@librechat/agents 3.3.0 → 3.3.2

package/dist/esm/llm/bedrock/index.mjs.map

@@ -1 +1 @@
-{"version":3,"file":"index.mjs","names":[],"sources":["../../../../src/llm/bedrock/index.ts"],"sourcesContent":["/**\n * Optimized ChatBedrockConverse wrapper that fixes content block merging for\n * streaming responses and adds support for latest @langchain/aws features:\n *\n * - Application Inference Profiles (PR #9129)\n * - Service Tiers (Priority/Standard/Flex) (PR #9785) - requires AWS SDK 3.966.0+\n *\n * Bedrock's `@langchain/aws` library does not include an `index` property on content\n * blocks (unlike Anthropic/OpenAI), which causes LangChain's `_mergeLists` to append\n * each streaming chunk as a separate array entry instead of merging by index.\n *\n * This wrapper takes full ownership of the stream by directly interfacing with the\n * AWS SDK client (`this.client`) and using custom handlers from `./utils/` that\n * include `contentBlockIndex` in response_metadata for every delta type. It then\n * promotes `contentBlockIndex` to an `index` property on each content block\n * (mirroring Anthropic's pattern) and strips it from metadata to avoid\n * `_mergeDicts` conflicts.\n *\n * When multiple content block types are present (e.g. reasoning + text), text deltas\n * are promoted from strings to array form with `index` so they merge correctly once\n * the accumulated content is already an array.\n */\n\nimport { ChatBedrockConverse } from '@langchain/aws';\nimport { AIMessageChunk } from '@langchain/core/messages';\nimport { ChatGenerationChunk, ChatResult } from '@langchain/core/outputs';\nimport {\n  ConverseStreamCommand,\n  type GuardrailConfiguration,\n  type GuardrailStreamConfiguration,\n} from '@aws-sdk/client-bedrock-runtime';\nimport type { CallbackManagerForLLMRun } from '@langchain/core/callbacks/manager';\nimport type { BaseMessage, ResponseMetadata } from '@langchain/core/messages';\nimport type { ChatBedrockConverseInput } from '@langchain/aws';\nimport {\n  convertToConverseMessages,\n  createConverseToolUseStopChunk,\n  handleConverseStreamContentBlockStart,\n  handleConverseStreamContentBlockDelta,\n  handleConverseStreamMetadata,\n} from './utils';\nimport type { ContentBlockDeltaEvent } from './types';\nimport {\n  resolveBedrockPromptCacheTtl,\n  supportsBedrockToolCache,\n  type PromptCacheTtl,\n} from '@/messages/cache';\nimport { applyCachePointsToConversePayload } from './cachePoints';\nimport { insertBedrockToolCachePoint } from './toolCache';\n\n/**\n * Service tier type for Bedrock invocations.\n * Requires AWS SDK >= 3.966.0 to actually work.\n * @see https://docs.aws.amazon.com/bedrock/latest/userguide/service-tiers-inference.html\n */\nexport type ServiceTierType = 'priority' | 'default' | 'flex' | 'reserved';\n\nexport type CustomGuardrailConfiguration = GuardrailConfiguration &\n  Pick<GuardrailStreamConfiguration, 'streamProcessingMode'>;\n\nconst MAX_STREAM_QUEUE_CHUNKS = 256;\nconst MAX_STREAM_QUEUE_TEXT_CHARS = 8192;\nconst STREAM_CHUNK_MIN_SIZE = 4;\nconst STREAM_BOUNDARIES = new Set([' ', '.', ',', '!', '?', ';', ':']);\n\ntype QueuedGenerationChunk = {\n  chunk: ChatGenerationChunk;\n  callbackChunk?: ChatGenerationChunk;\n  callbackToken: string;\n  smooth: boolean;\n  textLength: number;\n};\n\nfunction findStreamChunkBoundary(text: string, minSize: number): number {\n  if (minSize >= text.length) {\n    return text.length;\n  }\n\n  for (let position = minSize; position < text.length; position++) {\n    if (STREAM_BOUNDARIES.has(text[position])) {\n      return position + 1;\n    }\n  }\n\n  return text.length;\n}\n\nfunction splitStreamToken(text: string): string[] {\n  const chunks: string[] = [];\n  let currentIndex = 0;\n\n  while (currentIndex < text.length) {\n    const remainingText = text.slice(currentIndex);\n    const chunkSize = findStreamChunkBoundary(\n      remainingText,\n      STREAM_CHUNK_MIN_SIZE\n    );\n    chunks.push(text.slice(currentIndex, currentIndex + chunkSize));\n    currentIndex += chunkSize;\n  }\n\n  return chunks;\n}\n\nfunction getCadencedStreamDelay({\n  targetDelay,\n  lastVisibleContentAt,\n  now,\n}: {\n  targetDelay: number;\n  lastVisibleContentAt?: number;\n  now: number;\n}): number {\n  if (targetDelay <= 0 || lastVisibleContentAt == null) {\n    return 0;\n  }\n  return Math.max(0, targetDelay - (now - lastVisibleContentAt));\n}\n\nasync function waitForStreamDelay(\n  delay: number,\n  signal?: AbortSignal\n): Promise<void> {\n  if (delay <= 0 || isSignalAborted(signal)) {\n    return;\n  }\n  await new Promise<void>((resolve) => {\n    const timeoutRef: { current?: ReturnType<typeof setTimeout> } = {};\n    const onAbort = (): void => {\n      if (timeoutRef.current) {\n        clearTimeout(timeoutRef.current);\n      }\n      signal?.removeEventListener('abort', onAbort);\n      resolve();\n    };\n    timeoutRef.current = setTimeout(() => {\n      signal?.removeEventListener('abort', onAbort);\n      resolve();\n    }, delay);\n    signal?.addEventListener('abort', onAbort, { once: true });\n    if (isSignalAborted(signal)) {\n      onAbort();\n    }\n  });\n}\n\nfunction isSignalAborted(signal?: AbortSignal): boolean {\n  return signal?.aborted === true;\n}\n\n/**\n * Extended input interface with additional features:\n * - applicationInferenceProfile: Use an inference profile ARN instead of model ID\n * - serviceTier: Specify service tier (Priority, Standard, Flex, Reserved)\n */\nexport interface CustomChatBedrockConverseInput\n  extends ChatBedrockConverseInput {\n  /**\n   * Enables Bedrock prompt cache checkpoints for message and tool prefixes.\n   */\n  promptCache?: boolean;\n\n  /**\n   * Prompt-cache checkpoint TTL. Defaults to `'1h'` (extended cache) when\n   * `promptCache` is enabled; set `'5m'` for the legacy 5-minute behavior.\n   * Bedrock models that don't support the 1-hour TTL downgrade to 5m\n   * server-side (verified on Sonnet/Opus 4.6), so the default is safe to leave\n   * on; use `'5m'` for any model that rejects it.\n   */\n  promptCacheTtl?: PromptCacheTtl;\n\n  /**\n   * Minimum delay in milliseconds between visible streamed content deltas.\n   */\n  _lc_stream_delay?: number;\n\n  /**\n   * Guardrail configuration for Converse and ConverseStream invocations.\n   * `streamProcessingMode` is only used by ConverseStream.\n   */\n  guardrailConfig?: CustomGuardrailConfiguration;\n\n  /**\n   * Application Inference Profile ARN to use for the model.\n   * For example, \"arn:aws:bedrock:eu-west-1:123456789102:application-inference-profile/fm16bt65tzgx\"\n   * When provided, this ARN will be used for the actual inference calls instead of the model ID.\n   * Must still provide `model` as normal modelId to benefit from all the metadata.\n   * @see https://docs.aws.amazon.com/bedrock/latest/userguide/inference-profiles-create.html\n   */\n  applicationInferenceProfile?: string;\n\n  /**\n   * Service tier for model invocation.\n   * Specifies the processing tier type used for serving the request.\n   * Supported values are 'priority', 'default', 'flex', and 'reserved'.\n   *\n   * - 'priority': Prioritized processing for lower latency\n   * - 'default': Standard processing tier\n   * - 'flex': Flexible processing tier with lower cost\n   * - 'reserved': Reserved capacity for consistent performance\n   *\n   * If not provided, AWS uses the default tier.\n   * Note: Requires AWS SDK >= 3.966.0 to work.\n   * @see https://docs.aws.amazon.com/bedrock/latest/userguide/service-tiers-inference.html\n   */\n  serviceTier?: ServiceTierType;\n}\n\n/**\n * Extended call options with serviceTier override support.\n */\nexport interface CustomChatBedrockConverseCallOptions {\n  serviceTier?: ServiceTierType;\n  guardrailConfig?: CustomGuardrailConfiguration;\n}\n\nexport class CustomChatBedrockConverse extends ChatBedrockConverse {\n  _lc_stream_delay: number;\n\n  /**\n   * Whether to insert Bedrock prompt cache checkpoints when available.\n   */\n  promptCache?: boolean;\n\n  /**\n   * Prompt-cache checkpoint TTL (`'5m'` legacy or `'1h'` extended cache).\n   */\n  promptCacheTtl?: PromptCacheTtl;\n\n  /**\n   * Application Inference Profile ARN to use instead of model ID.\n   */\n  applicationInferenceProfile?: string;\n\n  /**\n   * Service tier for model invocation.\n   */\n  serviceTier?: ServiceTierType;\n\n  /**\n   * The configured model id, captured at construction so it survives the\n   * temporary `this.model` swap to an application-inference-profile ARN during\n   * generation. Used to gate the Bedrock tool cache point to Claude models\n   * (see {@link supportsBedrockToolCache}).\n   */\n  private readonly cacheModelId: string;\n\n  constructor(fields?: CustomChatBedrockConverseInput) {\n    super(fields);\n    this.promptCache = fields?.promptCache;\n    this.promptCacheTtl = fields?.promptCacheTtl;\n    this._lc_stream_delay = Math.max(0, fields?._lc_stream_delay ?? 0);\n    this.applicationInferenceProfile = fields?.applicationInferenceProfile;\n    this.serviceTier = fields?.serviceTier;\n    // `super(fields)` initializes `this.model` to LangChain's default Claude\n    // model when `fields.model` is omitted, so fall back to it rather than ''\n    // (which would treat the default Claude model as tool-cache-unsupported).\n    this.cacheModelId = fields?.model ?? this.model;\n  }\n\n  static lc_name(): string {\n    return 'LibreChatBedrockConverse';\n  }\n\n  /**\n   * Get the model ID to use for API calls.\n   * Returns applicationInferenceProfile if set, otherwise returns this.model.\n   */\n  protected getModelId(): string {\n    return this.applicationInferenceProfile ?? this.model;\n  }\n\n  /**\n   * Override invocationParams to add serviceTier support.\n   */\n  override invocationParams(\n    options?: this['ParsedCallOptions'] & CustomChatBedrockConverseCallOptions\n  ): ReturnType<ChatBedrockConverse['invocationParams']> & {\n    serviceTier?: { type: ServiceTierType };\n  } {\n    const baseParams = super.invocationParams(options);\n    const toolConfig =\n      this.promptCache === true && supportsBedrockToolCache(this.cacheModelId)\n        ? insertBedrockToolCachePoint(\n            baseParams.toolConfig,\n            true,\n            resolveBedrockPromptCacheTtl(this.promptCacheTtl, this.cacheModelId)\n          )\n        : baseParams.toolConfig;\n\n    /** Service tier from options or fall back to class-level setting */\n    const serviceTierType = options?.serviceTier ?? this.serviceTier;\n\n    return {\n      ...baseParams,\n      toolConfig,\n      serviceTier: serviceTierType ? { type: serviceTierType } : undefined,\n    };\n  }\n\n  /**\n   * Override _generateNonStreaming to use applicationInferenceProfile as modelId.\n   * Uses the same model-swapping pattern as streaming for consistency.\n   */\n  override async _generateNonStreaming(\n    messages: BaseMessage[],\n    options: this['ParsedCallOptions'] & CustomChatBedrockConverseCallOptions,\n    runManager?: CallbackManagerForLLMRun\n  ): Promise<ChatResult> {\n    const originalModel = this.model;\n    if (\n      this.applicationInferenceProfile != null &&\n      this.applicationInferenceProfile !== ''\n    ) {\n      this.model = this.applicationInferenceProfile;\n    }\n\n    try {\n      return await super._generateNonStreaming(messages, options, runManager);\n    } finally {\n      this.model = originalModel;\n    }\n  }\n\n  /**\n   * Own the stream end-to-end so we have direct access to every\n   * `contentBlockDelta.contentBlockIndex` from the AWS SDK.\n   *\n   * This replaces the parent's implementation which strips contentBlockIndex\n   * from text and reasoning deltas, making it impossible to merge correctly.\n   */\n  override async *_streamResponseChunks(\n    messages: BaseMessage[],\n    options: this['ParsedCallOptions'] & CustomChatBedrockConverseCallOptions,\n    runManager?: CallbackManagerForLLMRun\n  ): AsyncGenerator<ChatGenerationChunk> {\n    const { converseMessages, converseSystem } =\n      convertToConverseMessages(messages);\n    const params = this.invocationParams(options);\n\n    let { streamUsage } = this;\n    if ((options as Record<string, unknown>).streamUsage !== undefined) {\n      streamUsage = (options as Record<string, unknown>).streamUsage as boolean;\n    }\n\n    const modelId = this.getModelId();\n\n    applyCachePointsToConversePayload({\n      cacheControl: options.cache_control,\n      system: converseSystem,\n      messages: converseMessages,\n      params,\n      modelId,\n    });\n\n    const command = new ConverseStreamCommand({\n      modelId,\n      messages: converseMessages,\n      system: converseSystem,\n      ...(params as Record<string, unknown>),\n    });\n\n    const streamAbortController = new AbortController();\n    const abortStream = (): void => streamAbortController.abort();\n    options.signal?.addEventListener('abort', abortStream, { once: true });\n    if (isSignalAborted(options.signal)) {\n      abortStream();\n    }\n\n    try {\n      const response = await this.client.send(command, {\n        abortSignal: streamAbortController.signal,\n      });\n\n      const stream = response.stream;\n      if (!stream) {\n        return;\n      }\n\n      const seenBlockIndices = new Set<number>();\n      const toolUseBlockIndices = new Set<number>();\n      const queuedChunks: QueuedGenerationChunk[] = [];\n      const producerState: { done: boolean; error?: unknown } = { done: false };\n      let queuedChunkIndex = 0;\n      let bufferedTextLength = 0;\n      let consumerClosed = false;\n      let notifyConsumer: (() => void) | undefined;\n      let notifyProducer: (() => void) | undefined;\n      let hasEmittedVisibleContent = false;\n      let lastVisibleContentAt: number | undefined;\n\n      /**\n       * Guardrails can reject an already-streamed toolUse block at\n       * `messageStop` (`guardrail_intervened`), after `contentBlockStop` has\n       * passed. Only emit eager-execution seals when no guardrails are\n       * configured, so a later intervention can't race an eagerly started tool.\n       */\n      const sealToolUseOnStop =\n        options.guardrailConfig == null && this.guardrailConfig == null;\n\n      const notifyConsumerForChunk = (): void => {\n        notifyConsumer?.();\n        notifyConsumer = undefined;\n      };\n\n      const notifyProducerForSpace = (): void => {\n        notifyProducer?.();\n        notifyProducer = undefined;\n      };\n\n      const hasQueuedChunks = (): boolean =>\n        queuedChunkIndex < queuedChunks.length;\n\n      const getQueuedChunkCount = (): number =>\n        queuedChunks.length - queuedChunkIndex;\n\n      const isQueueAtCapacity = (): boolean =>\n        getQueuedChunkCount() >= MAX_STREAM_QUEUE_CHUNKS ||\n        bufferedTextLength >= MAX_STREAM_QUEUE_TEXT_CHARS;\n\n      const waitForNextChunk = async (): Promise<void> => {\n        if (\n          hasQueuedChunks() ||\n          producerState.done ||\n          producerState.error != null\n        ) {\n          return;\n        }\n        await new Promise<void>((resolve) => {\n          notifyConsumer = resolve;\n        });\n      };\n\n      const waitForQueueSpace = async (): Promise<void> => {\n        while (\n          isQueueAtCapacity() &&\n          !consumerClosed &&\n          !isSignalAborted(options.signal)\n        ) {\n          await new Promise<void>((resolve) => {\n            const signal = options.signal;\n            const onAbort = (): void => {\n              signal?.removeEventListener('abort', onAbort);\n              resolve();\n            };\n            const onSpace = (): void => {\n              signal?.removeEventListener('abort', onAbort);\n              resolve();\n            };\n            notifyProducer = onSpace;\n            signal?.addEventListener('abort', onAbort, { once: true });\n            if (isSignalAborted(signal)) {\n              onAbort();\n            }\n          });\n        }\n      };\n\n      const dequeue = (): QueuedGenerationChunk | undefined => {\n        if (!hasQueuedChunks()) {\n          return undefined;\n        }\n        const queuedChunk = queuedChunks[queuedChunkIndex];\n        queuedChunkIndex++;\n        if (\n          queuedChunkIndex > 128 &&\n          queuedChunkIndex * 2 >= queuedChunks.length\n        ) {\n          queuedChunks.splice(0, queuedChunkIndex);\n          queuedChunkIndex = 0;\n        }\n        return queuedChunk;\n      };\n\n      const enqueue = async (\n        queuedChunk: QueuedGenerationChunk\n      ): Promise<void> => {\n        await waitForQueueSpace();\n        if (consumerClosed || isSignalAborted(options.signal)) {\n          abortStream();\n          throw new Error('AbortError: User aborted the request.');\n        }\n        queuedChunks.push(queuedChunk);\n        if (queuedChunk.smooth) {\n          bufferedTextLength += queuedChunk.textLength;\n        }\n        notifyConsumerForChunk();\n      };\n\n      const enqueueChunk = async ({\n        chunk,\n        callbackChunk,\n        callbackToken = '',\n        smooth = false,\n        textLength = 0,\n      }: {\n        chunk: ChatGenerationChunk;\n        callbackChunk?: ChatGenerationChunk;\n        callbackToken?: string;\n        smooth?: boolean;\n        textLength?: number;\n      }): Promise<void> => {\n        await enqueue({\n          chunk,\n          callbackChunk,\n          callbackToken,\n          smooth,\n          textLength: smooth ? textLength : 0,\n        });\n      };\n\n      const enqueueDelta = async (\n        contentBlockDelta: ContentBlockDeltaEvent\n      ): Promise<void> => {\n        const delta = contentBlockDelta.delta;\n        if (delta == null) {\n          throw new Error('No delta found in content block.');\n        }\n\n        const idx = contentBlockDelta.contentBlockIndex;\n        if (idx != null) {\n          seenBlockIndices.add(idx);\n        }\n\n        const text = delta.text;\n        const reasoningContent = delta.reasoningContent;\n        const reasoningText = reasoningContent?.text;\n        const visibleText =\n          typeof text === 'string'\n            ? text\n            : typeof reasoningText === 'string'\n              ? reasoningText\n              : '';\n        const smooth = this._lc_stream_delay > 0 && visibleText !== '';\n        const tokenChunks = smooth\n          ? splitStreamToken(visibleText)\n          : [visibleText];\n\n        for (const token of tokenChunks) {\n          let splitDelta = contentBlockDelta;\n          if (typeof text === 'string') {\n            splitDelta = {\n              ...contentBlockDelta,\n              delta: { text: token },\n            };\n          } else if (\n            typeof reasoningText === 'string' &&\n            reasoningContent != null\n          ) {\n            splitDelta = {\n              ...contentBlockDelta,\n              delta: {\n                reasoningContent: {\n                  ...reasoningContent,\n                  text: token,\n                },\n              },\n            };\n          }\n\n          const deltaChunk = handleConverseStreamContentBlockDelta(splitDelta);\n          await enqueueChunk({\n            chunk: this.enrichChunk(deltaChunk, seenBlockIndices),\n            callbackChunk: deltaChunk,\n            callbackToken: deltaChunk.text,\n            smooth,\n            textLength: token.length,\n          });\n        }\n      };\n\n      const producer = (async (): Promise<void> => {\n        try {\n          for await (const event of stream) {\n            if (isSignalAborted(options.signal)) {\n              abortStream();\n              throw new Error('AbortError: User aborted the request.');\n            }\n\n            if (event.contentBlockStart != null) {\n              const startChunk = handleConverseStreamContentBlockStart(\n                event.contentBlockStart\n              );\n              if (startChunk != null) {\n                const idx = event.contentBlockStart.contentBlockIndex;\n                if (idx != null) {\n                  seenBlockIndices.add(idx);\n                  if (event.contentBlockStart.start?.toolUse != null) {\n                    toolUseBlockIndices.add(idx);\n                  }\n                }\n                await enqueueChunk({\n                  chunk: this.enrichChunk(startChunk, seenBlockIndices),\n                  callbackChunk: startChunk,\n                  callbackToken: startChunk.text,\n                });\n              }\n            } else if (event.contentBlockDelta != null) {\n              await enqueueDelta(event.contentBlockDelta);\n            } else if (event.metadata != null) {\n              await enqueueChunk({\n                chunk: handleConverseStreamMetadata(event.metadata, {\n                  streamUsage,\n                }),\n              });\n            } else if (event.contentBlockStop != null) {\n              const stopIdx = event.contentBlockStop.contentBlockIndex;\n              if (stopIdx != null) {\n                seenBlockIndices.add(stopIdx);\n                if (sealToolUseOnStop && toolUseBlockIndices.has(stopIdx)) {\n                  const sealChunk = createConverseToolUseStopChunk(stopIdx);\n                  await enqueueChunk({\n                    chunk: sealChunk,\n                    callbackChunk: sealChunk,\n                    callbackToken: sealChunk.text,\n                  });\n                }\n              }\n            } else {\n              await enqueueChunk({\n                chunk: new ChatGenerationChunk({\n                  text: '',\n                  message: new AIMessageChunk({\n                    content: '',\n                    response_metadata: { ...event } as ResponseMetadata,\n                  }),\n                }),\n              });\n            }\n          }\n        } catch (error) {\n          producerState.error = error;\n        } finally {\n          producerState.done = true;\n          notifyConsumerForChunk();\n        }\n      })();\n\n      try {\n        let keepStreaming = true;\n        while (keepStreaming) {\n          if (isSignalAborted(options.signal)) {\n            abortStream();\n            throw new Error('AbortError: User aborted the request.');\n          }\n\n          await waitForNextChunk();\n          const queuedChunk = dequeue();\n\n          if (!queuedChunk) {\n            if (producerState.error != null) {\n              throw producerState.error;\n            }\n            if (producerState.done) {\n              keepStreaming = false;\n            }\n            continue;\n          }\n\n          if (queuedChunk.smooth) {\n            bufferedTextLength = Math.max(\n              0,\n              bufferedTextLength - queuedChunk.textLength\n            );\n            notifyProducerForSpace();\n            await waitForStreamDelay(\n              getCadencedStreamDelay({\n                targetDelay: hasEmittedVisibleContent\n                  ? this._lc_stream_delay\n                  : 0,\n                lastVisibleContentAt,\n                now: Date.now(),\n              }),\n              options.signal\n            );\n            if (isSignalAborted(options.signal)) {\n              abortStream();\n              throw new Error('AbortError: User aborted the request.');\n            }\n            hasEmittedVisibleContent = true;\n            lastVisibleContentAt = Date.now();\n          } else {\n            notifyProducerForSpace();\n          }\n\n          yield queuedChunk.chunk;\n\n          if (queuedChunk.callbackChunk != null) {\n            await runManager?.handleLLMNewToken(\n              queuedChunk.callbackToken,\n              undefined,\n              undefined,\n              undefined,\n              undefined,\n              { chunk: queuedChunk.callbackChunk }\n            );\n          }\n        }\n      } finally {\n        consumerClosed = true;\n        if (!producerState.done) {\n          abortStream();\n          notifyProducerForSpace();\n        }\n        await producer;\n      }\n    } finally {\n      options.signal?.removeEventListener('abort', abortStream);\n      if (!streamAbortController.signal.aborted) {\n        streamAbortController.abort();\n      }\n    }\n  }\n\n  /**\n   * Inject `index` on content blocks for proper merge behaviour, then strip\n   * `contentBlockIndex` from response_metadata to prevent `_mergeDicts` conflicts.\n   *\n   * Text string content is promoted to array form only when the stream contains\n   * multiple content block indices (e.g. reasoning at index 0, text at index 1),\n   * ensuring text merges correctly with the already-array accumulated content.\n   */\n  private enrichChunk(\n    chunk: ChatGenerationChunk,\n    seenBlockIndices: Set<number>\n  ): ChatGenerationChunk {\n    const message = chunk.message;\n    if (!(message instanceof AIMessageChunk)) {\n      return chunk;\n    }\n\n    const metadata = message.response_metadata as Record<string, unknown>;\n    const blockIndex = this.extractContentBlockIndex(metadata);\n    const hasMetadataIndex = blockIndex != null;\n\n    let content: AIMessageChunk['content'] = message.content;\n    let contentModified = false;\n\n    if (Array.isArray(content) && blockIndex != null) {\n      content = content.map((block) =>\n        typeof block === 'object' && !('index' in block)\n          ? { ...block, index: blockIndex }\n          : block\n      );\n      contentModified = true;\n    } else if (\n      typeof content === 'string' &&\n      content !== '' &&\n      blockIndex != null &&\n      seenBlockIndices.size > 1\n    ) {\n      content = [{ type: 'text', text: content, index: blockIndex }];\n      contentModified = true;\n    }\n\n    if (!contentModified && !hasMetadataIndex) {\n      return chunk;\n    }\n\n    const cleanedMetadata = hasMetadataIndex\n      ? (this.removeContentBlockIndex(metadata) as Record<string, unknown>)\n      : metadata;\n\n    return new ChatGenerationChunk({\n      text: chunk.text,\n      message: new AIMessageChunk({\n        ...message,\n        content,\n        response_metadata: cleanedMetadata,\n      }),\n      generationInfo: chunk.generationInfo,\n    });\n  }\n\n  /**\n   * Extract `contentBlockIndex` from the top level of response_metadata.\n   * Our custom handlers always place it at the top level.\n   */\n  private extractContentBlockIndex(\n    metadata: Record<string, unknown>\n  ): number | undefined {\n    if (\n      'contentBlockIndex' in metadata &&\n      typeof metadata.contentBlockIndex === 'number'\n    ) {\n      return metadata.contentBlockIndex;\n    }\n    return undefined;\n  }\n\n  private removeContentBlockIndex(obj: unknown): unknown {\n    if (obj === null || obj === undefined) {\n      return obj;\n    }\n\n    if (Array.isArray(obj)) {\n      return obj.map((item) => this.removeContentBlockIndex(item));\n    }\n\n    if (typeof obj === 'object') {\n      const cleaned: Record<string, unknown> = {};\n      for (const [key, value] of Object.entries(obj)) {\n        if (key !== 'contentBlockIndex') {\n          cleaned[key] = this.removeContentBlockIndex(value);\n        }\n      }\n      return cleaned;\n    }\n\n    return obj;\n  }\n}\n\nexport type { ChatBedrockConverseInput };\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA4DA,MAAM,0BAA0B;AAChC,MAAM,8BAA8B;AACpC,MAAM,wBAAwB;AAC9B,MAAM,oBAAoB,IAAI,IAAI;CAAC;CAAK;CAAK;CAAK;CAAK;CAAK;CAAK;AAAG,CAAC;AAUrE,SAAS,wBAAwB,MAAc,SAAyB;CACtE,IAAI,WAAW,KAAK,QAClB,OAAO,KAAK;CAGd,KAAK,IAAI,WAAW,SAAS,WAAW,KAAK,QAAQ,YACnD,IAAI,kBAAkB,IAAI,KAAK,SAAS,GACtC,OAAO,WAAW;CAItB,OAAO,KAAK;AACd;AAEA,SAAS,iBAAiB,MAAwB;CAChD,MAAM,SAAmB,CAAC;CAC1B,IAAI,eAAe;CAEnB,OAAO,eAAe,KAAK,QAAQ;EAEjC,MAAM,YAAY,wBADI,KAAK,MAAM,YAEnB,GACZ,qBACF;EACA,OAAO,KAAK,KAAK,MAAM,cAAc,eAAe,SAAS,CAAC;EAC9D,gBAAgB;CAClB;CAEA,OAAO;AACT;AAEA,SAAS,uBAAuB,EAC9B,aACA,sBACA,OAKS;CACT,IAAI,eAAe,KAAK,wBAAwB,MAC9C,OAAO;CAET,OAAO,KAAK,IAAI,GAAG,eAAe,MAAM,qBAAqB;AAC/D;AAEA,eAAe,mBACb,OACA,QACe;CACf,IAAI,SAAS,KAAK,gBAAgB,MAAM,GACtC;CAEF,MAAM,IAAI,SAAe,YAAY;EACnC,MAAM,aAA0D,CAAC;EACjE,MAAM,gBAAsB;GAC1B,IAAI,WAAW,SACb,aAAa,WAAW,OAAO;GAEjC,QAAQ,oBAAoB,SAAS,OAAO;GAC5C,QAAQ;EACV;EACA,WAAW,UAAU,iBAAiB;GACpC,QAAQ,oBAAoB,SAAS,OAAO;GAC5C,QAAQ;EACV,GAAG,KAAK;EACR,QAAQ,iBAAiB,SAAS,SAAS,EAAE,MAAM,KAAK,CAAC;EACzD,IAAI,gBAAgB,MAAM,GACxB,QAAQ;CAEZ,CAAC;AACH;AAEA,SAAS,gBAAgB,QAA+B;CACtD,OAAO,QAAQ,YAAY;AAC7B;AAoEA,IAAa,4BAAb,cAA+C,oBAAoB;CACjE;;;;CAKA;;;;CAKA;;;;CAKA;;;;CAKA;;;;;;;CAQA;CAEA,YAAY,QAAyC;EACnD,MAAM,MAAM;EACZ,KAAK,cAAc,QAAQ;EAC3B,KAAK,iBAAiB,QAAQ;EAC9B,KAAK,mBAAmB,KAAK,IAAI,GAAG,QAAQ,oBAAoB,CAAC;EACjE,KAAK,8BAA8B,QAAQ;EAC3C,KAAK,cAAc,QAAQ;EAI3B,KAAK,eAAe,QAAQ,SAAS,KAAK;CAC5C;CAEA,OAAO,UAAkB;EACvB,OAAO;CACT;;;;;CAMA,aAA+B;EAC7B,OAAO,KAAK,+BAA+B,KAAK;CAClD;;;;CAKA,iBACE,SAGA;EACA,MAAM,aAAa,MAAM,iBAAiB,OAAO;EACjD,MAAM,aACJ,KAAK,gBAAgB,QAAQ,yBAAyB,KAAK,YAAY,IACnE,4BACE,WAAW,YACX,MACA,6BAA6B,KAAK,gBAAgB,KAAK,YAAY,CACrE,IACA,WAAW;;EAGjB,MAAM,kBAAkB,SAAS,eAAe,KAAK;EAErD,OAAO;GACL,GAAG;GACH;GACA,aAAa,kBAAkB,EAAE,MAAM,gBAAgB,IAAI,KAAA;EAC7D;CACF;;;;;CAMA,MAAe,sBACb,UACA,SACA,YACqB;EACrB,MAAM,gBAAgB,KAAK;EAC3B,IACE,KAAK,+BAA+B,QACpC,KAAK,gCAAgC,IAErC,KAAK,QAAQ,KAAK;EAGpB,IAAI;GACF,OAAO,MAAM,MAAM,sBAAsB,UAAU,SAAS,UAAU;EACxE,UAAU;GACR,KAAK,QAAQ;EACf;CACF;;;;;;;;CASA,OAAgB,sBACd,UACA,SACA,YACqC;EACrC,MAAM,EAAE,kBAAkB,mBACxB,0BAA0B,QAAQ;EACpC,MAAM,SAAS,KAAK,iBAAiB,OAAO;EAE5C,IAAI,EAAE,gBAAgB;EACtB,IAAK,QAAoC,gBAAgB,KAAA,GACvD,cAAe,QAAoC;EAGrD,MAAM,UAAU,KAAK,WAAW;EAEhC,kCAAkC;GAChC,cAAc,QAAQ;GACtB,QAAQ;GACR,UAAU;GACV;GACA;EACF,CAAC;EAED,MAAM,UAAU,IAAI,sBAAsB;GACxC;GACA,UAAU;GACV,QAAQ;GACR,GAAI;EACN,CAAC;EAED,MAAM,wBAAwB,IAAI,gBAAgB;EAClD,MAAM,oBAA0B,sBAAsB,MAAM;EAC5D,QAAQ,QAAQ,iBAAiB,SAAS,aAAa,EAAE,MAAM,KAAK,CAAC;EACrE,IAAI,gBAAgB,QAAQ,MAAM,GAChC,YAAY;EAGd,IAAI;GAKF,MAAM,UAAS,MAJQ,KAAK,OAAO,KAAK,SAAS,EAC/C,aAAa,sBAAsB,OACrC,CAAC,EAAA,CAEuB;GACxB,IAAI,CAAC,QACH;GAGF,MAAM,mCAAmB,IAAI,IAAY;GACzC,MAAM,sCAAsB,IAAI,IAAY;GAC5C,MAAM,eAAwC,CAAC;GAC/C,MAAM,gBAAoD,EAAE,MAAM,MAAM;GACxE,IAAI,mBAAmB;GACvB,IAAI,qBAAqB;GACzB,IAAI,iBAAiB;GACrB,IAAI;GACJ,IAAI;GACJ,IAAI,2BAA2B;GAC/B,IAAI;;;;;;;GAQJ,MAAM,oBACJ,QAAQ,mBAAmB,QAAQ,KAAK,mBAAmB;GAE7D,MAAM,+BAAqC;IACzC,iBAAiB;IACjB,iBAAiB,KAAA;GACnB;GAEA,MAAM,+BAAqC;IACzC,iBAAiB;IACjB,iBAAiB,KAAA;GACnB;GAEA,MAAM,wBACJ,mBAAmB,aAAa;GAElC,MAAM,4BACJ,aAAa,SAAS;GAExB,MAAM,0BACJ,oBAAoB,KAAK,2BACzB,sBAAsB;GAExB,MAAM,mBAAmB,YAA2B;IAClD,IACE,gBAAgB,KAChB,cAAc,QACd,cAAc,SAAS,MAEvB;IAEF,MAAM,IAAI,SAAe,YAAY;KACnC,iBAAiB;IACnB,CAAC;GACH;GAEA,MAAM,oBAAoB,YAA2B;IACnD,OACE,kBAAkB,KAClB,CAAC,kBACD,CAAC,gBAAgB,QAAQ,MAAM,GAE/B,MAAM,IAAI,SAAe,YAAY;KACnC,MAAM,SAAS,QAAQ;KACvB,MAAM,gBAAsB;MAC1B,QAAQ,oBAAoB,SAAS,OAAO;MAC5C,QAAQ;KACV;KACA,MAAM,gBAAsB;MAC1B,QAAQ,oBAAoB,SAAS,OAAO;MAC5C,QAAQ;KACV;KACA,iBAAiB;KACjB,QAAQ,iBAAiB,SAAS,SAAS,EAAE,MAAM,KAAK,CAAC;KACzD,IAAI,gBAAgB,MAAM,GACxB,QAAQ;IAEZ,CAAC;GAEL;GAEA,MAAM,gBAAmD;IACvD,IAAI,CAAC,gBAAgB,GACnB;IAEF,MAAM,cAAc,aAAa;IACjC;IACA,IACE,mBAAmB,OACnB,mBAAmB,KAAK,aAAa,QACrC;KACA,aAAa,OAAO,GAAG,gBAAgB;KACvC,mBAAmB;IACrB;IACA,OAAO;GACT;GAEA,MAAM,UAAU,OACd,gBACkB;IAClB,MAAM,kBAAkB;IACxB,IAAI,kBAAkB,gBAAgB,QAAQ,MAAM,GAAG;KACrD,YAAY;KACZ,MAAM,IAAI,MAAM,uCAAuC;IACzD;IACA,aAAa,KAAK,WAAW;IAC7B,IAAI,YAAY,QACd,sBAAsB,YAAY;IAEpC,uBAAuB;GACzB;GAEA,MAAM,eAAe,OAAO,EAC1B,OACA,eACA,gBAAgB,IAChB,SAAS,OACT,aAAa,QAOM;IACnB,MAAM,QAAQ;KACZ;KACA;KACA;KACA;KACA,YAAY,SAAS,aAAa;IACpC,CAAC;GACH;GAEA,MAAM,eAAe,OACnB,sBACkB;IAClB,MAAM,QAAQ,kBAAkB;IAChC,IAAI,SAAS,MACX,MAAM,IAAI,MAAM,kCAAkC;IAGpD,MAAM,MAAM,kBAAkB;IAC9B,IAAI,OAAO,MACT,iBAAiB,IAAI,GAAG;IAG1B,MAAM,OAAO,MAAM;IACnB,MAAM,mBAAmB,MAAM;IAC/B,MAAM,gBAAgB,kBAAkB;IACxC,MAAM,cACJ,OAAO,SAAS,WACZ,OACA,OAAO,kBAAkB,WACvB,gBACA;IACR,MAAM,SAAS,KAAK,mBAAmB,KAAK,gBAAgB;IAC5D,MAAM,cAAc,SAChB,iBAAiB,WAAW,IAC5B,CAAC,WAAW;IAEhB,KAAK,MAAM,SAAS,aAAa;KAC/B,IAAI,aAAa;KACjB,IAAI,OAAO,SAAS,UAClB,aAAa;MACX,GAAG;MACH,OAAO,EAAE,MAAM,MAAM;KACvB;UACK,IACL,OAAO,kBAAkB,YACzB,oBAAoB,MAEpB,aAAa;MACX,GAAG;MACH,OAAO,EACL,kBAAkB;OAChB,GAAG;OACH,MAAM;MACR,EACF;KACF;KAGF,MAAM,aAAa,sCAAsC,UAAU;KACnE,MAAM,aAAa;MACjB,OAAO,KAAK,YAAY,YAAY,gBAAgB;MACpD,eAAe;MACf,eAAe,WAAW;MAC1B;MACA,YAAY,MAAM;KACpB,CAAC;IACH;GACF;GAEA,MAAM,YAAY,YAA2B;IAC3C,IAAI;KACF,WAAW,MAAM,SAAS,QAAQ;MAChC,IAAI,gBAAgB,QAAQ,MAAM,GAAG;OACnC,YAAY;OACZ,MAAM,IAAI,MAAM,uCAAuC;MACzD;MAEA,IAAI,MAAM,qBAAqB,MAAM;OACnC,MAAM,aAAa,sCACjB,MAAM,iBACR;OACA,IAAI,cAAc,MAAM;QACtB,MAAM,MAAM,MAAM,kBAAkB;QACpC,IAAI,OAAO,MAAM;SACf,iBAAiB,IAAI,GAAG;SACxB,IAAI,MAAM,kBAAkB,OAAO,WAAW,MAC5C,oBAAoB,IAAI,GAAG;QAE/B;QACA,MAAM,aAAa;SACjB,OAAO,KAAK,YAAY,YAAY,gBAAgB;SACpD,eAAe;SACf,eAAe,WAAW;QAC5B,CAAC;OACH;MACF,OAAO,IAAI,MAAM,qBAAqB,MACpC,MAAM,aAAa,MAAM,iBAAiB;WACrC,IAAI,MAAM,YAAY,MAC3B,MAAM,aAAa,EACjB,OAAO,6BAA6B,MAAM,UAAU,EAClD,YACF,CAAC,EACH,CAAC;WACI,IAAI,MAAM,oBAAoB,MAAM;OACzC,MAAM,UAAU,MAAM,iBAAiB;OACvC,IAAI,WAAW,MAAM;QACnB,iBAAiB,IAAI,OAAO;QAC5B,IAAI,qBAAqB,oBAAoB,IAAI,OAAO,GAAG;SACzD,MAAM,YAAY,+BAA+B,OAAO;SACxD,MAAM,aAAa;UACjB,OAAO;UACP,eAAe;UACf,eAAe,UAAU;SAC3B,CAAC;QACH;OACF;MACF,OACE,MAAM,aAAa,EACjB,OAAO,IAAI,oBAAoB;OAC7B,MAAM;OACN,SAAS,IAAI,eAAe;QAC1B,SAAS;QACT,mBAAmB,EAAE,GAAG,MAAM;OAChC,CAAC;MACH,CAAC,EACH,CAAC;KAEL;IACF,SAAS,OAAO;KACd,cAAc,QAAQ;IACxB,UAAU;KACR,cAAc,OAAO;KACrB,uBAAuB;IACzB;GACF,EAAA,CAAG;GAEH,IAAI;IACF,IAAI,gBAAgB;IACpB,OAAO,eAAe;KACpB,IAAI,gBAAgB,QAAQ,MAAM,GAAG;MACnC,YAAY;MACZ,MAAM,IAAI,MAAM,uCAAuC;KACzD;KAEA,MAAM,iBAAiB;KACvB,MAAM,cAAc,QAAQ;KAE5B,IAAI,CAAC,aAAa;MAChB,IAAI,cAAc,SAAS,MACzB,MAAM,cAAc;MAEtB,IAAI,cAAc,MAChB,gBAAgB;MAElB;KACF;KAEA,IAAI,YAAY,QAAQ;MACtB,qBAAqB,KAAK,IACxB,GACA,qBAAqB,YAAY,UACnC;MACA,uBAAuB;MACvB,MAAM,mBACJ,uBAAuB;OACrB,aAAa,2BACT,KAAK,mBACL;OACJ;OACA,KAAK,KAAK,IAAI;MAChB,CAAC,GACD,QAAQ,MACV;MACA,IAAI,gBAAgB,QAAQ,MAAM,GAAG;OACnC,YAAY;OACZ,MAAM,IAAI,MAAM,uCAAuC;MACzD;MACA,2BAA2B;MAC3B,uBAAuB,KAAK,IAAI;KAClC,OACE,uBAAuB;KAGzB,MAAM,YAAY;KAElB,IAAI,YAAY,iBAAiB,MAC/B,MAAM,YAAY,kBAChB,YAAY,eACZ,KAAA,GACA,KAAA,GACA,KAAA,GACA,KAAA,GACA,EAAE,OAAO,YAAY,cAAc,CACrC;IAEJ;GACF,UAAU;IACR,iBAAiB;IACjB,IAAI,CAAC,cAAc,MAAM;KACvB,YAAY;KACZ,uBAAuB;IACzB;IACA,MAAM;GACR;EACF,UAAU;GACR,QAAQ,QAAQ,oBAAoB,SAAS,WAAW;GACxD,IAAI,CAAC,sBAAsB,OAAO,SAChC,sBAAsB,MAAM;EAEhC;CACF;;;;;;;;;CAUA,YACE,OACA,kBACqB;EACrB,MAAM,UAAU,MAAM;EACtB,IAAI,EAAE,mBAAmB,iBACvB,OAAO;EAGT,MAAM,WAAW,QAAQ;EACzB,MAAM,aAAa,KAAK,yBAAyB,QAAQ;EACzD,MAAM,mBAAmB,cAAc;EAEvC,IAAI,UAAqC,QAAQ;EACjD,IAAI,kBAAkB;EAEtB,IAAI,MAAM,QAAQ,OAAO,KAAK,cAAc,MAAM;GAChD,UAAU,QAAQ,KAAK,UACrB,OAAO,UAAU,YAAY,EAAE,WAAW,SACtC;IAAE,GAAG;IAAO,OAAO;GAAW,IAC9B,KACN;GACA,kBAAkB;EACpB,OAAO,IACL,OAAO,YAAY,YACnB,YAAY,MACZ,cAAc,QACd,iBAAiB,OAAO,GACxB;GACA,UAAU,CAAC;IAAE,MAAM;IAAQ,MAAM;IAAS,OAAO;GAAW,CAAC;GAC7D,kBAAkB;EACpB;EAEA,IAAI,CAAC,mBAAmB,CAAC,kBACvB,OAAO;EAGT,MAAM,kBAAkB,mBACnB,KAAK,wBAAwB,QAAQ,IACtC;EAEJ,OAAO,IAAI,oBAAoB;GAC7B,MAAM,MAAM;GACZ,SAAS,IAAI,eAAe;IAC1B,GAAG;IACH;IACA,mBAAmB;GACrB,CAAC;GACD,gBAAgB,MAAM;EACxB,CAAC;CACH;;;;;CAMA,yBACE,UACoB;EACpB,IACE,uBAAuB,YACvB,OAAO,SAAS,sBAAsB,UAEtC,OAAO,SAAS;CAGpB;CAEA,wBAAgC,KAAuB;EACrD,IAAI,QAAQ,QAAQ,QAAQ,KAAA,GAC1B,OAAO;EAGT,IAAI,MAAM,QAAQ,GAAG,GACnB,OAAO,IAAI,KAAK,SAAS,KAAK,wBAAwB,IAAI,CAAC;EAG7D,IAAI,OAAO,QAAQ,UAAU;GAC3B,MAAM,UAAmC,CAAC;GAC1C,KAAK,MAAM,CAAC,KAAK,UAAU,OAAO,QAAQ,GAAG,GAC3C,IAAI,QAAQ,qBACV,QAAQ,OAAO,KAAK,wBAAwB,KAAK;GAGrD,OAAO;EACT;EAEA,OAAO;CACT;AACF"}
+{"version":3,"file":"index.mjs","names":[],"sources":["../../../../src/llm/bedrock/index.ts"],"sourcesContent":["/**\n * Optimized ChatBedrockConverse wrapper that fixes content block merging for\n * streaming responses and adds support for latest @langchain/aws features:\n *\n * - Application Inference Profiles (PR #9129)\n * - Service Tiers (Priority/Standard/Flex) (PR #9785) - requires AWS SDK 3.966.0+\n *\n * Bedrock's `@langchain/aws` library does not include an `index` property on content\n * blocks (unlike Anthropic/OpenAI), which causes LangChain's `_mergeLists` to append\n * each streaming chunk as a separate array entry instead of merging by index.\n *\n * This wrapper takes full ownership of the stream by directly interfacing with the\n * AWS SDK client (`this.client`) and using custom handlers from `./utils/` that\n * include `contentBlockIndex` in response_metadata for every delta type. It then\n * promotes `contentBlockIndex` to an `index` property on each content block\n * (mirroring Anthropic's pattern) and strips it from metadata to avoid\n * `_mergeDicts` conflicts.\n *\n * When multiple content block types are present (e.g. reasoning + text), text deltas\n * are promoted from strings to array form with `index` so they merge correctly once\n * the accumulated content is already an array.\n */\n\nimport { ChatBedrockConverse } from '@langchain/aws';\nimport { AIMessageChunk } from '@langchain/core/messages';\nimport { ChatGenerationChunk, ChatResult } from '@langchain/core/outputs';\nimport {\n  ConverseStreamCommand,\n  type ConverseStreamOutput,\n  type GuardrailConfiguration,\n  type GuardrailStreamConfiguration,\n} from '@aws-sdk/client-bedrock-runtime';\nimport type { CallbackManagerForLLMRun } from '@langchain/core/callbacks/manager';\nimport type { BaseMessage, ResponseMetadata } from '@langchain/core/messages';\nimport type { ChatBedrockConverseInput } from '@langchain/aws';\nimport type { ContentBlockDeltaEvent } from './types';\nimport {\n  convertToConverseMessages,\n  createConverseToolUseStopChunk,\n  handleConverseStreamContentBlockStart,\n  handleConverseStreamContentBlockDelta,\n  handleConverseStreamMetadata,\n} from './utils';\nimport {\n  resolveBedrockPromptCacheTtl,\n  supportsBedrockToolCache,\n  type PromptCacheTtl,\n} from '@/messages/cache';\nimport { applyCachePointsToConversePayload } from './cachePoints';\nimport { insertBedrockToolCachePoint } from './toolCache';\n\n/**\n * Service tier type for Bedrock invocations.\n * Requires AWS SDK >= 3.966.0 to actually work.\n * @see https://docs.aws.amazon.com/bedrock/latest/userguide/service-tiers-inference.html\n */\nexport type ServiceTierType = 'priority' | 'default' | 'flex' | 'reserved';\n\nexport type CustomGuardrailConfiguration = GuardrailConfiguration &\n  Pick<GuardrailStreamConfiguration, 'streamProcessingMode'>;\n\nconst MAX_STREAM_QUEUE_CHUNKS = 256;\nconst MAX_STREAM_QUEUE_TEXT_CHARS = 8192;\nconst STREAM_CHUNK_MIN_SIZE = 4;\nconst STREAM_BOUNDARIES = new Set([' ', '.', ',', '!', '?', ';', ':']);\n\ntype QueuedGenerationChunk = {\n  chunk: ChatGenerationChunk;\n  callbackChunk?: ChatGenerationChunk;\n  callbackToken: string;\n  smooth: boolean;\n  textLength: number;\n};\n\nfunction findStreamChunkBoundary(text: string, minSize: number): number {\n  if (minSize >= text.length) {\n    return text.length;\n  }\n\n  for (let position = minSize; position < text.length; position++) {\n    if (STREAM_BOUNDARIES.has(text[position])) {\n      return position + 1;\n    }\n  }\n\n  return text.length;\n}\n\nfunction splitStreamToken(text: string): string[] {\n  const chunks: string[] = [];\n  let currentIndex = 0;\n\n  while (currentIndex < text.length) {\n    const remainingText = text.slice(currentIndex);\n    const chunkSize = findStreamChunkBoundary(\n      remainingText,\n      STREAM_CHUNK_MIN_SIZE\n    );\n    chunks.push(text.slice(currentIndex, currentIndex + chunkSize));\n    currentIndex += chunkSize;\n  }\n\n  return chunks;\n}\n\n/**\n * Resolves the text a delta contributes to the smoothing cadence, preferring a\n * text delta over a reasoning delta and ignoring non-string payloads.\n */\nfunction resolveVisibleText(text?: string, reasoningText?: string): string {\n  if (typeof text === 'string') {\n    return text;\n  }\n\n  if (typeof reasoningText === 'string') {\n    return reasoningText;\n  }\n\n  return '';\n}\n\nfunction getCadencedStreamDelay({\n  targetDelay,\n  lastVisibleContentAt,\n  now,\n}: {\n  targetDelay: number;\n  lastVisibleContentAt?: number;\n  now: number;\n}): number {\n  if (targetDelay <= 0 || lastVisibleContentAt == null) {\n    return 0;\n  }\n  return Math.max(0, targetDelay - (now - lastVisibleContentAt));\n}\n\nasync function waitForStreamDelay(\n  delay: number,\n  signal?: AbortSignal\n): Promise<void> {\n  if (delay <= 0 || isSignalAborted(signal)) {\n    return;\n  }\n  await new Promise<void>((resolve) => {\n    const timeoutRef: { current?: ReturnType<typeof setTimeout> } = {};\n    const onAbort = (): void => {\n      if (timeoutRef.current) {\n        clearTimeout(timeoutRef.current);\n      }\n      signal?.removeEventListener('abort', onAbort);\n      resolve();\n    };\n    timeoutRef.current = setTimeout(() => {\n      signal?.removeEventListener('abort', onAbort);\n      resolve();\n    }, delay);\n    signal?.addEventListener('abort', onAbort, { once: true });\n    if (isSignalAborted(signal)) {\n      onAbort();\n    }\n  });\n}\n\nfunction isSignalAborted(signal?: AbortSignal): boolean {\n  return signal?.aborted === true;\n}\n\n/**\n * Extended input interface with additional features:\n * - applicationInferenceProfile: Use an inference profile ARN instead of model ID\n * - serviceTier: Specify service tier (Priority, Standard, Flex, Reserved)\n */\nexport interface CustomChatBedrockConverseInput\n  extends ChatBedrockConverseInput {\n  /**\n   * Enables Bedrock prompt cache checkpoints for message and tool prefixes.\n   */\n  promptCache?: boolean;\n\n  /**\n   * Prompt-cache checkpoint TTL. Defaults to `'1h'` (extended cache) when\n   * `promptCache` is enabled; set `'5m'` for the legacy 5-minute behavior.\n   * Bedrock models that don't support the 1-hour TTL downgrade to 5m\n   * server-side (verified on Sonnet/Opus 4.6), so the default is safe to leave\n   * on; use `'5m'` for any model that rejects it.\n   */\n  promptCacheTtl?: PromptCacheTtl;\n\n  /**\n   * Minimum delay in milliseconds between visible streamed content deltas.\n   */\n  _lc_stream_delay?: number;\n\n  /**\n   * Guardrail configuration for Converse and ConverseStream invocations.\n   * `streamProcessingMode` is only used by ConverseStream.\n   */\n  guardrailConfig?: CustomGuardrailConfiguration;\n\n  /**\n   * Application Inference Profile ARN to use for the model.\n   * For example, \"arn:aws:bedrock:eu-west-1:123456789102:application-inference-profile/fm16bt65tzgx\"\n   * When provided, this ARN will be used for the actual inference calls instead of the model ID.\n   * Must still provide `model` as normal modelId to benefit from all the metadata.\n   * @see https://docs.aws.amazon.com/bedrock/latest/userguide/inference-profiles-create.html\n   */\n  applicationInferenceProfile?: string;\n\n  /**\n   * Service tier for model invocation.\n   * Specifies the processing tier type used for serving the request.\n   * Supported values are 'priority', 'default', 'flex', and 'reserved'.\n   *\n   * - 'priority': Prioritized processing for lower latency\n   * - 'default': Standard processing tier\n   * - 'flex': Flexible processing tier with lower cost\n   * - 'reserved': Reserved capacity for consistent performance\n   *\n   * If not provided, AWS uses the default tier.\n   * Note: Requires AWS SDK >= 3.966.0 to work.\n   * @see https://docs.aws.amazon.com/bedrock/latest/userguide/service-tiers-inference.html\n   */\n  serviceTier?: ServiceTierType;\n}\n\n/**\n * Extended call options with serviceTier override support.\n */\nexport interface CustomChatBedrockConverseCallOptions {\n  serviceTier?: ServiceTierType;\n  guardrailConfig?: CustomGuardrailConfiguration;\n}\n\nexport class CustomChatBedrockConverse extends ChatBedrockConverse {\n  _lc_stream_delay: number;\n\n  /**\n   * Whether to insert Bedrock prompt cache checkpoints when available.\n   */\n  promptCache?: boolean;\n\n  /**\n   * Prompt-cache checkpoint TTL (`'5m'` legacy or `'1h'` extended cache).\n   */\n  promptCacheTtl?: PromptCacheTtl;\n\n  /**\n   * Application Inference Profile ARN to use instead of model ID.\n   */\n  applicationInferenceProfile?: string;\n\n  /**\n   * Service tier for model invocation.\n   */\n  serviceTier?: ServiceTierType;\n\n  /**\n   * The configured model id, captured at construction so it survives the\n   * temporary `this.model` swap to an application-inference-profile ARN during\n   * generation. Used to gate the Bedrock tool cache point to Claude models\n   * (see {@link supportsBedrockToolCache}).\n   */\n  private readonly cacheModelId: string;\n\n  constructor(fields?: CustomChatBedrockConverseInput) {\n    super(fields);\n    this.promptCache = fields?.promptCache;\n    this.promptCacheTtl = fields?.promptCacheTtl;\n    this._lc_stream_delay = Math.max(0, fields?._lc_stream_delay ?? 0);\n    this.applicationInferenceProfile = fields?.applicationInferenceProfile;\n    this.serviceTier = fields?.serviceTier;\n    // `super(fields)` initializes `this.model` to LangChain's default Claude\n    // model when `fields.model` is omitted, so fall back to it rather than ''\n    // (which would treat the default Claude model as tool-cache-unsupported).\n    this.cacheModelId = fields?.model ?? this.model;\n  }\n\n  static lc_name(): string {\n    return 'LibreChatBedrockConverse';\n  }\n\n  /**\n   * Get the model ID to use for API calls.\n   * Returns applicationInferenceProfile if set, otherwise returns this.model.\n   */\n  protected getModelId(): string {\n    return this.applicationInferenceProfile ?? this.model;\n  }\n\n  /**\n   * Override invocationParams to add serviceTier support.\n   */\n  override invocationParams(\n    options?: this['ParsedCallOptions'] & CustomChatBedrockConverseCallOptions\n  ): ReturnType<ChatBedrockConverse['invocationParams']> & {\n    serviceTier?: { type: ServiceTierType };\n  } {\n    const baseParams = super.invocationParams(options);\n    const toolConfig =\n      this.promptCache === true && supportsBedrockToolCache(this.cacheModelId)\n        ? insertBedrockToolCachePoint(\n          baseParams.toolConfig,\n          true,\n          resolveBedrockPromptCacheTtl(this.promptCacheTtl, this.cacheModelId)\n        )\n        : baseParams.toolConfig;\n\n    /** Service tier from options or fall back to class-level setting */\n    const serviceTierType = options?.serviceTier ?? this.serviceTier;\n\n    return {\n      ...baseParams,\n      toolConfig,\n      serviceTier: serviceTierType ? { type: serviceTierType } : undefined,\n    };\n  }\n\n  /**\n   * Override _generateNonStreaming to use applicationInferenceProfile as modelId.\n   * Uses the same model-swapping pattern as streaming for consistency.\n   */\n  override async _generateNonStreaming(\n    messages: BaseMessage[],\n    options: this['ParsedCallOptions'] & CustomChatBedrockConverseCallOptions,\n    runManager?: CallbackManagerForLLMRun\n  ): Promise<ChatResult> {\n    const originalModel = this.model;\n    if (\n      this.applicationInferenceProfile != null &&\n      this.applicationInferenceProfile !== ''\n    ) {\n      this.model = this.applicationInferenceProfile;\n    }\n\n    try {\n      return await super._generateNonStreaming(messages, options, runManager);\n    } finally {\n      this.model = originalModel;\n    }\n  }\n\n  /**\n   * Own the stream end-to-end so we have direct access to every\n   * `contentBlockDelta.contentBlockIndex` from the AWS SDK.\n   *\n   * This replaces the parent's implementation which strips contentBlockIndex\n   * from text and reasoning deltas, making it impossible to merge correctly.\n   */\n  override async *_streamResponseChunks(\n    messages: BaseMessage[],\n    options: this['ParsedCallOptions'] & CustomChatBedrockConverseCallOptions,\n    runManager?: CallbackManagerForLLMRun\n  ): AsyncGenerator<ChatGenerationChunk> {\n    const { converseMessages, converseSystem } =\n      convertToConverseMessages(messages);\n    const params = this.invocationParams(options);\n\n    let { streamUsage } = this;\n    if ((options as Record<string, unknown>).streamUsage !== undefined) {\n      streamUsage = (options as Record<string, unknown>).streamUsage as boolean;\n    }\n\n    const modelId = this.getModelId();\n\n    applyCachePointsToConversePayload({\n      cacheControl: options.cache_control,\n      system: converseSystem,\n      messages: converseMessages,\n      params,\n      modelId,\n    });\n\n    const command = new ConverseStreamCommand({\n      modelId,\n      messages: converseMessages,\n      system: converseSystem,\n      ...(params as Record<string, unknown>),\n    });\n\n    const streamAbortController = new AbortController();\n    const abortStream = (): void => streamAbortController.abort();\n    options.signal?.addEventListener('abort', abortStream, { once: true });\n    if (isSignalAborted(options.signal)) {\n      abortStream();\n    }\n\n    try {\n      const response = await this.client.send(command, {\n        abortSignal: streamAbortController.signal,\n      });\n\n      const stream: AsyncIterable<ConverseStreamOutput> | undefined =\n        response.stream;\n      if (!stream) {\n        return;\n      }\n\n      const seenBlockIndices = new Set<number>();\n      const toolUseBlockIndices = new Set<number>();\n      const queuedChunks: QueuedGenerationChunk[] = [];\n      const producerState: { done: boolean; error?: unknown } = { done: false };\n      let queuedChunkIndex = 0;\n      let bufferedTextLength = 0;\n      let consumerClosed = false;\n      let notifyConsumer: (() => void) | undefined;\n      let notifyProducer: (() => void) | undefined;\n      let hasEmittedVisibleContent = false;\n      let lastVisibleContentAt: number | undefined;\n\n      /**\n       * Guardrails can reject an already-streamed toolUse block at\n       * `messageStop` (`guardrail_intervened`), after `contentBlockStop` has\n       * passed. Only emit eager-execution seals when no guardrails are\n       * configured, so a later intervention can't race an eagerly started tool.\n       */\n      const sealToolUseOnStop =\n        options.guardrailConfig == null && this.guardrailConfig == null;\n\n      const notifyConsumerForChunk = (): void => {\n        notifyConsumer?.();\n        notifyConsumer = undefined;\n      };\n\n      const notifyProducerForSpace = (): void => {\n        notifyProducer?.();\n        notifyProducer = undefined;\n      };\n\n      const hasQueuedChunks = (): boolean =>\n        queuedChunkIndex < queuedChunks.length;\n\n      const getQueuedChunkCount = (): number =>\n        queuedChunks.length - queuedChunkIndex;\n\n      const isQueueAtCapacity = (): boolean =>\n        getQueuedChunkCount() >= MAX_STREAM_QUEUE_CHUNKS ||\n        bufferedTextLength >= MAX_STREAM_QUEUE_TEXT_CHARS;\n\n      const waitForNextChunk = async (): Promise<void> => {\n        if (\n          hasQueuedChunks() ||\n          producerState.done ||\n          producerState.error != null\n        ) {\n          return;\n        }\n        await new Promise<void>((resolve) => {\n          notifyConsumer = resolve;\n        });\n      };\n\n      const waitForQueueSpace = async (): Promise<void> => {\n        while (\n          isQueueAtCapacity() &&\n          !consumerClosed &&\n          !isSignalAborted(options.signal)\n        ) {\n          await new Promise<void>((resolve) => {\n            const signal = options.signal;\n            const onAbort = (): void => {\n              signal?.removeEventListener('abort', onAbort);\n              resolve();\n            };\n            const onSpace = (): void => {\n              signal?.removeEventListener('abort', onAbort);\n              resolve();\n            };\n            notifyProducer = onSpace;\n            signal?.addEventListener('abort', onAbort, { once: true });\n            if (isSignalAborted(signal)) {\n              onAbort();\n            }\n          });\n        }\n      };\n\n      const dequeue = (): QueuedGenerationChunk | undefined => {\n        if (!hasQueuedChunks()) {\n          return undefined;\n        }\n        const queuedChunk = queuedChunks[queuedChunkIndex];\n        queuedChunkIndex++;\n        if (\n          queuedChunkIndex > 128 &&\n          queuedChunkIndex * 2 >= queuedChunks.length\n        ) {\n          queuedChunks.splice(0, queuedChunkIndex);\n          queuedChunkIndex = 0;\n        }\n        return queuedChunk;\n      };\n\n      const enqueue = async (\n        queuedChunk: QueuedGenerationChunk\n      ): Promise<void> => {\n        await waitForQueueSpace();\n        if (consumerClosed || isSignalAborted(options.signal)) {\n          abortStream();\n          throw new Error('AbortError: User aborted the request.');\n        }\n        queuedChunks.push(queuedChunk);\n        if (queuedChunk.smooth) {\n          bufferedTextLength += queuedChunk.textLength;\n        }\n        notifyConsumerForChunk();\n      };\n\n      const enqueueChunk = async ({\n        chunk,\n        callbackChunk,\n        callbackToken = '',\n        smooth = false,\n        textLength = 0,\n      }: {\n        chunk: ChatGenerationChunk;\n        callbackChunk?: ChatGenerationChunk;\n        callbackToken?: string;\n        smooth?: boolean;\n        textLength?: number;\n      }): Promise<void> => {\n        await enqueue({\n          chunk,\n          callbackChunk,\n          callbackToken,\n          smooth,\n          textLength: smooth ? textLength : 0,\n        });\n      };\n\n      const enqueueDelta = async (\n        contentBlockDelta: ContentBlockDeltaEvent\n      ): Promise<void> => {\n        const delta = contentBlockDelta.delta;\n        if (delta == null) {\n          throw new Error('No delta found in content block.');\n        }\n\n        const idx = contentBlockDelta.contentBlockIndex;\n        if (idx != null) {\n          seenBlockIndices.add(idx);\n        }\n\n        const text = delta.text;\n        const reasoningContent = delta.reasoningContent;\n        const reasoningText = reasoningContent?.text;\n        const visibleText = resolveVisibleText(text, reasoningText);\n        const smooth = this._lc_stream_delay > 0 && visibleText !== '';\n        const tokenChunks = smooth\n          ? splitStreamToken(visibleText)\n          : [visibleText];\n\n        for (const token of tokenChunks) {\n          let splitDelta = contentBlockDelta;\n          if (typeof text === 'string') {\n            splitDelta = {\n              ...contentBlockDelta,\n              delta: { text: token },\n            };\n          } else if (\n            typeof reasoningText === 'string' &&\n            reasoningContent != null\n          ) {\n            splitDelta = {\n              ...contentBlockDelta,\n              delta: {\n                reasoningContent: {\n                  ...reasoningContent,\n                  text: token,\n                },\n              },\n            };\n          }\n\n          const deltaChunk = handleConverseStreamContentBlockDelta(splitDelta);\n          await enqueueChunk({\n            chunk: this.enrichChunk(deltaChunk, seenBlockIndices),\n            callbackChunk: deltaChunk,\n            callbackToken: deltaChunk.text,\n            smooth,\n            textLength: token.length,\n          });\n        }\n      };\n\n      const producer = (async (): Promise<void> => {\n        try {\n          for await (const event of stream) {\n            if (isSignalAborted(options.signal)) {\n              abortStream();\n              throw new Error('AbortError: User aborted the request.');\n            }\n\n            if (event.contentBlockStart != null) {\n              const startChunk = handleConverseStreamContentBlockStart(\n                event.contentBlockStart\n              );\n              if (startChunk != null) {\n                const idx = event.contentBlockStart.contentBlockIndex;\n                if (idx != null) {\n                  seenBlockIndices.add(idx);\n                  if (event.contentBlockStart.start?.toolUse != null) {\n                    toolUseBlockIndices.add(idx);\n                  }\n                }\n                await enqueueChunk({\n                  chunk: this.enrichChunk(startChunk, seenBlockIndices),\n                  callbackChunk: startChunk,\n                  callbackToken: startChunk.text,\n                });\n              }\n            } else if (event.contentBlockDelta != null) {\n              await enqueueDelta(event.contentBlockDelta);\n            } else if (event.metadata != null) {\n              await enqueueChunk({\n                chunk: handleConverseStreamMetadata(event.metadata, {\n                  streamUsage,\n                }),\n              });\n            } else if (event.contentBlockStop != null) {\n              const stopIdx = event.contentBlockStop.contentBlockIndex;\n              if (stopIdx != null) {\n                seenBlockIndices.add(stopIdx);\n                if (sealToolUseOnStop && toolUseBlockIndices.has(stopIdx)) {\n                  const sealChunk = createConverseToolUseStopChunk(stopIdx);\n                  await enqueueChunk({\n                    chunk: sealChunk,\n                    callbackChunk: sealChunk,\n                    callbackToken: sealChunk.text,\n                  });\n                }\n              }\n            } else {\n              await enqueueChunk({\n                chunk: new ChatGenerationChunk({\n                  text: '',\n                  message: new AIMessageChunk({\n                    content: '',\n                    response_metadata: { ...event } as ResponseMetadata,\n                  }),\n                }),\n              });\n            }\n          }\n        } catch (error) {\n          producerState.error = error;\n        } finally {\n          producerState.done = true;\n          notifyConsumerForChunk();\n        }\n      })();\n\n      try {\n        let keepStreaming = true;\n        while (keepStreaming) {\n          if (isSignalAborted(options.signal)) {\n            abortStream();\n            throw new Error('AbortError: User aborted the request.');\n          }\n\n          await waitForNextChunk();\n          const queuedChunk = dequeue();\n\n          if (!queuedChunk) {\n            if (producerState.error != null) {\n              throw producerState.error;\n            }\n            if (producerState.done) {\n              keepStreaming = false;\n            }\n            continue;\n          }\n\n          if (queuedChunk.smooth) {\n            bufferedTextLength = Math.max(\n              0,\n              bufferedTextLength - queuedChunk.textLength\n            );\n            notifyProducerForSpace();\n            await waitForStreamDelay(\n              getCadencedStreamDelay({\n                targetDelay: hasEmittedVisibleContent\n                  ? this._lc_stream_delay\n                  : 0,\n                lastVisibleContentAt,\n                now: Date.now(),\n              }),\n              options.signal\n            );\n            if (isSignalAborted(options.signal)) {\n              abortStream();\n              throw new Error('AbortError: User aborted the request.');\n            }\n            hasEmittedVisibleContent = true;\n            lastVisibleContentAt = Date.now();\n          } else {\n            notifyProducerForSpace();\n          }\n\n          yield queuedChunk.chunk;\n\n          if (queuedChunk.callbackChunk != null) {\n            await runManager?.handleLLMNewToken(\n              queuedChunk.callbackToken,\n              undefined,\n              undefined,\n              undefined,\n              undefined,\n              { chunk: queuedChunk.callbackChunk }\n            );\n          }\n        }\n      } finally {\n        consumerClosed = true;\n        if (!producerState.done) {\n          abortStream();\n          notifyProducerForSpace();\n        }\n        await producer;\n      }\n    } finally {\n      options.signal?.removeEventListener('abort', abortStream);\n      if (!streamAbortController.signal.aborted) {\n        streamAbortController.abort();\n      }\n    }\n  }\n\n  /**\n   * Inject `index` on content blocks for proper merge behaviour, then strip\n   * `contentBlockIndex` from response_metadata to prevent `_mergeDicts` conflicts.\n   *\n   * Text string content is promoted to array form only when the stream contains\n   * multiple content block indices (e.g. reasoning at index 0, text at index 1),\n   * ensuring text merges correctly with the already-array accumulated content.\n   */\n  private enrichChunk(\n    chunk: ChatGenerationChunk,\n    seenBlockIndices: Set<number>\n  ): ChatGenerationChunk {\n    const message = chunk.message;\n    if (!(message instanceof AIMessageChunk)) {\n      return chunk;\n    }\n\n    const metadata = message.response_metadata as Record<string, unknown>;\n    const blockIndex = this.extractContentBlockIndex(metadata);\n    const hasMetadataIndex = blockIndex != null;\n\n    let content: AIMessageChunk['content'] = message.content;\n    let contentModified = false;\n\n    if (Array.isArray(content) && blockIndex != null) {\n      content = content.map((block) =>\n        typeof block === 'object' && !('index' in block)\n          ? { ...block, index: blockIndex }\n          : block\n      );\n      contentModified = true;\n    } else if (\n      typeof content === 'string' &&\n      content !== '' &&\n      blockIndex != null &&\n      seenBlockIndices.size > 1\n    ) {\n      content = [{ type: 'text', text: content, index: blockIndex }];\n      contentModified = true;\n    }\n\n    if (!contentModified && !hasMetadataIndex) {\n      return chunk;\n    }\n\n    const cleanedMetadata = hasMetadataIndex\n      ? (this.removeContentBlockIndex(metadata) as Record<string, unknown>)\n      : metadata;\n\n    return new ChatGenerationChunk({\n      text: chunk.text,\n      message: new AIMessageChunk({\n        ...message,\n        content,\n        response_metadata: cleanedMetadata,\n      }),\n      generationInfo: chunk.generationInfo,\n    });\n  }\n\n  /**\n   * Extract `contentBlockIndex` from the top level of response_metadata.\n   * Our custom handlers always place it at the top level.\n   */\n  private extractContentBlockIndex(\n    metadata: Record<string, unknown>\n  ): number | undefined {\n    if (\n      'contentBlockIndex' in metadata &&\n      typeof metadata.contentBlockIndex === 'number'\n    ) {\n      return metadata.contentBlockIndex;\n    }\n    return undefined;\n  }\n\n  private removeContentBlockIndex(obj: unknown): unknown {\n    if (obj === null || obj === undefined) {\n      return obj;\n    }\n\n    if (Array.isArray(obj)) {\n      return obj.map((item) => this.removeContentBlockIndex(item));\n    }\n\n    if (typeof obj === 'object') {\n      const cleaned: Record<string, unknown> = {};\n      for (const [key, value] of Object.entries(obj)) {\n        if (key !== 'contentBlockIndex') {\n          cleaned[key] = this.removeContentBlockIndex(value);\n        }\n      }\n      return cleaned;\n    }\n\n    return obj;\n  }\n}\n\nexport type { ChatBedrockConverseInput };\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA6DA,MAAM,0BAA0B;AAChC,MAAM,8BAA8B;AACpC,MAAM,wBAAwB;AAC9B,MAAM,oBAAoB,IAAI,IAAI;CAAC;CAAK;CAAK;CAAK;CAAK;CAAK;CAAK;AAAG,CAAC;AAUrE,SAAS,wBAAwB,MAAc,SAAyB;CACtE,IAAI,WAAW,KAAK,QAClB,OAAO,KAAK;CAGd,KAAK,IAAI,WAAW,SAAS,WAAW,KAAK,QAAQ,YACnD,IAAI,kBAAkB,IAAI,KAAK,SAAS,GACtC,OAAO,WAAW;CAItB,OAAO,KAAK;AACd;AAEA,SAAS,iBAAiB,MAAwB;CAChD,MAAM,SAAmB,CAAC;CAC1B,IAAI,eAAe;CAEnB,OAAO,eAAe,KAAK,QAAQ;EAEjC,MAAM,YAAY,wBADI,KAAK,MAAM,YAEnB,GACZ,qBACF;EACA,OAAO,KAAK,KAAK,MAAM,cAAc,eAAe,SAAS,CAAC;EAC9D,gBAAgB;CAClB;CAEA,OAAO;AACT;;;;;AAMA,SAAS,mBAAmB,MAAe,eAAgC;CACzE,IAAI,OAAO,SAAS,UAClB,OAAO;CAGT,IAAI,OAAO,kBAAkB,UAC3B,OAAO;CAGT,OAAO;AACT;AAEA,SAAS,uBAAuB,EAC9B,aACA,sBACA,OAKS;CACT,IAAI,eAAe,KAAK,wBAAwB,MAC9C,OAAO;CAET,OAAO,KAAK,IAAI,GAAG,eAAe,MAAM,qBAAqB;AAC/D;AAEA,eAAe,mBACb,OACA,QACe;CACf,IAAI,SAAS,KAAK,gBAAgB,MAAM,GACtC;CAEF,MAAM,IAAI,SAAe,YAAY;EACnC,MAAM,aAA0D,CAAC;EACjE,MAAM,gBAAsB;GAC1B,IAAI,WAAW,SACb,aAAa,WAAW,OAAO;GAEjC,QAAQ,oBAAoB,SAAS,OAAO;GAC5C,QAAQ;EACV;EACA,WAAW,UAAU,iBAAiB;GACpC,QAAQ,oBAAoB,SAAS,OAAO;GAC5C,QAAQ;EACV,GAAG,KAAK;EACR,QAAQ,iBAAiB,SAAS,SAAS,EAAE,MAAM,KAAK,CAAC;EACzD,IAAI,gBAAgB,MAAM,GACxB,QAAQ;CAEZ,CAAC;AACH;AAEA,SAAS,gBAAgB,QAA+B;CACtD,OAAO,QAAQ,YAAY;AAC7B;AAoEA,IAAa,4BAAb,cAA+C,oBAAoB;CACjE;;;;CAKA;;;;CAKA;;;;CAKA;;;;CAKA;;;;;;;CAQA;CAEA,YAAY,QAAyC;EACnD,MAAM,MAAM;EACZ,KAAK,cAAc,QAAQ;EAC3B,KAAK,iBAAiB,QAAQ;EAC9B,KAAK,mBAAmB,KAAK,IAAI,GAAG,QAAQ,oBAAoB,CAAC;EACjE,KAAK,8BAA8B,QAAQ;EAC3C,KAAK,cAAc,QAAQ;EAI3B,KAAK,eAAe,QAAQ,SAAS,KAAK;CAC5C;CAEA,OAAO,UAAkB;EACvB,OAAO;CACT;;;;;CAMA,aAA+B;EAC7B,OAAO,KAAK,+BAA+B,KAAK;CAClD;;;;CAKA,iBACE,SAGA;EACA,MAAM,aAAa,MAAM,iBAAiB,OAAO;EACjD,MAAM,aACJ,KAAK,gBAAgB,QAAQ,yBAAyB,KAAK,YAAY,IACnE,4BACA,WAAW,YACX,MACA,6BAA6B,KAAK,gBAAgB,KAAK,YAAY,CACrE,IACE,WAAW;;EAGjB,MAAM,kBAAkB,SAAS,eAAe,KAAK;EAErD,OAAO;GACL,GAAG;GACH;GACA,aAAa,kBAAkB,EAAE,MAAM,gBAAgB,IAAI,KAAA;EAC7D;CACF;;;;;CAMA,MAAe,sBACb,UACA,SACA,YACqB;EACrB,MAAM,gBAAgB,KAAK;EAC3B,IACE,KAAK,+BAA+B,QACpC,KAAK,gCAAgC,IAErC,KAAK,QAAQ,KAAK;EAGpB,IAAI;GACF,OAAO,MAAM,MAAM,sBAAsB,UAAU,SAAS,UAAU;EACxE,UAAU;GACR,KAAK,QAAQ;EACf;CACF;;;;;;;;CASA,OAAgB,sBACd,UACA,SACA,YACqC;EACrC,MAAM,EAAE,kBAAkB,mBACxB,0BAA0B,QAAQ;EACpC,MAAM,SAAS,KAAK,iBAAiB,OAAO;EAE5C,IAAI,EAAE,gBAAgB;EACtB,IAAK,QAAoC,gBAAgB,KAAA,GACvD,cAAe,QAAoC;EAGrD,MAAM,UAAU,KAAK,WAAW;EAEhC,kCAAkC;GAChC,cAAc,QAAQ;GACtB,QAAQ;GACR,UAAU;GACV;GACA;EACF,CAAC;EAED,MAAM,UAAU,IAAI,sBAAsB;GACxC;GACA,UAAU;GACV,QAAQ;GACR,GAAI;EACN,CAAC;EAED,MAAM,wBAAwB,IAAI,gBAAgB;EAClD,MAAM,oBAA0B,sBAAsB,MAAM;EAC5D,QAAQ,QAAQ,iBAAiB,SAAS,aAAa,EAAE,MAAM,KAAK,CAAC;EACrE,IAAI,gBAAgB,QAAQ,MAAM,GAChC,YAAY;EAGd,IAAI;GAKF,MAAM,UACJ,MALqB,KAAK,OAAO,KAAK,SAAS,EAC/C,aAAa,sBAAsB,OACrC,CAAC,EAAA,CAGU;GACX,IAAI,CAAC,QACH;GAGF,MAAM,mCAAmB,IAAI,IAAY;GACzC,MAAM,sCAAsB,IAAI,IAAY;GAC5C,MAAM,eAAwC,CAAC;GAC/C,MAAM,gBAAoD,EAAE,MAAM,MAAM;GACxE,IAAI,mBAAmB;GACvB,IAAI,qBAAqB;GACzB,IAAI,iBAAiB;GACrB,IAAI;GACJ,IAAI;GACJ,IAAI,2BAA2B;GAC/B,IAAI;;;;;;;GAQJ,MAAM,oBACJ,QAAQ,mBAAmB,QAAQ,KAAK,mBAAmB;GAE7D,MAAM,+BAAqC;IACzC,iBAAiB;IACjB,iBAAiB,KAAA;GACnB;GAEA,MAAM,+BAAqC;IACzC,iBAAiB;IACjB,iBAAiB,KAAA;GACnB;GAEA,MAAM,wBACJ,mBAAmB,aAAa;GAElC,MAAM,4BACJ,aAAa,SAAS;GAExB,MAAM,0BACJ,oBAAoB,KAAK,2BACzB,sBAAsB;GAExB,MAAM,mBAAmB,YAA2B;IAClD,IACE,gBAAgB,KAChB,cAAc,QACd,cAAc,SAAS,MAEvB;IAEF,MAAM,IAAI,SAAe,YAAY;KACnC,iBAAiB;IACnB,CAAC;GACH;GAEA,MAAM,oBAAoB,YAA2B;IACnD,OACE,kBAAkB,KAClB,CAAC,kBACD,CAAC,gBAAgB,QAAQ,MAAM,GAE/B,MAAM,IAAI,SAAe,YAAY;KACnC,MAAM,SAAS,QAAQ;KACvB,MAAM,gBAAsB;MAC1B,QAAQ,oBAAoB,SAAS,OAAO;MAC5C,QAAQ;KACV;KACA,MAAM,gBAAsB;MAC1B,QAAQ,oBAAoB,SAAS,OAAO;MAC5C,QAAQ;KACV;KACA,iBAAiB;KACjB,QAAQ,iBAAiB,SAAS,SAAS,EAAE,MAAM,KAAK,CAAC;KACzD,IAAI,gBAAgB,MAAM,GACxB,QAAQ;IAEZ,CAAC;GAEL;GAEA,MAAM,gBAAmD;IACvD,IAAI,CAAC,gBAAgB,GACnB;IAEF,MAAM,cAAc,aAAa;IACjC;IACA,IACE,mBAAmB,OACnB,mBAAmB,KAAK,aAAa,QACrC;KACA,aAAa,OAAO,GAAG,gBAAgB;KACvC,mBAAmB;IACrB;IACA,OAAO;GACT;GAEA,MAAM,UAAU,OACd,gBACkB;IAClB,MAAM,kBAAkB;IACxB,IAAI,kBAAkB,gBAAgB,QAAQ,MAAM,GAAG;KACrD,YAAY;KACZ,MAAM,IAAI,MAAM,uCAAuC;IACzD;IACA,aAAa,KAAK,WAAW;IAC7B,IAAI,YAAY,QACd,sBAAsB,YAAY;IAEpC,uBAAuB;GACzB;GAEA,MAAM,eAAe,OAAO,EAC1B,OACA,eACA,gBAAgB,IAChB,SAAS,OACT,aAAa,QAOM;IACnB,MAAM,QAAQ;KACZ;KACA;KACA;KACA;KACA,YAAY,SAAS,aAAa;IACpC,CAAC;GACH;GAEA,MAAM,eAAe,OACnB,sBACkB;IAClB,MAAM,QAAQ,kBAAkB;IAChC,IAAI,SAAS,MACX,MAAM,IAAI,MAAM,kCAAkC;IAGpD,MAAM,MAAM,kBAAkB;IAC9B,IAAI,OAAO,MACT,iBAAiB,IAAI,GAAG;IAG1B,MAAM,OAAO,MAAM;IACnB,MAAM,mBAAmB,MAAM;IAC/B,MAAM,gBAAgB,kBAAkB;IACxC,MAAM,cAAc,mBAAmB,MAAM,aAAa;IAC1D,MAAM,SAAS,KAAK,mBAAmB,KAAK,gBAAgB;IAC5D,MAAM,cAAc,SAChB,iBAAiB,WAAW,IAC5B,CAAC,WAAW;IAEhB,KAAK,MAAM,SAAS,aAAa;KAC/B,IAAI,aAAa;KACjB,IAAI,OAAO,SAAS,UAClB,aAAa;MACX,GAAG;MACH,OAAO,EAAE,MAAM,MAAM;KACvB;UACK,IACL,OAAO,kBAAkB,YACzB,oBAAoB,MAEpB,aAAa;MACX,GAAG;MACH,OAAO,EACL,kBAAkB;OAChB,GAAG;OACH,MAAM;MACR,EACF;KACF;KAGF,MAAM,aAAa,sCAAsC,UAAU;KACnE,MAAM,aAAa;MACjB,OAAO,KAAK,YAAY,YAAY,gBAAgB;MACpD,eAAe;MACf,eAAe,WAAW;MAC1B;MACA,YAAY,MAAM;KACpB,CAAC;IACH;GACF;GAEA,MAAM,YAAY,YAA2B;IAC3C,IAAI;KACF,WAAW,MAAM,SAAS,QAAQ;MAChC,IAAI,gBAAgB,QAAQ,MAAM,GAAG;OACnC,YAAY;OACZ,MAAM,IAAI,MAAM,uCAAuC;MACzD;MAEA,IAAI,MAAM,qBAAqB,MAAM;OACnC,MAAM,aAAa,sCACjB,MAAM,iBACR;OACA,IAAI,cAAc,MAAM;QACtB,MAAM,MAAM,MAAM,kBAAkB;QACpC,IAAI,OAAO,MAAM;SACf,iBAAiB,IAAI,GAAG;SACxB,IAAI,MAAM,kBAAkB,OAAO,WAAW,MAC5C,oBAAoB,IAAI,GAAG;QAE/B;QACA,MAAM,aAAa;SACjB,OAAO,KAAK,YAAY,YAAY,gBAAgB;SACpD,eAAe;SACf,eAAe,WAAW;QAC5B,CAAC;OACH;MACF,OAAO,IAAI,MAAM,qBAAqB,MACpC,MAAM,aAAa,MAAM,iBAAiB;WACrC,IAAI,MAAM,YAAY,MAC3B,MAAM,aAAa,EACjB,OAAO,6BAA6B,MAAM,UAAU,EAClD,YACF,CAAC,EACH,CAAC;WACI,IAAI,MAAM,oBAAoB,MAAM;OACzC,MAAM,UAAU,MAAM,iBAAiB;OACvC,IAAI,WAAW,MAAM;QACnB,iBAAiB,IAAI,OAAO;QAC5B,IAAI,qBAAqB,oBAAoB,IAAI,OAAO,GAAG;SACzD,MAAM,YAAY,+BAA+B,OAAO;SACxD,MAAM,aAAa;UACjB,OAAO;UACP,eAAe;UACf,eAAe,UAAU;SAC3B,CAAC;QACH;OACF;MACF,OACE,MAAM,aAAa,EACjB,OAAO,IAAI,oBAAoB;OAC7B,MAAM;OACN,SAAS,IAAI,eAAe;QAC1B,SAAS;QACT,mBAAmB,EAAE,GAAG,MAAM;OAChC,CAAC;MACH,CAAC,EACH,CAAC;KAEL;IACF,SAAS,OAAO;KACd,cAAc,QAAQ;IACxB,UAAU;KACR,cAAc,OAAO;KACrB,uBAAuB;IACzB;GACF,EAAA,CAAG;GAEH,IAAI;IACF,IAAI,gBAAgB;IACpB,OAAO,eAAe;KACpB,IAAI,gBAAgB,QAAQ,MAAM,GAAG;MACnC,YAAY;MACZ,MAAM,IAAI,MAAM,uCAAuC;KACzD;KAEA,MAAM,iBAAiB;KACvB,MAAM,cAAc,QAAQ;KAE5B,IAAI,CAAC,aAAa;MAChB,IAAI,cAAc,SAAS,MACzB,MAAM,cAAc;MAEtB,IAAI,cAAc,MAChB,gBAAgB;MAElB;KACF;KAEA,IAAI,YAAY,QAAQ;MACtB,qBAAqB,KAAK,IACxB,GACA,qBAAqB,YAAY,UACnC;MACA,uBAAuB;MACvB,MAAM,mBACJ,uBAAuB;OACrB,aAAa,2BACT,KAAK,mBACL;OACJ;OACA,KAAK,KAAK,IAAI;MAChB,CAAC,GACD,QAAQ,MACV;MACA,IAAI,gBAAgB,QAAQ,MAAM,GAAG;OACnC,YAAY;OACZ,MAAM,IAAI,MAAM,uCAAuC;MACzD;MACA,2BAA2B;MAC3B,uBAAuB,KAAK,IAAI;KAClC,OACE,uBAAuB;KAGzB,MAAM,YAAY;KAElB,IAAI,YAAY,iBAAiB,MAC/B,MAAM,YAAY,kBAChB,YAAY,eACZ,KAAA,GACA,KAAA,GACA,KAAA,GACA,KAAA,GACA,EAAE,OAAO,YAAY,cAAc,CACrC;IAEJ;GACF,UAAU;IACR,iBAAiB;IACjB,IAAI,CAAC,cAAc,MAAM;KACvB,YAAY;KACZ,uBAAuB;IACzB;IACA,MAAM;GACR;EACF,UAAU;GACR,QAAQ,QAAQ,oBAAoB,SAAS,WAAW;GACxD,IAAI,CAAC,sBAAsB,OAAO,SAChC,sBAAsB,MAAM;EAEhC;CACF;;;;;;;;;CAUA,YACE,OACA,kBACqB;EACrB,MAAM,UAAU,MAAM;EACtB,IAAI,EAAE,mBAAmB,iBACvB,OAAO;EAGT,MAAM,WAAW,QAAQ;EACzB,MAAM,aAAa,KAAK,yBAAyB,QAAQ;EACzD,MAAM,mBAAmB,cAAc;EAEvC,IAAI,UAAqC,QAAQ;EACjD,IAAI,kBAAkB;EAEtB,IAAI,MAAM,QAAQ,OAAO,KAAK,cAAc,MAAM;GAChD,UAAU,QAAQ,KAAK,UACrB,OAAO,UAAU,YAAY,EAAE,WAAW,SACtC;IAAE,GAAG;IAAO,OAAO;GAAW,IAC9B,KACN;GACA,kBAAkB;EACpB,OAAO,IACL,OAAO,YAAY,YACnB,YAAY,MACZ,cAAc,QACd,iBAAiB,OAAO,GACxB;GACA,UAAU,CAAC;IAAE,MAAM;IAAQ,MAAM;IAAS,OAAO;GAAW,CAAC;GAC7D,kBAAkB;EACpB;EAEA,IAAI,CAAC,mBAAmB,CAAC,kBACvB,OAAO;EAGT,MAAM,kBAAkB,mBACnB,KAAK,wBAAwB,QAAQ,IACtC;EAEJ,OAAO,IAAI,oBAAoB;GAC7B,MAAM,MAAM;GACZ,SAAS,IAAI,eAAe;IAC1B,GAAG;IACH;IACA,mBAAmB;GACrB,CAAC;GACD,gBAAgB,MAAM;EACxB,CAAC;CACH;;;;;CAMA,yBACE,UACoB;EACpB,IACE,uBAAuB,YACvB,OAAO,SAAS,sBAAsB,UAEtC,OAAO,SAAS;CAGpB;CAEA,wBAAgC,KAAuB;EACrD,IAAI,QAAQ,QAAQ,QAAQ,KAAA,GAC1B,OAAO;EAGT,IAAI,MAAM,QAAQ,GAAG,GACnB,OAAO,IAAI,KAAK,SAAS,KAAK,wBAAwB,IAAI,CAAC;EAG7D,IAAI,OAAO,QAAQ,UAAU;GAC3B,MAAM,UAAmC,CAAC;GAC1C,KAAK,MAAM,CAAC,KAAK,UAAU,OAAO,QAAQ,GAAG,GAC3C,IAAI,QAAQ,qBACV,QAAQ,OAAO,KAAK,wBAAwB,KAAK;GAGrD,OAAO;EACT;EAEA,OAAO;CACT;AACF"}

package/dist/esm/llm/contextOverflowRecovery.mjs

@@ -0,0 +1,130 @@
+import { getContextOverflowInfo } from "../utils/errors.mjs";
+//#region src/llm/contextOverflowRecovery.ts
+/** Fraction of the previous budget used when the provider named no ceiling. */
+const BLIND_SHRINK_RATIO = .7;
+/** Slack left below a known ceiling so the retry is not sized to the edge. */
+const CEILING_HEADROOM_RATIO = .95;
+/**
+* Fallback floor, used only when the instruction size is unknown. When it is
+* known the floor is derived from it instead, so a genuinely small model — a
+* 4k window, where a 95%-of-ceiling budget lands below this constant — can
+* still recover.
+*/
+const MIN_RECOVERY_BUDGET_TOKENS = 4e3;
+/**
+* Room a corrected budget must leave above the instructions for the messages
+* themselves. Without it, a budget that merely clears the system prompt and
+* tool schemas is not a budget anything can be compacted into.
+*/
+const MIN_MESSAGE_HEADROOM_TOKENS = 2e3;
+function isUsable(value) {
+	return value != null && Number.isFinite(value) && value > 0;
+}
+/**
+* Converts a provider-space retry budget into the units consumed by a pruner
+* calibrated for another provider while preserving the same raw-token limit.
+*/
+function translateRecoveryBudget(budgetTokens, sourceCalibrationRatio, targetCalibrationRatio) {
+	if (!isUsable(budgetTokens) || !isUsable(sourceCalibrationRatio) || !isUsable(targetCalibrationRatio)) return budgetTokens;
+	return Math.floor(budgetTokens * targetCalibrationRatio / sourceCalibrationRatio);
+}
+/** Applies the conservative shrink used when no provider-space ceiling is usable. */
+function getBlindRecoveryBudget(maxContextTokens) {
+	return isUsable(maxContextTokens) ? Math.floor(maxContextTokens * BLIND_SHRINK_RATIO) : void 0;
+}
+/**
+* The completion allowance the provider counted against the same ceiling,
+* when it reported both the total and the prompt portion. The retry budget
+* governs the prompt only, so this has to come off the ceiling first —
+* otherwise a large `maxTokens` keeps the request over the limit no matter
+* how far the prompt is compacted.
+*/
+function reservedForCompletion(info, configuredCompletionTokens) {
+	if (isUsable(info.requestedTokens) && isUsable(info.promptTokens)) {
+		const difference = info.requestedTokens - info.promptTokens;
+		if (difference > 0) return difference;
+	}
+	/**
+	* No breakdown on offer. Fall back to what the caller configured, because
+	* the provider still counts it: targeting the whole ceiling would leave the
+	* retry at `prompt + maxTokens` and over the limit however far the prompt
+	* is compacted.
+	*/
+	return isUsable(configuredCompletionTokens) ? configuredCompletionTokens : 0;
+}
+function resolveTargetBudget(info, maxContextTokens, estimatedPromptTokens, configuredCompletionTokens) {
+	if (isUsable(info.limitTokens)) {
+		const promptCeiling = info.limitTokens - reservedForCompletion(info, configuredCompletionTokens);
+		/**
+		* A completion allowance at or above the ceiling leaves nothing for the
+		* prompt: even an empty one plus the requested output overruns the limit.
+		* Compaction cannot fix that, so declining surfaces the real problem
+		* instead of burning the recovery budget on retries that must fail.
+		*/
+		if (promptCeiling <= 0) return null;
+		return promptCeiling * CEILING_HEADROOM_RATIO;
+	}
+	if (isUsable(estimatedPromptTokens)) return estimatedPromptTokens * BLIND_SHRINK_RATIO;
+	if (isUsable(maxContextTokens)) return maxContextTokens * BLIND_SHRINK_RATIO;
+	return null;
+}
+/**
+* Decides whether a failed model call is a recoverable context overflow and,
+* if so, what budget the retry should be re-pruned against.
+*
+* Returns `null` when the error is something compaction cannot fix, or when
+* the per-run recovery budget is spent — in both cases the caller should let
+* its normal failure handling proceed.
+*/
+function planContextOverflowRecovery({ error, provider, maxContextTokens, estimatedPromptTokens, calibrationRatio, instructionTokens, canSummarize = false, configuredCompletionTokens, attemptsSoFar, maxAttempts = 2 }) {
+	if (attemptsSoFar >= maxAttempts) return null;
+	const info = getContextOverflowInfo(error, {
+		provider,
+		estimatedPromptTokens,
+		maxContextTokens
+	});
+	if (info == null) return null;
+	const currentCalibrationRatio = isUsable(calibrationRatio) ? calibrationRatio : 1;
+	const estimatedMessageTokens = isUsable(estimatedPromptTokens) && isUsable(instructionTokens) ? estimatedPromptTokens - instructionTokens : estimatedPromptTokens;
+	const observedMessageTokens = isUsable(info.promptTokens) && isUsable(instructionTokens) ? info.promptTokens - instructionTokens : info.promptTokens;
+	const observedCalibrationRatio = isUsable(observedMessageTokens) && isUsable(estimatedMessageTokens) ? observedMessageTokens / estimatedMessageTokens * currentCalibrationRatio : void 0;
+	const target = resolveTargetBudget(info, maxContextTokens, estimatedPromptTokens, configuredCompletionTokens);
+	if (target == null) return isUsable(info.limitTokens) || isUsable(maxContextTokens) || isUsable(estimatedPromptTokens) ? null : {
+		info,
+		observedCalibrationRatio,
+		budgetTokens: void 0
+	};
+	/**
+	* A retry that does not actually shrink the prompt would just reproduce the
+	* same rejection, so a ceiling that lands at or above the budget we already
+	* had is replaced by a blind shrink.
+	*/
+	const bounded = isUsable(maxContextTokens) && target >= maxContextTokens ? getBlindRecoveryBudget(maxContextTokens) ?? target : target;
+	const budgetTokens = Math.floor(bounded);
+	/**
+	* Below the floor there is no usable budget left: either nothing survives
+	* pruning, or the instructions alone fill the window, in which case the
+	* summarize node refuses to run and the detour would bounce between the
+	* agent and summarize nodes without ever shrinking the prompt. Declining
+	* lets the existing "instructions exceed context budget" guidance surface
+	* instead.
+	*/
+	let floorTokens = MIN_RECOVERY_BUDGET_TOKENS;
+	if (isUsable(instructionTokens)) floorTokens = instructionTokens + MIN_MESSAGE_HEADROOM_TOKENS;
+	else if (canSummarize) floorTokens = 1;
+	if (budgetTokens < floorTokens) return null;
+	/**
+	* Refuse to report a "recovery" that changes nothing: when the budget in
+	* force is already at or below the target, re-pruning cannot free space.
+	*/
+	if (isUsable(maxContextTokens) && budgetTokens >= maxContextTokens) return null;
+	return {
+		budgetTokens,
+		info,
+		observedCalibrationRatio
+	};
+}
+//#endregion
+export { getBlindRecoveryBudget, planContextOverflowRecovery, translateRecoveryBudget };
+
+//# sourceMappingURL=contextOverflowRecovery.mjs.map

package/dist/esm/llm/contextOverflowRecovery.mjs.map

@@ -0,0 +1 @@
+{"version":3,"file":"contextOverflowRecovery.mjs","names":[],"sources":["../../../src/llm/contextOverflowRecovery.ts"],"sourcesContent":["/**\n * Recovery policy for provider context-overflow rejections.\n *\n * Detection (`@/utils/errors`) answers \"was this an overflow, and what did\n * the provider disclose?\". This module answers the follow-up: \"what budget\n * should the retry target?\" — deliberately kept as pure functions so the\n * policy can be reasoned about and tested without a graph.\n *\n * On units: `maxContextTokens` is a **provider-space** budget. The pruner\n * converts it into its own raw estimate space by dividing by the\n * `calibrationRatio` it learns from reported usage. A provider-reported\n * ceiling is therefore applied verbatim — converting it here as well would\n * apply the same correction twice and prune toward roughly `limit / ratio²`,\n * silently discarding far more history than the overflow called for.\n *\n * `observedCalibrationRatio` is returned separately so the caller can seed\n * the pruner's conversion without folding the same correction into the\n * provider-space budget.\n */\nimport type { ContextOverflowInfo } from '@/utils/errors';\nimport type { Providers } from '@/common';\nimport { getContextOverflowInfo } from '@/utils/errors';\n\n/** Fraction of the previous budget used when the provider named no ceiling. */\nconst BLIND_SHRINK_RATIO = 0.7;\n\n/** Slack left below a known ceiling so the retry is not sized to the edge. */\nconst CEILING_HEADROOM_RATIO = 0.95;\n\n/**\n * Fallback floor, used only when the instruction size is unknown. When it is\n * known the floor is derived from it instead, so a genuinely small model — a\n * 4k window, where a 95%-of-ceiling budget lands below this constant — can\n * still recover.\n */\nconst MIN_RECOVERY_BUDGET_TOKENS = 4_000;\n\n/**\n * Room a corrected budget must leave above the instructions for the messages\n * themselves. Without it, a budget that merely clears the system prompt and\n * tool schemas is not a budget anything can be compacted into.\n */\nconst MIN_MESSAGE_HEADROOM_TOKENS = 2_000;\n\n/** Bound on forced-compaction retries per agent, per run. */\nexport const DEFAULT_MAX_OVERFLOW_RECOVERIES = 2;\n\nexport interface OverflowRecoveryPlan {\n  /** Budget the retry should target, in provider token units when known. */\n  budgetTokens?: number;\n  /** What the provider disclosed. Carried through for logging. */\n  info: ContextOverflowInfo;\n  /**\n   * Provider-reported message tokens divided by our own message estimate,\n   * when both are known. Greater than 1 means we under-count relative to this\n   * provider.\n   *\n   * Returned separately so the graph can seed the pruner's calibration;\n   * applying it to this plan's budget as well would double-count. Fixed\n   * instruction overhead is removed before deriving it from\n   * `info.promptTokens`, never `info.requestedTokens`, since several providers\n   * fold the completion allowance into the latter.\n   */\n  observedCalibrationRatio?: number;\n}\n\nexport interface OverflowRecoveryParams {\n  error: unknown;\n  provider: Providers;\n  /** Budget in force when the rejected prompt was built. */\n  maxContextTokens?: number;\n  /** Our own estimate of the prompt we actually sent. */\n  estimatedPromptTokens?: number;\n  /** Provider/local calibration already applied to the prompt estimate. */\n  calibrationRatio?: number;\n  /**\n   * System prompt plus tool schemas — the part of the budget compaction\n   * cannot touch. A corrected budget at or below this leaves no room for\n   * messages, and the summarize node refuses to run, so recovery is declined\n   * rather than entered.\n   */\n  instructionTokens?: number;\n  /** Whether a model-backed summary can compact messages without a pruner. */\n  canSummarize?: boolean;\n  /**\n   * Completion allowance the caller configured. Providers count it against\n   * the same ceiling, so it has to come off the top when the error itself did\n   * not break the total down.\n   */\n  configuredCompletionTokens?: number;\n  /** Recoveries already attempted for this agent in this run. */\n  attemptsSoFar: number;\n  maxAttempts?: number;\n}\n\nfunction isUsable(value: number | undefined): value is number {\n  return value != null && Number.isFinite(value) && value > 0;\n}\n\n/**\n * Converts a provider-space retry budget into the units consumed by a pruner\n * calibrated for another provider while preserving the same raw-token limit.\n */\nexport function translateRecoveryBudget(\n  budgetTokens: number | undefined,\n  sourceCalibrationRatio: number | undefined,\n  targetCalibrationRatio: number | undefined\n): number | undefined {\n  if (\n    !isUsable(budgetTokens) ||\n    !isUsable(sourceCalibrationRatio) ||\n    !isUsable(targetCalibrationRatio)\n  ) {\n    return budgetTokens;\n  }\n  return Math.floor(\n    (budgetTokens * targetCalibrationRatio) / sourceCalibrationRatio\n  );\n}\n\n/** Applies the conservative shrink used when no provider-space ceiling is usable. */\nexport function getBlindRecoveryBudget(\n  maxContextTokens: number | undefined\n): number | undefined {\n  return isUsable(maxContextTokens)\n    ? Math.floor(maxContextTokens * BLIND_SHRINK_RATIO)\n    : undefined;\n}\n\n/**\n * The completion allowance the provider counted against the same ceiling,\n * when it reported both the total and the prompt portion. The retry budget\n * governs the prompt only, so this has to come off the ceiling first —\n * otherwise a large `maxTokens` keeps the request over the limit no matter\n * how far the prompt is compacted.\n */\nfunction reservedForCompletion(\n  info: ContextOverflowInfo,\n  configuredCompletionTokens: number | undefined\n): number {\n  if (isUsable(info.requestedTokens) && isUsable(info.promptTokens)) {\n    const difference = info.requestedTokens - info.promptTokens;\n    if (difference > 0) {\n      return difference;\n    }\n  }\n  /**\n   * No breakdown on offer. Fall back to what the caller configured, because\n   * the provider still counts it: targeting the whole ceiling would leave the\n   * retry at `prompt + maxTokens` and over the limit however far the prompt\n   * is compacted.\n   */\n  return isUsable(configuredCompletionTokens) ? configuredCompletionTokens : 0;\n}\n\nfunction resolveTargetBudget(\n  info: ContextOverflowInfo,\n  maxContextTokens: number | undefined,\n  estimatedPromptTokens: number | undefined,\n  configuredCompletionTokens: number | undefined\n): number | null {\n  if (isUsable(info.limitTokens)) {\n    const promptCeiling =\n      info.limitTokens -\n      reservedForCompletion(info, configuredCompletionTokens);\n    /**\n     * A completion allowance at or above the ceiling leaves nothing for the\n     * prompt: even an empty one plus the requested output overruns the limit.\n     * Compaction cannot fix that, so declining surfaces the real problem\n     * instead of burning the recovery budget on retries that must fail.\n     */\n    if (promptCeiling <= 0) {\n      return null;\n    }\n\n    return promptCeiling * CEILING_HEADROOM_RATIO;\n  }\n  if (isUsable(estimatedPromptTokens)) {\n    return estimatedPromptTokens * BLIND_SHRINK_RATIO;\n  }\n  if (isUsable(maxContextTokens)) {\n    return maxContextTokens * BLIND_SHRINK_RATIO;\n  }\n  return null;\n}\n\n/**\n * Decides whether a failed model call is a recoverable context overflow and,\n * if so, what budget the retry should be re-pruned against.\n *\n * Returns `null` when the error is something compaction cannot fix, or when\n * the per-run recovery budget is spent — in both cases the caller should let\n * its normal failure handling proceed.\n */\nexport function planContextOverflowRecovery({\n  error,\n  provider,\n  maxContextTokens,\n  estimatedPromptTokens,\n  calibrationRatio,\n  instructionTokens,\n  canSummarize = false,\n  configuredCompletionTokens,\n  attemptsSoFar,\n  maxAttempts = DEFAULT_MAX_OVERFLOW_RECOVERIES,\n}: OverflowRecoveryParams): OverflowRecoveryPlan | null {\n  if (attemptsSoFar >= maxAttempts) {\n    return null;\n  }\n\n  const info = getContextOverflowInfo(error, {\n    provider,\n    estimatedPromptTokens,\n    maxContextTokens,\n  });\n  if (info == null) {\n    return null;\n  }\n\n  const currentCalibrationRatio = isUsable(calibrationRatio)\n    ? calibrationRatio\n    : 1;\n  const estimatedMessageTokens =\n    isUsable(estimatedPromptTokens) && isUsable(instructionTokens)\n      ? estimatedPromptTokens - instructionTokens\n      : estimatedPromptTokens;\n  const observedMessageTokens =\n    isUsable(info.promptTokens) && isUsable(instructionTokens)\n      ? info.promptTokens - instructionTokens\n      : info.promptTokens;\n  const observedCalibrationRatio =\n    isUsable(observedMessageTokens) && isUsable(estimatedMessageTokens)\n      ? (observedMessageTokens / estimatedMessageTokens) *\n        currentCalibrationRatio\n      : undefined;\n\n  const target = resolveTargetBudget(\n    info,\n    maxContextTokens,\n    estimatedPromptTokens,\n    configuredCompletionTokens\n  );\n  if (target == null) {\n    const hasNumericBasis =\n      isUsable(info.limitTokens) ||\n      isUsable(maxContextTokens) ||\n      isUsable(estimatedPromptTokens);\n    return hasNumericBasis\n      ? null\n      : { info, observedCalibrationRatio, budgetTokens: undefined };\n  }\n\n  /**\n   * A retry that does not actually shrink the prompt would just reproduce the\n   * same rejection, so a ceiling that lands at or above the budget we already\n   * had is replaced by a blind shrink.\n   */\n  const bounded =\n    isUsable(maxContextTokens) && target >= maxContextTokens\n      ? (getBlindRecoveryBudget(maxContextTokens) ?? target)\n      : target;\n\n  const budgetTokens = Math.floor(bounded);\n\n  /**\n   * Below the floor there is no usable budget left: either nothing survives\n   * pruning, or the instructions alone fill the window, in which case the\n   * summarize node refuses to run and the detour would bounce between the\n   * agent and summarize nodes without ever shrinking the prompt. Declining\n   * lets the existing \"instructions exceed context budget\" guidance surface\n   * instead.\n   */\n  let floorTokens = MIN_RECOVERY_BUDGET_TOKENS;\n  if (isUsable(instructionTokens)) {\n    floorTokens = instructionTokens + MIN_MESSAGE_HEADROOM_TOKENS;\n  } else if (canSummarize) {\n    floorTokens = 1;\n  }\n  if (budgetTokens < floorTokens) {\n    return null;\n  }\n\n  /**\n   * Refuse to report a \"recovery\" that changes nothing: when the budget in\n   * force is already at or below the target, re-pruning cannot free space.\n   */\n  if (isUsable(maxContextTokens) && budgetTokens >= maxContextTokens) {\n    return null;\n  }\n\n  return { budgetTokens, info, observedCalibrationRatio };\n}\n"],"mappings":";;;AAwBA,MAAM,qBAAqB;;AAG3B,MAAM,yBAAyB;;;;;;;AAQ/B,MAAM,6BAA6B;;;;;;AAOnC,MAAM,8BAA8B;AAqDpC,SAAS,SAAS,OAA4C;CAC5D,OAAO,SAAS,QAAQ,OAAO,SAAS,KAAK,KAAK,QAAQ;AAC5D;;;;;AAMA,SAAgB,wBACd,cACA,wBACA,wBACoB;CACpB,IACE,CAAC,SAAS,YAAY,KACtB,CAAC,SAAS,sBAAsB,KAChC,CAAC,SAAS,sBAAsB,GAEhC,OAAO;CAET,OAAO,KAAK,MACT,eAAe,yBAA0B,sBAC5C;AACF;;AAGA,SAAgB,uBACd,kBACoB;CACpB,OAAO,SAAS,gBAAgB,IAC5B,KAAK,MAAM,mBAAmB,kBAAkB,IAChD,KAAA;AACN;;;;;;;;AASA,SAAS,sBACP,MACA,4BACQ;CACR,IAAI,SAAS,KAAK,eAAe,KAAK,SAAS,KAAK,YAAY,GAAG;EACjE,MAAM,aAAa,KAAK,kBAAkB,KAAK;EAC/C,IAAI,aAAa,GACf,OAAO;CAEX;;;;;;;CAOA,OAAO,SAAS,0BAA0B,IAAI,6BAA6B;AAC7E;AAEA,SAAS,oBACP,MACA,kBACA,uBACA,4BACe;CACf,IAAI,SAAS,KAAK,WAAW,GAAG;EAC9B,MAAM,gBACJ,KAAK,cACL,sBAAsB,MAAM,0BAA0B;;;;;;;EAOxD,IAAI,iBAAiB,GACnB,OAAO;EAGT,OAAO,gBAAgB;CACzB;CACA,IAAI,SAAS,qBAAqB,GAChC,OAAO,wBAAwB;CAEjC,IAAI,SAAS,gBAAgB,GAC3B,OAAO,mBAAmB;CAE5B,OAAO;AACT;;;;;;;;;AAUA,SAAgB,4BAA4B,EAC1C,OACA,UACA,kBACA,uBACA,kBACA,mBACA,eAAe,OACf,4BACA,eACA,cAAA,KACsD;CACtD,IAAI,iBAAiB,aACnB,OAAO;CAGT,MAAM,OAAO,uBAAuB,OAAO;EACzC;EACA;EACA;CACF,CAAC;CACD,IAAI,QAAQ,MACV,OAAO;CAGT,MAAM,0BAA0B,SAAS,gBAAgB,IACrD,mBACA;CACJ,MAAM,yBACJ,SAAS,qBAAqB,KAAK,SAAS,iBAAiB,IACzD,wBAAwB,oBACxB;CACN,MAAM,wBACJ,SAAS,KAAK,YAAY,KAAK,SAAS,iBAAiB,IACrD,KAAK,eAAe,oBACpB,KAAK;CACX,MAAM,2BACJ,SAAS,qBAAqB,KAAK,SAAS,sBAAsB,IAC7D,wBAAwB,yBACzB,0BACA,KAAA;CAEN,MAAM,SAAS,oBACb,MACA,kBACA,uBACA,0BACF;CACA,IAAI,UAAU,MAKZ,OAHE,SAAS,KAAK,WAAW,KACzB,SAAS,gBAAgB,KACzB,SAAS,qBAAqB,IAE5B,OACA;EAAE;EAAM;EAA0B,cAAc,KAAA;CAAU;;;;;;CAQhE,MAAM,UACJ,SAAS,gBAAgB,KAAK,UAAU,mBACnC,uBAAuB,gBAAgB,KAAK,SAC7C;CAEN,MAAM,eAAe,KAAK,MAAM,OAAO;;;;;;;;;CAUvC,IAAI,cAAc;CAClB,IAAI,SAAS,iBAAiB,GAC5B,cAAc,oBAAoB;MAC7B,IAAI,cACT,cAAc;CAEhB,IAAI,eAAe,aACjB,OAAO;;;;;CAOT,IAAI,SAAS,gBAAgB,KAAK,gBAAgB,kBAChD,OAAO;CAGT,OAAO;EAAE;EAAc;EAAM;CAAyB;AACxD"}

package/dist/esm/llm/google/index.mjs

@@ -1,8 +1,8 @@
 import { convertBaseMessagesToContent, convertResponseContentToChatGenerationChunk, dropUnsupportedModelTurnPrefill, mapGenerateContentResultToChatResult } from "./utils/common.mjs";
 import { AIMessageChunk } from "@langchain/core/messages";
 import { ChatGenerationChunk } from "@langchain/core/outputs";
-import { getEnvironmentVariable } from "@langchain/core/utils/env";
 import { ChatGoogleGenerativeAI } from "@langchain/google-genai";
+import { getEnvironmentVariable } from "@langchain/core/utils/env";
 import { FunctionCallingMode, GoogleGenerativeAI } from "@google/generative-ai";
 //#region src/llm/google/index.ts
 var CustomChatGoogleGenerativeAI = class extends ChatGoogleGenerativeAI {

package/dist/esm/llm/invoke.mjs

@@ -4,6 +4,7 @@ import { modifyDeltaProperties } from "../messages/core.mjs";
 import "../messages/index.mjs";
 import { annotateMessagesForLLM } from "../tools/toolOutputReferences.mjs";
 import { ChatModelStreamHandler } from "../stream.mjs";
+import { getContextOverflowInfo } from "../utils/errors.mjs";
 import { assertNotTruncatedToolCall } from "./truncation.mjs";
 import { manualToolStreamProviders } from "./providers.mjs";
 import { initializeModel } from "./init.mjs";
@@ -146,6 +147,22 @@ async function attemptInvoke({ model, messages, provider, context, onChunk }, co
 	assertNotTruncatedToolCall(finalMessage, provider);
 	return { messages: [finalMessage] };
 }
+const fallbackErrorContexts = /* @__PURE__ */ new WeakMap();
+const fallbackOverflowCandidates = /* @__PURE__ */ new WeakMap();
+function attachFallbackErrorContext(error, fallbackContext) {
+	if (typeof error !== "object" || error === null) return;
+	fallbackErrorContexts.set(error, fallbackContext);
+}
+/** Reads back the fallback attribution attached by `tryFallbackProviders`. */
+function getFallbackErrorContext(error) {
+	if (typeof error !== "object" || error === null) return;
+	return fallbackErrorContexts.get(error);
+}
+/** Returns every fallback overflow retained from an exhausted provider chain. */
+function getFallbackOverflowCandidates(error) {
+	if (typeof error !== "object" || error === null) return [];
+	return [...fallbackOverflowCandidates.get(error) ?? []];
+}
 /**
 * Best-effort read of the configured model name from client options.
 * Providers disagree on the key (`model` vs `modelName`).
@@ -157,10 +174,24 @@ function extractClientOptionsModel(clientOptions) {
 }
 /**
 * Attempts each fallback provider in order until one succeeds.
-* Throws the last error if all fallbacks fail.
+*
+* When every fallback fails, a context overflow among them is thrown in
+* preference to whichever failure happened to come last. An overflow is the
+* one failure the caller can act on — it compacts and retries — and losing it
+* behind a later unrelated error would surface a dead end instead. Ordinary
+* failures still throw last-error-wins.
 */
-async function tryFallbackProviders({ fallbacks, tools, messages, config, primaryError, context, onChunk }) {
+async function tryFallbackProviders({ fallbacks, tools, messages, config, primaryError, context, onChunk, overflowContext }) {
+	const isOverflow = (error, contextOverride = overflowContext) => getContextOverflowInfo(error, contextOverride) != null;
 	let lastError = primaryError;
+	/**
+	* Tracked apart from the primary's overflow. A caller reaching this
+	* function with an overflowing primary has already failed to recover from
+	* it, so a fallback overflow — which may sit against a different window and
+	* output allowance — is the more useful of the two to surface.
+	*/
+	const overflowCandidates = [];
+	const primaryOverflowError = isOverflow(primaryError) ? primaryError : void 0;
 	for (const fb of fallbacks) try {
 		const fbModel = initializeModel({
 			provider: fb.provider,
@@ -191,11 +222,35 @@ async function tryFallbackProviders({ fallbacks, tools, messages, config, primar
 		}, fbConfig);
 	} catch (e) {
 		lastError = e;
+		if (isOverflow(e, {
+			provider: fb.provider,
+			maxContextTokens: fb.maxContextTokens,
+			...overflowContext?.provider === fb.provider ? { estimatedPromptTokens: overflowContext.estimatedPromptTokens } : {}
+		})) {
+			const errorContext = {
+				provider: fb.provider,
+				clientOptions: fb.clientOptions,
+				maxContextTokens: fb.maxContextTokens
+			};
+			attachFallbackErrorContext(e, errorContext);
+			overflowCandidates.push({
+				error: e,
+				context: errorContext
+			});
+		}
 		continue;
 	}
-	if (lastError !== void 0) throw lastError;
+	/**
+	* Preference order: a fallback overflow, then the primary's overflow, then
+	* whichever failure came last. An overflow is the only one of the three a
+	* caller can act on, and the fallback's carries the client attribution that
+	* makes a correct retry budget possible.
+	*/
+	const preferred = overflowCandidates[0]?.error ?? primaryOverflowError ?? lastError;
+	if (overflowCandidates.length > 0 && typeof preferred === "object" && preferred !== null) fallbackOverflowCandidates.set(preferred, overflowCandidates);
+	if (preferred !== void 0) throw preferred;
 }
 //#endregion
-export { attemptInvoke, tryFallbackProviders };
+export { attemptInvoke, getFallbackErrorContext, getFallbackOverflowCandidates, tryFallbackProviders };
 
 //# sourceMappingURL=invoke.mjs.map

package/dist/esm/llm/invoke.mjs.map

@@ -1 +1 @@
-{"version":3,"file":"invoke.mjs","names":[],"sources":["../../../src/llm/invoke.ts"],"sourcesContent":["import { concat } from '@langchain/core/utils/stream';\nimport { AIMessageChunk } from '@langchain/core/messages';\nimport type { RunnableConfig } from '@langchain/core/runnables';\nimport type { ToolCall } from '@langchain/core/messages/tool';\nimport type { BaseMessage } from '@langchain/core/messages';\nimport type { ToolOutputReferenceRegistry } from '@/tools/toolOutputReferences';\nimport type * as t from '@/types';\nimport { annotateMessagesForLLM } from '@/tools/toolOutputReferences';\nimport { assertNotTruncatedToolCall } from '@/llm/truncation';\nimport { Constants, GraphEvents, Providers } from '@/common';\nimport { manualToolStreamProviders } from '@/llm/providers';\nimport { modifyDeltaProperties } from '@/messages';\nimport { ChatModelStreamHandler } from '@/stream';\nimport { initializeModel } from '@/llm/init';\n\n/**\n * Context passed to `attemptInvoke`. Matches the subset of Graph that\n * `ChatModelStreamHandler.handle` needs *plus* the explicit\n * `getOrCreateToolOutputRegistry()` accessor that `attemptInvoke`\n * itself calls to pull the run-scoped tool-output registry off the\n * graph and project each relevant ToolMessage into a transient\n * annotated copy before the provider call.\n *\n * The intersection is intentional: `Parameters<...>[3]` resolves\n * indirectly through the stream handler's signature (which returns\n * `StandardGraph` and already exposes the accessor since #117), but\n * stating it explicitly here surfaces the contract at the call site —\n * a developer reading `attemptInvoke` doesn't have to chase the\n * upstream handler's parameter list to discover that\n * `context?.getOrCreateToolOutputRegistry()` is a real thing. Single\n * optional chain only — the method itself is required on the\n * `StandardGraph` branch of the intersection, so the second `?.` is\n * unnecessary at the call site.\n *\n * `NonNullable<...>` strips `undefined` from the upstream parameter\n * type so the intersection doesn't collapse to `never` on the\n * undefined branch; callers express optionality via `context?:\n * InvokeContext` on the function signature instead.\n *\n * Callers without a registry (e.g. summarization) simply pass no\n * `context` and the transform safely no-ops.\n */\nexport type InvokeContext = NonNullable<\n  Parameters<ChatModelStreamHandler['handle']>[3]\n> & {\n  getOrCreateToolOutputRegistry?(): ToolOutputReferenceRegistry | undefined;\n};\n\n/**\n * Per-chunk callback for custom stream processing.\n * When provided, replaces the default `ChatModelStreamHandler`.\n */\nexport type OnChunk = (chunk: AIMessageChunk) => void | Promise<void>;\n\nfunction getRegisteredDefaultChatStreamHandler(\n  context?: InvokeContext\n): ChatModelStreamHandler | undefined {\n  const handler = context?.handlerRegistry?.getHandler(\n    GraphEvents.CHAT_MODEL_STREAM\n  );\n  return handler instanceof ChatModelStreamHandler ? handler : undefined;\n}\n\nfunction hasReasoningDetails(chunk: AIMessageChunk): boolean {\n  const reasoningDetails = chunk.additional_kwargs.reasoning_details;\n  return Array.isArray(reasoningDetails) && reasoningDetails.length > 0;\n}\n\nfunction removeOpenRouterFinalReasoningReplayContent({\n  current,\n  next,\n  provider,\n}: {\n  current?: AIMessageChunk;\n  next: AIMessageChunk;\n  provider: Providers;\n}): AIMessageChunk {\n  const content = getOpenRouterFinalReasoningContent({\n    current,\n    next,\n    provider,\n  });\n  if (content == null || content === next.content) {\n    return next;\n  }\n\n  return new AIMessageChunk(\n    Object.assign({}, next, {\n      content,\n    })\n  );\n}\n\nfunction getOpenRouterFinalReasoningContent({\n  current,\n  next,\n  provider,\n}: {\n  current?: AIMessageChunk;\n  next: AIMessageChunk;\n  provider: Providers;\n}): string | undefined {\n  if (\n    provider !== Providers.OPENROUTER ||\n    current == null ||\n    !hasReasoningDetails(next) ||\n    typeof current.content !== 'string' ||\n    current.content === '' ||\n    typeof next.content !== 'string' ||\n    next.content === ''\n  ) {\n    return undefined;\n  }\n  if (!next.content.startsWith(current.content)) {\n    return next.content;\n  }\n  return next.content.slice(current.content.length);\n}\n\nfunction removeReasoningDetails(\n  additionalKwargs: AIMessageChunk['additional_kwargs']\n): AIMessageChunk['additional_kwargs'] {\n  return Object.fromEntries(\n    Object.entries(additionalKwargs).filter(\n      ([key]) => key !== 'reasoning_details'\n    )\n  );\n}\n\nfunction getStreamHandlingChunk({\n  current,\n  next,\n  provider,\n}: {\n  current?: AIMessageChunk;\n  next: AIMessageChunk;\n  provider: Providers;\n}): AIMessageChunk | undefined {\n  const content = getOpenRouterFinalReasoningContent({\n    current,\n    next,\n    provider,\n  });\n  if (content == null) {\n    return next;\n  }\n  if (content === '') {\n    return undefined;\n  }\n  return new AIMessageChunk(\n    Object.assign({}, next, {\n      content,\n      additional_kwargs: removeReasoningDetails(next.additional_kwargs),\n    })\n  );\n}\n\nfunction appendStreamChunk({\n  current,\n  next,\n  provider,\n}: {\n  current?: AIMessageChunk;\n  next: AIMessageChunk;\n  provider: Providers;\n}): AIMessageChunk {\n  if (current == null) {\n    return next;\n  }\n  return concat(\n    current,\n    removeOpenRouterFinalReasoningReplayContent({ current, next, provider })\n  );\n}\n\n/**\n * Invokes a chat model with the given messages, handling both streaming and\n * non-streaming paths.\n *\n * By default, stream chunks are processed through a `ChatModelStreamHandler`\n * that dispatches run steps (MESSAGE_CREATION, TOOL_CALLS) for the graph.\n * Pass an `onChunk` callback to override this with custom chunk processing\n * (e.g. summarization delta events).\n */\nexport async function attemptInvoke(\n  {\n    model,\n    messages,\n    provider,\n    context,\n    onChunk,\n  }: {\n    model: t.ChatModel;\n    messages: BaseMessage[];\n    provider: Providers;\n    context?: InvokeContext;\n    onChunk?: OnChunk;\n  },\n  config?: RunnableConfig\n): Promise<Partial<t.BaseGraphState>> {\n  /**\n   * Pull the run-scoped tool output registry off the graph (when one\n   * exists) and project ToolMessages carrying ref metadata into a\n   * transient annotated copy. The original `messages` array stays\n   * untouched so the graph state never sees `[ref: …]` / `_ref`\n   * payload.\n   */\n  const registry = context?.getOrCreateToolOutputRegistry();\n  const runId = config?.configurable?.run_id as string | undefined;\n  const messagesForProvider = annotateMessagesForLLM(messages, registry, runId);\n\n  /**\n   * Stamp the provider that is ACTUALLY serving this invocation onto the\n   * callback metadata. `attemptInvoke` is the single funnel for primary,\n   * fallback, and summarization model calls, so consumers that need\n   * provider attribution per call (the subagent usage-capture handler)\n   * read this key instead of trusting static agent config — which is\n   * wrong for fallback-served calls — or `ls_provider` — which derived\n   * providers inherit from their base class.\n   */\n  config = {\n    ...config,\n    metadata: {\n      ...(config?.metadata ?? {}),\n      [Constants.INVOKED_PROVIDER]: provider,\n    },\n  };\n\n  if (model.stream) {\n    const stream = await model.stream(messagesForProvider, config);\n    let finalChunk: AIMessageChunk | undefined;\n    const registeredStreamHandler =\n      getRegisteredDefaultChatStreamHandler(context);\n\n    if (onChunk) {\n      for await (const chunk of stream) {\n        await onChunk(chunk);\n        finalChunk = appendStreamChunk({\n          current: finalChunk,\n          next: chunk,\n          provider,\n        });\n      }\n    } else if (registeredStreamHandler == null) {\n      const metadata = config.metadata as Record<string, unknown> | undefined;\n      const streamHandler = new ChatModelStreamHandler();\n      for await (const chunk of stream) {\n        const handlingChunk = getStreamHandlingChunk({\n          current: finalChunk,\n          next: chunk,\n          provider,\n        });\n        if (handlingChunk != null) {\n          await streamHandler.handle(\n            GraphEvents.CHAT_MODEL_STREAM,\n            { chunk: handlingChunk },\n            metadata,\n            context\n          );\n        }\n        finalChunk = appendStreamChunk({\n          current: finalChunk,\n          next: chunk,\n          provider,\n        });\n      }\n    } else {\n      const metadata = config.metadata as Record<string, unknown> | undefined;\n      for await (const chunk of stream) {\n        const handlingChunk = getStreamHandlingChunk({\n          current: finalChunk,\n          next: chunk,\n          provider,\n        });\n        if (handlingChunk != null && handlingChunk !== chunk) {\n          await registeredStreamHandler.handle(\n            GraphEvents.CHAT_MODEL_STREAM,\n            { chunk: handlingChunk },\n            metadata,\n            context\n          );\n        }\n        finalChunk = appendStreamChunk({\n          current: finalChunk,\n          next: chunk,\n          provider,\n        });\n      }\n    }\n\n    if (manualToolStreamProviders.has(provider)) {\n      finalChunk = modifyDeltaProperties(provider, finalChunk);\n    }\n\n    if ((finalChunk?.tool_calls?.length ?? 0) > 0) {\n      finalChunk!.tool_calls = finalChunk!.tool_calls?.filter(\n        (tool_call: ToolCall) => !!tool_call.name\n      );\n    }\n\n    assertNotTruncatedToolCall(finalChunk, provider);\n    return { messages: [finalChunk as AIMessageChunk] };\n  }\n\n  const finalMessage = await model.invoke(messagesForProvider, config);\n  if ((finalMessage.tool_calls?.length ?? 0) > 0) {\n    finalMessage.tool_calls = finalMessage.tool_calls?.filter(\n      (tool_call: ToolCall) => !!tool_call.name\n    );\n  }\n  assertNotTruncatedToolCall(finalMessage, provider);\n  return { messages: [finalMessage] };\n}\n\n/**\n * Best-effort read of the configured model name from client options.\n * Providers disagree on the key (`model` vs `modelName`).\n */\nfunction extractClientOptionsModel(\n  clientOptions: t.ClientOptions | undefined\n): string | undefined {\n  const options = clientOptions as\n    | { model?: unknown; modelName?: unknown }\n    | undefined;\n  if (typeof options?.model === 'string' && options.model !== '') {\n    return options.model;\n  }\n  if (typeof options?.modelName === 'string' && options.modelName !== '') {\n    return options.modelName;\n  }\n  return undefined;\n}\n\n/**\n * Attempts each fallback provider in order until one succeeds.\n * Throws the last error if all fallbacks fail.\n */\nexport async function tryFallbackProviders({\n  fallbacks,\n  tools,\n  messages,\n  config,\n  primaryError,\n  context,\n  onChunk,\n}: {\n  fallbacks: Array<{ provider: Providers; clientOptions?: t.ClientOptions }>;\n  tools?: t.GraphTools;\n  messages: BaseMessage[];\n  config?: RunnableConfig;\n  primaryError: unknown;\n  context?: InvokeContext;\n  onChunk?: OnChunk;\n}): Promise<Partial<t.BaseGraphState> | undefined> {\n  let lastError: unknown = primaryError;\n  for (const fb of fallbacks) {\n    try {\n      const fbModel = initializeModel({\n        provider: fb.provider,\n        clientOptions: fb.clientOptions,\n        tools,\n      });\n      /**\n       * Stamp the fallback's configured model onto callback metadata so\n       * per-call attribution (subagent usage capture) doesn't fall back to\n       * the PRIMARY config's model when the provider reports no\n       * `ls_model_name`. The serving provider is stamped uniformly by\n       * `attemptInvoke` (`INVOKED_PROVIDER`).\n       */\n      const fbModelName = extractClientOptionsModel(fb.clientOptions);\n      const fbConfig: RunnableConfig | undefined =\n        fbModelName == null\n          ? config\n          : {\n            ...config,\n            metadata: {\n              ...(config?.metadata ?? {}),\n              [Constants.INVOKED_MODEL]: fbModelName,\n            },\n          };\n      const result = await attemptInvoke(\n        {\n          model: fbModel as t.ChatModel,\n          messages,\n          provider: fb.provider,\n          context,\n          onChunk,\n        },\n        fbConfig\n      );\n      return result;\n    } catch (e) {\n      lastError = e;\n      continue;\n    }\n  }\n  if (lastError !== undefined) {\n    throw lastError;\n  }\n  return undefined;\n}\n"],"mappings":";;;;;;;;;;;;AAsDA,SAAS,sCACP,SACoC;CACpC,MAAM,UAAU,SAAS,iBAAiB,WAAA,sBAE1C;CACA,OAAO,mBAAmB,yBAAyB,UAAU,KAAA;AAC/D;AAEA,SAAS,oBAAoB,OAAgC;CAC3D,MAAM,mBAAmB,MAAM,kBAAkB;CACjD,OAAO,MAAM,QAAQ,gBAAgB,KAAK,iBAAiB,SAAS;AACtE;AAEA,SAAS,4CAA4C,EACnD,SACA,MACA,YAKiB;CACjB,MAAM,UAAU,mCAAmC;EACjD;EACA;EACA;CACF,CAAC;CACD,IAAI,WAAW,QAAQ,YAAY,KAAK,SACtC,OAAO;CAGT,OAAO,IAAI,eACT,OAAO,OAAO,CAAC,GAAG,MAAM,EACtB,QACF,CAAC,CACH;AACF;AAEA,SAAS,mCAAmC,EAC1C,SACA,MACA,YAKqB;CACrB,IACE,aAAA,gBACA,WAAW,QACX,CAAC,oBAAoB,IAAI,KACzB,OAAO,QAAQ,YAAY,YAC3B,QAAQ,YAAY,MACpB,OAAO,KAAK,YAAY,YACxB,KAAK,YAAY,IAEjB;CAEF,IAAI,CAAC,KAAK,QAAQ,WAAW,QAAQ,OAAO,GAC1C,OAAO,KAAK;CAEd,OAAO,KAAK,QAAQ,MAAM,QAAQ,QAAQ,MAAM;AAClD;AAEA,SAAS,uBACP,kBACqC;CACrC,OAAO,OAAO,YACZ,OAAO,QAAQ,gBAAgB,CAAC,CAAC,QAC9B,CAAC,SAAS,QAAQ,mBACrB,CACF;AACF;AAEA,SAAS,uBAAuB,EAC9B,SACA,MACA,YAK6B;CAC7B,MAAM,UAAU,mCAAmC;EACjD;EACA;EACA;CACF,CAAC;CACD,IAAI,WAAW,MACb,OAAO;CAET,IAAI,YAAY,IACd;CAEF,OAAO,IAAI,eACT,OAAO,OAAO,CAAC,GAAG,MAAM;EACtB;EACA,mBAAmB,uBAAuB,KAAK,iBAAiB;CAClE,CAAC,CACH;AACF;AAEA,SAAS,kBAAkB,EACzB,SACA,MACA,YAKiB;CACjB,IAAI,WAAW,MACb,OAAO;CAET,OAAO,OACL,SACA,4CAA4C;EAAE;EAAS;EAAM;CAAS,CAAC,CACzE;AACF;;;;;;;;;;AAWA,eAAsB,cACpB,EACE,OACA,UACA,UACA,SACA,WAQF,QACoC;;;;;;;;CAQpC,MAAM,WAAW,SAAS,8BAA8B;CACxD,MAAM,QAAQ,QAAQ,cAAc;CACpC,MAAM,sBAAsB,uBAAuB,UAAU,UAAU,KAAK;;;;;;;;;;CAW5E,SAAS;EACP,GAAG;EACH,UAAU;GACR,GAAI,QAAQ,YAAY,CAAC;2BACK;EAChC;CACF;CAEA,IAAI,MAAM,QAAQ;EAChB,MAAM,SAAS,MAAM,MAAM,OAAO,qBAAqB,MAAM;EAC7D,IAAI;EACJ,MAAM,0BACJ,sCAAsC,OAAO;EAE/C,IAAI,SACF,WAAW,MAAM,SAAS,QAAQ;GAChC,MAAM,QAAQ,KAAK;GACnB,aAAa,kBAAkB;IAC7B,SAAS;IACT,MAAM;IACN;GACF,CAAC;EACH;OACK,IAAI,2BAA2B,MAAM;GAC1C,MAAM,WAAW,OAAO;GACxB,MAAM,gBAAgB,IAAI,uBAAuB;GACjD,WAAW,MAAM,SAAS,QAAQ;IAChC,MAAM,gBAAgB,uBAAuB;KAC3C,SAAS;KACT,MAAM;KACN;IACF,CAAC;IACD,IAAI,iBAAiB,MACnB,MAAM,cAAc,OAAA,wBAElB,EAAE,OAAO,cAAc,GACvB,UACA,OACF;IAEF,aAAa,kBAAkB;KAC7B,SAAS;KACT,MAAM;KACN;IACF,CAAC;GACH;EACF,OAAO;GACL,MAAM,WAAW,OAAO;GACxB,WAAW,MAAM,SAAS,QAAQ;IAChC,MAAM,gBAAgB,uBAAuB;KAC3C,SAAS;KACT,MAAM;KACN;IACF,CAAC;IACD,IAAI,iBAAiB,QAAQ,kBAAkB,OAC7C,MAAM,wBAAwB,OAAA,wBAE5B,EAAE,OAAO,cAAc,GACvB,UACA,OACF;IAEF,aAAa,kBAAkB;KAC7B,SAAS;KACT,MAAM;KACN;IACF,CAAC;GACH;EACF;EAEA,IAAI,0BAA0B,IAAI,QAAQ,GACxC,aAAa,sBAAsB,UAAU,UAAU;EAGzD,KAAK,YAAY,YAAY,UAAU,KAAK,GAC1C,WAAY,aAAa,WAAY,YAAY,QAC9C,cAAwB,CAAC,CAAC,UAAU,IACvC;EAGF,2BAA2B,YAAY,QAAQ;EAC/C,OAAO,EAAE,UAAU,CAAC,UAA4B,EAAE;CACpD;CAEA,MAAM,eAAe,MAAM,MAAM,OAAO,qBAAqB,MAAM;CACnE,KAAK,aAAa,YAAY,UAAU,KAAK,GAC3C,aAAa,aAAa,aAAa,YAAY,QAChD,cAAwB,CAAC,CAAC,UAAU,IACvC;CAEF,2BAA2B,cAAc,QAAQ;CACjD,OAAO,EAAE,UAAU,CAAC,YAAY,EAAE;AACpC;;;;;AAMA,SAAS,0BACP,eACoB;CACpB,MAAM,UAAU;CAGhB,IAAI,OAAO,SAAS,UAAU,YAAY,QAAQ,UAAU,IAC1D,OAAO,QAAQ;CAEjB,IAAI,OAAO,SAAS,cAAc,YAAY,QAAQ,cAAc,IAClE,OAAO,QAAQ;AAGnB;;;;;AAMA,eAAsB,qBAAqB,EACzC,WACA,OACA,UACA,QACA,cACA,SACA,WASiD;CACjD,IAAI,YAAqB;CACzB,KAAK,MAAM,MAAM,WACf,IAAI;EACF,MAAM,UAAU,gBAAgB;GAC9B,UAAU,GAAG;GACb,eAAe,GAAG;GAClB;EACF,CAAC;;;;;;;;EAQD,MAAM,cAAc,0BAA0B,GAAG,aAAa;EAC9D,MAAM,WACJ,eAAe,OACX,SACA;GACA,GAAG;GACH,UAAU;IACR,GAAI,QAAQ,YAAY,CAAC;yBACE;GAC7B;EACF;EAWJ,OAAO,MAVc,cACnB;GACE,OAAO;GACP;GACA,UAAU,GAAG;GACb;GACA;EACF,GACA,QACF;CAEF,SAAS,GAAG;EACV,YAAY;EACZ;CACF;CAEF,IAAI,cAAc,KAAA,GAChB,MAAM;AAGV"}
+{"version":3,"file":"invoke.mjs","names":[],"sources":["../../../src/llm/invoke.ts"],"sourcesContent":["import { concat } from '@langchain/core/utils/stream';\nimport { AIMessageChunk } from '@langchain/core/messages';\nimport type { RunnableConfig } from '@langchain/core/runnables';\nimport type { ToolCall } from '@langchain/core/messages/tool';\nimport type { BaseMessage } from '@langchain/core/messages';\nimport type { ToolOutputReferenceRegistry } from '@/tools/toolOutputReferences';\nimport type { ContextOverflowContext } from '@/utils/errors';\nimport type * as t from '@/types';\nimport { annotateMessagesForLLM } from '@/tools/toolOutputReferences';\nimport { assertNotTruncatedToolCall } from '@/llm/truncation';\nimport { Constants, GraphEvents, Providers } from '@/common';\nimport { manualToolStreamProviders } from '@/llm/providers';\nimport { getContextOverflowInfo } from '@/utils/errors';\nimport { modifyDeltaProperties } from '@/messages';\nimport { ChatModelStreamHandler } from '@/stream';\nimport { initializeModel } from '@/llm/init';\n\n/**\n * Context passed to `attemptInvoke`. Matches the subset of Graph that\n * `ChatModelStreamHandler.handle` needs *plus* the explicit\n * `getOrCreateToolOutputRegistry()` accessor that `attemptInvoke`\n * itself calls to pull the run-scoped tool-output registry off the\n * graph and project each relevant ToolMessage into a transient\n * annotated copy before the provider call.\n *\n * The intersection is intentional: `Parameters<...>[3]` resolves\n * indirectly through the stream handler's signature (which returns\n * `StandardGraph` and already exposes the accessor since #117), but\n * stating it explicitly here surfaces the contract at the call site —\n * a developer reading `attemptInvoke` doesn't have to chase the\n * upstream handler's parameter list to discover that\n * `context?.getOrCreateToolOutputRegistry()` is a real thing. Single\n * optional chain only — the method itself is required on the\n * `StandardGraph` branch of the intersection, so the second `?.` is\n * unnecessary at the call site.\n *\n * `NonNullable<...>` strips `undefined` from the upstream parameter\n * type so the intersection doesn't collapse to `never` on the\n * undefined branch; callers express optionality via `context?:\n * InvokeContext` on the function signature instead.\n *\n * Callers without a registry (e.g. summarization) simply pass no\n * `context` and the transform safely no-ops.\n */\nexport type InvokeContext = NonNullable<\n  Parameters<ChatModelStreamHandler['handle']>[3]\n> & {\n  getOrCreateToolOutputRegistry?(): ToolOutputReferenceRegistry | undefined;\n};\n\n/**\n * Per-chunk callback for custom stream processing.\n * When provided, replaces the default `ChatModelStreamHandler`.\n */\nexport type OnChunk = (chunk: AIMessageChunk) => void | Promise<void>;\n\nfunction getRegisteredDefaultChatStreamHandler(\n  context?: InvokeContext\n): ChatModelStreamHandler | undefined {\n  const handler = context?.handlerRegistry?.getHandler(\n    GraphEvents.CHAT_MODEL_STREAM\n  );\n  return handler instanceof ChatModelStreamHandler ? handler : undefined;\n}\n\nfunction hasReasoningDetails(chunk: AIMessageChunk): boolean {\n  const reasoningDetails = chunk.additional_kwargs.reasoning_details;\n  return Array.isArray(reasoningDetails) && reasoningDetails.length > 0;\n}\n\nfunction removeOpenRouterFinalReasoningReplayContent({\n  current,\n  next,\n  provider,\n}: {\n  current?: AIMessageChunk;\n  next: AIMessageChunk;\n  provider: Providers;\n}): AIMessageChunk {\n  const content = getOpenRouterFinalReasoningContent({\n    current,\n    next,\n    provider,\n  });\n  if (content == null || content === next.content) {\n    return next;\n  }\n\n  return new AIMessageChunk(\n    Object.assign({}, next, {\n      content,\n    })\n  );\n}\n\nfunction getOpenRouterFinalReasoningContent({\n  current,\n  next,\n  provider,\n}: {\n  current?: AIMessageChunk;\n  next: AIMessageChunk;\n  provider: Providers;\n}): string | undefined {\n  if (\n    provider !== Providers.OPENROUTER ||\n    current == null ||\n    !hasReasoningDetails(next) ||\n    typeof current.content !== 'string' ||\n    current.content === '' ||\n    typeof next.content !== 'string' ||\n    next.content === ''\n  ) {\n    return undefined;\n  }\n  if (!next.content.startsWith(current.content)) {\n    return next.content;\n  }\n  return next.content.slice(current.content.length);\n}\n\nfunction removeReasoningDetails(\n  additionalKwargs: AIMessageChunk['additional_kwargs']\n): AIMessageChunk['additional_kwargs'] {\n  return Object.fromEntries(\n    Object.entries(additionalKwargs).filter(\n      ([key]) => key !== 'reasoning_details'\n    )\n  );\n}\n\nfunction getStreamHandlingChunk({\n  current,\n  next,\n  provider,\n}: {\n  current?: AIMessageChunk;\n  next: AIMessageChunk;\n  provider: Providers;\n}): AIMessageChunk | undefined {\n  const content = getOpenRouterFinalReasoningContent({\n    current,\n    next,\n    provider,\n  });\n  if (content == null) {\n    return next;\n  }\n  if (content === '') {\n    return undefined;\n  }\n  return new AIMessageChunk(\n    Object.assign({}, next, {\n      content,\n      additional_kwargs: removeReasoningDetails(next.additional_kwargs),\n    })\n  );\n}\n\nfunction appendStreamChunk({\n  current,\n  next,\n  provider,\n}: {\n  current?: AIMessageChunk;\n  next: AIMessageChunk;\n  provider: Providers;\n}): AIMessageChunk {\n  if (current == null) {\n    return next;\n  }\n  return concat(\n    current,\n    removeOpenRouterFinalReasoningReplayContent({ current, next, provider })\n  );\n}\n\n/**\n * Invokes a chat model with the given messages, handling both streaming and\n * non-streaming paths.\n *\n * By default, stream chunks are processed through a `ChatModelStreamHandler`\n * that dispatches run steps (MESSAGE_CREATION, TOOL_CALLS) for the graph.\n * Pass an `onChunk` callback to override this with custom chunk processing\n * (e.g. summarization delta events).\n */\nexport async function attemptInvoke(\n  {\n    model,\n    messages,\n    provider,\n    context,\n    onChunk,\n  }: {\n    model: t.ChatModel;\n    messages: BaseMessage[];\n    provider: Providers;\n    context?: InvokeContext;\n    onChunk?: OnChunk;\n  },\n  config?: RunnableConfig\n): Promise<Partial<t.BaseGraphState>> {\n  /**\n   * Pull the run-scoped tool output registry off the graph (when one\n   * exists) and project ToolMessages carrying ref metadata into a\n   * transient annotated copy. The original `messages` array stays\n   * untouched so the graph state never sees `[ref: …]` / `_ref`\n   * payload.\n   */\n  const registry = context?.getOrCreateToolOutputRegistry();\n  const runId = config?.configurable?.run_id as string | undefined;\n  const messagesForProvider = annotateMessagesForLLM(messages, registry, runId);\n\n  /**\n   * Stamp the provider that is ACTUALLY serving this invocation onto the\n   * callback metadata. `attemptInvoke` is the single funnel for primary,\n   * fallback, and summarization model calls, so consumers that need\n   * provider attribution per call (the subagent usage-capture handler)\n   * read this key instead of trusting static agent config — which is\n   * wrong for fallback-served calls — or `ls_provider` — which derived\n   * providers inherit from their base class.\n   */\n  config = {\n    ...config,\n    metadata: {\n      ...(config?.metadata ?? {}),\n      [Constants.INVOKED_PROVIDER]: provider,\n    },\n  };\n\n  if (model.stream) {\n    const stream = await model.stream(messagesForProvider, config);\n    let finalChunk: AIMessageChunk | undefined;\n    const registeredStreamHandler =\n      getRegisteredDefaultChatStreamHandler(context);\n\n    if (onChunk) {\n      for await (const chunk of stream) {\n        await onChunk(chunk);\n        finalChunk = appendStreamChunk({\n          current: finalChunk,\n          next: chunk,\n          provider,\n        });\n      }\n    } else if (registeredStreamHandler == null) {\n      const metadata = config.metadata as Record<string, unknown> | undefined;\n      const streamHandler = new ChatModelStreamHandler();\n      for await (const chunk of stream) {\n        const handlingChunk = getStreamHandlingChunk({\n          current: finalChunk,\n          next: chunk,\n          provider,\n        });\n        if (handlingChunk != null) {\n          await streamHandler.handle(\n            GraphEvents.CHAT_MODEL_STREAM,\n            { chunk: handlingChunk },\n            metadata,\n            context\n          );\n        }\n        finalChunk = appendStreamChunk({\n          current: finalChunk,\n          next: chunk,\n          provider,\n        });\n      }\n    } else {\n      const metadata = config.metadata as Record<string, unknown> | undefined;\n      for await (const chunk of stream) {\n        const handlingChunk = getStreamHandlingChunk({\n          current: finalChunk,\n          next: chunk,\n          provider,\n        });\n        if (handlingChunk != null && handlingChunk !== chunk) {\n          await registeredStreamHandler.handle(\n            GraphEvents.CHAT_MODEL_STREAM,\n            { chunk: handlingChunk },\n            metadata,\n            context\n          );\n        }\n        finalChunk = appendStreamChunk({\n          current: finalChunk,\n          next: chunk,\n          provider,\n        });\n      }\n    }\n\n    if (manualToolStreamProviders.has(provider)) {\n      finalChunk = modifyDeltaProperties(provider, finalChunk);\n    }\n\n    if ((finalChunk?.tool_calls?.length ?? 0) > 0) {\n      finalChunk!.tool_calls = finalChunk!.tool_calls?.filter(\n        (tool_call: ToolCall) => !!tool_call.name\n      );\n    }\n\n    assertNotTruncatedToolCall(finalChunk, provider);\n    return { messages: [finalChunk as AIMessageChunk] };\n  }\n\n  const finalMessage = await model.invoke(messagesForProvider, config);\n  if ((finalMessage.tool_calls?.length ?? 0) > 0) {\n    finalMessage.tool_calls = finalMessage.tool_calls?.filter(\n      (tool_call: ToolCall) => !!tool_call.name\n    );\n  }\n  assertNotTruncatedToolCall(finalMessage, provider);\n  return { messages: [finalMessage] };\n}\n\n/**\n * Identifies which fallback produced an error, so a caller planning a\n * recovery can reason about the client that actually failed rather than the\n * primary's configuration — their context windows and output allowances\n * differ, which is the whole reason a fallback exists.\n */\nexport interface FallbackErrorContext {\n  provider: Providers;\n  clientOptions?: t.ClientOptions;\n  maxContextTokens?: number;\n}\n\nexport interface FallbackOverflowCandidate {\n  error: unknown;\n  context: FallbackErrorContext;\n}\n\nconst fallbackErrorContexts = new WeakMap<object, FallbackErrorContext>();\nconst fallbackOverflowCandidates = new WeakMap<\n  object,\n  FallbackOverflowCandidate[]\n>();\n\nfunction attachFallbackErrorContext(\n  error: unknown,\n  fallbackContext: FallbackErrorContext\n): void {\n  if (typeof error !== 'object' || error === null) {\n    return;\n  }\n  fallbackErrorContexts.set(error, fallbackContext);\n}\n\n/** Reads back the fallback attribution attached by `tryFallbackProviders`. */\nexport function getFallbackErrorContext(\n  error: unknown\n): FallbackErrorContext | undefined {\n  if (typeof error !== 'object' || error === null) {\n    return undefined;\n  }\n  return fallbackErrorContexts.get(error);\n}\n\n/** Returns every fallback overflow retained from an exhausted provider chain. */\nexport function getFallbackOverflowCandidates(\n  error: unknown\n): FallbackOverflowCandidate[] {\n  if (typeof error !== 'object' || error === null) {\n    return [];\n  }\n  return [...(fallbackOverflowCandidates.get(error) ?? [])];\n}\n\n/**\n * Best-effort read of the configured model name from client options.\n * Providers disagree on the key (`model` vs `modelName`).\n */\nfunction extractClientOptionsModel(\n  clientOptions: t.ClientOptions | undefined\n): string | undefined {\n  const options = clientOptions as\n    | { model?: unknown; modelName?: unknown }\n    | undefined;\n  if (typeof options?.model === 'string' && options.model !== '') {\n    return options.model;\n  }\n  if (typeof options?.modelName === 'string' && options.modelName !== '') {\n    return options.modelName;\n  }\n  return undefined;\n}\n\n/**\n * Attempts each fallback provider in order until one succeeds.\n *\n * When every fallback fails, a context overflow among them is thrown in\n * preference to whichever failure happened to come last. An overflow is the\n * one failure the caller can act on — it compacts and retries — and losing it\n * behind a later unrelated error would surface a dead end instead. Ordinary\n * failures still throw last-error-wins.\n */\nexport async function tryFallbackProviders({\n  fallbacks,\n  tools,\n  messages,\n  config,\n  primaryError,\n  context,\n  onChunk,\n  overflowContext,\n}: {\n  fallbacks: t.FallbackConfig[];\n  tools?: t.GraphTools;\n  messages: BaseMessage[];\n  config?: RunnableConfig;\n  primaryError: unknown;\n  context?: InvokeContext;\n  onChunk?: OnChunk;\n  /**\n   * Prompt-size corroboration for signatures that are not self-describing.\n   * Vertex AI's overflow is a bare `400` with no reason, so without this a\n   * fallback that overflows is indistinguishable from any other 400 and would\n   * be dropped in favour of whichever failure came last.\n   */\n  overflowContext?: ContextOverflowContext;\n}): Promise<Partial<t.BaseGraphState> | undefined> {\n  const isOverflow = (\n    error: unknown,\n    contextOverride = overflowContext\n  ): boolean => getContextOverflowInfo(error, contextOverride) != null;\n  let lastError: unknown = primaryError;\n  /**\n   * Tracked apart from the primary's overflow. A caller reaching this\n   * function with an overflowing primary has already failed to recover from\n   * it, so a fallback overflow — which may sit against a different window and\n   * output allowance — is the more useful of the two to surface.\n   */\n  const overflowCandidates: FallbackOverflowCandidate[] = [];\n  const primaryOverflowError: unknown = isOverflow(primaryError)\n    ? primaryError\n    : undefined;\n  for (const fb of fallbacks) {\n    try {\n      const fbModel = initializeModel({\n        provider: fb.provider,\n        clientOptions: fb.clientOptions,\n        tools,\n      });\n      /**\n       * Stamp the fallback's configured model onto callback metadata so\n       * per-call attribution (subagent usage capture) doesn't fall back to\n       * the PRIMARY config's model when the provider reports no\n       * `ls_model_name`. The serving provider is stamped uniformly by\n       * `attemptInvoke` (`INVOKED_PROVIDER`).\n       */\n      const fbModelName = extractClientOptionsModel(fb.clientOptions);\n      const fbConfig: RunnableConfig | undefined =\n        fbModelName == null\n          ? config\n          : {\n            ...config,\n            metadata: {\n              ...(config?.metadata ?? {}),\n              [Constants.INVOKED_MODEL]: fbModelName,\n            },\n          };\n      const result = await attemptInvoke(\n        {\n          model: fbModel as t.ChatModel,\n          messages,\n          provider: fb.provider,\n          context,\n          onChunk,\n        },\n        fbConfig\n      );\n      return result;\n    } catch (e) {\n      lastError = e;\n      const fallbackOverflowContext: ContextOverflowContext = {\n        provider: fb.provider,\n        maxContextTokens: fb.maxContextTokens,\n        ...(overflowContext?.provider === fb.provider\n          ? {\n            estimatedPromptTokens: overflowContext.estimatedPromptTokens,\n          }\n          : {}),\n      };\n      if (isOverflow(e, fallbackOverflowContext)) {\n        const errorContext: FallbackErrorContext = {\n          provider: fb.provider,\n          clientOptions: fb.clientOptions,\n          maxContextTokens: fb.maxContextTokens,\n        };\n        attachFallbackErrorContext(e, errorContext);\n        overflowCandidates.push({ error: e, context: errorContext });\n      }\n      continue;\n    }\n  }\n  /**\n   * Preference order: a fallback overflow, then the primary's overflow, then\n   * whichever failure came last. An overflow is the only one of the three a\n   * caller can act on, and the fallback's carries the client attribution that\n   * makes a correct retry budget possible.\n   */\n  const preferred =\n    overflowCandidates[0]?.error ?? primaryOverflowError ?? lastError;\n  if (\n    overflowCandidates.length > 0 &&\n    typeof preferred === 'object' &&\n    preferred !== null\n  ) {\n    fallbackOverflowCandidates.set(preferred, overflowCandidates);\n  }\n  if (preferred !== undefined) {\n    throw preferred;\n  }\n  return undefined;\n}\n"],"mappings":";;;;;;;;;;;;;AAwDA,SAAS,sCACP,SACoC;CACpC,MAAM,UAAU,SAAS,iBAAiB,WAAA,sBAE1C;CACA,OAAO,mBAAmB,yBAAyB,UAAU,KAAA;AAC/D;AAEA,SAAS,oBAAoB,OAAgC;CAC3D,MAAM,mBAAmB,MAAM,kBAAkB;CACjD,OAAO,MAAM,QAAQ,gBAAgB,KAAK,iBAAiB,SAAS;AACtE;AAEA,SAAS,4CAA4C,EACnD,SACA,MACA,YAKiB;CACjB,MAAM,UAAU,mCAAmC;EACjD;EACA;EACA;CACF,CAAC;CACD,IAAI,WAAW,QAAQ,YAAY,KAAK,SACtC,OAAO;CAGT,OAAO,IAAI,eACT,OAAO,OAAO,CAAC,GAAG,MAAM,EACtB,QACF,CAAC,CACH;AACF;AAEA,SAAS,mCAAmC,EAC1C,SACA,MACA,YAKqB;CACrB,IACE,aAAA,gBACA,WAAW,QACX,CAAC,oBAAoB,IAAI,KACzB,OAAO,QAAQ,YAAY,YAC3B,QAAQ,YAAY,MACpB,OAAO,KAAK,YAAY,YACxB,KAAK,YAAY,IAEjB;CAEF,IAAI,CAAC,KAAK,QAAQ,WAAW,QAAQ,OAAO,GAC1C,OAAO,KAAK;CAEd,OAAO,KAAK,QAAQ,MAAM,QAAQ,QAAQ,MAAM;AAClD;AAEA,SAAS,uBACP,kBACqC;CACrC,OAAO,OAAO,YACZ,OAAO,QAAQ,gBAAgB,CAAC,CAAC,QAC9B,CAAC,SAAS,QAAQ,mBACrB,CACF;AACF;AAEA,SAAS,uBAAuB,EAC9B,SACA,MACA,YAK6B;CAC7B,MAAM,UAAU,mCAAmC;EACjD;EACA;EACA;CACF,CAAC;CACD,IAAI,WAAW,MACb,OAAO;CAET,IAAI,YAAY,IACd;CAEF,OAAO,IAAI,eACT,OAAO,OAAO,CAAC,GAAG,MAAM;EACtB;EACA,mBAAmB,uBAAuB,KAAK,iBAAiB;CAClE,CAAC,CACH;AACF;AAEA,SAAS,kBAAkB,EACzB,SACA,MACA,YAKiB;CACjB,IAAI,WAAW,MACb,OAAO;CAET,OAAO,OACL,SACA,4CAA4C;EAAE;EAAS;EAAM;CAAS,CAAC,CACzE;AACF;;;;;;;;;;AAWA,eAAsB,cACpB,EACE,OACA,UACA,UACA,SACA,WAQF,QACoC;;;;;;;;CAQpC,MAAM,WAAW,SAAS,8BAA8B;CACxD,MAAM,QAAQ,QAAQ,cAAc;CACpC,MAAM,sBAAsB,uBAAuB,UAAU,UAAU,KAAK;;;;;;;;;;CAW5E,SAAS;EACP,GAAG;EACH,UAAU;GACR,GAAI,QAAQ,YAAY,CAAC;2BACK;EAChC;CACF;CAEA,IAAI,MAAM,QAAQ;EAChB,MAAM,SAAS,MAAM,MAAM,OAAO,qBAAqB,MAAM;EAC7D,IAAI;EACJ,MAAM,0BACJ,sCAAsC,OAAO;EAE/C,IAAI,SACF,WAAW,MAAM,SAAS,QAAQ;GAChC,MAAM,QAAQ,KAAK;GACnB,aAAa,kBAAkB;IAC7B,SAAS;IACT,MAAM;IACN;GACF,CAAC;EACH;OACK,IAAI,2BAA2B,MAAM;GAC1C,MAAM,WAAW,OAAO;GACxB,MAAM,gBAAgB,IAAI,uBAAuB;GACjD,WAAW,MAAM,SAAS,QAAQ;IAChC,MAAM,gBAAgB,uBAAuB;KAC3C,SAAS;KACT,MAAM;KACN;IACF,CAAC;IACD,IAAI,iBAAiB,MACnB,MAAM,cAAc,OAAA,wBAElB,EAAE,OAAO,cAAc,GACvB,UACA,OACF;IAEF,aAAa,kBAAkB;KAC7B,SAAS;KACT,MAAM;KACN;IACF,CAAC;GACH;EACF,OAAO;GACL,MAAM,WAAW,OAAO;GACxB,WAAW,MAAM,SAAS,QAAQ;IAChC,MAAM,gBAAgB,uBAAuB;KAC3C,SAAS;KACT,MAAM;KACN;IACF,CAAC;IACD,IAAI,iBAAiB,QAAQ,kBAAkB,OAC7C,MAAM,wBAAwB,OAAA,wBAE5B,EAAE,OAAO,cAAc,GACvB,UACA,OACF;IAEF,aAAa,kBAAkB;KAC7B,SAAS;KACT,MAAM;KACN;IACF,CAAC;GACH;EACF;EAEA,IAAI,0BAA0B,IAAI,QAAQ,GACxC,aAAa,sBAAsB,UAAU,UAAU;EAGzD,KAAK,YAAY,YAAY,UAAU,KAAK,GAC1C,WAAY,aAAa,WAAY,YAAY,QAC9C,cAAwB,CAAC,CAAC,UAAU,IACvC;EAGF,2BAA2B,YAAY,QAAQ;EAC/C,OAAO,EAAE,UAAU,CAAC,UAA4B,EAAE;CACpD;CAEA,MAAM,eAAe,MAAM,MAAM,OAAO,qBAAqB,MAAM;CACnE,KAAK,aAAa,YAAY,UAAU,KAAK,GAC3C,aAAa,aAAa,aAAa,YAAY,QAChD,cAAwB,CAAC,CAAC,UAAU,IACvC;CAEF,2BAA2B,cAAc,QAAQ;CACjD,OAAO,EAAE,UAAU,CAAC,YAAY,EAAE;AACpC;AAmBA,MAAM,wCAAwB,IAAI,QAAsC;AACxE,MAAM,6CAA6B,IAAI,QAGrC;AAEF,SAAS,2BACP,OACA,iBACM;CACN,IAAI,OAAO,UAAU,YAAY,UAAU,MACzC;CAEF,sBAAsB,IAAI,OAAO,eAAe;AAClD;;AAGA,SAAgB,wBACd,OACkC;CAClC,IAAI,OAAO,UAAU,YAAY,UAAU,MACzC;CAEF,OAAO,sBAAsB,IAAI,KAAK;AACxC;;AAGA,SAAgB,8BACd,OAC6B;CAC7B,IAAI,OAAO,UAAU,YAAY,UAAU,MACzC,OAAO,CAAC;CAEV,OAAO,CAAC,GAAI,2BAA2B,IAAI,KAAK,KAAK,CAAC,CAAE;AAC1D;;;;;AAMA,SAAS,0BACP,eACoB;CACpB,MAAM,UAAU;CAGhB,IAAI,OAAO,SAAS,UAAU,YAAY,QAAQ,UAAU,IAC1D,OAAO,QAAQ;CAEjB,IAAI,OAAO,SAAS,cAAc,YAAY,QAAQ,cAAc,IAClE,OAAO,QAAQ;AAGnB;;;;;;;;;;AAWA,eAAsB,qBAAqB,EACzC,WACA,OACA,UACA,QACA,cACA,SACA,SACA,mBAgBiD;CACjD,MAAM,cACJ,OACA,kBAAkB,oBACN,uBAAuB,OAAO,eAAe,KAAK;CAChE,IAAI,YAAqB;;;;;;;CAOzB,MAAM,qBAAkD,CAAC;CACzD,MAAM,uBAAgC,WAAW,YAAY,IACzD,eACA,KAAA;CACJ,KAAK,MAAM,MAAM,WACf,IAAI;EACF,MAAM,UAAU,gBAAgB;GAC9B,UAAU,GAAG;GACb,eAAe,GAAG;GAClB;EACF,CAAC;;;;;;;;EAQD,MAAM,cAAc,0BAA0B,GAAG,aAAa;EAC9D,MAAM,WACJ,eAAe,OACX,SACA;GACA,GAAG;GACH,UAAU;IACR,GAAI,QAAQ,YAAY,CAAC;yBACE;GAC7B;EACF;EAWJ,OAAO,MAVc,cACnB;GACE,OAAO;GACP;GACA,UAAU,GAAG;GACb;GACA;EACF,GACA,QACF;CAEF,SAAS,GAAG;EACV,YAAY;EAUZ,IAAI,WAAW,GAAG;GARhB,UAAU,GAAG;GACb,kBAAkB,GAAG;GACrB,GAAI,iBAAiB,aAAa,GAAG,WACjC,EACA,uBAAuB,gBAAgB,sBACzC,IACE,CAAC;EAEiC,CAAC,GAAG;GAC1C,MAAM,eAAqC;IACzC,UAAU,GAAG;IACb,eAAe,GAAG;IAClB,kBAAkB,GAAG;GACvB;GACA,2BAA2B,GAAG,YAAY;GAC1C,mBAAmB,KAAK;IAAE,OAAO;IAAG,SAAS;GAAa,CAAC;EAC7D;EACA;CACF;;;;;;;CAQF,MAAM,YACJ,mBAAmB,EAAE,EAAE,SAAS,wBAAwB;CAC1D,IACE,mBAAmB,SAAS,KAC5B,OAAO,cAAc,YACrB,cAAc,MAEd,2BAA2B,IAAI,WAAW,kBAAkB;CAE9D,IAAI,cAAc,KAAA,GAChB,MAAM;AAGV"}