@illuma-ai/agents 1.1.0 → 1.1.1

package/dist/types/graphs/Graph.d.ts

@@ -80,6 +80,15 @@ export declare class StandardGraph extends Graph<t.BaseGraphState, t.GraphNode>
     private _pruneCalibration;
     /** Run-scoped tool discovery cache — avoids re-parsing conversation history on every iteration */
     private _toolDiscoveryCache;
+    /**
+     * SCALE: Tracks whether a summary call is already in-flight for this Graph instance.
+     * Prevents multiple concurrent summary LLM calls when rapid tool iterations each
+     * trigger pruning. At 2000 users with 3+ tool calls per turn, this prevents
+     * 6000+ summary calls/turn from becoming 2000.
+     */
+    private _summaryInFlight;
+    /** Messages accumulated across tool iterations while a summary call is in-flight */
+    private _pendingMessagesToRefine;
     /** Map of agent contexts by agent ID */
     agentContexts: Map<string, AgentContext>;
     /** Default agent ID to use */

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@illuma-ai/agents",
-  "version": "1.1.0",
+  "version": "1.1.1",
   "main": "./dist/cjs/main.cjs",
   "module": "./dist/esm/main.mjs",
   "types": "./dist/types/index.d.ts",

package/src/graphs/Graph.ts

@@ -218,6 +218,15 @@ export class StandardGraph extends Graph<t.BaseGraphState, t.GraphNode> {
   private _pruneCalibration: PruneCalibrationState;
   /** Run-scoped tool discovery cache — avoids re-parsing conversation history on every iteration */
   private _toolDiscoveryCache: ToolDiscoveryCache;
+  /**
+   * SCALE: Tracks whether a summary call is already in-flight for this Graph instance.
+   * Prevents multiple concurrent summary LLM calls when rapid tool iterations each
+   * trigger pruning. At 2000 users with 3+ tool calls per turn, this prevents
+   * 6000+ summary calls/turn from becoming 2000.
+   */
+  private _summaryInFlight: boolean = false;
+  /** Messages accumulated across tool iterations while a summary call is in-flight */
+  private _pendingMessagesToRefine: BaseMessage[] = [];
   /** Map of agent contexts by agent ID */
   agentContexts: Map<string, AgentContext> = new Map();
   /** Default agent ID to use */
@@ -301,9 +310,11 @@ export class StandardGraph extends Graph<t.BaseGraphState, t.GraphNode> {
       new Map()
     );
     this.invokedToolIds = resetIfNotEmpty(this.invokedToolIds, undefined);
-    // Reset EMA calibration and tool discovery cache for fresh run
+    // Reset EMA calibration, tool discovery cache, and summary debounce for fresh run
     this._pruneCalibration = createPruneCalibration();
     this._toolDiscoveryCache.reset();
+    this._summaryInFlight = false;
+    this._pendingMessagesToRefine = [];
     for (const context of this.agentContexts.values()) {
       context.reset();
     }
@@ -1598,20 +1609,39 @@ export class StandardGraph extends Graph<t.BaseGraphState, t.GraphNode> {
               `[Graph:ContextMgmt] Pruned ${messages.length}→${context.length} msgs (${messagesToRefine.length} discarded) | summary=${summarySource}${summary ? ` (len=${summary.length})` : ''} | calibration=${this._pruneCalibration.ratio.toFixed(3)}(${this._pruneCalibration.iterations})`
             );
 
-            // Fire background summarization — updates cache for next iteration/turn
-            agentContext
-              .summarizeCallback(messagesToRefine)
-              .then((updated) => {
-                if (updated != null && updated !== '') {
-                  this._cachedRunSummary = updated;
-                }
-              })
-              .catch((err) => {
-                console.error(
-                  '[Graph] Background summary failed (non-fatal):',
-                  err
-                );
-              });
+            // SCALE: Debounce background summarization — if a summary call is already
+            // in-flight (from a prior tool iteration), accumulate messages instead of
+            // firing another concurrent LLM call. At 2000 users with 3+ tool calls
+            // per turn, this prevents 3x summary call volume.
+            if (this._summaryInFlight) {
+              this._pendingMessagesToRefine.push(...messagesToRefine);
+              console.debug(
+                `[Graph:ContextMgmt] Summary in-flight, queued ${messagesToRefine.length} msgs (pending=${this._pendingMessagesToRefine.length})`
+              );
+            } else {
+              this._summaryInFlight = true;
+              const allMessages = this._pendingMessagesToRefine.length > 0
+                ? [...this._pendingMessagesToRefine, ...messagesToRefine]
+                : messagesToRefine;
+              this._pendingMessagesToRefine = [];
+
+              agentContext
+                .summarizeCallback(allMessages)
+                .then((updated) => {
+                  if (updated != null && updated !== '') {
+                    this._cachedRunSummary = updated;
+                  }
+                })
+                .catch((err) => {
+                  console.error(
+                    '[Graph] Background summary failed (non-fatal):',
+                    err
+                  );
+                })
+                .finally(() => {
+                  this._summaryInFlight = false;
+                });
+            }
 
             if (summary != null && summary !== '') {
               hasSummary = true;