@librechat/agents 3.2.68 → 3.3.1

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

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

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

@@ -1,8 +1,8 @@
 import { OPENAI_CHAT_SEQUENTIAL_STREAMED_TOOL_CALL_ADAPTER, STREAMED_TOOL_CALL_ADAPTER_METADATA_KEY } from "../../tools/streamedToolCallSeals.mjs";
 import { _convertMessagesToOpenAIParams, isReasoningModel } from "./utils/index.mjs";
 import { dropRepeatedScalarMetadata } from "./streamMetadata.mjs";
-import { AIMessage, AIMessageChunk, isAIMessage } from "@langchain/core/messages";
 import { AzureChatOpenAI, AzureChatOpenAICompletions, AzureChatOpenAIResponses, ChatOpenAI, ChatOpenAICompletions, ChatOpenAIResponses, OpenAIClient, getEndpoint, getHeadersWithUserAgent } from "@langchain/openai";
+import { AIMessage, AIMessageChunk, isAIMessage } from "@langchain/core/messages";
 import { AzureOpenAI } from "openai";
 import { ChatXAI } from "@langchain/xai";
 import { ChatGenerationChunk } from "@langchain/core/outputs";

package/dist/esm/main.mjs

@@ -8,15 +8,15 @@ import { getMessageId } from "./messages/ids.mjs";
 import { HARD_MAX_TOOL_RESULT_CHARS, HARD_MAX_TOTAL_TOOL_OUTPUT_SIZE, calculateMaxToolResultChars, calculateMaxTotalToolOutputSize, truncateToolInput, truncateToolResultContent } from "./utils/truncation.mjs";
 import { DEFAULT_CONTEXT_PRUNING_SETTINGS, resolveContextPruningSettings } from "./messages/contextPruningSettings.mjs";
 import { applyContextPruning } from "./messages/contextPruning.mjs";
-import { DEFAULT_RESERVE_RATIO, ORIGINAL_CONTENT_MAX_CHARS, calculateTotalTokens, checkValidNumber, createPruneMessages, enforceOriginalContentCap, getMessagesWithinTokenLimit, maskConsumedToolResults, preFlightTruncateToolCallInputs, preFlightTruncateToolResults, repairOrphanedToolMessages, sanitizeOrphanToolBlocks } from "./messages/prune.mjs";
+import { CALIBRATION_RATIO_MAX, CALIBRATION_RATIO_MIN, DEFAULT_RESERVE_RATIO, ORIGINAL_CONTENT_MAX_CHARS, calculateTotalTokens, checkValidNumber, clampCalibrationRatio, createPruneMessages, enforceOriginalContentCap, getMessagesWithinTokenLimit, maskConsumedToolResults, preFlightTruncateToolCallInputs, preFlightTruncateToolResults, repairOrphanedToolMessages, sanitizeOrphanToolBlocks } from "./messages/prune.mjs";
 import { syncBudgetDerivedFields } from "./messages/budget.mjs";
-import { ensureThinkingBlockInMessages, formatAgentMessages, formatFromLangChain, formatLangChainMessages, formatMediaMessage, formatMessage, labelContentByAgent, shiftIndexTokenCountMap, withMessageRole } from "./messages/format.mjs";
+import { ensureThinkingBlockInMessages, foldToolBlocksForToollessAgent, formatAgentMessages, formatFromLangChain, formatLangChainMessages, formatMediaMessage, formatMessage, labelContentByAgent, shiftIndexTokenCountMap, withMessageRole } from "./messages/format.mjs";
 import { DEFAULT_PROMPT_CACHE_TTL, addBedrockCacheControl, addBedrockTailCacheControl, addCacheControl, addCacheControlToStablePrefixMessages, addTailCacheControl, buildAnthropicCacheControl, buildBedrockCachePoint, cloneMessage, resolveBedrockPromptCacheTtl, resolvePromptCacheTtl, stripAnthropicCacheControl, stripBedrockCacheControl, supportsBedrockToolCache } from "./messages/cache.mjs";
 import { makeIsDeferred, partitionAndMarkAnthropicToolCache } from "./messages/anthropicToolCache.mjs";
 import { formatContentStrings, isLegacyConvertible } from "./messages/content.mjs";
 import { extractToolDiscoveries, hasToolSearchInCurrentTurn } from "./messages/tools.mjs";
 import { REMOVE_ALL_MESSAGES, createRemoveAllMessage, messagesStateReducer } from "./messages/reducer.mjs";
-import { splitAtRecencyBoundary } from "./messages/recency.mjs";
+import { DEFAULT_RETAIN_RECENT_TURNS, splitAtRecencyBoundary } from "./messages/recency.mjs";
 import "./messages/index.mjs";
 import { joinKeys, resetIfNotEmpty } from "./utils/graph.mjs";
 import { isAnthropicLike, isGoogleLike, isOpenAILike } from "./utils/llm.mjs";
@@ -26,15 +26,19 @@ import { HandlerRegistry, LLMStreamHandler, ModelEndHandler, TestChatStreamHandl
 import { createHandlers } from "./utils/handlers.mjs";
 import { RunnableCallable, sleep } from "./utils/run.mjs";
 import { isZodSchema, toJsonSchema } from "./utils/schema.mjs";
-import { extractErrorMessage, isContextOverflowError, isLikelyContextOverflowError } from "./utils/errors.mjs";
+import { extractErrorMessage, getContextOverflowInfo, isContextOverflowError, isLikelyContextOverflowError } from "./utils/errors.mjs";
 import "./utils/index.mjs";
 import { CustomOpenAIClient } from "./llm/openai/index.mjs";
+import { ChatOpenRouter } from "./llm/openrouter/index.mjs";
+import { getChatModelClass } from "./llm/providers.mjs";
+import { initializeModel } from "./llm/init.mjs";
+import { attemptInvoke, tryFallbackProviders } from "./llm/invoke.mjs";
 import { bashAstFindingsToErrors, runBashAstChecks } from "./tools/local/bashAst.mjs";
 import { LOCAL_SPAWN_TIMEOUT_MS, _resetLocalEngineWarningsForTests, buildSandboxRuntimeConfig, executeLocalBash, executeLocalBashWithArgs, executeLocalCode, getLocalCwd, getLocalSessionId, getReadRoots, getSpawn, getWorkspaceFS, getWorkspaceRoots, getWriteRoots, resolveLocalExecutionConfig, resolveWorkspacePath, resolveWorkspacePathSafe, shellQuote, spawnLocalProcess, truncateLocalOutput, validateBashCommand } from "./tools/local/LocalExecutionEngine.mjs";
 import { CompileCheckToolName, createCompileCheckTool, createCompileCheckToolDefinition } from "./tools/local/CompileCheckTool.mjs";
 import { LocalFileCheckpointerImpl, createLocalFileCheckpointer } from "./tools/local/FileCheckpointer.mjs";
 import { appendCodeSessionFileSummary, stripCodeSessionFileSummary } from "./tools/CodeSessionFileSummary.mjs";
-import { BASH_SHELL_GUIDANCE, CODE_ARTIFACT_PATH_GUIDANCE, CodeExecutionToolDefinition, CodeExecutionToolDescription, CodeExecutionToolName, CodeExecutionToolSchema, FAILED_EXECUTION_FILE_REMINDER, STATEFUL_ENV_NOTE, StatefulCodeExecutionToolDescription, TMP_SCRATCH_OUTPUT_REMINDER, appendFailedExecutionFileReminder, appendTmpScratchReminder, buildCodeApiHttpErrorMessage, buildCodeExecutionToolDescription, buildCodeExecutionToolSchema, createCodeExecutionTool, emptyOutputMessage, getCodeBaseURL, resolveCodeApiAuthHeaders } from "./tools/CodeExecutor.mjs";
+import { BASH_SHELL_GUIDANCE, CODE_API_AUTHORIZATION_ERROR_MESSAGE, CODE_API_EXECUTION_FAILED_ERROR_MESSAGE, CODE_API_INVALID_REQUEST_ERROR_MESSAGE, CODE_API_RATE_LIMITED_ERROR_MESSAGE, CODE_API_UNAVAILABLE_ERROR_MESSAGE, CODE_ARTIFACT_PATH_GUIDANCE, CodeApiRequestError, CodeExecutionToolDefinition, CodeExecutionToolDescription, CodeExecutionToolName, CodeExecutionToolSchema, FAILED_EXECUTION_FILE_REMINDER, STATEFUL_ENV_NOTE, StatefulCodeExecutionToolDescription, TMP_SCRATCH_OUTPUT_REMINDER, appendFailedExecutionFileReminder, appendTmpScratchReminder, buildCodeApiExecutionErrorMessage, buildCodeApiHttpErrorMessage, buildCodeExecutionToolDescription, buildCodeExecutionToolSchema, createCodeExecutionTool, emptyOutputMessage, getCodeBaseURL, normalizeCodeApiRequestError, resolveCodeApiAuthHeaders } from "./tools/CodeExecutor.mjs";
 import { ProgrammaticToolCallingDefinition, ProgrammaticToolCallingDescription, ProgrammaticToolCallingName, ProgrammaticToolCallingSchema, createProgrammaticToolCallingSchema, createProgrammaticToolCallingTool, executeTools, extractUsedToolNames, fetchSessionFiles, filterToolsByUsage, formatCompletedResponse, makeRequest, normalizeToPythonIdentifier, unwrapToolResponse } from "./tools/ProgrammaticToolCalling.mjs";
 import { BashProgrammaticToolCallingDefinition, BashProgrammaticToolCallingDescription, BashProgrammaticToolCallingName, BashProgrammaticToolCallingSchema, createBashProgrammaticToolCallingSchema, createBashProgrammaticToolCallingTool, extractUsedBashToolNames, filterBashToolsByUsage, normalizeBashToolResultsForReplay, normalizeToBashIdentifier } from "./tools/BashProgrammaticToolCalling.mjs";
 import { HookRegistry } from "./hooks/HookRegistry.mjs";
@@ -62,10 +66,6 @@ import "./tools/local/index.mjs";
 import { ToolNode, toolsCondition } from "./tools/ToolNode.mjs";
 import { SubagentExecutor, buildChildInputs, filterSubagentResult, resolveSubagentConfigs, summarizeEvent } from "./tools/subagent/SubagentExecutor.mjs";
 import "./tools/subagent/index.mjs";
-import { ChatOpenRouter } from "./llm/openrouter/index.mjs";
-import { getChatModelClass } from "./llm/providers.mjs";
-import { initializeModel } from "./llm/init.mjs";
-import { attemptInvoke, tryFallbackProviders } from "./llm/invoke.mjs";
 import { SubagentToolDefinition, SubagentToolDescription, SubagentToolName, SubagentToolSchema, buildSubagentToolParams, createSubagentToolDefinition } from "./tools/SubagentTool.mjs";
 import { _resetUnrecognizedTriggerWarnings, shouldTriggerSummarization } from "./summarization/index.mjs";
 import { getMaxOutputTokensKey, isThinkingEnabled } from "./llm/request.mjs";
@@ -98,4 +98,4 @@ import { Runnable, RunnableLambda, RunnableSequence } from "./langchain/runnable
 import { DynamicStructuredTool, StructuredTool, Tool, tool } from "./langchain/tools.mjs";
 import "./langchain/index.mjs";
 import { BaseCheckpointSaver, Command, INTERRUPT, MemorySaver, interrupt, isInterrupted } from "@langchain/langgraph";
-export { AIMessage, AIMessageChunk, ANTHROPIC_TOOL_TOKEN_MULTIPLIER, AgentSession, BASH_SHELL_GUIDANCE, BaseCheckpointSaver, BaseMessage, BaseMessageChunk, BashExecutionToolDefinition, BashExecutionToolDescription, BashExecutionToolName, BashExecutionToolSchema, BashProgrammaticToolCallingDefinition, BashProgrammaticToolCallingDescription, BashProgrammaticToolCallingName, BashProgrammaticToolCallingSchema, BashToolOutputReferencesGuide, CLOUDFLARE_BASH_CODING_TOOL_NAMES, CLOUDFLARE_CODING_TOOL_NAMES, CODE_ARTIFACT_PATH_GUIDANCE, CODE_EXECUTION_TOOLS, Calculator, CalculatorSchema, CalculatorToolDefinition, CalculatorToolDescription, CalculatorToolName, Callback, ChatModelStreamHandler, ChatOpenRouter, CloudflareBashExecutionToolDescription, CloudflareCodeExecutionToolDescription, CodeExecutionToolDefinition, CodeExecutionToolDescription, CodeExecutionToolName, CodeExecutionToolSchema, Command, CommonEvents, CompileCheckToolName, Constants, ContentTypes, CustomOpenAIClient, DATE_RANGE, DEFAULT_CONTEXT_PRUNING_SETTINGS, DEFAULT_COUNTRY_DESCRIPTION, DEFAULT_HOOK_TIMEOUT_MS, DEFAULT_PROMPT_CACHE_TTL, DEFAULT_QUERY_DESCRIPTION, DEFAULT_RESERVE_RATIO, DEFAULT_TOOL_TOKEN_MULTIPLIER, DynamicStructuredTool, EnvVar, FAILED_EXECUTION_FILE_REMINDER, FakeChatModel, Graph, GraphEvents, GraphNodeActions, GraphNodeKeys, HARD_MAX_TOOL_RESULT_CHARS, HARD_MAX_TOTAL_TOOL_OUTPUT_SIZE, HOOK_EVENTS, HOOK_INJECTED_MESSAGES_CAPABLE, HandlerRegistry, HookRegistry, HumanMessage, IMAGE_TOKEN_SAFETY_MARGIN, INTERRUPT, JsonlSessionStore, LLMStreamHandler, LOCAL_CODING_BUNDLE_NAMES, LOCAL_CODING_TOOL_NAMES, LOCAL_SPAWN_TIMEOUT_MS, LocalBashExecutionToolDescription, LocalCodeExecutionToolDescription, LocalEditFileToolName, LocalEditFileToolSchema, LocalFileCheckpointerImpl, LocalGlobSearchToolName, LocalGlobSearchToolSchema, LocalGrepSearchToolName, LocalGrepSearchToolSchema, LocalListDirectoryToolName, LocalListDirectoryToolSchema, LocalReadFileToolSchema, LocalWriteFileToolName, LocalWriteFileToolSchema, MAX_CACHE_SIZE, MAX_PATTERN_LENGTH, MemorySaver, ModelEndHandler, MultiAgentGraph, ORIGINAL_CONTENT_MAX_CHARS, ProgrammaticToolCallingDefinition, ProgrammaticToolCallingDescription, ProgrammaticToolCallingName, ProgrammaticToolCallingSchema, PromptTemplate, Providers, REMOVE_ALL_MESSAGES, ReadFileToolDefinition, ReadFileToolDescription, ReadFileToolName, ReadFileToolSchema, Run, Runnable, RunnableCallable, RunnableLambda, RunnableSequence, SEPARATORS, STATEFUL_BASH_NOTE, STATEFUL_ENV_NOTE, SessionManager, SkillToolDefinition, SkillToolDescription, SkillToolName, SkillToolSchema, SplitStreamHandler, StandardGraph, StatefulBashExecutionToolDescription, StatefulCodeExecutionToolDescription, StepTypes, StructuredTool, SubagentExecutor, SubagentToolDefinition, SubagentToolDescription, SubagentToolName, SubagentToolSchema, SystemMessage, TMP_SCRATCH_OUTPUT_REMINDER, TestChatStreamHandler, TestLLMStreamHandler, TitleMethod, TokenEncoderManager, Tool, ToolCallTypes, ToolEndHandler, ToolMessage, ToolNode, ToolSearchToolDefinition, ToolSearchToolDescription, ToolSearchToolName, ToolSearchToolSchema, WebSearchToolDefinition, WebSearchToolDescription, WebSearchToolName, WebSearchToolSchema, _createBashProgramForTests, _resetLocalEngineWarningsForTests, _resetRipgrepCacheForTests, _resetSyntaxCheckProbeCacheForTests, _resetUnrecognizedTriggerWarnings, addBedrockCacheControl, addBedrockTailCacheControl, addCacheControl, addCacheControlToStablePrefixMessages, addTailCacheControl, appendCodeSessionFileSummary, appendFailedExecutionFileReminder, appendTmpScratchReminder, applyContextPruning, applyEdit, applyPreToolUseHooksForBridge, apportionTokenCounts, askUserQuestion, attemptInvoke, bashAstFindingsToErrors, buildAnthropicCacheControl, buildBashExecutionToolDescription, buildBashExecutionToolSchema, buildBedrockCachePoint, buildChildInputs, buildCodeApiHttpErrorMessage, buildCodeExecutionToolDescription, buildCodeExecutionToolSchema, buildSandboxRuntimeConfig, buildSubagentToolParams, calculateMaxToolResultChars, calculateMaxTotalToolOutputSize, calculateTotalTokens, checkValidNumber, classifyAttachment, clientExecTimeoutMs, clientFsTimeoutMs, cloneMessage, composeEventHandlers, convertMessagesToContent, countNestedGroups, countrySchema, createAgentSession, createBashExecutionTool, createBashProgrammaticToolCallingSchema, createBashProgrammaticToolCallingTool, createCloudflareBashExecutionTool, createCloudflareBashProgrammaticToolCallingTool, createCloudflareBridgeRuntime, createCloudflareCodeExecutionTool, createCloudflareCodingToolBundle, createCloudflareCodingTools, createCloudflareExecutionTool, createCloudflareLocalExecutionConfig, createCloudflareProgrammaticToolCallingTool, createCloudflareWorkspaceFS, createCodeExecutionTool, createCompileCheckTool, createCompileCheckToolDefinition, createContentAggregator, createFakeStreamingLLM, createHandlers, createLocalBashExecutionTool, createLocalBashProgrammaticToolCallingTool, createLocalCodeExecutionTool, createLocalCodingToolBundle, createLocalCodingToolDefinitions, createLocalCodingToolRegistry, createLocalCodingTools, createLocalEditFileTool, createLocalFileCheckpointer, createLocalGlobSearchTool, createLocalGrepSearchTool, createLocalListDirectoryTool, createLocalProgrammaticToolCallingTool, createLocalReadFileTool, createLocalWriteFileTool, createMetadataAggregator, createProgrammaticToolCallingSchema, createProgrammaticToolCallingTool, createPruneMessages, createRemoveAllMessage, createRunHandlers, createSchemaOnlyTool, createSchemaOnlyTools, createSearchTool, createSubagentToolDefinition, createTokenCounter, createToolPolicyHook, createToolSearch, createWorkspacePolicyHook, dateSchema, decodeFile, defaultOmitOptions, deserializeMessage, emptyOutputMessage, encodeFile, encodingForModel, enforceOriginalContentCap, ensureThinkingBlockInMessages, escapeRegexSpecialChars, estimateAnthropicImageTokens, estimateDocumentBlockTokens, estimateImageBlockTokens, estimateOpenAIImageTokens, estimateTimedMediaBlockTokens, execWithClientTimeout, executeCloudflareBash, executeCloudflareCode, executeHooks, executeLocalBash, executeLocalBashWithArgs, executeLocalCode, executeParallelSearches, executeTools, extractErrorMessage, extractImageDimensions, extractMcpServerName, extractTextFromContent, extractToolDiscoveries, extractUsedBashToolNames, extractUsedToolNames, fetchSessionFiles, filterBashToolsByUsage, filterSubagentResult, filterToolsByUsage, findLastIndex, formatAgentMessages, formatAnthropicArtifactContent, formatAnthropicMessage, formatArtifactPayload, formatCloudflareOutput, formatCompletedResponse, formatContentStrings, formatFromLangChain, formatLangChainMessages, formatMediaMessage, formatMessage, formatServerListing, formatSkillCatalog, getAvailableMcpServers, getBaseToolName, getBufferString, getChatModelClass, getChunkContent, getCloudflareWorkspaceRoot, getCodeBaseURL, getConverseOverrideMessage, getDeferredToolsListing, getLocalCwd, getLocalSessionId, getMaxOutputTokensKey, getMessageId, getMessagesWithinTokenLimit, getReadRoots, getSpawn, getTokenCountForMessage, getWorkspaceFS, getWorkspaceRoots, getWriteRoots, handleServerToolResult, handleToolCallChunks, handleToolCalls, hasNestedQuantifier, hasNestedQuantifiers, hasToolSearchInCurrentTurn, imageAttachmentContent, imagesSchema, initializeModel, interrupt, isAIMessage, isAnthropicLike, isBaseMessage, isContextOverflowError, isDangerousPattern, isFromAnyMcpServer, isFromMcpServer, isGoogleLike, isInterrupted, isLegacyConvertible, isLikelyContextOverflowError, isOpenAILike, isPresent, isThinkingEnabled, isToolMessage, isZodSchema, joinKeys, labelContentByAgent, locateEdit, makeIsDeferred, makeRequest, maskConsumedToolResults, matchesQuery, messagesStateReducer, modifyDeltaProperties, newsSchema, normalizeBashToolResultsForReplay, normalizeServerFilter, normalizeToBashIdentifier, normalizeToPythonIdentifier, parseBooleanEnv, partitionAndMarkAnthropicToolCache, performLocalSearch, preFlightTruncateToolCallInputs, preFlightTruncateToolResults, projectAgentContextUsage, querySchema, repairOrphanedToolMessages, resetIfNotEmpty, resolveBedrockPromptCacheTtl, resolveCloudflareSandbox, resolveCodeApiAuthHeaders, resolveContextPruningSettings, resolveLocalExecutionConfig, resolveLocalExecutionTools, resolveLocalToolRegistry, resolveLocalToolsForBinding, resolvePromptCacheTtl, resolveSubagentConfigs, resolveWorkspacePath, resolveWorkspacePathSafe, runBashAstChecks, runPostEditSyntaxCheck, sanitizeOrphanToolBlocks, sanitizeRegex, serializeMessage, shellQuote, shiftIndexTokenCountMap, shouldTriggerSummarization, sleep, spawnLocalProcess, splitAtRecencyBoundary, stripAnthropicCacheControl, stripBedrockCacheControl, stripCodeSessionFileSummary, summarizeEvent, supportsBedrockToolCache, syncBudgetDerivedFields, toJsonSchema, tool, toolResultTypes, toolsCondition, truncateLocalOutput, truncateToolInput, truncateToolResultContent, tryFallbackProviders, unescapeObject, unwrapToolResponse, validateBashCommand, validateCloudflareBashCommand, videosSchema, withClientTimeout, withMessageRole };
+export { AIMessage, AIMessageChunk, ANTHROPIC_TOOL_TOKEN_MULTIPLIER, AgentSession, BASH_SHELL_GUIDANCE, BaseCheckpointSaver, BaseMessage, BaseMessageChunk, BashExecutionToolDefinition, BashExecutionToolDescription, BashExecutionToolName, BashExecutionToolSchema, BashProgrammaticToolCallingDefinition, BashProgrammaticToolCallingDescription, BashProgrammaticToolCallingName, BashProgrammaticToolCallingSchema, BashToolOutputReferencesGuide, CALIBRATION_RATIO_MAX, CALIBRATION_RATIO_MIN, CLOUDFLARE_BASH_CODING_TOOL_NAMES, CLOUDFLARE_CODING_TOOL_NAMES, CODE_API_AUTHORIZATION_ERROR_MESSAGE, CODE_API_EXECUTION_FAILED_ERROR_MESSAGE, CODE_API_INVALID_REQUEST_ERROR_MESSAGE, CODE_API_RATE_LIMITED_ERROR_MESSAGE, CODE_API_UNAVAILABLE_ERROR_MESSAGE, CODE_ARTIFACT_PATH_GUIDANCE, CODE_EXECUTION_TOOLS, Calculator, CalculatorSchema, CalculatorToolDefinition, CalculatorToolDescription, CalculatorToolName, Callback, ChatModelStreamHandler, ChatOpenRouter, CloudflareBashExecutionToolDescription, CloudflareCodeExecutionToolDescription, CodeApiRequestError, CodeExecutionToolDefinition, CodeExecutionToolDescription, CodeExecutionToolName, CodeExecutionToolSchema, Command, CommonEvents, CompileCheckToolName, Constants, ContentTypes, CustomOpenAIClient, DATE_RANGE, DEFAULT_CONTEXT_PRUNING_SETTINGS, DEFAULT_COUNTRY_DESCRIPTION, DEFAULT_HOOK_TIMEOUT_MS, DEFAULT_PROMPT_CACHE_TTL, DEFAULT_QUERY_DESCRIPTION, DEFAULT_RESERVE_RATIO, DEFAULT_RETAIN_RECENT_TURNS, DEFAULT_TOOL_TOKEN_MULTIPLIER, DynamicStructuredTool, EnvVar, FAILED_EXECUTION_FILE_REMINDER, FakeChatModel, Graph, GraphEvents, GraphNodeActions, GraphNodeKeys, HARD_MAX_TOOL_RESULT_CHARS, HARD_MAX_TOTAL_TOOL_OUTPUT_SIZE, HOOK_EVENTS, HOOK_INJECTED_MESSAGES_CAPABLE, HandlerRegistry, HookRegistry, HumanMessage, IMAGE_TOKEN_SAFETY_MARGIN, INTERRUPT, JsonlSessionStore, LLMStreamHandler, LOCAL_CODING_BUNDLE_NAMES, LOCAL_CODING_TOOL_NAMES, LOCAL_SPAWN_TIMEOUT_MS, LocalBashExecutionToolDescription, LocalCodeExecutionToolDescription, LocalEditFileToolName, LocalEditFileToolSchema, LocalFileCheckpointerImpl, LocalGlobSearchToolName, LocalGlobSearchToolSchema, LocalGrepSearchToolName, LocalGrepSearchToolSchema, LocalListDirectoryToolName, LocalListDirectoryToolSchema, LocalReadFileToolSchema, LocalWriteFileToolName, LocalWriteFileToolSchema, MAX_CACHE_SIZE, MAX_PATTERN_LENGTH, MemorySaver, ModelEndHandler, MultiAgentGraph, ORIGINAL_CONTENT_MAX_CHARS, ProgrammaticToolCallingDefinition, ProgrammaticToolCallingDescription, ProgrammaticToolCallingName, ProgrammaticToolCallingSchema, PromptTemplate, Providers, REMOVE_ALL_MESSAGES, ReadFileToolDefinition, ReadFileToolDescription, ReadFileToolName, ReadFileToolSchema, Run, Runnable, RunnableCallable, RunnableLambda, RunnableSequence, SEPARATORS, STATEFUL_BASH_NOTE, STATEFUL_ENV_NOTE, SessionManager, SkillToolDefinition, SkillToolDescription, SkillToolName, SkillToolSchema, SplitStreamHandler, StandardGraph, StatefulBashExecutionToolDescription, StatefulCodeExecutionToolDescription, StepTypes, StructuredTool, SubagentExecutor, SubagentToolDefinition, SubagentToolDescription, SubagentToolName, SubagentToolSchema, SystemMessage, TMP_SCRATCH_OUTPUT_REMINDER, TestChatStreamHandler, TestLLMStreamHandler, TitleMethod, TokenEncoderManager, Tool, ToolCallTypes, ToolEndHandler, ToolMessage, ToolNode, ToolSearchToolDefinition, ToolSearchToolDescription, ToolSearchToolName, ToolSearchToolSchema, WebSearchToolDefinition, WebSearchToolDescription, WebSearchToolName, WebSearchToolSchema, _createBashProgramForTests, _resetLocalEngineWarningsForTests, _resetRipgrepCacheForTests, _resetSyntaxCheckProbeCacheForTests, _resetUnrecognizedTriggerWarnings, addBedrockCacheControl, addBedrockTailCacheControl, addCacheControl, addCacheControlToStablePrefixMessages, addTailCacheControl, appendCodeSessionFileSummary, appendFailedExecutionFileReminder, appendTmpScratchReminder, applyContextPruning, applyEdit, applyPreToolUseHooksForBridge, apportionTokenCounts, askUserQuestion, attemptInvoke, bashAstFindingsToErrors, buildAnthropicCacheControl, buildBashExecutionToolDescription, buildBashExecutionToolSchema, buildBedrockCachePoint, buildChildInputs, buildCodeApiExecutionErrorMessage, buildCodeApiHttpErrorMessage, buildCodeExecutionToolDescription, buildCodeExecutionToolSchema, buildSandboxRuntimeConfig, buildSubagentToolParams, calculateMaxToolResultChars, calculateMaxTotalToolOutputSize, calculateTotalTokens, checkValidNumber, clampCalibrationRatio, classifyAttachment, clientExecTimeoutMs, clientFsTimeoutMs, cloneMessage, composeEventHandlers, convertMessagesToContent, countNestedGroups, countrySchema, createAgentSession, createBashExecutionTool, createBashProgrammaticToolCallingSchema, createBashProgrammaticToolCallingTool, createCloudflareBashExecutionTool, createCloudflareBashProgrammaticToolCallingTool, createCloudflareBridgeRuntime, createCloudflareCodeExecutionTool, createCloudflareCodingToolBundle, createCloudflareCodingTools, createCloudflareExecutionTool, createCloudflareLocalExecutionConfig, createCloudflareProgrammaticToolCallingTool, createCloudflareWorkspaceFS, createCodeExecutionTool, createCompileCheckTool, createCompileCheckToolDefinition, createContentAggregator, createFakeStreamingLLM, createHandlers, createLocalBashExecutionTool, createLocalBashProgrammaticToolCallingTool, createLocalCodeExecutionTool, createLocalCodingToolBundle, createLocalCodingToolDefinitions, createLocalCodingToolRegistry, createLocalCodingTools, createLocalEditFileTool, createLocalFileCheckpointer, createLocalGlobSearchTool, createLocalGrepSearchTool, createLocalListDirectoryTool, createLocalProgrammaticToolCallingTool, createLocalReadFileTool, createLocalWriteFileTool, createMetadataAggregator, createProgrammaticToolCallingSchema, createProgrammaticToolCallingTool, createPruneMessages, createRemoveAllMessage, createRunHandlers, createSchemaOnlyTool, createSchemaOnlyTools, createSearchTool, createSubagentToolDefinition, createTokenCounter, createToolPolicyHook, createToolSearch, createWorkspacePolicyHook, dateSchema, decodeFile, defaultOmitOptions, deserializeMessage, emptyOutputMessage, encodeFile, encodingForModel, enforceOriginalContentCap, ensureThinkingBlockInMessages, escapeRegexSpecialChars, estimateAnthropicImageTokens, estimateDocumentBlockTokens, estimateImageBlockTokens, estimateOpenAIImageTokens, estimateTimedMediaBlockTokens, execWithClientTimeout, executeCloudflareBash, executeCloudflareCode, executeHooks, executeLocalBash, executeLocalBashWithArgs, executeLocalCode, executeParallelSearches, executeTools, extractErrorMessage, extractImageDimensions, extractMcpServerName, extractTextFromContent, extractToolDiscoveries, extractUsedBashToolNames, extractUsedToolNames, fetchSessionFiles, filterBashToolsByUsage, filterSubagentResult, filterToolsByUsage, findLastIndex, foldToolBlocksForToollessAgent, formatAgentMessages, formatAnthropicArtifactContent, formatAnthropicMessage, formatArtifactPayload, formatCloudflareOutput, formatCompletedResponse, formatContentStrings, formatFromLangChain, formatLangChainMessages, formatMediaMessage, formatMessage, formatServerListing, formatSkillCatalog, getAvailableMcpServers, getBaseToolName, getBufferString, getChatModelClass, getChunkContent, getCloudflareWorkspaceRoot, getCodeBaseURL, getContextOverflowInfo, getConverseOverrideMessage, getDeferredToolsListing, getLocalCwd, getLocalSessionId, getMaxOutputTokensKey, getMessageId, getMessagesWithinTokenLimit, getReadRoots, getSpawn, getTokenCountForMessage, getWorkspaceFS, getWorkspaceRoots, getWriteRoots, handleServerToolResult, handleToolCallChunks, handleToolCalls, hasNestedQuantifier, hasNestedQuantifiers, hasToolSearchInCurrentTurn, imageAttachmentContent, imagesSchema, initializeModel, interrupt, isAIMessage, isAnthropicLike, isBaseMessage, isContextOverflowError, isDangerousPattern, isFromAnyMcpServer, isFromMcpServer, isGoogleLike, isInterrupted, isLegacyConvertible, isLikelyContextOverflowError, isOpenAILike, isPresent, isThinkingEnabled, isToolMessage, isZodSchema, joinKeys, labelContentByAgent, locateEdit, makeIsDeferred, makeRequest, maskConsumedToolResults, matchesQuery, messagesStateReducer, modifyDeltaProperties, newsSchema, normalizeBashToolResultsForReplay, normalizeCodeApiRequestError, normalizeServerFilter, normalizeToBashIdentifier, normalizeToPythonIdentifier, parseBooleanEnv, partitionAndMarkAnthropicToolCache, performLocalSearch, preFlightTruncateToolCallInputs, preFlightTruncateToolResults, projectAgentContextUsage, querySchema, repairOrphanedToolMessages, resetIfNotEmpty, resolveBedrockPromptCacheTtl, resolveCloudflareSandbox, resolveCodeApiAuthHeaders, resolveContextPruningSettings, resolveLocalExecutionConfig, resolveLocalExecutionTools, resolveLocalToolRegistry, resolveLocalToolsForBinding, resolvePromptCacheTtl, resolveSubagentConfigs, resolveWorkspacePath, resolveWorkspacePathSafe, runBashAstChecks, runPostEditSyntaxCheck, sanitizeOrphanToolBlocks, sanitizeRegex, serializeMessage, shellQuote, shiftIndexTokenCountMap, shouldTriggerSummarization, sleep, spawnLocalProcess, splitAtRecencyBoundary, stripAnthropicCacheControl, stripBedrockCacheControl, stripCodeSessionFileSummary, summarizeEvent, supportsBedrockToolCache, syncBudgetDerivedFields, toJsonSchema, tool, toolResultTypes, toolsCondition, truncateLocalOutput, truncateToolInput, truncateToolResultContent, tryFallbackProviders, unescapeObject, unwrapToolResponse, validateBashCommand, validateCloudflareBashCommand, videosSchema, withClientTimeout, withMessageRole };

package/dist/esm/messages/format.mjs

@@ -152,7 +152,7 @@ function getTextContent(part) {
 }
 function hasMeaningfulAssistantContent(part) {
 	if (part.type === "text") return getTextContent(part).trim().length > 0;
-	if (part.type === "tool_call" || part.type === "error" || part.type === "agent_update" || part.type === "summary") return false;
+	if (part.type === "tool_call" || part.type === "error" || part.type === "agent_update" || part.type === "summary" || part.type === "activity_label") return false;
 	if (part.type === "think" || part.type === "thinking" || part.type === "reasoning" || part.type === "reasoning_content" || part.type === "redacted_thinking") return extractReasoningContent(part).trim().length > 0;
 	return part.type != null && part.type !== "";
 }
@@ -357,7 +357,7 @@ function formatAssistantMessage(message, options) {
 				*  flushed above or intentionally dropped when not preserving). */
 				hasReasoning = false;
 				pendingReasoningContent = "";
-			} else if (part.type === "error" || part.type === "agent_update" || part.type === "summary") continue;
+			} else if (part.type === "error" || part.type === "agent_update" || part.type === "summary" || part.type === "activity_label") continue;
 			else {
 				if (part.type === "text" && !getTextContent(part).trim()) continue;
 				currentContent.push(part);
@@ -422,6 +422,12 @@ function labelAllAgentContent(contentParts, agentIdMap, agentNames) {
 	};
 	for (let i = 0; i < contentParts.length; i++) {
 		const part = contentParts[i];
+		/** UI-only progress headers are not agent content and must not disturb
+		*  agent state: a label with no `agentIdMap` entry would otherwise read
+		*  as an agent change and flush the buffer mid-agent, splitting one
+		*  agent's contiguous content into two labeled blocks. Skipped before
+		*  any state transition below (mirrors the transfer path). */
+		if (part.type === "activity_label") continue;
 		const agentId = agentIdMap[i];
 		if (agentId !== currentAgentId && currentAgentId !== void 0) flushAgentBuffer();
 		currentAgentId = agentId;
@@ -487,6 +493,12 @@ const labelContentByAgent = (contentParts, agentIdMap, agentNames, options) => {
 	};
 	for (let i = 0; i < contentParts.length; i++) {
 		const part = contentParts[i];
+		/** UI-only progress headers are not agent content and must not disturb
+		*  agent state: a label with no `agentIdMap` entry would otherwise look
+		*  like an agent change, flushing the buffer and resetting an open
+		*  transfer capture so the transferred agent's following chunks lose
+		*  their frame. Skipped before any state transition below. */
+		if (part.type === "activity_label") continue;
 		const agentId = agentIdMap[i];
 		const isTransferTool = (part.type === "tool_call" && part.tool_call?.name?.startsWith("lc_transfer_to_")) ?? false;
 		if (agentId !== currentAgentId && currentAgentId !== void 0) flushAgentBuffer();
@@ -902,6 +914,34 @@ function appendMessageContent(msg, role, textChunks, parts) {
 			textChunks.push(`${role}: [tool_use] ${String(block.name ?? "")} ${JSON.stringify(block.input ?? {})}`);
 			continue;
 		}
+		if (block.type === "tool_call") {
+			hasToolUseBlock = true;
+			const nested = block.tool_call;
+			const name = String(nested?.name ?? block.name ?? "");
+			const rawArgs = nested?.args ?? block.args ?? {};
+			const argsText = typeof rawArgs === "string" ? rawArgs : JSON.stringify(rawArgs);
+			textChunks.push(`${role}: [tool_use] ${name} ${argsText}`.trimEnd());
+			const output = nested?.output;
+			if (output != null && output !== "") textChunks.push(`Tool: ${String(output)}`);
+			continue;
+		}
+		if (block.type === "tool_result") {
+			hasToolUseBlock = true;
+			const inner = block.content;
+			if (typeof inner === "string") {
+				if (inner) textChunks.push(`${role}: [tool_result] ${inner}`);
+			} else if (Array.isArray(inner)) for (const innerBlock of inner) if (typeof innerBlock === "string") {
+				if (innerBlock) textChunks.push(`${role}: [tool_result] ${innerBlock}`);
+			} else if (IMAGE_BLOCK_TYPES.has(innerBlock.type ?? "")) {
+				flushTextChunks(textChunks, parts);
+				parts.push({ ...innerBlock });
+			} else {
+				const innerText = innerBlock.text ?? innerBlock.input;
+				textChunks.push(`${role}: [tool_result] ${typeof innerText === "string" && innerText ? innerText : JSON.stringify(innerBlock)}`);
+			}
+			else if (inner != null) textChunks.push(`${role}: [tool_result] ${JSON.stringify(inner)}`);
+			continue;
+		}
 		const text = block.text ?? block.input;
 		if (typeof text === "string" && text) {
 			textChunks.push(`${role}: ${text}`);
@@ -914,8 +954,17 @@ function appendMessageContent(msg, role, textChunks, parts) {
 function appendToolCalls(msg, role, textChunks) {
 	if (role !== "AI") return;
 	const aiMsg = msg;
-	if (!aiMsg.tool_calls || aiMsg.tool_calls.length === 0) return;
-	for (const tc of aiMsg.tool_calls) textChunks.push(`AI: [tool_call] ${tc.name}(${JSON.stringify(tc.args)})`);
+	if (aiMsg.tool_calls && aiMsg.tool_calls.length > 0) {
+		for (const tc of aiMsg.tool_calls) textChunks.push(`AI: [tool_call] ${tc.name}(${JSON.stringify(tc.args)})`);
+		return;
+	}
+	const rawToolCalls = aiMsg.additional_kwargs.tool_calls;
+	if (!Array.isArray(rawToolCalls)) return;
+	for (const tc of rawToolCalls) {
+		const fn = tc.function;
+		if (fn == null) continue;
+		textChunks.push(`AI: [tool_call] ${String(fn.name ?? "")}(${String(fn.arguments ?? "")})`);
+	}
 }
 /**
 * Ensures compatibility when switching from a non-thinking agent to a thinking-enabled agent.
@@ -1005,6 +1054,88 @@ function ensureThinkingBlockInMessages(messages, _provider, config, runStartInde
 	}
 	return result;
 }
+/** Whether a message carries tool content a tool-less agent cannot legally
+*  send. Covers every representation a provider converter will serialize back
+*  into a request: a ToolMessage, parsed `AIMessage.tool_calls`, raw
+*  `additional_kwargs.tool_calls` (OpenAI keeps calls here when the parsed
+*  array is empty), and `tool_use` / `tool_call` / `tool_result` content
+*  blocks (`@langchain/aws` and the Anthropic converter map these to Converse
+*  `toolUse` / `toolResult`). Missing the parent AI message is not just a
+*  passthrough: folding its ToolMessage alone would leave an orphan
+*  `assistant(tool_calls) -> user(...)` sequence. */
+function messageHasToolContent(msg) {
+	if (isToolMessage$1(msg)) return true;
+	const aiMsg = msg;
+	if (aiMsg.tool_calls != null && aiMsg.tool_calls.length > 0) return true;
+	const rawToolCalls = aiMsg.additional_kwargs.tool_calls;
+	if (Array.isArray(rawToolCalls) && rawToolCalls.length > 0) return true;
+	if (Array.isArray(msg.content)) {
+		for (const block of msg.content) if (typeof block === "object" && (block.type === "tool_use" || block.type === "tool_call" || block.type === "tool_result")) return true;
+	}
+	return false;
+}
+/** Whether a message carries a tool RESULT: a ToolMessage, or a message whose
+*  content includes a `tool_result` block (the shape when a call/result pair is
+*  split as `AIMessage(tool_call)` + `HumanMessage(tool_result)`). Such a result
+*  belongs with the preceding tool call, so it is absorbed into the same fold
+*  and labelled as tool output. */
+function isToolResultMessage(msg) {
+	if (isToolMessage$1(msg)) return true;
+	if (Array.isArray(msg.content)) return msg.content.some((block) => typeof block === "object" && block.type === "tool_result");
+	return false;
+}
+/**
+* Folds tool_use / tool_result content into plain text for an agent that binds
+* no tools.
+*
+* In a multi-agent graph, a tool-less destination still inherits the prior
+* agent's conversation history, which can contain toolUse/toolResult blocks.
+* Because it binds no tools, the model is invoked with no tool schema — and
+* Bedrock's Converse API rejects any request that carries toolUse/toolResult
+* blocks without a top-level toolConfig ("The toolConfig field must be defined
+* when using toolUse and toolResult content blocks"). Adding a dummy toolConfig
+* is not an option: AWS requires at least one tool, and it would expose a
+* capability the destination was intentionally denied.
+*
+* Each tool-call turn plus its trailing tool results (ToolMessages or
+* `tool_result` content blocks) is collapsed into a single `[Previous tool
+* interaction]` HumanMessage that preserves the tool name, arguments and result
+* as text (image blocks are kept as-is). Runs in a single pass: non-tool
+* messages pass through, `result` is allocated lazily on the first fold, and the
+* original array is returned unchanged when it holds no tool content (the common
+* fresh-tool-less-agent case).
+*/
+function foldToolBlocksForToollessAgent(messages, config) {
+	let result = null;
+	let foldedCount = 0;
+	let i = 0;
+	while (i < messages.length) {
+		const msg = messages[i];
+		if (!messageHasToolContent(msg)) {
+			result?.push(msg);
+			i++;
+			continue;
+		}
+		/** First fold — copy the untouched prefix once, then append from here. */
+		if (result === null) result = messages.slice(0, i);
+		const parts = [];
+		const textChunks = ["[Previous tool interaction]"];
+		appendMessageContent(msg, isToolResultMessage(msg) ? "Tool" : "AI", textChunks, parts);
+		foldedCount++;
+		let j = i + 1;
+		while (j < messages.length && isToolResultMessage(messages[j])) {
+			appendMessageContent(messages[j], "Tool", textChunks, parts);
+			foldedCount++;
+			j++;
+		}
+		flushTextChunks(textChunks, parts);
+		result.push(withMessageRole(new HumanMessage({ content: toLangChainContent(parts) }), "user"));
+		i = j;
+	}
+	if (result === null) return messages;
+	emitAgentLog(config, "warn", "format", `foldToolBlocksForToollessAgent: folded ${foldedCount} tool message(s) into text for a tool-less agent`);
+	return result;
+}
 /**
 * Walks backwards from `currentIndex` through the message array to check
 * whether an earlier AI message in the same "chain" (no HumanMessage boundary)
@@ -1031,6 +1162,6 @@ function chainHasThinkingBlock(messages, currentIndex) {
 	return false;
 }
 //#endregion
-export { ensureThinkingBlockInMessages, formatAgentMessages, formatFromLangChain, formatLangChainMessages, formatMediaMessage, formatMessage, labelContentByAgent, shiftIndexTokenCountMap, withMessageRole };
+export { ensureThinkingBlockInMessages, foldToolBlocksForToollessAgent, formatAgentMessages, formatFromLangChain, formatLangChainMessages, formatMediaMessage, formatMessage, labelContentByAgent, shiftIndexTokenCountMap, withMessageRole };
 
 //# sourceMappingURL=format.mjs.map