@dexto/core 1.5.3 → 1.5.4

package/dist/llm/executor/stream-processor.cjs

@@ -49,6 +49,7 @@ class StreamProcessor {
   actualTokens = { inputTokens: 0, outputTokens: 0, totalTokens: 0 };
   finishReason = "unknown";
   reasoningText = "";
+  reasoningMetadata;
   accumulatedText = "";
   logger;
   /**
@@ -81,6 +82,9 @@ class StreamProcessor {
             break;
           case "reasoning-delta":
             this.reasoningText += event.text;
+            if (event.providerMetadata) {
+              this.reasoningMetadata = event.providerMetadata;
+            }
             if (this.streaming) {
               this.eventBus.emit("llm:chunk", {
                 chunkType: "reasoning",
@@ -210,7 +214,12 @@ class StreamProcessor {
               await this.contextManager.updateAssistantMessage(
                 this.assistantMessageId,
                 {
-                  tokenUsage: usage
+                  tokenUsage: usage,
+                  // Persist reasoning text and metadata for round-tripping
+                  ...this.reasoningText && { reasoning: this.reasoningText },
+                  ...this.reasoningMetadata && {
+                    reasoningMetadata: this.reasoningMetadata
+                  }
                 }
               );
             }
@@ -222,6 +231,9 @@ class StreamProcessor {
                 provider: this.config.provider,
                 model: this.config.model,
                 tokenUsage: usage,
+                ...this.config.estimatedInputTokens !== void 0 && {
+                  estimatedInputTokens: this.config.estimatedInputTokens
+                },
                 finishReason: this.finishReason
               });
             }
@@ -283,6 +295,9 @@ class StreamProcessor {
               provider: this.config.provider,
               model: this.config.model,
               tokenUsage: this.actualTokens,
+              ...this.config.estimatedInputTokens !== void 0 && {
+                estimatedInputTokens: this.config.estimatedInputTokens
+              },
               finishReason: "cancelled"
             });
             return {
@@ -304,6 +319,9 @@ class StreamProcessor {
           provider: this.config.provider,
           model: this.config.model,
           tokenUsage: this.actualTokens,
+          ...this.config.estimatedInputTokens !== void 0 && {
+            estimatedInputTokens: this.config.estimatedInputTokens
+          },
           finishReason: "cancelled"
         });
         return {

package/dist/llm/executor/stream-processor.d.ts

@@ -8,6 +8,8 @@ import { LLMProvider } from '../types.js';
 export interface StreamProcessorConfig {
     provider: LLMProvider;
     model: string;
+    /** Estimated input tokens before LLM call (for analytics/calibration) */
+    estimatedInputTokens?: number;
 }
 export declare class StreamProcessor {
     private contextManager;
@@ -21,6 +23,7 @@ export declare class StreamProcessor {
     private actualTokens;
     private finishReason;
     private reasoningText;
+    private reasoningMetadata;
     private accumulatedText;
     private logger;
     /**

package/dist/llm/executor/stream-processor.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"stream-processor.d.ts","sourceRoot":"","sources":["../../../src/llm/executor/stream-processor.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,gBAAgB,EAAE,OAAO,IAAI,aAAa,EAAE,MAAM,IAAI,CAAC;AAChE,OAAO,EAAE,cAAc,EAAE,MAAM,0BAA0B,CAAC;AAC1D,OAAO,EAAE,eAAe,EAAmB,MAAM,uBAAuB,CAAC;AACzE,OAAO,EAAE,eAAe,EAAE,MAAM,0BAA0B,CAAC;AAE3D,OAAO,EAAE,qBAAqB,EAAE,MAAM,YAAY,CAAC;AAGnD,OAAO,EAAE,YAAY,EAAE,MAAM,0BAA0B,CAAC;AAExD,OAAO,EAAE,WAAW,EAAc,MAAM,aAAa,CAAC;AAEtD,MAAM,WAAW,qBAAqB;IAClC,QAAQ,EAAE,WAAW,CAAC;IACtB,KAAK,EAAE,MAAM,CAAC;CACjB;AAED,qBAAa,eAAe;IAwBpB,OAAO,CAAC,cAAc;IACtB,OAAO,CAAC,QAAQ;IAChB,OAAO,CAAC,eAAe;IACvB,OAAO,CAAC,WAAW;IACnB,OAAO,CAAC,MAAM;IAEd,OAAO,CAAC,SAAS;IACjB,OAAO,CAAC,gBAAgB,CAAC;IA9B7B,OAAO,CAAC,kBAAkB,CAAuB;IACjD,OAAO,CAAC,YAAY,CAAmE;IACvF,OAAO,CAAC,YAAY,CAA8B;IAClD,OAAO,CAAC,aAAa,CAAc;IACnC,OAAO,CAAC,eAAe,CAAc;IACrC,OAAO,CAAC,MAAM,CAAe;IAC7B;;;OAGG;IACH,OAAO,CAAC,gBAAgB,CAAgD;IAExE;;;;;;;;;OASG;gBAES,cAAc,EAAE,cAAc,EAC9B,QAAQ,EAAE,eAAe,EACzB,eAAe,EAAE,eAAe,EAChC,WAAW,EAAE,WAAW,EACxB,MAAM,EAAE,qBAAqB,EACrC,MAAM,EAAE,YAAY,EACZ,SAAS,GAAE,OAAc,EACzB,gBAAgB,CAAC,EAAE,GAAG,CAC1B,MAAM,EACN;QAAE,eAAe,EAAE,OAAO,CAAC;QAAC,cAAc,CAAC,EAAE,UAAU,GAAG,UAAU,CAAA;KAAE,CACzE,YAAA;IAKC,OAAO,CACT,QAAQ,EAAE,MAAM,gBAAgB,CAAC,aAAa,EAAE,OAAO,CAAC,GACzD,OAAO,CAAC,qBAAqB,CAAC;YAmanB,sBAAsB;YAKtB,gBAAgB;IAO9B;;;;OAIG;YACW,2BAA2B;CAmC5C"}
+{"version":3,"file":"stream-processor.d.ts","sourceRoot":"","sources":["../../../src/llm/executor/stream-processor.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,gBAAgB,EAAE,OAAO,IAAI,aAAa,EAAE,MAAM,IAAI,CAAC;AAChE,OAAO,EAAE,cAAc,EAAE,MAAM,0BAA0B,CAAC;AAC1D,OAAO,EAAE,eAAe,EAAmB,MAAM,uBAAuB,CAAC;AACzE,OAAO,EAAE,eAAe,EAAE,MAAM,0BAA0B,CAAC;AAE3D,OAAO,EAAE,qBAAqB,EAAE,MAAM,YAAY,CAAC;AAGnD,OAAO,EAAE,YAAY,EAAE,MAAM,0BAA0B,CAAC;AAExD,OAAO,EAAE,WAAW,EAAc,MAAM,aAAa,CAAC;AAEtD,MAAM,WAAW,qBAAqB;IAClC,QAAQ,EAAE,WAAW,CAAC;IACtB,KAAK,EAAE,MAAM,CAAC;IACd,yEAAyE;IACzE,oBAAoB,CAAC,EAAE,MAAM,CAAC;CACjC;AAED,qBAAa,eAAe;IAyBpB,OAAO,CAAC,cAAc;IACtB,OAAO,CAAC,QAAQ;IAChB,OAAO,CAAC,eAAe;IACvB,OAAO,CAAC,WAAW;IACnB,OAAO,CAAC,MAAM;IAEd,OAAO,CAAC,SAAS;IACjB,OAAO,CAAC,gBAAgB,CAAC;IA/B7B,OAAO,CAAC,kBAAkB,CAAuB;IACjD,OAAO,CAAC,YAAY,CAAmE;IACvF,OAAO,CAAC,YAAY,CAA8B;IAClD,OAAO,CAAC,aAAa,CAAc;IACnC,OAAO,CAAC,iBAAiB,CAAsC;IAC/D,OAAO,CAAC,eAAe,CAAc;IACrC,OAAO,CAAC,MAAM,CAAe;IAC7B;;;OAGG;IACH,OAAO,CAAC,gBAAgB,CAAgD;IAExE;;;;;;;;;OASG;gBAES,cAAc,EAAE,cAAc,EAC9B,QAAQ,EAAE,eAAe,EACzB,eAAe,EAAE,eAAe,EAChC,WAAW,EAAE,WAAW,EACxB,MAAM,EAAE,qBAAqB,EACrC,MAAM,EAAE,YAAY,EACZ,SAAS,GAAE,OAAc,EACzB,gBAAgB,CAAC,EAAE,GAAG,CAC1B,MAAM,EACN;QAAE,eAAe,EAAE,OAAO,CAAC;QAAC,cAAc,CAAC,EAAE,UAAU,GAAG,UAAU,CAAA;KAAE,CACzE,YAAA;IAKC,OAAO,CACT,QAAQ,EAAE,MAAM,gBAAgB,CAAC,aAAa,EAAE,OAAO,CAAC,GACzD,OAAO,CAAC,qBAAqB,CAAC;YAubnB,sBAAsB;YAKtB,gBAAgB;IAO9B;;;;OAIG;YACW,2BAA2B;CAmC5C"}

package/dist/llm/executor/stream-processor.js

@@ -27,6 +27,7 @@ class StreamProcessor {
   actualTokens = { inputTokens: 0, outputTokens: 0, totalTokens: 0 };
   finishReason = "unknown";
   reasoningText = "";
+  reasoningMetadata;
   accumulatedText = "";
   logger;
   /**
@@ -59,6 +60,9 @@ class StreamProcessor {
             break;
           case "reasoning-delta":
             this.reasoningText += event.text;
+            if (event.providerMetadata) {
+              this.reasoningMetadata = event.providerMetadata;
+            }
             if (this.streaming) {
               this.eventBus.emit("llm:chunk", {
                 chunkType: "reasoning",
@@ -188,7 +192,12 @@ class StreamProcessor {
               await this.contextManager.updateAssistantMessage(
                 this.assistantMessageId,
                 {
-                  tokenUsage: usage
+                  tokenUsage: usage,
+                  // Persist reasoning text and metadata for round-tripping
+                  ...this.reasoningText && { reasoning: this.reasoningText },
+                  ...this.reasoningMetadata && {
+                    reasoningMetadata: this.reasoningMetadata
+                  }
                 }
               );
             }
@@ -200,6 +209,9 @@ class StreamProcessor {
                 provider: this.config.provider,
                 model: this.config.model,
                 tokenUsage: usage,
+                ...this.config.estimatedInputTokens !== void 0 && {
+                  estimatedInputTokens: this.config.estimatedInputTokens
+                },
                 finishReason: this.finishReason
               });
             }
@@ -261,6 +273,9 @@ class StreamProcessor {
               provider: this.config.provider,
               model: this.config.model,
               tokenUsage: this.actualTokens,
+              ...this.config.estimatedInputTokens !== void 0 && {
+                estimatedInputTokens: this.config.estimatedInputTokens
+              },
               finishReason: "cancelled"
             });
             return {
@@ -282,6 +297,9 @@ class StreamProcessor {
           provider: this.config.provider,
           model: this.config.model,
           tokenUsage: this.actualTokens,
+          ...this.config.estimatedInputTokens !== void 0 && {
+            estimatedInputTokens: this.config.estimatedInputTokens
+          },
           finishReason: "cancelled"
         });
         return {

package/dist/llm/executor/turn-executor.cjs

@@ -91,7 +91,7 @@ var import_reactive_overflow = require("../../context/compaction/strategies/reac
 const toolSupportCache = /* @__PURE__ */ new Map();
 const LOCAL_PROVIDERS = ["ollama", "local"];
 class TurnExecutor {
-  constructor(model, toolManager, contextManager, eventBus, resourceManager, sessionId, config, llmContext, logger, messageQueue, modelLimits, externalSignal, compactionStrategy) {
+  constructor(model, toolManager, contextManager, eventBus, resourceManager, sessionId, config, llmContext, logger, messageQueue, modelLimits, externalSignal, compactionStrategy, compactionThresholdPercent = 1) {
     this.model = model;
     this.toolManager = toolManager;
     this.contextManager = contextManager;
@@ -103,6 +103,7 @@ class TurnExecutor {
     this.messageQueue = messageQueue;
     this.modelLimits = modelLimits;
     this.externalSignal = externalSignal;
+    this.compactionThresholdPercent = compactionThresholdPercent;
     this.logger = logger.createChild(import_types4.DextoLogComponent.EXECUTOR);
     this.stepAbortController = new AbortController();
     if (compactionStrategy !== void 0) {
@@ -123,13 +124,32 @@ class TurnExecutor {
    * Used to pass approval info from tool execution to result persistence.
    */
   approvalMetadata = /* @__PURE__ */ new Map();
+  /**
+   * Tracks whether compaction occurred during this turn.
+   * Used to signal to the caller that session continuation may be needed.
+   */
+  compactionOccurred = false;
+  /**
+   * Compaction data captured during this turn.
+   * Contains summary text and preserved messages for session continuation.
+   * This data is passed up the call chain (NOT persisted to original session).
+   */
+  compactionData = null;
+  /**
+   * Virtual context for remaining iterations after compaction.
+   * When set, the main loop uses this instead of calling getFormattedMessagesForLLM().
+   * This provides reduced context to the LLM without persisting to the original session.
+   */
+  virtualContext = null;
   /**
    * Get StreamProcessor config from TurnExecutor state.
+   * @param estimatedInputTokens Optional estimated input tokens for analytics
    */
-  getStreamProcessorConfig() {
+  getStreamProcessorConfig(estimatedInputTokens) {
     return {
       provider: this.llmContext.provider,
-      model: this.llmContext.model
+      model: this.llmContext.model,
+      ...estimatedInputTokens !== void 0 && { estimatedInputTokens }
     };
   }
   /**
@@ -185,13 +205,34 @@ class TurnExecutor {
           if (coalesced) {
             await this.injectQueuedMessages(coalesced);
           }
-          if (lastStepTokens && this.checkAndHandleOverflow(lastStepTokens)) {
-            await this.compress(lastStepTokens.inputTokens ?? 0);
-          }
-          const prepared = await this.contextManager.getFormattedMessagesWithCompression(
+          await this.pruneOldToolOutputs();
+          let prepared = this.virtualContext ? await this.buildMessagesFromVirtualContext(contributorContext) : await this.contextManager.getFormattedMessagesForLLM(
             contributorContext,
             this.llmContext
           );
+          const toolDefinitions = supportsTools ? await this.toolManager.getAllTools() : {};
+          let estimatedTokens = await this.contextManager.getEstimatedNextInputTokens(
+            prepared.systemPrompt,
+            prepared.preparedHistory,
+            toolDefinitions
+          );
+          if (this.shouldCompact(estimatedTokens)) {
+            this.logger.debug(
+              `Pre-check: estimated ${estimatedTokens} tokens exceeds threshold, compacting`
+            );
+            await this.compactToVirtualContext(estimatedTokens);
+            if (this.virtualContext) {
+              prepared = await this.buildMessagesFromVirtualContext(contributorContext);
+              estimatedTokens = await this.contextManager.getEstimatedNextInputTokens(
+                prepared.systemPrompt,
+                prepared.preparedHistory,
+                toolDefinitions
+              );
+              this.logger.debug(
+                `Post-compaction: recomputed estimate is ${estimatedTokens} tokens`
+              );
+            }
+          }
           this.logger.debug(`Step ${stepCount}: Starting`);
           const tools = supportsTools ? await this.createTools() : {};
           const streamProcessor = new import_stream_processor.StreamProcessor(
@@ -199,7 +240,7 @@ class TurnExecutor {
             this.eventBus,
             this.resourceManager,
             this.stepAbortController.signal,
-            this.getStreamProcessorConfig(),
+            this.getStreamProcessorConfig(estimatedTokens),
             this.logger,
             streaming,
             this.approvalMetadata
@@ -238,6 +279,28 @@ class TurnExecutor {
           this.logger.debug(
             `Step ${stepCount}: Finished with reason="${result.finishReason}", tokens=${JSON.stringify(result.usage)}`
           );
+          if (result.finishReason === "cancelled") {
+            this.logger.info(
+              `Context estimation (cancelled): keeping last known actuals, partial response (${result.text.length} chars) will be estimated`
+            );
+          } else if (result.usage?.inputTokens !== void 0) {
+            const diff = estimatedTokens - result.usage.inputTokens;
+            const diffPercent = result.usage.inputTokens > 0 ? (diff / result.usage.inputTokens * 100).toFixed(1) : "0.0";
+            this.logger.info(
+              `Context estimation accuracy: estimated=${estimatedTokens}, actual=${result.usage.inputTokens}, error=${diff} (${diffPercent}%)`
+            );
+            this.contextManager.setLastActualInputTokens(result.usage.inputTokens);
+            if (result.usage?.outputTokens !== void 0) {
+              this.contextManager.setLastActualOutputTokens(result.usage.outputTokens);
+            }
+            await this.contextManager.recordLastCallMessageCount();
+          }
+          if (!this.virtualContext && result.usage?.inputTokens && this.shouldCompactFromActual(result.usage.inputTokens)) {
+            this.logger.debug(
+              `Post-response: actual ${result.usage.inputTokens} tokens exceeds threshold, compacting`
+            );
+            await this.compactToVirtualContext(result.usage.inputTokens);
+          }
           if (result.finishReason !== "tool-calls") {
             const queuedOnTerminate = this.messageQueue.dequeueAll();
             if (queuedOnTerminate) {
@@ -261,7 +324,6 @@ class TurnExecutor {
             lastFinishReason = "max-steps";
             break;
           }
-          await this.pruneOldToolOutputs();
         }
       } catch (error) {
         const mappedError = this.mapProviderError(error);
@@ -291,7 +353,14 @@ class TurnExecutor {
         text: lastText,
         stepCount,
         usage: lastStepTokens,
-        finishReason: lastFinishReason
+        finishReason: lastFinishReason,
+        // Signal to caller that compaction occurred during this turn
+        // Caller can use this to trigger session-native continuation
+        didCompact: this.compactionOccurred,
+        // Pass compaction data up the chain (NOT persisted to original session)
+        // Caller uses this to create the continuation session with summary
+        // Use spread to conditionally include only when data exists (exactOptionalPropertyTypes)
+        ...this.compactionData && { compaction: this.compactionData }
       };
     } catch (_2) {
       var _error = _2, _hasError = true;
@@ -589,7 +658,7 @@ class TurnExecutor {
   /**
    * Prunes old tool outputs by marking them with compactedAt timestamp.
    * Does NOT modify content - transformation happens at format time in
-   * ContextManager.getFormattedMessagesWithCompression().
+   * ContextManager.prepareHistory().
    *
    * Algorithm:
    * 1. Go backwards through history (most recent first)
@@ -658,54 +727,174 @@ class TurnExecutor {
     this.messageQueue.clear();
   }
   /**
-   * Check if context has overflowed based on actual token usage from API.
+   * Check if context should be compacted based on estimated token count.
+   * Uses the threshold percentage from compaction config to trigger earlier (e.g., at 90%).
+   *
+   * @param estimatedTokens Estimated token count from the current context
+   * @returns true if compaction is needed before making the LLM call
+   */
+  shouldCompact(estimatedTokens) {
+    if (!this.modelLimits || !this.compactionStrategy) {
+      return false;
+    }
+    return (0, import_overflow.isOverflow)(
+      { inputTokens: estimatedTokens },
+      this.modelLimits,
+      this.compactionThresholdPercent
+    );
+  }
+  /**
+   * Check if context should be compacted based on actual token count from API response.
+   * This is a post-response check using real token counts rather than estimates.
+   *
+   * @param actualTokens Actual input token count from the API response
+   * @returns true if compaction is needed
    */
-  checkAndHandleOverflow(tokens) {
+  shouldCompactFromActual(actualTokens) {
     if (!this.modelLimits || !this.compactionStrategy) {
       return false;
     }
-    return (0, import_overflow.isOverflow)(tokens, this.modelLimits);
+    return (0, import_overflow.isOverflow)(
+      { inputTokens: actualTokens },
+      this.modelLimits,
+      this.compactionThresholdPercent
+    );
   }
   /**
-   * Compress context using ReactiveOverflowStrategy.
+   * Compact context using ReactiveOverflowStrategy WITHOUT persisting to original session.
+   *
+   * Key design: Creates a virtual context (summary + preserved messages) that will be used
+   * for the remaining iterations of this turn. The compaction data is passed up the call chain
+   * so the caller can create a continuation session with the summary.
    *
-   * Generates a summary of older messages and adds it to history.
-   * The actual token reduction happens at read-time via filterCompacted()
-   * in getFormattedMessagesWithCompression().
+   * The original session remains UNTOUCHED - no messages are added or modified.
    *
-   * @param originalTokens The actual input token count from API that triggered overflow
+   * @param originalTokens The estimated input token count that triggered overflow
    */
-  async compress(originalTokens) {
+  async compactToVirtualContext(originalTokens) {
     if (!this.compactionStrategy) {
       return;
     }
     this.logger.info(
-      `Context overflow detected (${originalTokens} tokens), running compression`
+      `Context overflow detected (${originalTokens} tokens), checking if compression is possible`
     );
     const history = await this.contextManager.getHistory();
+    if (history.length < 4) {
+      this.logger.debug("Compaction skipped: history too short to summarize");
+      return;
+    }
+    this.eventBus.emit("context:compacting", {
+      estimatedTokens: originalTokens
+    });
     const summaryMessages = await this.compactionStrategy.compact(history);
     if (summaryMessages.length === 0) {
-      this.logger.debug("Compaction returned no summary (history too short)");
+      this.logger.debug(
+        "Compaction skipped: strategy returned no summary (likely already compacted or nothing to summarize)"
+      );
+      this.eventBus.emit("context:compacted", {
+        originalTokens,
+        compactedTokens: originalTokens,
+        // No change
+        originalMessages: history.length,
+        compactedMessages: history.length,
+        // No change
+        strategy: this.compactionStrategy.name,
+        reason: "overflow"
+      });
+      return;
+    }
+    const summaryMessage = summaryMessages[0];
+    if (!summaryMessage) {
+      this.logger.warn("Compaction returned empty summary message array");
       return;
     }
-    for (const summary of summaryMessages) {
-      await this.contextManager.addMessage(summary);
+    const summaryText = this.extractSummaryText(summaryMessage);
+    const summarizedCount = summaryMessage.metadata?.originalMessageCount ?? 0;
+    let existingSummaryIndex = -1;
+    for (let i = history.length - 1; i >= 0; i--) {
+      const msg = history[i];
+      if (msg?.metadata?.isSummary === true || msg?.metadata?.isSessionSummary === true) {
+        existingSummaryIndex = i;
+        break;
+      }
     }
-    const { filterCompacted, estimateMessagesTokens } = await import("../../context/utils.js");
-    const updatedHistory = await this.contextManager.getHistory();
-    const filteredHistory = filterCompacted(updatedHistory);
-    const compactedTokens = estimateMessagesTokens(filteredHistory);
+    const baseIndex = existingSummaryIndex >= 0 ? existingSummaryIndex + 1 : 0;
+    const preservedMessages = history.slice(baseIndex + summarizedCount);
+    const firstTimestamp = summaryMessage.metadata?.originalFirstTimestamp;
+    const lastTimestamp = summaryMessage.metadata?.originalLastTimestamp;
+    this.compactionData = {
+      summaryText,
+      preservedMessages: [...preservedMessages],
+      // Copy to avoid mutation
+      summarizedCount,
+      ...firstTimestamp !== void 0 && { originalFirstTimestamp: firstTimestamp },
+      ...lastTimestamp !== void 0 && { originalLastTimestamp: lastTimestamp }
+    };
+    this.virtualContext = {
+      summaryMessage,
+      preservedMessages: [...preservedMessages]
+    };
+    this.compactionOccurred = true;
+    this.contextManager.resetActualTokenTracking();
+    const { estimateMessagesTokens: estimateTokens } = await import("../../context/utils.js");
+    const virtualMessages = [summaryMessage, ...preservedMessages];
+    const compactedTokens = estimateTokens(virtualMessages);
     this.eventBus.emit("context:compacted", {
       originalTokens,
       compactedTokens,
       originalMessages: history.length,
-      compactedMessages: filteredHistory.length,
+      compactedMessages: virtualMessages.length,
       strategy: this.compactionStrategy.name,
       reason: "overflow"
     });
     this.logger.info(
-      `Compaction complete: ${originalTokens} \u2192 ~${compactedTokens} tokens (${history.length} \u2192 ${filteredHistory.length} messages after filtering)`
+      `Compaction complete (virtual context): ${originalTokens} \u2192 ~${compactedTokens} tokens (${history.length} \u2192 ${virtualMessages.length} messages). Original session unchanged - summary will be passed to continuation session.`
+    );
+  }
+  /**
+   * Extract the summary text from a summary message.
+   */
+  extractSummaryText(summaryMessage) {
+    if (typeof summaryMessage.content === "string") {
+      return summaryMessage.content;
+    }
+    if (Array.isArray(summaryMessage.content)) {
+      return summaryMessage.content.filter((part) => part.type === "text").map((part) => part.text).join("\n");
+    }
+    return "";
+  }
+  /**
+   * Build formatted messages from virtual context (after compaction).
+   *
+   * This creates LLM-ready messages using:
+   * - System prompt (same as normal flow)
+   * - Summary message (as first message)
+   * - Preserved messages (formatted for LLM)
+   *
+   * Uses the same formatting pipeline as getFormattedMessagesForLLM()
+   * but with our virtual history instead of the stored history.
+   *
+   * @param contributorContext Context for system prompt contributors
+   * @returns Formatted messages ready for LLM call, matching getFormattedMessagesForLLM return type
+   */
+  async buildMessagesFromVirtualContext(contributorContext) {
+    if (!this.virtualContext) {
+      throw new Error("buildMessagesFromVirtualContext called without virtual context");
+    }
+    const { summaryMessage, preservedMessages } = this.virtualContext;
+    const systemPrompt = await this.contextManager.getSystemPrompt(contributorContext);
+    const virtualHistory = [summaryMessage, ...preservedMessages];
+    const formattedMessages = await this.contextManager.getFormattedMessages(
+      contributorContext,
+      this.llmContext,
+      systemPrompt,
+      virtualHistory
     );
+    return {
+      formattedMessages,
+      systemPrompt,
+      preparedHistory: virtualHistory
+    };
   }
   /**
    * Set telemetry span attributes for token usage.

package/dist/llm/executor/turn-executor.d.ts

@@ -34,6 +34,7 @@ export declare class TurnExecutor {
     private messageQueue;
     private modelLimits?;
     private externalSignal?;
+    private compactionThresholdPercent;
     private logger;
     /**
      * Per-step abort controller. Created fresh for each iteration of the loop.
@@ -46,15 +47,33 @@ export declare class TurnExecutor {
      * Used to pass approval info from tool execution to result persistence.
      */
     private approvalMetadata;
+    /**
+     * Tracks whether compaction occurred during this turn.
+     * Used to signal to the caller that session continuation may be needed.
+     */
+    private compactionOccurred;
+    /**
+     * Compaction data captured during this turn.
+     * Contains summary text and preserved messages for session continuation.
+     * This data is passed up the call chain (NOT persisted to original session).
+     */
+    private compactionData;
+    /**
+     * Virtual context for remaining iterations after compaction.
+     * When set, the main loop uses this instead of calling getFormattedMessagesForLLM().
+     * This provides reduced context to the LLM without persisting to the original session.
+     */
+    private virtualContext;
     constructor(model: LanguageModel, toolManager: ToolManager, contextManager: ContextManager<ModelMessage>, eventBus: SessionEventBus, resourceManager: ResourceManager, sessionId: string, config: {
         maxSteps?: number | undefined;
         maxOutputTokens?: number | undefined;
         temperature?: number | undefined;
         baseURL?: string | undefined;
         reasoningEffort?: 'none' | 'minimal' | 'low' | 'medium' | 'high' | 'xhigh' | undefined;
-    }, llmContext: LLMContext, logger: IDextoLogger, messageQueue: MessageQueueService, modelLimits?: ModelLimits | undefined, externalSignal?: AbortSignal | undefined, compactionStrategy?: ICompactionStrategy | null);
+    }, llmContext: LLMContext, logger: IDextoLogger, messageQueue: MessageQueueService, modelLimits?: ModelLimits | undefined, externalSignal?: AbortSignal | undefined, compactionStrategy?: ICompactionStrategy | null, compactionThresholdPercent?: number);
     /**
      * Get StreamProcessor config from TurnExecutor state.
+     * @param estimatedInputTokens Optional estimated input tokens for analytics
      */
     private getStreamProcessorConfig;
     /**
@@ -119,7 +138,7 @@ export declare class TurnExecutor {
     /**
      * Prunes old tool outputs by marking them with compactedAt timestamp.
      * Does NOT modify content - transformation happens at format time in
-     * ContextManager.getFormattedMessagesWithCompression().
+     * ContextManager.prepareHistory().
      *
      * Algorithm:
      * 1. Go backwards through history (most recent first)
@@ -143,19 +162,52 @@ export declare class TurnExecutor {
      */
     private cleanup;
     /**
-     * Check if context has overflowed based on actual token usage from API.
+     * Check if context should be compacted based on estimated token count.
+     * Uses the threshold percentage from compaction config to trigger earlier (e.g., at 90%).
+     *
+     * @param estimatedTokens Estimated token count from the current context
+     * @returns true if compaction is needed before making the LLM call
+     */
+    private shouldCompact;
+    /**
+     * Check if context should be compacted based on actual token count from API response.
+     * This is a post-response check using real token counts rather than estimates.
+     *
+     * @param actualTokens Actual input token count from the API response
+     * @returns true if compaction is needed
      */
-    private checkAndHandleOverflow;
+    private shouldCompactFromActual;
     /**
-     * Compress context using ReactiveOverflowStrategy.
+     * Compact context using ReactiveOverflowStrategy WITHOUT persisting to original session.
      *
-     * Generates a summary of older messages and adds it to history.
-     * The actual token reduction happens at read-time via filterCompacted()
-     * in getFormattedMessagesWithCompression().
+     * Key design: Creates a virtual context (summary + preserved messages) that will be used
+     * for the remaining iterations of this turn. The compaction data is passed up the call chain
+     * so the caller can create a continuation session with the summary.
      *
-     * @param originalTokens The actual input token count from API that triggered overflow
+     * The original session remains UNTOUCHED - no messages are added or modified.
+     *
+     * @param originalTokens The estimated input token count that triggered overflow
+     */
+    private compactToVirtualContext;
+    /**
+     * Extract the summary text from a summary message.
+     */
+    private extractSummaryText;
+    /**
+     * Build formatted messages from virtual context (after compaction).
+     *
+     * This creates LLM-ready messages using:
+     * - System prompt (same as normal flow)
+     * - Summary message (as first message)
+     * - Preserved messages (formatted for LLM)
+     *
+     * Uses the same formatting pipeline as getFormattedMessagesForLLM()
+     * but with our virtual history instead of the stored history.
+     *
+     * @param contributorContext Context for system prompt contributors
+     * @returns Formatted messages ready for LLM call, matching getFormattedMessagesForLLM return type
      */
-    private compress;
+    private buildMessagesFromVirtualContext;
     /**
      * Set telemetry span attributes for token usage.
      */

package/dist/llm/executor/turn-executor.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"turn-executor.d.ts","sourceRoot":"","sources":["../../../src/llm/executor/turn-executor.ts"],"names":[],"mappings":"AAAA,OAAO,EACH,aAAa,EAMb,KAAK,YAAY,EAEpB,MAAM,IAAI,CAAC;AAEZ,OAAO,EAAE,cAAc,EAAE,MAAM,0BAA0B,CAAC;AAE1D,OAAO,EAAE,WAAW,EAAE,MAAM,6BAA6B,CAAC;AAG1D,OAAO,EAAE,cAAc,EAAE,MAAM,YAAY,CAAC;AAG5C,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,0BAA0B,CAAC;AAE7D,OAAO,KAAK,EAAE,eAAe,EAAmB,MAAM,uBAAuB,CAAC;AAC9E,OAAO,KAAK,EAAE,eAAe,EAAE,MAAM,0BAA0B,CAAC;AAChE,OAAO,EAAE,yBAAyB,EAAE,MAAM,6BAA6B,CAAC;AACxE,OAAO,EAAE,UAAU,EAAoB,MAAM,aAAa,CAAC;AAC3D,OAAO,KAAK,EAAE,mBAAmB,EAAE,MAAM,gCAAgC,CAAC;AAQ1E,OAAO,EAAc,KAAK,WAAW,EAAE,MAAM,sCAAsC,CAAC;AAEpF,OAAO,KAAK,EAAE,mBAAmB,EAAE,MAAM,mCAAmC,CAAC;AAc7E;;;;;;;;;;;GAWG;AACH,qBAAa,YAAY;IAkBjB,OAAO,CAAC,KAAK;IACb,OAAO,CAAC,WAAW;IACnB,OAAO,CAAC,cAAc;IACtB,OAAO,CAAC,QAAQ;IAChB,OAAO,CAAC,eAAe;IACvB,OAAO,CAAC,SAAS;IACjB,OAAO,CAAC,MAAM;IAQd,OAAO,CAAC,UAAU;IAElB,OAAO,CAAC,YAAY;IACpB,OAAO,CAAC,WAAW,CAAC;IACpB,OAAO,CAAC,cAAc,CAAC;IAnC3B,OAAO,CAAC,MAAM,CAAe;IAC7B;;;OAGG;IACH,OAAO,CAAC,mBAAmB,CAAkB;IAC7C,OAAO,CAAC,kBAAkB,CAAoC;IAC9D;;;OAGG;IACH,OAAO,CAAC,gBAAgB,CAGpB;gBAGQ,KAAK,EAAE,aAAa,EACpB,WAAW,EAAE,WAAW,EACxB,cAAc,EAAE,cAAc,CAAC,YAAY,CAAC,EAC5C,QAAQ,EAAE,eAAe,EACzB,eAAe,EAAE,eAAe,EAChC,SAAS,EAAE,MAAM,EACjB,MAAM,EAAE;QACZ,QAAQ,CAAC,EAAE,MAAM,GAAG,SAAS,CAAC;QAC9B,eAAe,CAAC,EAAE,MAAM,GAAG,SAAS,CAAC;QACrC,WAAW,CAAC,EAAE,MAAM,GAAG,SAAS,CAAC;QACjC,OAAO,CAAC,EAAE,MAAM,GAAG,SAAS,CAAC;QAE7B,eAAe,CAAC,EAAE,MAAM,GAAG,SAAS,GAAG,KAAK,GAAG,QAAQ,GAAG,MAAM,GAAG,OAAO,GAAG,SAAS,CAAC;KAC1F,EACO,UAAU,EAAE,UAAU,EAC9B,MAAM,EAAE,YAAY,EACZ,YAAY,EAAE,mBAAmB,EACjC,WAAW,CAAC,EAAE,WAAW,YAAA,EACzB,cAAc,CAAC,EAAE,WAAW,YAAA,EACpC,kBAAkB,CAAC,EAAE,mBAAmB,GAAG,IAAI;IAqBnD;;OAEG;IACH,OAAO,CAAC,wBAAwB;IAOhC;;;;;;OAMG;IACG,OAAO,CACT,kBAAkB,EAAE,yBAAyB,EAC7C,SAAS,GAAE,OAAc,GAC1B,OAAO,CAAC,cAAc,CAAC;IAuN1B;;;OAGG;IACH,KAAK,IAAI,IAAI;IAIb;;;OAGG;YACW,oBAAoB;IAmBlC;;;;;;OAMG;YACW,mBAAmB;IA0EjC;;;;;;;OAOG;YACW,WAAW;IA0GzB;;;;;OAKG;IACH,OAAO,CAAC,sBAAsB;IAuF9B;;OAEG;IACH,OAAO,CAAC,gBAAgB;IAyBxB;;OAEG;IACH,OAAO,CAAC,eAAe;IAavB;;OAEG;IACH,OAAO,CAAC,oBAAoB;IAS5B;;OAEG;IACH,OAAO,CAAC,MAAM,CAAC,QAAQ,CAAC,aAAa,CAAU;IAC/C,OAAO,CAAC,MAAM,CAAC,QAAQ,CAAC,aAAa,CAAU;IAE/C;;;;;;;;;;;OAWG;YACW,mBAAmB;IAkDjC;;;;;;OAMG;IACH,OAAO,CAAC,kBAAkB;IAgB1B;;;OAGG;IACH,OAAO,CAAC,OAAO;IAYf;;OAEG;IACH,OAAO,CAAC,sBAAsB;IAO9B;;;;;;;;OAQG;YACW,QAAQ;IA8CtB;;OAEG;IACH,OAAO,CAAC,sBAAsB;IAoB9B;;OAEG;IACH,OAAO,CAAC,gBAAgB;CA0D3B"}
+{"version":3,"file":"turn-executor.d.ts","sourceRoot":"","sources":["../../../src/llm/executor/turn-executor.ts"],"names":[],"mappings":"AAAA,OAAO,EACH,aAAa,EAMb,KAAK,YAAY,EAEpB,MAAM,IAAI,CAAC;AAEZ,OAAO,EAAE,cAAc,EAAE,MAAM,0BAA0B,CAAC;AAQ1D,OAAO,EAAE,WAAW,EAAE,MAAM,6BAA6B,CAAC;AAG1D,OAAO,EAAE,cAAc,EAAkB,MAAM,YAAY,CAAC;AAG5D,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,0BAA0B,CAAC;AAE7D,OAAO,KAAK,EAAE,eAAe,EAAmB,MAAM,uBAAuB,CAAC;AAC9E,OAAO,KAAK,EAAE,eAAe,EAAE,MAAM,0BAA0B,CAAC;AAChE,OAAO,EAAE,yBAAyB,EAAE,MAAM,6BAA6B,CAAC;AACxE,OAAO,EAAE,UAAU,EAAoB,MAAM,aAAa,CAAC;AAC3D,OAAO,KAAK,EAAE,mBAAmB,EAAE,MAAM,gCAAgC,CAAC;AAQ1E,OAAO,EAAc,KAAK,WAAW,EAAE,MAAM,sCAAsC,CAAC;AAEpF,OAAO,KAAK,EAAE,mBAAmB,EAAE,MAAM,mCAAmC,CAAC;AAc7E;;;;;;;;;;;GAWG;AACH,qBAAa,YAAY;IAyCjB,OAAO,CAAC,KAAK;IACb,OAAO,CAAC,WAAW;IACnB,OAAO,CAAC,cAAc;IACtB,OAAO,CAAC,QAAQ;IAChB,OAAO,CAAC,eAAe;IACvB,OAAO,CAAC,SAAS;IACjB,OAAO,CAAC,MAAM;IAQd,OAAO,CAAC,UAAU;IAElB,OAAO,CAAC,YAAY;IACpB,OAAO,CAAC,WAAW,CAAC;IACpB,OAAO,CAAC,cAAc,CAAC;IAEvB,OAAO,CAAC,0BAA0B;IA5DtC,OAAO,CAAC,MAAM,CAAe;IAC7B;;;OAGG;IACH,OAAO,CAAC,mBAAmB,CAAkB;IAC7C,OAAO,CAAC,kBAAkB,CAAoC;IAC9D;;;OAGG;IACH,OAAO,CAAC,gBAAgB,CAGpB;IAEJ;;;OAGG;IACH,OAAO,CAAC,kBAAkB,CAAS;IAEnC;;;;OAIG;IACH,OAAO,CAAC,cAAc,CAA+B;IAErD;;;;OAIG;IACH,OAAO,CAAC,cAAc,CAGN;gBAGJ,KAAK,EAAE,aAAa,EACpB,WAAW,EAAE,WAAW,EACxB,cAAc,EAAE,cAAc,CAAC,YAAY,CAAC,EAC5C,QAAQ,EAAE,eAAe,EACzB,eAAe,EAAE,eAAe,EAChC,SAAS,EAAE,MAAM,EACjB,MAAM,EAAE;QACZ,QAAQ,CAAC,EAAE,MAAM,GAAG,SAAS,CAAC;QAC9B,eAAe,CAAC,EAAE,MAAM,GAAG,SAAS,CAAC;QACrC,WAAW,CAAC,EAAE,MAAM,GAAG,SAAS,CAAC;QACjC,OAAO,CAAC,EAAE,MAAM,GAAG,SAAS,CAAC;QAE7B,eAAe,CAAC,EAAE,MAAM,GAAG,SAAS,GAAG,KAAK,GAAG,QAAQ,GAAG,MAAM,GAAG,OAAO,GAAG,SAAS,CAAC;KAC1F,EACO,UAAU,EAAE,UAAU,EAC9B,MAAM,EAAE,YAAY,EACZ,YAAY,EAAE,mBAAmB,EACjC,WAAW,CAAC,EAAE,WAAW,YAAA,EACzB,cAAc,CAAC,EAAE,WAAW,YAAA,EACpC,kBAAkB,CAAC,EAAE,mBAAmB,GAAG,IAAI,EACvC,0BAA0B,GAAE,MAAY;IAqBpD;;;OAGG;IACH,OAAO,CAAC,wBAAwB;IAQhC;;;;;;OAMG;IACG,OAAO,CACT,kBAAkB,EAAE,yBAAyB,EAC7C,SAAS,GAAE,OAAc,GAC1B,OAAO,CAAC,cAAc,CAAC;IAwT1B;;;OAGG;IACH,KAAK,IAAI,IAAI;IAIb;;;OAGG;YACW,oBAAoB;IAmBlC;;;;;;OAMG;YACW,mBAAmB;IA0EjC;;;;;;;OAOG;YACW,WAAW;IA0GzB;;;;;OAKG;IACH,OAAO,CAAC,sBAAsB;IAuF9B;;OAEG;IACH,OAAO,CAAC,gBAAgB;IAyBxB;;OAEG;IACH,OAAO,CAAC,eAAe;IAavB;;OAEG;IACH,OAAO,CAAC,oBAAoB;IAS5B;;OAEG;IACH,OAAO,CAAC,MAAM,CAAC,QAAQ,CAAC,aAAa,CAAU;IAC/C,OAAO,CAAC,MAAM,CAAC,QAAQ,CAAC,aAAa,CAAU;IAE/C;;;;;;;;;;;OAWG;YACW,mBAAmB;IAkDjC;;;;;;OAMG;IACH,OAAO,CAAC,kBAAkB;IAgB1B;;;OAGG;IACH,OAAO,CAAC,OAAO;IAYf;;;;;;OAMG;IACH,OAAO,CAAC,aAAa;IAYrB;;;;;;OAMG;IACH,OAAO,CAAC,uBAAuB;IAY/B;;;;;;;;;;OAUG;YACW,uBAAuB;IA4HrC;;OAEG;IACH,OAAO,CAAC,kBAAkB;IAa1B;;;;;;;;;;;;;OAaG;YACW,+BAA+B;IAmC7C;;OAEG;IACH,OAAO,CAAC,sBAAsB;IAoB9B;;OAEG;IACH,OAAO,CAAC,gBAAgB;CA0D3B"}