@librechat/agents 3.2.54 → 3.2.55

package/dist/cjs/llm/invoke.cjs

@@ -4,6 +4,7 @@ const require_core = require("../messages/core.cjs");
 require("../messages/index.cjs");
 const require_toolOutputReferences = require("../tools/toolOutputReferences.cjs");
 const require_stream = require("../stream.cjs");
+const require_truncation = require("./truncation.cjs");
 const require_providers = require("./providers.cjs");
 const require_init = require("./init.cjs");
 let _langchain_core_messages = require("@langchain/core/messages");
@@ -137,10 +138,12 @@ async function attemptInvoke({ model, messages, provider, context, onChunk }, co
 		}
 		if (require_providers.manualToolStreamProviders.has(provider)) finalChunk = require_core.modifyDeltaProperties(provider, finalChunk);
 		if ((finalChunk?.tool_calls?.length ?? 0) > 0) finalChunk.tool_calls = finalChunk.tool_calls?.filter((tool_call) => !!tool_call.name);
+		require_truncation.assertNotTruncatedToolCall(finalChunk, provider);
 		return { messages: [finalChunk] };
 	}
 	const finalMessage = await model.invoke(messagesForProvider, config);
 	if ((finalMessage.tool_calls?.length ?? 0) > 0) finalMessage.tool_calls = finalMessage.tool_calls?.filter((tool_call) => !!tool_call.name);
+	require_truncation.assertNotTruncatedToolCall(finalMessage, provider);
 	return { messages: [finalMessage] };
 }
 /**

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

@@ -1 +1 @@
-{"version":3,"file":"invoke.cjs","names":["ChatModelStreamHandler","AIMessageChunk","annotateMessagesForLLM","manualToolStreamProviders","modifyDeltaProperties","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 * as t from '@/types';\nimport { annotateMessagesForLLM } from '@/tools/toolOutputReferences';\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    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  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":";;;;;;;;;;;AAqDA,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,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,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,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;EACZ;CACF;CAEF,IAAI,cAAc,KAAA,GAChB,MAAM;AAGV"}
+{"version":3,"file":"invoke.cjs","names":["ChatModelStreamHandler","AIMessageChunk","annotateMessagesForLLM","manualToolStreamProviders","modifyDeltaProperties","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 * 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,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;;;;;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,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;EACZ;CACF;CAEF,IAAI,cAAc,KAAA,GAChB,MAAM;AAGV"}

package/dist/cjs/llm/truncation.cjs

@@ -0,0 +1,110 @@
+require("../common/enum.cjs");
+require("../common/index.cjs");
+//#region src/llm/truncation.ts
+/**
+* Providers whose adapters deliver tool calls as complete, atomic objects
+* rather than streamed argument deltas. Google/Vertex (GenAI) emit each
+* `functionCall` whole and seal it on arrival, so a `MAX_TOKENS` finish does
+* NOT imply the arguments were cut off — the truncation guard must skip them.
+*/
+const ATOMIC_TOOL_CALL_ARG_PROVIDERS = new Set(["google", "vertexai"]);
+const MAX_TOKEN_VALUES = new Set([
+	"max_tokens",
+	"max_token",
+	"maxtokens",
+	"max_output_tokens"
+]);
+const LENGTH_VALUES = new Set(["length"]);
+function normalizeStopValue(value) {
+	if (typeof value !== "string") return null;
+	const normalized = value.trim().toLowerCase();
+	if (MAX_TOKEN_VALUES.has(normalized)) return "max_tokens";
+	if (LENGTH_VALUES.has(normalized)) return "length";
+	return null;
+}
+/**
+* Reads the truncation stop reason off a message, covering every provider
+* shape we stream:
+* - `response_metadata.stopReason` (Bedrock Converse, non-streaming)
+* - `response_metadata.messageStop.stopReason` (Bedrock Converse streaming)
+* - `response_metadata.stop_reason` (Anthropic, non-streaming)
+* - `additional_kwargs.stop_reason` (Anthropic streaming `message_delta`)
+* - `response_metadata.finish_reason` (OpenAI chat-completions / compatible)
+* - `response_metadata.incomplete_details.reason` (OpenAI Responses API)
+* - `response_metadata.finishReason` (Google)
+*
+* Returns the normalized reason when the model stopped because it hit the
+* output token ceiling, otherwise `null`.
+*/
+function getTruncationStopReason(message) {
+	const meta = message?.response_metadata;
+	const additionalKwargs = message?.additional_kwargs;
+	if (meta == null && additionalKwargs == null) return null;
+	const messageStop = meta?.messageStop;
+	const incompleteDetails = meta?.incomplete_details;
+	const candidates = [
+		meta?.stopReason,
+		meta?.stop_reason,
+		meta?.finish_reason,
+		meta?.finishReason,
+		messageStop?.stopReason,
+		incompleteDetails?.reason,
+		additionalKwargs?.stop_reason
+	];
+	for (const candidate of candidates) {
+		const normalized = normalizeStopValue(candidate);
+		if (normalized != null) return normalized;
+	}
+	return null;
+}
+function hasOpenToolCall(message) {
+	const m = message;
+	return (m.tool_calls?.length ?? 0) > 0 || (m.tool_call_chunks?.length ?? 0) > 0 || (m.invalid_tool_calls?.length ?? 0) > 0;
+}
+/**
+* Error raised when a model turn is cut off by the output token limit while it
+* was still emitting a tool call. The arguments are necessarily incomplete, so
+* executing or re-prompting the call would loop on a malformed request. Failing
+* fast with this surfaces an actionable message to the caller instead.
+*/
+var OutputTruncationError = class extends Error {
+	stopReason;
+	toolCallNames;
+	constructor(stopReason, toolCallNames) {
+		const named = toolCallNames.length > 0 ? ` (tool call: ${toolCallNames.join(", ")})` : "";
+		super(`The model response was truncated at the maximum output token limit before the tool call arguments were complete${named}. Increase the model's max output tokens, or have the model produce smaller arguments — for large files, write a lean main file and move bulky content into separate files.`);
+		this.name = "OutputTruncationError";
+		this.stopReason = stopReason;
+		this.toolCallNames = toolCallNames;
+	}
+};
+function collectToolCallNames(message) {
+	const m = message;
+	const names = [
+		...m.tool_calls ?? [],
+		...m.tool_call_chunks ?? [],
+		...m.invalid_tool_calls ?? []
+	].map((tc) => tc.name).filter((name) => typeof name === "string" && name !== "");
+	return [...new Set(names)];
+}
+/**
+* Throws {@link OutputTruncationError} when `message` was truncated by the
+* output token limit AND still carries a tool call. For providers that stream
+* tool arguments incrementally (Anthropic, Bedrock, OpenAI), a tool call
+* emitted under truncation has incomplete arguments — the tool-use block is the
+* last thing the model streams — so letting it through makes the agent loop on a
+* malformed call. No-ops for normal completions, truncated plain-text turns, and
+* providers that deliver complete tool calls atomically (Google/Vertex), where
+* `MAX_TOKENS` does not imply the arguments were cut off.
+*/
+function assertNotTruncatedToolCall(message, provider) {
+	if (message == null) return;
+	if (provider != null && ATOMIC_TOOL_CALL_ARG_PROVIDERS.has(provider)) return;
+	const stopReason = getTruncationStopReason(message);
+	if (stopReason == null || !hasOpenToolCall(message)) return;
+	throw new OutputTruncationError(stopReason, collectToolCallNames(message));
+}
+//#endregion
+exports.assertNotTruncatedToolCall = assertNotTruncatedToolCall;
+
+//# sourceMappingURL=truncation.cjs.map

package/dist/cjs/llm/truncation.cjs.map

@@ -0,0 +1 @@
+{"version":3,"file":"truncation.cjs","names":[],"sources":["../../../src/llm/truncation.ts"],"sourcesContent":["import type { AIMessageChunk, BaseMessage } from '@langchain/core/messages';\nimport { Providers } from '@/common';\n\n/**\n * Providers whose adapters deliver tool calls as complete, atomic objects\n * rather than streamed argument deltas. Google/Vertex (GenAI) emit each\n * `functionCall` whole and seal it on arrival, so a `MAX_TOKENS` finish does\n * NOT imply the arguments were cut off — the truncation guard must skip them.\n */\nconst ATOMIC_TOOL_CALL_ARG_PROVIDERS = new Set<Providers>([\n  Providers.GOOGLE,\n  Providers.VERTEXAI,\n]);\n\n/**\n * Normalized truncation reasons across providers. Providers disagree on the\n * key and the value: Anthropic/Bedrock use `max_tokens`, OpenAI uses `length`,\n * Google uses `MAX_TOKENS`.\n */\nexport type TruncationStopReason = 'max_tokens' | 'length';\n\nconst MAX_TOKEN_VALUES = new Set([\n  'max_tokens',\n  'max_token',\n  'maxtokens',\n  'max_output_tokens',\n]);\nconst LENGTH_VALUES = new Set(['length']);\n\nfunction normalizeStopValue(value: unknown): TruncationStopReason | null {\n  if (typeof value !== 'string') {\n    return null;\n  }\n  const normalized = value.trim().toLowerCase();\n  if (MAX_TOKEN_VALUES.has(normalized)) {\n    return 'max_tokens';\n  }\n  if (LENGTH_VALUES.has(normalized)) {\n    return 'length';\n  }\n  return null;\n}\n\n/**\n * Reads the truncation stop reason off a message, covering every provider\n * shape we stream:\n * - `response_metadata.stopReason` (Bedrock Converse, non-streaming)\n * - `response_metadata.messageStop.stopReason` (Bedrock Converse streaming)\n * - `response_metadata.stop_reason` (Anthropic, non-streaming)\n * - `additional_kwargs.stop_reason` (Anthropic streaming `message_delta`)\n * - `response_metadata.finish_reason` (OpenAI chat-completions / compatible)\n * - `response_metadata.incomplete_details.reason` (OpenAI Responses API)\n * - `response_metadata.finishReason` (Google)\n *\n * Returns the normalized reason when the model stopped because it hit the\n * output token ceiling, otherwise `null`.\n */\nexport function getTruncationStopReason(\n  message:\n    | Pick<BaseMessage, 'response_metadata' | 'additional_kwargs'>\n    | undefined\n    | null\n): TruncationStopReason | null {\n  const meta = message?.response_metadata as\n    | Record<string, unknown>\n    | undefined;\n  const additionalKwargs = message?.additional_kwargs as\n    | Record<string, unknown>\n    | undefined;\n  if (meta == null && additionalKwargs == null) {\n    return null;\n  }\n\n  const messageStop = meta?.messageStop as { stopReason?: unknown } | undefined;\n  const incompleteDetails = meta?.incomplete_details as\n    | { reason?: unknown }\n    | undefined;\n  const candidates: unknown[] = [\n    meta?.stopReason,\n    meta?.stop_reason,\n    meta?.finish_reason,\n    meta?.finishReason,\n    messageStop?.stopReason,\n    incompleteDetails?.reason,\n    additionalKwargs?.stop_reason,\n  ];\n\n  for (const candidate of candidates) {\n    const normalized = normalizeStopValue(candidate);\n    if (normalized != null) {\n      return normalized;\n    }\n  }\n  return null;\n}\n\nfunction hasOpenToolCall(message: AIMessageChunk | BaseMessage): boolean {\n  const m = message as AIMessageChunk & {\n    tool_calls?: unknown[];\n    tool_call_chunks?: unknown[];\n    invalid_tool_calls?: unknown[];\n  };\n  return (\n    (m.tool_calls?.length ?? 0) > 0 ||\n    (m.tool_call_chunks?.length ?? 0) > 0 ||\n    (m.invalid_tool_calls?.length ?? 0) > 0\n  );\n}\n\n/**\n * Error raised when a model turn is cut off by the output token limit while it\n * was still emitting a tool call. The arguments are necessarily incomplete, so\n * executing or re-prompting the call would loop on a malformed request. Failing\n * fast with this surfaces an actionable message to the caller instead.\n */\nexport class OutputTruncationError extends Error {\n  readonly stopReason: TruncationStopReason;\n  readonly toolCallNames: string[];\n\n  constructor(stopReason: TruncationStopReason, toolCallNames: string[]) {\n    const named =\n      toolCallNames.length > 0\n        ? ` (tool call: ${toolCallNames.join(', ')})`\n        : '';\n    super(\n      'The model response was truncated at the maximum output token limit before ' +\n        `the tool call arguments were complete${named}. Increase the model's max ` +\n        'output tokens, or have the model produce smaller arguments — for large ' +\n        'files, write a lean main file and move bulky content into separate files.'\n    );\n    this.name = 'OutputTruncationError';\n    this.stopReason = stopReason;\n    this.toolCallNames = toolCallNames;\n  }\n}\n\nfunction collectToolCallNames(message: AIMessageChunk | BaseMessage): string[] {\n  const m = message as AIMessageChunk & {\n    tool_calls?: Array<{ name?: string }>;\n    tool_call_chunks?: Array<{ name?: string }>;\n    invalid_tool_calls?: Array<{ name?: string }>;\n  };\n  const names = [\n    ...(m.tool_calls ?? []),\n    ...(m.tool_call_chunks ?? []),\n    ...(m.invalid_tool_calls ?? []),\n  ]\n    .map((tc) => tc.name)\n    .filter((name): name is string => typeof name === 'string' && name !== '');\n  return [...new Set(names)];\n}\n\n/**\n * Throws {@link OutputTruncationError} when `message` was truncated by the\n * output token limit AND still carries a tool call. For providers that stream\n * tool arguments incrementally (Anthropic, Bedrock, OpenAI), a tool call\n * emitted under truncation has incomplete arguments — the tool-use block is the\n * last thing the model streams — so letting it through makes the agent loop on a\n * malformed call. No-ops for normal completions, truncated plain-text turns, and\n * providers that deliver complete tool calls atomically (Google/Vertex), where\n * `MAX_TOKENS` does not imply the arguments were cut off.\n */\nexport function assertNotTruncatedToolCall(\n  message: AIMessageChunk | BaseMessage | undefined | null,\n  provider?: Providers\n): void {\n  if (message == null) {\n    return;\n  }\n  if (provider != null && ATOMIC_TOOL_CALL_ARG_PROVIDERS.has(provider)) {\n    return;\n  }\n  const stopReason = getTruncationStopReason(message);\n  if (stopReason == null || !hasOpenToolCall(message)) {\n    return;\n  }\n  throw new OutputTruncationError(stopReason, collectToolCallNames(message));\n}\n"],"mappings":";;;;;;;;;AASA,MAAM,iCAAiC,IAAI,IAAe,CAAA,UAAA,UAG1D,CAAC;AASD,MAAM,mBAAmB,IAAI,IAAI;CAC/B;CACA;CACA;CACA;AACF,CAAC;AACD,MAAM,gBAAgB,IAAI,IAAI,CAAC,QAAQ,CAAC;AAExC,SAAS,mBAAmB,OAA6C;CACvE,IAAI,OAAO,UAAU,UACnB,OAAO;CAET,MAAM,aAAa,MAAM,KAAK,CAAC,CAAC,YAAY;CAC5C,IAAI,iBAAiB,IAAI,UAAU,GACjC,OAAO;CAET,IAAI,cAAc,IAAI,UAAU,GAC9B,OAAO;CAET,OAAO;AACT;;;;;;;;;;;;;;;AAgBA,SAAgB,wBACd,SAI6B;CAC7B,MAAM,OAAO,SAAS;CAGtB,MAAM,mBAAmB,SAAS;CAGlC,IAAI,QAAQ,QAAQ,oBAAoB,MACtC,OAAO;CAGT,MAAM,cAAc,MAAM;CAC1B,MAAM,oBAAoB,MAAM;CAGhC,MAAM,aAAwB;EAC5B,MAAM;EACN,MAAM;EACN,MAAM;EACN,MAAM;EACN,aAAa;EACb,mBAAmB;EACnB,kBAAkB;CACpB;CAEA,KAAK,MAAM,aAAa,YAAY;EAClC,MAAM,aAAa,mBAAmB,SAAS;EAC/C,IAAI,cAAc,MAChB,OAAO;CAEX;CACA,OAAO;AACT;AAEA,SAAS,gBAAgB,SAAgD;CACvE,MAAM,IAAI;CAKV,QACG,EAAE,YAAY,UAAU,KAAK,MAC7B,EAAE,kBAAkB,UAAU,KAAK,MACnC,EAAE,oBAAoB,UAAU,KAAK;AAE1C;;;;;;;AAQA,IAAa,wBAAb,cAA2C,MAAM;CAC/C;CACA;CAEA,YAAY,YAAkC,eAAyB;EACrE,MAAM,QACJ,cAAc,SAAS,IACnB,gBAAgB,cAAc,KAAK,IAAI,EAAE,KACzC;EACN,MACE,kHAC0C,MAAM,4KAGlD;EACA,KAAK,OAAO;EACZ,KAAK,aAAa;EAClB,KAAK,gBAAgB;CACvB;AACF;AAEA,SAAS,qBAAqB,SAAiD;CAC7E,MAAM,IAAI;CAKV,MAAM,QAAQ;EACZ,GAAI,EAAE,cAAc,CAAC;EACrB,GAAI,EAAE,oBAAoB,CAAC;EAC3B,GAAI,EAAE,sBAAsB,CAAC;CAC/B,CAAC,CACE,KAAK,OAAO,GAAG,IAAI,CAAC,CACpB,QAAQ,SAAyB,OAAO,SAAS,YAAY,SAAS,EAAE;CAC3E,OAAO,CAAC,GAAG,IAAI,IAAI,KAAK,CAAC;AAC3B;;;;;;;;;;;AAYA,SAAgB,2BACd,SACA,UACM;CACN,IAAI,WAAW,MACb;CAEF,IAAI,YAAY,QAAQ,+BAA+B,IAAI,QAAQ,GACjE;CAEF,MAAM,aAAa,wBAAwB,OAAO;CAClD,IAAI,cAAc,QAAQ,CAAC,gBAAgB,OAAO,GAChD;CAEF,MAAM,IAAI,sBAAsB,YAAY,qBAAqB,OAAO,CAAC;AAC3E"}

package/dist/esm/llm/invoke.mjs

@@ -4,6 +4,7 @@ import { modifyDeltaProperties } from "../messages/core.mjs";
 import "../messages/index.mjs";
 import { annotateMessagesForLLM } from "../tools/toolOutputReferences.mjs";
 import { ChatModelStreamHandler } from "../stream.mjs";
+import { assertNotTruncatedToolCall } from "./truncation.mjs";
 import { manualToolStreamProviders } from "./providers.mjs";
 import { initializeModel } from "./init.mjs";
 import { AIMessageChunk } from "@langchain/core/messages";
@@ -137,10 +138,12 @@ async function attemptInvoke({ model, messages, provider, context, onChunk }, co
 		}
 		if (manualToolStreamProviders.has(provider)) finalChunk = modifyDeltaProperties(provider, finalChunk);
 		if ((finalChunk?.tool_calls?.length ?? 0) > 0) finalChunk.tool_calls = finalChunk.tool_calls?.filter((tool_call) => !!tool_call.name);
+		assertNotTruncatedToolCall(finalChunk, provider);
 		return { messages: [finalChunk] };
 	}
 	const finalMessage = await model.invoke(messagesForProvider, config);
 	if ((finalMessage.tool_calls?.length ?? 0) > 0) finalMessage.tool_calls = finalMessage.tool_calls?.filter((tool_call) => !!tool_call.name);
+	assertNotTruncatedToolCall(finalMessage, provider);
 	return { messages: [finalMessage] };
 }
 /**

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 { 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    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  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":";;;;;;;;;;;AAqDA,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,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,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 * 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"}

package/dist/esm/llm/truncation.mjs

@@ -0,0 +1,110 @@
+import "../common/enum.mjs";
+import "../common/index.mjs";
+//#region src/llm/truncation.ts
+/**
+* Providers whose adapters deliver tool calls as complete, atomic objects
+* rather than streamed argument deltas. Google/Vertex (GenAI) emit each
+* `functionCall` whole and seal it on arrival, so a `MAX_TOKENS` finish does
+* NOT imply the arguments were cut off — the truncation guard must skip them.
+*/
+const ATOMIC_TOOL_CALL_ARG_PROVIDERS = new Set(["google", "vertexai"]);
+const MAX_TOKEN_VALUES = new Set([
+	"max_tokens",
+	"max_token",
+	"maxtokens",
+	"max_output_tokens"
+]);
+const LENGTH_VALUES = new Set(["length"]);
+function normalizeStopValue(value) {
+	if (typeof value !== "string") return null;
+	const normalized = value.trim().toLowerCase();
+	if (MAX_TOKEN_VALUES.has(normalized)) return "max_tokens";
+	if (LENGTH_VALUES.has(normalized)) return "length";
+	return null;
+}
+/**
+* Reads the truncation stop reason off a message, covering every provider
+* shape we stream:
+* - `response_metadata.stopReason` (Bedrock Converse, non-streaming)
+* - `response_metadata.messageStop.stopReason` (Bedrock Converse streaming)
+* - `response_metadata.stop_reason` (Anthropic, non-streaming)
+* - `additional_kwargs.stop_reason` (Anthropic streaming `message_delta`)
+* - `response_metadata.finish_reason` (OpenAI chat-completions / compatible)
+* - `response_metadata.incomplete_details.reason` (OpenAI Responses API)
+* - `response_metadata.finishReason` (Google)
+*
+* Returns the normalized reason when the model stopped because it hit the
+* output token ceiling, otherwise `null`.
+*/
+function getTruncationStopReason(message) {
+	const meta = message?.response_metadata;
+	const additionalKwargs = message?.additional_kwargs;
+	if (meta == null && additionalKwargs == null) return null;
+	const messageStop = meta?.messageStop;
+	const incompleteDetails = meta?.incomplete_details;
+	const candidates = [
+		meta?.stopReason,
+		meta?.stop_reason,
+		meta?.finish_reason,
+		meta?.finishReason,
+		messageStop?.stopReason,
+		incompleteDetails?.reason,
+		additionalKwargs?.stop_reason
+	];
+	for (const candidate of candidates) {
+		const normalized = normalizeStopValue(candidate);
+		if (normalized != null) return normalized;
+	}
+	return null;
+}
+function hasOpenToolCall(message) {
+	const m = message;
+	return (m.tool_calls?.length ?? 0) > 0 || (m.tool_call_chunks?.length ?? 0) > 0 || (m.invalid_tool_calls?.length ?? 0) > 0;
+}
+/**
+* Error raised when a model turn is cut off by the output token limit while it
+* was still emitting a tool call. The arguments are necessarily incomplete, so
+* executing or re-prompting the call would loop on a malformed request. Failing
+* fast with this surfaces an actionable message to the caller instead.
+*/
+var OutputTruncationError = class extends Error {
+	stopReason;
+	toolCallNames;
+	constructor(stopReason, toolCallNames) {
+		const named = toolCallNames.length > 0 ? ` (tool call: ${toolCallNames.join(", ")})` : "";
+		super(`The model response was truncated at the maximum output token limit before the tool call arguments were complete${named}. Increase the model's max output tokens, or have the model produce smaller arguments — for large files, write a lean main file and move bulky content into separate files.`);
+		this.name = "OutputTruncationError";
+		this.stopReason = stopReason;
+		this.toolCallNames = toolCallNames;
+	}
+};
+function collectToolCallNames(message) {
+	const m = message;
+	const names = [
+		...m.tool_calls ?? [],
+		...m.tool_call_chunks ?? [],
+		...m.invalid_tool_calls ?? []
+	].map((tc) => tc.name).filter((name) => typeof name === "string" && name !== "");
+	return [...new Set(names)];
+}
+/**
+* Throws {@link OutputTruncationError} when `message` was truncated by the
+* output token limit AND still carries a tool call. For providers that stream
+* tool arguments incrementally (Anthropic, Bedrock, OpenAI), a tool call
+* emitted under truncation has incomplete arguments — the tool-use block is the
+* last thing the model streams — so letting it through makes the agent loop on a
+* malformed call. No-ops for normal completions, truncated plain-text turns, and
+* providers that deliver complete tool calls atomically (Google/Vertex), where
+* `MAX_TOKENS` does not imply the arguments were cut off.
+*/
+function assertNotTruncatedToolCall(message, provider) {
+	if (message == null) return;
+	if (provider != null && ATOMIC_TOOL_CALL_ARG_PROVIDERS.has(provider)) return;
+	const stopReason = getTruncationStopReason(message);
+	if (stopReason == null || !hasOpenToolCall(message)) return;
+	throw new OutputTruncationError(stopReason, collectToolCallNames(message));
+}
+//#endregion
+export { assertNotTruncatedToolCall };
+
+//# sourceMappingURL=truncation.mjs.map

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

@@ -0,0 +1 @@
+{"version":3,"file":"truncation.mjs","names":[],"sources":["../../../src/llm/truncation.ts"],"sourcesContent":["import type { AIMessageChunk, BaseMessage } from '@langchain/core/messages';\nimport { Providers } from '@/common';\n\n/**\n * Providers whose adapters deliver tool calls as complete, atomic objects\n * rather than streamed argument deltas. Google/Vertex (GenAI) emit each\n * `functionCall` whole and seal it on arrival, so a `MAX_TOKENS` finish does\n * NOT imply the arguments were cut off — the truncation guard must skip them.\n */\nconst ATOMIC_TOOL_CALL_ARG_PROVIDERS = new Set<Providers>([\n  Providers.GOOGLE,\n  Providers.VERTEXAI,\n]);\n\n/**\n * Normalized truncation reasons across providers. Providers disagree on the\n * key and the value: Anthropic/Bedrock use `max_tokens`, OpenAI uses `length`,\n * Google uses `MAX_TOKENS`.\n */\nexport type TruncationStopReason = 'max_tokens' | 'length';\n\nconst MAX_TOKEN_VALUES = new Set([\n  'max_tokens',\n  'max_token',\n  'maxtokens',\n  'max_output_tokens',\n]);\nconst LENGTH_VALUES = new Set(['length']);\n\nfunction normalizeStopValue(value: unknown): TruncationStopReason | null {\n  if (typeof value !== 'string') {\n    return null;\n  }\n  const normalized = value.trim().toLowerCase();\n  if (MAX_TOKEN_VALUES.has(normalized)) {\n    return 'max_tokens';\n  }\n  if (LENGTH_VALUES.has(normalized)) {\n    return 'length';\n  }\n  return null;\n}\n\n/**\n * Reads the truncation stop reason off a message, covering every provider\n * shape we stream:\n * - `response_metadata.stopReason` (Bedrock Converse, non-streaming)\n * - `response_metadata.messageStop.stopReason` (Bedrock Converse streaming)\n * - `response_metadata.stop_reason` (Anthropic, non-streaming)\n * - `additional_kwargs.stop_reason` (Anthropic streaming `message_delta`)\n * - `response_metadata.finish_reason` (OpenAI chat-completions / compatible)\n * - `response_metadata.incomplete_details.reason` (OpenAI Responses API)\n * - `response_metadata.finishReason` (Google)\n *\n * Returns the normalized reason when the model stopped because it hit the\n * output token ceiling, otherwise `null`.\n */\nexport function getTruncationStopReason(\n  message:\n    | Pick<BaseMessage, 'response_metadata' | 'additional_kwargs'>\n    | undefined\n    | null\n): TruncationStopReason | null {\n  const meta = message?.response_metadata as\n    | Record<string, unknown>\n    | undefined;\n  const additionalKwargs = message?.additional_kwargs as\n    | Record<string, unknown>\n    | undefined;\n  if (meta == null && additionalKwargs == null) {\n    return null;\n  }\n\n  const messageStop = meta?.messageStop as { stopReason?: unknown } | undefined;\n  const incompleteDetails = meta?.incomplete_details as\n    | { reason?: unknown }\n    | undefined;\n  const candidates: unknown[] = [\n    meta?.stopReason,\n    meta?.stop_reason,\n    meta?.finish_reason,\n    meta?.finishReason,\n    messageStop?.stopReason,\n    incompleteDetails?.reason,\n    additionalKwargs?.stop_reason,\n  ];\n\n  for (const candidate of candidates) {\n    const normalized = normalizeStopValue(candidate);\n    if (normalized != null) {\n      return normalized;\n    }\n  }\n  return null;\n}\n\nfunction hasOpenToolCall(message: AIMessageChunk | BaseMessage): boolean {\n  const m = message as AIMessageChunk & {\n    tool_calls?: unknown[];\n    tool_call_chunks?: unknown[];\n    invalid_tool_calls?: unknown[];\n  };\n  return (\n    (m.tool_calls?.length ?? 0) > 0 ||\n    (m.tool_call_chunks?.length ?? 0) > 0 ||\n    (m.invalid_tool_calls?.length ?? 0) > 0\n  );\n}\n\n/**\n * Error raised when a model turn is cut off by the output token limit while it\n * was still emitting a tool call. The arguments are necessarily incomplete, so\n * executing or re-prompting the call would loop on a malformed request. Failing\n * fast with this surfaces an actionable message to the caller instead.\n */\nexport class OutputTruncationError extends Error {\n  readonly stopReason: TruncationStopReason;\n  readonly toolCallNames: string[];\n\n  constructor(stopReason: TruncationStopReason, toolCallNames: string[]) {\n    const named =\n      toolCallNames.length > 0\n        ? ` (tool call: ${toolCallNames.join(', ')})`\n        : '';\n    super(\n      'The model response was truncated at the maximum output token limit before ' +\n        `the tool call arguments were complete${named}. Increase the model's max ` +\n        'output tokens, or have the model produce smaller arguments — for large ' +\n        'files, write a lean main file and move bulky content into separate files.'\n    );\n    this.name = 'OutputTruncationError';\n    this.stopReason = stopReason;\n    this.toolCallNames = toolCallNames;\n  }\n}\n\nfunction collectToolCallNames(message: AIMessageChunk | BaseMessage): string[] {\n  const m = message as AIMessageChunk & {\n    tool_calls?: Array<{ name?: string }>;\n    tool_call_chunks?: Array<{ name?: string }>;\n    invalid_tool_calls?: Array<{ name?: string }>;\n  };\n  const names = [\n    ...(m.tool_calls ?? []),\n    ...(m.tool_call_chunks ?? []),\n    ...(m.invalid_tool_calls ?? []),\n  ]\n    .map((tc) => tc.name)\n    .filter((name): name is string => typeof name === 'string' && name !== '');\n  return [...new Set(names)];\n}\n\n/**\n * Throws {@link OutputTruncationError} when `message` was truncated by the\n * output token limit AND still carries a tool call. For providers that stream\n * tool arguments incrementally (Anthropic, Bedrock, OpenAI), a tool call\n * emitted under truncation has incomplete arguments — the tool-use block is the\n * last thing the model streams — so letting it through makes the agent loop on a\n * malformed call. No-ops for normal completions, truncated plain-text turns, and\n * providers that deliver complete tool calls atomically (Google/Vertex), where\n * `MAX_TOKENS` does not imply the arguments were cut off.\n */\nexport function assertNotTruncatedToolCall(\n  message: AIMessageChunk | BaseMessage | undefined | null,\n  provider?: Providers\n): void {\n  if (message == null) {\n    return;\n  }\n  if (provider != null && ATOMIC_TOOL_CALL_ARG_PROVIDERS.has(provider)) {\n    return;\n  }\n  const stopReason = getTruncationStopReason(message);\n  if (stopReason == null || !hasOpenToolCall(message)) {\n    return;\n  }\n  throw new OutputTruncationError(stopReason, collectToolCallNames(message));\n}\n"],"mappings":";;;;;;;;;AASA,MAAM,iCAAiC,IAAI,IAAe,CAAA,UAAA,UAG1D,CAAC;AASD,MAAM,mBAAmB,IAAI,IAAI;CAC/B;CACA;CACA;CACA;AACF,CAAC;AACD,MAAM,gBAAgB,IAAI,IAAI,CAAC,QAAQ,CAAC;AAExC,SAAS,mBAAmB,OAA6C;CACvE,IAAI,OAAO,UAAU,UACnB,OAAO;CAET,MAAM,aAAa,MAAM,KAAK,CAAC,CAAC,YAAY;CAC5C,IAAI,iBAAiB,IAAI,UAAU,GACjC,OAAO;CAET,IAAI,cAAc,IAAI,UAAU,GAC9B,OAAO;CAET,OAAO;AACT;;;;;;;;;;;;;;;AAgBA,SAAgB,wBACd,SAI6B;CAC7B,MAAM,OAAO,SAAS;CAGtB,MAAM,mBAAmB,SAAS;CAGlC,IAAI,QAAQ,QAAQ,oBAAoB,MACtC,OAAO;CAGT,MAAM,cAAc,MAAM;CAC1B,MAAM,oBAAoB,MAAM;CAGhC,MAAM,aAAwB;EAC5B,MAAM;EACN,MAAM;EACN,MAAM;EACN,MAAM;EACN,aAAa;EACb,mBAAmB;EACnB,kBAAkB;CACpB;CAEA,KAAK,MAAM,aAAa,YAAY;EAClC,MAAM,aAAa,mBAAmB,SAAS;EAC/C,IAAI,cAAc,MAChB,OAAO;CAEX;CACA,OAAO;AACT;AAEA,SAAS,gBAAgB,SAAgD;CACvE,MAAM,IAAI;CAKV,QACG,EAAE,YAAY,UAAU,KAAK,MAC7B,EAAE,kBAAkB,UAAU,KAAK,MACnC,EAAE,oBAAoB,UAAU,KAAK;AAE1C;;;;;;;AAQA,IAAa,wBAAb,cAA2C,MAAM;CAC/C;CACA;CAEA,YAAY,YAAkC,eAAyB;EACrE,MAAM,QACJ,cAAc,SAAS,IACnB,gBAAgB,cAAc,KAAK,IAAI,EAAE,KACzC;EACN,MACE,kHAC0C,MAAM,4KAGlD;EACA,KAAK,OAAO;EACZ,KAAK,aAAa;EAClB,KAAK,gBAAgB;CACvB;AACF;AAEA,SAAS,qBAAqB,SAAiD;CAC7E,MAAM,IAAI;CAKV,MAAM,QAAQ;EACZ,GAAI,EAAE,cAAc,CAAC;EACrB,GAAI,EAAE,oBAAoB,CAAC;EAC3B,GAAI,EAAE,sBAAsB,CAAC;CAC/B,CAAC,CACE,KAAK,OAAO,GAAG,IAAI,CAAC,CACpB,QAAQ,SAAyB,OAAO,SAAS,YAAY,SAAS,EAAE;CAC3E,OAAO,CAAC,GAAG,IAAI,IAAI,KAAK,CAAC;AAC3B;;;;;;;;;;;AAYA,SAAgB,2BACd,SACA,UACM;CACN,IAAI,WAAW,MACb;CAEF,IAAI,YAAY,QAAQ,+BAA+B,IAAI,QAAQ,GACjE;CAEF,MAAM,aAAa,wBAAwB,OAAO;CAClD,IAAI,cAAc,QAAQ,CAAC,gBAAgB,OAAO,GAChD;CAEF,MAAM,IAAI,sBAAsB,YAAY,qBAAqB,OAAO,CAAC;AAC3E"}

package/dist/types/llm/truncation.d.ts

@@ -0,0 +1,45 @@
+import type { AIMessageChunk, BaseMessage } from '@langchain/core/messages';
+import { Providers } from '@/common';
+/**
+ * Normalized truncation reasons across providers. Providers disagree on the
+ * key and the value: Anthropic/Bedrock use `max_tokens`, OpenAI uses `length`,
+ * Google uses `MAX_TOKENS`.
+ */
+export type TruncationStopReason = 'max_tokens' | 'length';
+/**
+ * Reads the truncation stop reason off a message, covering every provider
+ * shape we stream:
+ * - `response_metadata.stopReason` (Bedrock Converse, non-streaming)
+ * - `response_metadata.messageStop.stopReason` (Bedrock Converse streaming)
+ * - `response_metadata.stop_reason` (Anthropic, non-streaming)
+ * - `additional_kwargs.stop_reason` (Anthropic streaming `message_delta`)
+ * - `response_metadata.finish_reason` (OpenAI chat-completions / compatible)
+ * - `response_metadata.incomplete_details.reason` (OpenAI Responses API)
+ * - `response_metadata.finishReason` (Google)
+ *
+ * Returns the normalized reason when the model stopped because it hit the
+ * output token ceiling, otherwise `null`.
+ */
+export declare function getTruncationStopReason(message: Pick<BaseMessage, 'response_metadata' | 'additional_kwargs'> | undefined | null): TruncationStopReason | null;
+/**
+ * Error raised when a model turn is cut off by the output token limit while it
+ * was still emitting a tool call. The arguments are necessarily incomplete, so
+ * executing or re-prompting the call would loop on a malformed request. Failing
+ * fast with this surfaces an actionable message to the caller instead.
+ */
+export declare class OutputTruncationError extends Error {
+    readonly stopReason: TruncationStopReason;
+    readonly toolCallNames: string[];
+    constructor(stopReason: TruncationStopReason, toolCallNames: string[]);
+}
+/**
+ * Throws {@link OutputTruncationError} when `message` was truncated by the
+ * output token limit AND still carries a tool call. For providers that stream
+ * tool arguments incrementally (Anthropic, Bedrock, OpenAI), a tool call
+ * emitted under truncation has incomplete arguments — the tool-use block is the
+ * last thing the model streams — so letting it through makes the agent loop on a
+ * malformed call. No-ops for normal completions, truncated plain-text turns, and
+ * providers that deliver complete tool calls atomically (Google/Vertex), where
+ * `MAX_TOKENS` does not imply the arguments were cut off.
+ */
+export declare function assertNotTruncatedToolCall(message: AIMessageChunk | BaseMessage | undefined | null, provider?: Providers): void;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@librechat/agents",
-  "version": "3.2.54",
+  "version": "3.2.55",
   "main": "./dist/cjs/main.cjs",
   "module": "./dist/esm/main.mjs",
   "types": "./dist/types/index.d.ts",

package/src/llm/invoke.ts

@@ -6,6 +6,7 @@ import type { BaseMessage } from '@langchain/core/messages';
 import type { ToolOutputReferenceRegistry } from '@/tools/toolOutputReferences';
 import type * as t from '@/types';
 import { annotateMessagesForLLM } from '@/tools/toolOutputReferences';
+import { assertNotTruncatedToolCall } from '@/llm/truncation';
 import { Constants, GraphEvents, Providers } from '@/common';
 import { manualToolStreamProviders } from '@/llm/providers';
 import { modifyDeltaProperties } from '@/messages';
@@ -297,6 +298,7 @@ export async function attemptInvoke(
       );
     }
 
+    assertNotTruncatedToolCall(finalChunk, provider);
     return { messages: [finalChunk as AIMessageChunk] };
   }
 
@@ -306,6 +308,7 @@ export async function attemptInvoke(
       (tool_call: ToolCall) => !!tool_call.name
     );
   }
+  assertNotTruncatedToolCall(finalMessage, provider);
   return { messages: [finalMessage] };
 }
 

package/src/llm/truncation.test.ts

@@ -0,0 +1,242 @@
+import { describe, expect, it } from '@jest/globals';
+import { AIMessage, AIMessageChunk } from '@langchain/core/messages';
+import {
+  OutputTruncationError,
+  assertNotTruncatedToolCall,
+  getTruncationStopReason,
+} from '@/llm/truncation';
+import { Providers } from '@/common';
+
+describe('getTruncationStopReason', () => {
+  it('returns null when there is no response metadata', () => {
+    expect(getTruncationStopReason(new AIMessage('hi'))).toBeNull();
+    expect(getTruncationStopReason(undefined)).toBeNull();
+  });
+
+  it('detects Bedrock Converse streaming shape (messageStop.stopReason)', () => {
+    const message = new AIMessageChunk({
+      content: '',
+      response_metadata: { messageStop: { stopReason: 'max_tokens' } },
+    });
+    expect(getTruncationStopReason(message)).toBe('max_tokens');
+  });
+
+  it('detects Bedrock Converse non-streaming shape (stopReason)', () => {
+    const message = new AIMessage({
+      content: '',
+      response_metadata: { stopReason: 'max_tokens' },
+    });
+    expect(getTruncationStopReason(message)).toBe('max_tokens');
+  });
+
+  it('detects Anthropic shape (stop_reason)', () => {
+    const message = new AIMessage({
+      content: '',
+      response_metadata: { stop_reason: 'max_tokens' },
+    });
+    expect(getTruncationStopReason(message)).toBe('max_tokens');
+  });
+
+  it('detects OpenAI shape (finish_reason: length)', () => {
+    const message = new AIMessage({
+      content: '',
+      response_metadata: { finish_reason: 'length' },
+    });
+    expect(getTruncationStopReason(message)).toBe('length');
+  });
+
+  it('detects Google shape (finishReason: MAX_TOKENS, case-insensitive)', () => {
+    const message = new AIMessage({
+      content: '',
+      response_metadata: { finishReason: 'MAX_TOKENS' },
+    });
+    expect(getTruncationStopReason(message)).toBe('max_tokens');
+  });
+
+  it('detects Anthropic streaming shape (additional_kwargs.stop_reason)', () => {
+    const message = new AIMessageChunk({
+      content: '',
+      additional_kwargs: { stop_reason: 'max_tokens' },
+      response_metadata: { model_provider: 'anthropic' },
+    });
+    expect(getTruncationStopReason(message)).toBe('max_tokens');
+  });
+
+  it('detects OpenAI Responses shape (incomplete_details.reason)', () => {
+    const message = new AIMessage({
+      content: '',
+      response_metadata: {
+        status: 'incomplete',
+        incomplete_details: { reason: 'max_output_tokens' },
+      },
+    });
+    expect(getTruncationStopReason(message)).toBe('max_tokens');
+  });
+
+  it('returns null for a non-token incomplete reason (content_filter)', () => {
+    const message = new AIMessage({
+      content: '',
+      response_metadata: {
+        status: 'incomplete',
+        incomplete_details: { reason: 'content_filter' },
+      },
+    });
+    expect(getTruncationStopReason(message)).toBeNull();
+  });
+
+  it('returns null for normal stop reasons', () => {
+    expect(
+      getTruncationStopReason(
+        new AIMessage({
+          content: 'x',
+          response_metadata: { stopReason: 'end_turn' },
+        })
+      )
+    ).toBeNull();
+    expect(
+      getTruncationStopReason(
+        new AIMessage({
+          content: 'x',
+          response_metadata: { finish_reason: 'stop' },
+        })
+      )
+    ).toBeNull();
+    expect(
+      getTruncationStopReason(
+        new AIMessage({
+          content: 'x',
+          response_metadata: { stopReason: 'tool_use' },
+        })
+      )
+    ).toBeNull();
+  });
+});
+
+describe('assertNotTruncatedToolCall', () => {
+  it('no-ops for a normal completion', () => {
+    expect(() =>
+      assertNotTruncatedToolCall(
+        new AIMessage({
+          content: 'done',
+          response_metadata: { stopReason: 'end_turn' },
+        })
+      )
+    ).not.toThrow();
+  });
+
+  it('no-ops for a truncated plain-text turn (no tool calls)', () => {
+    expect(() =>
+      assertNotTruncatedToolCall(
+        new AIMessage({
+          content: 'partial...',
+          response_metadata: { stopReason: 'max_tokens' },
+        })
+      )
+    ).not.toThrow();
+  });
+
+  it('no-ops for a complete tool call (normal stop reason)', () => {
+    const message = new AIMessage({
+      content: '',
+      tool_calls: [
+        { id: '1', name: 'create_file', args: { path: 'a', content: 'b' } },
+      ],
+      response_metadata: { stopReason: 'tool_use' },
+    });
+    expect(() => assertNotTruncatedToolCall(message)).not.toThrow();
+  });
+
+  it('throws when a tool call is present and the turn was truncated', () => {
+    const message = new AIMessage({
+      content: '',
+      tool_calls: [{ id: '1', name: 'create_file', args: { path: 'a' } }],
+      response_metadata: { messageStop: { stopReason: 'max_tokens' } },
+    });
+    expect(() => assertNotTruncatedToolCall(message)).toThrow(
+      OutputTruncationError
+    );
+    try {
+      assertNotTruncatedToolCall(message);
+    } catch (err) {
+      expect(err).toBeInstanceOf(OutputTruncationError);
+      expect((err as OutputTruncationError).stopReason).toBe('max_tokens');
+      expect((err as OutputTruncationError).toolCallNames).toContain(
+        'create_file'
+      );
+      expect((err as OutputTruncationError).message).toMatch(
+        /max output tokens/i
+      );
+    }
+  });
+
+  it('throws when only incomplete tool_call_chunks survived truncation', () => {
+    const message = new AIMessageChunk({
+      content: '',
+      tool_call_chunks: [
+        {
+          name: 'create_file',
+          args: '{"path":"a"',
+          index: 0,
+          type: 'tool_call_chunk',
+        },
+      ],
+      response_metadata: { messageStop: { stopReason: 'max_tokens' } },
+    });
+    expect(() => assertNotTruncatedToolCall(message)).toThrow(
+      OutputTruncationError
+    );
+  });
+
+  it('still throws for a streaming-arg provider (Anthropic)', () => {
+    const message = new AIMessage({
+      content: '',
+      tool_calls: [{ id: '1', name: 'create_file', args: { path: 'a' } }],
+      response_metadata: { stop_reason: 'max_tokens' },
+    });
+    expect(() =>
+      assertNotTruncatedToolCall(message, Providers.ANTHROPIC)
+    ).toThrow(OutputTruncationError);
+  });
+
+  it('throws for an Anthropic streaming tool call truncated via additional_kwargs', () => {
+    const message = new AIMessageChunk({
+      content: '',
+      tool_calls: [{ id: '1', name: 'create_file', args: { path: 'a' } }],
+      additional_kwargs: { stop_reason: 'max_tokens' },
+      response_metadata: { model_provider: 'anthropic' },
+    });
+    expect(() =>
+      assertNotTruncatedToolCall(message, Providers.ANTHROPIC)
+    ).toThrow(OutputTruncationError);
+  });
+
+  it('throws for an OpenAI Responses tool call truncated via incomplete_details', () => {
+    const message = new AIMessage({
+      content: '',
+      tool_calls: [{ id: '1', name: 'create_file', args: { path: 'a' } }],
+      response_metadata: {
+        status: 'incomplete',
+        incomplete_details: { reason: 'max_output_tokens' },
+      },
+    });
+    expect(() => assertNotTruncatedToolCall(message, Providers.OPENAI)).toThrow(
+      OutputTruncationError
+    );
+  });
+
+  it('does not throw for providers that deliver complete tool calls (Google/Vertex)', () => {
+    const message = new AIMessage({
+      content: '',
+      tool_calls: [
+        { id: '1', name: 'create_file', args: { path: 'a', content: 'b' } },
+      ],
+      response_metadata: { finishReason: 'MAX_TOKENS' },
+    });
+    expect(() =>
+      assertNotTruncatedToolCall(message, Providers.GOOGLE)
+    ).not.toThrow();
+    expect(() =>
+      assertNotTruncatedToolCall(message, Providers.VERTEXAI)
+    ).not.toThrow();
+  });
+});

package/src/llm/truncation.ts

@@ -0,0 +1,178 @@
+import type { AIMessageChunk, BaseMessage } from '@langchain/core/messages';
+import { Providers } from '@/common';
+
+/**
+ * Providers whose adapters deliver tool calls as complete, atomic objects
+ * rather than streamed argument deltas. Google/Vertex (GenAI) emit each
+ * `functionCall` whole and seal it on arrival, so a `MAX_TOKENS` finish does
+ * NOT imply the arguments were cut off — the truncation guard must skip them.
+ */
+const ATOMIC_TOOL_CALL_ARG_PROVIDERS = new Set<Providers>([
+  Providers.GOOGLE,
+  Providers.VERTEXAI,
+]);
+
+/**
+ * Normalized truncation reasons across providers. Providers disagree on the
+ * key and the value: Anthropic/Bedrock use `max_tokens`, OpenAI uses `length`,
+ * Google uses `MAX_TOKENS`.
+ */
+export type TruncationStopReason = 'max_tokens' | 'length';
+
+const MAX_TOKEN_VALUES = new Set([
+  'max_tokens',
+  'max_token',
+  'maxtokens',
+  'max_output_tokens',
+]);
+const LENGTH_VALUES = new Set(['length']);
+
+function normalizeStopValue(value: unknown): TruncationStopReason | null {
+  if (typeof value !== 'string') {
+    return null;
+  }
+  const normalized = value.trim().toLowerCase();
+  if (MAX_TOKEN_VALUES.has(normalized)) {
+    return 'max_tokens';
+  }
+  if (LENGTH_VALUES.has(normalized)) {
+    return 'length';
+  }
+  return null;
+}
+
+/**
+ * Reads the truncation stop reason off a message, covering every provider
+ * shape we stream:
+ * - `response_metadata.stopReason` (Bedrock Converse, non-streaming)
+ * - `response_metadata.messageStop.stopReason` (Bedrock Converse streaming)
+ * - `response_metadata.stop_reason` (Anthropic, non-streaming)
+ * - `additional_kwargs.stop_reason` (Anthropic streaming `message_delta`)
+ * - `response_metadata.finish_reason` (OpenAI chat-completions / compatible)
+ * - `response_metadata.incomplete_details.reason` (OpenAI Responses API)
+ * - `response_metadata.finishReason` (Google)
+ *
+ * Returns the normalized reason when the model stopped because it hit the
+ * output token ceiling, otherwise `null`.
+ */
+export function getTruncationStopReason(
+  message:
+    | Pick<BaseMessage, 'response_metadata' | 'additional_kwargs'>
+    | undefined
+    | null
+): TruncationStopReason | null {
+  const meta = message?.response_metadata as
+    | Record<string, unknown>
+    | undefined;
+  const additionalKwargs = message?.additional_kwargs as
+    | Record<string, unknown>
+    | undefined;
+  if (meta == null && additionalKwargs == null) {
+    return null;
+  }
+
+  const messageStop = meta?.messageStop as { stopReason?: unknown } | undefined;
+  const incompleteDetails = meta?.incomplete_details as
+    | { reason?: unknown }
+    | undefined;
+  const candidates: unknown[] = [
+    meta?.stopReason,
+    meta?.stop_reason,
+    meta?.finish_reason,
+    meta?.finishReason,
+    messageStop?.stopReason,
+    incompleteDetails?.reason,
+    additionalKwargs?.stop_reason,
+  ];
+
+  for (const candidate of candidates) {
+    const normalized = normalizeStopValue(candidate);
+    if (normalized != null) {
+      return normalized;
+    }
+  }
+  return null;
+}
+
+function hasOpenToolCall(message: AIMessageChunk | BaseMessage): boolean {
+  const m = message as AIMessageChunk & {
+    tool_calls?: unknown[];
+    tool_call_chunks?: unknown[];
+    invalid_tool_calls?: unknown[];
+  };
+  return (
+    (m.tool_calls?.length ?? 0) > 0 ||
+    (m.tool_call_chunks?.length ?? 0) > 0 ||
+    (m.invalid_tool_calls?.length ?? 0) > 0
+  );
+}
+
+/**
+ * Error raised when a model turn is cut off by the output token limit while it
+ * was still emitting a tool call. The arguments are necessarily incomplete, so
+ * executing or re-prompting the call would loop on a malformed request. Failing
+ * fast with this surfaces an actionable message to the caller instead.
+ */
+export class OutputTruncationError extends Error {
+  readonly stopReason: TruncationStopReason;
+  readonly toolCallNames: string[];
+
+  constructor(stopReason: TruncationStopReason, toolCallNames: string[]) {
+    const named =
+      toolCallNames.length > 0
+        ? ` (tool call: ${toolCallNames.join(', ')})`
+        : '';
+    super(
+      'The model response was truncated at the maximum output token limit before ' +
+        `the tool call arguments were complete${named}. Increase the model's max ` +
+        'output tokens, or have the model produce smaller arguments — for large ' +
+        'files, write a lean main file and move bulky content into separate files.'
+    );
+    this.name = 'OutputTruncationError';
+    this.stopReason = stopReason;
+    this.toolCallNames = toolCallNames;
+  }
+}
+
+function collectToolCallNames(message: AIMessageChunk | BaseMessage): string[] {
+  const m = message as AIMessageChunk & {
+    tool_calls?: Array<{ name?: string }>;
+    tool_call_chunks?: Array<{ name?: string }>;
+    invalid_tool_calls?: Array<{ name?: string }>;
+  };
+  const names = [
+    ...(m.tool_calls ?? []),
+    ...(m.tool_call_chunks ?? []),
+    ...(m.invalid_tool_calls ?? []),
+  ]
+    .map((tc) => tc.name)
+    .filter((name): name is string => typeof name === 'string' && name !== '');
+  return [...new Set(names)];
+}
+
+/**
+ * Throws {@link OutputTruncationError} when `message` was truncated by the
+ * output token limit AND still carries a tool call. For providers that stream
+ * tool arguments incrementally (Anthropic, Bedrock, OpenAI), a tool call
+ * emitted under truncation has incomplete arguments — the tool-use block is the
+ * last thing the model streams — so letting it through makes the agent loop on a
+ * malformed call. No-ops for normal completions, truncated plain-text turns, and
+ * providers that deliver complete tool calls atomically (Google/Vertex), where
+ * `MAX_TOKENS` does not imply the arguments were cut off.
+ */
+export function assertNotTruncatedToolCall(
+  message: AIMessageChunk | BaseMessage | undefined | null,
+  provider?: Providers
+): void {
+  if (message == null) {
+    return;
+  }
+  if (provider != null && ATOMIC_TOOL_CALL_ARG_PROVIDERS.has(provider)) {
+    return;
+  }
+  const stopReason = getTruncationStopReason(message);
+  if (stopReason == null || !hasOpenToolCall(message)) {
+    return;
+  }
+  throw new OutputTruncationError(stopReason, collectToolCallNames(message));
+}

package/src/specs/bedrock-truncation.live.test.ts

@@ -0,0 +1,191 @@
+/**
+ * Live Bedrock tool-call truncation verification.
+ *
+ * Reproduces the production failure where a large tool argument (`content`)
+ * gets cut off by the max output token limit, leaving the handler with a
+ * well-formed-but-incomplete tool call and no truncation signal.
+ *
+ * Run with:
+ * RUN_BEDROCK_TRUNCATION_LIVE=1 npm test -- bedrock-truncation.live.test.ts --runInBand
+ */
+import { config as dotenvConfig } from 'dotenv';
+dotenvConfig();
+
+import { z } from 'zod';
+import { HumanMessage } from '@langchain/core/messages';
+import { DynamicStructuredTool } from '@langchain/core/tools';
+import { describe, expect, it, jest } from '@jest/globals';
+import type * as t from '@/types';
+import { GraphEvents, Providers } from '@/common';
+import { ChatModelStreamHandler } from '@/stream';
+import { ToolEndHandler, ModelEndHandler } from '@/events';
+import { Run } from '@/run';
+
+const shouldRunLive =
+  process.env.RUN_BEDROCK_TRUNCATION_LIVE === '1' &&
+  (process.env.BEDROCK_AWS_ACCESS_KEY_ID ?? '') !== '' &&
+  (process.env.BEDROCK_AWS_SECRET_ACCESS_KEY ?? '') !== '';
+
+const describeIfLive = shouldRunLive ? describe : describe.skip;
+
+const REGION =
+  [
+    process.env.BEDROCK_AWS_REGION,
+    process.env.BEDROCK_AWS_DEFAULT_REGION,
+    process.env.AWS_DEFAULT_REGION,
+  ].find(
+    (value): value is string => typeof value === 'string' && value !== ''
+  ) ?? 'us-east-1';
+const MODEL = 'us.anthropic.claude-sonnet-4-5-20250929-v1:0';
+const MAX_TOKENS = Number(process.env.REPRO_MAX_TOKENS ?? 350);
+
+const schema = z.object({
+  path: z.string().describe('Path to write.'),
+  content: z.string().describe('Complete file contents.'),
+  overwrite: z.boolean().optional(),
+});
+
+describeIfLive('Bedrock tool-call truncation (live)', () => {
+  jest.setTimeout(120_000);
+
+  it('fails fast with a truncation error instead of looping', async () => {
+    const received: Array<{ keys: string[]; contentLen: number | null }> = [];
+    const stepToolCalls: string[][] = [];
+    let invocations = 0;
+
+    const createFile = new DynamicStructuredTool({
+      name: 'create_file',
+      description: 'Create a new file. Requires path and content.',
+      schema,
+      func: async (input: z.infer<typeof schema>) => {
+        invocations += 1;
+        const hasContent =
+          typeof input.content === 'string' && input.content.length > 0;
+        received.push({
+          keys: Object.keys(input as Record<string, unknown>),
+          contentLen: hasContent ? input.content.length : null,
+        });
+        if (!hasContent) {
+          return 'Error: content is required\n Please fix your mistakes.';
+        }
+        return `Wrote ${input.content.length} bytes to ${input.path}`;
+      },
+    });
+
+    const llmConfig = {
+      provider: Providers.BEDROCK,
+      model: MODEL,
+      region: REGION,
+      credentials: {
+        accessKeyId: process.env.BEDROCK_AWS_ACCESS_KEY_ID!,
+        secretAccessKey: process.env.BEDROCK_AWS_SECRET_ACCESS_KEY!,
+      },
+      maxTokens: MAX_TOKENS,
+      streaming: true,
+      streamUsage: true,
+    } as t.LLMConfig;
+
+    const run = await Run.create<t.IState>({
+      runId: 'bedrock-truncation-live',
+      graphConfig: {
+        type: 'standard',
+        llmConfig,
+        tools: [createFile],
+        instructions: 'You are a ClickHouse assistant. Use tools when asked.',
+        maxContextTokens: 89000,
+      },
+      returnContent: true,
+      skipCleanup: true,
+      customHandlers: {
+        [GraphEvents.TOOL_END]: new ToolEndHandler(
+          async (toolEndData: t.ToolEndData) => {
+            const out = toolEndData.output as
+              | { content?: unknown; name?: string; status?: string }
+              | undefined;
+            const content =
+              typeof out?.content === 'string'
+                ? out.content.slice(0, 120)
+                : JSON.stringify(out?.content).slice(0, 120);
+            // eslint-disable-next-line no-console
+            console.log(
+              `TOOL_END name=${out?.name} status=${out?.status} content=${content}`
+            );
+          }
+        ),
+        [GraphEvents.CHAT_MODEL_END]: new ModelEndHandler(),
+        [GraphEvents.CHAT_MODEL_STREAM]: new ChatModelStreamHandler(),
+        [GraphEvents.ON_RUN_STEP]: {
+          handle: (_event: string, data: t.StreamEventData): void => {
+            const step = data as unknown as {
+              stepDetails?: {
+                type?: string;
+                tool_calls?: Array<{ name?: string; args?: unknown }>;
+              };
+            };
+            const tcs = step.stepDetails?.tool_calls;
+            if ((tcs?.length ?? 0) > 0 && tcs != null) {
+              for (const tc of tcs) {
+                if (tc.name === 'create_file') {
+                  stepToolCalls.push(
+                    typeof tc.args === 'string'
+                      ? ['<raw-string>']
+                      : Object.keys((tc.args ?? {}) as Record<string, unknown>)
+                  );
+                }
+              }
+              // eslint-disable-next-line no-console
+              console.log(
+                'ON_RUN_STEP tool_calls=',
+                JSON.stringify(
+                  tcs.map((tc) => ({
+                    name: tc.name,
+                    args:
+                      typeof tc.args === 'string'
+                        ? tc.args.slice(0, 80)
+                        : Object.keys(
+                            (tc.args ?? {}) as Record<string, unknown>
+                        ),
+                  }))
+                )
+              );
+            }
+          },
+        },
+      },
+    });
+
+    const prompt =
+      'Call create_file once: path="skills/clickhouse-demo/SKILL.md", content=a complete ' +
+      '3000-word ClickHouse guide (schema design, MergeTree, partitioning, 10 example queries). ' +
+      'Put the entire guide in the content argument.';
+
+    let loopError: string | null = null;
+    try {
+      await run.processStream(
+        { messages: [new HumanMessage(prompt)] },
+        {
+          configurable: { provider: Providers.BEDROCK, thread_id: 'trunc' },
+          version: 'v2',
+          recursionLimit: 8,
+        }
+      );
+    } catch (err) {
+      loopError = err instanceof Error ? err.message : String(err);
+    }
+
+    // eslint-disable-next-line no-console
+    console.log('LIVE RESULT', {
+      invocations,
+      received,
+      stepToolCalls,
+      loopError,
+    });
+
+    // The run fails fast with the actionable truncation error...
+    expect(loopError).toMatch(/truncated at the maximum output token limit/i);
+    // ...instead of looping until the recursion limit.
+    expect(loopError).not.toMatch(/Recursion limit/i);
+    // No truncated create_file call ever carried a usable `content` arg.
+    expect(stepToolCalls.every((keys) => !keys.includes('content'))).toBe(true);
+  });
+});