@librechat/agents 3.6.14 → 3.6.16

package/src/messages/format.ts

@@ -27,6 +27,7 @@ import type {
   ToolCallPart,
   TPayload,
   TMessage,
+  ProviderName,
 } from '@/types';
 import type {
   ProviderMessageAttribution,
@@ -366,7 +367,7 @@ export const formatFromLangChain = (
 interface FormatAssistantMessageOptions {
   preserveUnpairedServerToolUses?: boolean;
   preserveReasoningContent?: boolean;
-  provider?: Providers;
+  provider?: ProviderName;
   sourceMessageId?: string;
   sourceContentPartOffset?: number;
   sourceContentPartIndices?: readonly SourceContentPartIndices[];
@@ -376,7 +377,7 @@ interface FormatAssistantMessageOptions {
 type SourceContentPartIndices = number | readonly number[];
 
 interface FormatAgentMessagesOptions {
-  provider?: Providers;
+  provider?: ProviderName;
   /** Reconstruct hidden `reasoning_content` from `THINK` parts onto prior
    *  tool-call messages. Explicit opt-in for OpenAI-compatible endpoints that
    *  replay reasoning across turns; defaults to on for DeepSeek thinking-mode. */
@@ -478,10 +479,7 @@ function collectTrustedToolResultSourceContentPartIndices(
   let trusted: Set<number> | undefined;
   let previousPart: MessageContentComplex | null | undefined;
   for (let index = 0; index < content.length; index++) {
-    const part = content[index] as
-      | MessageContentComplex
-      | null
-      | undefined;
+    const part = content[index] as MessageContentComplex | null | undefined;
     if (part == null) {
       previousPart = part;
       continue;
@@ -509,13 +507,18 @@ function sourceContentPartIndicesAreTrustedToolResult(
   trustedToolSourceContentPartIndices: ReadonlySet<number> | undefined
 ): boolean {
   if (typeof sourceContentPartIndices === 'number') {
-    return trustedToolSourceContentPartIndices?.has(sourceContentPartIndices) === true;
+    return (
+      trustedToolSourceContentPartIndices?.has(sourceContentPartIndices) ===
+      true
+    );
   }
   if (sourceContentPartIndices.length === 0) {
     return false;
   }
   for (const sourceContentPartIndex of sourceContentPartIndices) {
-    if (trustedToolSourceContentPartIndices?.has(sourceContentPartIndex) !== true) {
+    if (
+      trustedToolSourceContentPartIndices?.has(sourceContentPartIndex) !== true
+    ) {
       return false;
     }
   }
@@ -1377,10 +1380,7 @@ function stampSourceMessageIdentity(
       },
     ];
   }
-  setProviderMessageProvenance(
-    message,
-    partsToStamp
-  );
+  setProviderMessageProvenance(message, partsToStamp);
   if (sourceMessageId == null || derivedIndex !== 0) {
     return;
   }
@@ -2759,9 +2759,7 @@ function getSyntheticProviderContextProvenanceParts(
 ): ProviderMessageProvenancePart[] | null {
   /** Fold labels are generated context, while retained source bytes keep their
    * original attribution so downstream policy can still route them exactly. */
-  const parts: ProviderMessageProvenancePart[] = [
-    { attribution: 'synthetic' },
-  ];
+  const parts: ProviderMessageProvenancePart[] = [{ attribution: 'synthetic' }];
   for (const source of sourceMessages) {
     const {
       message: sourceMessage,
@@ -3311,7 +3309,7 @@ function appendToolCalls(
  */
 export function ensureThinkingBlockInMessages(
   messages: BaseMessage[],
-  _provider: Providers,
+  _provider: ProviderName,
   config?: RunnableConfig,
   runStartIndex?: number
 ): BaseMessage[] {

package/src/messages/prune.ts

@@ -13,6 +13,7 @@ import type {
 } from '@/types/stream';
 import type { ContextPruningConfig } from '@/types/graph';
 import type { TokenCounter } from '@/types/run';
+import type { ProviderName } from '@/types';
 import {
   cloneToolMessageWithContent,
   compactToolContent,
@@ -29,6 +30,7 @@ import {
 import { resolveContextPruningSettings } from './contextPruningSettings';
 import { hasUnsafeStructuredSerialization } from '@/utils/tokens';
 import { ContentTypes, Providers, Constants } from '@/common';
+import { getProviderFamily } from '@/llm/providerRegistry';
 import { applyContextPruning } from './contextPruning';
 import { toLangChainContent } from './langchain';
 
@@ -130,7 +132,7 @@ export function clampCalibrationRatio(ratio: number): number {
 }
 
 export type PruneMessagesFactoryParams = {
-  provider?: Providers;
+  provider?: ProviderName;
   maxTokens: number;
   /** Per-tool-result character cap applied while reconciling cached counts. */
   maxToolResultChars?: number;
@@ -2065,6 +2067,15 @@ type ThinkingBlocks = {
 };
 
 export function createPruneMessages(factoryParams: PruneMessagesFactoryParams) {
+  const providerFamily =
+    factoryParams.provider == null
+      ? undefined
+      : getProviderFamily(factoryParams.provider);
+  const usesBedrockThinking =
+    factoryParams.provider === Providers.BEDROCK ||
+    providerFamily === 'bedrock';
+  const usesOpenAIThinking =
+    factoryParams.provider === Providers.OPENAI || providerFamily === 'openai';
   const indexTokenCountMap = { ...factoryParams.indexTokenCountMap };
   let lastTurnStartIndex = factoryParams.startIndex;
   let lastCutOffIndex = 0;
@@ -2152,10 +2163,7 @@ export function createPruneMessages(factoryParams: PruneMessagesFactoryParams) {
       };
     }
 
-    if (
-      factoryParams.provider === Providers.OPENAI &&
-      factoryParams.thinkingEnabled === true
-    ) {
+    if (usesOpenAIThinking && factoryParams.thinkingEnabled === true) {
       for (let i = lastTurnStartIndex; i < params.messages.length; i++) {
         const m = params.messages[i];
         if (
@@ -2729,10 +2737,9 @@ export function createPruneMessages(factoryParams: PruneMessagesFactoryParams) {
       thinkingEnabled: factoryParams.thinkingEnabled,
       tokenCounter: factoryParams.tokenCounter,
       instructionTokens: rawSpaceInstructionTokens,
-      reasoningType:
-        factoryParams.provider === Providers.BEDROCK
-          ? ContentTypes.REASONING_CONTENT
-          : ContentTypes.THINKING,
+      reasoningType: usesBedrockThinking
+        ? ContentTypes.REASONING_CONTENT
+        : ContentTypes.THINKING,
       thinkingStartIndex:
         factoryParams.thinkingEnabled === true
           ? runThinkingStartIndex
@@ -2826,10 +2833,9 @@ export function createPruneMessages(factoryParams: PruneMessagesFactoryParams) {
         thinkingEnabled: factoryParams.thinkingEnabled,
         tokenCounter: factoryParams.tokenCounter,
         instructionTokens: currentInstructionTokens,
-        reasoningType:
-          factoryParams.provider === Providers.BEDROCK
-            ? ContentTypes.REASONING_CONTENT
-            : ContentTypes.THINKING,
+        reasoningType: usesBedrockThinking
+          ? ContentTypes.REASONING_CONTENT
+          : ContentTypes.THINKING,
         thinkingStartIndex:
           factoryParams.thinkingEnabled === true
             ? runThinkingStartIndex
@@ -2963,10 +2969,9 @@ export function createPruneMessages(factoryParams: PruneMessagesFactoryParams) {
           thinkingEnabled: factoryParams.thinkingEnabled,
           tokenCounter: factoryParams.tokenCounter,
           instructionTokens: currentInstructionTokens,
-          reasoningType:
-            factoryParams.provider === Providers.BEDROCK
-              ? ContentTypes.REASONING_CONTENT
-              : ContentTypes.THINKING,
+          reasoningType: usesBedrockThinking
+            ? ContentTypes.REASONING_CONTENT
+            : ContentTypes.THINKING,
           thinkingStartIndex:
             factoryParams.thinkingEnabled === true
               ? runThinkingStartIndex

package/src/provider-registration.ts

@@ -0,0 +1,29 @@
+import type { BaseChatModel } from '@langchain/core/language_models/chat_models';
+
+declare const CUSTOM_PROVIDER_OPTIONS_TYPE: unique symbol;
+
+/** Declaration-merge this map to type host-registered provider options. */
+export interface CustomProviderOptionsMap {
+  readonly [CUSTOM_PROVIDER_OPTIONS_TYPE]?: never;
+}
+
+export type ProviderFamily =
+  | 'openai'
+  | 'anthropic'
+  | 'bedrock'
+  | 'google'
+  | 'mistral'
+  | 'generic';
+
+export interface ProviderRegistrationOptions<
+  TOptions extends object,
+  TModel extends BaseChatModel,
+> {
+  provider: string;
+  model: new (config: TOptions) => TModel;
+  family?: ProviderFamily;
+  manualToolStream?: boolean;
+  strictAlternation?: boolean;
+}
+
+export { registerProvider } from './llm/providers';

package/src/run.ts

@@ -452,7 +452,7 @@ export class Run<_T extends t.BaseGraphState> {
         provider,
         clientOptions,
         agentId: 'default',
-      };
+      } as t.AgentInputs;
       signal = legacySignal;
     }
 

package/src/session/AgentSession.ts

@@ -588,7 +588,7 @@ function createAgentInputFromGraphConfig(
       provider,
       clientOptions,
       agentId: 'default',
-    };
+    } as t.AgentInputs;
   }
   const summarizationConfig: t.SummarizationConfig = {
     ...(agent.summarizationConfig ?? {}),

package/src/stream.ts

@@ -820,12 +820,11 @@ function startEagerToolExecutions(args: {
       toolCalls: entries.map((entry) => entry.request),
       userId: graph.config?.configurable?.user_id as string | undefined,
       agentId: agentContext?.agentId,
-      callerCapabilityProjection:
-        (
-          agentContext as
-            | Partial<Pick<AgentContext, 'getCallerCapabilityProjectionSnapshot'>>
-            | undefined
-        )?.getCallerCapabilityProjectionSnapshot?.(),
+      callerCapabilityProjection: (
+        agentContext as
+          | Partial<Pick<AgentContext, 'getCallerCapabilityProjectionSnapshot'>>
+          | undefined
+      )?.getCallerCapabilityProjectionSnapshot?.(),
       configurable: graph.config?.configurable as
         | Record<string, unknown>
         | undefined,
@@ -1327,7 +1326,7 @@ export function getChunkContent({
   reasoningKey,
 }: {
   chunk?: Partial<AIMessageChunk>;
-  provider?: Providers;
+  provider?: t.ProviderName;
   reasoningKey: 'reasoning_content' | 'reasoning';
 }): string | t.MessageContentComplex[] | undefined {
   if (
@@ -1851,7 +1850,7 @@ export class ChatModelStreamHandler implements t.EventHandler {
 
     if (Array.isArray(content) && content.every(isTextContentPart)) {
       const contentGroups = splitAssistantTextContentByPhase(content);
-      const currentStepId = graph.stepKeyIds?.get(stepKey)?.at(-1);
+      const currentStepId = graph.stepKeyIds.get(stepKey)?.at(-1);
       const currentStep =
         currentStepId == null ? undefined : graph.getRunStep(currentStepId);
       const currentPhase =

package/src/summarization/node.ts

@@ -657,7 +657,7 @@ async function executeSummarizationWithFallback(params: {
      * `log('error', ...)` path below rather than bubbling up silently.
      */
     const summarizationModel = initializeModel({
-      provider: clientConfig.provider as Providers,
+      provider: clientConfig.provider,
       clientOptions: clientConfig.clientOptions as t.ClientOptions,
       tools: agentContext.getToolsForBinding(),
     }) as t.ChatModel;
@@ -670,13 +670,13 @@ async function executeSummarizationWithFallback(params: {
       priorSummaryText,
       config: summarizeConfig,
       stepId,
-      provider: clientConfig.provider as Providers,
+      provider: clientConfig.provider,
       reasoningKey: agentContext.reasoningKey,
       graph,
       usePromptCache,
       promptCacheTtl:
-        (clientConfig.provider as Providers) === Providers.ANTHROPIC ||
-        (clientConfig.provider as Providers) === Providers.OPENROUTER
+        clientConfig.provider === Providers.ANTHROPIC ||
+        clientConfig.provider === Providers.OPENROUTER
           ? resolvePromptCacheTtl(
             (
                 clientConfig.clientOptions as {
@@ -737,7 +737,7 @@ async function executeSummarizationWithFallback(params: {
         const onChunk = createSummarizationChunkHandler({
           stepId,
           config: traceConfig(summarizeConfig, 'cache_hit_compaction'),
-          provider: clientConfig.provider as Providers,
+          provider: clientConfig.provider,
           reasoningKey: agentContext.reasoningKey,
           graph,
         });
@@ -1509,7 +1509,7 @@ export function createSummarizationChunkHandler({
 }: {
   stepId?: string;
   config?: RunnableConfig;
-  provider?: Providers;
+  provider?: t.ProviderName;
   reasoningKey?: 'reasoning_content' | 'reasoning';
   graph?: StreamLimitState & { getBreakerController?: () => AbortController };
 }): OnChunk | undefined {
@@ -1622,7 +1622,7 @@ async function summarizeWithCacheHit({
   priorSummaryText: string;
   config?: RunnableConfig;
   stepId?: string;
-  provider: Providers;
+  provider: t.ProviderName;
   reasoningKey?: 'reasoning_content' | 'reasoning';
   graph?: StreamLimitState & { getBreakerSignal?: () => AbortSignal };
   usePromptCache?: boolean;

package/src/types/activityLabel.ts

@@ -1,7 +1,6 @@
 import type { RunnableConfig } from '@langchain/core/runnables';
 import type { AssistantTextPhase } from '@/types/assistantPhase';
-import type { ClientOptions } from '@/types/llm';
-import type { Providers } from '@/common';
+import type { ClientOptions, ProviderName } from '@/types/llm';
 
 /** One tool call's contribution to the label payload (host-assembled). */
 export type ActivityLabelToolEntry = {
@@ -25,7 +24,7 @@ export type ActivityLabelToolEntry = {
  * `activity_label` content type's formatter exclusions.
  */
 export type RunActivityLabelOptions = {
-  provider: Providers;
+  provider: ProviderName;
   clientOptions?: ClientOptions;
   /**
    * Agent that executed the labeled batch. Selects that agent's Langfuse
@@ -85,7 +84,7 @@ export type ActivityPhaseEntry = {
 
 /** Options for `Run.generateActivityPhaseLabel`. */
 export type RunActivityPhaseLabelOptions = {
-  provider: Providers;
+  provider: ProviderName;
   clientOptions?: ClientOptions;
   /**
    * Logical activities in run order. The SDK requires at least two so hosts

package/src/types/graph.ts

@@ -40,10 +40,10 @@ import type {
   StreamPreemption,
   TokenBudgetBreakdown,
 } from '@/types/run';
-import type { Providers, Callback, GraphNodeKeys } from '@/common';
 import type { SubagentTaskConfig } from '@/types/subagentTasks';
 import type { StandardGraph, MultiAgentGraph } from '@/graphs';
-import type { ClientOptions } from '@/types/llm';
+import type { ProviderClientOptionsConfig } from '@/types/llm';
+import type { Callback, GraphNodeKeys } from '@/common';
 
 /** Interface for bound model with stream and invoke methods */
 export interface ChatModel {
@@ -824,7 +824,7 @@ export interface LangfuseConfig {
   deterministicTraceId?: boolean;
 }
 
-export interface AgentInputs {
+interface AgentInputFields {
   agentId: string;
   /**
    * Partition key for transient code-session ids and file refs. Agents with
@@ -839,12 +839,10 @@ export interface AgentInputs {
   toolEnd?: boolean;
   toolMap?: ToolMap;
   tools?: GraphTools;
-  provider: Providers;
   /** Stable/cacheable system instructions. */
   instructions?: string;
   streamBuffer?: number;
   maxContextTokens?: number;
-  clientOptions?: ClientOptions;
   /** Per-agent Langfuse tracing configuration. */
   langfuse?: LangfuseConfig;
   /** Dynamic system tail appended after stable instructions without provider cache markers. */
@@ -914,6 +912,8 @@ export interface AgentInputs {
   graphTools?: GenericTool[];
 }
 
+export type AgentInputs = AgentInputFields & ProviderClientOptionsConfig;
+
 export interface ContextPruningConfig {
   enabled?: boolean;
   keepLastAssistants?: number;

package/src/types/llm.ts

@@ -8,6 +8,7 @@ import type {
 } from '@langchain/openai';
 import type {
   BindToolsInput,
+  BaseChatModel,
   BaseChatModelParams,
 } from '@langchain/core/language_models/chat_models';
 import type { GoogleGenerativeAIChatInput } from '@langchain/google-genai';
@@ -21,6 +22,7 @@ import type { AnthropicInput } from '@langchain/anthropic';
 import type { Runnable } from '@langchain/core/runnables';
 import type { OpenAI as OpenAIClient } from 'openai';
 import type { ChatXAIInput } from '@langchain/xai';
+import type { CustomProviderOptionsMap } from '../provider-registration';
 import type { ChatOpenRouterCallOptions } from '@/llm/openrouter';
 import type { PromptCacheTtl } from '@/messages/cache';
 import {
@@ -138,7 +140,7 @@ export type DeepSeekClientOptions = Partial<ChatDeepSeekInput> &
   StreamSmoothingOptions;
 export type XAIClientOptions = ChatXAIInput & StreamSmoothingOptions;
 
-export type ClientOptions =
+export type BuiltInClientOptions =
   | OpenAIClientOptions
   | AzureClientOptions
   | AnthropicClientOptions
@@ -149,24 +151,86 @@ export type ClientOptions =
   | DeepSeekClientOptions
   | XAIClientOptions;
 
-export type SharedLLMConfig = {
-  provider: Providers;
+type CustomProviderName = Extract<keyof CustomProviderOptionsMap, string>;
+
+type LooseRuntimeProviderName = string & {
+  readonly __runtimeProviderName?: never;
+};
+
+export type ProviderName =
+  | keyof ProviderOptionsMap
+  | CustomProviderName
+  | LooseRuntimeProviderName;
+
+declare const RUNTIME_PROVIDER_NAME: unique symbol;
+
+/** A runtime provider without declaration-merged option types. */
+export type RuntimeProviderName = string & {
+  readonly [RUNTIME_PROVIDER_NAME]: true;
+};
+
+export type ClientOptions =
+  | BuiltInClientOptions
+  | CustomProviderOptionsMap[CustomProviderName];
+
+export type SharedLLMConfig<
+  P extends
+    | keyof ProviderOptionsMap
+    | CustomProviderName
+    | RuntimeProviderName = keyof ProviderOptionsMap,
+> = {
+  provider: P;
+  model?: string;
   _lc_stream_delay?: number;
 };
 
-export interface FallbackConfig {
-  provider: Providers;
-  clientOptions?: ClientOptions;
+type CustomProviderClientOptionsConfig<P extends CustomProviderName> = {
+  provider: P;
+} & (object extends CustomProviderOptionsMap[P]
+  ? { clientOptions?: CustomProviderOptionsMap[P] }
+  : { clientOptions: CustomProviderOptionsMap[P] });
+
+export type ProviderClientOptionsConfig =
+  | {
+      provider: keyof ProviderOptionsMap;
+      clientOptions?: BuiltInClientOptions;
+    }
+  | {
+      [P in CustomProviderName]: CustomProviderClientOptionsConfig<P>;
+    }[CustomProviderName]
+  | {
+      provider: RuntimeProviderName;
+      clientOptions: ClientOptions;
+    };
+
+export type FallbackConfig = ProviderClientOptionsConfig & {
   /** Context window used to corroborate ambiguous fallback overflow errors. */
   maxContextTokens?: number;
-}
+};
+
+type LLMConfigFor<P extends CustomProviderName> = SharedLLMConfig<P> &
+  CustomProviderOptionsMap[P] & {
+    /** Optional provider fallbacks in order of attempt */
+    fallbacks?: FallbackConfig[];
+  };
 
-export type LLMConfig = SharedLLMConfig &
-  ClientOptions & {
+export type BuiltInLLMConfig = SharedLLMConfig &
+  BuiltInClientOptions & {
     /** Optional provider fallbacks in order of attempt */
     fallbacks?: FallbackConfig[];
   };
 
+export type LLMConfig =
+  | BuiltInLLMConfig
+  | {
+      [P in CustomProviderName]: LLMConfigFor<P>;
+    }[CustomProviderName]
+  | (SharedLLMConfig<RuntimeProviderName> &
+      ClientOptions & {
+        /** Optional provider fallbacks in order of attempt */
+        fallbacks?: FallbackConfig[];
+      });
+
 export type ProviderOptionsMap = {
   [Providers.AZURE]: AzureClientOptions;
   [Providers.OPENAI]: OpenAIClientOptions;
@@ -182,7 +246,7 @@ export type ProviderOptionsMap = {
   [Providers.MOONSHOT]: OpenAIClientOptions;
 };
 
-export type ChatModelMap = {
+export interface ChatModelMap {
   [Providers.XAI]: ChatXAI;
   [Providers.OPENAI]: ChatOpenAI;
   [Providers.AZURE]: AzureChatOpenAI;
@@ -195,14 +259,32 @@ export type ChatModelMap = {
   [Providers.BEDROCK]: CustomChatBedrockConverse;
   [Providers.GOOGLE]: CustomChatGoogleGenerativeAI;
   [Providers.MOONSHOT]: ChatMoonshot;
-};
+}
+
+export type ProviderOptionsFor<P extends ProviderName> =
+  P extends keyof ProviderOptionsMap
+    ? ProviderOptionsMap[P]
+    : P extends CustomProviderName
+      ? CustomProviderOptionsMap[P]
+      : BuiltInClientOptions;
+
+export type ProviderModelFor<P extends ProviderName> =
+  P extends keyof ChatModelMap
+    ? ChatModelMap[P] & BaseChatModel
+    : BaseChatModel;
+
+export type ProviderModelConstructor<P extends ProviderName> = new (
+  config: ProviderOptionsFor<P>
+) => ProviderModelFor<P>;
 
 export type ChatModelConstructorMap = {
   [P in Providers]: new (config: ProviderOptionsMap[P]) => ChatModelMap[P];
 };
 
-export type ChatModelInstance = ChatModelMap[Providers];
+export type ChatModelInstance = BaseChatModel;
 
-export type ModelWithTools = ChatModelInstance & {
+export type ModelWithTools = BaseChatModel & {
   bindTools(tools: CommonToolType[]): Runnable;
 };
+
+export type { CustomProviderOptionsMap } from '../provider-registration';

package/src/types/reasoningLabel.ts

@@ -1,7 +1,6 @@
 import type { RunnableConfig } from '@langchain/core/runnables';
 import type { UsageMetadata } from '@langchain/core/messages';
-import type { ClientOptions } from '@/types/llm';
-import type { Providers } from '@/common';
+import type { ClientOptions, ProviderName } from '@/types/llm';
 
 /** Lifecycle state of the visible reasoning snapshot being labeled. */
 export type ReasoningLabelStatus = 'streaming' | 'complete';
@@ -15,7 +14,7 @@ export type ReasoningLabelResult = {
 
 /** Options for `Run.generateReasoningLabel`. */
 export type RunReasoningLabelOptions = {
-  provider: Providers;
+  provider: ProviderName;
   clientOptions?: ClientOptions;
   /**
    * Complete user-visible reasoning accumulated for this step so far. Hidden

package/src/types/run.ts

@@ -22,7 +22,7 @@ import type * as l from '@/types/llm';
 export type ZodObjectAny = z.ZodObject<any, any, any, any>;
 export type BaseGraphConfig = {
   llmConfig: l.LLMConfig;
-  provider?: e.Providers;
+  provider?: l.ProviderName;
   clientOptions?: l.ClientOptions;
   /** Optional compile options for workflow.compile() */
   compileOptions?: g.CompileOptions;
@@ -47,7 +47,7 @@ export type SupervisedGraphConfig = BaseGraphConfig & {
   routingPolicies?: Array<{
     stage: string;
     agents?: string[];
-    model?: e.Providers;
+    model?: l.ProviderName;
     parallel?: boolean;
     /** Optional simple condition on content/tools */
     when?:
@@ -70,7 +70,7 @@ export type SupervisedGraphConfig = BaseGraphConfig & {
 
 export type RunTitleOptions = {
   inputText: string;
-  provider: e.Providers;
+  provider: l.ProviderName;
   contentParts: (s.MessageContentComplex | undefined)[];
   titlePrompt?: string;
   skipLanguage?: boolean;

package/src/types/summarize.ts

@@ -1,5 +1,5 @@
 import type { SummaryContentBlock } from '@/types/stream';
-import type { Providers } from '@/common';
+import type { ProviderName } from '@/types/llm';
 
 export type SummarizationTrigger = {
   type:
@@ -35,7 +35,7 @@ export type RetainRecentConfig = {
 };
 
 export type SummarizationConfig = {
-  provider?: Providers;
+  provider?: ProviderName;
   model?: string;
   parameters?: Record<string, unknown>;
   prompt?: string;

package/src/utils/errors.ts

@@ -12,7 +12,7 @@
  * into a forced summarization pass instead of surfacing the error.
  */
 import { ContextOverflowError } from '@langchain/core/errors';
-import type { Providers } from '@/common';
+import type { ProviderName } from '@/types';
 
 /**
  * Why the request was rejected. Both kinds are fixed by shrinking the
@@ -49,7 +49,7 @@ export interface ContextOverflowInfo {
   promptTokens?: number;
   /** Which layer produced the verdict. Surfaced in logs and asserted in tests. */
   source: 'langchain' | 'pattern';
-  provider?: Providers;
+  provider?: ProviderName;
 }
 
 interface OverflowPattern {
@@ -80,7 +80,7 @@ interface OverflowPattern {
 const CONTEXT_PRESSURE_RATIO = 0.8;
 
 export interface ContextOverflowContext {
-  provider?: Providers;
+  provider?: ProviderName;
   /** Our own estimate of the prompt size for the call that failed. */
   estimatedPromptTokens?: number;
   /** The budget we believed applied when we built that prompt. */