@illuma-ai/agents 1.1.0 → 1.1.1

package/dist/esm/graphs/Graph.mjs

@@ -100,6 +100,15 @@ class StandardGraph extends Graph {
     _pruneCalibration;
     /** Run-scoped tool discovery cache — avoids re-parsing conversation history on every iteration */
     _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.
+     */
+    _summaryInFlight = false;
+    /** Messages accumulated across tool iterations while a summary call is in-flight */
+    _pendingMessagesToRefine = [];
     /** Map of agent contexts by agent ID */
     agentContexts = new Map();
     /** Default agent ID to use */
@@ -155,9 +164,11 @@ class StandardGraph extends Graph {
         this.messageStepHasToolCalls = resetIfNotEmpty(this.messageStepHasToolCalls, new Map());
         this.prelimMessageIdsByStepKey = resetIfNotEmpty(this.prelimMessageIdsByStepKey, 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();
         }
@@ -1119,17 +1130,34 @@ class StandardGraph extends Graph {
                         }
                         // Single consolidated log for the entire prune+summarize decision
                         console.debug(`[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;
                             const summaryMsg = new SystemMessage(`[Conversation Summary]\n${summary}`);