take-2

Author: joshualitt

packages/core/src/context/config/profiles.ts

@@ -78,6 +78,7 @@ export const generalistProfile: ContextProfile = {
     budget: {
       retainedTokens: 65000,
       maxTokens: 150000,
+      coalescingThresholdTokens: 5000,
     },
   },
 
@@ -117,14 +118,14 @@ export const generalistProfile: ContextProfile = {
             'NodeDistillation',
             env,
             resolveProcessorOptions(config, 'NodeDistillation', {
-              nodeThresholdTokens: 1000,
+              nodeThresholdTokens: 3000,
             }),
           ),
           createNodeTruncationProcessor(
             'NodeTruncation',
             env,
             resolveProcessorOptions(config, 'NodeTruncation', {
-              maxTokensPerNode: 1200,
+              maxTokensPerNode: 4000,
             }),
           ),
         ],

packages/core/src/context/config/schema.ts

@@ -42,6 +42,11 @@ export function getContextManagementConfigSchema(
             description:
               'The absolute maximum token count allowed before synchronous truncation kicks in.',
           },
+          coalescingThresholdTokens: {
+            type: 'number',
+            description:
+              'Only trigger background consolidation (snapshots) when at least this many tokens have aged out. Prevents "turn-by-turn" utility model churn.',
+          },
         },
       },
       processorOptions: {

packages/core/src/context/config/types.ts

@@ -29,6 +29,11 @@ export interface AsyncPipelineDef {
 export interface ContextBudget {
   retainedTokens: number;
   maxTokens: number;
+  /**
+   * Only trigger background consolidation (snapshots) when at least this many
+   * tokens have aged out. Prevents "turn-by-turn" utility model churn.
+   */
+  coalescingThresholdTokens?: number;
 }
 
 /**

packages/core/src/context/contextManager.ts

@@ -141,15 +141,23 @@ export class ContextManager {
       }
 
       if (agedOutNodes.size > 0) {
-        this.env.tokenCalculator.garbageCollectCache(
-          new Set(this.buffer.nodes.map((n) => n.id)),
-        );
-        this.eventBus.emitConsolidationNeeded({
-          nodes: this.buffer.nodes,
-          targetDeficit:
-            currentTokens - this.sidecar.config.budget.retainedTokens,
-          targetNodeIds: agedOutNodes,
-        });
+        const targetDeficit =
+          currentTokens - this.sidecar.config.budget.retainedTokens;
+
+        // Respect coalescing threshold for background work
+        const threshold =
+          this.sidecar.config.budget.coalescingThresholdTokens || 0;
+
+        if (targetDeficit >= threshold) {
+          this.env.tokenCalculator.garbageCollectCache(
+            new Set(this.buffer.nodes.map((n) => n.id)),
+          );
+          this.eventBus.emitConsolidationNeeded({
+            nodes: this.buffer.nodes,
+            targetDeficit,
+            targetNodeIds: agedOutNodes,
+          });
+        }
       }
     }
   }

packages/core/src/context/processors/nodeDistillationProcessor.ts

@@ -97,9 +97,6 @@ export function createNodeDistillationProcessor(
               ]);
 
               if (newTokens < oldTokens) {
-                console.error(
-                  `[NodeDistillationProcessor] Distilled text node ${node.id} (${oldTokens} -> ${newTokens} tokens)`,
-                );
                 const distilledPayload = updatePart(payload, { text: summary });
 
                 returnedNodes.push({
@@ -154,9 +151,6 @@ export function createNodeDistillationProcessor(
                 );
 
                 if (newObsTokens < oldObsTokens) {
-                  console.error(
-                    `[NodeDistillationProcessor] Distilled tool observation for ${payload.functionResponse.name} in node ${node.id} (${oldObsTokens} -> ${newObsTokens} tokens)`,
-                  );
                   const newFR = cloneFunctionResponse(payload.functionResponse);
                   newFR.response = newObsObject;
 

packages/core/src/context/processors/stateSnapshotAsyncProcessor.ts

@@ -92,10 +92,6 @@ export function createStateSnapshotAsyncProcessor(
           options.systemInstruction,
         );
 
-        console.error(
-          `[StateSnapshotAsyncProcessor] Generated async snapshot of ${nodesToSummarize.length} nodes (${snapshotText.length} chars).`,
-        );
-
         const newConsumedIds = [
           ...previousConsumedIds,
           ...targets.map((t) => t.id),

packages/core/src/context/processors/stateSnapshotProcessor.ts

@@ -146,10 +146,6 @@ export function createStateSnapshotProcessor(
           nodesToSummarize,
           options.systemInstruction,
         );
-
-        console.error(
-          `[StateSnapshotProcessor] Generated sync snapshot of ${nodesToSummarize.length} nodes (${snapshotText.length} chars).`,
-        );
         const newId = randomUUID();
         const snapshotNode: Snapshot = {
           id: newId,

packages/core/src/context/utils/snapshotGenerator.ts

@@ -19,8 +19,6 @@ export class SnapshotGenerator {
       `You are an expert Context Memory Manager. You will be provided with a raw transcript of older conversation turns between a user and an AI assistant.
 Your task is to synthesize these turns into a single, dense, factual snapshot that preserves all critical context, preferences, active tasks, and factual knowledge.
 
-CRITICAL: You must explicitly preserve the assistant's "Current Operating Mode". If the assistant has entered "Plan Mode" or any other specialized state, this must be clearly stated in the snapshot.
-
 Discard conversational filler, pleasantries, and redundant back-and-forth iterations. Output ONLY the raw factual snapshot, formatted compactly. Do not include markdown wrappers, prefixes like "Here is the snapshot", or conversational elements.`;
 
     let userPromptText = 'TRANSCRIPT TO SNAPSHOT:\n\n';

packages/core/src/core/client.test.ts

@@ -289,6 +289,7 @@ describe('Gemini Client (client.ts)', () => {
       resetTurn: vi.fn(),
 
       isAutoDistillationEnabled: vi.fn().mockReturnValue(false),
+      isContextManagementEnabled: vi.fn().mockReturnValue(false),
       getContextManagementConfig: vi.fn().mockReturnValue({ enabled: false }),
       getModelAvailabilityService: vi
         .fn()

packages/core/src/core/geminiChat.ts

@@ -769,23 +769,17 @@ export class GeminiChat {
    *
    * @param curated - whether to return the curated history or the comprehensive
    * history.
-   * @param skipScrubbing - if true, returns the raw history without removing internal metadata.
    * @return History contents alternating between user and model for the entire
    * chat session.
    */
-  getHistory(
-    curated: boolean = false,
-    skipScrubbing: boolean = false,
-  ): readonly Content[] {
+  getHistory(curated: boolean = false): readonly Content[] {
     const history = curated
       ? extractCuratedHistory([...this.agentHistory.get()])
       : this.agentHistory.get();
 
-    if (skipScrubbing) {
-      return [...history];
-    }
-
-    return scrubHistory([...history]);
+    return this.context.config.isContextManagementEnabled()
+      ? scrubHistory([...history])
+      : [...history];
   }
 
   /**