@librechat/agents 3.8.0 → 3.8.1

package/dist/types/types/graph.d.ts

@@ -30,6 +30,13 @@ export type SystemCallbacks = {
 export type BaseGraphState = {
     messages: BaseMessage[];
     runStepState?: RunStepResumeState;
+    /**
+     * The summary a summarize-only run produced. Kept in state because such a
+     * run has no assistant reply: trace roots report it as the run's output
+     * instead of the retained tail or the raw state. Empty once any later run
+     * on the same checkpointed thread starts, so it never outlives its run.
+     */
+    manualSummary?: string;
 };
 export type AgentSubgraphState = BaseGraphState & {
     summarizationRequest?: SummarizationNodeInput;
@@ -725,6 +732,19 @@ interface AgentInputFields {
      */
     discoveredTools?: string[];
     summarizationEnabled?: boolean;
+    /**
+     * Runs this agent as a summarize-only pass: the first model step requests
+     * summarization outright instead of consulting the trigger, and the step
+     * after the summary emits its context snapshot and ends the run without a
+     * model call. Hosts use it for user-initiated compaction, where the summary
+     * is the whole response. Requires `summarizationEnabled`: a run that cannot
+     * attempt the summary rejects with `ManualSummarizationSkippedError`
+     * rather than ending with nothing, and one whose provider calls all fail
+     * keeps the history and reports the failure on the summary step. A
+     * multi-agent workflow compiles to the agent that opted in alone; no other
+     * agent runs.
+     */
+    summarizeOnly?: boolean;
     summarizationConfig?: SummarizationConfig;
     /**
      * Optional host-supplied, user-visible guidance for compaction. The SDK

package/dist/types/types/run.d.ts

@@ -453,6 +453,10 @@ export type TokenBudgetBreakdown = {
     toolTokenCounts?: Record<string, number>;
     /** Names of counted tools that are deferred (`defer_loading`) and discovered. */
     deferredToolNames?: string[];
+    /** Calibrated retained tool traffic: tool results plus assistant turns made only of tool calls, inline provider tool results and reasoning. A subset of messageTokens, not an additional budget category. */
+    toolMessageTokens?: number;
+    /** Per-tool result-message share; excludes assistant invocation overhead. */
+    toolMessageTokenCounts?: Record<string, number>;
 };
 export type EventStreamOptions = {
     callbacks?: g.ClientCallbacks;

package/dist/types/types/summarize.d.ts

@@ -104,8 +104,11 @@ export interface SummarizationNodeInput {
      *   is compacting to recover. When summarization is not enabled, this
      *   variant performs no model call — the corrected budget alone is what the
      *   retry needs.
+     * - `manual`: the host asked for a summary outright (a summarize-only run).
+     *   The recency window is not applied unless the host configured
+     *   `retainRecent` explicitly, so the summary replaces the whole history.
      */
-    reason?: 'trigger' | 'overflow';
+    reason?: 'trigger' | 'overflow' | 'manual';
     /**
      * Whether an overflow recovery may spend a summarization model call.
      *

package/dist/types/utils/tokens.d.ts

@@ -1,6 +1,6 @@
 import type { BaseMessage } from '@langchain/core/messages';
 export type EncodingName = 'o200k_base' | 'claude';
-export type UnsafeTokenMeasurementReason = 'message_proxy' | 'content_proxy' | 'metadata_proxy' | 'metadata_accessor' | 'invalid_count';
+export type UnsafeTokenMeasurementReason = 'message_proxy' | 'content_proxy' | 'metadata_proxy' | 'metadata_accessor' | 'metadata_limit' | 'invalid_count';
 export declare class UnsafeTokenMeasurementError extends Error {
     readonly type = "unsafe_token_measurement";
     readonly reason: UnsafeTokenMeasurementReason;
@@ -61,6 +61,7 @@ export declare function encodingForModel(model: string): EncodingName;
  * unsafe so the caller can normalize or reject them conservatively.
  */
 export declare function hasUnsafeStructuredSerialization(value: unknown): boolean;
+export declare function readTokenMetadataProperty(value: object, key: PropertyKey, path: string): unknown;
 export declare function getTokenCountForMessage(message: BaseMessage, getTokenCount: (text: string) => number, encoding?: EncodingName): number;
 /**
  * Largest-remainder apportionment: scales each count by `multiplier` and

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@librechat/agents",
-  "version": "3.8.0",
+  "version": "3.8.1",
   "reova": {
     "enabled": true,
     "endpoint": "https://telemetry.reo.dev/data"

package/src/agents/AgentContext.ts

@@ -123,6 +123,7 @@ export class AgentContext {
       useLegacyContent,
       discoveredTools,
       summarizationEnabled,
+      summarizeOnly,
       summarizationConfig,
       compactionSemanticIndex,
       initialSummary,
@@ -158,6 +159,7 @@ export class AgentContext {
       useLegacyContent,
       discoveredTools,
       summarizationEnabled,
+      summarizeOnly,
       summarizationConfig,
       compactionSemanticIndex,
       contextPruningConfig,
@@ -383,6 +385,10 @@ export class AgentContext {
   useLegacyContent: boolean = false;
   /** Enables graph-level summarization for this agent */
   summarizationEnabled?: boolean;
+  /** Summarize-only run: request a summary on the first model step and end after it */
+  summarizeOnly?: boolean;
+  /** Whether the summarize-only run has already issued its one request */
+  private _manualSummarizationRequested: boolean = false;
   /** Summarization runtime settings used by graph pruning hooks */
   summarizationConfig?: t.SummarizationConfig;
   /** Host-supplied advisory guidance consumed only when compaction runs. */
@@ -480,6 +486,7 @@ export class AgentContext {
     useLegacyContent,
     discoveredTools,
     summarizationEnabled,
+    summarizeOnly,
     summarizationConfig,
     compactionSemanticIndex,
     contextPruningConfig,
@@ -508,6 +515,7 @@ export class AgentContext {
     useLegacyContent?: boolean;
     discoveredTools?: string[];
     summarizationEnabled?: boolean;
+    summarizeOnly?: boolean;
     summarizationConfig?: t.SummarizationConfig;
     compactionSemanticIndex?: t.CompactionSemanticIndex;
     contextPruningConfig?: t.ContextPruningConfig;
@@ -549,6 +557,7 @@ export class AgentContext {
 
     this.useLegacyContent = useLegacyContent ?? false;
     this.summarizationEnabled = summarizationEnabled;
+    this.summarizeOnly = summarizeOnly;
     this.summarizationConfig = summarizationConfig;
     if (compactionSemanticIndex != null) {
       this.compactionSemanticIndex = snapshotCompactionSemanticIndex(
@@ -1268,6 +1277,7 @@ export class AgentContext {
     this.summaryPrecedesMessages = this.durableSummaryPrecedesMessages;
     this._lastSummarizationMsgCount = 0;
     this._summarizationFailures = 0;
+    this._manualSummarizationRequested = false;
     this.lastCallUsage = undefined;
     this.totalTokensFresh = false;
     this.restoreContextBudgetAfterOverflow();
@@ -1617,6 +1627,20 @@ export class AgentContext {
     this._lastSummarizationMsgCount = msgCount;
   }
 
+  /**
+   * Claims the single summarization request a summarize-only run makes.
+   * True exactly once per run, on the first model step; the step that gets
+   * `false` is the one after the summary, which reports usage and ends
+   * without a model call. Always false when `summarizeOnly` is off.
+   */
+  claimManualSummarization(): boolean {
+    if (this.summarizeOnly !== true || this._manualSummarizationRequested) {
+      return false;
+    }
+    this._manualSummarizationRequested = true;
+    return true;
+  }
+
   /**
    * Records a summarization attempt that produced no usable summary — an
    * empty model response, or a provider failure the run declined to paper
@@ -1767,8 +1791,9 @@ export class AgentContext {
         const recounted: Record<string, number> = {};
         for (let i = 0; i < messages.length; i++) {
           const localCount = this.tokenCounter(messages[i]);
-          const conservativeFloor =
-            this.providerProjectionRecountFloors?.get(messages[i]);
+          const conservativeFloor = this.providerProjectionRecountFloors?.get(
+            messages[i]
+          );
           recounted[i] =
             conservativeFloor == null
               ? localCount
@@ -2112,7 +2137,13 @@ export class AgentContext {
       remainingContextTokens,
       calibrationRatio,
     };
-    syncBudgetDerivedFields(usage);
+    syncBudgetDerivedFields(
+      usage,
+      context,
+      this.contextPressureTokenCounts?.count ?? tokenCounter,
+      undefined,
+      this.provider
+    );
     return usage;
   }