@librechat/agents 3.3.10 → 3.3.12

package/dist/cjs/llm/invoke.cjs.map

@@ -1 +1 @@
-{"version":3,"file":"invoke.cjs","names":["isOpenAILike","projectToolStreamContentForProvider","projectOpenAIResponsesToolMessageContent","stripAnthropicCacheControl","stripBedrockCacheControl","projectComputerCallOutputsToText","projectOpenRouterToolMessageContent","projectOpenAIChatToolMessageContent","projectSingleTextToolOutputsToText","projectCacheControlledToolOutputsToText","projectStructuredToolOutputsToText","dispatchesChatModelStream","AIMessageChunk","CallbackManager","CallbackManagerForLLMRun","safeDispatchCustomEvent","annotateMessagesForLLM","isAnthropicLike","appendPredecessorHandoffCue","removePredecessorHandoffCue","strictAlternationProviders","coalesceAdjacentUserTurns","appendCallbacks","ChatModelStreamHandler","canSealPreempt","manualToolStreamProviders","modifyDeltaProperties","getContextOverflowInfo","initializeModel"],"sources":["../../../src/llm/invoke.ts"],"sourcesContent":["import { concat } from '@langchain/core/utils/stream';\nimport { AIMessageChunk } from '@langchain/core/messages';\nimport {\n  CallbackManager,\n  CallbackManagerForLLMRun,\n  type Callbacks,\n} from '@langchain/core/callbacks/manager';\nimport { getCallbackManagerForConfig } from '@langchain/core/runnables';\nimport type { Serialized } from '@langchain/core/load/serializable';\nimport type { ChatGeneration } from '@langchain/core/outputs';\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 {\n  projectCacheControlledToolOutputsToText,\n  projectComputerCallOutputsToText,\n  projectOpenAIChatToolMessageContent,\n  projectOpenAIResponsesToolMessageContent,\n  projectOpenRouterToolMessageContent,\n  projectSingleTextToolOutputsToText,\n  projectStructuredToolOutputsToText,\n  projectToolStreamContentForProvider,\n} from '@/messages/core';\nimport {\n  stripAnthropicCacheControl,\n  stripBedrockCacheControl,\n} from '@/messages/cache';\nimport { annotateMessagesForLLM } from '@/tools/toolOutputReferences';\nimport { assertNotTruncatedToolCall } from '@/llm/truncation';\nimport { Constants, ContentTypes, GraphEvents, Providers } from '@/common';\nimport { manualToolStreamProviders } from '@/llm/providers';\nimport { appendCallbacks } from '@/utils/callbacks';\nimport { safeDispatchCustomEvent } from '@/utils/events';\nimport { getContextOverflowInfo } from '@/utils/errors';\nimport {\n  modifyDeltaProperties,\n  coalesceAdjacentUserTurns,\n  strictAlternationProviders,\n  appendPredecessorHandoffCue,\n  removePredecessorHandoffCue,\n} from '@/messages';\nimport { canSealPreempt } from '@/llm/preempt';\nimport { ChatModelStreamHandler, dispatchesChatModelStream } from '@/stream';\nimport { initializeModel } from '@/llm/init';\nimport { isAnthropicLike, isOpenAILike } from '@/utils/llm';\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\nexport function usesNativeOpenAIResponses(\n  model: t.ChatModel,\n  provider: Providers,\n  callOptions?: unknown\n): boolean {\n  if (!isOpenAILike(provider)) {\n    return false;\n  }\n  let candidate: unknown = model;\n  let effectiveCallOptions = callOptions;\n  const seen = new Set<object>();\n  for (let depth = 0; depth < 20; depth++) {\n    if (candidate == null || typeof candidate !== 'object') {\n      return false;\n    }\n    if (seen.has(candidate)) {\n      return false;\n    }\n    seen.add(candidate);\n    const runnable = candidate as {\n      _useResponsesApi?: (options?: unknown) => boolean;\n      bound?: unknown;\n      defaultOptions?: unknown;\n      last?: unknown;\n      constructor?: { name?: unknown };\n    };\n    try {\n      if (\n        runnable.defaultOptions != null &&\n        typeof runnable.defaultOptions === 'object' &&\n        !Array.isArray(runnable.defaultOptions) &&\n        effectiveCallOptions != null &&\n        typeof effectiveCallOptions === 'object' &&\n        !Array.isArray(effectiveCallOptions)\n      ) {\n        effectiveCallOptions = {\n          ...(runnable.defaultOptions as Record<string, unknown>),\n          ...(effectiveCallOptions as Record<string, unknown>),\n        };\n      } else if (effectiveCallOptions == null) {\n        effectiveCallOptions = runnable.defaultOptions;\n      }\n      if (\n        runnable._useResponsesApi?.(effectiveCallOptions) === true ||\n        runnable._useResponsesApi?.(undefined) === true\n      ) {\n        return true;\n      }\n    } catch {\n      // Continue through RunnableSequence/RunnableBinding wrappers.\n    }\n    if (\n      typeof runnable.constructor?.name === 'string' &&\n      runnable.constructor.name.includes('Responses')\n    ) {\n      return true;\n    }\n    if (runnable.last != null && typeof runnable.last === 'object') {\n      candidate = runnable.last;\n      continue;\n    }\n    if (runnable.bound != null && typeof runnable.bound === 'object') {\n      candidate = runnable.bound;\n      continue;\n    }\n    return false;\n  }\n  return false;\n}\n\n/**\n * Produces the exact provider-facing message representation before a model\n * adapter serializes it. This is shared by invocation and Graph's final budget\n * guard so structured tool output cannot grow after the payload was measured.\n */\nexport function projectMessagesForProvider({\n  model,\n  messages,\n  provider,\n  maxToolResultChars,\n  callOptions,\n}: {\n  model: t.ChatModel;\n  messages: BaseMessage[];\n  provider: Providers;\n  maxToolResultChars?: number;\n  callOptions?: unknown;\n}): BaseMessage[] {\n  const providerInputMessages = projectToolStreamContentForProvider(messages);\n  if (usesNativeOpenAIResponses(model, provider, callOptions)) {\n    return projectOpenAIResponsesToolMessageContent(\n      stripAnthropicCacheControl(\n        stripBedrockCacheControl(providerInputMessages)\n      ),\n      maxToolResultChars\n    );\n  }\n  if (provider === Providers.OPENROUTER) {\n    return projectComputerCallOutputsToText(\n      projectOpenRouterToolMessageContent(\n        stripBedrockCacheControl(providerInputMessages),\n        maxToolResultChars\n      )\n    );\n  }\n  if (isOpenAILike(provider)) {\n    return projectComputerCallOutputsToText(\n      projectOpenAIChatToolMessageContent(\n        stripAnthropicCacheControl(\n          stripBedrockCacheControl(providerInputMessages)\n        ),\n        maxToolResultChars\n      )\n    );\n  }\n  if (provider === Providers.ANTHROPIC) {\n    return projectComputerCallOutputsToText(\n      projectSingleTextToolOutputsToText(\n        stripBedrockCacheControl(providerInputMessages),\n        maxToolResultChars\n      )\n    );\n  }\n  if (provider === Providers.BEDROCK) {\n    return stripAnthropicCacheControl(\n      projectComputerCallOutputsToText(\n        projectCacheControlledToolOutputsToText(\n          providerInputMessages,\n          maxToolResultChars\n        )\n      )\n    );\n  }\n  return projectComputerCallOutputsToText(\n    projectStructuredToolOutputsToText(\n      projectSingleTextToolOutputsToText(\n        stripAnthropicCacheControl(\n          stripBedrockCacheControl(providerInputMessages)\n        ),\n        maxToolResultChars\n      ),\n      maxToolResultChars\n    )\n  );\n}\n\n/**\n * The registered handler that owns content-part dispatch, if any.\n *\n * Detected by brand rather than by `instanceof`: a host that registers\n * `new ChatModelStreamHandler()` to opt out of sealing gets wrapped by\n * `createRunHandlers` on every `AgentSession` run, and by\n * `composeEventHandlers` on a key collision. Both wrappers forward to the same\n * dispatcher while failing an identity check, so an identity test would\n * silently revoke the opt-out documented on `StreamPreemption`.\n */\nfunction getRegisteredDefaultChatStreamHandler(\n  context?: InvokeContext\n): t.EventHandler | undefined {\n  const handler = context?.handlerRegistry?.getHandler(\n    GraphEvents.CHAT_MODEL_STREAM\n  );\n  return dispatchesChatModelStream(handler) ? 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\n/**\n * Best-effort output-token count for a sealed turn, used only when the\n * provider never got to send its usage chunk.\n */\nfunction countSealedTokens(\n  context: InvokeContext | undefined,\n  metadata: Record<string, unknown> | undefined,\n  messages: BaseMessage[]\n): number | undefined {\n  try {\n    const counter = context?.getAgentContext(metadata).tokenCounter;\n    if (counter == null) {\n      return undefined;\n    }\n    let total = 0;\n    for (const message of messages) {\n      total += counter(message);\n    }\n    return total;\n  } catch {\n    return undefined;\n  }\n}\n\n/**\n * Instruction overhead the provider processed but that never appears in the\n * message array: `createCallModel` pipes the model through\n * `agentContext.systemRunnable` and binds tool schemas AFTER `messages` is\n * formed, so the system prompt, dynamic instructions, summary and tool\n * schemas are all billed yet invisible here.\n *\n * Read per-node via `getAgentContext(metadata)` rather than the graph-level\n * accessor, which is hardcoded to `defaultAgentId` and would report the wrong\n * agent's overhead in a `MultiAgentGraph`.\n */\nfunction sealedInstructionOverhead(\n  context: InvokeContext | undefined,\n  metadata: Record<string, unknown> | undefined\n): number {\n  try {\n    const agentContext = context?.getAgentContext(metadata);\n    return (\n      agentContext?.resolvedInstructionOverhead ??\n      agentContext?.instructionTokens ??\n      0\n    );\n  } catch {\n    return 0;\n  }\n}\n\n/**\n * Best-effort usage for a turn the provider never got to bill us for.\n *\n * The prompt matters as much as the completion: the provider processed the\n * ENTIRE prompt — messages plus instruction overhead — before we sealed, and\n * every resume re-sends it, so under-counting input hides the expensive half\n * of a preempted run.\n *\n * ESTIMATE, NOT MEASUREMENT. Messages are counted with the host's tokenizer\n * rather than the provider's, and `toolSchemaTokens` applies a heuristic\n * multiplier. It is also an over-count on the fallback path, where\n * `tryFallbackProviders` builds a bare model with no `systemRunnable` pipe so\n * the system prompt genuinely is not sent. Accepted rather than threaded\n * through a flag: only the fallback-plus-seal combination is affected, and an\n * over-count is safer than the previous fabricated `input_tokens: 0`.\n *\n * Marked `estimated_usage` so calibration can refuse to learn from it — a\n * ratio derived from the same counter that produced the estimate is\n * self-consistent by construction and would drag a provider's real\n * calibration toward 1.0.\n */\nfunction synthesizeSealedUsage(\n  context: InvokeContext | undefined,\n  chunk: AIMessageChunk,\n  prompt: BaseMessage[],\n  metadata: Record<string, unknown> | undefined\n): void {\n  if (chunk.usage_metadata != null) {\n    return;\n  }\n  const outputTokens = countSealedTokens(context, metadata, [chunk]);\n  if (outputTokens == null) {\n    return;\n  }\n  const inputTokens =\n    (countSealedTokens(context, metadata, prompt) ?? 0) +\n    sealedInstructionOverhead(context, metadata);\n  chunk.usage_metadata = {\n    input_tokens: inputTokens,\n    output_tokens: outputTokens,\n    total_tokens: inputTokens + outputTokens,\n  };\n  chunk.response_metadata = {\n    ...chunk.response_metadata,\n    estimated_usage: true,\n  };\n}\n\nfunction getMessageText(chunk: AIMessageChunk): string {\n  if (typeof chunk.content === 'string') {\n    return chunk.content;\n  }\n  let text = '';\n  for (const block of chunk.content) {\n    if (block.type === ContentTypes.TEXT) {\n      const value = block[ContentTypes.TEXT];\n      if (typeof value === 'string') {\n        text += value;\n      }\n    }\n  }\n  return text;\n}\n\n/**\n * Ends the real model run for a turn that was sealed mid-stream.\n *\n * Mandatory, not cosmetic. `@langchain/core`'s `_streamIterator` calls\n * `handleLLMEnd` after its try/catch with no `finally`, so breaking out of the\n * consumer's `for await` produces a *return* completion that fires neither\n * `handleLLMError` nor `handleLLMEnd`. The run would stay open in every\n * callback handler: the host records no usage — and since each seal re-sends\n * the whole prompt, N preemptions cost N unrecorded prompts — while LangSmith\n * and Langfuse hold a span that never closes.\n *\n * `runId` cannot be dictated from here (the bound runnable consumes\n * `config.runId` for its own run and hands the chat model a fresh one), but it\n * can be OBSERVED: the capture handler installed at the `model.stream` call\n * records it from `handleChatModelStart`, which fires before the first chunk.\n * Rebuilding the manager against that id closes the real run, and the host's\n * `on_chat_model_end` then arrives through the ordinary `streamEvents` path.\n *\n * Falls back to a custom-event dispatch if the id was never observed, so the\n * host still records usage even when the native close is unavailable.\n */\n/**\n * Every callbacks source the real model run would compose beyond the per-call\n * config. `model` here is whatever `createCallModel` produced — with tools\n * that is `bindTools(...)`'s `RunnableBinding`, and a system runnable pipes a\n * `RunnableSequence` on top — while `clientOptions.callbacks` lives on the\n * chat model at the BOTTOM of that stack. Walks `bound` (bindings) and\n * `last`/`steps` (sequences), collecting each wrapper's own `callbacks` and\n * any binding-config callbacks along the way, since the binding merges its\n * config into the call before the chat model composes.\n */\nfunction collectModelCallbackSources(model: unknown): Callbacks[] {\n  const sources: Callbacks[] = [];\n  const seen = new Set<unknown>();\n  let current: unknown = model;\n  while (current != null && typeof current === 'object' && !seen.has(current)) {\n    seen.add(current);\n    const wrapper = current as {\n      callbacks?: Callbacks;\n      config?: { callbacks?: Callbacks };\n      bound?: unknown;\n      last?: unknown;\n      steps?: unknown[];\n    };\n    if (wrapper.callbacks != null) {\n      sources.push(wrapper.callbacks);\n    }\n    if (wrapper.config?.callbacks != null) {\n      sources.push(wrapper.config.callbacks);\n    }\n    current =\n      wrapper.bound ??\n      wrapper.last ??\n      (Array.isArray(wrapper.steps)\n        ? wrapper.steps[wrapper.steps.length - 1]\n        : undefined);\n  }\n  return sources;\n}\n\n/**\n * The serving model's id, read through the same wrapper stack\n * `collectModelCallbackSources` walks — `bindTools` returns a\n * `RunnableBinding` and a system runnable pipes a `RunnableSequence`, and\n * neither exposes the chat model's `model` at the top level.\n */\nexport function resolveServingModelId(model: unknown): string | undefined {\n  const seen = new Set<unknown>();\n  let current: unknown = model;\n  while (current != null && typeof current === 'object' && !seen.has(current)) {\n    seen.add(current);\n    const wrapper = current as {\n      model?: unknown;\n      bound?: unknown;\n      last?: unknown;\n      steps?: unknown[];\n    };\n    if (typeof wrapper.model === 'string' && wrapper.model !== '') {\n      return wrapper.model;\n    }\n    current =\n      wrapper.bound ??\n      wrapper.last ??\n      (Array.isArray(wrapper.steps)\n        ? wrapper.steps[wrapper.steps.length - 1]\n        : undefined);\n  }\n  return undefined;\n}\n\nasync function endSealedModelRun(\n  context: InvokeContext | undefined,\n  chunk: AIMessageChunk,\n  prompt: BaseMessage[],\n  llmRunId: string | undefined,\n  config?: RunnableConfig,\n  model?: t.ChatModel\n): Promise<void> {\n  const metadata = config?.metadata as Record<string, unknown> | undefined;\n  synthesizeSealedUsage(context, chunk, prompt, metadata);\n  if (llmRunId != null) {\n    try {\n      let callbackManager = await getCallbackManagerForConfig(config);\n      /**\n       * The real model run composes the per-call config's callbacks WITH the\n       * model's own (`CallbackManager.configure(config.callbacks,\n       * this.callbacks, …)` in `@langchain/core`'s base chat model), so a\n       * handler supplied via `clientOptions.callbacks` received\n       * `handleChatModelStart` for this run. Rebuilding from the config alone\n       * would close the run for every handler EXCEPT those — leaving their\n       * span open forever. Composed the same way the real run composes:\n       * model callbacks appended non-inheritable, parent run id preserved by\n       * `copy`, tracers deduped by `configure`.\n       */\n      for (const source of collectModelCallbackSources(model)) {\n        callbackManager =\n          CallbackManager.configure(callbackManager ?? undefined, source) ??\n          callbackManager;\n      }\n      if (callbackManager != null) {\n        const runManager = new CallbackManagerForLLMRun(\n          llmRunId,\n          callbackManager.handlers,\n          callbackManager.inheritableHandlers,\n          callbackManager.tags,\n          callbackManager.inheritableTags,\n          callbackManager.metadata,\n          callbackManager.inheritableMetadata,\n          callbackManager.getParentRunId()\n        );\n        const generation: ChatGeneration = {\n          text: getMessageText(chunk),\n          message: chunk,\n        };\n        await runManager.handleLLMEnd({\n          generations: [[generation]],\n          llmOutput: {},\n        });\n        return;\n      }\n    } catch (e) {\n      /**\n       * A sealed answer that reaches the user is worth more than a tidy\n       * trace. Fall through to the custom event rather than failing the run.\n       */\n      // eslint-disable-next-line no-console\n      console.warn(\n        '[attemptInvoke] Native close of the sealed model run failed; falling back to a custom event:',\n        e instanceof Error ? e.message : e\n      );\n    }\n  }\n  await safeDispatchCustomEvent(\n    GraphEvents.CHAT_MODEL_END,\n    { output: chunk },\n    config\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 invocationMessages = projectMessagesForProvider({\n    model,\n    messages,\n    provider,\n    callOptions: config,\n  });\n  const registry = context?.getOrCreateToolOutputRegistry();\n  const runId = config?.configurable?.run_id as string | undefined;\n  const annotated = annotateMessagesForLLM(\n    invocationMessages,\n    registry,\n    runId\n  );\n  /**\n   * Keyed on the provider ACTUALLY serving this call, not the agent's primary.\n   * `createCallModel` normalizes for the primary, but `tryFallbackProviders`\n   * re-sends the same array — so an OpenAI primary that fails after a boundary\n   * injected two human turns would hand a Bedrock or Mistral fallback the\n   * consecutive user turns those APIs reject, and the recovery request would\n   * fail for a reason unrelated to the original failure.\n   *\n   * `attemptInvoke` is the single funnel for primary, fallback and\n   * summarization calls, so applying it here covers all three. Idempotent, so\n   * the primary simply re-runs a no-op over already-coalesced messages.\n   */\n  /**\n   * Serving-provider re-keying for the predecessor handoff cue (#345). The\n   * PRIMARY's cue is baked in createCallModel's measured transform stage —\n   * appending after measurement could push a just-fits prompt over budget —\n   * so this funnel only corrects for fallbacks crossing provider families:\n   * a tolerant primary falling back to a Claude surface gains the cue here,\n   * and an Anthropic primary falling back to OpenAI/Mistral/Nova has the\n   * Claude-only synthetic turn stripped. Both helpers are identity on their\n   * no-op paths, so the primary's own pass re-runs for free.\n   *\n   * The serving model id is read through the wrapper stack (`bindTools`'\n   * binding, a system runnable's sequence) — a wrapper's top-level `.model`\n   * is undefined, and `isAnthropicLike` would otherwise default a wrapped\n   * Bedrock-Nova model to Claude. The context cast is widened deliberately:\n   * the type says every context is a full Graph, but summarization passes\n   * none and long-standing tests pass partial stubs.\n   */\n  const isRunProduced = (\n    context as\n      | { isRunProducedMessage?: (message: BaseMessage) => boolean }\n      | undefined\n  )?.isRunProducedMessage;\n  const cued = isAnthropicLike(provider, {\n    model: resolveServingModelId(model),\n  })\n    ? appendPredecessorHandoffCue(\n      annotated,\n      isRunProduced == null\n        ? undefined\n        : (message): boolean => isRunProduced.call(context, message)\n    )\n    : removePredecessorHandoffCue(annotated);\n  const messagesForProvider = strictAlternationProviders.has(provider)\n    ? coalesceAdjacentUserTurns(cued)\n    : cued;\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    /**\n     * Observed, not dictated. `handleChatModelStart` fires with the chat\n     * model's real run id before the first chunk, which is the only way to\n     * name the run a seal has to close — pinning `config.runId` does not\n     * survive the bound runnable. Installed only when preemption is\n     * configured, so a run that cannot seal carries no extra handler.\n     */\n    let sealedRunId: string | undefined;\n    const streamConfig =\n      context?.preemption == null\n        ? config\n        : {\n          ...config,\n          callbacks: appendCallbacks(config.callbacks, [\n            {\n              handleChatModelStart: (\n                _llm: Serialized,\n                _messages: BaseMessage[][],\n                runId: string\n              ): void => {\n                sealedRunId ??= runId;\n              },\n            },\n          ]),\n        };\n    const stream = await model.stream(messagesForProvider, streamConfig);\n    let finalChunk: AIMessageChunk | undefined;\n    let preempted = false;\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         * Only this loop may seal. The registered-handler branch below\n         * dispatches through `run.ts`'s decoupled `streamEvents` consumer,\n         * which can lag the accumulated chunk — sealing there would let the\n         * host index a content part the user has not been shown yet.\n         */\n        /**\n         * Cheap poll first, shape check second, budget claim last. The claim\n         * is what makes this safe under a parallel `MultiAgentGraph`: several\n         * agents share one graph and can each see the poll as true, but only\n         * one can take the slot, and a chunk that cannot seal never spends it.\n         */\n        if (\n          context?.shouldPreemptStream() === true &&\n          canSealPreempt(finalChunk) &&\n          context.claimPreemptSeal()\n        ) {\n          preempted = true;\n          break;\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 (preempted && finalChunk != null) {\n      finalChunk.response_metadata = {\n        ...finalChunk.response_metadata,\n        preempted: true,\n      };\n      await endSealedModelRun(\n        context,\n        finalChunk,\n        messagesForProvider,\n        sealedRunId,\n        config,\n        model\n      );\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  prepareProviderMessages,\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  /**\n   * Optional final payload guard used by Graph. It receives the initialized,\n   * tool-bound fallback model so Responses-vs-Chat projection is exact before\n   * the fallback request is measured and sent.\n   */\n  prepareProviderMessages?: (input: {\n    model: t.ChatModel;\n    messages: BaseMessage[];\n    provider: Providers;\n    clientOptions?: t.ClientOptions;\n    maxContextTokens?: number;\n    config?: RunnableConfig;\n  }) => BaseMessage[] | Promise<BaseMessage[]>;\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 fallbackMessages =\n        (await prepareProviderMessages?.({\n          model: fbModel as t.ChatModel,\n          messages,\n          provider: fb.provider,\n          clientOptions: fb.clientOptions,\n          maxContextTokens: fb.maxContextTokens,\n          config: fbConfig,\n        })) ?? messages;\n      const result = await attemptInvoke(\n        {\n          model: fbModel as t.ChatModel,\n          messages: fallbackMessages,\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":";;;;;;;;;;;;;;;;;;;;;;AAwFA,SAAgB,0BACd,OACA,UACA,aACS;CACT,IAAI,CAACA,YAAAA,aAAa,QAAQ,GACxB,OAAO;CAET,IAAI,YAAqB;CACzB,IAAI,uBAAuB;CAC3B,MAAM,uBAAO,IAAI,IAAY;CAC7B,KAAK,IAAI,QAAQ,GAAG,QAAQ,IAAI,SAAS;EACvC,IAAI,aAAa,QAAQ,OAAO,cAAc,UAC5C,OAAO;EAET,IAAI,KAAK,IAAI,SAAS,GACpB,OAAO;EAET,KAAK,IAAI,SAAS;EAClB,MAAM,WAAW;EAOjB,IAAI;GACF,IACE,SAAS,kBAAkB,QAC3B,OAAO,SAAS,mBAAmB,YACnC,CAAC,MAAM,QAAQ,SAAS,cAAc,KACtC,wBAAwB,QACxB,OAAO,yBAAyB,YAChC,CAAC,MAAM,QAAQ,oBAAoB,GAEnC,uBAAuB;IACrB,GAAI,SAAS;IACb,GAAI;GACN;QACK,IAAI,wBAAwB,MACjC,uBAAuB,SAAS;GAElC,IACE,SAAS,mBAAmB,oBAAoB,MAAM,QACtD,SAAS,mBAAmB,KAAA,CAAS,MAAM,MAE3C,OAAO;EAEX,QAAQ,CAER;EACA,IACE,OAAO,SAAS,aAAa,SAAS,YACtC,SAAS,YAAY,KAAK,SAAS,WAAW,GAE9C,OAAO;EAET,IAAI,SAAS,QAAQ,QAAQ,OAAO,SAAS,SAAS,UAAU;GAC9D,YAAY,SAAS;GACrB;EACF;EACA,IAAI,SAAS,SAAS,QAAQ,OAAO,SAAS,UAAU,UAAU;GAChE,YAAY,SAAS;GACrB;EACF;EACA,OAAO;CACT;CACA,OAAO;AACT;;;;;;AAOA,SAAgB,2BAA2B,EACzC,OACA,UACA,UACA,oBACA,eAOgB;CAChB,MAAM,wBAAwBC,aAAAA,oCAAoC,QAAQ;CAC1E,IAAI,0BAA0B,OAAO,UAAU,WAAW,GACxD,OAAOC,aAAAA,yCACLC,cAAAA,2BACEC,cAAAA,yBAAyB,qBAAqB,CAChD,GACA,kBACF;CAEF,IAAI,aAAA,cACF,OAAOC,aAAAA,iCACLC,aAAAA,oCACEF,cAAAA,yBAAyB,qBAAqB,GAC9C,kBACF,CACF;CAEF,IAAIJ,YAAAA,aAAa,QAAQ,GACvB,OAAOK,aAAAA,iCACLE,aAAAA,oCACEJ,cAAAA,2BACEC,cAAAA,yBAAyB,qBAAqB,CAChD,GACA,kBACF,CACF;CAEF,IAAI,aAAA,aACF,OAAOC,aAAAA,iCACLG,aAAAA,mCACEJ,cAAAA,yBAAyB,qBAAqB,GAC9C,kBACF,CACF;CAEF,IAAI,aAAA,WACF,OAAOD,cAAAA,2BACLE,aAAAA,iCACEI,aAAAA,wCACE,uBACA,kBACF,CACF,CACF;CAEF,OAAOJ,aAAAA,iCACLK,aAAAA,mCACEF,aAAAA,mCACEL,cAAAA,2BACEC,cAAAA,yBAAyB,qBAAqB,CAChD,GACA,kBACF,GACA,kBACF,CACF;AACF;;;;;;;;;;;AAYA,SAAS,sCACP,SAC4B;CAC5B,MAAM,UAAU,SAAS,iBAAiB,WAAA,sBAE1C;CACA,OAAOO,eAAAA,0BAA0B,OAAO,IAAI,UAAU,KAAA;AACxD;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,IAAIC,yBAAAA,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,IAAIA,yBAAAA,eACT,OAAO,OAAO,CAAC,GAAG,MAAM;EACtB;EACA,mBAAmB,uBAAuB,KAAK,iBAAiB;CAClE,CAAC,CACH;AACF;;;;;AAMA,SAAS,kBACP,SACA,UACA,UACoB;CACpB,IAAI;EACF,MAAM,UAAU,SAAS,gBAAgB,QAAQ,CAAC,CAAC;EACnD,IAAI,WAAW,MACb;EAEF,IAAI,QAAQ;EACZ,KAAK,MAAM,WAAW,UACpB,SAAS,QAAQ,OAAO;EAE1B,OAAO;CACT,QAAQ;EACN;CACF;AACF;;;;;;;;;;;;AAaA,SAAS,0BACP,SACA,UACQ;CACR,IAAI;EACF,MAAM,eAAe,SAAS,gBAAgB,QAAQ;EACtD,OACE,cAAc,+BACd,cAAc,qBACd;CAEJ,QAAQ;EACN,OAAO;CACT;AACF;;;;;;;;;;;;;;;;;;;;;;AAuBA,SAAS,sBACP,SACA,OACA,QACA,UACM;CACN,IAAI,MAAM,kBAAkB,MAC1B;CAEF,MAAM,eAAe,kBAAkB,SAAS,UAAU,CAAC,KAAK,CAAC;CACjE,IAAI,gBAAgB,MAClB;CAEF,MAAM,eACH,kBAAkB,SAAS,UAAU,MAAM,KAAK,KACjD,0BAA0B,SAAS,QAAQ;CAC7C,MAAM,iBAAiB;EACrB,cAAc;EACd,eAAe;EACf,cAAc,cAAc;CAC9B;CACA,MAAM,oBAAoB;EACxB,GAAG,MAAM;EACT,iBAAiB;CACnB;AACF;AAEA,SAAS,eAAe,OAA+B;CACrD,IAAI,OAAO,MAAM,YAAY,UAC3B,OAAO,MAAM;CAEf,IAAI,OAAO;CACX,KAAK,MAAM,SAAS,MAAM,SACxB,IAAI,MAAM,SAAA,QAA4B;EACpC,MAAM,QAAQ,MAAA;EACd,IAAI,OAAO,UAAU,UACnB,QAAQ;CAEZ;CAEF,OAAO;AACT;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAiCA,SAAS,4BAA4B,OAA6B;CAChE,MAAM,UAAuB,CAAC;CAC9B,MAAM,uBAAO,IAAI,IAAa;CAC9B,IAAI,UAAmB;CACvB,OAAO,WAAW,QAAQ,OAAO,YAAY,YAAY,CAAC,KAAK,IAAI,OAAO,GAAG;EAC3E,KAAK,IAAI,OAAO;EAChB,MAAM,UAAU;EAOhB,IAAI,QAAQ,aAAa,MACvB,QAAQ,KAAK,QAAQ,SAAS;EAEhC,IAAI,QAAQ,QAAQ,aAAa,MAC/B,QAAQ,KAAK,QAAQ,OAAO,SAAS;EAEvC,UACE,QAAQ,SACR,QAAQ,SACP,MAAM,QAAQ,QAAQ,KAAK,IACxB,QAAQ,MAAM,QAAQ,MAAM,SAAS,KACrC,KAAA;CACR;CACA,OAAO;AACT;;;;;;;AAQA,SAAgB,sBAAsB,OAAoC;CACxE,MAAM,uBAAO,IAAI,IAAa;CAC9B,IAAI,UAAmB;CACvB,OAAO,WAAW,QAAQ,OAAO,YAAY,YAAY,CAAC,KAAK,IAAI,OAAO,GAAG;EAC3E,KAAK,IAAI,OAAO;EAChB,MAAM,UAAU;EAMhB,IAAI,OAAO,QAAQ,UAAU,YAAY,QAAQ,UAAU,IACzD,OAAO,QAAQ;EAEjB,UACE,QAAQ,SACR,QAAQ,SACP,MAAM,QAAQ,QAAQ,KAAK,IACxB,QAAQ,MAAM,QAAQ,MAAM,SAAS,KACrC,KAAA;CACR;AAEF;AAEA,eAAe,kBACb,SACA,OACA,QACA,UACA,QACA,OACe;CACf,MAAM,WAAW,QAAQ;CACzB,sBAAsB,SAAS,OAAO,QAAQ,QAAQ;CACtD,IAAI,YAAY,MACd,IAAI;EACF,IAAI,kBAAkB,OAAA,GAAA,0BAAA,4BAAA,CAAkC,MAAM;;;;;;;;;;;;EAY9D,KAAK,MAAM,UAAU,4BAA4B,KAAK,GACpD,kBACEC,kCAAAA,gBAAgB,UAAU,mBAAmB,KAAA,GAAW,MAAM,KAC9D;EAEJ,IAAI,mBAAmB,MAAM;GAC3B,MAAM,aAAa,IAAIC,kCAAAA,yBACrB,UACA,gBAAgB,UAChB,gBAAgB,qBAChB,gBAAgB,MAChB,gBAAgB,iBAChB,gBAAgB,UAChB,gBAAgB,qBAChB,gBAAgB,eAAe,CACjC;GACA,MAAM,aAA6B;IACjC,MAAM,eAAe,KAAK;IAC1B,SAAS;GACX;GACA,MAAM,WAAW,aAAa;IAC5B,aAAa,CAAC,CAAC,UAAU,CAAC;IAC1B,WAAW,CAAC;GACd,CAAC;GACD;EACF;CACF,SAAS,GAAG;;;;;EAMV,QAAQ,KACN,gGACA,aAAa,QAAQ,EAAE,UAAU,CACnC;CACF;CAEF,MAAMC,eAAAA,wBAAAA,qBAEJ,EAAE,QAAQ,MAAM,GAChB,MACF;AACF;AAEA,SAAS,kBAAkB,EACzB,SACA,MACA,YAKiB;CACjB,IAAI,WAAW,MACb,OAAO;CAET,QAAA,GAAA,6BAAA,OAAA,CACE,SACA,4CAA4C;EAAE;EAAS;EAAM;CAAS,CAAC,CACzE;AACF;;;;;;;;;;AAWA,eAAsB,cACpB,EACE,OACA,UACA,UACA,SACA,WAQF,QACoC;;;;;;;;CAQpC,MAAM,qBAAqB,2BAA2B;EACpD;EACA;EACA;EACA,aAAa;CACf,CAAC;CACD,MAAM,WAAW,SAAS,8BAA8B;CACxD,MAAM,QAAQ,QAAQ,cAAc;CACpC,MAAM,YAAYC,6BAAAA,uBAChB,oBACA,UACA,KACF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;CA8BA,MAAM,gBACJ,SAGC;CACH,MAAM,OAAOC,YAAAA,gBAAgB,UAAU,EACrC,OAAO,sBAAsB,KAAK,EACpC,CAAC,IACGC,mBAAAA,4BACA,WACA,iBAAiB,OACb,KAAA,KACC,YAAqB,cAAc,KAAK,SAAS,OAAO,CAC/D,IACEC,mBAAAA,4BAA4B,SAAS;CACzC,MAAM,sBAAsBC,oBAAAA,2BAA2B,IAAI,QAAQ,IAC/DC,oBAAAA,0BAA0B,IAAI,IAC9B;;;;;;;;;;CAWJ,SAAS;EACP,GAAG;EACH,UAAU;GACR,GAAI,QAAQ,YAAY,CAAC;2BACK;EAChC;CACF;CAEA,IAAI,MAAM,QAAQ;;;;;;;;EAQhB,IAAI;EACJ,MAAM,eACJ,SAAS,cAAc,OACnB,SACA;GACA,GAAG;GACH,WAAWC,kBAAAA,gBAAgB,OAAO,WAAW,CAC3C,EACE,uBACE,MACA,WACA,UACS;IACT,gBAAgB;GAClB,EACF,CACF,CAAC;EACH;EACJ,MAAM,SAAS,MAAM,MAAM,OAAO,qBAAqB,YAAY;EACnE,IAAI;EACJ,IAAI,YAAY;EAChB,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,IAAIC,eAAAA,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;;;;;;;;;;;;;IAaD,IACE,SAAS,oBAAoB,MAAM,QACnCC,gBAAAA,eAAe,UAAU,KACzB,QAAQ,iBAAiB,GACzB;KACA,YAAY;KACZ;IACF;GACF;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,IAAIC,kBAAAA,0BAA0B,IAAI,QAAQ,GACxC,aAAaC,aAAAA,sBAAsB,UAAU,UAAU;EAGzD,IAAI,aAAa,cAAc,MAAM;GACnC,WAAW,oBAAoB;IAC7B,GAAG,WAAW;IACd,WAAW;GACb;GACA,MAAM,kBACJ,SACA,YACA,qBACA,aACA,QACA,KACF;EACF;EAEA,KAAK,YAAY,YAAY,UAAU,KAAK,GAC1C,WAAY,aAAa,WAAY,YAAY,QAC9C,cAAwB,CAAC,CAAC,UAAU,IACvC;EAGF,mBAAA,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,mBAAA,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,iBACA,2BA6BiD;CACjD,MAAM,cACJ,OACA,kBAAkB,oBACNC,eAAAA,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,UAAUC,aAAAA,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;EAoBJ,OAAO,MAVc,cACnB;GACE,OAAO;GACP,UAXD,MAAM,0BAA0B;IAC/B,OAAO;IACP;IACA,UAAU,GAAG;IACb,eAAe,GAAG;IAClB,kBAAkB,GAAG;IACrB,QAAQ;GACV,CAAC,KAAM;GAKL,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"}
+{"version":3,"file":"invoke.cjs","names":["isOpenAILike","projectToolStreamContentForProvider","projectOpenAIResponsesToolMessageContent","stripAnthropicCacheControl","stripBedrockCacheControl","projectComputerCallOutputsToText","projectOpenRouterToolMessageContent","projectOpenAIChatToolMessageContent","projectSingleTextToolOutputsToText","projectCacheControlledToolOutputsToText","projectStructuredToolOutputsToText","dispatchesChatModelStream","AIMessageChunk","CallbackManager","CallbackManagerForLLMRun","safeDispatchCustomEvent","STREAM_LIMIT_ATTEMPT_KEY","streamLimitAccountingEnabled","resolveGenerationKey","annotateMessagesForLLM","isAnthropicLike","appendPredecessorHandoffCue","removePredecessorHandoffCue","strictAlternationProviders","coalesceAdjacentUserTurns","appendCallbacks","StreamLimitExceededError","ChatModelStreamHandler","canSealPreempt","STREAM_LIMIT_REDISPATCH_KEY","manualToolStreamProviders","modifyDeltaProperties","getContextOverflowInfo","initializeModel"],"sources":["../../../src/llm/invoke.ts"],"sourcesContent":["import { concat } from '@langchain/core/utils/stream';\nimport { AIMessageChunk } from '@langchain/core/messages';\nimport { getCallbackManagerForConfig } from '@langchain/core/runnables';\nimport {\n  CallbackManager,\n  CallbackManagerForLLMRun,\n  type Callbacks,\n} from '@langchain/core/callbacks/manager';\nimport type { Serialized } from '@langchain/core/load/serializable';\nimport type { RunnableConfig } from '@langchain/core/runnables';\nimport type { ChatGeneration } from '@langchain/core/outputs';\nimport type { ToolCall } from '@langchain/core/messages/tool';\nimport type { BaseMessage } from '@langchain/core/messages';\nimport type { ToolOutputReferenceRegistry } from '@/tools/toolOutputReferences';\nimport type { StreamLimitState } from '@/llm/streamLimits';\nimport type { ContextOverflowContext } from '@/utils/errors';\nimport type * as t from '@/types';\nimport {\n  projectCacheControlledToolOutputsToText,\n  projectComputerCallOutputsToText,\n  projectOpenAIChatToolMessageContent,\n  projectOpenAIResponsesToolMessageContent,\n  projectOpenRouterToolMessageContent,\n  projectSingleTextToolOutputsToText,\n  projectStructuredToolOutputsToText,\n  projectToolStreamContentForProvider,\n} from '@/messages/core';\nimport {\n  modifyDeltaProperties,\n  coalesceAdjacentUserTurns,\n  strictAlternationProviders,\n  appendPredecessorHandoffCue,\n  removePredecessorHandoffCue,\n} from '@/messages';\nimport {\n  stripAnthropicCacheControl,\n  stripBedrockCacheControl,\n} from '@/messages/cache';\nimport { ChatModelStreamHandler, dispatchesChatModelStream } from '@/stream';\nimport { Constants, ContentTypes, GraphEvents, Providers } from '@/common';\nimport {\n  enforceStreamLimitsForWireChunk,\n  registerActiveStreamLimitGeneration,\n  releaseStreamLimitGeneration,\n  resolveGenerationKey,\n  streamLimitAccountingEnabled,\n  StreamLimitExceededError,\n  STREAM_LIMIT_REDISPATCH_KEY,\n  STREAM_LIMIT_ATTEMPT_KEY,\n} from '@/llm/streamLimits';\nimport { annotateMessagesForLLM } from '@/tools/toolOutputReferences';\nimport { assertNotTruncatedToolCall } from '@/llm/truncation';\nimport { manualToolStreamProviders } from '@/llm/providers';\nimport { isAnthropicLike, isOpenAILike } from '@/utils/llm';\nimport { safeDispatchCustomEvent } from '@/utils/events';\nimport { getContextOverflowInfo } from '@/utils/errors';\nimport { appendCallbacks } from '@/utils/callbacks';\nimport { canSealPreempt } from '@/llm/preempt';\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 *\n * `metadata` is the attempt's callback metadata (carrying the provider and\n * stream-limit attempt stamps), so consumers that count against the stream\n * limits key each model attempt separately.\n */\nexport type OnChunk = (\n  chunk: AIMessageChunk,\n  metadata?: Record<string, unknown>\n) => void | Promise<void>;\n\n/** Unique per-model-attempt sequence; see the stamp in `attemptInvoke`. */\nlet streamLimitAttemptSeq = 0;\n\nexport function usesNativeOpenAIResponses(\n  model: t.ChatModel,\n  provider: Providers,\n  callOptions?: unknown\n): boolean {\n  if (!isOpenAILike(provider)) {\n    return false;\n  }\n  let candidate: unknown = model;\n  let effectiveCallOptions = callOptions;\n  const seen = new Set<object>();\n  for (let depth = 0; depth < 20; depth++) {\n    if (candidate == null || typeof candidate !== 'object') {\n      return false;\n    }\n    if (seen.has(candidate)) {\n      return false;\n    }\n    seen.add(candidate);\n    const runnable = candidate as {\n      _useResponsesApi?: (options?: unknown) => boolean;\n      bound?: unknown;\n      defaultOptions?: unknown;\n      last?: unknown;\n      constructor?: { name?: unknown };\n    };\n    try {\n      if (\n        runnable.defaultOptions != null &&\n        typeof runnable.defaultOptions === 'object' &&\n        !Array.isArray(runnable.defaultOptions) &&\n        effectiveCallOptions != null &&\n        typeof effectiveCallOptions === 'object' &&\n        !Array.isArray(effectiveCallOptions)\n      ) {\n        effectiveCallOptions = {\n          ...(runnable.defaultOptions as Record<string, unknown>),\n          ...(effectiveCallOptions as Record<string, unknown>),\n        };\n      } else if (effectiveCallOptions == null) {\n        effectiveCallOptions = runnable.defaultOptions;\n      }\n      if (\n        runnable._useResponsesApi?.(effectiveCallOptions) === true ||\n        runnable._useResponsesApi?.(undefined) === true\n      ) {\n        return true;\n      }\n    } catch {\n      // Continue through RunnableSequence/RunnableBinding wrappers.\n    }\n    if (\n      typeof runnable.constructor?.name === 'string' &&\n      runnable.constructor.name.includes('Responses')\n    ) {\n      return true;\n    }\n    if (runnable.last != null && typeof runnable.last === 'object') {\n      candidate = runnable.last;\n      continue;\n    }\n    if (runnable.bound != null && typeof runnable.bound === 'object') {\n      candidate = runnable.bound;\n      continue;\n    }\n    return false;\n  }\n  return false;\n}\n\n/**\n * Produces the exact provider-facing message representation before a model\n * adapter serializes it. This is shared by invocation and Graph's final budget\n * guard so structured tool output cannot grow after the payload was measured.\n */\nexport function projectMessagesForProvider({\n  model,\n  messages,\n  provider,\n  maxToolResultChars,\n  callOptions,\n}: {\n  model: t.ChatModel;\n  messages: BaseMessage[];\n  provider: Providers;\n  maxToolResultChars?: number;\n  callOptions?: unknown;\n}): BaseMessage[] {\n  const nativeOpenAIResponses = usesNativeOpenAIResponses(\n    model,\n    provider,\n    callOptions\n  );\n  const providerInputMessages = projectToolStreamContentForProvider(\n    messages,\n    nativeOpenAIResponses ? 'native' : 'fallback',\n    maxToolResultChars\n  );\n  if (nativeOpenAIResponses) {\n    return projectOpenAIResponsesToolMessageContent(\n      stripAnthropicCacheControl(\n        stripBedrockCacheControl(providerInputMessages)\n      ),\n      maxToolResultChars\n    );\n  }\n  if (provider === Providers.OPENROUTER) {\n    return projectComputerCallOutputsToText(\n      projectOpenRouterToolMessageContent(\n        stripBedrockCacheControl(providerInputMessages),\n        maxToolResultChars\n      )\n    );\n  }\n  if (isOpenAILike(provider)) {\n    return projectComputerCallOutputsToText(\n      projectOpenAIChatToolMessageContent(\n        stripAnthropicCacheControl(\n          stripBedrockCacheControl(providerInputMessages)\n        ),\n        maxToolResultChars\n      )\n    );\n  }\n  if (provider === Providers.ANTHROPIC) {\n    return projectComputerCallOutputsToText(\n      projectSingleTextToolOutputsToText(\n        stripBedrockCacheControl(providerInputMessages),\n        maxToolResultChars\n      )\n    );\n  }\n  if (provider === Providers.BEDROCK) {\n    return stripAnthropicCacheControl(\n      projectComputerCallOutputsToText(\n        projectCacheControlledToolOutputsToText(\n          providerInputMessages,\n          maxToolResultChars\n        )\n      )\n    );\n  }\n  return projectComputerCallOutputsToText(\n    projectStructuredToolOutputsToText(\n      projectSingleTextToolOutputsToText(\n        stripAnthropicCacheControl(\n          stripBedrockCacheControl(providerInputMessages)\n        ),\n        maxToolResultChars\n      ),\n      maxToolResultChars\n    )\n  );\n}\n\n/**\n * The registered handler that owns content-part dispatch, if any.\n *\n * Detected by brand rather than by `instanceof`: a host that registers\n * `new ChatModelStreamHandler()` to opt out of sealing gets wrapped by\n * `createRunHandlers` on every `AgentSession` run, and by\n * `composeEventHandlers` on a key collision. Both wrappers forward to the same\n * dispatcher while failing an identity check, so an identity test would\n * silently revoke the opt-out documented on `StreamPreemption`.\n */\nfunction getRegisteredDefaultChatStreamHandler(\n  context?: InvokeContext\n): t.EventHandler | undefined {\n  const handler = context?.handlerRegistry?.getHandler(\n    GraphEvents.CHAT_MODEL_STREAM\n  );\n  return dispatchesChatModelStream(handler) ? 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\n/**\n * Best-effort output-token count for a sealed turn, used only when the\n * provider never got to send its usage chunk.\n */\nfunction countSealedTokens(\n  context: InvokeContext | undefined,\n  metadata: Record<string, unknown> | undefined,\n  messages: BaseMessage[]\n): number | undefined {\n  try {\n    const counter = context?.getAgentContext(metadata).tokenCounter;\n    if (counter == null) {\n      return undefined;\n    }\n    let total = 0;\n    for (const message of messages) {\n      total += counter(message);\n    }\n    return total;\n  } catch {\n    return undefined;\n  }\n}\n\n/**\n * Instruction overhead the provider processed but that never appears in the\n * message array: `createCallModel` pipes the model through\n * `agentContext.systemRunnable` and binds tool schemas AFTER `messages` is\n * formed, so the system prompt, dynamic instructions, summary and tool\n * schemas are all billed yet invisible here.\n *\n * Read per-node via `getAgentContext(metadata)` rather than the graph-level\n * accessor, which is hardcoded to `defaultAgentId` and would report the wrong\n * agent's overhead in a `MultiAgentGraph`.\n */\nfunction sealedInstructionOverhead(\n  context: InvokeContext | undefined,\n  metadata: Record<string, unknown> | undefined\n): number {\n  try {\n    const agentContext = context?.getAgentContext(metadata);\n    return (\n      agentContext?.resolvedInstructionOverhead ??\n      agentContext?.instructionTokens ??\n      0\n    );\n  } catch {\n    return 0;\n  }\n}\n\n/**\n * Best-effort usage for a turn the provider never got to bill us for.\n *\n * The prompt matters as much as the completion: the provider processed the\n * ENTIRE prompt — messages plus instruction overhead — before we sealed, and\n * every resume re-sends it, so under-counting input hides the expensive half\n * of a preempted run.\n *\n * ESTIMATE, NOT MEASUREMENT. Messages are counted with the host's tokenizer\n * rather than the provider's, and `toolSchemaTokens` applies a heuristic\n * multiplier. It is also an over-count on the fallback path, where\n * `tryFallbackProviders` builds a bare model with no `systemRunnable` pipe so\n * the system prompt genuinely is not sent. Accepted rather than threaded\n * through a flag: only the fallback-plus-seal combination is affected, and an\n * over-count is safer than the previous fabricated `input_tokens: 0`.\n *\n * Marked `estimated_usage` so calibration can refuse to learn from it — a\n * ratio derived from the same counter that produced the estimate is\n * self-consistent by construction and would drag a provider's real\n * calibration toward 1.0.\n */\nfunction synthesizeSealedUsage(\n  context: InvokeContext | undefined,\n  chunk: AIMessageChunk,\n  prompt: BaseMessage[],\n  metadata: Record<string, unknown> | undefined\n): void {\n  if (chunk.usage_metadata != null) {\n    return;\n  }\n  const outputTokens = countSealedTokens(context, metadata, [chunk]);\n  if (outputTokens == null) {\n    return;\n  }\n  const inputTokens =\n    (countSealedTokens(context, metadata, prompt) ?? 0) +\n    sealedInstructionOverhead(context, metadata);\n  const usageMetadata = {\n    input_tokens: inputTokens,\n    output_tokens: outputTokens,\n    total_tokens: inputTokens + outputTokens,\n  };\n  chunk.usage_metadata = usageMetadata;\n  chunk.lc_kwargs.usage_metadata = usageMetadata;\n  const responseMetadata = {\n    ...chunk.response_metadata,\n    estimated_usage: true,\n  };\n  chunk.response_metadata = responseMetadata;\n  chunk.lc_kwargs.response_metadata = responseMetadata;\n}\n\nfunction getMessageText(chunk: AIMessageChunk): string {\n  if (typeof chunk.content === 'string') {\n    return chunk.content;\n  }\n  let text = '';\n  for (const block of chunk.content) {\n    if (block.type === ContentTypes.TEXT) {\n      const value = block[ContentTypes.TEXT];\n      if (typeof value === 'string') {\n        text += value;\n      }\n    }\n  }\n  return text;\n}\n\n/**\n * Ends the real model run for a turn that was sealed mid-stream.\n *\n * Mandatory, not cosmetic. `@langchain/core`'s `_streamIterator` calls\n * `handleLLMEnd` after its try/catch with no `finally`, so breaking out of the\n * consumer's `for await` produces a *return* completion that fires neither\n * `handleLLMError` nor `handleLLMEnd`. The run would stay open in every\n * callback handler: the host records no usage — and since each seal re-sends\n * the whole prompt, N preemptions cost N unrecorded prompts — while LangSmith\n * and Langfuse hold a span that never closes.\n *\n * `runId` cannot be dictated from here (the bound runnable consumes\n * `config.runId` for its own run and hands the chat model a fresh one), but it\n * can be OBSERVED: the capture handler installed at the `model.stream` call\n * records it from `handleChatModelStart`, which fires before the first chunk.\n * Rebuilding the manager against that id closes the real run, and the host's\n * `on_chat_model_end` then arrives through the ordinary `streamEvents` path.\n *\n * Falls back to a custom-event dispatch if the id was never observed, so the\n * host still records usage even when the native close is unavailable.\n */\n/**\n * Every callbacks source the real model run would compose beyond the per-call\n * config. `model` here is whatever `createCallModel` produced — with tools\n * that is `bindTools(...)`'s `RunnableBinding`, and a system runnable pipes a\n * `RunnableSequence` on top — while `clientOptions.callbacks` lives on the\n * chat model at the BOTTOM of that stack. Walks `bound` (bindings) and\n * `last`/`steps` (sequences), collecting each wrapper's own `callbacks` and\n * any binding-config callbacks along the way, since the binding merges its\n * config into the call before the chat model composes.\n */\nfunction collectModelCallbackSources(model: unknown): Callbacks[] {\n  const sources: Callbacks[] = [];\n  const seen = new Set<unknown>();\n  let current: unknown = model;\n  while (current != null && typeof current === 'object' && !seen.has(current)) {\n    seen.add(current);\n    const wrapper = current as {\n      callbacks?: Callbacks;\n      config?: { callbacks?: Callbacks };\n      bound?: unknown;\n      last?: unknown;\n      steps?: unknown[];\n    };\n    if (wrapper.callbacks != null) {\n      sources.push(wrapper.callbacks);\n    }\n    if (wrapper.config?.callbacks != null) {\n      sources.push(wrapper.config.callbacks);\n    }\n    current =\n      wrapper.bound ??\n      wrapper.last ??\n      (Array.isArray(wrapper.steps)\n        ? wrapper.steps[wrapper.steps.length - 1]\n        : undefined);\n  }\n  return sources;\n}\n\n/**\n * The serving model's id, read through the same wrapper stack\n * `collectModelCallbackSources` walks — `bindTools` returns a\n * `RunnableBinding` and a system runnable pipes a `RunnableSequence`, and\n * neither exposes the chat model's `model` at the top level.\n */\nexport function resolveServingModelId(model: unknown): string | undefined {\n  const seen = new Set<unknown>();\n  let current: unknown = model;\n  while (current != null && typeof current === 'object' && !seen.has(current)) {\n    seen.add(current);\n    const wrapper = current as {\n      model?: unknown;\n      bound?: unknown;\n      last?: unknown;\n      steps?: unknown[];\n    };\n    if (typeof wrapper.model === 'string' && wrapper.model !== '') {\n      return wrapper.model;\n    }\n    current =\n      wrapper.bound ??\n      wrapper.last ??\n      (Array.isArray(wrapper.steps)\n        ? wrapper.steps[wrapper.steps.length - 1]\n        : undefined);\n  }\n  return undefined;\n}\n\nasync function endSealedModelRun(\n  context: InvokeContext | undefined,\n  chunk: AIMessageChunk,\n  prompt: BaseMessage[],\n  llmRunId: string | undefined,\n  config?: RunnableConfig,\n  model?: t.ChatModel\n): Promise<void> {\n  const metadata = config?.metadata as Record<string, unknown> | undefined;\n  synthesizeSealedUsage(context, chunk, prompt, metadata);\n  if (llmRunId != null) {\n    try {\n      let callbackManager = await getCallbackManagerForConfig(config);\n      /**\n       * The real model run composes the per-call config's callbacks WITH the\n       * model's own (`CallbackManager.configure(config.callbacks,\n       * this.callbacks, …)` in `@langchain/core`'s base chat model), so a\n       * handler supplied via `clientOptions.callbacks` received\n       * `handleChatModelStart` for this run. Rebuilding from the config alone\n       * would close the run for every handler EXCEPT those — leaving their\n       * span open forever. Composed the same way the real run composes:\n       * model callbacks appended non-inheritable, parent run id preserved by\n       * `copy`, tracers deduped by `configure`.\n       */\n      for (const source of collectModelCallbackSources(model)) {\n        callbackManager =\n          CallbackManager.configure(callbackManager ?? undefined, source) ??\n          callbackManager;\n      }\n      if (callbackManager != null) {\n        const runManager = new CallbackManagerForLLMRun(\n          llmRunId,\n          callbackManager.handlers,\n          callbackManager.inheritableHandlers,\n          callbackManager.tags,\n          callbackManager.inheritableTags,\n          callbackManager.metadata,\n          callbackManager.inheritableMetadata,\n          callbackManager.getParentRunId()\n        );\n        const generation: ChatGeneration = {\n          text: getMessageText(chunk),\n          message: chunk,\n        };\n        await runManager.handleLLMEnd({\n          generations: [[generation]],\n          llmOutput: {},\n        });\n        return;\n      }\n    } catch (e) {\n      /**\n       * A sealed answer that reaches the user is worth more than a tidy\n       * trace. Fall through to the custom event rather than failing the run.\n       */\n      // eslint-disable-next-line no-console\n      console.warn(\n        '[attemptInvoke] Native close of the sealed model run failed; falling back to a custom event:',\n        e instanceof Error ? e.message : e\n      );\n    }\n  }\n  await safeDispatchCustomEvent(\n    GraphEvents.CHAT_MODEL_END,\n    { output: chunk },\n    config\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 */\ninterface AttemptInvokeParams {\n  model: t.ChatModel;\n  messages: BaseMessage[];\n  provider: Providers;\n  context?: InvokeContext;\n  onChunk?: OnChunk;\n  /** Accounting owner for callers that deliberately pass no `context`\n   * (summarization) — used ONLY for the attempt's accounting lease, never\n   * for charge claims. */\n  streamLimitState?: StreamLimitState;\n}\n\n/**\n * One model attempt. Stamps the attempt identity into callback metadata\n * (see the generation-key notes in `streamLimits.ts`), leases the attempt's\n * accounting for its LIFETIME, and releases both from `finally`: retention\n * must follow the attempt — a cancellation-ignoring straggler keeps its\n * original budget no matter how many runs start and reset while it drains.\n */\nexport async function attemptInvoke(\n  params: AttemptInvokeParams,\n  config?: RunnableConfig\n): Promise<Partial<t.BaseGraphState>> {\n  const stampedConfig: RunnableConfig = {\n    ...config,\n    metadata: {\n      ...(config?.metadata ?? {}),\n      [Constants.INVOKED_PROVIDER]: params.provider,\n      /**\n       * One `attemptInvoke` call is one model attempt; primary, fallback,\n       * and retry attempts within a node otherwise share the same langgraph\n       * metadata, so without a unique attempt stamp a fallback re-streaming\n       * a tool call from scratch would be charged the failed primary's\n       * partial bytes (or a same-named sibling fallback's) and could\n       * falsely trip the stream limits. The stamp rides the same metadata\n       * rebuild that already attributes the serving provider.\n       */\n      [STREAM_LIMIT_ATTEMPT_KEY]: ++streamLimitAttemptSeq,\n    },\n  };\n  const rawLeaseTarget = params.context ?? params.streamLimitState;\n  /** No lease when no guard can fire: the lease only protects accounting\n   * entries, and fully disabled guards must allocate no bookkeeping at\n   * all — per-attempt included. */\n  const leaseTarget =\n    rawLeaseTarget != null && streamLimitAccountingEnabled(rawLeaseTarget)\n      ? rawLeaseTarget\n      : undefined;\n  const generationKey =\n    leaseTarget != null\n      ? resolveGenerationKey(\n        stampedConfig.metadata as Record<string, unknown>\n      )\n      : undefined;\n  if (leaseTarget != null && generationKey != null) {\n    registerActiveStreamLimitGeneration(leaseTarget, generationKey);\n  }\n  try {\n    return await attemptInvokeBody(params, stampedConfig);\n  } finally {\n    if (leaseTarget != null && generationKey != null) {\n      releaseStreamLimitGeneration(leaseTarget, generationKey);\n    }\n  }\n}\n\nasync function attemptInvokeBody(\n  {\n    model,\n    messages,\n    provider,\n    context,\n    onChunk,\n  }: AttemptInvokeParams,\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 invocationMessages = projectMessagesForProvider({\n    model,\n    messages,\n    provider,\n    callOptions: config,\n  });\n  const registry = context?.getOrCreateToolOutputRegistry();\n  const runId = config.configurable?.run_id as string | undefined;\n  const annotated = annotateMessagesForLLM(invocationMessages, registry, runId);\n  /**\n   * Keyed on the provider ACTUALLY serving this call, not the agent's primary.\n   * `createCallModel` normalizes for the primary, but `tryFallbackProviders`\n   * re-sends the same array — so an OpenAI primary that fails after a boundary\n   * injected two human turns would hand a Bedrock or Mistral fallback the\n   * consecutive user turns those APIs reject, and the recovery request would\n   * fail for a reason unrelated to the original failure.\n   *\n   * `attemptInvoke` is the single funnel for primary, fallback and\n   * summarization calls, so applying it here covers all three. Idempotent, so\n   * the primary simply re-runs a no-op over already-coalesced messages.\n   */\n  /**\n   * Serving-provider re-keying for the predecessor handoff cue (#345). The\n   * PRIMARY's cue is baked in createCallModel's measured transform stage —\n   * appending after measurement could push a just-fits prompt over budget —\n   * so this funnel only corrects for fallbacks crossing provider families:\n   * a tolerant primary falling back to a Claude surface gains the cue here,\n   * and an Anthropic primary falling back to OpenAI/Mistral/Nova has the\n   * Claude-only synthetic turn stripped. Both helpers are identity on their\n   * no-op paths, so the primary's own pass re-runs for free.\n   *\n   * The serving model id is read through the wrapper stack (`bindTools`'\n   * binding, a system runnable's sequence) — a wrapper's top-level `.model`\n   * is undefined, and `isAnthropicLike` would otherwise default a wrapped\n   * Bedrock-Nova model to Claude. The context cast is widened deliberately:\n   * the type says every context is a full Graph, but summarization passes\n   * none and long-standing tests pass partial stubs.\n   */\n  const isRunProduced = (\n    context as\n      | { isRunProducedMessage?: (message: BaseMessage) => boolean }\n      | undefined\n  )?.isRunProducedMessage;\n  const cued = isAnthropicLike(provider, {\n    model: resolveServingModelId(model),\n  })\n    ? appendPredecessorHandoffCue(\n      annotated,\n      isRunProduced == null\n        ? undefined\n        : (message): boolean => isRunProduced.call(context, message)\n    )\n    : removePredecessorHandoffCue(annotated);\n  const messagesForProvider = strictAlternationProviders.has(provider)\n    ? coalesceAdjacentUserTurns(cued)\n    : cued;\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  if (model.stream) {\n    /**\n     * Observed, not dictated. `handleChatModelStart` fires with the chat\n     * model's real run id before the first chunk, which is the only way to\n     * name the run a seal has to close — pinning `config.runId` does not\n     * survive the bound runnable. Installed only when preemption is\n     * configured, so a run that cannot seal carries no extra handler.\n     */\n    let sealedRunId: string | undefined;\n    const streamConfig =\n      context?.preemption == null\n        ? config\n        : {\n          ...config,\n          callbacks: appendCallbacks(config.callbacks, [\n            {\n              handleChatModelStart: (\n                _llm: Serialized,\n                _messages: BaseMessage[][],\n                runId: string\n              ): void => {\n                sealedRunId ??= runId;\n              },\n            },\n          ]),\n        };\n    const stream = await model.stream(messagesForProvider, streamConfig);\n    let finalChunk: AIMessageChunk | undefined;\n    let preempted = false;\n    const registeredStreamHandler =\n      getRegisteredDefaultChatStreamHandler(context);\n    /** A sibling's trip aborts the composed signal, but an adapter that\n     * ignores cancellation keeps yielding — and text-only chunks with the\n     * event cap off never throw in enforcement, so nothing else would stop\n     * the drain. Checked on every yielded chunk in all three loops;\n     * throwing closes the iterator and tears down the provider stream. */\n    const throwIfBreakerTripped = (): void => {\n      const signal = config.signal;\n      if (\n        signal?.aborted === true &&\n        signal.reason instanceof StreamLimitExceededError\n      ) {\n        throw signal.reason;\n      }\n    };\n\n    if (onChunk) {\n      const attemptMetadata = config.metadata as\n        | Record<string, unknown>\n        | undefined;\n      for await (const chunk of stream) {\n        throwIfBreakerTripped();\n        /** An onChunk consumer replaces the stream handler entirely, so\n         * stream limits are enforced here for every such caller — public\n         * package consumers get no other accounting. The internal\n         * summarization onChunk charges producer-side itself and passes no\n         * context, precisely so this claim and its own never stack. */\n        if (context != null) {\n          enforceStreamLimitsForWireChunk({\n            graph: context,\n            metadata: attemptMetadata,\n            chunk,\n          });\n        }\n        await onChunk(chunk, attemptMetadata);\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        throwIfBreakerTripped();\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        } else if (context != null) {\n          /**\n           * A replay-skipped chunk yields no handling chunk, and in this\n           * local branch no `streamEvents` consumer judges the wire event\n           * either — yet a cumulative OpenRouter replay can still carry\n           * `tool_call_chunks` or complete `tool_calls` that are appended\n           * below. Charge the full limits (event budget and argument bytes)\n           * directly so neither cap can be bypassed. Consumer side: the\n           * local handler.handle call above claims as consumer, and one\n           * reused chunk object can alternate between these two arms.\n           */\n          enforceStreamLimitsForWireChunk({\n            graph: context,\n            metadata,\n            chunk,\n            side: 'consumer',\n          });\n        }\n        finalChunk = appendStreamChunk({\n          current: finalChunk,\n          next: chunk,\n          provider,\n        });\n        /**\n         * Only this loop may seal. The registered-handler branch below\n         * dispatches through `run.ts`'s decoupled `streamEvents` consumer,\n         * which can lag the accumulated chunk — sealing there would let the\n         * host index a content part the user has not been shown yet.\n         */\n        /**\n         * Cheap poll first, shape check second, budget claim last. The claim\n         * is what makes this safe under a parallel `MultiAgentGraph`: several\n         * agents share one graph and can each see the poll as true, but only\n         * one can take the slot, and a chunk that cannot seal never spends it.\n         */\n        if (\n          context?.shouldPreemptStream() === true &&\n          canSealPreempt(finalChunk) &&\n          context.claimPreemptSeal()\n        ) {\n          preempted = true;\n          break;\n        }\n      }\n    } else {\n      const metadata = config.metadata as Record<string, unknown> | undefined;\n      /**\n       * The original wire chunk still reaches the registered handler through\n       * `streamEvents` (where the late-reasoning skip discards it AFTER the\n       * event guard counts it), so this inline re-dispatch of the transformed\n       * chunk is marked to not consume a second event-budget slot. Allocated\n       * once per attempt, only when a transformation occurs.\n       */\n      let redispatchMetadata: Record<string, unknown> | undefined;\n      for await (const chunk of stream) {\n        throwIfBreakerTripped();\n        /**\n         * Charged synchronously, ahead of the decoupled `streamEvents`\n         * reader that will echo this same chunk to the registered handler:\n         * a lagging reader would otherwise let an oversized complete call\n         * return to LangGraph and reach ToolNode before the queued handler\n         * throws. The chunk is marked so the echo skips accounting.\n         */\n        if (context != null) {\n          enforceStreamLimitsForWireChunk({ graph: context, metadata, chunk });\n        }\n        const handlingChunk = getStreamHandlingChunk({\n          current: finalChunk,\n          next: chunk,\n          provider,\n        });\n        if (handlingChunk != null && handlingChunk !== chunk) {\n          redispatchMetadata ??= {\n            ...(metadata ?? {}),\n            [STREAM_LIMIT_REDISPATCH_KEY]: true,\n          };\n          await registeredStreamHandler.handle(\n            GraphEvents.CHAT_MODEL_STREAM,\n            { chunk: handlingChunk },\n            redispatchMetadata,\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 (preempted && finalChunk != null) {\n      const responseMetadata = {\n        ...finalChunk.response_metadata,\n        preempted: true,\n      };\n      finalChunk.response_metadata = responseMetadata;\n      finalChunk.lc_kwargs.response_metadata = responseMetadata;\n      await endSealedModelRun(\n        context,\n        finalChunk,\n        messagesForProvider,\n        sealedRunId,\n        config,\n        model\n      );\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  streamLimitState,\n  overflowContext,\n  prepareProviderMessages,\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  /** Accounting-lease owner forwarded to each fallback attempt (see\n   * `AttemptInvokeParams.streamLimitState`). */\n  streamLimitState?: StreamLimitState;\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  /**\n   * Optional final payload guard used by Graph. It receives the initialized,\n   * tool-bound fallback model so Responses-vs-Chat projection is exact before\n   * the fallback request is measured and sent.\n   */\n  prepareProviderMessages?: (input: {\n    model: t.ChatModel;\n    messages: BaseMessage[];\n    provider: Providers;\n    clientOptions?: t.ClientOptions;\n    maxContextTokens?: number;\n    config?: RunnableConfig;\n  }) => BaseMessage[] | Promise<BaseMessage[]>;\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 fallbackMessages =\n        (await prepareProviderMessages?.({\n          model: fbModel as t.ChatModel,\n          messages,\n          provider: fb.provider,\n          clientOptions: fb.clientOptions,\n          maxContextTokens: fb.maxContextTokens,\n          config: fbConfig,\n        })) ?? messages;\n      /** A sibling can trip the breaker while the preparation above is\n       * awaited — and the catch below only sees attempts that THROW, so a\n       * provider that ignores an aborted signal and succeeds would resolve\n       * a run that must reject. Check before every fallback invocation. */\n      if (\n        config?.signal?.aborted === true &&\n        config.signal.reason instanceof StreamLimitExceededError\n      ) {\n        throw config.signal.reason;\n      }\n      const result = await attemptInvoke(\n        {\n          model: fbModel as t.ChatModel,\n          messages: fallbackMessages,\n          provider: fb.provider,\n          context,\n          onChunk,\n          streamLimitState,\n        },\n        fbConfig\n      );\n      return result;\n    } catch (e) {\n      /**\n       * A tripped stream circuit breaker is a deliberate abort, not a\n       * provider failure. Continuing would try the remaining fallbacks and a\n       * succeeding one would resolve a run that must reject.\n       */\n      if (e instanceof StreamLimitExceededError) {\n        throw e;\n      }\n      /** A parallel sibling's trip aborts this branch's composed signal, and\n       * a provider can surface that as a generic abort error; advancing to\n       * the next fallback would start new provider work after the safety\n       * abort. Rethrow the breaker's own reason instead. */\n      if (\n        config?.signal?.aborted === true &&\n        config.signal.reason instanceof StreamLimitExceededError\n      ) {\n        throw config.signal.reason;\n      }\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":";;;;;;;;;;;;;;;;;;;;;;;;AA2GA,IAAI,wBAAwB;AAE5B,SAAgB,0BACd,OACA,UACA,aACS;CACT,IAAI,CAACA,YAAAA,aAAa,QAAQ,GACxB,OAAO;CAET,IAAI,YAAqB;CACzB,IAAI,uBAAuB;CAC3B,MAAM,uBAAO,IAAI,IAAY;CAC7B,KAAK,IAAI,QAAQ,GAAG,QAAQ,IAAI,SAAS;EACvC,IAAI,aAAa,QAAQ,OAAO,cAAc,UAC5C,OAAO;EAET,IAAI,KAAK,IAAI,SAAS,GACpB,OAAO;EAET,KAAK,IAAI,SAAS;EAClB,MAAM,WAAW;EAOjB,IAAI;GACF,IACE,SAAS,kBAAkB,QAC3B,OAAO,SAAS,mBAAmB,YACnC,CAAC,MAAM,QAAQ,SAAS,cAAc,KACtC,wBAAwB,QACxB,OAAO,yBAAyB,YAChC,CAAC,MAAM,QAAQ,oBAAoB,GAEnC,uBAAuB;IACrB,GAAI,SAAS;IACb,GAAI;GACN;QACK,IAAI,wBAAwB,MACjC,uBAAuB,SAAS;GAElC,IACE,SAAS,mBAAmB,oBAAoB,MAAM,QACtD,SAAS,mBAAmB,KAAA,CAAS,MAAM,MAE3C,OAAO;EAEX,QAAQ,CAER;EACA,IACE,OAAO,SAAS,aAAa,SAAS,YACtC,SAAS,YAAY,KAAK,SAAS,WAAW,GAE9C,OAAO;EAET,IAAI,SAAS,QAAQ,QAAQ,OAAO,SAAS,SAAS,UAAU;GAC9D,YAAY,SAAS;GACrB;EACF;EACA,IAAI,SAAS,SAAS,QAAQ,OAAO,SAAS,UAAU,UAAU;GAChE,YAAY,SAAS;GACrB;EACF;EACA,OAAO;CACT;CACA,OAAO;AACT;;;;;;AAOA,SAAgB,2BAA2B,EACzC,OACA,UACA,UACA,oBACA,eAOgB;CAChB,MAAM,wBAAwB,0BAC5B,OACA,UACA,WACF;CACA,MAAM,wBAAwBC,aAAAA,oCAC5B,UACA,wBAAwB,WAAW,YACnC,kBACF;CACA,IAAI,uBACF,OAAOC,aAAAA,yCACLC,cAAAA,2BACEC,cAAAA,yBAAyB,qBAAqB,CAChD,GACA,kBACF;CAEF,IAAI,aAAA,cACF,OAAOC,aAAAA,iCACLC,aAAAA,oCACEF,cAAAA,yBAAyB,qBAAqB,GAC9C,kBACF,CACF;CAEF,IAAIJ,YAAAA,aAAa,QAAQ,GACvB,OAAOK,aAAAA,iCACLE,aAAAA,oCACEJ,cAAAA,2BACEC,cAAAA,yBAAyB,qBAAqB,CAChD,GACA,kBACF,CACF;CAEF,IAAI,aAAA,aACF,OAAOC,aAAAA,iCACLG,aAAAA,mCACEJ,cAAAA,yBAAyB,qBAAqB,GAC9C,kBACF,CACF;CAEF,IAAI,aAAA,WACF,OAAOD,cAAAA,2BACLE,aAAAA,iCACEI,aAAAA,wCACE,uBACA,kBACF,CACF,CACF;CAEF,OAAOJ,aAAAA,iCACLK,aAAAA,mCACEF,aAAAA,mCACEL,cAAAA,2BACEC,cAAAA,yBAAyB,qBAAqB,CAChD,GACA,kBACF,GACA,kBACF,CACF;AACF;;;;;;;;;;;AAYA,SAAS,sCACP,SAC4B;CAC5B,MAAM,UAAU,SAAS,iBAAiB,WAAA,sBAE1C;CACA,OAAOO,eAAAA,0BAA0B,OAAO,IAAI,UAAU,KAAA;AACxD;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,IAAIC,yBAAAA,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,IAAIA,yBAAAA,eACT,OAAO,OAAO,CAAC,GAAG,MAAM;EACtB;EACA,mBAAmB,uBAAuB,KAAK,iBAAiB;CAClE,CAAC,CACH;AACF;;;;;AAMA,SAAS,kBACP,SACA,UACA,UACoB;CACpB,IAAI;EACF,MAAM,UAAU,SAAS,gBAAgB,QAAQ,CAAC,CAAC;EACnD,IAAI,WAAW,MACb;EAEF,IAAI,QAAQ;EACZ,KAAK,MAAM,WAAW,UACpB,SAAS,QAAQ,OAAO;EAE1B,OAAO;CACT,QAAQ;EACN;CACF;AACF;;;;;;;;;;;;AAaA,SAAS,0BACP,SACA,UACQ;CACR,IAAI;EACF,MAAM,eAAe,SAAS,gBAAgB,QAAQ;EACtD,OACE,cAAc,+BACd,cAAc,qBACd;CAEJ,QAAQ;EACN,OAAO;CACT;AACF;;;;;;;;;;;;;;;;;;;;;;AAuBA,SAAS,sBACP,SACA,OACA,QACA,UACM;CACN,IAAI,MAAM,kBAAkB,MAC1B;CAEF,MAAM,eAAe,kBAAkB,SAAS,UAAU,CAAC,KAAK,CAAC;CACjE,IAAI,gBAAgB,MAClB;CAEF,MAAM,eACH,kBAAkB,SAAS,UAAU,MAAM,KAAK,KACjD,0BAA0B,SAAS,QAAQ;CAC7C,MAAM,gBAAgB;EACpB,cAAc;EACd,eAAe;EACf,cAAc,cAAc;CAC9B;CACA,MAAM,iBAAiB;CACvB,MAAM,UAAU,iBAAiB;CACjC,MAAM,mBAAmB;EACvB,GAAG,MAAM;EACT,iBAAiB;CACnB;CACA,MAAM,oBAAoB;CAC1B,MAAM,UAAU,oBAAoB;AACtC;AAEA,SAAS,eAAe,OAA+B;CACrD,IAAI,OAAO,MAAM,YAAY,UAC3B,OAAO,MAAM;CAEf,IAAI,OAAO;CACX,KAAK,MAAM,SAAS,MAAM,SACxB,IAAI,MAAM,SAAA,QAA4B;EACpC,MAAM,QAAQ,MAAA;EACd,IAAI,OAAO,UAAU,UACnB,QAAQ;CAEZ;CAEF,OAAO;AACT;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAiCA,SAAS,4BAA4B,OAA6B;CAChE,MAAM,UAAuB,CAAC;CAC9B,MAAM,uBAAO,IAAI,IAAa;CAC9B,IAAI,UAAmB;CACvB,OAAO,WAAW,QAAQ,OAAO,YAAY,YAAY,CAAC,KAAK,IAAI,OAAO,GAAG;EAC3E,KAAK,IAAI,OAAO;EAChB,MAAM,UAAU;EAOhB,IAAI,QAAQ,aAAa,MACvB,QAAQ,KAAK,QAAQ,SAAS;EAEhC,IAAI,QAAQ,QAAQ,aAAa,MAC/B,QAAQ,KAAK,QAAQ,OAAO,SAAS;EAEvC,UACE,QAAQ,SACR,QAAQ,SACP,MAAM,QAAQ,QAAQ,KAAK,IACxB,QAAQ,MAAM,QAAQ,MAAM,SAAS,KACrC,KAAA;CACR;CACA,OAAO;AACT;;;;;;;AAQA,SAAgB,sBAAsB,OAAoC;CACxE,MAAM,uBAAO,IAAI,IAAa;CAC9B,IAAI,UAAmB;CACvB,OAAO,WAAW,QAAQ,OAAO,YAAY,YAAY,CAAC,KAAK,IAAI,OAAO,GAAG;EAC3E,KAAK,IAAI,OAAO;EAChB,MAAM,UAAU;EAMhB,IAAI,OAAO,QAAQ,UAAU,YAAY,QAAQ,UAAU,IACzD,OAAO,QAAQ;EAEjB,UACE,QAAQ,SACR,QAAQ,SACP,MAAM,QAAQ,QAAQ,KAAK,IACxB,QAAQ,MAAM,QAAQ,MAAM,SAAS,KACrC,KAAA;CACR;AAEF;AAEA,eAAe,kBACb,SACA,OACA,QACA,UACA,QACA,OACe;CACf,MAAM,WAAW,QAAQ;CACzB,sBAAsB,SAAS,OAAO,QAAQ,QAAQ;CACtD,IAAI,YAAY,MACd,IAAI;EACF,IAAI,kBAAkB,OAAA,GAAA,0BAAA,4BAAA,CAAkC,MAAM;;;;;;;;;;;;EAY9D,KAAK,MAAM,UAAU,4BAA4B,KAAK,GACpD,kBACEC,kCAAAA,gBAAgB,UAAU,mBAAmB,KAAA,GAAW,MAAM,KAC9D;EAEJ,IAAI,mBAAmB,MAAM;GAC3B,MAAM,aAAa,IAAIC,kCAAAA,yBACrB,UACA,gBAAgB,UAChB,gBAAgB,qBAChB,gBAAgB,MAChB,gBAAgB,iBAChB,gBAAgB,UAChB,gBAAgB,qBAChB,gBAAgB,eAAe,CACjC;GACA,MAAM,aAA6B;IACjC,MAAM,eAAe,KAAK;IAC1B,SAAS;GACX;GACA,MAAM,WAAW,aAAa;IAC5B,aAAa,CAAC,CAAC,UAAU,CAAC;IAC1B,WAAW,CAAC;GACd,CAAC;GACD;EACF;CACF,SAAS,GAAG;;;;;EAMV,QAAQ,KACN,gGACA,aAAa,QAAQ,EAAE,UAAU,CACnC;CACF;CAEF,MAAMC,eAAAA,wBAAAA,qBAEJ,EAAE,QAAQ,MAAM,GAChB,MACF;AACF;AAEA,SAAS,kBAAkB,EACzB,SACA,MACA,YAKiB;CACjB,IAAI,WAAW,MACb,OAAO;CAET,QAAA,GAAA,6BAAA,OAAA,CACE,SACA,4CAA4C;EAAE;EAAS;EAAM;CAAS,CAAC,CACzE;AACF;;;;;;;;AA8BA,eAAsB,cACpB,QACA,QACoC;CACpC,MAAM,gBAAgC;EACpC,GAAG;EACH,UAAU;GACR,GAAI,QAAQ,YAAY,CAAC;2BACK,OAAO;;;;;;;;;;IAUpCC,qBAAAA,2BAA2B,EAAE;EAChC;CACF;CACA,MAAM,iBAAiB,OAAO,WAAW,OAAO;;;;CAIhD,MAAM,cACJ,kBAAkB,QAAQC,qBAAAA,6BAA6B,cAAc,IACjE,iBACA,KAAA;CACN,MAAM,gBACJ,eAAe,OACXC,qBAAAA,qBACA,cAAc,QAChB,IACE,KAAA;CACN,IAAI,eAAe,QAAQ,iBAAiB,MAC1C,qBAAA,oCAAoC,aAAa,aAAa;CAEhE,IAAI;EACF,OAAO,MAAM,kBAAkB,QAAQ,aAAa;CACtD,UAAU;EACR,IAAI,eAAe,QAAQ,iBAAiB,MAC1C,qBAAA,6BAA6B,aAAa,aAAa;CAE3D;AACF;AAEA,eAAe,kBACb,EACE,OACA,UACA,UACA,SACA,WAEF,QACoC;;;;;;;;CAQpC,MAAM,qBAAqB,2BAA2B;EACpD;EACA;EACA;EACA,aAAa;CACf,CAAC;CACD,MAAM,WAAW,SAAS,8BAA8B;CACxD,MAAM,QAAQ,OAAO,cAAc;CACnC,MAAM,YAAYC,6BAAAA,uBAAuB,oBAAoB,UAAU,KAAK;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;CA8B5E,MAAM,gBACJ,SAGC;CACH,MAAM,OAAOC,YAAAA,gBAAgB,UAAU,EACrC,OAAO,sBAAsB,KAAK,EACpC,CAAC,IACGC,mBAAAA,4BACA,WACA,iBAAiB,OACb,KAAA,KACC,YAAqB,cAAc,KAAK,SAAS,OAAO,CAC/D,IACEC,mBAAAA,4BAA4B,SAAS;CACzC,MAAM,sBAAsBC,oBAAAA,2BAA2B,IAAI,QAAQ,IAC/DC,oBAAAA,0BAA0B,IAAI,IAC9B;;;;;;;;;;CAWJ,IAAI,MAAM,QAAQ;;;;;;;;EAQhB,IAAI;EACJ,MAAM,eACJ,SAAS,cAAc,OACnB,SACA;GACA,GAAG;GACH,WAAWC,kBAAAA,gBAAgB,OAAO,WAAW,CAC3C,EACE,uBACE,MACA,WACA,UACS;IACT,gBAAgB;GAClB,EACF,CACF,CAAC;EACH;EACJ,MAAM,SAAS,MAAM,MAAM,OAAO,qBAAqB,YAAY;EACnE,IAAI;EACJ,IAAI,YAAY;EAChB,MAAM,0BACJ,sCAAsC,OAAO;;;;;;EAM/C,MAAM,8BAAoC;GACxC,MAAM,SAAS,OAAO;GACtB,IACE,QAAQ,YAAY,QACpB,OAAO,kBAAkBC,qBAAAA,0BAEzB,MAAM,OAAO;EAEjB;EAEA,IAAI,SAAS;GACX,MAAM,kBAAkB,OAAO;GAG/B,WAAW,MAAM,SAAS,QAAQ;IAChC,sBAAsB;;;;;;IAMtB,IAAI,WAAW,MACb,qBAAA,gCAAgC;KAC9B,OAAO;KACP,UAAU;KACV;IACF,CAAC;IAEH,MAAM,QAAQ,OAAO,eAAe;IACpC,aAAa,kBAAkB;KAC7B,SAAS;KACT,MAAM;KACN;IACF,CAAC;GACH;EACF,OAAO,IAAI,2BAA2B,MAAM;GAC1C,MAAM,WAAW,OAAO;GACxB,MAAM,gBAAgB,IAAIC,eAAAA,uBAAuB;GACjD,WAAW,MAAM,SAAS,QAAQ;IAChC,sBAAsB;IACtB,MAAM,gBAAgB,uBAAuB;KAC3C,SAAS;KACT,MAAM;KACN;IACF,CAAC;IACD,IAAI,iBAAiB,MACnB,MAAM,cAAc,OAAA,wBAElB,EAAE,OAAO,cAAc,GACvB,UACA,OACF;SACK,IAAI,WAAW;;;;;;;;;;;IAWpB,qBAAA,gCAAgC;KAC9B,OAAO;KACP;KACA;KACA,MAAM;IACR,CAAC;IAEH,aAAa,kBAAkB;KAC7B,SAAS;KACT,MAAM;KACN;IACF,CAAC;;;;;;;;;;;;;IAaD,IACE,SAAS,oBAAoB,MAAM,QACnCC,gBAAAA,eAAe,UAAU,KACzB,QAAQ,iBAAiB,GACzB;KACA,YAAY;KACZ;IACF;GACF;EACF,OAAO;GACL,MAAM,WAAW,OAAO;;;;;;;;GAQxB,IAAI;GACJ,WAAW,MAAM,SAAS,QAAQ;IAChC,sBAAsB;;;;;;;;IAQtB,IAAI,WAAW,MACb,qBAAA,gCAAgC;KAAE,OAAO;KAAS;KAAU;IAAM,CAAC;IAErE,MAAM,gBAAgB,uBAAuB;KAC3C,SAAS;KACT,MAAM;KACN;IACF,CAAC;IACD,IAAI,iBAAiB,QAAQ,kBAAkB,OAAO;KACpD,uBAAuB;MACrB,GAAI,YAAY,CAAC;OAChBC,qBAAAA,8BAA8B;KACjC;KACA,MAAM,wBAAwB,OAAA,wBAE5B,EAAE,OAAO,cAAc,GACvB,oBACA,OACF;IACF;IACA,aAAa,kBAAkB;KAC7B,SAAS;KACT,MAAM;KACN;IACF,CAAC;GACH;EACF;EAEA,IAAIC,kBAAAA,0BAA0B,IAAI,QAAQ,GACxC,aAAaC,aAAAA,sBAAsB,UAAU,UAAU;EAGzD,IAAI,aAAa,cAAc,MAAM;GACnC,MAAM,mBAAmB;IACvB,GAAG,WAAW;IACd,WAAW;GACb;GACA,WAAW,oBAAoB;GAC/B,WAAW,UAAU,oBAAoB;GACzC,MAAM,kBACJ,SACA,YACA,qBACA,aACA,QACA,KACF;EACF;EAEA,KAAK,YAAY,YAAY,UAAU,KAAK,GAC1C,WAAY,aAAa,WAAY,YAAY,QAC9C,cAAwB,CAAC,CAAC,UAAU,IACvC;EAGF,mBAAA,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,mBAAA,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,kBACA,iBACA,2BAgCiD;CACjD,MAAM,cACJ,OACA,kBAAkB,oBACNC,eAAAA,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,UAAUC,aAAAA,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;EACJ,MAAM,mBACH,MAAM,0BAA0B;GAC/B,OAAO;GACP;GACA,UAAU,GAAG;GACb,eAAe,GAAG;GAClB,kBAAkB,GAAG;GACrB,QAAQ;EACV,CAAC,KAAM;;;;;EAKT,IACE,QAAQ,QAAQ,YAAY,QAC5B,OAAO,OAAO,kBAAkBP,qBAAAA,0BAEhC,MAAM,OAAO,OAAO;EAatB,OAAO,MAXc,cACnB;GACE,OAAO;GACP,UAAU;GACV,UAAU,GAAG;GACb;GACA;GACA;EACF,GACA,QACF;CAEF,SAAS,GAAG;;;;;;EAMV,IAAI,aAAaA,qBAAAA,0BACf,MAAM;;;;;EAMR,IACE,QAAQ,QAAQ,YAAY,QAC5B,OAAO,OAAO,kBAAkBA,qBAAAA,0BAEhC,MAAM,OAAO,OAAO;EAEtB,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"}

package/dist/cjs/llm/openai/index.cjs

@@ -1,7 +1,7 @@
 const require_cache = require("../../messages/cache.cjs");
 const require_core = require("../../messages/core.cjs");
-const require_intentArg = require("../../tools/intentArg.cjs");
 const require_streamedToolCallSeals = require("../../tools/streamedToolCallSeals.cjs");
+const require_intentArg = require("../../tools/intentArg.cjs");
 const require_index = require("./utils/index.cjs");
 const require_streamMetadata = require("./streamMetadata.cjs");
 let _langchain_openai = require("@langchain/openai");
@@ -123,6 +123,9 @@ function addChatCacheBreakpoints(messages) {
 function isResponseMessage(item) {
 	return item.type === "message";
 }
+function isResponseInputRole(role) {
+	return role === "system" || role === "developer" || role === "user";
+}
 /** Only `input_text`/`input_image`/`input_file` accept a Responses breakpoint;
 *  `output_text`/`refusal` (replayed assistant blocks) are rejected with a 400. */
 function isCacheableResponsePart(part) {
@@ -165,7 +168,7 @@ function addResponseCacheBreakpoints(input) {
 		/** Only input roles take a Responses breakpoint. Assistant/tool turns
 		*  carry output content (string or output_text) that the API rejects
 		*  under an input marker, so they're never eligible. */
-		if (item.role !== "system" && item.role !== "developer" && item.role !== "user") return false;
+		if (!isResponseInputRole(item.role)) return false;
 		const content = item.content;
 		return typeof content === "string" ? content.length > 0 : Array.isArray(content) && content.some((part) => isCacheableResponsePart(part));
 	})));
@@ -178,7 +181,7 @@ function shouldIncludeEncryptedReasoning(model, params) {
 }
 function getCacheWriteTokens(message) {
 	const responseMetadata = message.response_metadata;
-	const reported = responseMetadata.usage?.input_tokens_details.cache_write_tokens;
+	const reported = responseMetadata.usage?.input_tokens_details?.cache_write_tokens;
 	if (reported != null) return reported;
 	const serialized = responseMetadata.metadata?.[CACHE_WRITE_METADATA_KEY];
 	if (serialized == null) return;
@@ -202,7 +205,7 @@ function attachCacheWriteUsage(message) {
 	};
 }
 function attachCacheWriteMetadata(response) {
-	const cacheWriteTokens = response.usage?.input_tokens_details.cache_write_tokens;
+	const cacheWriteTokens = response.usage?.input_tokens_details?.cache_write_tokens;
 	if (cacheWriteTokens == null) return response;
 	return {
 		...response,
@@ -215,6 +218,165 @@ function attachCacheWriteMetadata(response) {
 function isResponsesStream(result) {
 	return Symbol.asyncIterator in result;
 }
+const RESPONSES_REPLAY_OUTPUT_ITEM_TYPES = new Set([
+	"local_shell_call_output",
+	"shell_call_output",
+	"apply_patch_call_output",
+	"program_output"
+]);
+function isResponsesReplayOutputItem(item) {
+	return typeof item === "object" && item != null && "type" in item && typeof item.type === "string" && RESPONSES_REPLAY_OUTPUT_ITEM_TYPES.has(item.type);
+}
+/**
+* LangChain's Responses converter places the authoritative terminal output in
+* response_metadata.output. Its chunk merge has no way to delete provisional
+* tool_outputs or replay-position sidecars, so remove those preemption-only
+* captures once that terminal output arrives. An interrupted stream has no
+* terminal chunk and keeps the captures for replay.
+*/
+var ResponsesReplayAIMessageChunk = class extends _langchain_core_messages.AIMessageChunk {
+	get lc_id() {
+		return [...this.lc_namespace, _langchain_core_messages.AIMessageChunk.lc_name()];
+	}
+	concat(chunk) {
+		const combined = super.concat(chunk);
+		if (!Array.isArray(chunk.response_metadata.output)) return combined;
+		delete combined.additional_kwargs[require_core.OPENAI_RESPONSES_REPLAY_POSITIONS_KEY];
+		const toolOutputs = combined.additional_kwargs.tool_outputs;
+		if (!Array.isArray(toolOutputs)) return combined;
+		const retainedToolOutputs = toolOutputs.filter((item) => !isResponsesReplayOutputItem(item));
+		if (retainedToolOutputs.length === toolOutputs.length) return combined;
+		if (retainedToolOutputs.length > 0) combined.additional_kwargs.tool_outputs = retainedToolOutputs;
+		else delete combined.additional_kwargs.tool_outputs;
+		return combined;
+	}
+};
+function makeResponsesReplayAggregationSafe(chunk) {
+	if (!_langchain_core_messages.AIMessageChunk.isInstance(chunk.message)) return chunk;
+	const message = chunk.message;
+	chunk.message = new ResponsesReplayAIMessageChunk({
+		id: message.id,
+		name: message.name,
+		content: message.content,
+		additional_kwargs: message.additional_kwargs,
+		response_metadata: message.response_metadata,
+		tool_calls: message.tool_calls,
+		invalid_tool_calls: message.invalid_tool_calls,
+		tool_call_chunks: message.tool_call_chunks,
+		usage_metadata: message.usage_metadata
+	});
+	return chunk;
+}
+function remapResponsesTextBlockIndex(chunk, event, textBlockIndices) {
+	const position = iife(() => {
+		if (event.type === "response.output_text.delta" || event.type === "response.output_text.annotation.added") return {
+			contentIndex: event.content_index,
+			outputIndex: event.output_index
+		};
+		if (event.type === "response.output_item.added" && event.item.type === "message") return {
+			contentIndex: 0,
+			outputIndex: event.output_index
+		};
+	});
+	if (position == null || !Array.isArray(chunk.message.content)) return;
+	const key = `${position.outputIndex}:${position.contentIndex}`;
+	let blockIndex = textBlockIndices.get(key);
+	if (blockIndex == null) {
+		blockIndex = textBlockIndices.size;
+		textBlockIndices.set(key, blockIndex);
+	}
+	const content = chunk.message.content.map((block) => typeof block === "object" && block.type === "text" ? {
+		...block,
+		index: blockIndex
+	} : block);
+	chunk.message.content = content;
+	chunk.message.lc_kwargs.content = content;
+}
+function convertDroppedResponsesReplayOutput(event) {
+	if (event.type !== "response.output_item.done") return null;
+	if (event.item.type === "reasoning") return new _langchain_core_outputs.ChatGenerationChunk({
+		text: "",
+		message: new _langchain_core_messages.AIMessageChunk({
+			content: [],
+			additional_kwargs: { reasoning: {
+				status: event.item.status,
+				...typeof event.item.encrypted_content === "string" ? { encrypted_content: event.item.encrypted_content } : {}
+			} },
+			response_metadata: { model_provider: "openai" }
+		})
+	});
+	if (!RESPONSES_REPLAY_OUTPUT_ITEM_TYPES.has(event.item.type)) return null;
+	return new _langchain_core_outputs.ChatGenerationChunk({
+		text: "",
+		message: new _langchain_core_messages.AIMessageChunk({
+			content: [],
+			additional_kwargs: { tool_outputs: [event.item] },
+			response_metadata: { model_provider: "openai" }
+		})
+	});
+}
+function attachResponsesReplayPosition(chunk, event, seenPositions) {
+	let position;
+	if (event.type === "response.output_text.delta" && event.delta.length > 0) position = {
+		contentIndex: event.content_index,
+		itemId: event.item_id,
+		kind: "text",
+		outputIndex: event.output_index
+	};
+	else if (event.type === "response.output_item.added" && event.item.type === "message" && typeof event.item.id === "string" && event.item.id.length > 0) position = {
+		itemId: event.item.id,
+		kind: "message",
+		outputIndex: event.output_index
+	};
+	else if (event.type === "response.output_item.added" && event.item.type === "reasoning" && typeof event.item.id === "string" && event.item.id.length > 0) position = {
+		itemId: event.item.id,
+		kind: "reasoning",
+		outputIndex: event.output_index
+	};
+	else if (event.type === "response.output_item.done" && (RESPONSES_REPLAY_OUTPUT_ITEM_TYPES.has(event.item.type) || Array.isArray(chunk.message.additional_kwargs.tool_outputs))) {
+		let itemId;
+		if (typeof event.item.id === "string" && event.item.id.length > 0) itemId = event.item.id;
+		else if ("call_id" in event.item && typeof event.item.call_id === "string" && event.item.call_id.length > 0) itemId = event.item.call_id;
+		if (itemId != null) position = {
+			itemId,
+			kind: "output",
+			outputIndex: event.output_index
+		};
+	}
+	if (position == null) return;
+	const positionKey = `${position.kind}:${position.itemId}:${position.outputIndex}:${position.contentIndex ?? ""}`;
+	if (seenPositions.has(positionKey)) return;
+	seenPositions.add(positionKey);
+	const existing = chunk.message.additional_kwargs[require_core.OPENAI_RESPONSES_REPLAY_POSITIONS_KEY];
+	const additionalKwargs = {
+		...chunk.message.additional_kwargs,
+		[require_core.OPENAI_RESPONSES_REPLAY_POSITIONS_KEY]: [...Array.isArray(existing) ? existing : [], position]
+	};
+	chunk.message.additional_kwargs = additionalKwargs;
+	chunk.message.lc_kwargs.additional_kwargs = additionalKwargs;
+}
+async function* convertLibreChatResponsesStream(stream, options, runManager) {
+	const seenReplayPositions = /* @__PURE__ */ new Set();
+	const responsesTextBlockIndices = /* @__PURE__ */ new Map();
+	try {
+		for await (const event of stream) {
+			options.signal?.throwIfAborted();
+			const convertedChunk = (0, _langchain_openai.convertResponsesDeltaToChatGenerationChunk)(event) ?? convertDroppedResponsesReplayOutput(event);
+			if (convertedChunk == null) continue;
+			const chunk = makeResponsesReplayAggregationSafe(convertedChunk);
+			remapResponsesTextBlockIndex(chunk, event, responsesTextBlockIndices);
+			attachResponsesReplayPosition(chunk, event, seenReplayPositions);
+			attachCacheWriteUsage(chunk.message);
+			await runManager?.handleLLMNewToken(chunk.text || "", {
+				prompt: options.promptIndex ?? 0,
+				completion: 0
+			}, void 0, void 0, void 0, { chunk });
+			yield chunk;
+		}
+	} catch (e) {
+		throw (0, _langchain_openai.wrapOpenAIClientError)(e);
+	}
+}
 function createUsageMetadata(usage) {
 	const usageMetadata = {
 		input_tokens: usage?.prompt_tokens ?? 0,
@@ -817,10 +979,16 @@ var LibreChatOpenAIResponses = class extends _langchain_openai.ChatOpenAIRespons
 		return result;
 	}
 	async *_streamResponseChunks(messages, options, runManager) {
-		for await (const chunk of super._streamResponseChunks(projectOpenAIResponsesProviderMessages(messages), options, runManager)) {
-			attachCacheWriteUsage(chunk.message);
-			yield chunk;
-		}
+		const projectedMessages = projectOpenAIResponsesProviderMessages(messages);
+		yield* convertLibreChatResponsesStream(await this.completionWithRetry({
+			...this.invocationParams(options),
+			input: (0, _langchain_openai.convertMessagesToResponsesInput)({
+				messages: projectedMessages,
+				zdrEnabled: this.zdrEnabled ?? false,
+				model: this.model
+			}),
+			stream: true
+		}, options), options, runManager);
 	}
 	async *_streamChatModelEvents(messages, options, runManager) {
 		yield* super._streamChatModelEvents(projectOpenAIResponsesProviderMessages(messages), options, runManager);
@@ -935,15 +1103,24 @@ var LibreChatAzureOpenAIResponses = class extends _langchain_openai.AzureChatOpe
 		return isResponsesStream(result) ? result : attachCacheWriteMetadata(result);
 	}
 	async _generate(messages, options, runManager) {
-		const result = await super._generate(messages, options, runManager);
+		const result = await super._generate(projectOpenAIResponsesProviderMessages(messages), options, runManager);
 		for (const generation of result.generations) attachCacheWriteUsage(generation.message);
 		return result;
 	}
 	async *_streamResponseChunks(messages, options, runManager) {
-		for await (const chunk of super._streamResponseChunks(messages, options, runManager)) {
-			attachCacheWriteUsage(chunk.message);
-			yield chunk;
-		}
+		const projectedMessages = projectOpenAIResponsesProviderMessages(messages);
+		yield* convertLibreChatResponsesStream(await this.completionWithRetry({
+			...this.invocationParams(options),
+			input: (0, _langchain_openai.convertMessagesToResponsesInput)({
+				messages: projectedMessages,
+				zdrEnabled: this.zdrEnabled ?? false,
+				model: this.model
+			}),
+			stream: true
+		}, options), options, runManager);
+	}
+	async *_streamChatModelEvents(messages, options, runManager) {
+		yield* super._streamChatModelEvents(projectOpenAIResponsesProviderMessages(messages), options, runManager);
 	}
 	_getReasoningParams(options) {
 		return getGatedReasoningParams(this.model, this.reasoning, options);