@librechat/agents 3.6.15 → 3.7.0

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@librechat/agents",
-  "version": "3.6.15",
+  "version": "3.7.0",
   "reova": {
     "enabled": true,
     "endpoint": "https://telemetry.reo.dev/data"
@@ -14,6 +14,11 @@
       "import": "./dist/esm/main.mjs",
       "require": "./dist/cjs/main.cjs"
     },
+    "./provider-registration": {
+      "types": "./dist/types/provider-registration.d.ts",
+      "import": "./dist/esm/provider-registration.mjs",
+      "require": "./dist/cjs/provider-registration.cjs"
+    },
     "./openai": {
       "types": "./dist/types/openai/index.d.ts",
       "import": "./dist/esm/openai/index.mjs",
@@ -73,10 +78,48 @@
       "types": "./dist/types/langchain/utils/env.d.ts",
       "import": "./dist/esm/langchain/utils/env.mjs",
       "require": "./dist/cjs/langchain/utils/env.cjs"
+    },
+    "./llm/openai": {
+      "types": "./dist/types/llm/openai/index.d.ts",
+      "import": "./dist/esm/llm/openai/index.mjs",
+      "require": "./dist/cjs/llm/openai/index.cjs"
+    },
+    "./llm/mistral": {
+      "types": "./dist/types/llm/mistral/index.d.ts",
+      "import": "./dist/esm/llm/mistral/index.mjs",
+      "require": "./dist/cjs/llm/mistral/index.cjs"
+    },
+    "./llm/anthropic": {
+      "types": "./dist/types/llm/anthropic/index.d.ts",
+      "import": "./dist/esm/llm/anthropic/index.mjs",
+      "require": "./dist/cjs/llm/anthropic/index.cjs"
+    },
+    "./llm/google": {
+      "types": "./dist/types/llm/google/index.d.ts",
+      "import": "./dist/esm/llm/google/index.mjs",
+      "require": "./dist/cjs/llm/google/index.cjs"
+    },
+    "./llm/bedrock": {
+      "types": "./dist/types/llm/bedrock/index.d.ts",
+      "import": "./dist/esm/llm/bedrock/index.mjs",
+      "require": "./dist/cjs/llm/bedrock/index.cjs"
+    },
+    "./llm/vertexai": {
+      "types": "./dist/types/llm/vertexai/index.d.ts",
+      "import": "./dist/esm/llm/vertexai/index.mjs",
+      "require": "./dist/cjs/llm/vertexai/index.cjs"
+    },
+    "./llm/openrouter": {
+      "types": "./dist/types/llm/openrouter/index.d.ts",
+      "import": "./dist/esm/llm/openrouter/index.mjs",
+      "require": "./dist/cjs/llm/openrouter/index.cjs"
     }
   },
   "typesVersions": {
     "*": {
+      "provider-registration": [
+        "dist/types/provider-registration.d.ts"
+      ],
       "langchain": [
         "dist/types/langchain/index.d.ts"
       ],
@@ -88,6 +131,27 @@
       ],
       "responses": [
         "dist/types/responses/index.d.ts"
+      ],
+      "llm/openai": [
+        "./dist/types/llm/openai/index.d.ts"
+      ],
+      "llm/mistral": [
+        "./dist/types/llm/mistral/index.d.ts"
+      ],
+      "llm/anthropic": [
+        "./dist/types/llm/anthropic/index.d.ts"
+      ],
+      "llm/google": [
+        "./dist/types/llm/google/index.d.ts"
+      ],
+      "llm/bedrock": [
+        "./dist/types/llm/bedrock/index.d.ts"
+      ],
+      "llm/vertexai": [
+        "./dist/types/llm/vertexai/index.d.ts"
+      ],
+      "llm/openrouter": [
+        "./dist/types/llm/openrouter/index.d.ts"
       ]
     }
   },
@@ -154,6 +218,7 @@
     "tool_search": "node -r dotenv/config --loader ./tsconfig-paths-bootstrap.mjs --experimental-specifier-resolution=node ./src/scripts/tool_search.ts",
     "bench:cache": "node --loader ./tsconfig-paths-bootstrap.mjs --experimental-specifier-resolution=node ./src/scripts/bench-prompt-cache.ts",
     "bench:context-pressure": "node --loader ./tsconfig-paths-bootstrap.mjs --experimental-specifier-resolution=node ./src/scripts/bench-context-pressure-cache.ts",
+    "bench:provider-derivation": "node --loader ./tsconfig-paths-bootstrap.mjs --experimental-specifier-resolution=node ./src/scripts/bench-provider-derivation.ts",
     "bench:execution-world": "node --loader ./tsconfig-paths-bootstrap.mjs --experimental-specifier-resolution=node ./src/scripts/bench-execution-world.ts",
     "probe:overflow": "node -r dotenv/config --loader ./tsconfig-paths-bootstrap.mjs --experimental-specifier-resolution=node ./src/scripts/context-overflow-probe.ts",
     "subagent": "node -r dotenv/config --loader ./tsconfig-paths-bootstrap.mjs --experimental-specifier-resolution=node ./src/scripts/multi-agent-subagent.ts",

package/src/agents/AgentContext.ts

@@ -7,8 +7,18 @@ import type {
   BaseMessageFields,
 } from '@langchain/core/messages';
 import type { RunnableConfig, Runnable } from '@langchain/core/runnables';
-import type * as t from '@/types';
 import type { ExactTokenCountCache } from '@/llm/contextPressureMeter';
+import type * as t from '@/types';
+import {
+  type CallerCapabilityProjection,
+  allowsToolCaller,
+  applyCallerCapabilityDefinitionOverrides,
+  createCallerCapabilityProjectionSnapshot,
+  isToolDefinitionActive,
+  isProgrammaticControlTool,
+  mergeCallerCapabilityDefinitions,
+  resolveCallerCapabilityProjection,
+} from '@/tools/CallerCapabilities';
 import {
   addTailCacheControl,
   addCacheControlToStablePrefixMessages,
@@ -19,6 +29,12 @@ import {
   cloneMessage,
   type PromptCacheTtl,
 } from '@/messages/cache';
+import {
+  isProgrammaticRunnerAutoBound,
+  isProgrammaticRunnerResolvedDirectly,
+  resolveLocalImplementationNames,
+  resolveLocalToolRegistry,
+} from '@/tools/local/resolveLocalExecutionTools';
 import {
   DEFAULT_RESERVE_RATIO,
   ORIGINAL_CONTENT_MAX_CHARS,
@@ -33,27 +49,11 @@ import {
   Constants,
   Providers,
 } from '@/common';
-import {
-  isProgrammaticRunnerAutoBound,
-  isProgrammaticRunnerResolvedDirectly,
-  resolveLocalImplementationNames,
-  resolveLocalToolRegistry,
-} from '@/tools/local/resolveLocalExecutionTools';
-import {
-  type CallerCapabilityProjection,
-  allowsToolCaller,
-  applyCallerCapabilityDefinitionOverrides,
-  createCallerCapabilityProjectionSnapshot,
-  isToolDefinitionActive,
-  isProgrammaticControlTool,
-  mergeCallerCapabilityDefinitions,
-  resolveCallerCapabilityProjection,
-} from '@/tools/CallerCapabilities';
+import { isTokenCounterCacheCompatible } from '@/llm/tokenCounterCacheCompatibility';
+import { createExactTokenCountCache } from '@/llm/contextPressureMeter';
 import { createSchemaOnlyTools } from '@/tools/schema';
 import { apportionTokenCounts } from '@/utils/tokens';
 import { isThinkingEnabled } from '@/llm/request';
-import { createExactTokenCountCache } from '@/llm/contextPressureMeter';
-import { isTokenCounterCacheCompatible } from '@/llm/tokenCounterCacheCompatibility';
 import { toJsonSchema } from '@/utils/schema';
 
 type AgentSystemTextBlock = {
@@ -201,7 +201,7 @@ export class AgentContext {
   /** Human-readable name for this agent (used in handoff context). Falls back to agentId if not provided. */
   name?: string;
   /** Provider for this specific agent */
-  provider: Providers;
+  provider: t.ProviderName;
   /** Client options for this agent */
   clientOptions?: t.ClientOptions;
   /** Per-agent Langfuse tracing configuration. */
@@ -433,7 +433,7 @@ export class AgentContext {
     agentId: string;
     codeSessionKey?: string;
     name?: string;
-    provider: Providers;
+    provider: t.ProviderName;
     clientOptions?: t.ClientOptions;
     langfuse?: t.LangfuseConfig;
     maxContextTokens?: number;
@@ -570,7 +570,9 @@ export class AgentContext {
       directOnlyToolNames.length > 0
         ? `\nCall these tools directly; never list them in the \`tool_manifest\` or reference them inside ${programmaticRunnerNames}: ${directOnlyToolNames
           .map((name) => `\`${name}\``)
-          .join(', ')}. Every ${programmaticRunnerNames} call must include a \`tool_manifest\` containing the exact registered names used by its code; the manifest is validated before execution starts.`
+          .join(
+            ', '
+          )}. Every ${programmaticRunnerNames} call must include a \`tool_manifest\` containing the exact registered names used by its code; the manifest is validated before execution starts.`
         : '';
     const boundary =
       '\n\n' +
@@ -658,9 +660,8 @@ export class AgentContext {
         this.toolExecution,
         this.toolDefinitions?.some((toolDef) => toolDef.name === name) === true
       ) ||
-      this.graphTools?.some(
-        (tool) => 'name' in tool && tool.name === name
-      ) === true
+      this.graphTools?.some((tool) => 'name' in tool && tool.name === name) ===
+        true
     );
   }
 
@@ -1618,7 +1619,7 @@ export class AgentContext {
 
   /** Applies token calibration only when the observation came from this provider. */
   applyObservedOverflowCalibration(
-    provider: Providers | undefined,
+    provider: t.ProviderName | undefined,
     observedCalibrationRatio: number | undefined
   ): void {
     if (

package/src/graphs/Graph.ts

@@ -53,9 +53,8 @@ import {
   cloneMessage,
   CALIBRATION_RATIO_MAX,
   createPruneMessages,
-  projectToolCallInputs,
+  projectToolMessagesForProvider,
   calculateMaxToolCallInputChars,
-  projectToolStreamContentForProvider,
   syncBudgetDerivedFields,
   addTailCacheControl,
   resolvePromptCacheTtl,
@@ -72,7 +71,6 @@ import {
   splitAtRecencyBoundary,
   convertInjectedMessages,
   coalesceAdjacentUserTurns,
-  strictAlternationProviders,
   appendPredecessorHandoffCue,
   stampSyntheticProviderMessage,
 } from '@/messages';
@@ -98,14 +96,6 @@ import {
   joinKeys,
   sleep,
 } from '@/utils';
-import {
-  attemptInvoke,
-  tryFallbackProviders,
-  getFallbackErrorContext,
-  getFallbackOverflowCandidates,
-} from '@/llm/invoke';
-import { prepareProviderRequest } from '@/llm/prepareProviderRequest';
-import { createContextPressureMeter } from '@/llm/contextPressureMeter';
 import {
   resolveStreamLimits,
   StreamLimitExceededError,
@@ -130,6 +120,12 @@ import {
   planContextOverflowRecovery,
   translateRecoveryBudget,
 } from '@/llm/contextOverflowRecovery';
+import {
+  attemptInvoke,
+  tryFallbackProviders,
+  getFallbackErrorContext,
+  getFallbackOverflowCandidates,
+} from '@/llm/invoke';
 import {
   hasToolOutputTracingConfig,
   resolveLangfuseConfig,
@@ -159,8 +155,11 @@ import { createLocalCodingToolBundle } from '@/tools/local/LocalCodingTools';
 import { SUBAGENT_REPLAY_CONTROLLER } from '@/tools/subagent/SubagentReplay';
 import { applyGraphRuntimeConfig } from '@/graphs/applyGraphRuntimeConfig';
 import { partitionAndMarkBedrockToolCache } from '@/llm/bedrock/toolCache';
+import { createContextPressureMeter } from '@/llm/contextPressureMeter';
 import { safeDispatchCustomEvent, emitAgentLog } from '@/utils/events';
+import { prepareProviderRequest } from '@/llm/prepareProviderRequest';
 import { createCloudflareCodingToolBundle } from '@/tools/cloudflare';
+import { providerRequiresStrictAlternation } from '@/llm/providers';
 import { buildSubagentToolParams } from '@/tools/SubagentTool';
 import { initializeLangfuseTracing } from '@/instrumentation';
 import { shouldTriggerSummarization } from '@/summarization';
@@ -324,7 +323,7 @@ function isGoogleServerSideToolMessageContentPart(
 }
 
 function hasGoogleServerSideToolDeltaContent(
-  provider: Providers | undefined,
+  provider: t.ProviderName | undefined,
   content: t.MessageDelta['content']
 ): content is t.MessageContentComplex[] {
   return (
@@ -337,7 +336,7 @@ function hasGoogleServerSideToolDeltaContent(
 }
 
 function getMessageDeltaContent(
-  provider: Providers | undefined,
+  provider: t.ProviderName | undefined,
   content: MessageContent | undefined
 ): t.MessageDelta['content'] | undefined {
   if (content == null) {
@@ -550,7 +549,7 @@ async function dispatchTextMessageContent({
 }: {
   graph: Graph<t.BaseGraphState>;
   stepKey: string;
-  provider?: Providers;
+  provider?: t.ProviderName;
   content: t.MessageDelta['content'];
   metadata: Record<string, unknown>;
 }): Promise<boolean> {
@@ -3178,15 +3177,10 @@ export class StandardGraph extends Graph<t.BaseGraphState, t.GraphNode> {
       const maxProviderToolResultChars =
         agentContext.maxToolResultChars ??
         calculateMaxToolResultChars(agentContext.maxContextTokens);
-      const beforeToolStreamProjection = finalMessages;
-      finalMessages = trackProviderMessageOrigins(
-        beforeToolStreamProjection,
-        projectToolStreamContentForProvider(beforeToolStreamProjection)
-      );
       const beforeToolInputProjection = finalMessages;
       finalMessages = trackProviderMessageOrigins(
         beforeToolInputProjection,
-        projectToolCallInputs(
+        projectToolMessagesForProvider(
           beforeToolInputProjection,
           calculateMaxToolCallInputChars(agentContext.maxContextTokens)
         )
@@ -3237,7 +3231,7 @@ export class StandardGraph extends Graph<t.BaseGraphState, t.GraphNode> {
         info,
       }: {
         projection: ReturnType<typeof measureProviderPayload>;
-        provider?: Providers;
+        provider?: t.ProviderName;
         info: string;
       }): ContextOverflowError => {
         const error = new ContextOverflowError(
@@ -3508,7 +3502,7 @@ export class StandardGraph extends Graph<t.BaseGraphState, t.GraphNode> {
        * drop messages, so coalescing has to see its output, and it is the
        * last shaping step before the cache breakpoint is chosen.
        */
-      if (strictAlternationProviders.has(agentContext.provider)) {
+      if (providerRequiresStrictAlternation(agentContext.provider)) {
         /**
          * Wrapped like every other provider transform: the merged message is
          * a NEW object, and without re-attachment the final pre-invoke

package/src/index.ts

@@ -70,15 +70,20 @@ export type { Interrupt } from '@langchain/langgraph';
 
 /* LLM */
 export { markTokenCounterCacheCompatible } from './llm/tokenCounterCacheCompatibility';
-export { CustomOpenAIClient } from './llm/openai';
-export { ChatOpenRouter } from './llm/openrouter';
+/** Provider chat-model classes moved off the root barrel: importing any of them here
+ *  forced every host to pay that provider SDK's module init at boot. They remain
+ *  available from their own entries, e.g. `@librechat/agents/llm/openai`. */
 export type {
   OpenRouterReasoning,
   OpenRouterReasoningEffort,
   ChatOpenRouterCallOptions,
 } from './llm/openrouter';
 export { getChatModelClass } from './llm/providers';
-export { CustomChatMistralAI } from './llm/mistral';
+export { registerProvider } from './provider-registration';
+export type {
+  ProviderFamily,
+  ProviderRegistrationOptions,
+} from './provider-registration';
 export {
   smoothStream,
   resolveStreamDelay,

package/src/lazyRequire.ts

@@ -0,0 +1,67 @@
+import { createRequire } from 'node:module';
+
+/**
+ * Synchronous on-demand module loading: provider SDKs and other heavy dependencies
+ * load with their first request instead of at import time. This is the only module
+ * that touches `import.meta`; jest maps it to `test/stubs/lazyRequire.ts` so suites
+ * resolve source modules through their own resolver (the same precedent as the
+ * `@langchain/mistralai` stub).
+ *
+ * Internal modules load as format-matched siblings — `.cjs` neighbors from the CJS
+ * build and `.mjs` neighbors from the ESM build (Node's `require(esm)`, safe on the
+ * declared `>=24` engine) — so a lazily resolved provider shares one LangChain class
+ * graph with the code that requested it. Under a source-mode runner such as `tsx`,
+ * the seam resolves the TypeScript source directly through the active loader.
+ */
+const moduleUrl = import.meta.url;
+const requireModule = createRequire(moduleUrl);
+
+function detectBuildExtension(url: string): '.cjs' | '.mjs' | null {
+  if (url.endsWith('.mjs')) {
+    return '.mjs';
+  }
+  if (url.endsWith('.cjs')) {
+    return '.cjs';
+  }
+  return null;
+}
+
+const buildExtension = detectBuildExtension(moduleUrl);
+
+/** False when the package runs from TypeScript source, where dist siblings do
+ *  not exist and lazily loadable modules must be registered up front. */
+export function isBuiltRuntime(): boolean {
+  return buildExtension != null;
+}
+
+/** Source-mode commands run the TypeScript through an ESM loader that a synchronous
+ *  CJS `require` can neither reach nor share module identity with, so lazily loadable
+ *  modules are provided up front instead: `@/llm/providers.eager` imports them through
+ *  the active loader and registers them here. */
+const sourceModeModules = new Map<string, unknown>();
+
+export function registerSourceModeModules(modules: Record<string, unknown>): void {
+  for (const [relativePath, moduleExports] of Object.entries(modules)) {
+    sourceModeModules.set(relativePath, moduleExports);
+  }
+}
+
+/** Loads a module of this package by its src-relative path, e.g. `llm/openai/index`. */
+export function requireInternalModule<T>(relativePath: string): T {
+  if (buildExtension != null) {
+    return requireModule(`./${relativePath}${buildExtension}`) as T;
+  }
+  const provided = sourceModeModules.get(relativePath);
+  if (provided == null) {
+    throw new Error(
+      `Lazily loaded module "${relativePath}" is unavailable when running from source; ` +
+        'import \'@/llm/providers.eager\' at the entrypoint before any model is used.'
+    );
+  }
+  return provided as T;
+}
+
+/** Loads a third-party package on first use; identity-sensitive callers must not use this. */
+export function requireLazyModule<T>(specifier: string): T {
+  return requireModule(specifier) as T;
+}

package/src/llm/bedrock/toolCache.ts

@@ -4,7 +4,11 @@ import type { OpenAIClient } from '@langchain/openai';
 import type { DocumentType } from '@smithy/types';
 import type { GraphTools } from '@/types';
 import { buildBedrockCachePoint, type PromptCacheTtl } from '@/messages/cache';
-import { _convertToOpenAITool } from '@/llm/openai';
+import { requireInternalModule } from '@/lazyRequire';
+
+/** Loads with the first Bedrock request rather than alongside every graph. */
+const convertToOpenAITool: typeof import('@/llm/openai')._convertToOpenAITool = (...args) =>
+  requireInternalModule<typeof import('@/llm/openai')>('llm/openai/index')._convertToOpenAITool(...args);
 
 const BEDROCK_TOOL_CACHE_MARKER = '__lc_bedrock_cache_point_after';
 const BEDROCK_TOOL_CACHE_DISABLED_MARKER = '__lc_bedrock_skip_tool_cache';
@@ -77,7 +81,7 @@ function toBedrockTool(tool: unknown): BedrockToolWithCacheMarker {
   }
 
   return openAIToBedrockTool(
-    _convertToOpenAITool(tool as BindToolsInput) as OpenAIFunctionTool
+    convertToOpenAITool(tool as BindToolsInput) as OpenAIFunctionTool
   ) as BedrockToolWithCacheMarker;
 }
 

package/src/llm/contextOverflowRecovery.ts

@@ -18,7 +18,7 @@
  * provider-space budget.
  */
 import type { ContextOverflowInfo } from '@/utils/errors';
-import type { Providers } from '@/common';
+import type { ProviderName } from '@/types';
 import { getContextOverflowInfo } from '@/utils/errors';
 
 /** Fraction of the previous budget used when the provider named no ceiling. */
@@ -66,7 +66,7 @@ export interface OverflowRecoveryPlan {
 
 export interface OverflowRecoveryParams {
   error: unknown;
-  provider: Providers;
+  provider: ProviderName;
   /** Budget in force when the rejected prompt was built. */
   maxContextTokens?: number;
   /** Our own estimate of the prompt we actually sent. */

package/src/llm/init.ts

@@ -1,39 +1,63 @@
-import { ChatVertexAI } from '@langchain/google-vertexai';
 import type { Runnable } from '@langchain/core/runnables';
 import type * as t from '@/types';
-import { ChatOpenAI, AzureChatOpenAI } from '@/llm/openai';
 import { getChatModelClass } from '@/llm/providers';
-import { isOpenAILike } from '@/utils';
+import { isOpenAILike, isLibreChatOpenAIModel, constructorChainHasLcName } from '@/utils';
 import { Providers } from '@/common';
 
+type InitializeModelParams<P extends t.ProviderName> = {
+  provider: P;
+  tools?: t.GraphTools;
+} & (
+  | {
+      override: t.ChatModelInstance;
+      clientOptions?: t.ProviderOptionsFor<P>;
+    }
+  | ([P] extends [keyof t.ProviderOptionsMap]
+      ? {
+          override?: t.ChatModelInstance;
+          clientOptions?: t.ProviderOptionsFor<P>;
+        }
+      : object extends t.ProviderOptionsFor<P>
+        ? {
+            override?: t.ChatModelInstance;
+            clientOptions?: t.ProviderOptionsFor<P>;
+          }
+        : {
+            override?: undefined;
+            clientOptions: t.ProviderOptionsFor<P>;
+          })
+);
+
+const VERTEX_LC_NAMES: ReadonlySet<string> = new Set(['ChatVertexAI']);
+
+/** Structural stand-in for `instanceof ChatVertexAI`: matches both builds of
+ *  `@langchain/google-vertexai`, while this package's own vertex class reports
+ *  `LibreChatVertexAI` and keeps failing this guard exactly as it did under
+ *  `instanceof`. */
+function isLangchainVertexModel(
+  model: unknown
+): model is import('@langchain/google-vertexai').ChatVertexAI {
+  return constructorChainHasLcName(model, VERTEX_LC_NAMES);
+}
+
 /**
- * Creates a chat model instance for a given provider, applies provider-specific
- * field assignments, and optionally binds tools.
- *
- * This is the single entry point for model creation across the codebase — used
- * by both the agent graph (main LLM) and the summarization node (compaction LLM).
- * An optional `override` model can be passed to skip construction entirely
- * (useful for cached/reused model instances or test fakes).
+ * Creates a chat model instance for a given built-in or host-registered
+ * provider, applies provider-specific field assignments, and optionally binds
+ * tools.
  */
-export function initializeModel({
+export function initializeModel<P extends t.ProviderName>({
   provider,
   clientOptions,
   tools,
   override,
-}: {
-  provider: Providers;
-  clientOptions?: t.ClientOptions;
-  tools?: t.GraphTools;
-  override?: t.ChatModelInstance;
-}): Runnable {
+}: InitializeModelParams<P>): Runnable {
   const model =
     override ??
-    new (getChatModelClass(provider))(clientOptions ?? ({} as never));
+    new (getChatModelClass(provider))(
+      (clientOptions ?? {}) as t.ProviderOptionsFor<P>
+    );
 
-  if (
-    isOpenAILike(provider) &&
-    (model instanceof ChatOpenAI || model instanceof AzureChatOpenAI)
-  ) {
+  if (isOpenAILike(provider) && isLibreChatOpenAIModel(model)) {
     const opts = clientOptions as t.OpenAIClientOptions | undefined;
     if (opts) {
       model.temperature = opts.temperature as number;
@@ -42,7 +66,7 @@ export function initializeModel({
       model.presencePenalty = opts.presencePenalty as number;
       model.n = opts.n as number;
     }
-  } else if (provider === Providers.VERTEXAI && model instanceof ChatVertexAI) {
+  } else if (provider === Providers.VERTEXAI && isLangchainVertexModel(model)) {
     const opts = clientOptions as t.VertexAIClientOptions | undefined;
     if (opts) {
       model.temperature = opts.temperature as number;
@@ -56,8 +80,14 @@ export function initializeModel({
   }
 
   if (!tools || tools.length === 0) {
-    return model as unknown as Runnable;
+    return model;
+  }
+
+  if (!('bindTools' in model) || typeof model.bindTools !== 'function') {
+    throw new TypeError(
+      `LLM provider does not support tool binding: ${provider}`
+    );
   }
 
-  return (model as t.ModelWithTools).bindTools(tools);
+  return model.bindTools(tools);
 }