@librechat/agents 3.2.34 → 3.2.35

package/dist/esm/llm/invoke.mjs

@@ -75,6 +75,22 @@ async function attemptInvoke({ model, messages, provider, context, onChunk }, co
 	const registry = context?.getOrCreateToolOutputRegistry();
 	const runId = config?.configurable?.run_id;
 	const messagesForProvider = annotateMessagesForLLM(messages, registry, runId);
+	/**
+	* Stamp the provider that is ACTUALLY serving this invocation onto the
+	* callback metadata. `attemptInvoke` is the single funnel for primary,
+	* fallback, and summarization model calls, so consumers that need
+	* provider attribution per call (the subagent usage-capture handler)
+	* read this key instead of trusting static agent config — which is
+	* wrong for fallback-served calls — or `ls_provider` — which derived
+	* providers inherit from their base class.
+	*/
+	config = {
+		...config,
+		metadata: {
+			...config?.metadata ?? {},
+			["__invoked_provider"]: provider
+		}
+	};
 	if (model.stream) {
 		const stream = await model.stream(messagesForProvider, config);
 		let finalChunk;
@@ -88,7 +104,7 @@ async function attemptInvoke({ model, messages, provider, context, onChunk }, co
 			});
 		}
 		else if (registeredStreamHandler == null) {
-			const metadata = config?.metadata;
+			const metadata = config.metadata;
 			const streamHandler = new ChatModelStreamHandler();
 			for await (const chunk of stream) {
 				const handlingChunk = getStreamHandlingChunk({
@@ -104,7 +120,7 @@ async function attemptInvoke({ model, messages, provider, context, onChunk }, co
 				});
 			}
 		} else {
-			const metadata = config?.metadata;
+			const metadata = config.metadata;
 			for await (const chunk of stream) {
 				const handlingChunk = getStreamHandlingChunk({
 					current: finalChunk,
@@ -128,23 +144,48 @@ async function attemptInvoke({ model, messages, provider, context, onChunk }, co
 	return { messages: [finalMessage] };
 }
 /**
+* Best-effort read of the configured model name from client options.
+* Providers disagree on the key (`model` vs `modelName`).
+*/
+function extractClientOptionsModel(clientOptions) {
+	const options = clientOptions;
+	if (typeof options?.model === "string" && options.model !== "") return options.model;
+	if (typeof options?.modelName === "string" && options.modelName !== "") return options.modelName;
+}
+/**
 * Attempts each fallback provider in order until one succeeds.
 * Throws the last error if all fallbacks fail.
 */
 async function tryFallbackProviders({ fallbacks, tools, messages, config, primaryError, context, onChunk }) {
 	let lastError = primaryError;
 	for (const fb of fallbacks) try {
+		const fbModel = initializeModel({
+			provider: fb.provider,
+			clientOptions: fb.clientOptions,
+			tools
+		});
+		/**
+		* Stamp the fallback's configured model onto callback metadata so
+		* per-call attribution (subagent usage capture) doesn't fall back to
+		* the PRIMARY config's model when the provider reports no
+		* `ls_model_name`. The serving provider is stamped uniformly by
+		* `attemptInvoke` (`INVOKED_PROVIDER`).
+		*/
+		const fbModelName = extractClientOptionsModel(fb.clientOptions);
+		const fbConfig = fbModelName == null ? config : {
+			...config,
+			metadata: {
+				...config?.metadata ?? {},
+				["__invoked_model"]: fbModelName
+			}
+		};
 		return await attemptInvoke({
-			model: initializeModel({
-				provider: fb.provider,
-				clientOptions: fb.clientOptions,
-				tools
-			}),
+			model: fbModel,
 			messages,
 			provider: fb.provider,
 			context,
 			onChunk
-		}, config);
+		}, fbConfig);
 	} catch (e) {
 		lastError = e;
 		continue;

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 { manualToolStreamProviders } from '@/llm/providers';\nimport { modifyDeltaProperties } from '@/messages';\nimport { ChatModelStreamHandler } from '@/stream';\nimport { GraphEvents, Providers } from '@/common';\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  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 * 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      const result = await attemptInvoke(\n        {\n          model: fbModel as t.ChatModel,\n          messages,\n          provider: fb.provider,\n          context,\n          onChunk,\n        },\n        config\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;CAE5E,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,QAAQ;GACzB,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,QAAQ;GACzB,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,eAAsB,qBAAqB,EACzC,WACA,OACA,UACA,QACA,cACA,SACA,WASiD;CACjD,IAAI,YAAqB;CACzB,KAAK,MAAM,MAAM,WACf,IAAI;EAgBF,OAAO,MAVc,cACnB;GACE,OAPY,gBAAgB;IAC9B,UAAU,GAAG;IACb,eAAe,GAAG;IAClB;GACF,CAGiB;GACb;GACA,UAAU,GAAG;GACb;GACA;EACF,GACA,MACF;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 { 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"}

package/dist/esm/main.mjs

@@ -2,7 +2,7 @@ import { isPresent, unescapeObject } from "./utils/misc.mjs";
 import { ANTHROPIC_TOOL_TOKEN_MULTIPLIER, DEFAULT_TOOL_TOKEN_MULTIPLIER } from "./common/constants.mjs";
 import { CODE_EXECUTION_TOOLS, Callback, CommonEvents, Constants, ContentTypes, EnvVar, GraphEvents, GraphNodeActions, GraphNodeKeys, LOCAL_CODING_BUNDLE_NAMES, LOCAL_CODING_TOOL_NAMES, Providers, StepTypes, TitleMethod, ToolCallTypes } from "./common/enum.mjs";
 import "./common/index.mjs";
-import { TokenEncoderManager, createTokenCounter, encodingForModel, estimateAnthropicImageTokens, estimateOpenAIImageTokens, extractImageDimensions, getTokenCountForMessage } from "./utils/tokens.mjs";
+import { TokenEncoderManager, apportionTokenCounts, createTokenCounter, encodingForModel, estimateAnthropicImageTokens, estimateOpenAIImageTokens, extractImageDimensions, getTokenCountForMessage } from "./utils/tokens.mjs";
 import { convertMessagesToContent, findLastIndex, formatAnthropicArtifactContent, formatAnthropicMessage, formatArtifactPayload, getConverseOverrideMessage, modifyDeltaProperties } from "./messages/core.mjs";
 import { getMessageId } from "./messages/ids.mjs";
 import { HARD_MAX_TOOL_RESULT_CHARS, HARD_MAX_TOTAL_TOOL_OUTPUT_SIZE, calculateMaxToolResultChars, calculateMaxTotalToolOutputSize, truncateToolInput, truncateToolResultContent } from "./utils/truncation.mjs";
@@ -12,7 +12,7 @@ import { DEFAULT_RESERVE_RATIO, ORIGINAL_CONTENT_MAX_CHARS, calculateTotalTokens
 import { ensureThinkingBlockInMessages, formatAgentMessages, formatFromLangChain, formatLangChainMessages, formatMediaMessage, formatMessage, labelContentByAgent, shiftIndexTokenCountMap, withMessageRole } from "./messages/format.mjs";
 import { addBedrockCacheControl, addCacheControl, addCacheControlToStablePrefixMessages, stripAnthropicCacheControl, stripBedrockCacheControl } from "./messages/cache.mjs";
 import { makeIsDeferred, partitionAndMarkAnthropicToolCache } from "./messages/anthropicToolCache.mjs";
-import { formatContentStrings } from "./messages/content.mjs";
+import { formatContentStrings, isLegacyConvertible } from "./messages/content.mjs";
 import { extractToolDiscoveries, hasToolSearchInCurrentTurn } from "./messages/tools.mjs";
 import { REMOVE_ALL_MESSAGES, createRemoveAllMessage, messagesStateReducer } from "./messages/reducer.mjs";
 import { splitAtRecencyBoundary } from "./messages/recency.mjs";
@@ -96,4 +96,4 @@ import { Runnable, RunnableLambda, RunnableSequence } from "./langchain/runnable
 import { DynamicStructuredTool, StructuredTool, Tool, tool } from "./langchain/tools.mjs";
 import "./langchain/index.mjs";
 import { BaseCheckpointSaver, Command, INTERRUPT, MemorySaver, interrupt, isInterrupted } from "@langchain/langgraph";
-export { AIMessage, AIMessageChunk, ANTHROPIC_TOOL_TOKEN_MULTIPLIER, AgentSession, BASH_SHELL_GUIDANCE, BaseCheckpointSaver, BaseMessage, BaseMessageChunk, BashExecutionToolDefinition, BashExecutionToolDescription, BashExecutionToolName, BashExecutionToolSchema, BashProgrammaticToolCallingDefinition, BashProgrammaticToolCallingDescription, BashProgrammaticToolCallingName, BashProgrammaticToolCallingSchema, BashToolOutputReferencesGuide, CLOUDFLARE_BASH_CODING_TOOL_NAMES, CLOUDFLARE_CODING_TOOL_NAMES, CODE_ARTIFACT_PATH_GUIDANCE, CODE_EXECUTION_TOOLS, Calculator, CalculatorSchema, CalculatorToolDefinition, CalculatorToolDescription, CalculatorToolName, Callback, ChatModelStreamHandler, ChatOpenRouter, CloudflareBashExecutionToolDescription, CloudflareCodeExecutionToolDescription, CodeExecutionToolDefinition, CodeExecutionToolDescription, CodeExecutionToolName, CodeExecutionToolSchema, Command, CommonEvents, CompileCheckToolName, Constants, ContentTypes, CustomOpenAIClient, DATE_RANGE, DEFAULT_CONTEXT_PRUNING_SETTINGS, DEFAULT_COUNTRY_DESCRIPTION, DEFAULT_HOOK_TIMEOUT_MS, DEFAULT_QUERY_DESCRIPTION, DEFAULT_RESERVE_RATIO, DEFAULT_TOOL_TOKEN_MULTIPLIER, DynamicStructuredTool, EnvVar, FAILED_EXECUTION_FILE_REMINDER, FakeChatModel, Graph, GraphEvents, GraphNodeActions, GraphNodeKeys, HARD_MAX_TOOL_RESULT_CHARS, HARD_MAX_TOTAL_TOOL_OUTPUT_SIZE, HOOK_EVENTS, HandlerRegistry, HookRegistry, HumanMessage, INTERRUPT, JsonlSessionStore, LLMStreamHandler, LOCAL_CODING_BUNDLE_NAMES, LOCAL_CODING_TOOL_NAMES, LOCAL_SPAWN_TIMEOUT_MS, LocalBashExecutionToolDescription, LocalCodeExecutionToolDescription, LocalEditFileToolName, LocalEditFileToolSchema, LocalFileCheckpointerImpl, LocalGlobSearchToolName, LocalGlobSearchToolSchema, LocalGrepSearchToolName, LocalGrepSearchToolSchema, LocalListDirectoryToolName, LocalListDirectoryToolSchema, LocalReadFileToolSchema, LocalWriteFileToolName, LocalWriteFileToolSchema, MAX_CACHE_SIZE, MAX_PATTERN_LENGTH, MemorySaver, ModelEndHandler, MultiAgentGraph, ORIGINAL_CONTENT_MAX_CHARS, ProgrammaticToolCallingDefinition, ProgrammaticToolCallingDescription, ProgrammaticToolCallingName, ProgrammaticToolCallingSchema, PromptTemplate, Providers, REMOVE_ALL_MESSAGES, ReadFileToolDefinition, ReadFileToolDescription, ReadFileToolName, ReadFileToolSchema, Run, Runnable, RunnableCallable, RunnableLambda, RunnableSequence, SEPARATORS, SessionManager, SkillToolDefinition, SkillToolDescription, SkillToolName, SkillToolSchema, SplitStreamHandler, StandardGraph, StepTypes, StructuredTool, SubagentExecutor, SubagentToolDefinition, SubagentToolDescription, SubagentToolName, SubagentToolSchema, SystemMessage, TMP_SCRATCH_OUTPUT_REMINDER, TestChatStreamHandler, TestLLMStreamHandler, TitleMethod, TokenEncoderManager, Tool, ToolCallTypes, ToolEndHandler, ToolMessage, ToolNode, ToolSearchToolDefinition, ToolSearchToolDescription, ToolSearchToolName, ToolSearchToolSchema, WebSearchToolDefinition, WebSearchToolDescription, WebSearchToolName, WebSearchToolSchema, _createBashProgramForTests, _resetLocalEngineWarningsForTests, _resetRipgrepCacheForTests, _resetSyntaxCheckProbeCacheForTests, _resetUnrecognizedTriggerWarnings, addBedrockCacheControl, addCacheControl, addCacheControlToStablePrefixMessages, appendCodeSessionFileSummary, appendFailedExecutionFileReminder, appendTmpScratchReminder, applyContextPruning, applyEdit, applyPreToolUseHooksForBridge, askUserQuestion, attemptInvoke, bashAstFindingsToErrors, buildBashExecutionToolDescription, buildChildInputs, buildCodeApiHttpErrorMessage, buildSandboxRuntimeConfig, buildSubagentToolParams, calculateMaxToolResultChars, calculateMaxTotalToolOutputSize, calculateTotalTokens, checkValidNumber, classifyAttachment, composeEventHandlers, convertMessagesToContent, countNestedGroups, countrySchema, createAgentSession, createBashExecutionTool, createBashProgrammaticToolCallingSchema, createBashProgrammaticToolCallingTool, createCloudflareBashExecutionTool, createCloudflareBashProgrammaticToolCallingTool, createCloudflareBridgeRuntime, createCloudflareCodeExecutionTool, createCloudflareCodingToolBundle, createCloudflareCodingTools, createCloudflareExecutionTool, createCloudflareLocalExecutionConfig, createCloudflareProgrammaticToolCallingTool, createCloudflareWorkspaceFS, createCodeExecutionTool, createCompileCheckTool, createCompileCheckToolDefinition, createContentAggregator, createFakeStreamingLLM, createHandlers, createLocalBashExecutionTool, createLocalBashProgrammaticToolCallingTool, createLocalCodeExecutionTool, createLocalCodingToolBundle, createLocalCodingToolDefinitions, createLocalCodingToolRegistry, createLocalCodingTools, createLocalEditFileTool, createLocalFileCheckpointer, createLocalGlobSearchTool, createLocalGrepSearchTool, createLocalListDirectoryTool, createLocalProgrammaticToolCallingTool, createLocalReadFileTool, createLocalWriteFileTool, createMetadataAggregator, createProgrammaticToolCallingSchema, createProgrammaticToolCallingTool, createPruneMessages, createRemoveAllMessage, createRunHandlers, createSchemaOnlyTool, createSchemaOnlyTools, createSearchTool, createSubagentToolDefinition, createTokenCounter, createToolPolicyHook, createToolSearch, createWorkspacePolicyHook, dateSchema, decodeFile, defaultOmitOptions, deserializeMessage, emptyOutputMessage, encodeFile, encodingForModel, enforceOriginalContentCap, ensureThinkingBlockInMessages, escapeRegexSpecialChars, estimateAnthropicImageTokens, estimateOpenAIImageTokens, executeCloudflareBash, executeCloudflareCode, executeHooks, executeLocalBash, executeLocalBashWithArgs, executeLocalCode, executeParallelSearches, executeTools, extractErrorMessage, extractImageDimensions, extractMcpServerName, extractTextFromContent, extractToolDiscoveries, extractUsedBashToolNames, extractUsedToolNames, fetchSessionFiles, filterBashToolsByUsage, filterSubagentResult, filterToolsByUsage, findLastIndex, formatAgentMessages, formatAnthropicArtifactContent, formatAnthropicMessage, formatArtifactPayload, formatCloudflareOutput, formatCompletedResponse, formatContentStrings, formatFromLangChain, formatLangChainMessages, formatMediaMessage, formatMessage, formatServerListing, formatSkillCatalog, getAvailableMcpServers, getBaseToolName, getBufferString, getChatModelClass, getChunkContent, getCloudflareWorkspaceRoot, getCodeBaseURL, getConverseOverrideMessage, getDeferredToolsListing, getLocalCwd, getLocalSessionId, getMaxOutputTokensKey, getMessageId, getMessagesWithinTokenLimit, getReadRoots, getSpawn, getTokenCountForMessage, getWorkspaceFS, getWorkspaceRoots, getWriteRoots, handleServerToolResult, handleToolCallChunks, handleToolCalls, hasNestedQuantifier, hasNestedQuantifiers, hasToolSearchInCurrentTurn, imageAttachmentContent, imagesSchema, initializeModel, interrupt, isAIMessage, isAnthropicLike, isBaseMessage, isContextOverflowError, isDangerousPattern, isFromAnyMcpServer, isFromMcpServer, isGoogleLike, isInterrupted, isLikelyContextOverflowError, isOpenAILike, isPresent, isThinkingEnabled, isToolMessage, isZodSchema, joinKeys, labelContentByAgent, locateEdit, makeIsDeferred, makeRequest, maskConsumedToolResults, matchesQuery, messagesStateReducer, modifyDeltaProperties, newsSchema, normalizeBashToolResultsForReplay, normalizeServerFilter, normalizeToBashIdentifier, normalizeToPythonIdentifier, partitionAndMarkAnthropicToolCache, performLocalSearch, preFlightTruncateToolCallInputs, preFlightTruncateToolResults, querySchema, repairOrphanedToolMessages, resetIfNotEmpty, resolveCloudflareSandbox, resolveCodeApiAuthHeaders, resolveContextPruningSettings, resolveLocalExecutionConfig, resolveLocalExecutionTools, resolveLocalToolRegistry, resolveLocalToolsForBinding, resolveSubagentConfigs, resolveWorkspacePath, resolveWorkspacePathSafe, runBashAstChecks, runPostEditSyntaxCheck, sanitizeOrphanToolBlocks, sanitizeRegex, serializeMessage, shellQuote, shiftIndexTokenCountMap, shouldTriggerSummarization, sleep, spawnLocalProcess, splitAtRecencyBoundary, stripAnthropicCacheControl, stripBedrockCacheControl, stripCodeSessionFileSummary, summarizeEvent, toJsonSchema, tool, toolResultTypes, toolsCondition, truncateLocalOutput, truncateToolInput, truncateToolResultContent, tryFallbackProviders, unescapeObject, unwrapToolResponse, validateBashCommand, validateCloudflareBashCommand, videosSchema, withMessageRole };
+export { AIMessage, AIMessageChunk, ANTHROPIC_TOOL_TOKEN_MULTIPLIER, AgentSession, BASH_SHELL_GUIDANCE, BaseCheckpointSaver, BaseMessage, BaseMessageChunk, BashExecutionToolDefinition, BashExecutionToolDescription, BashExecutionToolName, BashExecutionToolSchema, BashProgrammaticToolCallingDefinition, BashProgrammaticToolCallingDescription, BashProgrammaticToolCallingName, BashProgrammaticToolCallingSchema, BashToolOutputReferencesGuide, CLOUDFLARE_BASH_CODING_TOOL_NAMES, CLOUDFLARE_CODING_TOOL_NAMES, CODE_ARTIFACT_PATH_GUIDANCE, CODE_EXECUTION_TOOLS, Calculator, CalculatorSchema, CalculatorToolDefinition, CalculatorToolDescription, CalculatorToolName, Callback, ChatModelStreamHandler, ChatOpenRouter, CloudflareBashExecutionToolDescription, CloudflareCodeExecutionToolDescription, CodeExecutionToolDefinition, CodeExecutionToolDescription, CodeExecutionToolName, CodeExecutionToolSchema, Command, CommonEvents, CompileCheckToolName, Constants, ContentTypes, CustomOpenAIClient, DATE_RANGE, DEFAULT_CONTEXT_PRUNING_SETTINGS, DEFAULT_COUNTRY_DESCRIPTION, DEFAULT_HOOK_TIMEOUT_MS, DEFAULT_QUERY_DESCRIPTION, DEFAULT_RESERVE_RATIO, DEFAULT_TOOL_TOKEN_MULTIPLIER, DynamicStructuredTool, EnvVar, FAILED_EXECUTION_FILE_REMINDER, FakeChatModel, Graph, GraphEvents, GraphNodeActions, GraphNodeKeys, HARD_MAX_TOOL_RESULT_CHARS, HARD_MAX_TOTAL_TOOL_OUTPUT_SIZE, HOOK_EVENTS, HandlerRegistry, HookRegistry, HumanMessage, INTERRUPT, JsonlSessionStore, LLMStreamHandler, LOCAL_CODING_BUNDLE_NAMES, LOCAL_CODING_TOOL_NAMES, LOCAL_SPAWN_TIMEOUT_MS, LocalBashExecutionToolDescription, LocalCodeExecutionToolDescription, LocalEditFileToolName, LocalEditFileToolSchema, LocalFileCheckpointerImpl, LocalGlobSearchToolName, LocalGlobSearchToolSchema, LocalGrepSearchToolName, LocalGrepSearchToolSchema, LocalListDirectoryToolName, LocalListDirectoryToolSchema, LocalReadFileToolSchema, LocalWriteFileToolName, LocalWriteFileToolSchema, MAX_CACHE_SIZE, MAX_PATTERN_LENGTH, MemorySaver, ModelEndHandler, MultiAgentGraph, ORIGINAL_CONTENT_MAX_CHARS, ProgrammaticToolCallingDefinition, ProgrammaticToolCallingDescription, ProgrammaticToolCallingName, ProgrammaticToolCallingSchema, PromptTemplate, Providers, REMOVE_ALL_MESSAGES, ReadFileToolDefinition, ReadFileToolDescription, ReadFileToolName, ReadFileToolSchema, Run, Runnable, RunnableCallable, RunnableLambda, RunnableSequence, SEPARATORS, SessionManager, SkillToolDefinition, SkillToolDescription, SkillToolName, SkillToolSchema, SplitStreamHandler, StandardGraph, StepTypes, StructuredTool, SubagentExecutor, SubagentToolDefinition, SubagentToolDescription, SubagentToolName, SubagentToolSchema, SystemMessage, TMP_SCRATCH_OUTPUT_REMINDER, TestChatStreamHandler, TestLLMStreamHandler, TitleMethod, TokenEncoderManager, Tool, ToolCallTypes, ToolEndHandler, ToolMessage, ToolNode, ToolSearchToolDefinition, ToolSearchToolDescription, ToolSearchToolName, ToolSearchToolSchema, WebSearchToolDefinition, WebSearchToolDescription, WebSearchToolName, WebSearchToolSchema, _createBashProgramForTests, _resetLocalEngineWarningsForTests, _resetRipgrepCacheForTests, _resetSyntaxCheckProbeCacheForTests, _resetUnrecognizedTriggerWarnings, addBedrockCacheControl, addCacheControl, addCacheControlToStablePrefixMessages, appendCodeSessionFileSummary, appendFailedExecutionFileReminder, appendTmpScratchReminder, applyContextPruning, applyEdit, applyPreToolUseHooksForBridge, apportionTokenCounts, askUserQuestion, attemptInvoke, bashAstFindingsToErrors, buildBashExecutionToolDescription, buildChildInputs, buildCodeApiHttpErrorMessage, buildSandboxRuntimeConfig, buildSubagentToolParams, calculateMaxToolResultChars, calculateMaxTotalToolOutputSize, calculateTotalTokens, checkValidNumber, classifyAttachment, composeEventHandlers, convertMessagesToContent, countNestedGroups, countrySchema, createAgentSession, createBashExecutionTool, createBashProgrammaticToolCallingSchema, createBashProgrammaticToolCallingTool, createCloudflareBashExecutionTool, createCloudflareBashProgrammaticToolCallingTool, createCloudflareBridgeRuntime, createCloudflareCodeExecutionTool, createCloudflareCodingToolBundle, createCloudflareCodingTools, createCloudflareExecutionTool, createCloudflareLocalExecutionConfig, createCloudflareProgrammaticToolCallingTool, createCloudflareWorkspaceFS, createCodeExecutionTool, createCompileCheckTool, createCompileCheckToolDefinition, createContentAggregator, createFakeStreamingLLM, createHandlers, createLocalBashExecutionTool, createLocalBashProgrammaticToolCallingTool, createLocalCodeExecutionTool, createLocalCodingToolBundle, createLocalCodingToolDefinitions, createLocalCodingToolRegistry, createLocalCodingTools, createLocalEditFileTool, createLocalFileCheckpointer, createLocalGlobSearchTool, createLocalGrepSearchTool, createLocalListDirectoryTool, createLocalProgrammaticToolCallingTool, createLocalReadFileTool, createLocalWriteFileTool, createMetadataAggregator, createProgrammaticToolCallingSchema, createProgrammaticToolCallingTool, createPruneMessages, createRemoveAllMessage, createRunHandlers, createSchemaOnlyTool, createSchemaOnlyTools, createSearchTool, createSubagentToolDefinition, createTokenCounter, createToolPolicyHook, createToolSearch, createWorkspacePolicyHook, dateSchema, decodeFile, defaultOmitOptions, deserializeMessage, emptyOutputMessage, encodeFile, encodingForModel, enforceOriginalContentCap, ensureThinkingBlockInMessages, escapeRegexSpecialChars, estimateAnthropicImageTokens, estimateOpenAIImageTokens, executeCloudflareBash, executeCloudflareCode, executeHooks, executeLocalBash, executeLocalBashWithArgs, executeLocalCode, executeParallelSearches, executeTools, extractErrorMessage, extractImageDimensions, extractMcpServerName, extractTextFromContent, extractToolDiscoveries, extractUsedBashToolNames, extractUsedToolNames, fetchSessionFiles, filterBashToolsByUsage, filterSubagentResult, filterToolsByUsage, findLastIndex, formatAgentMessages, formatAnthropicArtifactContent, formatAnthropicMessage, formatArtifactPayload, formatCloudflareOutput, formatCompletedResponse, formatContentStrings, formatFromLangChain, formatLangChainMessages, formatMediaMessage, formatMessage, formatServerListing, formatSkillCatalog, getAvailableMcpServers, getBaseToolName, getBufferString, getChatModelClass, getChunkContent, getCloudflareWorkspaceRoot, getCodeBaseURL, getConverseOverrideMessage, getDeferredToolsListing, getLocalCwd, getLocalSessionId, getMaxOutputTokensKey, getMessageId, getMessagesWithinTokenLimit, getReadRoots, getSpawn, getTokenCountForMessage, getWorkspaceFS, getWorkspaceRoots, getWriteRoots, handleServerToolResult, handleToolCallChunks, handleToolCalls, hasNestedQuantifier, hasNestedQuantifiers, hasToolSearchInCurrentTurn, imageAttachmentContent, imagesSchema, initializeModel, interrupt, isAIMessage, isAnthropicLike, isBaseMessage, isContextOverflowError, isDangerousPattern, isFromAnyMcpServer, isFromMcpServer, isGoogleLike, isInterrupted, isLegacyConvertible, isLikelyContextOverflowError, isOpenAILike, isPresent, isThinkingEnabled, isToolMessage, isZodSchema, joinKeys, labelContentByAgent, locateEdit, makeIsDeferred, makeRequest, maskConsumedToolResults, matchesQuery, messagesStateReducer, modifyDeltaProperties, newsSchema, normalizeBashToolResultsForReplay, normalizeServerFilter, normalizeToBashIdentifier, normalizeToPythonIdentifier, partitionAndMarkAnthropicToolCache, performLocalSearch, preFlightTruncateToolCallInputs, preFlightTruncateToolResults, querySchema, repairOrphanedToolMessages, resetIfNotEmpty, resolveCloudflareSandbox, resolveCodeApiAuthHeaders, resolveContextPruningSettings, resolveLocalExecutionConfig, resolveLocalExecutionTools, resolveLocalToolRegistry, resolveLocalToolsForBinding, resolveSubagentConfigs, resolveWorkspacePath, resolveWorkspacePathSafe, runBashAstChecks, runPostEditSyntaxCheck, sanitizeOrphanToolBlocks, sanitizeRegex, serializeMessage, shellQuote, shiftIndexTokenCountMap, shouldTriggerSummarization, sleep, spawnLocalProcess, splitAtRecencyBoundary, stripAnthropicCacheControl, stripBedrockCacheControl, stripCodeSessionFileSummary, summarizeEvent, toJsonSchema, tool, toolResultTypes, toolsCondition, truncateLocalOutput, truncateToolInput, truncateToolResultContent, tryFallbackProviders, unescapeObject, unwrapToolResponse, validateBashCommand, validateCloudflareBashCommand, videosSchema, withMessageRole };

package/dist/esm/messages/content.mjs

@@ -2,6 +2,16 @@ import "../common/enum.mjs";
 import "../common/index.mjs";
 //#region src/messages/content.ts
 /**
+* Whether {@link formatContentStrings} will flatten this message's content:
+* a human/ai/system message whose content is an array of text-only blocks.
+*/
+const isLegacyConvertible = (message) => {
+	const messageType = message.getType();
+	if (!(messageType === "human" || messageType === "ai" || messageType === "system")) return false;
+	if (!Array.isArray(message.content)) return false;
+	return message.content.every((block) => block.type === "text");
+};
+/**
 * Formats an array of messages for LangChain, making sure all content fields are strings
 * @param {Array<HumanMessage | AIMessage | SystemMessage | ToolMessage>} payload - The array of messages to format.
 * @returns {Array<HumanMessage | AIMessage | SystemMessage | ToolMessage>} - The array of formatted LangChain messages, including ToolMessages for tool calls.
@@ -9,20 +19,7 @@ import "../common/index.mjs";
 const formatContentStrings = (payload) => {
 	const result = [];
 	for (const message of payload) {
-		const messageType = message.getType();
-		if (!(messageType === "human" || messageType === "ai" || messageType === "system")) {
-			result.push(message);
-			continue;
-		}
-		if (typeof message.content === "string") {
-			result.push(message);
-			continue;
-		}
-		if (!Array.isArray(message.content)) {
-			result.push(message);
-			continue;
-		}
-		if (!message.content.every((block) => block.type === "text")) {
+		if (!isLegacyConvertible(message)) {
 			result.push(message);
 			continue;
 		}
@@ -35,6 +32,6 @@ const formatContentStrings = (payload) => {
 	return result;
 };
 //#endregion
-export { formatContentStrings };
+export { formatContentStrings, isLegacyConvertible };
 
 //# sourceMappingURL=content.mjs.map

package/dist/esm/messages/content.mjs.map

@@ -1 +1 @@
-{"version":3,"file":"content.mjs","names":[],"sources":["../../../src/messages/content.ts"],"sourcesContent":["import type { BaseMessage } from '@langchain/core/messages';\nimport { ContentTypes } from '@/common';\n\n/**\n * Formats an array of messages for LangChain, making sure all content fields are strings\n * @param {Array<HumanMessage | AIMessage | SystemMessage | ToolMessage>} payload - The array of messages to format.\n * @returns {Array<HumanMessage | AIMessage | SystemMessage | ToolMessage>} - The array of formatted LangChain messages, including ToolMessages for tool calls.\n */\nexport const formatContentStrings = (\n  payload: Array<BaseMessage>\n): Array<BaseMessage> => {\n  // Create a new array to store the processed messages\n  const result: Array<BaseMessage> = [];\n\n  for (const message of payload) {\n    const messageType = message.getType();\n    const isValidMessage =\n      messageType === 'human' ||\n      messageType === 'ai' ||\n      messageType === 'system';\n\n    if (!isValidMessage) {\n      result.push(message);\n      continue;\n    }\n\n    // If content is already a string, add as-is\n    if (typeof message.content === 'string') {\n      result.push(message);\n      continue;\n    }\n\n    // If content is not an array, add as-is\n    if (!Array.isArray(message.content)) {\n      result.push(message);\n      continue;\n    }\n\n    // Check if all content blocks are text type\n    const allTextBlocks = message.content.every(\n      (block) => block.type === ContentTypes.TEXT\n    );\n\n    // Only convert to string if all blocks are text type\n    if (!allTextBlocks) {\n      result.push(message);\n      continue;\n    }\n\n    // Reduce text types to a single string\n    const content = message.content.reduce((acc, curr) => {\n      if (curr.type === ContentTypes.TEXT) {\n        return `${acc}${curr[ContentTypes.TEXT] || ''}\\n`;\n      }\n      return acc;\n    }, '');\n\n    message.content = content.trim();\n    result.push(message);\n  }\n\n  return result;\n};\n"],"mappings":";;;;;;;;AAQA,MAAa,wBACX,YACuB;CAEvB,MAAM,SAA6B,CAAC;CAEpC,KAAK,MAAM,WAAW,SAAS;EAC7B,MAAM,cAAc,QAAQ,QAAQ;EAMpC,IAAI,EAJF,gBAAgB,WAChB,gBAAgB,QAChB,gBAAgB,WAEG;GACnB,OAAO,KAAK,OAAO;GACnB;EACF;EAGA,IAAI,OAAO,QAAQ,YAAY,UAAU;GACvC,OAAO,KAAK,OAAO;GACnB;EACF;EAGA,IAAI,CAAC,MAAM,QAAQ,QAAQ,OAAO,GAAG;GACnC,OAAO,KAAK,OAAO;GACnB;EACF;EAQA,IAAI,CALkB,QAAQ,QAAQ,OACnC,UAAU,MAAM,SAAA,MAIF,GAAG;GAClB,OAAO,KAAK,OAAO;GACnB;EACF;EAUA,QAAQ,UAPQ,QAAQ,QAAQ,QAAQ,KAAK,SAAS;GACpD,IAAI,KAAK,SAAA,QACP,OAAO,GAAG,MAAM,KAAA,WAA2B,GAAG;GAEhD,OAAO;EACT,GAAG,EAEqB,CAAC,CAAC,KAAK;EAC/B,OAAO,KAAK,OAAO;CACrB;CAEA,OAAO;AACT"}
+{"version":3,"file":"content.mjs","names":[],"sources":["../../../src/messages/content.ts"],"sourcesContent":["import type {\n  BaseMessage,\n  MessageContentComplex,\n} from '@langchain/core/messages';\nimport { ContentTypes } from '@/common';\n\n/**\n * Whether {@link formatContentStrings} will flatten this message's content:\n * a human/ai/system message whose content is an array of text-only blocks.\n */\nexport const isLegacyConvertible = (message: BaseMessage): boolean => {\n  const messageType = message.getType();\n  const isValidMessage =\n    messageType === 'human' || messageType === 'ai' || messageType === 'system';\n  if (!isValidMessage) {\n    return false;\n  }\n  if (!Array.isArray(message.content)) {\n    return false;\n  }\n  return message.content.every((block) => block.type === ContentTypes.TEXT);\n};\n\n/**\n * Formats an array of messages for LangChain, making sure all content fields are strings\n * @param {Array<HumanMessage | AIMessage | SystemMessage | ToolMessage>} payload - The array of messages to format.\n * @returns {Array<HumanMessage | AIMessage | SystemMessage | ToolMessage>} - The array of formatted LangChain messages, including ToolMessages for tool calls.\n */\nexport const formatContentStrings = (\n  payload: Array<BaseMessage>\n): Array<BaseMessage> => {\n  // Create a new array to store the processed messages\n  const result: Array<BaseMessage> = [];\n\n  for (const message of payload) {\n    if (!isLegacyConvertible(message)) {\n      result.push(message);\n      continue;\n    }\n\n    // Reduce text types to a single string\n    const blocks = message.content as MessageContentComplex[];\n    const content = blocks.reduce((acc, curr) => {\n      if (curr.type === ContentTypes.TEXT) {\n        return `${acc}${curr[ContentTypes.TEXT] || ''}\\n`;\n      }\n      return acc;\n    }, '');\n\n    message.content = content.trim();\n    result.push(message);\n  }\n\n  return result;\n};\n"],"mappings":";;;;;;;AAUA,MAAa,uBAAuB,YAAkC;CACpE,MAAM,cAAc,QAAQ,QAAQ;CAGpC,IAAI,EADF,gBAAgB,WAAW,gBAAgB,QAAQ,gBAAgB,WAEnE,OAAO;CAET,IAAI,CAAC,MAAM,QAAQ,QAAQ,OAAO,GAChC,OAAO;CAET,OAAO,QAAQ,QAAQ,OAAO,UAAU,MAAM,SAAA,MAA0B;AAC1E;;;;;;AAOA,MAAa,wBACX,YACuB;CAEvB,MAAM,SAA6B,CAAC;CAEpC,KAAK,MAAM,WAAW,SAAS;EAC7B,IAAI,CAAC,oBAAoB,OAAO,GAAG;GACjC,OAAO,KAAK,OAAO;GACnB;EACF;EAWA,QAAQ,UARO,QAAQ,QACA,QAAQ,KAAK,SAAS;GAC3C,IAAI,KAAK,SAAA,QACP,OAAO,GAAG,MAAM,KAAA,WAA2B,GAAG;GAEhD,OAAO;EACT,GAAG,EAEqB,CAAC,CAAC,KAAK;EAC/B,OAAO,KAAK,OAAO;CACrB;CAEA,OAAO;AACT"}

package/dist/esm/messages/prune.mjs

@@ -685,15 +685,24 @@ function createPruneMessages(factoryParams) {
 	let originalToolContentSize = 0;
 	const contextPruningSettings = resolveContextPruningSettings(factoryParams.contextPruningConfig);
 	return function pruneMessages(params) {
-		if (params.messages.length === 0) return {
-			context: [],
-			indexTokenCountMap,
-			messagesToRefine: [],
-			prePruneContextTokens: 0,
-			remainingContextTokens: factoryParams.maxTokens,
-			calibrationRatio,
-			resolvedInstructionOverhead: bestInstructionOverhead
-		};
+		if (params.messages.length === 0) {
+			/** Post-compaction calls still invoke the model — report the same
+			*  reserve-adjusted budget fields as the populated paths */
+			const emptyInstructionTokens = factoryParams.getInstructionTokens?.() ?? 0;
+			const emptyReserveRatio = factoryParams.reserveRatio ?? .05;
+			const emptyBudget = factoryParams.maxTokens - (emptyReserveRatio > 0 && emptyReserveRatio < 1 ? Math.round(factoryParams.maxTokens * emptyReserveRatio) : 0);
+			return {
+				context: [],
+				indexTokenCountMap,
+				messagesToRefine: [],
+				prePruneContextTokens: 0,
+				remainingContextTokens: Math.max(0, emptyBudget - emptyInstructionTokens),
+				calibrationRatio,
+				resolvedInstructionOverhead: bestInstructionOverhead,
+				contextBudget: emptyBudget,
+				effectiveInstructionTokens: emptyInstructionTokens
+			};
+		}
 		if (factoryParams.provider === "openAI" && factoryParams.thinkingEnabled === true) for (let i = lastTurnStartIndex; i < params.messages.length; i++) {
 			const m = params.messages[i];
 			if (m.getType() === "ai" && typeof m.additional_kwargs.reasoning_content === "string" && Array.isArray(m.additional_kwargs.provider_specific_fields?.thinking_blocks) && m.tool_calls && (m.tool_calls?.length ?? 0) > 0) {
@@ -799,7 +808,9 @@ function createPruneMessages(factoryParams) {
 				remainingContextTokens: 0,
 				contextPressure: pruningBudget > 0 ? calibratedTotalTokens / pruningBudget : 0,
 				calibrationRatio,
-				resolvedInstructionOverhead: bestInstructionOverhead
+				resolvedInstructionOverhead: bestInstructionOverhead,
+				contextBudget: pruningBudget,
+				effectiveInstructionTokens: currentInstructionTokens
 			};
 		}
 		totalTokens = sumTokenCounts(indexTokenCountMap, params.messages.length);
@@ -892,7 +903,9 @@ function createPruneMessages(factoryParams) {
 			contextPressure,
 			originalToolContent: originalToolContent.size > 0 ? originalToolContent : void 0,
 			calibrationRatio,
-			resolvedInstructionOverhead: bestInstructionOverhead
+			resolvedInstructionOverhead: bestInstructionOverhead,
+			contextBudget: pruningBudget,
+			effectiveInstructionTokens: currentInstructionTokens
 		};
 		const rawSpaceBudget = calibrationRatio > 0 ? Math.round(pruningBudget / calibrationRatio) : pruningBudget;
 		const rawSpaceInstructionTokens = calibrationRatio > 0 ? Math.round(currentInstructionTokens / calibrationRatio) : currentInstructionTokens;
@@ -1085,7 +1098,10 @@ function createPruneMessages(factoryParams) {
 				for (const [key, value] of Object.entries(preEmergencyTokenCounts)) indexTokenCountMap[key] = value;
 			}
 		}
-		const remainingContextTokens = Math.max(0, Math.min(pruningBudget, initialRemainingContextTokens + reclaimedTokens));
+		/** Scale raw-space remaining back to calibrated/provider units so it is
+		*  directly comparable with pruningBudget and prePruneContextTokens */
+		const rawRemaining = Math.max(0, initialRemainingContextTokens + reclaimedTokens);
+		const remainingContextTokens = Math.max(0, Math.min(pruningBudget, calibrationRatio > 0 ? Math.round(rawRemaining * calibrationRatio) : rawRemaining));
 		runThinkingStartIndex = thinkingStartIndex ?? -1;
 		/** The index is the first value of `context`, index relative to `params.messages` */
 		lastCutOffIndex = Math.max(params.messages.length - (context.length - (context[0]?.getType() === "system" ? 1 : 0)), 0);
@@ -1098,7 +1114,9 @@ function createPruneMessages(factoryParams) {
 			contextPressure,
 			originalToolContent: originalToolContent.size > 0 ? originalToolContent : void 0,
 			calibrationRatio,
-			resolvedInstructionOverhead: bestInstructionOverhead
+			resolvedInstructionOverhead: bestInstructionOverhead,
+			contextBudget: pruningBudget,
+			effectiveInstructionTokens: currentInstructionTokens
 		};
 	};
 }