@librechat/agents 3.6.15 → 3.6.16

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

@@ -72,7 +72,6 @@ import {
   splitAtRecencyBoundary,
   convertInjectedMessages,
   coalesceAdjacentUserTurns,
-  strictAlternationProviders,
   appendPredecessorHandoffCue,
   stampSyntheticProviderMessage,
 } from '@/messages';
@@ -98,14 +97,6 @@ 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 +121,12 @@ import {
   planContextOverflowRecovery,
   translateRecoveryBudget,
 } from '@/llm/contextOverflowRecovery';
+import {
+  attemptInvoke,
+  tryFallbackProviders,
+  getFallbackErrorContext,
+  getFallbackOverflowCandidates,
+} from '@/llm/invoke';
 import {
   hasToolOutputTracingConfig,
   resolveLangfuseConfig,
@@ -159,8 +156,11 @@ 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 +324,7 @@ function isGoogleServerSideToolMessageContentPart(
 }
 
 function hasGoogleServerSideToolDeltaContent(
-  provider: Providers | undefined,
+  provider: t.ProviderName | undefined,
   content: t.MessageDelta['content']
 ): content is t.MessageContentComplex[] {
   return (
@@ -337,7 +337,7 @@ function hasGoogleServerSideToolDeltaContent(
 }
 
 function getMessageDeltaContent(
-  provider: Providers | undefined,
+  provider: t.ProviderName | undefined,
   content: MessageContent | undefined
 ): t.MessageDelta['content'] | undefined {
   if (content == null) {
@@ -550,7 +550,7 @@ async function dispatchTextMessageContent({
 }: {
   graph: Graph<t.BaseGraphState>;
   stepKey: string;
-  provider?: Providers;
+  provider?: t.ProviderName;
   content: t.MessageDelta['content'];
   metadata: Record<string, unknown>;
 }): Promise<boolean> {
@@ -3237,7 +3237,7 @@ 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 +3508,7 @@ 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

@@ -78,6 +78,11 @@ export type {
   ChatOpenRouterCallOptions,
 } from './llm/openrouter';
 export { getChatModelClass } from './llm/providers';
+export { registerProvider } from './provider-registration';
+export type {
+  ProviderFamily,
+  ProviderRegistrationOptions,
+} from './provider-registration';
 export { CustomChatMistralAI } from './llm/mistral';
 export {
   smoothStream,

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

@@ -6,29 +6,46 @@ import { getChatModelClass } from '@/llm/providers';
 import { isOpenAILike } 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>;
+          })
+);
+
 /**
- * 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) &&
@@ -56,8 +73,14 @@ 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);
 }

package/src/llm/invoke.ts

@@ -13,14 +13,9 @@ import type { ToolCall } from '@langchain/core/messages/tool';
 import type { BaseMessage } from '@langchain/core/messages';
 import type { ToolOutputReferenceRegistry } from '@/tools/toolOutputReferences';
 import type { PreparedProviderRequest } from '@/llm/prepareProviderRequest';
-import type { StreamLimitState } from '@/llm/streamLimits';
 import type { ContextOverflowContext } from '@/utils/errors';
+import type { StreamLimitState } from '@/llm/streamLimits';
 import type * as t from '@/types';
-import {
-  modifyDeltaProperties,
-} from '@/messages';
-import { ChatModelStreamHandler, dispatchesChatModelStream } from '@/stream';
-import { Constants, ContentTypes, GraphEvents, Providers } from '@/common';
 import {
   enforceStreamLimitsForWireChunk,
   registerActiveStreamLimitGeneration,
@@ -35,11 +30,17 @@ import {
   assertPreparedProviderRequestFor,
   prepareProviderRequest,
 } from '@/llm/prepareProviderRequest';
+import {
+  getProviderFamily,
+  providerUsesManualToolStream,
+} from '@/llm/providers';
+import { ChatModelStreamHandler, dispatchesChatModelStream } from '@/stream';
+import { Constants, ContentTypes, GraphEvents, Providers } from '@/common';
 import { assertNotTruncatedToolCall } from '@/llm/truncation';
-import { manualToolStreamProviders } from '@/llm/providers';
 import { safeDispatchCustomEvent } from '@/utils/events';
 import { getContextOverflowInfo } from '@/utils/errors';
 import { appendCallbacks } from '@/utils/callbacks';
+import { modifyDeltaProperties } from '@/messages';
 import { canSealPreempt } from '@/llm/preempt';
 import { initializeModel } from '@/llm/init';
 
@@ -105,6 +106,19 @@ export type OnChunk = (
 /** Unique per-model-attempt sequence; see the stamp in `attemptInvoke`. */
 let streamLimitAttemptSeq = 0;
 
+function getManualToolStreamNormalizationProvider(
+  provider: t.ProviderName
+): t.ProviderName {
+  const family = getProviderFamily(provider);
+  if (family === 'anthropic') {
+    return Providers.ANTHROPIC;
+  }
+  if (family === 'bedrock') {
+    return Providers.BEDROCK;
+  }
+  return provider;
+}
+
 /**
  * The registered handler that owns content-part dispatch, if any.
  *
@@ -136,7 +150,7 @@ function removeOpenRouterFinalReasoningReplayContent({
 }: {
   current?: AIMessageChunk;
   next: AIMessageChunk;
-  provider: Providers;
+  provider: t.ProviderName;
 }): AIMessageChunk {
   const content = getOpenRouterFinalReasoningContent({
     current,
@@ -161,7 +175,7 @@ function getOpenRouterFinalReasoningContent({
 }: {
   current?: AIMessageChunk;
   next: AIMessageChunk;
-  provider: Providers;
+  provider: t.ProviderName;
 }): string | undefined {
   if (
     provider !== Providers.OPENROUTER ||
@@ -197,7 +211,7 @@ function getStreamHandlingChunk({
 }: {
   current?: AIMessageChunk;
   next: AIMessageChunk;
-  provider: Providers;
+  provider: t.ProviderName;
 }): AIMessageChunk | undefined {
   const content = getOpenRouterFinalReasoningContent({
     current,
@@ -473,7 +487,7 @@ function appendStreamChunk({
 }: {
   current?: AIMessageChunk;
   next: AIMessageChunk;
-  provider: Providers;
+  provider: t.ProviderName;
 }): AIMessageChunk {
   if (current == null) {
     return next;
@@ -514,11 +528,11 @@ type AttemptInvokeParams = AttemptInvokeCommonParams &
         request?: never;
         model: t.ChatModel;
         messages: BaseMessage[];
-        provider: Providers;
+        provider: t.ProviderName;
       }
   );
 
-function resolveAttemptProvider(params: AttemptInvokeParams): Providers {
+function resolveAttemptProvider(params: AttemptInvokeParams): t.ProviderName {
   if (params.request != null) {
     return params.request.provider;
   }
@@ -586,9 +600,7 @@ export async function attemptInvoke(
       : undefined;
   const generationKey =
     leaseTarget != null
-      ? resolveGenerationKey(
-        stampedConfig.metadata as Record<string, unknown>
-      )
+      ? resolveGenerationKey(stampedConfig.metadata as Record<string, unknown>)
       : undefined;
   if (leaseTarget != null && generationKey != null) {
     registerActiveStreamLimitGeneration(leaseTarget, generationKey);
@@ -809,8 +821,11 @@ async function attemptInvokeBody(
       }
     }
 
-    if (manualToolStreamProviders.has(provider)) {
-      finalChunk = modifyDeltaProperties(provider, finalChunk);
+    if (providerUsesManualToolStream(provider)) {
+      finalChunk = modifyDeltaProperties(
+        getManualToolStreamNormalizationProvider(provider),
+        finalChunk
+      );
     }
 
     if (preempted && finalChunk != null) {
@@ -857,7 +872,7 @@ async function attemptInvokeBody(
  * differ, which is the whole reason a fallback exists.
  */
 export interface FallbackErrorContext {
-  provider: Providers;
+  provider: t.ProviderName;
   clientOptions?: t.ClientOptions;
   maxContextTokens?: number;
 }
@@ -968,7 +983,7 @@ export async function tryFallbackProviders({
   prepareProviderRequest?: (input: {
     model: t.ChatModel;
     messages: BaseMessage[];
-    provider: Providers;
+    provider: t.ProviderName;
     clientOptions?: t.ClientOptions;
     maxContextTokens?: number;
     config?: RunnableConfig;
@@ -977,7 +992,7 @@ export async function tryFallbackProviders({
   prepareProviderMessages?: (input: {
     model: t.ChatModel;
     messages: BaseMessage[];
-    provider: Providers;
+    provider: t.ProviderName;
     clientOptions?: t.ClientOptions;
     maxContextTokens?: number;
     config?: RunnableConfig;

package/src/llm/prepareProviderRequest.ts

@@ -14,7 +14,6 @@ import {
 } from '@/messages/core';
 import {
   coalesceAdjacentUserTurns,
-  strictAlternationProviders,
   appendPredecessorHandoffCue,
   removePredecessorHandoffCue,
 } from '@/messages';
@@ -23,8 +22,10 @@ import {
   stripBedrockCacheControl,
 } from '@/messages/cache';
 import { annotateMessagesForLLM } from '@/tools/toolOutputReferences';
-import { Providers } from '@/common';
+import { providerRequiresStrictAlternation } from '@/llm/providers';
 import { isAnthropicLike, isOpenAILike } from '@/utils/llm';
+import { getProviderFamily } from '@/llm/providerRegistry';
+import { Providers } from '@/common';
 
 const preparedProviderRequestBrand = Symbol('PreparedProviderRequest');
 
@@ -43,7 +44,7 @@ export interface ProviderPayloadMeasurement {
 export interface PreparedProviderRequest {
   readonly model: t.ChatModel;
   readonly modelId?: string;
-  readonly provider: Providers;
+  readonly provider: t.ProviderName;
   readonly projectionMode: ProviderMessageProjectionMode;
   readonly messages: BaseMessage[];
   readonly measurement?: ProviderPayloadMeasurement;
@@ -63,7 +64,7 @@ export interface ProviderRequestContext {
 export interface PrepareProviderRequestParams {
   model: t.ChatModel;
   messages: BaseMessage[];
-  provider: Providers;
+  provider: t.ProviderName;
   context?: ProviderRequestContext;
   config?: RunnableConfig;
   maxToolResultChars?: number;
@@ -72,7 +73,7 @@ export interface PrepareProviderRequestParams {
 
 export function usesNativeOpenAIResponses(
   model: t.ChatModel,
-  provider: Providers,
+  provider: t.ProviderName,
   callOptions?: unknown
 ): boolean {
   if (!isOpenAILike(provider)) {
@@ -142,7 +143,7 @@ export function usesNativeOpenAIResponses(
 
 function resolveProviderMessageProjectionMode(
   model: t.ChatModel,
-  provider: Providers,
+  provider: t.ProviderName,
   callOptions?: unknown
 ): ProviderMessageProjectionMode {
   return usesNativeOpenAIResponses(model, provider, callOptions)
@@ -153,18 +154,16 @@ function resolveProviderMessageProjectionMode(
 interface ProjectMessagesForProviderParams {
   model: t.ChatModel;
   messages: BaseMessage[];
-  provider: Providers;
+  provider: t.ProviderName;
   maxToolResultChars?: number;
   callOptions?: unknown;
 }
 
-function projectMessagesForProviderMode({
-  messages,
-  provider,
-  maxToolResultChars,
-}: ProjectMessagesForProviderParams,
-projectionMode: ProviderMessageProjectionMode
+function projectMessagesForProviderMode(
+  { messages, provider, maxToolResultChars }: ProjectMessagesForProviderParams,
+  projectionMode: ProviderMessageProjectionMode
 ): BaseMessage[] {
+  const providerFamily = getProviderFamily(provider);
   const nativeOpenAIResponses = projectionMode === 'openai-responses';
   const providerInputMessages = projectToolStreamContentForProvider(
     messages,
@@ -197,7 +196,7 @@ projectionMode: ProviderMessageProjectionMode
       )
     );
   }
-  if (provider === Providers.ANTHROPIC) {
+  if (provider === Providers.ANTHROPIC || providerFamily === 'anthropic') {
     return projectComputerCallOutputsToText(
       projectSingleTextToolOutputsToText(
         stripBedrockCacheControl(providerInputMessages),
@@ -205,7 +204,7 @@ projectionMode: ProviderMessageProjectionMode
       )
     );
   }
-  if (provider === Providers.BEDROCK) {
+  if (provider === Providers.BEDROCK || providerFamily === 'bedrock') {
     return stripAnthropicCacheControl(
       projectComputerCallOutputsToText(
         projectCacheControlledToolOutputsToText(
@@ -310,7 +309,7 @@ export function prepareProviderRequest({
         : (message): boolean => isRunProduced.call(context, message)
     )
     : removePredecessorHandoffCue(annotated);
-  const preparedMessages = strictAlternationProviders.has(provider)
+  const preparedMessages = providerRequiresStrictAlternation(provider)
     ? coalesceAdjacentUserTurns(cued)
     : cued;
 
@@ -332,7 +331,7 @@ export function prepareProviderRequest({
 export function assertPreparedProviderRequestFor(
   request: PreparedProviderRequest,
   model: t.ChatModel,
-  provider: Providers,
+  provider: t.ProviderName,
   config?: RunnableConfig
 ): void {
   if (
@@ -344,7 +343,9 @@ export function assertPreparedProviderRequestFor(
     throw new Error('Prepared provider request does not match serving model');
   }
   if (request.provider !== provider) {
-    throw new Error('Prepared provider request does not match serving provider');
+    throw new Error(
+      'Prepared provider request does not match serving provider'
+    );
   }
   if (
     request.projectionMode !==