,\n      signal: options.signal,\n      timeout: options.timeout,\n      reasoning: options.reasoning,\n      usageTags: options.usageTags,\n      continueOnLength: options.continueOnLength,\n    });\n  }\n\n  async message(text: string, options: MessageOptions = {}): Promise<string> {\n    const messages: AIMessage[] = [\n      ...(options.history || []),\n      { role: options.role || 'user', content: text },\n    ];\n    const response = await this.chat(messages, {\n      model: options.model,\n      maxTokens: options.maxTokens,\n      temperature: options.temperature,\n      topP: options.topP,\n      stop: options.stop,\n      stream: options.stream,\n      frequencyPenalty: options.frequencyPenalty,\n      presencePenalty: options.presencePenalty,\n      responseFormat: options.responseFormat,\n      responseSchema: options.responseSchema,\n      seed: options.seed,\n      tools: options.tools,\n      toolChoice: options.toolChoice,\n      onProgress: options.onProgress,\n      signal: options.signal,\n      timeout: options.timeout,\n      reasoning: options.reasoning,\n      usageTags: options.usageTags,\n      continueOnLength: options.continueOnLength,\n    });\n    return response.content;\n  }\n\n  /**\n   * Streams a chat completion as text chunks. Like the other providers, a\n   * stream yields text only; use `chat()` for rounds that may call tools.\n   * Breaking out of the loop interrupts generation on the engine.\n   */\n  async *stream(\n    messages: AIMessage[],\n    options: ChatOptions = {},\n  ): AsyncIterable<string> {\n    const startTime = Date.now();\n    let controls: PreparedRequestControls | undefined;\n    let detach: (() => void) | undefined;\n    let engine: WebLLMEngineLike | undefined;\n    let release: (() => void) | undefined;\n    let iterator: AsyncIterator<WebLLMChunk> | undefined;\n    let iteratorDone = false;\n    let holdRelease: Promise<unknown> | undefined;\n    const model =\n      options.model || this.options.defaultModel || DEFAULT_WEBLLM_MODEL;\n    try {\n      options = normalizeChatOptions(this.options, options, PROVIDER, model);\n      engine = await this.acquireEngine(model, options.signal);\n      controls = prepareRequestControls(this.options, options);\n      release = await this.lockEngine(engine, controls.signal);\n      detach = this.bindInterrupt(engine, controls.signal);\n\n      const pending: Promise<AsyncIterable<WebLLMChunk>> =\n        engine.chat.completions.create(\n          this.buildRequest(\n            messages,\n            options as ChatOptions & { maxTokens: number },\n            this.requestModel(model, options),\n            true,\n            false,\n          ),\n        );\n      let chunks: AsyncIterable<WebLLMChunk>;\n      try {\n        chunks = await abortable(pending, controls.signal);\n      } catch (error) {\n        // The caller is gone but the engine may still start this request.\n        // Keep our queue closed until it has been interrupted and drained, or\n        // the next request would wait behind it and could interrupt it.\n        holdRelease = pending.then(\n          (late) => abandonStream(late[Symbol.asyncIterator](), engine),\n          () => {},\n        );\n        throw error;\n      }\n      // Iterated by hand: WebLLM releases its per-model lock only when its\n      // generator runs to completion, never on an early `return()`.\n      iterator = chunks[Symbol.asyncIterator]();\n\n      let finishReason: string | null | undefined;\n      let usage: TokenUsage | undefined;\n      for (;;) {\n        const step = await iterator.next();\n        if (step.done) {\n          iteratorDone = true;\n          break;\n        }\n        const chunk = step.value;\n        const choice = chunk.choices?.[0];\n        finishReason = choice?.finish_reason ?? finishReason;\n        if (chunk.usage) usage = this.mapUsage(chunk.usage);\n        const content = choice?.delta?.content;\n        if (content) {\n          options.onProgress?.(content);\n          yield content;\n        }\n      }\n      if (controls.signal.aborted) {\n        throw (\n          this.abortError(controls, options, model) ?? controls.signal.reason\n        );\n      }\n      this.assertNotInterrupted(finishReason, model);\n      options.onFinishReason?.(this.mapFinishReason(finishReason));\n      emitUsage(\n        this.options,\n        PROVIDER,\n        'stream',\n        model,\n        usage,\n        startTime,\n        options.usageTags,\n      );\n    } catch (error) {\n      throw (\n        this.abortError(controls, options, model) ?? this.mapError(error, model)\n      );\n    } finally {\n      if (iterator && !iteratorDone && engine) {\n        // Consumer stopped early (or the loop threw): stop the GPU, then let\n        // the engine's generator finish so its lock is released before the\n        // next request is admitted.\n        try {\n          void Promise.resolve(engine.interruptGenerate()).catch(() => {});\n        } catch {\n          // Best effort.\n        }\n        await drainIterator(iterator);\n      }\n      detach?.();\n      releaseWhenIdle(release, holdRelease);\n      controls?.cleanup();\n    }\n  }\n\n  async countTokens(text: string): Promise<number> {\n    // WebLLM exposes no standalone tokenizer; ~4 characters per token.\n    return Math.ceil(text.length / 4);\n  }\n\n  // --------------------------------------------------------------------------\n  // Models and capabilities\n  // --------------------------------------------------------------------------\n\n  async getModels(): Promise<AIModel[]> {\n    const webllm = await importWebLLM();\n    const list =\n      (\n        this.options.appConfig as\n          | { model_list?: WebLLMModelRecord[] }\n          | undefined\n      )?.model_list ?? webllm.prebuiltAppConfig.model_list;\n    const functionCalling = new Set(webllm.functionCallingModelIds ?? []);\n    return list\n      .filter((record) => record.model_type !== 1) // embedding models\n      .map((record) => {\n        const vision = record.model_type === 2;\n        const functions = functionCalling.has(record.model_id);\n        return {\n          id: record.model_id,\n          name: record.model_id,\n          description: record.vram_required_MB\n            ? `Runs on-device via WebGPU (~${Math.round(record.vram_required_MB)} MB VRAM)`\n            : 'Runs on-device via WebGPU',\n          contextLength:\n            record.overrides?.context_window_size ?? DEFAULT_CONTEXT_LENGTH,\n          capabilities: [\n            'text',\n            'chat',\n            ...(functions ? ['functions'] : []),\n            ...(vision ? ['vision'] : []),\n          ],\n          supportsFunctions: functions,\n          supportsVision: vision,\n        };\n      });\n  }\n\n  async getCapabilities(): Promise<AICapabilities> {\n    return { ...WEBLLM_CAPABILITIES };\n  }\n\n  // --------------------------------------------------------------------------\n  // No in-browser equivalent\n  // --------------------------------------------------------------------------\n\n  async embed(\n    _text: string | string[],\n    _options: EmbeddingOptions = {},\n  ): Promise<EmbeddingResponse> {\n    return unsupported('Embeddings');\n  }\n\n  async embedImage(\n    _image: string | Buffer,\n    _options: ImageEmbeddingOptions = {},\n  ): Promise<EmbeddingResponse> {\n    return unsupported('Image embeddings');\n  }\n\n  async describeImage(\n    _image: string | Buffer,\n    _prompt?: string,\n    _options: ImageDescriptionOptions = {},\n  ): Promise<string> {\n    return unsupported('Image description');\n  }\n\n  async generateImage(\n    _prompt: string,\n    _options: ImageGenerationOptions = {},\n  ): Promise<ImageGenerationResponse> {\n    return unsupported('Image generation');\n  }\n\n  async submitVideoGenerationJob(\n    _options: VideoGenerationOptions,\n  ): Promise<VideoGenerationJob> {\n    return unsupported('Video generation');\n  }\n\n  async getVideoGenerationJob(\n    _handle: VideoGenerationJob,\n  ): Promise<VideoGenerationStatusResult> {\n    return unsupported('Video generation');\n  }\n\n  async fetchVideoGenerationResult(\n    _handle: VideoGenerationJob,\n  ): Promise<VideoGenerationResult> {\n    return unsupported('Video generation');\n  }\n\n  async cancelVideoGenerationJob(_handle: VideoGenerationJob): Promise<void> {\n    return unsupported('Video generation');\n  }\n\n  async validateVideoGenerationAccess(): Promise<boolean> {\n    return unsupported('Video generation');\n  }\n\n  async synthesizeSpeech(\n    _text: string,\n    _options?: TTSOptions,\n  ): Promise<TTSResponse> {\n    return unsupported('Text-to-speech');\n  }\n\n  streamSpeech(_text: string, _options?: TTSOptions): AsyncIterable<Buffer> {\n    const error = new AIError(\n      'Text-to-speech streaming is not supported by the webllm provider.',\n      'NOT_SUPPORTED',\n      PROVIDER,\n    );\n    return {\n      [Symbol.asyncIterator]: () => ({\n        next: () => Promise.reject(error),\n      }),\n    };\n  }\n\n  async cloneVoice(_options: VoiceCloneOptions): Promise<Voice> {\n    return unsupported('Voice cloning');\n  }\n\n  async designVoice(_options: VoiceDesignOptions): Promise<Voice> {\n    return unsupported('Voice design');\n  }\n\n  async getVoices(_options?: VoiceListOptions): Promise<Voice[]> {\n    return unsupported('Voice listing');\n  }\n}\n\n/** One streamed WebLLM chunk (structural subset). */\ninterface WebLLMChunk {\n  choices?: Array<{\n    delta?: { content?: string | null };\n    finish_reason?: string | null;\n  }>;\n  usage?: unknown;\n}\n\n/** Runs a WebLLM stream to its end, ignoring errors; used after interrupting it. */\nasync function drainIterator(\n  iterator: AsyncIterator<WebLLMChunk>,\n): Promise<void> {\n  try {\n    while (!(await iterator.next()).done) {\n      // discard\n    }\n  } catch {\n    // The engine already stopped; nothing more to release.\n  }\n}\n\n/** Releases an engine queue slot now, or once `idle` has settled. */\nfunction releaseWhenIdle(\n  release: (() => void) | undefined,\n  idle?: Promise<unknown>,\n): void {\n  if (!release) return;\n  if (idle) void idle.then(release, release);\n  else release();\n}\n\n/**\n * Finishes a stream nobody is reading. WebLLM resets its interrupt flag on the\n * first `next()`, so interrupt only after pulling one chunk, then drain it so\n * the engine's lock is released.\n */\nasync function abandonStream(\n  iterator: AsyncIterator<WebLLMChunk>,\n  engine: WebLLMEngineLike | undefined,\n): Promise<void> {\n  try {\n    if ((await iterator.next()).done) return;\n    await Promise.resolve(engine?.interruptGenerate());\n  } catch {\n    return;\n  }\n  await drainIterator(iterator);\n}\n\n/** Rejects with the signal's reason if it fires first; the underlying work continues. */\nfunction abortable<T>(promise: Promise<T>, signal?: AbortSignal): Promise<T> {\n  if (!signal) return promise;\n  if (signal.aborted) {\n    return Promise.reject(\n      new AIError('AI request aborted by caller', 'AI_ABORTED', PROVIDER),\n    );\n  }\n  return new Promise<T>((resolve, reject) => {\n    const onAbort = () =>\n      reject(\n        new AIError('AI request aborted by caller', 'AI_ABORTED', PROVIDER),\n      );\n    signal.addEventListener('abort', onAbort, { once: true });\n    promise.then(\n      (value) => {\n        signal.removeEventListener('abort', onAbort);\n        resolve(value);\n      },\n      (error) => {\n        signal.removeEventListener('abort', onAbort);\n        reject(error);\n      },\n    );\n  });\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;AA8DA,IAAM,WAAW;;AAGjB,IAAa,uBAAuB;;AAGpC,IAAM,yBAAyB;AAE/B,IAAM,sBAAsC;CAC1C,MAAM;CACN,YAAY;CACZ,YAAY;CACZ,WAAW;CACX,WAAW;CACX,QAAQ;CACR,YAAY;CACZ,iBAAiB;CACjB,iBAAiB;CACjB,iBAAiB;CACjB,KAAK;CACL,cAAc;CACd,aAAa;CACb,kBAAkB;CAClB,qBAAqB;EAAC;EAAQ;EAAc;EAAa;CAAW;AACtE;;;;;;AAkCA,IAAM,+BAAe,IAAI,QAA+B;;AAGxD,IAAM,gCAAgB,IAAI,IAAyB;;;;;AAMnD,eAAsB,eAAsC;CAC1D,IAAI;EACF,MAAM,SAAU,MAAM,OAAO;EAG7B,MAAM,SACJ,OAAO,QAAQ,oBAAoB,aAAa,SAAS,QAAQ;EACnE,IAAI,OAAO,QAAQ,oBAAoB,YACrC,MAAM,IAAI,MAAM,gDAAgD;EAElE,OAAO;CACT,SAAS,OAAO;EACd,MAAM,IAAI,uBAAuB,KAAK;CACxC;AACF;;AAGA,SAAgB,oBAA6B;CAC3C,MAAM,MAAO,WAAiD;CAC9D,OAAO,QAAQ,KAAK,GAAG;AACzB;;;;;AAMA,eAAsB,uBAAsC;CAC1D,MAAM,UAAU,CAAC,GAAG,cAAc,OAAO,CAAC;CAC1C,cAAc,MAAM;CACpB,MAAM,QAAQ,WACZ,QAAQ,IAAI,OAAO,WAAW,MAAM,MAAM,QAAA,CAAS,SAAS,CAAC,CAC/D;AACF;AAEA,SAAS,YAAY,SAAwB;CAC3C,MAAM,IAAI,QACR,GAAG,QAAQ,4CACX,iBACA,QACF;AACF;AAEA,SAAS,eAAkD,OAAa;CACtE,OAAO,OAAO,YACZ,OAAO,QAAQ,KAAK,CAAC,CAAC,QAAQ,GAAG,WAAW,UAAU,KAAA,CAAS,CACjE;AACF;;;;AAKA,IAAa,iBAAb,MAAmD;CACjD;;CAEA;CACA;CACA,+BAAgC,IAAI,IAAyB;CAE7D,YAAY,SAAwB;EAClC,KAAK,gBAAgB,QAAQ,gBAAgB,QAAQ;EACrD,KAAK,UAAU,uBAAuB;GACpC,GAAG;GACH,cAAc,KAAK,iBAAA;EACrB,CAAC;CACH;CAMA,MAAc,cACZ,OACA,QAC2B;EAC3B,MAAM,WAAW,KAAK,QAAQ;EAC9B,IAAI,UAAU;GACZ,IAAI,OAAO,aAAa,YAAY,OAAO;GAC3C,KAAK,kBAAkB,SAAS,CAAC,CAAC,OAAO,UAAU;IACjD,KAAK,gBAAgB,KAAA;IACrB,MAAM;GACR,CAAC;GACD,OAAO,UAAU,KAAK,eAAe,MAAM;EAC7C;EAEA,IAAI,CAAC,kBAAkB,GACrB,MAAM,IAAI,uBAAuB,QAAQ;EAK3C,MAAM,QAAQ,KAAK,QAAQ,YAAY,KAAK,eAAe;EAC3D,IAAI,QAAQ,MAAM,IAAI,KAAK;EAC3B,IAAI,CAAC,OAAO;GACV,MAAM,4BAAY,IAAI,IAA0C;GAChE,MAAM,UAAuB;IAC3B;IACA,SAAS,KAAK,aAAa,OAAO,SAAS;GAC7C;GACA,QAAQ,QAAQ,YAAY;IAC1B,IAAI,MAAM,IAAI,KAAK,MAAM,SAAS,MAAM,OAAO,KAAK;GACtD,CAAC;GACD,MAAM,IAAI,OAAO,OAAO;GACxB,QAAQ;EACV;EAEA,MAAM,aAAa,KAAK,QAAQ;EAChC,IAAI,YAAY,MAAM,UAAU,IAAI,UAAU;EAC9C,IAAI;GACF,OAAO,MAAM,UAAU,MAAM,SAAS,MAAM;EAC9C,UAAU;GACR,IAAI,YAAY,MAAM,UAAU,OAAO,UAAU;EACnD;CACF;CAEA,MAAc,aACZ,OACA,WAC2B;EAE3B,QAAO,MADc,aAAa,EAAA,CACpB,gBAAgB,OAAO;GACnC,GAAI,KAAK,QAAQ,YAAY,EAAE,WAAW,KAAK,QAAQ,UAAU,IAAI,CAAC;GACtE,uBAAuB,WAAW;IAChC,KAAK,MAAM,YAAY,WACrB,IAAI;KACF,SAAS,MAAM;IACjB,QAAQ,CAER;GAEJ;EACF,CAAC;CACH;CAMA,YAAoB,UAAkD;EACpE,OAAO,SAAS,KAAK,YAAY;GAC/B,IAAI,QAAQ,SAAS,UAAU,MAAM,QAAQ,QAAQ,OAAO,GAC1D,OAAO;IACL,MAAM;IACN,SAAS,QAAQ,QAAQ,KAAK,SAC5B,KAAK,SAAS,SACV;KAAE,MAAM;KAAQ,MAAM,KAAK;IAAK,IAChC;KAAE,MAAM;KAAa,WAAW,EAAE,KAAK,KAAK,UAAU,IAAI;IAAE,CAClE;IACA,GAAI,QAAQ,OAAO,EAAE,MAAM,QAAQ,KAAK,IAAI,CAAC;GAC/C;GAEF,MAAM,UAAU,mBAAmB,QAAQ,OAAO;GAClD,IAAI,QAAQ,SAAS,UAAU,QAAQ,SAAS,YAC9C,OAAO;IACL,MAAM;IACN;IACA,cAAc,QAAQ,gBAAgB,QAAQ,QAAQ;GACxD;GAEF,IAAI,QAAQ,SAAS,aACnB,OAAO,eAAe;IACpB,MAAM;IACN,SAAS,YAAY,QAAQ,YAAY,SAAS,OAAO;IACzD,YAAY,QAAQ,YAAY,KAAK,UAAU;KAC7C,IAAI,KAAK;KACT,MAAM,KAAK;KACX,UAAU;MACR,MAAM,KAAK,SAAS;MACpB,WAAW,KAAK,SAAS;KAC3B;IACF,EAAE;GACJ,CAAC;GAEH,OAAO;IACL,MAAM,QAAQ;IACd;IACA,GAAI,QAAQ,OAAO,EAAE,MAAM,QAAQ,KAAK,IAAI,CAAC;GAC/C;EACF,CAAC;CACH;CAEA,kBACE,SACqC;EACrC,IAAI,QAAQ,mBAAmB,KAAA,GAO7B,OAAO;GAAE,MAAM;GAAe,QAL5B,OAAO,QAAQ,mBAAmB,WAC9B,QAAQ,iBACR,KAAK,UAAU,QAAQ,cAAc;EAGN;EAEvC,OAAO,QAAQ,gBAAgB,SAAS,gBACpC,EAAE,MAAM,cAAc,IACtB,KAAA;CACN;CAEA,aACE,UACA,SACA,OACA,QACA,cACyB;EACzB,MAAM,QACJ,gBAAgB,QAAQ,OAAO,SAC3B,QAAQ,MAAM,KAAK,UAAU;GAC3B,MAAM;GACN,UAAU;IACR,MAAM,KAAK,SAAS;IACpB,aAAa,KAAK,SAAS;IAC3B,YAAY,KAAK,SAAS;GAC5B;EACF,EAAE,IACF,KAAA;EACN,OAAO,eAAe;GACpB;GACA,UAAU,KAAK,YAAY,QAAQ;GACnC,YAAY,QAAQ;GACpB,aAAa,QAAQ;GACrB,OAAO,QAAQ;GACf,GAAG,QAAQ;GACX,MAAM,QAAQ;GACd,mBAAmB,QAAQ;GAC3B,kBAAkB,QAAQ;GAC1B,MAAM,QAAQ;GACd;GACA,aAAa,QAAQ,QAAQ,aAAa,KAAA;GAC1C,iBAAiB,KAAK,kBAAkB,OAAO;GAC/C;GACA,gBAAgB,SAAS,EAAE,eAAe,KAAK,IAAI,KAAA;EACrD,CAAC;CACH;;CAGA,aACE,OACA,SACoB;EACpB,IAAI,CAAC,KAAK,QAAQ,QAAQ,OAAO;EACjC,OAAO,QAAQ,SAAS,KAAK;CAC/B;CAEA,SAAiB,OAAwC;EACvD,MAAM,MAAM;EAQZ,IAAI,CAAC,KAAK,OAAO,KAAA;EACjB,OAAO;GACL,cAAc,IAAI,iBAAiB;GACnC,kBAAkB,IAAI,qBAAqB;GAC3C,aAAa,IAAI,gBAAgB;EACnC;CACF;CAEA,gBAAwB,QAAmC;EACzD,QAAQ,QAAR;GACE,KAAK,UACH,OAAO;GACT,KAAK,cACH,OAAO;GACT,SACE,OAAO;EACX;CACF;;;;;;;CAQA,qBACE,QACA,OACM;EACN,IAAI,WAAW,SACb,MAAM,IAAI,QACR,iDACA,kBACA,UACA,OACA,IACF;CAEJ;CAEA,SAAiB,OAAgB,OAAyB;EACxD,IAAI,iBAAiB,SAAS,OAAO;EACrC,MAAM,OAAQ,OAAyC,QAAQ;EAC/D,MAAM,UAAU,iBAAiB,QAAQ,MAAM,UAAU,OAAO,KAAK;EACrE,IAAI,SAAS,6BAA6B,SAAS,uBACjD,OAAO,IAAI,uBAAuB,UAAU,OAAO;EAErD,IACE,SAAS,oCACT,SAAS,yCAET,OAAO,IAAI,mBAAmB,UAAU,KAAK;EAE/C,IACE,SAAS,wBACT,SAAS,iCACT,SAAS,6BACT,SAAS,uBAET,OAAO,IAAI,mBAAmB,SAAS,SAAS,QAAQ;EAE1D,MAAM,SAAS,IAAI,QACjB,SACA,SAAS,oBAAoB,uBAAuB,aACpD,UACA,KACF;EACA,OAAO,QAAQ;EACf,OAAO;CACT;;CAGA,WACE,UACA,SACA,OACqB;EACrB,IAAI,UAAU,WAAW,GACvB,OAAO,IAAI,QACT,8BAA8B,SAAS,QAAQ,KAC/C,cACA,UACA,KACF;EAEF,IAAI,QAAQ,QAAQ,SAClB,OAAO,IAAI,QACT,gCACA,cACA,UACA,KACF;CAGJ;;;;;;CAOA,MAAc,WACZ,QACA,QACqB;EACrB,MAAM,WAAW,aAAa,IAAI,MAAM,KAAK,QAAQ,QAAQ;EAC7D,IAAI,gBAA4B,CAAC;EACjC,MAAM,OAAO,IAAI,SAAe,YAAY;GAC1C,UAAU;EACZ,CAAC;EACD,aAAa,IACX,QACA,SAAS,WAAW,IAAI,CAC1B;EACA,IAAI;GACF,MAAM,UAAU,UAAU,MAAM;EAClC,SAAS,OAAO;GACd,QAAQ;GACR,MAAM;EACR;EACA,OAAO;CACT;;CAGA,cACE,QACA,QACY;EACZ,MAAM,kBAAkB;GACtB,IAAI;IACF,QAAa,QAAQ,OAAO,kBAAkB,CAAC,CAAC,CAAC,YAAY,CAAC,CAAC;GACjE,QAAQ,CAER;EACF;EACA,IAAI,OAAO,SAAS,UAAU;EAC9B,OAAO,iBAAiB,SAAS,WAAW,EAAE,MAAM,KAAK,CAAC;EAC1D,aAAa,OAAO,oBAAoB,SAAS,SAAS;CAC5D;CAMA,MAAM,KACJ,UACA,UAAuB,CAAC,GACH;EACrB,MAAM,YAAY,KAAK,IAAI;EAC3B,IAAI;EACJ,IAAI;EACJ,IAAI;EACJ,IAAI;EACJ,IAAI;EACJ,MAAM,QACJ,QAAQ,SAAS,KAAK,QAAQ,gBAAA;EAChC,IAAI;GACF,UAAU,qBAAqB,KAAK,SAAS,SAAS,UAAU,KAAK;GAGrE,MAAM,SAAS,MAAM,KAAK,cAAc,OAAO,QAAQ,MAAM;GAC7D,WAAW,uBAAuB,KAAK,SAAS,OAAO;GACvD,UAAU,MAAM,KAAK,WAAW,QAAQ,SAAS,MAAM;GACvD,SAAS,KAAK,cAAc,QAAQ,SAAS,MAAM;GAEnD,UAAU,OAAO,KAAK,YAAY,OAChC,KAAK,aACH,UACA,SACA,KAAK,aAAa,OAAO,OAAO,GAChC,OACA,IACF,CACF;GACA,MAAM,UAA4B;GAClC,IAAI,UAAU;GACd,MAAM,oBAAoB;IACxB,UAAU;GACZ;GACA,QAAQ,KAAK,aAAa,WAAW;GACrC,oBAAqB,UAAU,KAAA,IAAY;GAC3C,MAAM,WAAW,MAAM,UAAU,SAAS,SAAS,MAAM;GACzD,IAAI,SAAS,OAAO,SAClB,MACE,KAAK,WAAW,UAAU,SAAS,KAAK,KAAK,SAAS,OAAO;GAIjE,MAAM,SAAS,UAAU,UAAU;GACnC,KAAK,qBAAqB,QAAQ,eAAe,KAAK;GACtD,IAAI,CAAC,QACH,MAAM,IAAI,QACR,mCACA,cACA,QACF;GAEF,MAAM,QAAQ,KAAK,SAAS,SAAS,KAAK;GAC1C,UACE,KAAK,SACL,UACA,QACA,SAAS,SAAS,OAClB,OACA,WACA,QAAQ,SACV;GACA,MAAM,YACJ,OAAO,SAAS,YAAY,KACzB,UAGM;IACL,IAAI,KAAK;IACT,MAAM;IACN,UAAU;KACR,MAAM,KAAK,SAAS;KACpB,WAAW,KAAK,SAAS;IAC3B;GACF,EACF;GACF,OAAO;IACL,SAAS,OAAO,SAAS,WAAW;IACpC;IACA,OAAO,SAAS,SAAS;IACzB,cAAc,KAAK,gBAAgB,OAAO,aAAa;IACvD,GAAI,WAAW,SAAS,EAAE,UAAU,IAAI,CAAC;GAC3C;EACF,SAAS,OAAO;GACd,MACE,KAAK,WAAW,UAAU,SAAS,KAAK,KAAK,KAAK,SAAS,OAAO,KAAK;EAE3E,UAAU;GACR,SAAS;GACT,gBAAgB,SAAS,cAAc,CAAC;GACxC,UAAU,QAAQ;EACpB;CACF;CAEA,MAAM,SACJ,QACA,UAA6B,CAAC,GACT;EACrB,OAAO,KAAK,KAAK,CAAC;GAAE,MAAM;GAAQ,SAAS;EAAO,CAAC,GAAG;GACpD,OAAO,QAAQ;GACf,WAAW,QAAQ;GACnB,aAAa,QAAQ;GACrB,MAAM,QAAQ;GACd,GAAG,QAAQ;GACX,MAAM,QAAQ;GACd,QAAQ,QAAQ;GAChB,YAAY,QAAQ;GACpB,QAAQ,QAAQ;GAChB,SAAS,QAAQ;GACjB,WAAW,QAAQ;GACnB,WAAW,QAAQ;GACnB,kBAAkB,QAAQ;EAC5B,CAAC;CACH;CAEA,MAAM,QAAQ,MAAc,UAA0B,CAAC,GAAoB;EACzE,MAAM,WAAwB,CAC5B,GAAI,QAAQ,WAAW,CAAC,GACxB;GAAE,MAAM,QAAQ,QAAQ;GAAQ,SAAS;EAAK,CAChD;EAsBA,QAAO,MArBgB,KAAK,KAAK,UAAU;GACzC,OAAO,QAAQ;GACf,WAAW,QAAQ;GACnB,aAAa,QAAQ;GACrB,MAAM,QAAQ;GACd,MAAM,QAAQ;GACd,QAAQ,QAAQ;GAChB,kBAAkB,QAAQ;GAC1B,iBAAiB,QAAQ;GACzB,gBAAgB,QAAQ;GACxB,gBAAgB,QAAQ;GACxB,MAAM,QAAQ;GACd,OAAO,QAAQ;GACf,YAAY,QAAQ;GACpB,YAAY,QAAQ;GACpB,QAAQ,QAAQ;GAChB,SAAS,QAAQ;GACjB,WAAW,QAAQ;GACnB,WAAW,QAAQ;GACnB,kBAAkB,QAAQ;EAC5B,CAAC,EAAA,CACe;CAClB;;;;;;CAOA,OAAO,OACL,UACA,UAAuB,CAAC,GACD;EACvB,MAAM,YAAY,KAAK,IAAI;EAC3B,IAAI;EACJ,IAAI;EACJ,IAAI;EACJ,IAAI;EACJ,IAAI;EACJ,IAAI,eAAe;EACnB,IAAI;EACJ,MAAM,QACJ,QAAQ,SAAS,KAAK,QAAQ,gBAAA;EAChC,IAAI;GACF,UAAU,qBAAqB,KAAK,SAAS,SAAS,UAAU,KAAK;GACrE,SAAS,MAAM,KAAK,cAAc,OAAO,QAAQ,MAAM;GACvD,WAAW,uBAAuB,KAAK,SAAS,OAAO;GACvD,UAAU,MAAM,KAAK,WAAW,QAAQ,SAAS,MAAM;GACvD,SAAS,KAAK,cAAc,QAAQ,SAAS,MAAM;GAEnD,MAAM,UACJ,OAAO,KAAK,YAAY,OACtB,KAAK,aACH,UACA,SACA,KAAK,aAAa,OAAO,OAAO,GAChC,MACA,KACF,CACF;GACF,IAAI;GACJ,IAAI;IACF,SAAS,MAAM,UAAU,SAAS,SAAS,MAAM;GACnD,SAAS,OAAO;IAId,cAAc,QAAQ,MACnB,SAAS,cAAc,KAAK,OAAO,cAAc,CAAC,GAAG,MAAM,SACtD,CAAC,CACT;IACA,MAAM;GACR;GAGA,WAAW,OAAO,OAAO,cAAc,CAAC;GAExC,IAAI;GACJ,IAAI;GACJ,SAAS;IACP,MAAM,OAAO,MAAM,SAAS,KAAK;IACjC,IAAI,KAAK,MAAM;KACb,eAAe;KACf;IACF;IACA,MAAM,QAAQ,KAAK;IACnB,MAAM,SAAS,MAAM,UAAU;IAC/B,eAAe,QAAQ,iBAAiB;IACxC,IAAI,MAAM,OAAO,QAAQ,KAAK,SAAS,MAAM,KAAK;IAClD,MAAM,UAAU,QAAQ,OAAO;IAC/B,IAAI,SAAS;KACX,QAAQ,aAAa,OAAO;KAC5B,MAAM;IACR;GACF;GACA,IAAI,SAAS,OAAO,SAClB,MACE,KAAK,WAAW,UAAU,SAAS,KAAK,KAAK,SAAS,OAAO;GAGjE,KAAK,qBAAqB,cAAc,KAAK;GAC7C,QAAQ,iBAAiB,KAAK,gBAAgB,YAAY,CAAC;GAC3D,UACE,KAAK,SACL,UACA,UACA,OACA,OACA,WACA,QAAQ,SACV;EACF,SAAS,OAAO;GACd,MACE,KAAK,WAAW,UAAU,SAAS,KAAK,KAAK,KAAK,SAAS,OAAO,KAAK;EAE3E,UAAU;GACR,IAAI,YAAY,CAAC,gBAAgB,QAAQ;IAIvC,IAAI;KACF,QAAa,QAAQ,OAAO,kBAAkB,CAAC,CAAC,CAAC,YAAY,CAAC,CAAC;IACjE,QAAQ,CAER;IACA,MAAM,cAAc,QAAQ;GAC9B;GACA,SAAS;GACT,gBAAgB,SAAS,WAAW;GACpC,UAAU,QAAQ;EACpB;CACF;CAEA,MAAM,YAAY,MAA+B;EAE/C,OAAO,KAAK,KAAK,KAAK,SAAS,CAAC;CAClC;CAMA,MAAM,YAAgC;EACpC,MAAM,SAAS,MAAM,aAAa;EAClC,MAAM,OAEF,KAAK,QAAQ,WAGZ,cAAc,OAAO,kBAAkB;EAC5C,MAAM,kBAAkB,IAAI,IAAI,OAAO,2BAA2B,CAAC,CAAC;EACpE,OAAO,KACJ,QAAQ,WAAW,OAAO,eAAe,CAAC,CAAC,CAC3C,KAAK,WAAW;GACf,MAAM,SAAS,OAAO,eAAe;GACrC,MAAM,YAAY,gBAAgB,IAAI,OAAO,QAAQ;GACrD,OAAO;IACL,IAAI,OAAO;IACX,MAAM,OAAO;IACb,aAAa,OAAO,mBAChB,+BAA+B,KAAK,MAAM,OAAO,gBAAgB,EAAE,aACnE;IACJ,eACE,OAAO,WAAW,uBAAuB;IAC3C,cAAc;KACZ;KACA;KACA,GAAI,YAAY,CAAC,WAAW,IAAI,CAAC;KACjC,GAAI,SAAS,CAAC,QAAQ,IAAI,CAAC;IAC7B;IACA,mBAAmB;IACnB,gBAAgB;GAClB;EACF,CAAC;CACL;CAEA,MAAM,kBAA2C;EAC/C,OAAO,EAAE,GAAG,oBAAoB;CAClC;CAMA,MAAM,MACJ,OACA,WAA6B,CAAC,GACF;EAC5B,OAAO,YAAY,YAAY;CACjC;CAEA,MAAM,WACJ,QACA,WAAkC,CAAC,GACP;EAC5B,OAAO,YAAY,kBAAkB;CACvC;CAEA,MAAM,cACJ,QACA,SACA,WAAoC,CAAC,GACpB;EACjB,OAAO,YAAY,mBAAmB;CACxC;CAEA,MAAM,cACJ,SACA,WAAmC,CAAC,GACF;EAClC,OAAO,YAAY,kBAAkB;CACvC;CAEA,MAAM,yBACJ,UAC6B;EAC7B,OAAO,YAAY,kBAAkB;CACvC;CAEA,MAAM,sBACJ,SACsC;EACtC,OAAO,YAAY,kBAAkB;CACvC;CAEA,MAAM,2BACJ,SACgC;EAChC,OAAO,YAAY,kBAAkB;CACvC;CAEA,MAAM,yBAAyB,SAA4C;EACzE,OAAO,YAAY,kBAAkB;CACvC;CAEA,MAAM,gCAAkD;EACtD,OAAO,YAAY,kBAAkB;CACvC;CAEA,MAAM,iBACJ,OACA,UACsB;EACtB,OAAO,YAAY,gBAAgB;CACrC;CAEA,aAAa,OAAe,UAA8C;EACxE,MAAM,QAAQ,IAAI,QAChB,qEACA,iBACA,QACF;EACA,OAAO,GACJ,OAAO,uBAAuB,EAC7B,YAAY,QAAQ,OAAO,KAAK,EAClC,GACF;CACF;CAEA,MAAM,WAAW,UAA6C;EAC5D,OAAO,YAAY,eAAe;CACpC;CAEA,MAAM,YAAY,UAA8C;EAC9D,OAAO,YAAY,cAAc;CACnC;CAEA,MAAM,UAAU,UAA+C;EAC7D,OAAO,YAAY,eAAe;CACpC;AACF;;AAYA,eAAe,cACb,UACe;CACf,IAAI;EACF,OAAO,EAAE,MAAM,SAAS,KAAK,EAAA,CAAG;CAGlC,QAAQ,CAER;AACF;;AAGA,SAAS,gBACP,SACA,MACM;CACN,IAAI,CAAC,SAAS;CACd,IAAI,MAAM,KAAU,KAAK,SAAS,OAAO;MACpC,QAAQ;AACf;;;;;;AAOA,eAAe,cACb,UACA,QACe;CACf,IAAI;EACF,KAAK,MAAM,SAAS,KAAK,EAAA,CAAG,MAAM;EAClC,MAAM,QAAQ,QAAQ,QAAQ,kBAAkB,CAAC;CACnD,QAAQ;EACN;CACF;CACA,MAAM,cAAc,QAAQ;AAC9B;;AAGA,SAAS,UAAa,SAAqB,QAAkC;CAC3E,IAAI,CAAC,QAAQ,OAAO;CACpB,IAAI,OAAO,SACT,OAAO,QAAQ,OACb,IAAI,QAAQ,gCAAgC,cAAc,QAAQ,CACpE;CAEF,OAAO,IAAI,SAAY,SAAS,WAAW;EACzC,MAAM,gBACJ,OACE,IAAI,QAAQ,gCAAgC,cAAc,QAAQ,CACpE;EACF,OAAO,iBAAiB,SAAS,SAAS,EAAE,MAAM,KAAK,CAAC;EACxD,QAAQ,MACL,UAAU;GACT,OAAO,oBAAoB,SAAS,OAAO;GAC3C,QAAQ,KAAK;EACf,IACC,UAAU;GACT,OAAO,oBAAoB,SAAS,OAAO;GAC3C,OAAO,KAAK;EACd,CACF;CACF,CAAC;AACH"}
+{"version":3,"file":"webllm-DrwR5V-y.js","names":[],"sources":["../../src/shared/providers/webllm.ts"],"sourcesContent":["// biome-ignore-all lint/style/useNamingConvention: WebLLM's OpenAI-shaped wire format and model records are snake_case.\n/**\n * WebLLM provider: runs a chat model entirely in the browser on WebGPU.\n *\n * Backed by `@mlc-ai/web-llm`, an optional peer dependency that is imported\n * lazily on first use, so neither the root entry nor `getAI()` for other\n * provider types ever loads it. Reachable through `getAI({ type: 'webllm' })`\n * and the `@happyvertical/ai/local` subpath.\n *\n * WebLLM's engine API is OpenAI-shaped, so requests and responses map almost\n * one to one. Structured output (`responseSchema`) becomes WebLLM's\n * grammar-constrained `response_format`, which keeps a small model from\n * emitting a value outside the schema.\n */\n\nimport {\n  normalizeBaseAIOptions,\n  normalizeChatOptions,\n  type PreparedRequestControls,\n  prepareRequestControls,\n} from '../safety';\nimport type {\n  AICapabilities,\n  AIInterface,\n  AIMessage,\n  AIModel,\n  AIResponse,\n  AIToolCall,\n  ChatOptions,\n  CompletionOptions,\n  EmbeddingOptions,\n  EmbeddingResponse,\n  ImageDescriptionOptions,\n  ImageEmbeddingOptions,\n  ImageGenerationOptions,\n  ImageGenerationResponse,\n  MessageOptions,\n  TokenUsage,\n  TTSOptions,\n  TTSResponse,\n  VideoGenerationJob,\n  VideoGenerationOptions,\n  VideoGenerationResult,\n  VideoGenerationStatusResult,\n  Voice,\n  VoiceCloneOptions,\n  VoiceDesignOptions,\n  VoiceListOptions,\n  WebLLMEngineLike,\n  WebLLMLoadProgress,\n  WebLLMOptions,\n} from '../types';\nimport {\n  AIError,\n  ContextLengthError,\n  extractTextContent,\n  ModelNotFoundError,\n  WebGPUUnavailableError,\n  WebLLMPeerMissingError,\n} from '../types';\nimport { emitUsage } from './usage';\n\nconst PROVIDER = 'webllm';\n\n/** Small enough for most GPUs (about 0.9 GB of VRAM); override with `defaultModel`. */\nexport const DEFAULT_WEBLLM_MODEL = 'Llama-3.2-1B-Instruct-q4f16_1-MLC';\n\n/** Context window assumed when a model record does not declare one. */\nconst DEFAULT_CONTEXT_LENGTH = 4096;\n\nconst WEBLLM_CAPABILITIES: AICapabilities = {\n  chat: true,\n  completion: true,\n  embeddings: false,\n  streaming: true,\n  functions: true,\n  vision: false,\n  fineTuning: false,\n  imageEmbeddings: false,\n  imageGeneration: false,\n  videoGeneration: false,\n  tts: false,\n  voiceCloning: false,\n  voiceDesign: false,\n  maxContextLength: DEFAULT_CONTEXT_LENGTH,\n  supportedOperations: ['chat', 'completion', 'streaming', 'functions'],\n};\n\n/** Structural subset of the `@mlc-ai/web-llm` module this provider uses. */\ninterface WebLLMModule {\n  CreateMLCEngine(\n    model: string,\n    config?: {\n      appConfig?: unknown;\n      initProgressCallback?: (report: WebLLMLoadProgress) => void;\n    },\n  ): Promise<WebLLMEngineLike>;\n  prebuiltAppConfig: { model_list: WebLLMModelRecord[] };\n  functionCallingModelIds?: string[];\n}\n\ninterface WebLLMModelRecord {\n  model_id: string;\n  vram_required_MB?: number;\n  low_resource_required?: boolean;\n  /** WebLLM `ModelType`: 0 or absent = LLM, 1 = embedding, 2 = VLM. */\n  model_type?: number;\n  overrides?: { context_window_size?: number };\n}\n\ninterface EngineEntry {\n  promise: Promise<WebLLMEngineLike>;\n  listeners: Set<(report: WebLLMLoadProgress) => void>;\n}\n\n/**\n * Per-engine request queues. WebLLM runs one generation at a time per engine\n * and `interruptGenerate()` stops whichever one is running, so requests are\n * serialized here and only the running request may interrupt.\n */\nconst engineQueues = new WeakMap<object, Promise<void>>();\n\n/** Engines shared by every provider instance that uses the prebuilt model list. */\nconst sharedEngines = new Map<string, EngineEntry>();\n\n/**\n * Imports the optional peer. Isolated so a missing install becomes a typed\n * {@link WebLLMPeerMissingError} instead of a bare module-resolution failure.\n */\nexport async function importWebLLM(): Promise<WebLLMModule> {\n  try {\n    const loaded = (await import('@mlc-ai/web-llm')) as unknown as\n      | (WebLLMModule & { default?: WebLLMModule })\n      | undefined;\n    const module =\n      typeof loaded?.CreateMLCEngine === 'function' ? loaded : loaded?.default;\n    if (typeof module?.CreateMLCEngine !== 'function') {\n      throw new Error('@mlc-ai/web-llm did not export CreateMLCEngine');\n    }\n    return module;\n  } catch (error) {\n    throw new WebLLMPeerMissingError(error);\n  }\n}\n\n/** True when this runtime exposes WebGPU (`navigator.gpu`). */\nexport function isWebGPUAvailable(): boolean {\n  const nav = (globalThis as { navigator?: { gpu?: unknown } }).navigator;\n  return Boolean(nav?.gpu);\n}\n\n/**\n * Unloads every engine this module created and forgets them. Caller-supplied\n * engines are never touched.\n */\nexport async function disposeWebLLMEngines(): Promise<void> {\n  const entries = [...sharedEngines.values()];\n  sharedEngines.clear();\n  await Promise.allSettled(\n    entries.map(async (entry) => (await entry.promise).unload?.()),\n  );\n}\n\nfunction unsupported(feature: string): never {\n  throw new AIError(\n    `${feature} is not supported by the webllm provider.`,\n    'NOT_SUPPORTED',\n    PROVIDER,\n  );\n}\n\nfunction stripUndefined<T extends Record<string, unknown>>(value: T): T {\n  return Object.fromEntries(\n    Object.entries(value).filter(([, entry]) => entry !== undefined),\n  ) as T;\n}\n\n/**\n * In-browser chat provider on WebLLM (WebGPU). See the module header.\n */\nexport class WebLLMProvider implements AIInterface {\n  private options: WebLLMOptions;\n  /** Model the caller named, if any; a supplied engine is loaded with it already. */\n  private readonly explicitModel?: string;\n  private factoryEngine?: Promise<WebLLMEngineLike>;\n  private readonly localEngines = new Map<string, EngineEntry>();\n\n  constructor(options: WebLLMOptions) {\n    this.explicitModel = options.defaultModel ?? options.model;\n    this.options = normalizeBaseAIOptions({\n      ...options,\n      defaultModel: this.explicitModel ?? DEFAULT_WEBLLM_MODEL,\n    });\n  }\n\n  // --------------------------------------------------------------------------\n  // Engine lifecycle\n  // --------------------------------------------------------------------------\n\n  private async acquireEngine(\n    model: string,\n    signal?: AbortSignal,\n  ): Promise<WebLLMEngineLike> {\n    const supplied = this.options.engine;\n    if (supplied) {\n      if (typeof supplied !== 'function') return supplied;\n      this.factoryEngine ??= supplied().catch((error) => {\n        this.factoryEngine = undefined;\n        throw error;\n      });\n      return abortable(this.factoryEngine, signal);\n    }\n\n    if (!isWebGPUAvailable()) {\n      throw new WebGPUUnavailableError(PROVIDER);\n    }\n\n    // A custom appConfig changes what a model id means, so it must not share\n    // an engine with other callers.\n    const cache = this.options.appConfig ? this.localEngines : sharedEngines;\n    let entry = cache.get(model);\n    if (!entry) {\n      const listeners = new Set<(report: WebLLMLoadProgress) => void>();\n      const created: EngineEntry = {\n        listeners,\n        promise: this.createEngine(model, listeners),\n      };\n      created.promise.catch(() => {\n        if (cache.get(model) === created) cache.delete(model);\n      });\n      cache.set(model, created);\n      entry = created;\n    }\n\n    const onProgress = this.options.onLoadProgress;\n    if (onProgress) entry.listeners.add(onProgress);\n    try {\n      return await abortable(entry.promise, signal);\n    } finally {\n      if (onProgress) entry.listeners.delete(onProgress);\n    }\n  }\n\n  private async createEngine(\n    model: string,\n    listeners: Set<(report: WebLLMLoadProgress) => void>,\n  ): Promise<WebLLMEngineLike> {\n    const webllm = await importWebLLM();\n    return webllm.CreateMLCEngine(model, {\n      ...(this.options.appConfig ? { appConfig: this.options.appConfig } : {}),\n      initProgressCallback: (report) => {\n        for (const listener of listeners) {\n          try {\n            listener(report);\n          } catch {\n            // A progress listener must never break model loading.\n          }\n        }\n      },\n    });\n  }\n\n  // --------------------------------------------------------------------------\n  // Request / response mapping\n  // --------------------------------------------------------------------------\n\n  private mapMessages(messages: AIMessage[]): Record<string, unknown>[] {\n    return messages.map((message) => {\n      if (message.role === 'user' && Array.isArray(message.content)) {\n        return {\n          role: 'user',\n          content: message.content.map((part) =>\n            part.type === 'text'\n              ? { type: 'text', text: part.text }\n              : { type: 'image_url', image_url: { url: part.image_url.url } },\n          ),\n          ...(message.name ? { name: message.name } : {}),\n        };\n      }\n      const content = extractTextContent(message.content);\n      if (message.role === 'tool' || message.role === 'function') {\n        return {\n          role: 'tool',\n          content,\n          tool_call_id: message.tool_call_id ?? message.name ?? '',\n        };\n      }\n      if (message.role === 'assistant') {\n        return stripUndefined({\n          role: 'assistant',\n          content: content || (message.tool_calls?.length ? null : ''),\n          tool_calls: message.tool_calls?.map((call) => ({\n            id: call.id,\n            type: call.type,\n            function: {\n              name: call.function.name,\n              arguments: call.function.arguments,\n            },\n          })),\n        });\n      }\n      return {\n        role: message.role,\n        content,\n        ...(message.name ? { name: message.name } : {}),\n      };\n    });\n  }\n\n  private mapResponseFormat(\n    options: Pick<ChatOptions, 'responseFormat' | 'responseSchema'>,\n  ): Record<string, unknown> | undefined {\n    if (options.responseSchema !== undefined) {\n      const schema =\n        typeof options.responseSchema === 'string'\n          ? options.responseSchema\n          : JSON.stringify(options.responseSchema);\n      // WebLLM compiles the schema to a grammar and masks logits, so the\n      // output cannot leave it.\n      return { type: 'json_object', schema };\n    }\n    return options.responseFormat?.type === 'json_object'\n      ? { type: 'json_object' }\n      : undefined;\n  }\n\n  private buildRequest(\n    messages: AIMessage[],\n    options: ChatOptions & { maxTokens: number },\n    model: string | undefined,\n    stream: boolean,\n    includeTools: boolean,\n  ): Record<string, unknown> {\n    const tools =\n      includeTools && options.tools?.length\n        ? options.tools.map((tool) => ({\n            type: 'function' as const,\n            function: {\n              name: tool.function.name,\n              description: tool.function.description,\n              parameters: tool.function.parameters,\n            },\n          }))\n        : undefined;\n    return stripUndefined({\n      model,\n      messages: this.mapMessages(messages),\n      max_tokens: options.maxTokens,\n      temperature: options.temperature,\n      top_p: options.topP,\n      n: options.n,\n      stop: options.stop,\n      frequency_penalty: options.frequencyPenalty,\n      presence_penalty: options.presencePenalty,\n      seed: options.seed,\n      tools,\n      tool_choice: tools ? options.toolChoice : undefined,\n      response_format: this.mapResponseFormat(options),\n      stream,\n      stream_options: stream ? { include_usage: true } : undefined,\n    });\n  }\n\n  /** Model id to send. A caller-supplied engine already has its model loaded. */\n  private requestModel(\n    model: string,\n    options: ChatOptions,\n  ): string | undefined {\n    if (!this.options.engine) return model;\n    return options.model ?? this.explicitModel;\n  }\n\n  private mapUsage(usage: unknown): TokenUsage | undefined {\n    const raw = usage as\n      | {\n          prompt_tokens?: number;\n          completion_tokens?: number;\n          total_tokens?: number;\n        }\n      | undefined\n      | null;\n    if (!raw) return undefined;\n    return {\n      promptTokens: raw.prompt_tokens ?? 0,\n      completionTokens: raw.completion_tokens ?? 0,\n      totalTokens: raw.total_tokens ?? 0,\n    };\n  }\n\n  private mapFinishReason(reason: string | null | undefined) {\n    switch (reason) {\n      case 'length':\n        return 'length' as const;\n      case 'tool_calls':\n        return 'tool_calls' as const;\n      default:\n        return 'stop' as const;\n    }\n  }\n\n  /**\n   * WebLLM reports `abort` when generation was interrupted. The provider's own\n   * aborts are raised before this point, so reaching it means someone else\n   * interrupted the engine (for example a caller sharing a supplied engine);\n   * a truncated reply must not look like a normal stop.\n   */\n  private assertNotInterrupted(\n    reason: string | null | undefined,\n    model: string,\n  ): void {\n    if (reason === 'abort') {\n      throw new AIError(\n        'Generation was interrupted before it finished',\n        'AI_INTERRUPTED',\n        PROVIDER,\n        model,\n        true,\n      );\n    }\n  }\n\n  private mapError(error: unknown, model?: string): AIError {\n    if (error instanceof AIError) return error;\n    const name = (error as { name?: string } | undefined)?.name ?? '';\n    const message = error instanceof Error ? error.message : String(error);\n    if (name === 'WebGPUNotAvailableError' || name === 'WebGPUNotFoundError') {\n      return new WebGPUUnavailableError(PROVIDER, message);\n    }\n    if (\n      name === 'ContextWindowSizeExceededError' ||\n      name === 'PrefillChunkSizeSmallerThanImageError'\n    ) {\n      return new ContextLengthError(PROVIDER, model);\n    }\n    if (\n      name === 'ModelNotFoundError' ||\n      name === 'SpecifiedModelNotFoundError' ||\n      name === 'UnsupportedModelIdError' ||\n      name === 'ModelNotLoadedError'\n    ) {\n      return new ModelNotFoundError(model ?? message, PROVIDER);\n    }\n    const mapped = new AIError(\n      message,\n      name === 'DeviceLostError' ? 'WEBGPU_DEVICE_LOST' : 'API_ERROR',\n      PROVIDER,\n      model,\n    );\n    mapped.cause = error;\n    return mapped;\n  }\n\n  /** Throws the right AI error when the request ended by timeout or caller abort. */\n  private abortError(\n    controls: PreparedRequestControls | undefined,\n    options: ChatOptions,\n    model?: string,\n  ): AIError | undefined {\n    if (controls?.didTimeout()) {\n      return new AIError(\n        `AI request timed out after ${controls.timeout}ms`,\n        'AI_TIMEOUT',\n        PROVIDER,\n        model,\n      );\n    }\n    if (options.signal?.aborted) {\n      return new AIError(\n        'AI request aborted by caller',\n        'AI_ABORTED',\n        PROVIDER,\n        model,\n      );\n    }\n    return undefined;\n  }\n\n  /**\n   * Waits for this request's turn on the engine; resolves to a release fn. A\n   * queued request that is aborted or times out leaves the queue at once and\n   * never interrupts the generation ahead of it.\n   */\n  private async lockEngine(\n    engine: WebLLMEngineLike,\n    signal: AbortSignal,\n  ): Promise<() => void> {\n    const previous = engineQueues.get(engine) ?? Promise.resolve();\n    let release: () => void = () => {};\n    const mine = new Promise<void>((resolve) => {\n      release = resolve;\n    });\n    engineQueues.set(\n      engine,\n      previous.then(() => mine),\n    );\n    try {\n      await abortable(previous, signal);\n    } catch (error) {\n      release();\n      throw error;\n    }\n    return release;\n  }\n\n  /** Interrupts generation when the composed signal fires; returns a detach fn. */\n  private bindInterrupt(\n    engine: WebLLMEngineLike,\n    signal: AbortSignal,\n  ): () => void {\n    const interrupt = () => {\n      try {\n        void Promise.resolve(engine.interruptGenerate()).catch(() => {});\n      } catch {\n        // Best effort: the request is already being abandoned.\n      }\n    };\n    if (signal.aborted) interrupt();\n    signal.addEventListener('abort', interrupt, { once: true });\n    return () => signal.removeEventListener('abort', interrupt);\n  }\n\n  // --------------------------------------------------------------------------\n  // Chat\n  // --------------------------------------------------------------------------\n\n  async chat(\n    messages: AIMessage[],\n    options: ChatOptions = {},\n  ): Promise<AIResponse> {\n    const startTime = Date.now();\n    let controls: PreparedRequestControls | undefined;\n    let detach: (() => void) | undefined;\n    let release: (() => void) | undefined;\n    let pending: Promise<any> | undefined;\n    let holdRelease: (() => Promise<unknown> | undefined) | undefined;\n    const model =\n      options.model || this.options.defaultModel || DEFAULT_WEBLLM_MODEL;\n    try {\n      options = normalizeChatOptions(this.options, options, PROVIDER, model);\n      // Load before arming the request timeout: a multi-GB download must not\n      // consume the generation budget.\n      const engine = await this.acquireEngine(model, options.signal);\n      controls = prepareRequestControls(this.options, options);\n      release = await this.lockEngine(engine, controls.signal);\n      detach = this.bindInterrupt(engine, controls.signal);\n\n      pending = engine.chat.completions.create(\n        this.buildRequest(\n          messages,\n          options as ChatOptions & { maxTokens: number },\n          this.requestModel(model, options),\n          false,\n          true,\n        ),\n      );\n      const started: Promise<unknown> = pending;\n      let settled = false;\n      const markSettled = () => {\n        settled = true;\n      };\n      started.then(markSettled, markSettled);\n      holdRelease = () => (settled ? undefined : started);\n      const response = await abortable(pending, controls.signal);\n      if (controls.signal.aborted) {\n        throw (\n          this.abortError(controls, options, model) ?? controls.signal.reason\n        );\n      }\n\n      const choice = response?.choices?.[0];\n      this.assertNotInterrupted(choice?.finish_reason, model);\n      if (!choice) {\n        throw new AIError(\n          'No choices returned from webllm',\n          'NO_CHOICES',\n          PROVIDER,\n        );\n      }\n      const usage = this.mapUsage(response.usage);\n      emitUsage(\n        this.options,\n        PROVIDER,\n        'chat',\n        response.model || model,\n        usage,\n        startTime,\n        options.usageTags,\n      );\n      const toolCalls: AIToolCall[] | undefined =\n        choice.message?.tool_calls?.map(\n          (call: {\n            id: string;\n            function: { name: string; arguments: string };\n          }) => ({\n            id: call.id,\n            type: 'function' as const,\n            function: {\n              name: call.function.name,\n              arguments: call.function.arguments,\n            },\n          }),\n        );\n      return {\n        content: choice.message?.content || '',\n        usage,\n        model: response.model || model,\n        finishReason: this.mapFinishReason(choice.finish_reason),\n        ...(toolCalls?.length ? { toolCalls } : {}),\n      };\n    } catch (error) {\n      throw (\n        this.abortError(controls, options, model) ?? this.mapError(error, model)\n      );\n    } finally {\n      detach?.();\n      releaseWhenIdle(release, holdRelease?.());\n      controls?.cleanup();\n    }\n  }\n\n  async complete(\n    prompt: string,\n    options: CompletionOptions = {},\n  ): Promise<AIResponse> {\n    return this.chat([{ role: 'user', content: prompt }], {\n      model: options.model,\n      maxTokens: options.maxTokens,\n      temperature: options.temperature,\n      topP: options.topP,\n      n: options.n,\n      stop: options.stop,\n      stream: options.stream,\n      onProgress: options.onProgress,\n      signal: options.signal,\n      timeout: options.timeout,\n      reasoning: options.reasoning,\n      usageTags: options.usageTags,\n      continueOnLength: options.continueOnLength,\n    });\n  }\n\n  async message(text: string, options: MessageOptions = {}): Promise<string> {\n    const messages: AIMessage[] = [\n      ...(options.history || []),\n      { role: options.role || 'user', content: text },\n    ];\n    const response = await this.chat(messages, {\n      model: options.model,\n      maxTokens: options.maxTokens,\n      temperature: options.temperature,\n      topP: options.topP,\n      stop: options.stop,\n      stream: options.stream,\n      frequencyPenalty: options.frequencyPenalty,\n      presencePenalty: options.presencePenalty,\n      responseFormat: options.responseFormat,\n      responseSchema: options.responseSchema,\n      seed: options.seed,\n      tools: options.tools,\n      toolChoice: options.toolChoice,\n      onProgress: options.onProgress,\n      signal: options.signal,\n      timeout: options.timeout,\n      reasoning: options.reasoning,\n      usageTags: options.usageTags,\n      continueOnLength: options.continueOnLength,\n    });\n    return response.content;\n  }\n\n  /**\n   * Streams a chat completion as text chunks. Like the other providers, a\n   * stream yields text only; use `chat()` for rounds that may call tools.\n   * Breaking out of the loop interrupts generation on the engine.\n   */\n  async *stream(\n    messages: AIMessage[],\n    options: ChatOptions = {},\n  ): AsyncIterable<string> {\n    const startTime = Date.now();\n    let controls: PreparedRequestControls | undefined;\n    let detach: (() => void) | undefined;\n    let engine: WebLLMEngineLike | undefined;\n    let release: (() => void) | undefined;\n    let iterator: AsyncIterator<WebLLMChunk> | undefined;\n    let iteratorDone = false;\n    let holdRelease: Promise<unknown> | undefined;\n    const model =\n      options.model || this.options.defaultModel || DEFAULT_WEBLLM_MODEL;\n    try {\n      options = normalizeChatOptions(this.options, options, PROVIDER, model);\n      engine = await this.acquireEngine(model, options.signal);\n      controls = prepareRequestControls(this.options, options);\n      release = await this.lockEngine(engine, controls.signal);\n      detach = this.bindInterrupt(engine, controls.signal);\n\n      const pending: Promise<AsyncIterable<WebLLMChunk>> =\n        engine.chat.completions.create(\n          this.buildRequest(\n            messages,\n            options as ChatOptions & { maxTokens: number },\n            this.requestModel(model, options),\n            true,\n            false,\n          ),\n        );\n      let chunks: AsyncIterable<WebLLMChunk>;\n      try {\n        chunks = await abortable(pending, controls.signal);\n      } catch (error) {\n        // The caller is gone but the engine may still start this request.\n        // Keep our queue closed until it has been interrupted and drained, or\n        // the next request would wait behind it and could interrupt it.\n        holdRelease = pending.then(\n          (late) => abandonStream(late[Symbol.asyncIterator](), engine),\n          () => {},\n        );\n        throw error;\n      }\n      // Iterated by hand: WebLLM releases its per-model lock only when its\n      // generator runs to completion, never on an early `return()`.\n      iterator = chunks[Symbol.asyncIterator]();\n\n      let finishReason: string | null | undefined;\n      let usage: TokenUsage | undefined;\n      for (;;) {\n        const step = await iterator.next();\n        if (step.done) {\n          iteratorDone = true;\n          break;\n        }\n        const chunk = step.value;\n        const choice = chunk.choices?.[0];\n        finishReason = choice?.finish_reason ?? finishReason;\n        if (chunk.usage) usage = this.mapUsage(chunk.usage);\n        const content = choice?.delta?.content;\n        if (content) {\n          options.onProgress?.(content);\n          yield content;\n        }\n      }\n      if (controls.signal.aborted) {\n        throw (\n          this.abortError(controls, options, model) ?? controls.signal.reason\n        );\n      }\n      this.assertNotInterrupted(finishReason, model);\n      options.onFinishReason?.(this.mapFinishReason(finishReason));\n      emitUsage(\n        this.options,\n        PROVIDER,\n        'stream',\n        model,\n        usage,\n        startTime,\n        options.usageTags,\n      );\n    } catch (error) {\n      throw (\n        this.abortError(controls, options, model) ?? this.mapError(error, model)\n      );\n    } finally {\n      if (iterator && !iteratorDone && engine) {\n        // Consumer stopped early (or the loop threw): stop the GPU, then let\n        // the engine's generator finish so its lock is released before the\n        // next request is admitted.\n        try {\n          void Promise.resolve(engine.interruptGenerate()).catch(() => {});\n        } catch {\n          // Best effort.\n        }\n        await drainIterator(iterator);\n      }\n      detach?.();\n      releaseWhenIdle(release, holdRelease);\n      controls?.cleanup();\n    }\n  }\n\n  async countTokens(text: string): Promise<number> {\n    // WebLLM exposes no standalone tokenizer; ~4 characters per token.\n    return Math.ceil(text.length / 4);\n  }\n\n  // --------------------------------------------------------------------------\n  // Models and capabilities\n  // --------------------------------------------------------------------------\n\n  async getModels(): Promise<AIModel[]> {\n    const webllm = await importWebLLM();\n    const list =\n      (\n        this.options.appConfig as\n          | { model_list?: WebLLMModelRecord[] }\n          | undefined\n      )?.model_list ?? webllm.prebuiltAppConfig.model_list;\n    const functionCalling = new Set(webllm.functionCallingModelIds ?? []);\n    return list\n      .filter((record) => record.model_type !== 1) // embedding models\n      .map((record) => {\n        const vision = record.model_type === 2;\n        const functions = functionCalling.has(record.model_id);\n        return {\n          id: record.model_id,\n          name: record.model_id,\n          description: record.vram_required_MB\n            ? `Runs on-device via WebGPU (~${Math.round(record.vram_required_MB)} MB VRAM)`\n            : 'Runs on-device via WebGPU',\n          contextLength:\n            record.overrides?.context_window_size ?? DEFAULT_CONTEXT_LENGTH,\n          capabilities: [\n            'text',\n            'chat',\n            ...(functions ? ['functions'] : []),\n            ...(vision ? ['vision'] : []),\n          ],\n          supportsFunctions: functions,\n          supportsVision: vision,\n        };\n      });\n  }\n\n  async getCapabilities(): Promise<AICapabilities> {\n    return { ...WEBLLM_CAPABILITIES };\n  }\n\n  // --------------------------------------------------------------------------\n  // No in-browser equivalent\n  // --------------------------------------------------------------------------\n\n  async embed(\n    _text: string | string[],\n    _options: EmbeddingOptions = {},\n  ): Promise<EmbeddingResponse> {\n    return unsupported('Embeddings');\n  }\n\n  async embedImage(\n    _image: string | Buffer,\n    _options: ImageEmbeddingOptions = {},\n  ): Promise<EmbeddingResponse> {\n    return unsupported('Image embeddings');\n  }\n\n  async describeImage(\n    _image: string | Buffer,\n    _prompt?: string,\n    _options: ImageDescriptionOptions = {},\n  ): Promise<string> {\n    return unsupported('Image description');\n  }\n\n  async generateImage(\n    _prompt: string,\n    _options: ImageGenerationOptions = {},\n  ): Promise<ImageGenerationResponse> {\n    return unsupported('Image generation');\n  }\n\n  async submitVideoGenerationJob(\n    _options: VideoGenerationOptions,\n  ): Promise<VideoGenerationJob> {\n    return unsupported('Video generation');\n  }\n\n  async getVideoGenerationJob(\n    _handle: VideoGenerationJob,\n  ): Promise<VideoGenerationStatusResult> {\n    return unsupported('Video generation');\n  }\n\n  async fetchVideoGenerationResult(\n    _handle: VideoGenerationJob,\n  ): Promise<VideoGenerationResult> {\n    return unsupported('Video generation');\n  }\n\n  async cancelVideoGenerationJob(_handle: VideoGenerationJob): Promise<void> {\n    return unsupported('Video generation');\n  }\n\n  async validateVideoGenerationAccess(): Promise<boolean> {\n    return unsupported('Video generation');\n  }\n\n  async synthesizeSpeech(\n    _text: string,\n    _options?: TTSOptions,\n  ): Promise<TTSResponse> {\n    return unsupported('Text-to-speech');\n  }\n\n  streamSpeech(_text: string, _options?: TTSOptions): AsyncIterable<Buffer> {\n    const error = new AIError(\n      'Text-to-speech streaming is not supported by the webllm provider.',\n      'NOT_SUPPORTED',\n      PROVIDER,\n    );\n    return {\n      [Symbol.asyncIterator]: () => ({\n        next: () => Promise.reject(error),\n      }),\n    };\n  }\n\n  async cloneVoice(_options: VoiceCloneOptions): Promise<Voice> {\n    return unsupported('Voice cloning');\n  }\n\n  async designVoice(_options: VoiceDesignOptions): Promise<Voice> {\n    return unsupported('Voice design');\n  }\n\n  async getVoices(_options?: VoiceListOptions): Promise<Voice[]> {\n    return unsupported('Voice listing');\n  }\n}\n\n/** One streamed WebLLM chunk (structural subset). */\ninterface WebLLMChunk {\n  choices?: Array<{\n    delta?: { content?: string | null };\n    finish_reason?: string | null;\n  }>;\n  usage?: unknown;\n}\n\n/** Runs a WebLLM stream to its end, ignoring errors; used after interrupting it. */\nasync function drainIterator(\n  iterator: AsyncIterator<WebLLMChunk>,\n): Promise<void> {\n  try {\n    while (!(await iterator.next()).done) {\n      // discard\n    }\n  } catch {\n    // The engine already stopped; nothing more to release.\n  }\n}\n\n/** Releases an engine queue slot now, or once `idle` has settled. */\nfunction releaseWhenIdle(\n  release: (() => void) | undefined,\n  idle?: Promise<unknown>,\n): void {\n  if (!release) return;\n  if (idle) void idle.then(release, release);\n  else release();\n}\n\n/**\n * Finishes a stream nobody is reading. WebLLM resets its interrupt flag on the\n * first `next()`, so interrupt only after pulling one chunk, then drain it so\n * the engine's lock is released.\n */\nasync function abandonStream(\n  iterator: AsyncIterator<WebLLMChunk>,\n  engine: WebLLMEngineLike | undefined,\n): Promise<void> {\n  try {\n    if ((await iterator.next()).done) return;\n    await Promise.resolve(engine?.interruptGenerate());\n  } catch {\n    return;\n  }\n  await drainIterator(iterator);\n}\n\n/** Rejects with the signal's reason if it fires first; the underlying work continues. */\nfunction abortable<T>(promise: Promise<T>, signal?: AbortSignal): Promise<T> {\n  if (!signal) return promise;\n  if (signal.aborted) {\n    return Promise.reject(\n      new AIError('AI request aborted by caller', 'AI_ABORTED', PROVIDER),\n    );\n  }\n  return new Promise<T>((resolve, reject) => {\n    const onAbort = () =>\n      reject(\n        new AIError('AI request aborted by caller', 'AI_ABORTED', PROVIDER),\n      );\n    signal.addEventListener('abort', onAbort, { once: true });\n    promise.then(\n      (value) => {\n        signal.removeEventListener('abort', onAbort);\n        resolve(value);\n      },\n      (error) => {\n        signal.removeEventListener('abort', onAbort);\n        reject(error);\n      },\n    );\n  });\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;AA8DA,IAAM,WAAW;;AAGjB,IAAa,uBAAuB;;AAGpC,IAAM,yBAAyB;AAE/B,IAAM,sBAAsC;CAC1C,MAAM;CACN,YAAY;CACZ,YAAY;CACZ,WAAW;CACX,WAAW;CACX,QAAQ;CACR,YAAY;CACZ,iBAAiB;CACjB,iBAAiB;CACjB,iBAAiB;CACjB,KAAK;CACL,cAAc;CACd,aAAa;CACb,kBAAkB;CAClB,qBAAqB;EAAC;EAAQ;EAAc;EAAa;CAAW;AACtE;;;;;;AAkCA,IAAM,+BAAe,IAAI,QAA+B;;AAGxD,IAAM,gCAAgB,IAAI,IAAyB;;;;;AAMnD,eAAsB,eAAsC;CAC1D,IAAI;EACF,MAAM,SAAU,MAAM,OAAO;EAG7B,MAAM,SACJ,OAAO,QAAQ,oBAAoB,aAAa,SAAS,QAAQ;EACnE,IAAI,OAAO,QAAQ,oBAAoB,YACrC,MAAM,IAAI,MAAM,gDAAgD;EAElE,OAAO;CACT,SAAS,OAAO;EACd,MAAM,IAAI,uBAAuB,KAAK;CACxC;AACF;;AAGA,SAAgB,oBAA6B;CAC3C,MAAM,MAAO,WAAiD;CAC9D,OAAO,QAAQ,KAAK,GAAG;AACzB;;;;;AAMA,eAAsB,uBAAsC;CAC1D,MAAM,UAAU,CAAC,GAAG,cAAc,OAAO,CAAC;CAC1C,cAAc,MAAM;CACpB,MAAM,QAAQ,WACZ,QAAQ,IAAI,OAAO,WAAW,MAAM,MAAM,QAAA,CAAS,SAAS,CAAC,CAC/D;AACF;AAEA,SAAS,YAAY,SAAwB;CAC3C,MAAM,IAAI,QACR,GAAG,QAAQ,4CACX,iBACA,QACF;AACF;AAEA,SAAS,eAAkD,OAAa;CACtE,OAAO,OAAO,YACZ,OAAO,QAAQ,KAAK,CAAC,CAAC,QAAQ,GAAG,WAAW,UAAU,KAAA,CAAS,CACjE;AACF;;;;AAKA,IAAa,iBAAb,MAAmD;CACjD;;CAEA;CACA;CACA,+BAAgC,IAAI,IAAyB;CAE7D,YAAY,SAAwB;EAClC,KAAK,gBAAgB,QAAQ,gBAAgB,QAAQ;EACrD,KAAK,UAAU,uBAAuB;GACpC,GAAG;GACH,cAAc,KAAK,iBAAA;EACrB,CAAC;CACH;CAMA,MAAc,cACZ,OACA,QAC2B;EAC3B,MAAM,WAAW,KAAK,QAAQ;EAC9B,IAAI,UAAU;GACZ,IAAI,OAAO,aAAa,YAAY,OAAO;GAC3C,KAAK,kBAAkB,SAAS,CAAC,CAAC,OAAO,UAAU;IACjD,KAAK,gBAAgB,KAAA;IACrB,MAAM;GACR,CAAC;GACD,OAAO,UAAU,KAAK,eAAe,MAAM;EAC7C;EAEA,IAAI,CAAC,kBAAkB,GACrB,MAAM,IAAI,uBAAuB,QAAQ;EAK3C,MAAM,QAAQ,KAAK,QAAQ,YAAY,KAAK,eAAe;EAC3D,IAAI,QAAQ,MAAM,IAAI,KAAK;EAC3B,IAAI,CAAC,OAAO;GACV,MAAM,4BAAY,IAAI,IAA0C;GAChE,MAAM,UAAuB;IAC3B;IACA,SAAS,KAAK,aAAa,OAAO,SAAS;GAC7C;GACA,QAAQ,QAAQ,YAAY;IAC1B,IAAI,MAAM,IAAI,KAAK,MAAM,SAAS,MAAM,OAAO,KAAK;GACtD,CAAC;GACD,MAAM,IAAI,OAAO,OAAO;GACxB,QAAQ;EACV;EAEA,MAAM,aAAa,KAAK,QAAQ;EAChC,IAAI,YAAY,MAAM,UAAU,IAAI,UAAU;EAC9C,IAAI;GACF,OAAO,MAAM,UAAU,MAAM,SAAS,MAAM;EAC9C,UAAU;GACR,IAAI,YAAY,MAAM,UAAU,OAAO,UAAU;EACnD;CACF;CAEA,MAAc,aACZ,OACA,WAC2B;EAE3B,QAAO,MADc,aAAa,EAAA,CACpB,gBAAgB,OAAO;GACnC,GAAI,KAAK,QAAQ,YAAY,EAAE,WAAW,KAAK,QAAQ,UAAU,IAAI,CAAC;GACtE,uBAAuB,WAAW;IAChC,KAAK,MAAM,YAAY,WACrB,IAAI;KACF,SAAS,MAAM;IACjB,QAAQ,CAER;GAEJ;EACF,CAAC;CACH;CAMA,YAAoB,UAAkD;EACpE,OAAO,SAAS,KAAK,YAAY;GAC/B,IAAI,QAAQ,SAAS,UAAU,MAAM,QAAQ,QAAQ,OAAO,GAC1D,OAAO;IACL,MAAM;IACN,SAAS,QAAQ,QAAQ,KAAK,SAC5B,KAAK,SAAS,SACV;KAAE,MAAM;KAAQ,MAAM,KAAK;IAAK,IAChC;KAAE,MAAM;KAAa,WAAW,EAAE,KAAK,KAAK,UAAU,IAAI;IAAE,CAClE;IACA,GAAI,QAAQ,OAAO,EAAE,MAAM,QAAQ,KAAK,IAAI,CAAC;GAC/C;GAEF,MAAM,UAAU,mBAAmB,QAAQ,OAAO;GAClD,IAAI,QAAQ,SAAS,UAAU,QAAQ,SAAS,YAC9C,OAAO;IACL,MAAM;IACN;IACA,cAAc,QAAQ,gBAAgB,QAAQ,QAAQ;GACxD;GAEF,IAAI,QAAQ,SAAS,aACnB,OAAO,eAAe;IACpB,MAAM;IACN,SAAS,YAAY,QAAQ,YAAY,SAAS,OAAO;IACzD,YAAY,QAAQ,YAAY,KAAK,UAAU;KAC7C,IAAI,KAAK;KACT,MAAM,KAAK;KACX,UAAU;MACR,MAAM,KAAK,SAAS;MACpB,WAAW,KAAK,SAAS;KAC3B;IACF,EAAE;GACJ,CAAC;GAEH,OAAO;IACL,MAAM,QAAQ;IACd;IACA,GAAI,QAAQ,OAAO,EAAE,MAAM,QAAQ,KAAK,IAAI,CAAC;GAC/C;EACF,CAAC;CACH;CAEA,kBACE,SACqC;EACrC,IAAI,QAAQ,mBAAmB,KAAA,GAO7B,OAAO;GAAE,MAAM;GAAe,QAL5B,OAAO,QAAQ,mBAAmB,WAC9B,QAAQ,iBACR,KAAK,UAAU,QAAQ,cAAc;EAGN;EAEvC,OAAO,QAAQ,gBAAgB,SAAS,gBACpC,EAAE,MAAM,cAAc,IACtB,KAAA;CACN;CAEA,aACE,UACA,SACA,OACA,QACA,cACyB;EACzB,MAAM,QACJ,gBAAgB,QAAQ,OAAO,SAC3B,QAAQ,MAAM,KAAK,UAAU;GAC3B,MAAM;GACN,UAAU;IACR,MAAM,KAAK,SAAS;IACpB,aAAa,KAAK,SAAS;IAC3B,YAAY,KAAK,SAAS;GAC5B;EACF,EAAE,IACF,KAAA;EACN,OAAO,eAAe;GACpB;GACA,UAAU,KAAK,YAAY,QAAQ;GACnC,YAAY,QAAQ;GACpB,aAAa,QAAQ;GACrB,OAAO,QAAQ;GACf,GAAG,QAAQ;GACX,MAAM,QAAQ;GACd,mBAAmB,QAAQ;GAC3B,kBAAkB,QAAQ;GAC1B,MAAM,QAAQ;GACd;GACA,aAAa,QAAQ,QAAQ,aAAa,KAAA;GAC1C,iBAAiB,KAAK,kBAAkB,OAAO;GAC/C;GACA,gBAAgB,SAAS,EAAE,eAAe,KAAK,IAAI,KAAA;EACrD,CAAC;CACH;;CAGA,aACE,OACA,SACoB;EACpB,IAAI,CAAC,KAAK,QAAQ,QAAQ,OAAO;EACjC,OAAO,QAAQ,SAAS,KAAK;CAC/B;CAEA,SAAiB,OAAwC;EACvD,MAAM,MAAM;EAQZ,IAAI,CAAC,KAAK,OAAO,KAAA;EACjB,OAAO;GACL,cAAc,IAAI,iBAAiB;GACnC,kBAAkB,IAAI,qBAAqB;GAC3C,aAAa,IAAI,gBAAgB;EACnC;CACF;CAEA,gBAAwB,QAAmC;EACzD,QAAQ,QAAR;GACE,KAAK,UACH,OAAO;GACT,KAAK,cACH,OAAO;GACT,SACE,OAAO;EACX;CACF;;;;;;;CAQA,qBACE,QACA,OACM;EACN,IAAI,WAAW,SACb,MAAM,IAAI,QACR,iDACA,kBACA,UACA,OACA,IACF;CAEJ;CAEA,SAAiB,OAAgB,OAAyB;EACxD,IAAI,iBAAiB,SAAS,OAAO;EACrC,MAAM,OAAQ,OAAyC,QAAQ;EAC/D,MAAM,UAAU,iBAAiB,QAAQ,MAAM,UAAU,OAAO,KAAK;EACrE,IAAI,SAAS,6BAA6B,SAAS,uBACjD,OAAO,IAAI,uBAAuB,UAAU,OAAO;EAErD,IACE,SAAS,oCACT,SAAS,yCAET,OAAO,IAAI,mBAAmB,UAAU,KAAK;EAE/C,IACE,SAAS,wBACT,SAAS,iCACT,SAAS,6BACT,SAAS,uBAET,OAAO,IAAI,mBAAmB,SAAS,SAAS,QAAQ;EAE1D,MAAM,SAAS,IAAI,QACjB,SACA,SAAS,oBAAoB,uBAAuB,aACpD,UACA,KACF;EACA,OAAO,QAAQ;EACf,OAAO;CACT;;CAGA,WACE,UACA,SACA,OACqB;EACrB,IAAI,UAAU,WAAW,GACvB,OAAO,IAAI,QACT,8BAA8B,SAAS,QAAQ,KAC/C,cACA,UACA,KACF;EAEF,IAAI,QAAQ,QAAQ,SAClB,OAAO,IAAI,QACT,gCACA,cACA,UACA,KACF;CAGJ;;;;;;CAOA,MAAc,WACZ,QACA,QACqB;EACrB,MAAM,WAAW,aAAa,IAAI,MAAM,KAAK,QAAQ,QAAQ;EAC7D,IAAI,gBAA4B,CAAC;EACjC,MAAM,OAAO,IAAI,SAAe,YAAY;GAC1C,UAAU;EACZ,CAAC;EACD,aAAa,IACX,QACA,SAAS,WAAW,IAAI,CAC1B;EACA,IAAI;GACF,MAAM,UAAU,UAAU,MAAM;EAClC,SAAS,OAAO;GACd,QAAQ;GACR,MAAM;EACR;EACA,OAAO;CACT;;CAGA,cACE,QACA,QACY;EACZ,MAAM,kBAAkB;GACtB,IAAI;IACF,QAAa,QAAQ,OAAO,kBAAkB,CAAC,CAAC,CAAC,YAAY,CAAC,CAAC;GACjE,QAAQ,CAER;EACF;EACA,IAAI,OAAO,SAAS,UAAU;EAC9B,OAAO,iBAAiB,SAAS,WAAW,EAAE,MAAM,KAAK,CAAC;EAC1D,aAAa,OAAO,oBAAoB,SAAS,SAAS;CAC5D;CAMA,MAAM,KACJ,UACA,UAAuB,CAAC,GACH;EACrB,MAAM,YAAY,KAAK,IAAI;EAC3B,IAAI;EACJ,IAAI;EACJ,IAAI;EACJ,IAAI;EACJ,IAAI;EACJ,MAAM,QACJ,QAAQ,SAAS,KAAK,QAAQ,gBAAA;EAChC,IAAI;GACF,UAAU,qBAAqB,KAAK,SAAS,SAAS,UAAU,KAAK;GAGrE,MAAM,SAAS,MAAM,KAAK,cAAc,OAAO,QAAQ,MAAM;GAC7D,WAAW,uBAAuB,KAAK,SAAS,OAAO;GACvD,UAAU,MAAM,KAAK,WAAW,QAAQ,SAAS,MAAM;GACvD,SAAS,KAAK,cAAc,QAAQ,SAAS,MAAM;GAEnD,UAAU,OAAO,KAAK,YAAY,OAChC,KAAK,aACH,UACA,SACA,KAAK,aAAa,OAAO,OAAO,GAChC,OACA,IACF,CACF;GACA,MAAM,UAA4B;GAClC,IAAI,UAAU;GACd,MAAM,oBAAoB;IACxB,UAAU;GACZ;GACA,QAAQ,KAAK,aAAa,WAAW;GACrC,oBAAqB,UAAU,KAAA,IAAY;GAC3C,MAAM,WAAW,MAAM,UAAU,SAAS,SAAS,MAAM;GACzD,IAAI,SAAS,OAAO,SAClB,MACE,KAAK,WAAW,UAAU,SAAS,KAAK,KAAK,SAAS,OAAO;GAIjE,MAAM,SAAS,UAAU,UAAU;GACnC,KAAK,qBAAqB,QAAQ,eAAe,KAAK;GACtD,IAAI,CAAC,QACH,MAAM,IAAI,QACR,mCACA,cACA,QACF;GAEF,MAAM,QAAQ,KAAK,SAAS,SAAS,KAAK;GAC1C,UACE,KAAK,SACL,UACA,QACA,SAAS,SAAS,OAClB,OACA,WACA,QAAQ,SACV;GACA,MAAM,YACJ,OAAO,SAAS,YAAY,KACzB,UAGM;IACL,IAAI,KAAK;IACT,MAAM;IACN,UAAU;KACR,MAAM,KAAK,SAAS;KACpB,WAAW,KAAK,SAAS;IAC3B;GACF,EACF;GACF,OAAO;IACL,SAAS,OAAO,SAAS,WAAW;IACpC;IACA,OAAO,SAAS,SAAS;IACzB,cAAc,KAAK,gBAAgB,OAAO,aAAa;IACvD,GAAI,WAAW,SAAS,EAAE,UAAU,IAAI,CAAC;GAC3C;EACF,SAAS,OAAO;GACd,MACE,KAAK,WAAW,UAAU,SAAS,KAAK,KAAK,KAAK,SAAS,OAAO,KAAK;EAE3E,UAAU;GACR,SAAS;GACT,gBAAgB,SAAS,cAAc,CAAC;GACxC,UAAU,QAAQ;EACpB;CACF;CAEA,MAAM,SACJ,QACA,UAA6B,CAAC,GACT;EACrB,OAAO,KAAK,KAAK,CAAC;GAAE,MAAM;GAAQ,SAAS;EAAO,CAAC,GAAG;GACpD,OAAO,QAAQ;GACf,WAAW,QAAQ;GACnB,aAAa,QAAQ;GACrB,MAAM,QAAQ;GACd,GAAG,QAAQ;GACX,MAAM,QAAQ;GACd,QAAQ,QAAQ;GAChB,YAAY,QAAQ;GACpB,QAAQ,QAAQ;GAChB,SAAS,QAAQ;GACjB,WAAW,QAAQ;GACnB,WAAW,QAAQ;GACnB,kBAAkB,QAAQ;EAC5B,CAAC;CACH;CAEA,MAAM,QAAQ,MAAc,UAA0B,CAAC,GAAoB;EACzE,MAAM,WAAwB,CAC5B,GAAI,QAAQ,WAAW,CAAC,GACxB;GAAE,MAAM,QAAQ,QAAQ;GAAQ,SAAS;EAAK,CAChD;EAsBA,QAAO,MArBgB,KAAK,KAAK,UAAU;GACzC,OAAO,QAAQ;GACf,WAAW,QAAQ;GACnB,aAAa,QAAQ;GACrB,MAAM,QAAQ;GACd,MAAM,QAAQ;GACd,QAAQ,QAAQ;GAChB,kBAAkB,QAAQ;GAC1B,iBAAiB,QAAQ;GACzB,gBAAgB,QAAQ;GACxB,gBAAgB,QAAQ;GACxB,MAAM,QAAQ;GACd,OAAO,QAAQ;GACf,YAAY,QAAQ;GACpB,YAAY,QAAQ;GACpB,QAAQ,QAAQ;GAChB,SAAS,QAAQ;GACjB,WAAW,QAAQ;GACnB,WAAW,QAAQ;GACnB,kBAAkB,QAAQ;EAC5B,CAAC,EAAA,CACe;CAClB;;;;;;CAOA,OAAO,OACL,UACA,UAAuB,CAAC,GACD;EACvB,MAAM,YAAY,KAAK,IAAI;EAC3B,IAAI;EACJ,IAAI;EACJ,IAAI;EACJ,IAAI;EACJ,IAAI;EACJ,IAAI,eAAe;EACnB,IAAI;EACJ,MAAM,QACJ,QAAQ,SAAS,KAAK,QAAQ,gBAAA;EAChC,IAAI;GACF,UAAU,qBAAqB,KAAK,SAAS,SAAS,UAAU,KAAK;GACrE,SAAS,MAAM,KAAK,cAAc,OAAO,QAAQ,MAAM;GACvD,WAAW,uBAAuB,KAAK,SAAS,OAAO;GACvD,UAAU,MAAM,KAAK,WAAW,QAAQ,SAAS,MAAM;GACvD,SAAS,KAAK,cAAc,QAAQ,SAAS,MAAM;GAEnD,MAAM,UACJ,OAAO,KAAK,YAAY,OACtB,KAAK,aACH,UACA,SACA,KAAK,aAAa,OAAO,OAAO,GAChC,MACA,KACF,CACF;GACF,IAAI;GACJ,IAAI;IACF,SAAS,MAAM,UAAU,SAAS,SAAS,MAAM;GACnD,SAAS,OAAO;IAId,cAAc,QAAQ,MACnB,SAAS,cAAc,KAAK,OAAO,cAAc,CAAC,GAAG,MAAM,SACtD,CAAC,CACT;IACA,MAAM;GACR;GAGA,WAAW,OAAO,OAAO,cAAc,CAAC;GAExC,IAAI;GACJ,IAAI;GACJ,SAAS;IACP,MAAM,OAAO,MAAM,SAAS,KAAK;IACjC,IAAI,KAAK,MAAM;KACb,eAAe;KACf;IACF;IACA,MAAM,QAAQ,KAAK;IACnB,MAAM,SAAS,MAAM,UAAU;IAC/B,eAAe,QAAQ,iBAAiB;IACxC,IAAI,MAAM,OAAO,QAAQ,KAAK,SAAS,MAAM,KAAK;IAClD,MAAM,UAAU,QAAQ,OAAO;IAC/B,IAAI,SAAS;KACX,QAAQ,aAAa,OAAO;KAC5B,MAAM;IACR;GACF;GACA,IAAI,SAAS,OAAO,SAClB,MACE,KAAK,WAAW,UAAU,SAAS,KAAK,KAAK,SAAS,OAAO;GAGjE,KAAK,qBAAqB,cAAc,KAAK;GAC7C,QAAQ,iBAAiB,KAAK,gBAAgB,YAAY,CAAC;GAC3D,UACE,KAAK,SACL,UACA,UACA,OACA,OACA,WACA,QAAQ,SACV;EACF,SAAS,OAAO;GACd,MACE,KAAK,WAAW,UAAU,SAAS,KAAK,KAAK,KAAK,SAAS,OAAO,KAAK;EAE3E,UAAU;GACR,IAAI,YAAY,CAAC,gBAAgB,QAAQ;IAIvC,IAAI;KACF,QAAa,QAAQ,OAAO,kBAAkB,CAAC,CAAC,CAAC,YAAY,CAAC,CAAC;IACjE,QAAQ,CAER;IACA,MAAM,cAAc,QAAQ;GAC9B;GACA,SAAS;GACT,gBAAgB,SAAS,WAAW;GACpC,UAAU,QAAQ;EACpB;CACF;CAEA,MAAM,YAAY,MAA+B;EAE/C,OAAO,KAAK,KAAK,KAAK,SAAS,CAAC;CAClC;CAMA,MAAM,YAAgC;EACpC,MAAM,SAAS,MAAM,aAAa;EAClC,MAAM,OAEF,KAAK,QAAQ,WAGZ,cAAc,OAAO,kBAAkB;EAC5C,MAAM,kBAAkB,IAAI,IAAI,OAAO,2BAA2B,CAAC,CAAC;EACpE,OAAO,KACJ,QAAQ,WAAW,OAAO,eAAe,CAAC,CAAC,CAC3C,KAAK,WAAW;GACf,MAAM,SAAS,OAAO,eAAe;GACrC,MAAM,YAAY,gBAAgB,IAAI,OAAO,QAAQ;GACrD,OAAO;IACL,IAAI,OAAO;IACX,MAAM,OAAO;IACb,aAAa,OAAO,mBAChB,+BAA+B,KAAK,MAAM,OAAO,gBAAgB,EAAE,aACnE;IACJ,eACE,OAAO,WAAW,uBAAuB;IAC3C,cAAc;KACZ;KACA;KACA,GAAI,YAAY,CAAC,WAAW,IAAI,CAAC;KACjC,GAAI,SAAS,CAAC,QAAQ,IAAI,CAAC;IAC7B;IACA,mBAAmB;IACnB,gBAAgB;GAClB;EACF,CAAC;CACL;CAEA,MAAM,kBAA2C;EAC/C,OAAO,EAAE,GAAG,oBAAoB;CAClC;CAMA,MAAM,MACJ,OACA,WAA6B,CAAC,GACF;EAC5B,OAAO,YAAY,YAAY;CACjC;CAEA,MAAM,WACJ,QACA,WAAkC,CAAC,GACP;EAC5B,OAAO,YAAY,kBAAkB;CACvC;CAEA,MAAM,cACJ,QACA,SACA,WAAoC,CAAC,GACpB;EACjB,OAAO,YAAY,mBAAmB;CACxC;CAEA,MAAM,cACJ,SACA,WAAmC,CAAC,GACF;EAClC,OAAO,YAAY,kBAAkB;CACvC;CAEA,MAAM,yBACJ,UAC6B;EAC7B,OAAO,YAAY,kBAAkB;CACvC;CAEA,MAAM,sBACJ,SACsC;EACtC,OAAO,YAAY,kBAAkB;CACvC;CAEA,MAAM,2BACJ,SACgC;EAChC,OAAO,YAAY,kBAAkB;CACvC;CAEA,MAAM,yBAAyB,SAA4C;EACzE,OAAO,YAAY,kBAAkB;CACvC;CAEA,MAAM,gCAAkD;EACtD,OAAO,YAAY,kBAAkB;CACvC;CAEA,MAAM,iBACJ,OACA,UACsB;EACtB,OAAO,YAAY,gBAAgB;CACrC;CAEA,aAAa,OAAe,UAA8C;EACxE,MAAM,QAAQ,IAAI,QAChB,qEACA,iBACA,QACF;EACA,OAAO,GACJ,OAAO,uBAAuB,EAC7B,YAAY,QAAQ,OAAO,KAAK,EAClC,GACF;CACF;CAEA,MAAM,WAAW,UAA6C;EAC5D,OAAO,YAAY,eAAe;CACpC;CAEA,MAAM,YAAY,UAA8C;EAC9D,OAAO,YAAY,cAAc;CACnC;CAEA,MAAM,UAAU,UAA+C;EAC7D,OAAO,YAAY,eAAe;CACpC;AACF;;AAYA,eAAe,cACb,UACe;CACf,IAAI;EACF,OAAO,EAAE,MAAM,SAAS,KAAK,EAAA,CAAG;CAGlC,QAAQ,CAER;AACF;;AAGA,SAAS,gBACP,SACA,MACM;CACN,IAAI,CAAC,SAAS;CACd,IAAI,MAAM,KAAU,KAAK,SAAS,OAAO;MACpC,QAAQ;AACf;;;;;;AAOA,eAAe,cACb,UACA,QACe;CACf,IAAI;EACF,KAAK,MAAM,SAAS,KAAK,EAAA,CAAG,MAAM;EAClC,MAAM,QAAQ,QAAQ,QAAQ,kBAAkB,CAAC;CACnD,QAAQ;EACN;CACF;CACA,MAAM,cAAc,QAAQ;AAC9B;;AAGA,SAAS,UAAa,SAAqB,QAAkC;CAC3E,IAAI,CAAC,QAAQ,OAAO;CACpB,IAAI,OAAO,SACT,OAAO,QAAQ,OACb,IAAI,QAAQ,gCAAgC,cAAc,QAAQ,CACpE;CAEF,OAAO,IAAI,SAAY,SAAS,WAAW;EACzC,MAAM,gBACJ,OACE,IAAI,QAAQ,gCAAgC,cAAc,QAAQ,CACpE;EACF,OAAO,iBAAiB,SAAS,SAAS,EAAE,MAAM,KAAK,CAAC;EACxD,QAAQ,MACL,UAAU;GACT,OAAO,oBAAoB,SAAS,OAAO;GAC3C,QAAQ,KAAK;EACf,IACC,UAAU;GACT,OAAO,oBAAoB,SAAS,OAAO;GAC3C,OAAO,KAAK;EACd,CACF;CACF,CAAC;AACH"}

package/dist/index.js

@@ -1,3 +1,3 @@
-import { _ as RateLimitError, b as extractTextContent, d as AI_PROVIDER_TYPES, f as AuthenticationError, g as ModelNotFoundError, h as DEFAULT_RATE_LIMIT_MAX_RETRY_DELAY_MS, i as DEFAULT_AI_TIMEOUT_MS, m as ContextLengthError, n as DEFAULT_AI_GENERATION_LIMITS, p as ContentFilterError, r as DEFAULT_AI_MAX_RETRIES, u as AIError, v as WebGPUUnavailableError, x as isRateLimitRetryableSoon, y as WebLLMPeerMissingError } from "./chunks/usage-DS7uDhi1.js";
-import { a as ChatGPTPlanSessionRevokedError, c as ChatGPTPlanWorkspaceError, d as getAI, f as getAIAuto, g as getOpenAI, h as getAIClient, i as ChatGPTPlanProvider, l as AIMessage, m as OpenAIClient, n as AIThread, o as ChatGPTPlanUsageLimitError, p as AIClient, r as ChatGPTPlanError, s as ChatGPTPlanUsageUnavailableError, t as PACKAGE_VERSION_INITIALIZED } from "./chunks/src-ChaEcI7l.js";
+import { S as isRateLimitRetryableSoon, _ as ModelNotFoundError, a as DEFAULT_AI_TIMEOUT_MS, b as WebLLMPeerMissingError, d as AIError, f as AI_PROVIDER_TYPES, g as DEFAULT_RATE_LIMIT_MAX_RETRY_DELAY_MS, h as ContextLengthError, i as DEFAULT_AI_MAX_RETRIES, m as ContentFilterError, p as AuthenticationError, r as DEFAULT_AI_GENERATION_LIMITS, v as RateLimitError, x as extractTextContent, y as WebGPUUnavailableError } from "./chunks/usage-yIT5lpQm.js";
+import { a as ChatGPTPlanSessionRevokedError, c as ChatGPTPlanWorkspaceError, d as getAI, f as getAIAuto, g as getOpenAI, h as getAIClient, i as ChatGPTPlanProvider, l as AIMessage, m as OpenAIClient, n as AIThread, o as ChatGPTPlanUsageLimitError, p as AIClient, r as ChatGPTPlanError, s as ChatGPTPlanUsageUnavailableError, t as PACKAGE_VERSION_INITIALIZED } from "./chunks/src-mOOK32QB.js";
 export { AIClient, AIError, AIMessage as AIMessageClass, AIThread, AI_PROVIDER_TYPES, AuthenticationError, ChatGPTPlanError, ChatGPTPlanProvider, ChatGPTPlanSessionRevokedError, ChatGPTPlanUsageLimitError, ChatGPTPlanUsageUnavailableError, ChatGPTPlanWorkspaceError, ContentFilterError, ContextLengthError, DEFAULT_AI_GENERATION_LIMITS, DEFAULT_AI_MAX_RETRIES, DEFAULT_AI_TIMEOUT_MS, DEFAULT_RATE_LIMIT_MAX_RETRY_DELAY_MS, ModelNotFoundError, OpenAIClient, PACKAGE_VERSION_INITIALIZED, RateLimitError, WebGPUUnavailableError, WebLLMPeerMissingError, extractTextContent, getAI, getAIAuto, getAIClient, getOpenAI, isRateLimitRetryableSoon };

package/dist/local.js

@@ -1,3 +1,3 @@
-import { v as WebGPUUnavailableError, y as WebLLMPeerMissingError } from "./chunks/usage-DS7uDhi1.js";
-import { a as isWebGPUAvailable, i as importWebLLM, n as WebLLMProvider, r as disposeWebLLMEngines, t as DEFAULT_WEBLLM_MODEL } from "./chunks/webllm-DVDsgVT7.js";
+import { b as WebLLMPeerMissingError, y as WebGPUUnavailableError } from "./chunks/usage-yIT5lpQm.js";
+import { a as isWebGPUAvailable, i as importWebLLM, n as WebLLMProvider, r as disposeWebLLMEngines, t as DEFAULT_WEBLLM_MODEL } from "./chunks/webllm-DrwR5V-y.js";
 export { DEFAULT_WEBLLM_MODEL, WebGPUUnavailableError, WebLLMPeerMissingError, WebLLMProvider, disposeWebLLMEngines, importWebLLM, isWebGPUAvailable };

package/dist/node.js

@@ -1,5 +1,5 @@
-import { _ as RateLimitError, b as extractTextContent, d as AI_PROVIDER_TYPES, f as AuthenticationError, g as ModelNotFoundError, h as DEFAULT_RATE_LIMIT_MAX_RETRY_DELAY_MS, i as DEFAULT_AI_TIMEOUT_MS, m as ContextLengthError, n as DEFAULT_AI_GENERATION_LIMITS, p as ContentFilterError, r as DEFAULT_AI_MAX_RETRIES, u as AIError, v as WebGPUUnavailableError, x as isRateLimitRetryableSoon, y as WebLLMPeerMissingError } from "./chunks/usage-DS7uDhi1.js";
-import { a as ChatGPTPlanSessionRevokedError, c as ChatGPTPlanWorkspaceError, d as getAI, g as getOpenAI, h as getAIClient, i as ChatGPTPlanProvider, l as AIMessage, m as OpenAIClient, n as AIThread, o as ChatGPTPlanUsageLimitError, p as AIClient, r as ChatGPTPlanError, s as ChatGPTPlanUsageUnavailableError, t as PACKAGE_VERSION_INITIALIZED } from "./chunks/src-ChaEcI7l.js";
+import { S as isRateLimitRetryableSoon, _ as ModelNotFoundError, a as DEFAULT_AI_TIMEOUT_MS, b as WebLLMPeerMissingError, d as AIError, f as AI_PROVIDER_TYPES, g as DEFAULT_RATE_LIMIT_MAX_RETRY_DELAY_MS, h as ContextLengthError, i as DEFAULT_AI_MAX_RETRIES, m as ContentFilterError, p as AuthenticationError, r as DEFAULT_AI_GENERATION_LIMITS, v as RateLimitError, x as extractTextContent, y as WebGPUUnavailableError } from "./chunks/usage-yIT5lpQm.js";
+import { a as ChatGPTPlanSessionRevokedError, c as ChatGPTPlanWorkspaceError, d as getAI, g as getOpenAI, h as getAIClient, i as ChatGPTPlanProvider, l as AIMessage, m as OpenAIClient, n as AIThread, o as ChatGPTPlanUsageLimitError, p as AIClient, r as ChatGPTPlanError, s as ChatGPTPlanUsageUnavailableError, t as PACKAGE_VERSION_INITIALIZED } from "./chunks/src-mOOK32QB.js";
 import { ValidationError, loadEnvConfig } from "@happyvertical/utils";
 //#region src/node/factory.ts
 /**
@@ -104,7 +104,7 @@ async function getAIAuto(options = {}) {
 		return getAI(config);
 	}
 	try {
-		return await import("./chunks/src-ChaEcI7l.js").then((n) => n.u).then((m) => m.getAIAuto(config));
+		return await import("./chunks/src-mOOK32QB.js").then((n) => n.u).then((m) => m.getAIAuto(config));
 	} catch (_error) {}
 	const hasLiteLLMSignal = Boolean(process.env.LITELLM_BASE_URL || process.env.LITELLM_API_KEY);
 	if (Boolean(process.env.BIFROST_BASE_URL) && !config.type) return getAI({

package/dist/shared/chat-template-kwargs.d.ts

@@ -0,0 +1,17 @@
+/** Largest accepted `chatTemplateKwargs`, measured as serialized JSON bytes. */
+export declare const MAX_CHAT_TEMPLATE_KWARGS_BYTES = 8192;
+/** Deepest accepted nesting of objects and arrays (the top level is 1). */
+export declare const MAX_CHAT_TEMPLATE_KWARGS_DEPTH = 4;
+/** Most keys accepted in any one object, and items in any one array. */
+export declare const MAX_CHAT_TEMPLATE_KWARGS_ENTRIES = 32;
+/**
+ * Validate `chatTemplateKwargs` and return a detached JSON copy to send as
+ * `chat_template_kwargs`, or `undefined` when the option is absent or empty.
+ * Objects may have `Object.prototype` or a null prototype
+ * (`Object.create(null)`); the copy is always an ordinary object.
+ *
+ * @throws {AIError} `AI_INVALID_CHAT_TEMPLATE_KWARGS` when the value is not a
+ *   plain JSON object within the size, depth, and entry bounds.
+ */
+export declare function prepareChatTemplateKwargs(value: unknown, provider?: string, model?: string): Record<string, unknown> | undefined;
+//# sourceMappingURL=chat-template-kwargs.d.ts.map

package/dist/shared/chat-template-kwargs.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"chat-template-kwargs.d.ts","sourceRoot":"","sources":["../../src/shared/chat-template-kwargs.ts"],"names":[],"mappings":"AAcA,gFAAgF;AAChF,eAAO,MAAM,8BAA8B,OAAO,CAAC;AACnD,2EAA2E;AAC3E,eAAO,MAAM,8BAA8B,IAAI,CAAC;AAChD,wEAAwE;AACxE,eAAO,MAAM,gCAAgC,KAAK,CAAC;AAiInD;;;;;;;;GAQG;AACH,wBAAgB,yBAAyB,CACvC,KAAK,EAAE,OAAO,EACd,QAAQ,CAAC,EAAE,MAAM,EACjB,KAAK,CAAC,EAAE,MAAM,GACb,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,GAAG,SAAS,CAYrC"}

package/dist/shared/continuation.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"continuation.d.ts","sourceRoot":"","sources":["../../src/shared/continuation.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAEV,WAAW,EACX,SAAS,EACT,UAAU,EACV,aAAa,EACb,WAAW,EACX,sBAAsB,EAEvB,MAAM,YAAS,CAAC;AAGjB,oEAAoE;AACpE,eAAO,MAAM,yBAAyB,IAAI,CAAC;AAO3C,eAAO,MAAM,eAAe,iHACoF,CAAC;AAEjH,UAAU,oBAAoB;IAC5B,gBAAgB,EAAE,MAAM,CAAC;CAC1B;AAED;;;;GAIG;AACH,wBAAgB,mBAAmB,CACjC,eAAe,EAAE,aAAa,EAC9B,OAAO,EAAE,IAAI,CAAC,WAAW,EAAE,OAAO,GAAG,gBAAgB,GAAG,gBAAgB,CAAC,GAAG;IAC1E,gBAAgB,CAAC,EAAE,sBAAsB,CAAC;CAC3C,GACA,oBAAoB,GAAG,SAAS,CAuBlC;AAmBD;;;;GAIG;AACH,wBAAgB,oBAAoB,CAClC,WAAW,EAAE,MAAM,EACnB,IAAI,EAAE,MAAM,GACX,MAAM,CAUR;AAkCD;;;GAGG;AACH,wBAAsB,oBAAoB,CACxC,IAAI,EAAE,CAAC,QAAQ,EAAE,SAAS,EAAE,EAAE,OAAO,CAAC,EAAE,WAAW,KAAK,OAAO,CAAC,UAAU,CAAC,EAC3E,eAAe,EAAE,aAAa,EAC9B,QAAQ,EAAE,SAAS,EAAE,EACrB,OAAO,GAAE,WAAW,GAAG;IAAE,gBAAgB,CAAC,EAAE,sBAAsB,CAAA;CAAO,GACxE,OAAO,CAAC,UAAU,CAAC,CA6CrB;AAED;;;;GAIG;AACH,wBAAuB,sBAAsB,CAC3C,MAAM,EAAE,CACN,QAAQ,EAAE,SAAS,EAAE,EACrB,OAAO,CAAC,EAAE,WAAW,KAClB,aAAa,CAAC,MAAM,CAAC,EAC1B,eAAe,EAAE,aAAa,EAC9B,QAAQ,EAAE,SAAS,EAAE,EACrB,OAAO,GAAE,WAAW,GAAG;IAAE,gBAAgB,CAAC,EAAE,sBAAsB,CAAA;CAAO,GACxE,aAAa,CAAC,MAAM,CAAC,CAyDvB;AAED;;;;;;;;;;;GAWG;AACH,wBAAgB,kBAAkB,CAAC,CAAC,SAAS,WAAW,EACtD,MAAM,EAAE,CAAC,EACT,eAAe,EAAE,aAAa,EAC9B,OAAO,GAAE,WAAoB,GAC5B,CAAC,CAsDH"}
+{"version":3,"file":"continuation.d.ts","sourceRoot":"","sources":["../../src/shared/continuation.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EAEV,WAAW,EACX,SAAS,EACT,UAAU,EACV,aAAa,EACb,WAAW,EACX,sBAAsB,EAEvB,MAAM,YAAS,CAAC;AAGjB,oEAAoE;AACpE,eAAO,MAAM,yBAAyB,IAAI,CAAC;AAO3C,eAAO,MAAM,eAAe,iHACoF,CAAC;AAEjH,UAAU,oBAAoB;IAC5B,gBAAgB,EAAE,MAAM,CAAC;CAC1B;AAED;;;;GAIG;AACH,wBAAgB,mBAAmB,CACjC,eAAe,EAAE,aAAa,EAC9B,OAAO,EAAE,IAAI,CAAC,WAAW,EAAE,OAAO,GAAG,gBAAgB,GAAG,gBAAgB,CAAC,GAAG;IAC1E,gBAAgB,CAAC,EAAE,sBAAsB,CAAC;CAC3C,GACA,oBAAoB,GAAG,SAAS,CAuBlC;AAmBD;;;;GAIG;AACH,wBAAgB,oBAAoB,CAClC,WAAW,EAAE,MAAM,EACnB,IAAI,EAAE,MAAM,GACX,MAAM,CAUR;AAkCD;;;GAGG;AACH,wBAAsB,oBAAoB,CACxC,IAAI,EAAE,CAAC,QAAQ,EAAE,SAAS,EAAE,EAAE,OAAO,CAAC,EAAE,WAAW,KAAK,OAAO,CAAC,UAAU,CAAC,EAC3E,eAAe,EAAE,aAAa,EAC9B,QAAQ,EAAE,SAAS,EAAE,EACrB,OAAO,GAAE,WAAW,GAAG;IAAE,gBAAgB,CAAC,EAAE,sBAAsB,CAAA;CAAO,GACxE,OAAO,CAAC,UAAU,CAAC,CAiDrB;AAED;;;;GAIG;AACH,wBAAuB,sBAAsB,CAC3C,MAAM,EAAE,CACN,QAAQ,EAAE,SAAS,EAAE,EACrB,OAAO,CAAC,EAAE,WAAW,KAClB,aAAa,CAAC,MAAM,CAAC,EAC1B,eAAe,EAAE,aAAa,EAC9B,QAAQ,EAAE,SAAS,EAAE,EACrB,OAAO,GAAE,WAAW,GAAG;IAAE,gBAAgB,CAAC,EAAE,sBAAsB,CAAA;CAAO,GACxE,aAAa,CAAC,MAAM,CAAC,CAmEvB;AAED;;;;;;;;;;;GAWG;AACH,wBAAgB,kBAAkB,CAAC,CAAC,SAAS,WAAW,EACtD,MAAM,EAAE,CAAC,EACT,eAAe,EAAE,aAAa,EAC9B,OAAO,GAAE,WAAoB,GAC5B,CAAC,CAsDH"}

package/dist/shared/providers/connection-retry.d.ts

@@ -0,0 +1,34 @@
+/**
+ * One-shot retry for stale keep-alive connections in the OpenAI-compatible
+ * providers (#1467).
+ *
+ * A pooled socket the server closed while idle fails the next request
+ * (ECONNRESET, EPIPE, undici `UND_ERR_SOCKET`, "socket hang up"). Callers
+ * decide when a resend is safe: `chat()` only before response headers,
+ * `stream()` only before its first yielded chunk. HTTP status responses,
+ * timeouts, and caller aborts are never retried.
+ */
+import { type AIAttemptRecorder } from "../safety.js";
+/**
+ * True when `error` (or anything on its `cause` chain) is a stale
+ * connection failure: the OpenAI SDK's `APIConnectionError` (not its timeout
+ * subclass) or a socket error coded ECONNRESET / EPIPE / UND_ERR_SOCKET /
+ * "socket hang up". Any timeout, abort, or HTTP status on the chain wins and
+ * makes the error non-retryable.
+ */
+export declare function isStaleConnectionError(error: unknown): boolean;
+/**
+ * Decide whether a failed attempt gets its single transparent retry, and
+ * record it on the per-call attempt recorder when it does.
+ */
+export declare function shouldRetryConnection(error: unknown, state: {
+    enabled: boolean | undefined;
+    alreadyRetried: boolean;
+    signal: AbortSignal;
+    callerSignal?: AbortSignal;
+    didTimeout: () => boolean;
+    recorder?: AIAttemptRecorder;
+}): boolean;
+/** The per-call attempt recorder the observed client attaches, if any. */
+export declare function attemptRecorderFrom(options: object): AIAttemptRecorder | undefined;
+//# sourceMappingURL=connection-retry.d.ts.map

package/dist/shared/providers/connection-retry.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"connection-retry.d.ts","sourceRoot":"","sources":["../../../src/shared/providers/connection-retry.ts"],"names":[],"mappings":"AAAA;;;;;;;;;GASG;AAGH,OAAO,EAAuB,KAAK,iBAAiB,EAAE,MAAM,cAAW,CAAC;AAKxE;;;;;;GAMG;AACH,wBAAgB,sBAAsB,CAAC,KAAK,EAAE,OAAO,GAAG,OAAO,CAoC9D;AAED;;;GAGG;AACH,wBAAgB,qBAAqB,CACnC,KAAK,EAAE,OAAO,EACd,KAAK,EAAE;IACL,OAAO,EAAE,OAAO,GAAG,SAAS,CAAC;IAC7B,cAAc,EAAE,OAAO,CAAC;IACxB,MAAM,EAAE,WAAW,CAAC;IACpB,YAAY,CAAC,EAAE,WAAW,CAAC;IAC3B,UAAU,EAAE,MAAM,OAAO,CAAC;IAC1B,QAAQ,CAAC,EAAE,iBAAiB,CAAC;CAC9B,GACA,OAAO,CAaT;AAED,0EAA0E;AAC1E,wBAAgB,mBAAmB,CACjC,OAAO,EAAE,MAAM,GACd,iBAAiB,GAAG,SAAS,CAI/B"}

package/dist/shared/providers/ollama.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"ollama.d.ts","sourceRoot":"","sources":["../../../src/shared/providers/ollama.ts"],"names":[],"mappings":"AAAA;;;;;;GAMG;AAaH,OAAO,KAAK,EACV,cAAc,EACd,WAAW,EACX,SAAS,EACT,OAAO,EACP,UAAU,EACV,WAAW,EACX,iBAAiB,EACjB,gBAAgB,EAChB,iBAAiB,EACjB,uBAAuB,EACvB,qBAAqB,EACrB,sBAAsB,EACtB,uBAAuB,EACvB,cAAc,EACd,aAAa,EAEb,UAAU,EACV,WAAW,EACX,kBAAkB,EAClB,sBAAsB,EACtB,qBAAqB,EACrB,2BAA2B,EAC3B,KAAK,EACL,iBAAiB,EACjB,kBAAkB,EAClB,gBAAgB,EACjB,MAAM,aAAU,CAAC;AA4TlB,qBAAa,cAAe,YAAW,WAAW;IAChD,OAAO,CAAC,QAAQ,CAAC,OAAO,CAAgB;IACxC,OAAO,CAAC,QAAQ,CAAC,IAAI,CAAS;IAC9B,OAAO,CAAC,QAAQ,CAAC,sBAAsB,CAAC,CAAS;IACjD,OAAO,CAAC,gBAAgB,CAAC,CAAqB;IAC9C,OAAO,CAAC,QAAQ,CAAC,cAAc,CAG3B;IACJ,OAAO,CAAC,QAAQ,CAAC,kBAAkB,CAG/B;gBAEQ,OAAO,EAAE,aAAa;IASlC,OAAO,KAAK,aAAa,GAExB;IAED,OAAO,KAAK,oBAAoB,GAE/B;IAED,OAAO,CAAC,YAAY;YAaN,cAAc;IA+B5B,OAAO,CAAC,QAAQ;YAcF,WAAW;YAsBX,aAAa;YAmBZ,WAAW;YAiCZ,SAAS;YAeT,YAAY;YA4BZ,WAAW;IAyCzB,OAAO,CAAC,eAAe;YAQT,oBAAoB;YA0BpB,cAAc;IA0B5B,OAAO,CAAC,kBAAkB;YAaZ,mBAAmB;IA2DjC,OAAO,CAAC,QAAQ;IA6BhB,OAAO,CAAC,mBAAmB;IA8B3B,OAAO,CAAC,QAAQ;IAkBV,IAAI,CACR,QAAQ,EAAE,SAAS,EAAE,EACrB,OAAO,GAAE,WAAgB,GACxB,OAAO,CAAC,UAAU,CAAC;IAyFhB,QAAQ,CACZ,MAAM,EAAE,MAAM,EACd,OAAO,GAAE,iBAAsB,GAC9B,OAAO,CAAC,UAAU,CAAC;IAkFhB,OAAO,CAAC,IAAI,EAAE,MAAM,EAAE,OAAO,GAAE,cAAmB,GAAG,OAAO,CAAC,MAAM,CAAC;IA8BpE,KAAK,CACT,IAAI,EAAE,MAAM,GAAG,MAAM,EAAE,EACvB,OAAO,GAAE,gBAAqB,GAC7B,OAAO,CAAC,iBAAiB,CAAC;IA4CvB,UAAU,CACd,KAAK,EAAE,MAAM,GAAG,MAAM,EACtB,OAAO,GAAE,qBAA0B,GAClC,OAAO,CAAC,iBAAiB,CAAC;IAavB,aAAa,CACjB,KAAK,EAAE,MAAM,GAAG,MAAM,EACtB,MAAM,CAAC,EAAE,MAAM,EACf,OAAO,GAAE,uBAA4B,GACpC,OAAO,CAAC,MAAM,CAAC;IAqCZ,aAAa,CACjB,MAAM,EAAE,MAAM,EACd,OAAO,GAAE,sBAA2B,GACnC,OAAO,CAAC,uBAAuB,CAAC;IAgF5B,MAAM,CACX,QAAQ,EAAE,SAAS,EAAE,EACrB,OAAO,GAAE,WAAgB,GACxB,aAAa,CAAC,MAAM,CAAC;IA4FlB,WAAW,CAAC,IAAI,EAAE,MAAM,GAAG,OAAO,CAAC,MAAM,CAAC;IAI1C,SAAS,IAAI,OAAO,CAAC,OAAO,EAAE,CAAC;IAkD/B,eAAe,IAAI,OAAO,CAAC,cAAc,CAAC;IAS1C,wBAAwB,CAC5B,QAAQ,EAAE,sBAAsB,GAC/B,OAAO,CAAC,kBAAkB,CAAC;IAQxB,qBAAqB,CACzB,OAAO,EAAE,kBAAkB,GAC1B,OAAO,CAAC,2BAA2B,CAAC;IAQjC,0BAA0B,CAC9B,OAAO,EAAE,kBAAkB,GAC1B,OAAO,CAAC,qBAAqB,CAAC;IAQ3B,wBAAwB,CAAC,OAAO,EAAE,kBAAkB,GAAG,OAAO,CAAC,IAAI,CAAC;IAQpE,6BAA6B,IAAI,OAAO,CAAC,OAAO,CAAC;IAQjD,gBAAgB,CACpB,KAAK,EAAE,MAAM,EACb,QAAQ,CAAC,EAAE,UAAU,GACpB,OAAO,CAAC,WAAW,CAAC;IAQvB,YAAY,CAAC,KAAK,EAAE,MAAM,EAAE,QAAQ,CAAC,EAAE,UAAU,GAAG,aAAa,CAAC,MAAM,CAAC;IAanE,UAAU,CAAC,QAAQ,EAAE,iBAAiB,GAAG,OAAO,CAAC,KAAK,CAAC;IAQvD,WAAW,CAAC,QAAQ,EAAE,kBAAkB,GAAG,OAAO,CAAC,KAAK,CAAC;IAQzD,SAAS,CAAC,QAAQ,CAAC,EAAE,gBAAgB,GAAG,OAAO,CAAC,KAAK,EAAE,CAAC;CAO/D"}
+{"version":3,"file":"ollama.d.ts","sourceRoot":"","sources":["../../../src/shared/providers/ollama.ts"],"names":[],"mappings":"AAAA;;;;;;GAMG;AAcH,OAAO,KAAK,EACV,cAAc,EACd,WAAW,EACX,SAAS,EACT,OAAO,EACP,UAAU,EACV,WAAW,EACX,iBAAiB,EACjB,gBAAgB,EAChB,iBAAiB,EACjB,uBAAuB,EACvB,qBAAqB,EACrB,sBAAsB,EACtB,uBAAuB,EACvB,cAAc,EACd,aAAa,EAEb,UAAU,EACV,WAAW,EACX,kBAAkB,EAClB,sBAAsB,EACtB,qBAAqB,EACrB,2BAA2B,EAC3B,KAAK,EACL,iBAAiB,EACjB,kBAAkB,EAClB,gBAAgB,EACjB,MAAM,aAAU,CAAC;AA4TlB,qBAAa,cAAe,YAAW,WAAW;IAChD,OAAO,CAAC,QAAQ,CAAC,OAAO,CAAgB;IACxC,OAAO,CAAC,QAAQ,CAAC,IAAI,CAAS;IAC9B,OAAO,CAAC,QAAQ,CAAC,sBAAsB,CAAC,CAAS;IACjD,OAAO,CAAC,gBAAgB,CAAC,CAAqB;IAC9C,OAAO,CAAC,QAAQ,CAAC,cAAc,CAG3B;IACJ,OAAO,CAAC,QAAQ,CAAC,kBAAkB,CAG/B;gBAEQ,OAAO,EAAE,aAAa;IASlC,OAAO,KAAK,aAAa,GAExB;IAED,OAAO,KAAK,oBAAoB,GAE/B;IAED,OAAO,CAAC,YAAY;YAaN,cAAc;IA+B5B,OAAO,CAAC,QAAQ;YAcF,WAAW;YAsBX,aAAa;YAmBZ,WAAW;YAiCZ,SAAS;YAeT,YAAY;YA4BZ,WAAW;IAyCzB,OAAO,CAAC,eAAe;YAQT,oBAAoB;YA0BpB,cAAc;IA0B5B,OAAO,CAAC,kBAAkB;YAaZ,mBAAmB;IA2DjC,OAAO,CAAC,QAAQ;IA6BhB,OAAO,CAAC,mBAAmB;IA8B3B,OAAO,CAAC,QAAQ;IAkBV,IAAI,CACR,QAAQ,EAAE,SAAS,EAAE,EACrB,OAAO,GAAE,WAAgB,GACxB,OAAO,CAAC,UAAU,CAAC;IA4FhB,QAAQ,CACZ,MAAM,EAAE,MAAM,EACd,OAAO,GAAE,iBAAsB,GAC9B,OAAO,CAAC,UAAU,CAAC;IAmFhB,OAAO,CAAC,IAAI,EAAE,MAAM,EAAE,OAAO,GAAE,cAAmB,GAAG,OAAO,CAAC,MAAM,CAAC;IA8BpE,KAAK,CACT,IAAI,EAAE,MAAM,GAAG,MAAM,EAAE,EACvB,OAAO,GAAE,gBAAqB,GAC7B,OAAO,CAAC,iBAAiB,CAAC;IA4CvB,UAAU,CACd,KAAK,EAAE,MAAM,GAAG,MAAM,EACtB,OAAO,GAAE,qBAA0B,GAClC,OAAO,CAAC,iBAAiB,CAAC;IAavB,aAAa,CACjB,KAAK,EAAE,MAAM,GAAG,MAAM,EACtB,MAAM,CAAC,EAAE,MAAM,EACf,OAAO,GAAE,uBAA4B,GACpC,OAAO,CAAC,MAAM,CAAC;IAqCZ,aAAa,CACjB,MAAM,EAAE,MAAM,EACd,OAAO,GAAE,sBAA2B,GACnC,OAAO,CAAC,uBAAuB,CAAC;IAgF5B,MAAM,CACX,QAAQ,EAAE,SAAS,EAAE,EACrB,OAAO,GAAE,WAAgB,GACxB,aAAa,CAAC,MAAM,CAAC;IAqGlB,WAAW,CAAC,IAAI,EAAE,MAAM,GAAG,OAAO,CAAC,MAAM,CAAC;IAI1C,SAAS,IAAI,OAAO,CAAC,OAAO,EAAE,CAAC;IAkD/B,eAAe,IAAI,OAAO,CAAC,cAAc,CAAC;IAS1C,wBAAwB,CAC5B,QAAQ,EAAE,sBAAsB,GAC/B,OAAO,CAAC,kBAAkB,CAAC;IAQxB,qBAAqB,CACzB,OAAO,EAAE,kBAAkB,GAC1B,OAAO,CAAC,2BAA2B,CAAC;IAQjC,0BAA0B,CAC9B,OAAO,EAAE,kBAAkB,GAC1B,OAAO,CAAC,qBAAqB,CAAC;IAQ3B,wBAAwB,CAAC,OAAO,EAAE,kBAAkB,GAAG,OAAO,CAAC,IAAI,CAAC;IAQpE,6BAA6B,IAAI,OAAO,CAAC,OAAO,CAAC;IAQjD,gBAAgB,CACpB,KAAK,EAAE,MAAM,EACb,QAAQ,CAAC,EAAE,UAAU,GACpB,OAAO,CAAC,WAAW,CAAC;IAQvB,YAAY,CAAC,KAAK,EAAE,MAAM,EAAE,QAAQ,CAAC,EAAE,UAAU,GAAG,aAAa,CAAC,MAAM,CAAC;IAanE,UAAU,CAAC,QAAQ,EAAE,iBAAiB,GAAG,OAAO,CAAC,KAAK,CAAC;IAQvD,WAAW,CAAC,QAAQ,EAAE,kBAAkB,GAAG,OAAO,CAAC,KAAK,CAAC;IAQzD,SAAS,CAAC,QAAQ,CAAC,EAAE,gBAAgB,GAAG,OAAO,CAAC,KAAK,EAAE,CAAC;CAO/D"}

package/dist/shared/providers/openai.d.ts

@@ -82,6 +82,8 @@ export interface OpenAICompatibleOptions extends BaseAIOptions {
     apiKey?: string;
     baseUrl?: string;
     organization?: string;
+    /** See {@link OpenAIOptions.retryConnectionErrors}. @default true */
+    retryConnectionErrors?: boolean;
 }
 /**
  * OpenAI provider implementation that handles all interactions with OpenAI's API.
@@ -317,12 +319,30 @@ export declare class OpenAIProvider implements AIInterface {
      */
     private mapUsage;
     /**
-     * Maps OpenAI finish reason to internal finish reason format
+     * Reject `chatTemplateKwargs` on the hosted OpenAI API, which does not
+     * accept `chat_template_kwargs`. Throwing (rather than dropping the field)
+     * keeps a caller from silently losing a template switch it asked for, such
+     * as `enable_thinking: false`, and fails before the request is billed.
+     * The resolved client URL is checked, so `OPENAI_BASE_URL` pointing at a
+     * vLLM server still works when `baseUrl` is unset.
+     * @private
+     */
+    private assertChatTemplateKwargsSupported;
+    /**
+     * Maps OpenAI finish reason to internal finish reason format. An absent
+     * reason is `'stop'`; a value outside the normalised set (e.g. vLLM
+     * `abort`, or legacy `function_call`, which carries no `tool_calls`) is
+     * `'other'` so a partial reply is not mistaken for a complete one.
      * @param reason - OpenAI finish reason from API response
      * @returns Internal finish reason
      * @private
      */
     private mapFinishReason;
+    /**
+     * Whether a failed chat/stream attempt gets its one transparent
+     * stale-connection retry (#1467). Records the retry for `attempts`.
+     */
+    private retriesConnection;
     /**
      * Gets the context length for a given OpenAI model
      * @param modelId - The OpenAI model identifier

package/dist/shared/providers/openai.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"openai.d.ts","sourceRoot":"","sources":["../../../src/shared/providers/openai.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAWH,OAAO,KAAK,EACV,cAAc,EACd,WAAW,EACX,SAAS,EACT,OAAO,EACP,UAAU,EACV,aAAa,EACb,WAAW,EACX,iBAAiB,EAEjB,gBAAgB,EAChB,iBAAiB,EACjB,uBAAuB,EACvB,qBAAqB,EACrB,sBAAsB,EACtB,uBAAuB,EACvB,cAAc,EAEd,UAAU,EACV,WAAW,EACX,kBAAkB,EAClB,sBAAsB,EACtB,qBAAqB,EACrB,2BAA2B,EAC3B,KAAK,EACL,iBAAiB,EACjB,kBAAkB,EAClB,gBAAgB,EACjB,MAAM,aAAU,CAAC;AAUlB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAkCG;AACH,wBAAgB,wBAAwB,CAAC,KAAK,EAAE,MAAM,GAAG,SAAS,GAAG,OAAO,CAK3E;AAED;;;;;;;;;;;;;GAaG;AACH,wBAAgB,4BAA4B,CAC1C,KAAK,EAAE,MAAM,EACb,SAAS,EAAE,MAAM,GAAG,SAAS,EAC7B,WAAW,EAAE,MAAM,GAAG,SAAS,GAC9B;IACD,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,qBAAqB,CAAC,EAAE,MAAM,CAAC;IAC/B,WAAW,CAAC,EAAE,MAAM,CAAC;CACtB,CAQA;AAED;;GAEG;AACH,MAAM,WAAW,uBAAuB;IACtC,aAAa,EAAE,MAAM,CAAC;IACtB,YAAY,EAAE,MAAM,CAAC;IACrB,YAAY,EAAE,MAAM,CAAC;IACrB,YAAY,EAAE,cAAc,CAAC;IAC7B,aAAa,CAAC,OAAO,EAAE,MAAM,GAAG,MAAM,CAAC;IACvC,gBAAgB,CAAC,OAAO,EAAE,MAAM,GAAG,MAAM,CAAC;IAC1C,oBAAoB,CAAC,OAAO,EAAE,MAAM,GAAG,MAAM,EAAE,CAAC;IAChD,kBAAkB,CAAC,OAAO,EAAE,MAAM,GAAG,OAAO,CAAC;IAC7C,iBAAiB,CAAC,OAAO,EAAE,MAAM,GAAG,OAAO,CAAC;IAC5C,cAAc,CAAC,OAAO,EAAE,MAAM,GAAG,OAAO,CAAC;CAC1C;AA0ED,MAAM,WAAW,uBAAwB,SAAQ,aAAa;IAC5D,IAAI,CAAC,EAAE,MAAM,CAAC;IACd,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,YAAY,CAAC,EAAE,MAAM,CAAC;CACvB;AAED;;;GAGG;AACH,qBAAa,cAAe,YAAW,WAAW;IAChD,OAAO,CAAC,MAAM,CAAS;IACvB,SAAS,CAAC,OAAO,EAAE,uBAAuB,CAAC;IAC3C,SAAS,CAAC,QAAQ,CAAC,OAAO,EAAE,uBAAuB,CAAC;IAEpD;;;OAGG;gBAED,OAAO,EAAE,uBAAuB,EAChC,OAAO,GAAE,uBAAwC;IAkBnD;;;;;;;;;;;;;;;;;;;OAmBG;IACG,IAAI,CACR,QAAQ,EAAE,SAAS,EAAE,EACrB,OAAO,GAAE,WAAgB,GACxB,OAAO,CAAC,UAAU,CAAC;IAoItB;;;;;;;;;;;;;;;OAeG;IACG,QAAQ,CACZ,MAAM,EAAE,MAAM,EACd,OAAO,GAAE,iBAAsB,GAC9B,OAAO,CAAC,UAAU,CAAC;IAkBtB;;;;;;;;;;;;;;;;;;;;;;;;;;OA0BG;IACG,OAAO,CAAC,IAAI,EAAE,MAAM,EAAE,OAAO,GAAE,cAAmB,GAAG,OAAO,CAAC,MAAM,CAAC;IA+B1E;;;;;;;;;;;;;;;;;OAiBG;IACG,KAAK,CACT,IAAI,EAAE,MAAM,GAAG,MAAM,EAAE,EACvB,OAAO,GAAE,gBAAqB,GAC7B,OAAO,CAAC,iBAAiB,CAAC;IAkC7B;;;;;OAKG;YACW,aAAa;IAyB3B;;;;;;;;;;;OAWG;IACG,aAAa,CACjB,KAAK,EAAE,MAAM,GAAG,MAAM,EACtB,MAAM,CAAC,EAAE,MAAM,EACf,OAAO,GAAE,uBAA4B,GACpC,OAAO,CAAC,MAAM,CAAC;IA8DlB;;;;;;;;;;OAUG;IACG,UAAU,CACd,KAAK,EAAE,MAAM,GAAG,MAAM,EACtB,OAAO,GAAE,qBAA0B,GAClC,OAAO,CAAC,iBAAiB,CAAC;IAkB7B;;;;;;;;;;;OAWG;IACG,aAAa,CACjB,MAAM,EAAE,MAAM,EACd,OAAO,GAAE,sBAA2B,GACnC,OAAO,CAAC,uBAAuB,CAAC;IAyF7B,wBAAwB,CAC5B,QAAQ,EAAE,sBAAsB,GAC/B,OAAO,CAAC,kBAAkB,CAAC;IAQxB,qBAAqB,CACzB,OAAO,EAAE,kBAAkB,GAC1B,OAAO,CAAC,2BAA2B,CAAC;IAQjC,0BAA0B,CAC9B,OAAO,EAAE,kBAAkB,GAC1B,OAAO,CAAC,qBAAqB,CAAC;IAQ3B,wBAAwB,CAAC,OAAO,EAAE,kBAAkB,GAAG,OAAO,CAAC,IAAI,CAAC;IAQpE,6BAA6B,IAAI,OAAO,CAAC,OAAO,CAAC;IAQvD;;;;;;;;;;;;;;;OAeG;IACI,MAAM,CACX,QAAQ,EAAE,SAAS,EAAE,EACrB,OAAO,GAAE,WAAgB,GACxB,aAAa,CAAC,MAAM,CAAC;IAsGxB;;;;;;;;;;;;;;;OAeG;IACG,WAAW,CAAC,IAAI,EAAE,MAAM,GAAG,OAAO,CAAC,MAAM,CAAC;IAMhD;;;;;;;;;;OAUG;IACG,SAAS,IAAI,OAAO,CAAC,OAAO,EAAE,CAAC;IAmBrC;;;;;;;;;;;OAWG;IACG,eAAe,IAAI,OAAO,CAAC,cAAc,CAAC;IAQ1C,gBAAgB,CACpB,KAAK,EAAE,MAAM,EACb,QAAQ,CAAC,EAAE,UAAU,GACpB,OAAO,CAAC,WAAW,CAAC;IAQvB,YAAY,CAAC,KAAK,EAAE,MAAM,EAAE,QAAQ,CAAC,EAAE,UAAU,GAAG,aAAa,CAAC,MAAM,CAAC;IAanE,UAAU,CAAC,QAAQ,EAAE,iBAAiB,GAAG,OAAO,CAAC,KAAK,CAAC;IAQvD,WAAW,CAAC,QAAQ,EAAE,kBAAkB,GAAG,OAAO,CAAC,KAAK,CAAC;IAQzD,SAAS,CAAC,QAAQ,CAAC,EAAE,gBAAgB,GAAG,OAAO,CAAC,KAAK,EAAE,CAAC;IAQ9D;;;;;OAKG;IACH,OAAO,CAAC,mBAAmB;IAkE3B;;;;;OAKG;IACH,OAAO,CAAC,aAAa;IAcrB;;;;;OAKG;IACH,OAAO,CAAC,QAAQ;IAchB;;;;;OAKG;IACH,OAAO,CAAC,eAAe;IAevB;;;;;OAKG;IACH;;;;;OAKG;IAEH,OAAO,CAAC,QAAQ;CAgDjB"}
+{"version":3,"file":"openai.d.ts","sourceRoot":"","sources":["../../../src/shared/providers/openai.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAaH,OAAO,KAAK,EACV,cAAc,EAEd,WAAW,EACX,SAAS,EACT,OAAO,EACP,UAAU,EACV,aAAa,EACb,WAAW,EACX,iBAAiB,EAEjB,gBAAgB,EAChB,iBAAiB,EACjB,uBAAuB,EACvB,qBAAqB,EACrB,sBAAsB,EACtB,uBAAuB,EACvB,cAAc,EAEd,UAAU,EACV,WAAW,EACX,kBAAkB,EAClB,sBAAsB,EACtB,qBAAqB,EACrB,2BAA2B,EAC3B,KAAK,EACL,iBAAiB,EACjB,kBAAkB,EAClB,gBAAgB,EACjB,MAAM,aAAU,CAAC;AAWlB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAkCG;AACH,wBAAgB,wBAAwB,CAAC,KAAK,EAAE,MAAM,GAAG,SAAS,GAAG,OAAO,CAK3E;AAED;;;;;;;;;;;;;GAaG;AACH,wBAAgB,4BAA4B,CAC1C,KAAK,EAAE,MAAM,EACb,SAAS,EAAE,MAAM,GAAG,SAAS,EAC7B,WAAW,EAAE,MAAM,GAAG,SAAS,GAC9B;IACD,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,qBAAqB,CAAC,EAAE,MAAM,CAAC;IAC/B,WAAW,CAAC,EAAE,MAAM,CAAC;CACtB,CAQA;AAED;;GAEG;AACH,MAAM,WAAW,uBAAuB;IACtC,aAAa,EAAE,MAAM,CAAC;IACtB,YAAY,EAAE,MAAM,CAAC;IACrB,YAAY,EAAE,MAAM,CAAC;IACrB,YAAY,EAAE,cAAc,CAAC;IAC7B,aAAa,CAAC,OAAO,EAAE,MAAM,GAAG,MAAM,CAAC;IACvC,gBAAgB,CAAC,OAAO,EAAE,MAAM,GAAG,MAAM,CAAC;IAC1C,oBAAoB,CAAC,OAAO,EAAE,MAAM,GAAG,MAAM,EAAE,CAAC;IAChD,kBAAkB,CAAC,OAAO,EAAE,MAAM,GAAG,OAAO,CAAC;IAC7C,iBAAiB,CAAC,OAAO,EAAE,MAAM,GAAG,OAAO,CAAC;IAC5C,cAAc,CAAC,OAAO,EAAE,MAAM,GAAG,OAAO,CAAC;CAC1C;AA0ED,MAAM,WAAW,uBAAwB,SAAQ,aAAa;IAC5D,IAAI,CAAC,EAAE,MAAM,CAAC;IACd,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,qEAAqE;IACrE,qBAAqB,CAAC,EAAE,OAAO,CAAC;CACjC;AAED;;;GAGG;AACH,qBAAa,cAAe,YAAW,WAAW;IAChD,OAAO,CAAC,MAAM,CAAS;IACvB,SAAS,CAAC,OAAO,EAAE,uBAAuB,CAAC;IAC3C,SAAS,CAAC,QAAQ,CAAC,OAAO,EAAE,uBAAuB,CAAC;IAEpD;;;OAGG;gBAED,OAAO,EAAE,uBAAuB,EAChC,OAAO,GAAE,uBAAwC;IAkBnD;;;;;;;;;;;;;;;;;;;OAmBG;IACG,IAAI,CACR,QAAQ,EAAE,SAAS,EAAE,EACrB,OAAO,GAAE,WAAgB,GACxB,OAAO,CAAC,UAAU,CAAC;IAqLtB;;;;;;;;;;;;;;;OAeG;IACG,QAAQ,CACZ,MAAM,EAAE,MAAM,EACd,OAAO,GAAE,iBAAsB,GAC9B,OAAO,CAAC,UAAU,CAAC;IAkBtB;;;;;;;;;;;;;;;;;;;;;;;;;;OA0BG;IACG,OAAO,CAAC,IAAI,EAAE,MAAM,EAAE,OAAO,GAAE,cAAmB,GAAG,OAAO,CAAC,MAAM,CAAC;IA+B1E;;;;;;;;;;;;;;;;;OAiBG;IACG,KAAK,CACT,IAAI,EAAE,MAAM,GAAG,MAAM,EAAE,EACvB,OAAO,GAAE,gBAAqB,GAC7B,OAAO,CAAC,iBAAiB,CAAC;IAkC7B;;;;;OAKG;YACW,aAAa;IAyB3B;;;;;;;;;;;OAWG;IACG,aAAa,CACjB,KAAK,EAAE,MAAM,GAAG,MAAM,EACtB,MAAM,CAAC,EAAE,MAAM,EACf,OAAO,GAAE,uBAA4B,GACpC,OAAO,CAAC,MAAM,CAAC;IA8DlB;;;;;;;;;;OAUG;IACG,UAAU,CACd,KAAK,EAAE,MAAM,GAAG,MAAM,EACtB,OAAO,GAAE,qBAA0B,GAClC,OAAO,CAAC,iBAAiB,CAAC;IAkB7B;;;;;;;;;;;OAWG;IACG,aAAa,CACjB,MAAM,EAAE,MAAM,EACd,OAAO,GAAE,sBAA2B,GACnC,OAAO,CAAC,uBAAuB,CAAC;IAyF7B,wBAAwB,CAC5B,QAAQ,EAAE,sBAAsB,GAC/B,OAAO,CAAC,kBAAkB,CAAC;IAQxB,qBAAqB,CACzB,OAAO,EAAE,kBAAkB,GAC1B,OAAO,CAAC,2BAA2B,CAAC;IAQjC,0BAA0B,CAC9B,OAAO,EAAE,kBAAkB,GAC1B,OAAO,CAAC,qBAAqB,CAAC;IAQ3B,wBAAwB,CAAC,OAAO,EAAE,kBAAkB,GAAG,OAAO,CAAC,IAAI,CAAC;IAQpE,6BAA6B,IAAI,OAAO,CAAC,OAAO,CAAC;IAQvD;;;;;;;;;;;;;;;OAeG;IACI,MAAM,CACX,QAAQ,EAAE,SAAS,EAAE,EACrB,OAAO,GAAE,WAAgB,GACxB,aAAa,CAAC,MAAM,CAAC;IA6IxB;;;;;;;;;;;;;;;OAeG;IACG,WAAW,CAAC,IAAI,EAAE,MAAM,GAAG,OAAO,CAAC,MAAM,CAAC;IAMhD;;;;;;;;;;OAUG;IACG,SAAS,IAAI,OAAO,CAAC,OAAO,EAAE,CAAC;IAmBrC;;;;;;;;;;;OAWG;IACG,eAAe,IAAI,OAAO,CAAC,cAAc,CAAC;IAQ1C,gBAAgB,CACpB,KAAK,EAAE,MAAM,EACb,QAAQ,CAAC,EAAE,UAAU,GACpB,OAAO,CAAC,WAAW,CAAC;IAQvB,YAAY,CAAC,KAAK,EAAE,MAAM,EAAE,QAAQ,CAAC,EAAE,UAAU,GAAG,aAAa,CAAC,MAAM,CAAC;IAanE,UAAU,CAAC,QAAQ,EAAE,iBAAiB,GAAG,OAAO,CAAC,KAAK,CAAC;IAQvD,WAAW,CAAC,QAAQ,EAAE,kBAAkB,GAAG,OAAO,CAAC,KAAK,CAAC;IAQzD,SAAS,CAAC,QAAQ,CAAC,EAAE,gBAAgB,GAAG,OAAO,CAAC,KAAK,EAAE,CAAC;IAQ9D;;;;;OAKG;IACH,OAAO,CAAC,mBAAmB;IAkE3B;;;;;OAKG;IACH,OAAO,CAAC,aAAa;IAcrB;;;;;OAKG;IACH,OAAO,CAAC,QAAQ;IAchB;;;;;;;;OAQG;IACH,OAAO,CAAC,iCAAiC;IAkBzC;;;;;;;;OAQG;IACH,OAAO,CAAC,eAAe;IAiBvB;;;OAGG;IACH,OAAO,CAAC,iBAAiB;IAiBzB;;;;;OAKG;IACH;;;;;OAKG;IAEH,OAAO,CAAC,QAAQ;CAgDjB"}

package/dist/shared/reasoning.d.ts

@@ -0,0 +1,19 @@
+/**
+ * Reasoning-text helpers shared by the chat adapters and continuation.
+ */
+/**
+ * Reasoning text an OpenAI-compatible server returned beside the answer.
+ * vLLM's reasoning parsers emit `reasoning_content` (older releases) or
+ * `reasoning` (newer releases) on the message and on stream deltas; SGLang and
+ * several gateways use the same names. Only string values count, so a
+ * structured `reasoning` object from another gateway is ignored.
+ */
+export declare function readReasoningText(source: unknown): string | undefined;
+/**
+ * Deliver accumulated stream reasoning to `ChatOptions.onReasoning`. Called
+ * from a stream's `finally`, so it also runs when the stream fails, is
+ * aborted, or the consumer stops early; a throwing callback is swallowed so it
+ * can neither break the stream nor mask the stream's own error.
+ */
+export declare function reportReasoning(callback: ((reasoning: string) => void) | undefined, reasoning: string): void;
+//# sourceMappingURL=reasoning.d.ts.map

package/dist/shared/reasoning.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"reasoning.d.ts","sourceRoot":"","sources":["../../src/shared/reasoning.ts"],"names":[],"mappings":"AAAA;;GAEG;AAEH;;;;;;GAMG;AACH,wBAAgB,iBAAiB,CAAC,MAAM,EAAE,OAAO,GAAG,MAAM,GAAG,SAAS,CAQrE;AAED;;;;;GAKG;AACH,wBAAgB,eAAe,CAC7B,QAAQ,EAAE,CAAC,CAAC,SAAS,EAAE,MAAM,KAAK,IAAI,CAAC,GAAG,SAAS,EACnD,SAAS,EAAE,MAAM,GAChB,IAAI,CAON"}

package/dist/shared/safety.d.ts

@@ -1,6 +1,18 @@
 import type { AIGenerationLimits, AIInterface, AIReasoningOptions, AIRequestControls, AIRequestEvent, AIRequestOperation, AIRequestStatus, BaseAIOptions, ChatOptions, ImageGenerationOptions } from "./types.js";
 export declare const DEFAULT_AI_TIMEOUT_MS = 120000;
 export declare const DEFAULT_AI_MAX_RETRIES = 0;
+/**
+ * Internal per-call channel from a provider back to the observed client: a
+ * provider that transparently resends a request (the OpenAI-compatible
+ * stale-connection retry, #1467) bumps `connectionRetries` so the request
+ * event's `attempts` reflects it. Enumerable so option spreads keep it.
+ * @internal
+ */
+export declare const AI_ATTEMPT_RECORDER: unique symbol;
+/** @internal */
+export interface AIAttemptRecorder {
+    connectionRetries: number;
+}
 export declare const DEFAULT_AI_GENERATION_LIMITS: Readonly<AIGenerationLimits>;
 export interface NormalizedBaseAIOptions extends BaseAIOptions {
     timeout: number;

package/dist/shared/safety.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"safety.d.ts","sourceRoot":"","sources":["../../src/shared/safety.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EACV,kBAAkB,EAClB,WAAW,EACX,kBAAkB,EAClB,iBAAiB,EACjB,cAAc,EACd,kBAAkB,EAClB,eAAe,EACf,aAAa,EACb,WAAW,EACX,sBAAsB,EACvB,MAAM,YAAS,CAAC;AAGjB,eAAO,MAAM,qBAAqB,SAAU,CAAC;AAC7C,eAAO,MAAM,sBAAsB,IAAI,CAAC;AAExC,eAAO,MAAM,4BAA4B,EAAE,QAAQ,CAAC,kBAAkB,CAWlE,CAAC;AAEL,MAAM,WAAW,uBAAwB,SAAQ,aAAa;IAC5D,OAAO,EAAE,MAAM,CAAC;IAChB,UAAU,EAAE,MAAM,CAAC;IACnB,gBAAgB,EAAE,kBAAkB,CAAC;CACtC;AAED,MAAM,WAAW,uBAAuB;IACtC,MAAM,EAAE,WAAW,CAAC;IACpB,OAAO,EAAE,MAAM,CAAC;IAChB,OAAO,IAAI,IAAI,CAAC;IAChB,UAAU,IAAI,OAAO,CAAC;CACvB;AAiBD,wBAAgB,sBAAsB,CAAC,CAAC,SAAS,aAAa,EAC5D,OAAO,EAAE,CAAC,GACT,CAAC,GAAG,uBAAuB,CAwD7B;AAiDD,wBAAgB,oBAAoB,CAClC,CAAC,SAAS,iBAAiB,GAAG;IAC5B,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,aAAa,CAAC,EAAE,WAAW,CAAC,eAAe,CAAC,CAAC;IAC7C,eAAe,CAAC,EAAE,OAAO,CAAC;CAC3B,EAED,eAAe,EAAE,aAAa,EAC9B,OAAO,EAAE,CAAC,EACV,QAAQ,CAAC,EAAE,MAAM,EACjB,KAAK,CAAC,EAAE,MAAM,GACb,CAAC,GAAG;IAAE,SAAS,EAAE,MAAM,CAAC;IAAC,SAAS,EAAE,kBAAkB,CAAA;CAAE,CA8C1D;AAED,wBAAgB,+BAA+B,CAC7C,CAAC,SAAS,sBAAsB,EAEhC,eAAe,EAAE,aAAa,EAC9B,OAAO,EAAE,CAAC,EACV,QAAQ,CAAC,EAAE,MAAM,EACjB,KAAK,CAAC,EAAE,MAAM,GACb,CAAC,GAAG;IAAE,CAAC,EAAE,MAAM,CAAA;CAAE,CAcnB;AAED,wBAAgB,sBAAsB,CACpC,eAAe,EAAE,aAAa,EAC9B,cAAc,GAAE,iBAAsB,GACrC,uBAAuB,CAiCzB;AAWD,wBAAgB,gBAAgB,CAC9B,eAAe,EAAE,aAAa,EAC9B,KAAK,EAAE,IAAI,CAAC,cAAc,EAAE,MAAM,CAAC,GAAG;IACpC,IAAI,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;CAC/B,GACA,IAAI,CAUN;AAED,wBAAgB,sBAAsB,CACpC,KAAK,EAAE,OAAO,EACd,UAAU,UAAQ,GACjB;IAAE,MAAM,EAAE,eAAe,CAAC;IAAC,SAAS,CAAC,EAAE,MAAM,CAAA;CAAE,CAuBjD;AAED,wBAAgB,gBAAgB,CAAC,OAAO,EAAE;IACxC,eAAe,EAAE,aAAa,CAAC;IAC/B,QAAQ,EAAE,MAAM,CAAC;IACjB,KAAK,EAAE,MAAM,CAAC;IACd,SAAS,EAAE,kBAAkB,CAAC;IAC9B,WAAW,CAAC,EAAE;QAAE,SAAS,CAAC,EAAE,MAAM,CAAC;QAAC,SAAS,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAA;KAAE,CAAC;IACzE,wBAAwB,CAAC,EAAE,MAAM,CAAC;CACnC,GAAG,IAAI,CAAC,cAAc,EAAE,QAAQ,GAAG,UAAU,CAAC,GAAG;IAChD,IAAI,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;CAC/B,CAyBA;AA0GD;;;;GAIG;AACH,wBAAgB,gBAAgB,CAAC,CAAC,SAAS,WAAW,EACpD,MAAM,EAAE,CAAC,EACT,eAAe,EAAE,aAAa,GAC7B,CAAC,CA8GH"}
+{"version":3,"file":"safety.d.ts","sourceRoot":"","sources":["../../src/shared/safety.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EACV,kBAAkB,EAClB,WAAW,EACX,kBAAkB,EAClB,iBAAiB,EACjB,cAAc,EACd,kBAAkB,EAClB,eAAe,EACf,aAAa,EACb,WAAW,EACX,sBAAsB,EACvB,MAAM,YAAS,CAAC;AAGjB,eAAO,MAAM,qBAAqB,SAAU,CAAC;AAC7C,eAAO,MAAM,sBAAsB,IAAI,CAAC;AAExC;;;;;;GAMG;AACH,eAAO,MAAM,mBAAmB,eAE/B,CAAC;AAEF,gBAAgB;AAChB,MAAM,WAAW,iBAAiB;IAChC,iBAAiB,EAAE,MAAM,CAAC;CAC3B;AA6BD,eAAO,MAAM,4BAA4B,EAAE,QAAQ,CAAC,kBAAkB,CAWlE,CAAC;AAEL,MAAM,WAAW,uBAAwB,SAAQ,aAAa;IAC5D,OAAO,EAAE,MAAM,CAAC;IAChB,UAAU,EAAE,MAAM,CAAC;IACnB,gBAAgB,EAAE,kBAAkB,CAAC;CACtC;AAED,MAAM,WAAW,uBAAuB;IACtC,MAAM,EAAE,WAAW,CAAC;IACpB,OAAO,EAAE,MAAM,CAAC;IAChB,OAAO,IAAI,IAAI,CAAC;IAChB,UAAU,IAAI,OAAO,CAAC;CACvB;AAiBD,wBAAgB,sBAAsB,CAAC,CAAC,SAAS,aAAa,EAC5D,OAAO,EAAE,CAAC,GACT,CAAC,GAAG,uBAAuB,CAwD7B;AAiDD,wBAAgB,oBAAoB,CAClC,CAAC,SAAS,iBAAiB,GAAG;IAC5B,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,aAAa,CAAC,EAAE,WAAW,CAAC,eAAe,CAAC,CAAC;IAC7C,eAAe,CAAC,EAAE,OAAO,CAAC;CAC3B,EAED,eAAe,EAAE,aAAa,EAC9B,OAAO,EAAE,CAAC,EACV,QAAQ,CAAC,EAAE,MAAM,EACjB,KAAK,CAAC,EAAE,MAAM,GACb,CAAC,GAAG;IAAE,SAAS,EAAE,MAAM,CAAC;IAAC,SAAS,EAAE,kBAAkB,CAAA;CAAE,CA8C1D;AAED,wBAAgB,+BAA+B,CAC7C,CAAC,SAAS,sBAAsB,EAEhC,eAAe,EAAE,aAAa,EAC9B,OAAO,EAAE,CAAC,EACV,QAAQ,CAAC,EAAE,MAAM,EACjB,KAAK,CAAC,EAAE,MAAM,GACb,CAAC,GAAG;IAAE,CAAC,EAAE,MAAM,CAAA;CAAE,CAcnB;AAED,wBAAgB,sBAAsB,CACpC,eAAe,EAAE,aAAa,EAC9B,cAAc,GAAE,iBAAsB,GACrC,uBAAuB,CAiCzB;AAWD,wBAAgB,gBAAgB,CAC9B,eAAe,EAAE,aAAa,EAC9B,KAAK,EAAE,IAAI,CAAC,cAAc,EAAE,MAAM,CAAC,GAAG;IACpC,IAAI,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;CAC/B,GACA,IAAI,CAUN;AAED,wBAAgB,sBAAsB,CACpC,KAAK,EAAE,OAAO,EACd,UAAU,UAAQ,GACjB;IAAE,MAAM,EAAE,eAAe,CAAC;IAAC,SAAS,CAAC,EAAE,MAAM,CAAA;CAAE,CAuBjD;AAED,wBAAgB,gBAAgB,CAAC,OAAO,EAAE;IACxC,eAAe,EAAE,aAAa,CAAC;IAC/B,QAAQ,EAAE,MAAM,CAAC;IACjB,KAAK,EAAE,MAAM,CAAC;IACd,SAAS,EAAE,kBAAkB,CAAC;IAC9B,WAAW,CAAC,EAAE;QAAE,SAAS,CAAC,EAAE,MAAM,CAAC;QAAC,SAAS,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAA;KAAE,CAAC;IACzE,wBAAwB,CAAC,EAAE,MAAM,CAAC;CACnC,GAAG,IAAI,CAAC,cAAc,EAAE,QAAQ,GAAG,UAAU,CAAC,GAAG;IAChD,IAAI,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;CAC/B,CAyBA;AA0GD;;;;GAIG;AACH,wBAAgB,gBAAgB,CAAC,CAAC,SAAS,WAAW,EACpD,MAAM,EAAE,CAAC,EACT,eAAe,EAAE,aAAa,GAC7B,CAAC,CAwHH"}

package/dist/shared/types.d.ts

@@ -11,8 +11,13 @@ export type GeminiThinkingLevel = 'minimal' | 'low' | 'medium' | 'high';
  *
  * Provider instances merge partial overrides with the exported safe defaults.
  */
-/** Why a generation stopped, normalised across providers. */
-export type AIFinishReason = 'stop' | 'length' | 'tool_calls' | 'content_filter';
+/**
+ * Why a generation stopped, normalised across providers. `'other'` means the
+ * server reported a reason outside this set (e.g. vLLM `abort`): treat the
+ * reply as possibly incomplete. Only the OpenAI-compatible providers report
+ * it today; see {@link AIResponse.rawFinishReason} for the server's value.
+ */
+export type AIFinishReason = 'stop' | 'length' | 'tool_calls' | 'content_filter' | 'other';
 /**
  * Automatic continuation when a reply stops on its output limit. `true`
  * allows up to 3 extra requests; pass `{ maxContinuations }` to change that.
@@ -327,6 +332,30 @@ export interface ChatOptions extends AIRequestControls {
      * finish reason. Used for continuation; also useful to detect truncation.
      */
     onFinishReason?: (reason: AIFinishReason) => void;
+    /**
+     * Called by streaming adapters once when the stream ends, with the
+     * reasoning text accumulated from its reasoning deltas (vLLM
+     * `reasoning_content`/`reasoning`, Ollama `thinking`). It also runs when
+     * the stream fails, is aborted, or the consumer stops iterating early, with
+     * whatever arrived first, so it is a diagnostic, not a completion signal.
+     * Not called when no reasoning arrived. Errors it throws are swallowed.
+     * `chat()` returns the same text as {@link AIResponse.reasoning}.
+     */
+    onReasoning?: (reasoning: string) => void;
+    /**
+     * Extra chat-template arguments sent as `chat_template_kwargs` by the
+     * OpenAI-compatible providers (`openai`, `litellm`, `bifrost`), for servers
+     * such as vLLM and SGLang that apply them to the model's chat template, e.g.
+     * `{ enable_thinking: false }` to switch thinking off for one request.
+     * Must be a plain JSON object (at most 8192 bytes serialized, 4 levels
+     * deep, 32 entries per object or array); invalid values throw
+     * `AI_INVALID_CHAT_TEMPLATE_KWARGS` before any request is sent (the error
+     * locates the entry by position, never by key name or value). An empty
+     * object is not sent. The `openai` provider throws `AI_UNSUPPORTED_OPTION`
+     * when its resolved base URL is the hosted `api.openai.com`, which does not
+     * accept the field. Other providers ignore it.
+     */
+    chatTemplateKwargs?: Record<string, unknown>;
     /**
      * Continue a reply that stopped on its output limit and return the stitched
      * text. Off by default; ignored for tools and JSON output. Each part is a
@@ -1193,9 +1222,25 @@ export interface AIResponse {
      */
     model?: string;
     /**
-     * Finish reason
+     * Why generation stopped. The OpenAI-compatible providers always set it:
+     * an absent server value is `'stop'`, an unrecognised one is `'other'`, and
+     * a reply that carries tool calls but reports `'stop'` is `'tool_calls'`
+     * (`'length'` and `'content_filter'` are kept as reported).
      */
     finishReason?: AIFinishReason;
+    /**
+     * The server's own finish reason when `finishReason` is `'other'`, e.g.
+     * `'abort'` from vLLM or legacy `'function_call'`.
+     */
+    rawFinishReason?: string;
+    /**
+     * Reasoning ("thinking") text the server returned separately from
+     * `content`: vLLM/SGLang `reasoning_content` or `reasoning` on
+     * OpenAI-compatible providers, `thinking` on Ollama. Absent when the server
+     * sent none. A reply whose `content` is empty but whose `reasoning` is not
+     * usually means the answer was parsed into the reasoning channel.
+     */
+    reasoning?: string;
     /**
      * Tool calls made by the model
      */
@@ -1736,6 +1781,17 @@ export interface OpenAIOptions extends BaseAIOptions {
     apiKey?: string;
     baseUrl?: string;
     organization?: string;
+    /**
+     * Retry once, immediately, when a request fails on a stale keep-alive
+     * connection (ECONNRESET, EPIPE, `UND_ERR_SOCKET`, "socket hang up", or the
+     * OpenAI SDK's `APIConnectionError`): `chat()` only before response
+     * headers arrive; `stream()` until the first chunk is yielded, which may
+     * re-run a generation the server had started. Independent of `maxRetries`;
+     * never retries after an HTTP status, a timeout, a caller abort, or a
+     * streamed chunk.
+     * @default true
+     */
+    retryConnectionErrors?: boolean;
 }
 /**
  * Protected local session surface consumed by the ChatGPT plan provider.
@@ -1774,6 +1830,17 @@ export interface LiteLLMOptions extends BaseAIOptions {
     apiKey?: string;
     baseUrl?: string;
     organization?: string;
+    /**
+     * Retry once, immediately, when a request fails on a stale keep-alive
+     * connection (ECONNRESET, EPIPE, `UND_ERR_SOCKET`, "socket hang up", or the
+     * OpenAI SDK's `APIConnectionError`): `chat()` only before response
+     * headers arrive; `stream()` until the first chunk is yielded, which may
+     * re-run a generation the server had started. Independent of `maxRetries`;
+     * never retries after an HTTP status, a timeout, a caller abort, or a
+     * streamed chunk.
+     * @default true
+     */
+    retryConnectionErrors?: boolean;
     adminApiKey?: string;
     adminBaseUrl?: string;
     adminUrl?: string;
@@ -1790,6 +1857,17 @@ export interface BifrostOptions extends BaseAIOptions {
     apiKey?: string;
     baseUrl?: string;
     organization?: string;
+    /**
+     * Retry once, immediately, when a request fails on a stale keep-alive
+     * connection (ECONNRESET, EPIPE, `UND_ERR_SOCKET`, "socket hang up", or the
+     * OpenAI SDK's `APIConnectionError`): `chat()` only before response
+     * headers arrive; `stream()` until the first chunk is yielded, which may
+     * re-run a generation the server had started. Independent of `maxRetries`;
+     * never retries after an HTTP status, a timeout, a caller abort, or a
+     * streamed chunk.
+     * @default true
+     */
+    retryConnectionErrors?: boolean;
     /**
      * Optional virtual key for admin routes. Bifrost OSS admin APIs typically use
      * username/password Basic auth instead; use `adminUser` / `adminPassword`