@librechat/agents 3.3.2 → 3.3.4

package/dist/cjs/llm/invoke.cjs

@@ -1,7 +1,9 @@
 require("../common/enum.cjs");
 require("../common/index.cjs");
+const require_cache = require("../messages/cache.cjs");
 const require_core = require("../messages/core.cjs");
 require("../messages/index.cjs");
+const require_llm = require("../utils/llm.cjs");
 const require_toolOutputReferences = require("../tools/toolOutputReferences.cjs");
 const require_stream = require("../stream.cjs");
 const require_errors = require("../utils/errors.cjs");
@@ -11,6 +13,51 @@ const require_init = require("./init.cjs");
 let _langchain_core_messages = require("@langchain/core/messages");
 let _langchain_core_utils_stream = require("@langchain/core/utils/stream");
 //#region src/llm/invoke.ts
+function usesNativeOpenAIResponses(model, provider, callOptions) {
+	if (!require_llm.isOpenAILike(provider)) return false;
+	let candidate = model;
+	let effectiveCallOptions = callOptions;
+	const seen = /* @__PURE__ */ new Set();
+	for (let depth = 0; depth < 20; depth++) {
+		if (candidate == null || typeof candidate !== "object") return false;
+		if (seen.has(candidate)) return false;
+		seen.add(candidate);
+		const runnable = candidate;
+		try {
+			if (runnable.defaultOptions != null && typeof runnable.defaultOptions === "object" && !Array.isArray(runnable.defaultOptions) && effectiveCallOptions != null && typeof effectiveCallOptions === "object" && !Array.isArray(effectiveCallOptions)) effectiveCallOptions = {
+				...runnable.defaultOptions,
+				...effectiveCallOptions
+			};
+			else if (effectiveCallOptions == null) effectiveCallOptions = runnable.defaultOptions;
+			if (runnable._useResponsesApi?.(effectiveCallOptions) === true || runnable._useResponsesApi?.(void 0) === true) return true;
+		} catch {}
+		if (typeof runnable.constructor?.name === "string" && runnable.constructor.name.includes("Responses")) return true;
+		if (runnable.last != null && typeof runnable.last === "object") {
+			candidate = runnable.last;
+			continue;
+		}
+		if (runnable.bound != null && typeof runnable.bound === "object") {
+			candidate = runnable.bound;
+			continue;
+		}
+		return false;
+	}
+	return false;
+}
+/**
+* Produces the exact provider-facing message representation before a model
+* adapter serializes it. This is shared by invocation and Graph's final budget
+* guard so structured tool output cannot grow after the payload was measured.
+*/
+function projectMessagesForProvider({ model, messages, provider, maxToolResultChars, callOptions }) {
+	const providerInputMessages = require_core.projectToolStreamContentForProvider(messages);
+	if (usesNativeOpenAIResponses(model, provider, callOptions)) return require_core.projectOpenAIResponsesToolMessageContent(require_cache.stripAnthropicCacheControl(require_cache.stripBedrockCacheControl(providerInputMessages)), maxToolResultChars);
+	if (provider === "openrouter") return require_core.projectComputerCallOutputsToText(require_core.projectOpenRouterToolMessageContent(require_cache.stripBedrockCacheControl(providerInputMessages), maxToolResultChars));
+	if (require_llm.isOpenAILike(provider)) return require_core.projectComputerCallOutputsToText(require_core.projectOpenAIChatToolMessageContent(require_cache.stripAnthropicCacheControl(require_cache.stripBedrockCacheControl(providerInputMessages)), maxToolResultChars));
+	if (provider === "anthropic") return require_core.projectComputerCallOutputsToText(require_core.projectSingleTextToolOutputsToText(require_cache.stripBedrockCacheControl(providerInputMessages), maxToolResultChars));
+	if (provider === "bedrock") return require_cache.stripAnthropicCacheControl(require_core.projectComputerCallOutputsToText(require_core.projectCacheControlledToolOutputsToText(providerInputMessages, maxToolResultChars)));
+	return require_core.projectComputerCallOutputsToText(require_core.projectStructuredToolOutputsToText(require_core.projectSingleTextToolOutputsToText(require_cache.stripAnthropicCacheControl(require_cache.stripBedrockCacheControl(providerInputMessages)), maxToolResultChars), maxToolResultChars));
+}
 function getRegisteredDefaultChatStreamHandler(context) {
 	const handler = context?.handlerRegistry?.getHandler("on_chat_model_stream");
 	return handler instanceof require_stream.ChatModelStreamHandler ? handler : void 0;
@@ -74,9 +121,15 @@ async function attemptInvoke({ model, messages, provider, context, onChunk }, co
 	* untouched so the graph state never sees `[ref: …]` / `_ref`
 	* payload.
 	*/
+	const invocationMessages = projectMessagesForProvider({
+		model,
+		messages,
+		provider,
+		callOptions: config
+	});
 	const registry = context?.getOrCreateToolOutputRegistry();
 	const runId = config?.configurable?.run_id;
-	const messagesForProvider = require_toolOutputReferences.annotateMessagesForLLM(messages, registry, runId);
+	const messagesForProvider = require_toolOutputReferences.annotateMessagesForLLM(invocationMessages, registry, runId);
 	/**
 	* Stamp the provider that is ACTUALLY serving this invocation onto the
 	* callback metadata. `attemptInvoke` is the single funnel for primary,
@@ -181,7 +234,7 @@ function extractClientOptionsModel(clientOptions) {
 * behind a later unrelated error would surface a dead end instead. Ordinary
 * failures still throw last-error-wins.
 */
-async function tryFallbackProviders({ fallbacks, tools, messages, config, primaryError, context, onChunk, overflowContext }) {
+async function tryFallbackProviders({ fallbacks, tools, messages, config, primaryError, context, onChunk, overflowContext, prepareProviderMessages }) {
 	const isOverflow = (error, contextOverride = overflowContext) => require_errors.getContextOverflowInfo(error, contextOverride) != null;
 	let lastError = primaryError;
 	/**
@@ -215,7 +268,14 @@ async function tryFallbackProviders({ fallbacks, tools, messages, config, primar
 		};
 		return await attemptInvoke({
 			model: fbModel,
-			messages,
+			messages: await prepareProviderMessages?.({
+				model: fbModel,
+				messages,
+				provider: fb.provider,
+				clientOptions: fb.clientOptions,
+				maxContextTokens: fb.maxContextTokens,
+				config: fbConfig
+			}) ?? messages,
 			provider: fb.provider,
 			context,
 			onChunk
@@ -254,6 +314,7 @@ async function tryFallbackProviders({ fallbacks, tools, messages, config, primar
 exports.attemptInvoke = attemptInvoke;
 exports.getFallbackErrorContext = getFallbackErrorContext;
 exports.getFallbackOverflowCandidates = getFallbackOverflowCandidates;
+exports.projectMessagesForProvider = projectMessagesForProvider;
 exports.tryFallbackProviders = tryFallbackProviders;
 
 //# sourceMappingURL=invoke.cjs.map

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

@@ -1 +1 @@
-{"version":3,"file":"invoke.cjs","names":["ChatModelStreamHandler","AIMessageChunk","annotateMessagesForLLM","manualToolStreamProviders","modifyDeltaProperties","getContextOverflowInfo","initializeModel"],"sources":["../../../src/llm/invoke.ts"],"sourcesContent":["import { concat } from '@langchain/core/utils/stream';\nimport { AIMessageChunk } from '@langchain/core/messages';\nimport type { RunnableConfig } from '@langchain/core/runnables';\nimport type { ToolCall } from '@langchain/core/messages/tool';\nimport type { BaseMessage } from '@langchain/core/messages';\nimport type { ToolOutputReferenceRegistry } from '@/tools/toolOutputReferences';\nimport type { ContextOverflowContext } from '@/utils/errors';\nimport type * as t from '@/types';\nimport { annotateMessagesForLLM } from '@/tools/toolOutputReferences';\nimport { assertNotTruncatedToolCall } from '@/llm/truncation';\nimport { Constants, GraphEvents, Providers } from '@/common';\nimport { manualToolStreamProviders } from '@/llm/providers';\nimport { getContextOverflowInfo } from '@/utils/errors';\nimport { modifyDeltaProperties } from '@/messages';\nimport { ChatModelStreamHandler } from '@/stream';\nimport { initializeModel } from '@/llm/init';\n\n/**\n * Context passed to `attemptInvoke`. Matches the subset of Graph that\n * `ChatModelStreamHandler.handle` needs *plus* the explicit\n * `getOrCreateToolOutputRegistry()` accessor that `attemptInvoke`\n * itself calls to pull the run-scoped tool-output registry off the\n * graph and project each relevant ToolMessage into a transient\n * annotated copy before the provider call.\n *\n * The intersection is intentional: `Parameters<...>[3]` resolves\n * indirectly through the stream handler's signature (which returns\n * `StandardGraph` and already exposes the accessor since #117), but\n * stating it explicitly here surfaces the contract at the call site —\n * a developer reading `attemptInvoke` doesn't have to chase the\n * upstream handler's parameter list to discover that\n * `context?.getOrCreateToolOutputRegistry()` is a real thing. Single\n * optional chain only — the method itself is required on the\n * `StandardGraph` branch of the intersection, so the second `?.` is\n * unnecessary at the call site.\n *\n * `NonNullable<...>` strips `undefined` from the upstream parameter\n * type so the intersection doesn't collapse to `never` on the\n * undefined branch; callers express optionality via `context?:\n * InvokeContext` on the function signature instead.\n *\n * Callers without a registry (e.g. summarization) simply pass no\n * `context` and the transform safely no-ops.\n */\nexport type InvokeContext = NonNullable<\n  Parameters<ChatModelStreamHandler['handle']>[3]\n> & {\n  getOrCreateToolOutputRegistry?(): ToolOutputReferenceRegistry | undefined;\n};\n\n/**\n * Per-chunk callback for custom stream processing.\n * When provided, replaces the default `ChatModelStreamHandler`.\n */\nexport type OnChunk = (chunk: AIMessageChunk) => void | Promise<void>;\n\nfunction getRegisteredDefaultChatStreamHandler(\n  context?: InvokeContext\n): ChatModelStreamHandler | undefined {\n  const handler = context?.handlerRegistry?.getHandler(\n    GraphEvents.CHAT_MODEL_STREAM\n  );\n  return handler instanceof ChatModelStreamHandler ? handler : undefined;\n}\n\nfunction hasReasoningDetails(chunk: AIMessageChunk): boolean {\n  const reasoningDetails = chunk.additional_kwargs.reasoning_details;\n  return Array.isArray(reasoningDetails) && reasoningDetails.length > 0;\n}\n\nfunction removeOpenRouterFinalReasoningReplayContent({\n  current,\n  next,\n  provider,\n}: {\n  current?: AIMessageChunk;\n  next: AIMessageChunk;\n  provider: Providers;\n}): AIMessageChunk {\n  const content = getOpenRouterFinalReasoningContent({\n    current,\n    next,\n    provider,\n  });\n  if (content == null || content === next.content) {\n    return next;\n  }\n\n  return new AIMessageChunk(\n    Object.assign({}, next, {\n      content,\n    })\n  );\n}\n\nfunction getOpenRouterFinalReasoningContent({\n  current,\n  next,\n  provider,\n}: {\n  current?: AIMessageChunk;\n  next: AIMessageChunk;\n  provider: Providers;\n}): string | undefined {\n  if (\n    provider !== Providers.OPENROUTER ||\n    current == null ||\n    !hasReasoningDetails(next) ||\n    typeof current.content !== 'string' ||\n    current.content === '' ||\n    typeof next.content !== 'string' ||\n    next.content === ''\n  ) {\n    return undefined;\n  }\n  if (!next.content.startsWith(current.content)) {\n    return next.content;\n  }\n  return next.content.slice(current.content.length);\n}\n\nfunction removeReasoningDetails(\n  additionalKwargs: AIMessageChunk['additional_kwargs']\n): AIMessageChunk['additional_kwargs'] {\n  return Object.fromEntries(\n    Object.entries(additionalKwargs).filter(\n      ([key]) => key !== 'reasoning_details'\n    )\n  );\n}\n\nfunction getStreamHandlingChunk({\n  current,\n  next,\n  provider,\n}: {\n  current?: AIMessageChunk;\n  next: AIMessageChunk;\n  provider: Providers;\n}): AIMessageChunk | undefined {\n  const content = getOpenRouterFinalReasoningContent({\n    current,\n    next,\n    provider,\n  });\n  if (content == null) {\n    return next;\n  }\n  if (content === '') {\n    return undefined;\n  }\n  return new AIMessageChunk(\n    Object.assign({}, next, {\n      content,\n      additional_kwargs: removeReasoningDetails(next.additional_kwargs),\n    })\n  );\n}\n\nfunction appendStreamChunk({\n  current,\n  next,\n  provider,\n}: {\n  current?: AIMessageChunk;\n  next: AIMessageChunk;\n  provider: Providers;\n}): AIMessageChunk {\n  if (current == null) {\n    return next;\n  }\n  return concat(\n    current,\n    removeOpenRouterFinalReasoningReplayContent({ current, next, provider })\n  );\n}\n\n/**\n * Invokes a chat model with the given messages, handling both streaming and\n * non-streaming paths.\n *\n * By default, stream chunks are processed through a `ChatModelStreamHandler`\n * that dispatches run steps (MESSAGE_CREATION, TOOL_CALLS) for the graph.\n * Pass an `onChunk` callback to override this with custom chunk processing\n * (e.g. summarization delta events).\n */\nexport async function attemptInvoke(\n  {\n    model,\n    messages,\n    provider,\n    context,\n    onChunk,\n  }: {\n    model: t.ChatModel;\n    messages: BaseMessage[];\n    provider: Providers;\n    context?: InvokeContext;\n    onChunk?: OnChunk;\n  },\n  config?: RunnableConfig\n): Promise<Partial<t.BaseGraphState>> {\n  /**\n   * Pull the run-scoped tool output registry off the graph (when one\n   * exists) and project ToolMessages carrying ref metadata into a\n   * transient annotated copy. The original `messages` array stays\n   * untouched so the graph state never sees `[ref: …]` / `_ref`\n   * payload.\n   */\n  const registry = context?.getOrCreateToolOutputRegistry();\n  const runId = config?.configurable?.run_id as string | undefined;\n  const messagesForProvider = annotateMessagesForLLM(messages, registry, runId);\n\n  /**\n   * Stamp the provider that is ACTUALLY serving this invocation onto the\n   * callback metadata. `attemptInvoke` is the single funnel for primary,\n   * fallback, and summarization model calls, so consumers that need\n   * provider attribution per call (the subagent usage-capture handler)\n   * read this key instead of trusting static agent config — which is\n   * wrong for fallback-served calls — or `ls_provider` — which derived\n   * providers inherit from their base class.\n   */\n  config = {\n    ...config,\n    metadata: {\n      ...(config?.metadata ?? {}),\n      [Constants.INVOKED_PROVIDER]: provider,\n    },\n  };\n\n  if (model.stream) {\n    const stream = await model.stream(messagesForProvider, config);\n    let finalChunk: AIMessageChunk | undefined;\n    const registeredStreamHandler =\n      getRegisteredDefaultChatStreamHandler(context);\n\n    if (onChunk) {\n      for await (const chunk of stream) {\n        await onChunk(chunk);\n        finalChunk = appendStreamChunk({\n          current: finalChunk,\n          next: chunk,\n          provider,\n        });\n      }\n    } else if (registeredStreamHandler == null) {\n      const metadata = config.metadata as Record<string, unknown> | undefined;\n      const streamHandler = new ChatModelStreamHandler();\n      for await (const chunk of stream) {\n        const handlingChunk = getStreamHandlingChunk({\n          current: finalChunk,\n          next: chunk,\n          provider,\n        });\n        if (handlingChunk != null) {\n          await streamHandler.handle(\n            GraphEvents.CHAT_MODEL_STREAM,\n            { chunk: handlingChunk },\n            metadata,\n            context\n          );\n        }\n        finalChunk = appendStreamChunk({\n          current: finalChunk,\n          next: chunk,\n          provider,\n        });\n      }\n    } else {\n      const metadata = config.metadata as Record<string, unknown> | undefined;\n      for await (const chunk of stream) {\n        const handlingChunk = getStreamHandlingChunk({\n          current: finalChunk,\n          next: chunk,\n          provider,\n        });\n        if (handlingChunk != null && handlingChunk !== chunk) {\n          await registeredStreamHandler.handle(\n            GraphEvents.CHAT_MODEL_STREAM,\n            { chunk: handlingChunk },\n            metadata,\n            context\n          );\n        }\n        finalChunk = appendStreamChunk({\n          current: finalChunk,\n          next: chunk,\n          provider,\n        });\n      }\n    }\n\n    if (manualToolStreamProviders.has(provider)) {\n      finalChunk = modifyDeltaProperties(provider, finalChunk);\n    }\n\n    if ((finalChunk?.tool_calls?.length ?? 0) > 0) {\n      finalChunk!.tool_calls = finalChunk!.tool_calls?.filter(\n        (tool_call: ToolCall) => !!tool_call.name\n      );\n    }\n\n    assertNotTruncatedToolCall(finalChunk, provider);\n    return { messages: [finalChunk as AIMessageChunk] };\n  }\n\n  const finalMessage = await model.invoke(messagesForProvider, config);\n  if ((finalMessage.tool_calls?.length ?? 0) > 0) {\n    finalMessage.tool_calls = finalMessage.tool_calls?.filter(\n      (tool_call: ToolCall) => !!tool_call.name\n    );\n  }\n  assertNotTruncatedToolCall(finalMessage, provider);\n  return { messages: [finalMessage] };\n}\n\n/**\n * Identifies which fallback produced an error, so a caller planning a\n * recovery can reason about the client that actually failed rather than the\n * primary's configuration — their context windows and output allowances\n * differ, which is the whole reason a fallback exists.\n */\nexport interface FallbackErrorContext {\n  provider: Providers;\n  clientOptions?: t.ClientOptions;\n  maxContextTokens?: number;\n}\n\nexport interface FallbackOverflowCandidate {\n  error: unknown;\n  context: FallbackErrorContext;\n}\n\nconst fallbackErrorContexts = new WeakMap<object, FallbackErrorContext>();\nconst fallbackOverflowCandidates = new WeakMap<\n  object,\n  FallbackOverflowCandidate[]\n>();\n\nfunction attachFallbackErrorContext(\n  error: unknown,\n  fallbackContext: FallbackErrorContext\n): void {\n  if (typeof error !== 'object' || error === null) {\n    return;\n  }\n  fallbackErrorContexts.set(error, fallbackContext);\n}\n\n/** Reads back the fallback attribution attached by `tryFallbackProviders`. */\nexport function getFallbackErrorContext(\n  error: unknown\n): FallbackErrorContext | undefined {\n  if (typeof error !== 'object' || error === null) {\n    return undefined;\n  }\n  return fallbackErrorContexts.get(error);\n}\n\n/** Returns every fallback overflow retained from an exhausted provider chain. */\nexport function getFallbackOverflowCandidates(\n  error: unknown\n): FallbackOverflowCandidate[] {\n  if (typeof error !== 'object' || error === null) {\n    return [];\n  }\n  return [...(fallbackOverflowCandidates.get(error) ?? [])];\n}\n\n/**\n * Best-effort read of the configured model name from client options.\n * Providers disagree on the key (`model` vs `modelName`).\n */\nfunction extractClientOptionsModel(\n  clientOptions: t.ClientOptions | undefined\n): string | undefined {\n  const options = clientOptions as\n    | { model?: unknown; modelName?: unknown }\n    | undefined;\n  if (typeof options?.model === 'string' && options.model !== '') {\n    return options.model;\n  }\n  if (typeof options?.modelName === 'string' && options.modelName !== '') {\n    return options.modelName;\n  }\n  return undefined;\n}\n\n/**\n * Attempts each fallback provider in order until one succeeds.\n *\n * When every fallback fails, a context overflow among them is thrown in\n * preference to whichever failure happened to come last. An overflow is the\n * one failure the caller can act on — it compacts and retries — and losing it\n * behind a later unrelated error would surface a dead end instead. Ordinary\n * failures still throw last-error-wins.\n */\nexport async function tryFallbackProviders({\n  fallbacks,\n  tools,\n  messages,\n  config,\n  primaryError,\n  context,\n  onChunk,\n  overflowContext,\n}: {\n  fallbacks: t.FallbackConfig[];\n  tools?: t.GraphTools;\n  messages: BaseMessage[];\n  config?: RunnableConfig;\n  primaryError: unknown;\n  context?: InvokeContext;\n  onChunk?: OnChunk;\n  /**\n   * Prompt-size corroboration for signatures that are not self-describing.\n   * Vertex AI's overflow is a bare `400` with no reason, so without this a\n   * fallback that overflows is indistinguishable from any other 400 and would\n   * be dropped in favour of whichever failure came last.\n   */\n  overflowContext?: ContextOverflowContext;\n}): Promise<Partial<t.BaseGraphState> | undefined> {\n  const isOverflow = (\n    error: unknown,\n    contextOverride = overflowContext\n  ): boolean => getContextOverflowInfo(error, contextOverride) != null;\n  let lastError: unknown = primaryError;\n  /**\n   * Tracked apart from the primary's overflow. A caller reaching this\n   * function with an overflowing primary has already failed to recover from\n   * it, so a fallback overflow — which may sit against a different window and\n   * output allowance — is the more useful of the two to surface.\n   */\n  const overflowCandidates: FallbackOverflowCandidate[] = [];\n  const primaryOverflowError: unknown = isOverflow(primaryError)\n    ? primaryError\n    : undefined;\n  for (const fb of fallbacks) {\n    try {\n      const fbModel = initializeModel({\n        provider: fb.provider,\n        clientOptions: fb.clientOptions,\n        tools,\n      });\n      /**\n       * Stamp the fallback's configured model onto callback metadata so\n       * per-call attribution (subagent usage capture) doesn't fall back to\n       * the PRIMARY config's model when the provider reports no\n       * `ls_model_name`. The serving provider is stamped uniformly by\n       * `attemptInvoke` (`INVOKED_PROVIDER`).\n       */\n      const fbModelName = extractClientOptionsModel(fb.clientOptions);\n      const fbConfig: RunnableConfig | undefined =\n        fbModelName == null\n          ? config\n          : {\n            ...config,\n            metadata: {\n              ...(config?.metadata ?? {}),\n              [Constants.INVOKED_MODEL]: fbModelName,\n            },\n          };\n      const result = await attemptInvoke(\n        {\n          model: fbModel as t.ChatModel,\n          messages,\n          provider: fb.provider,\n          context,\n          onChunk,\n        },\n        fbConfig\n      );\n      return result;\n    } catch (e) {\n      lastError = e;\n      const fallbackOverflowContext: ContextOverflowContext = {\n        provider: fb.provider,\n        maxContextTokens: fb.maxContextTokens,\n        ...(overflowContext?.provider === fb.provider\n          ? {\n            estimatedPromptTokens: overflowContext.estimatedPromptTokens,\n          }\n          : {}),\n      };\n      if (isOverflow(e, fallbackOverflowContext)) {\n        const errorContext: FallbackErrorContext = {\n          provider: fb.provider,\n          clientOptions: fb.clientOptions,\n          maxContextTokens: fb.maxContextTokens,\n        };\n        attachFallbackErrorContext(e, errorContext);\n        overflowCandidates.push({ error: e, context: errorContext });\n      }\n      continue;\n    }\n  }\n  /**\n   * Preference order: a fallback overflow, then the primary's overflow, then\n   * whichever failure came last. An overflow is the only one of the three a\n   * caller can act on, and the fallback's carries the client attribution that\n   * makes a correct retry budget possible.\n   */\n  const preferred =\n    overflowCandidates[0]?.error ?? primaryOverflowError ?? lastError;\n  if (\n    overflowCandidates.length > 0 &&\n    typeof preferred === 'object' &&\n    preferred !== null\n  ) {\n    fallbackOverflowCandidates.set(preferred, overflowCandidates);\n  }\n  if (preferred !== undefined) {\n    throw preferred;\n  }\n  return undefined;\n}\n"],"mappings":";;;;;;;;;;;;;AAwDA,SAAS,sCACP,SACoC;CACpC,MAAM,UAAU,SAAS,iBAAiB,WAAA,sBAE1C;CACA,OAAO,mBAAmBA,eAAAA,yBAAyB,UAAU,KAAA;AAC/D;AAEA,SAAS,oBAAoB,OAAgC;CAC3D,MAAM,mBAAmB,MAAM,kBAAkB;CACjD,OAAO,MAAM,QAAQ,gBAAgB,KAAK,iBAAiB,SAAS;AACtE;AAEA,SAAS,4CAA4C,EACnD,SACA,MACA,YAKiB;CACjB,MAAM,UAAU,mCAAmC;EACjD;EACA;EACA;CACF,CAAC;CACD,IAAI,WAAW,QAAQ,YAAY,KAAK,SACtC,OAAO;CAGT,OAAO,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;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,WAAW,SAAS,8BAA8B;CACxD,MAAM,QAAQ,QAAQ,cAAc;CACpC,MAAM,sBAAsBC,6BAAAA,uBAAuB,UAAU,UAAU,KAAK;;;;;;;;;;CAW5E,SAAS;EACP,GAAG;EACH,UAAU;GACR,GAAI,QAAQ,YAAY,CAAC;2BACK;EAChC;CACF;CAEA,IAAI,MAAM,QAAQ;EAChB,MAAM,SAAS,MAAM,MAAM,OAAO,qBAAqB,MAAM;EAC7D,IAAI;EACJ,MAAM,0BACJ,sCAAsC,OAAO;EAE/C,IAAI,SACF,WAAW,MAAM,SAAS,QAAQ;GAChC,MAAM,QAAQ,KAAK;GACnB,aAAa,kBAAkB;IAC7B,SAAS;IACT,MAAM;IACN;GACF,CAAC;EACH;OACK,IAAI,2BAA2B,MAAM;GAC1C,MAAM,WAAW,OAAO;GACxB,MAAM,gBAAgB,IAAIF,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;GACH;EACF,OAAO;GACL,MAAM,WAAW,OAAO;GACxB,WAAW,MAAM,SAAS,QAAQ;IAChC,MAAM,gBAAgB,uBAAuB;KAC3C,SAAS;KACT,MAAM;KACN;IACF,CAAC;IACD,IAAI,iBAAiB,QAAQ,kBAAkB,OAC7C,MAAM,wBAAwB,OAAA,wBAE5B,EAAE,OAAO,cAAc,GACvB,UACA,OACF;IAEF,aAAa,kBAAkB;KAC7B,SAAS;KACT,MAAM;KACN;IACF,CAAC;GACH;EACF;EAEA,IAAIG,kBAAAA,0BAA0B,IAAI,QAAQ,GACxC,aAAaC,aAAAA,sBAAsB,UAAU,UAAU;EAGzD,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,mBAgBiD;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;EAWJ,OAAO,MAVc,cACnB;GACE,OAAO;GACP;GACA,UAAU,GAAG;GACb;GACA;EACF,GACA,QACF;CAEF,SAAS,GAAG;EACV,YAAY;EAUZ,IAAI,WAAW,GAAG;GARhB,UAAU,GAAG;GACb,kBAAkB,GAAG;GACrB,GAAI,iBAAiB,aAAa,GAAG,WACjC,EACA,uBAAuB,gBAAgB,sBACzC,IACE,CAAC;EAEiC,CAAC,GAAG;GAC1C,MAAM,eAAqC;IACzC,UAAU,GAAG;IACb,eAAe,GAAG;IAClB,kBAAkB,GAAG;GACvB;GACA,2BAA2B,GAAG,YAAY;GAC1C,mBAAmB,KAAK;IAAE,OAAO;IAAG,SAAS;GAAa,CAAC;EAC7D;EACA;CACF;;;;;;;CAQF,MAAM,YACJ,mBAAmB,EAAE,EAAE,SAAS,wBAAwB;CAC1D,IACE,mBAAmB,SAAS,KAC5B,OAAO,cAAc,YACrB,cAAc,MAEd,2BAA2B,IAAI,WAAW,kBAAkB;CAE9D,IAAI,cAAc,KAAA,GAChB,MAAM;AAGV"}
+{"version":3,"file":"invoke.cjs","names":["isOpenAILike","projectToolStreamContentForProvider","projectOpenAIResponsesToolMessageContent","stripAnthropicCacheControl","stripBedrockCacheControl","projectComputerCallOutputsToText","projectOpenRouterToolMessageContent","projectOpenAIChatToolMessageContent","projectSingleTextToolOutputsToText","projectCacheControlledToolOutputsToText","projectStructuredToolOutputsToText","ChatModelStreamHandler","AIMessageChunk","annotateMessagesForLLM","manualToolStreamProviders","modifyDeltaProperties","getContextOverflowInfo","initializeModel"],"sources":["../../../src/llm/invoke.ts"],"sourcesContent":["import { concat } from '@langchain/core/utils/stream';\nimport { AIMessageChunk } from '@langchain/core/messages';\nimport type { RunnableConfig } from '@langchain/core/runnables';\nimport type { ToolCall } from '@langchain/core/messages/tool';\nimport type { BaseMessage } from '@langchain/core/messages';\nimport type { ToolOutputReferenceRegistry } from '@/tools/toolOutputReferences';\nimport type { ContextOverflowContext } from '@/utils/errors';\nimport type * as t from '@/types';\nimport {\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, GraphEvents, Providers } from '@/common';\nimport { manualToolStreamProviders } from '@/llm/providers';\nimport { getContextOverflowInfo } from '@/utils/errors';\nimport { modifyDeltaProperties } from '@/messages';\nimport { ChatModelStreamHandler } from '@/stream';\nimport { initializeModel } from '@/llm/init';\nimport { 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\nfunction getRegisteredDefaultChatStreamHandler(\n  context?: InvokeContext\n): ChatModelStreamHandler | undefined {\n  const handler = context?.handlerRegistry?.getHandler(\n    GraphEvents.CHAT_MODEL_STREAM\n  );\n  return handler instanceof ChatModelStreamHandler ? handler : undefined;\n}\n\nfunction hasReasoningDetails(chunk: AIMessageChunk): boolean {\n  const reasoningDetails = chunk.additional_kwargs.reasoning_details;\n  return Array.isArray(reasoningDetails) && reasoningDetails.length > 0;\n}\n\nfunction removeOpenRouterFinalReasoningReplayContent({\n  current,\n  next,\n  provider,\n}: {\n  current?: AIMessageChunk;\n  next: AIMessageChunk;\n  provider: Providers;\n}): AIMessageChunk {\n  const content = getOpenRouterFinalReasoningContent({\n    current,\n    next,\n    provider,\n  });\n  if (content == null || content === next.content) {\n    return next;\n  }\n\n  return new AIMessageChunk(\n    Object.assign({}, next, {\n      content,\n    })\n  );\n}\n\nfunction getOpenRouterFinalReasoningContent({\n  current,\n  next,\n  provider,\n}: {\n  current?: AIMessageChunk;\n  next: AIMessageChunk;\n  provider: Providers;\n}): string | undefined {\n  if (\n    provider !== Providers.OPENROUTER ||\n    current == null ||\n    !hasReasoningDetails(next) ||\n    typeof current.content !== 'string' ||\n    current.content === '' ||\n    typeof next.content !== 'string' ||\n    next.content === ''\n  ) {\n    return undefined;\n  }\n  if (!next.content.startsWith(current.content)) {\n    return next.content;\n  }\n  return next.content.slice(current.content.length);\n}\n\nfunction removeReasoningDetails(\n  additionalKwargs: AIMessageChunk['additional_kwargs']\n): AIMessageChunk['additional_kwargs'] {\n  return Object.fromEntries(\n    Object.entries(additionalKwargs).filter(\n      ([key]) => key !== 'reasoning_details'\n    )\n  );\n}\n\nfunction getStreamHandlingChunk({\n  current,\n  next,\n  provider,\n}: {\n  current?: AIMessageChunk;\n  next: AIMessageChunk;\n  provider: Providers;\n}): AIMessageChunk | undefined {\n  const content = getOpenRouterFinalReasoningContent({\n    current,\n    next,\n    provider,\n  });\n  if (content == null) {\n    return next;\n  }\n  if (content === '') {\n    return undefined;\n  }\n  return new AIMessageChunk(\n    Object.assign({}, next, {\n      content,\n      additional_kwargs: removeReasoningDetails(next.additional_kwargs),\n    })\n  );\n}\n\nfunction appendStreamChunk({\n  current,\n  next,\n  provider,\n}: {\n  current?: AIMessageChunk;\n  next: AIMessageChunk;\n  provider: Providers;\n}): AIMessageChunk {\n  if (current == null) {\n    return next;\n  }\n  return concat(\n    current,\n    removeOpenRouterFinalReasoningReplayContent({ current, next, provider })\n  );\n}\n\n/**\n * Invokes a chat model with the given messages, handling both streaming and\n * non-streaming paths.\n *\n * By default, stream chunks are processed through a `ChatModelStreamHandler`\n * that dispatches run steps (MESSAGE_CREATION, TOOL_CALLS) for the graph.\n * Pass an `onChunk` callback to override this with custom chunk processing\n * (e.g. summarization delta events).\n */\nexport async function attemptInvoke(\n  {\n    model,\n    messages,\n    provider,\n    context,\n    onChunk,\n  }: {\n    model: t.ChatModel;\n    messages: BaseMessage[];\n    provider: Providers;\n    context?: InvokeContext;\n    onChunk?: OnChunk;\n  },\n  config?: RunnableConfig\n): Promise<Partial<t.BaseGraphState>> {\n  /**\n   * Pull the run-scoped tool output registry off the graph (when one\n   * exists) and project ToolMessages carrying ref metadata into a\n   * transient annotated copy. The original `messages` array stays\n   * untouched so the graph state never sees `[ref: …]` / `_ref`\n   * payload.\n   */\n  const 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 messagesForProvider = annotateMessagesForLLM(\n    invocationMessages,\n    registry,\n    runId\n  );\n\n  /**\n   * Stamp the provider that is ACTUALLY serving this invocation onto the\n   * callback metadata. `attemptInvoke` is the single funnel for primary,\n   * fallback, and summarization model calls, so consumers that need\n   * provider attribution per call (the subagent usage-capture handler)\n   * read this key instead of trusting static agent config — which is\n   * wrong for fallback-served calls — or `ls_provider` — which derived\n   * providers inherit from their base class.\n   */\n  config = {\n    ...config,\n    metadata: {\n      ...(config?.metadata ?? {}),\n      [Constants.INVOKED_PROVIDER]: provider,\n    },\n  };\n\n  if (model.stream) {\n    const stream = await model.stream(messagesForProvider, config);\n    let finalChunk: AIMessageChunk | undefined;\n    const registeredStreamHandler =\n      getRegisteredDefaultChatStreamHandler(context);\n\n    if (onChunk) {\n      for await (const chunk of stream) {\n        await onChunk(chunk);\n        finalChunk = appendStreamChunk({\n          current: finalChunk,\n          next: chunk,\n          provider,\n        });\n      }\n    } else if (registeredStreamHandler == null) {\n      const metadata = config.metadata as Record<string, unknown> | undefined;\n      const streamHandler = new ChatModelStreamHandler();\n      for await (const chunk of stream) {\n        const handlingChunk = getStreamHandlingChunk({\n          current: finalChunk,\n          next: chunk,\n          provider,\n        });\n        if (handlingChunk != null) {\n          await streamHandler.handle(\n            GraphEvents.CHAT_MODEL_STREAM,\n            { chunk: handlingChunk },\n            metadata,\n            context\n          );\n        }\n        finalChunk = appendStreamChunk({\n          current: finalChunk,\n          next: chunk,\n          provider,\n        });\n      }\n    } else {\n      const metadata = config.metadata as Record<string, unknown> | undefined;\n      for await (const chunk of stream) {\n        const handlingChunk = getStreamHandlingChunk({\n          current: finalChunk,\n          next: chunk,\n          provider,\n        });\n        if (handlingChunk != null && handlingChunk !== chunk) {\n          await registeredStreamHandler.handle(\n            GraphEvents.CHAT_MODEL_STREAM,\n            { chunk: handlingChunk },\n            metadata,\n            context\n          );\n        }\n        finalChunk = appendStreamChunk({\n          current: finalChunk,\n          next: chunk,\n          provider,\n        });\n      }\n    }\n\n    if (manualToolStreamProviders.has(provider)) {\n      finalChunk = modifyDeltaProperties(provider, finalChunk);\n    }\n\n    if ((finalChunk?.tool_calls?.length ?? 0) > 0) {\n      finalChunk!.tool_calls = finalChunk!.tool_calls?.filter(\n        (tool_call: ToolCall) => !!tool_call.name\n      );\n    }\n\n    assertNotTruncatedToolCall(finalChunk, provider);\n    return { messages: [finalChunk as AIMessageChunk] };\n  }\n\n  const finalMessage = await model.invoke(messagesForProvider, config);\n  if ((finalMessage.tool_calls?.length ?? 0) > 0) {\n    finalMessage.tool_calls = finalMessage.tool_calls?.filter(\n      (tool_call: ToolCall) => !!tool_call.name\n    );\n  }\n  assertNotTruncatedToolCall(finalMessage, provider);\n  return { messages: [finalMessage] };\n}\n\n/**\n * Identifies which fallback produced an error, so a caller planning a\n * recovery can reason about the client that actually failed rather than the\n * primary's configuration — their context windows and output allowances\n * differ, which is the whole reason a fallback exists.\n */\nexport interface FallbackErrorContext {\n  provider: Providers;\n  clientOptions?: t.ClientOptions;\n  maxContextTokens?: number;\n}\n\nexport interface FallbackOverflowCandidate {\n  error: unknown;\n  context: FallbackErrorContext;\n}\n\nconst fallbackErrorContexts = new WeakMap<object, FallbackErrorContext>();\nconst fallbackOverflowCandidates = new WeakMap<\n  object,\n  FallbackOverflowCandidate[]\n>();\n\nfunction attachFallbackErrorContext(\n  error: unknown,\n  fallbackContext: FallbackErrorContext\n): void {\n  if (typeof error !== 'object' || error === null) {\n    return;\n  }\n  fallbackErrorContexts.set(error, fallbackContext);\n}\n\n/** Reads back the fallback attribution attached by `tryFallbackProviders`. */\nexport function getFallbackErrorContext(\n  error: unknown\n): FallbackErrorContext | undefined {\n  if (typeof error !== 'object' || error === null) {\n    return undefined;\n  }\n  return fallbackErrorContexts.get(error);\n}\n\n/** Returns every fallback overflow retained from an exhausted provider chain. */\nexport function getFallbackOverflowCandidates(\n  error: unknown\n): FallbackOverflowCandidate[] {\n  if (typeof error !== 'object' || error === null) {\n    return [];\n  }\n  return [...(fallbackOverflowCandidates.get(error) ?? [])];\n}\n\n/**\n * Best-effort read of the configured model name from client options.\n * Providers disagree on the key (`model` vs `modelName`).\n */\nfunction extractClientOptionsModel(\n  clientOptions: t.ClientOptions | undefined\n): string | undefined {\n  const options = clientOptions as\n    | { model?: unknown; modelName?: unknown }\n    | undefined;\n  if (typeof options?.model === 'string' && options.model !== '') {\n    return options.model;\n  }\n  if (typeof options?.modelName === 'string' && options.modelName !== '') {\n    return options.modelName;\n  }\n  return undefined;\n}\n\n/**\n * Attempts each fallback provider in order until one succeeds.\n *\n * When every fallback fails, a context overflow among them is thrown in\n * preference to whichever failure happened to come last. An overflow is the\n * one failure the caller can act on — it compacts and retries — and losing it\n * behind a later unrelated error would surface a dead end instead. Ordinary\n * failures still throw last-error-wins.\n */\nexport async function tryFallbackProviders({\n  fallbacks,\n  tools,\n  messages,\n  config,\n  primaryError,\n  context,\n  onChunk,\n  overflowContext,\n  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":";;;;;;;;;;;;;;;AAuEA,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;AAEA,SAAS,sCACP,SACoC;CACpC,MAAM,UAAU,SAAS,iBAAiB,WAAA,sBAE1C;CACA,OAAO,mBAAmBO,eAAAA,yBAAyB,UAAU,KAAA;AAC/D;AAEA,SAAS,oBAAoB,OAAgC;CAC3D,MAAM,mBAAmB,MAAM,kBAAkB;CACjD,OAAO,MAAM,QAAQ,gBAAgB,KAAK,iBAAiB,SAAS;AACtE;AAEA,SAAS,4CAA4C,EACnD,SACA,MACA,YAKiB;CACjB,MAAM,UAAU,mCAAmC;EACjD;EACA;EACA;CACF,CAAC;CACD,IAAI,WAAW,QAAQ,YAAY,KAAK,SACtC,OAAO;CAGT,OAAO,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;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,sBAAsBC,6BAAAA,uBAC1B,oBACA,UACA,KACF;;;;;;;;;;CAWA,SAAS;EACP,GAAG;EACH,UAAU;GACR,GAAI,QAAQ,YAAY,CAAC;2BACK;EAChC;CACF;CAEA,IAAI,MAAM,QAAQ;EAChB,MAAM,SAAS,MAAM,MAAM,OAAO,qBAAqB,MAAM;EAC7D,IAAI;EACJ,MAAM,0BACJ,sCAAsC,OAAO;EAE/C,IAAI,SACF,WAAW,MAAM,SAAS,QAAQ;GAChC,MAAM,QAAQ,KAAK;GACnB,aAAa,kBAAkB;IAC7B,SAAS;IACT,MAAM;IACN;GACF,CAAC;EACH;OACK,IAAI,2BAA2B,MAAM;GAC1C,MAAM,WAAW,OAAO;GACxB,MAAM,gBAAgB,IAAIF,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;GACH;EACF,OAAO;GACL,MAAM,WAAW,OAAO;GACxB,WAAW,MAAM,SAAS,QAAQ;IAChC,MAAM,gBAAgB,uBAAuB;KAC3C,SAAS;KACT,MAAM;KACN;IACF,CAAC;IACD,IAAI,iBAAiB,QAAQ,kBAAkB,OAC7C,MAAM,wBAAwB,OAAA,wBAE5B,EAAE,OAAO,cAAc,GACvB,UACA,OACF;IAEF,aAAa,kBAAkB;KAC7B,SAAS;KACT,MAAM;KACN;IACF,CAAC;GACH;EACF;EAEA,IAAIG,kBAAAA,0BAA0B,IAAI,QAAQ,GACxC,aAAaC,aAAAA,sBAAsB,UAAU,UAAU;EAGzD,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"}

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

@@ -1,3 +1,5 @@
+const require_cache = require("../../messages/cache.cjs");
+const require_core = require("../../messages/core.cjs");
 const require_streamedToolCallSeals = require("../../tools/streamedToolCallSeals.cjs");
 const require_index = require("./utils/index.cjs");
 const require_streamMetadata = require("./streamMetadata.cjs");
@@ -36,6 +38,22 @@ function normalizeHeaders(headers) {
 	});
 	return Object.fromEntries(output.entries());
 }
+function stripResponsesToolCacheControl(tools) {
+	if (!Array.isArray(tools)) return tools;
+	return tools.map((tool) => {
+		if (tool == null || typeof tool !== "object") return tool;
+		const clone = { ...tool };
+		delete clone.cache_control;
+		if (clone.extras != null && typeof clone.extras === "object" && !Array.isArray(clone.extras)) {
+			clone.extras = { ...clone.extras };
+			delete clone.extras.cache_control;
+		}
+		return clone;
+	});
+}
+function projectOpenAIResponsesProviderMessages(messages) {
+	return require_core.projectOpenAIResponsesToolMessageContent(require_cache.stripAnthropicCacheControl(require_cache.stripBedrockCacheControl(require_core.projectToolStreamContentForProvider(messages))));
+}
 const CACHE_WRITE_METADATA_KEY = "__librechat_cache_write_tokens";
 function applyManagedRequestParams(params, fields) {
 	return {
@@ -504,6 +522,7 @@ var LibreChatOpenAICompletions = class extends _langchain_openai.ChatOpenAICompl
 	includeReasoningContent;
 	includeReasoningDetails;
 	convertReasoningDetailsToContent;
+	preserveToolCacheControl;
 	promptCacheExplicit;
 	safetyIdentifier;
 	constructor(fields) {
@@ -511,6 +530,7 @@ var LibreChatOpenAICompletions = class extends _langchain_openai.ChatOpenAICompl
 		this.includeReasoningContent = fields?.includeReasoningContent;
 		this.includeReasoningDetails = fields?.includeReasoningDetails;
 		this.convertReasoningDetailsToContent = fields?.convertReasoningDetailsToContent;
+		this.preserveToolCacheControl = fields?.preserveToolCacheControl;
 		this.promptCacheExplicit = fields?.promptCacheExplicit;
 		this.safetyIdentifier = fields?.safety_identifier;
 	}
@@ -548,7 +568,8 @@ var LibreChatOpenAICompletions = class extends _langchain_openai.ChatOpenAICompl
 		const messagesMapped = require_index._convertMessagesToOpenAIParams(messages, this.model, {
 			includeReasoningContent: this.includeReasoningContent,
 			includeReasoningDetails: this.includeReasoningDetails,
-			convertReasoningDetailsToContent: this.convertReasoningDetailsToContent
+			convertReasoningDetailsToContent: this.convertReasoningDetailsToContent,
+			preserveToolCacheControl: this.preserveToolCacheControl
 		});
 		if (params.stream === true) {
 			const stream = this._streamResponseChunks(messages, options, runManager);
@@ -632,7 +653,8 @@ var LibreChatOpenAICompletions = class extends _langchain_openai.ChatOpenAICompl
 		const messagesMapped = require_index._convertMessagesToOpenAIParams(messages, this.model, {
 			includeReasoningContent: this.includeReasoningContent,
 			includeReasoningDetails: this.includeReasoningDetails,
-			convertReasoningDetailsToContent: this.convertReasoningDetailsToContent
+			convertReasoningDetailsToContent: this.convertReasoningDetailsToContent,
+			preserveToolCacheControl: this.preserveToolCacheControl
 		});
 		const params = {
 			...this.invocationParams(options, { streaming: true }),
@@ -711,17 +733,28 @@ var LibreChatOpenAICompletions = class extends _langchain_openai.ChatOpenAICompl
 };
 var LibreChatOpenAIResponses = class extends _langchain_openai.ChatOpenAIResponses {
 	promptCacheExplicit;
+	responsesPromptCache;
+	responsesPromptCacheTtl;
 	safetyIdentifier;
 	constructor(fields) {
 		super(fields);
 		this.promptCacheExplicit = fields?.promptCacheExplicit;
+		this.responsesPromptCache = fields?.responsesPromptCache;
+		this.responsesPromptCacheTtl = fields?.responsesPromptCacheTtl;
 		this.safetyIdentifier = fields?.safety_identifier;
 	}
 	invocationParams(options) {
-		const params = applyManagedRequestParams(super.invocationParams(options), {
+		const cacheOptions = options;
+		const cacheControl = (cacheOptions?.promptCache ?? this.responsesPromptCache) === true ? require_cache.buildAnthropicCacheControl(require_cache.resolvePromptCacheTtl(cacheOptions?.promptCacheTtl ?? this.responsesPromptCacheTtl)) : void 0;
+		const baseParams = applyManagedRequestParams(super.invocationParams(options), {
 			promptCacheExplicit: this.promptCacheExplicit,
 			safetyIdentifier: this.safetyIdentifier
 		});
+		const params = {
+			...baseParams,
+			...baseParams.tools != null && { tools: stripResponsesToolCacheControl(baseParams.tools) },
+			...cacheControl != null && { cache_control: cacheControl }
+		};
 		if (shouldIncludeEncryptedReasoning(this.model, params)) params.include = [...new Set([...params.include ?? [], "reasoning.encrypted_content"])];
 		return params;
 	}
@@ -734,16 +767,19 @@ var LibreChatOpenAIResponses = class extends _langchain_openai.ChatOpenAIRespons
 		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)) {
+		for await (const chunk of super._streamResponseChunks(projectOpenAIResponsesProviderMessages(messages), options, runManager)) {
 			attachCacheWriteUsage(chunk.message);
 			yield chunk;
 		}
 	}
+	async *_streamChatModelEvents(messages, options, runManager) {
+		yield* super._streamChatModelEvents(projectOpenAIResponsesProviderMessages(messages), options, runManager);
+	}
 	_getReasoningParams(options) {
 		return getReasoningParams(this.reasoning, options);
 	}