blun-king-cli 9.1.219 → 9.1.221

package/bin/pending-token-estimate-policy.cjs

@@ -0,0 +1,41 @@
+'use strict';
+
+function createPendingTokenEstimateCache(estimateMessages) {
+  if (typeof estimateMessages !== 'function') {
+    throw new TypeError('estimateMessages must be a function');
+  }
+
+  let cachedHistory = null;
+  let cachedCoveredCount = -1;
+  let cachedRevision = -1;
+  let cachedEstimate = 0;
+
+  return {
+    read(history, coveredMessageCount, revision) {
+      const messages = Array.isArray(history) ? history : [];
+      const covered = Math.max(
+        0,
+        Math.min(messages.length, Math.floor(Number(coveredMessageCount) || 0)),
+      );
+      const currentRevision = Number.isSafeInteger(revision) ? revision : 0;
+
+      if (
+        messages === cachedHistory
+        && covered === cachedCoveredCount
+        && currentRevision === cachedRevision
+      ) {
+        return cachedEstimate;
+      }
+
+      cachedEstimate = estimateMessages(messages.slice(covered));
+      cachedHistory = messages;
+      cachedCoveredCount = covered;
+      cachedRevision = currentRevision;
+      return cachedEstimate;
+    },
+  };
+}
+
+module.exports = {
+  createPendingTokenEstimateCache,
+};

package/bin/repeated-assistant-response-policy.cjs

@@ -10,6 +10,7 @@ const INTRA_MESSAGE_MIN_BLOCK_CHARS = 40;
 const INTRA_MESSAGE_MIN_SAVED_CHARS = 1_000;
 const INTRA_MESSAGE_SIMILARITY_THRESHOLD = 0.88;
 const INTRA_MESSAGE_REPEAT_PLACEHOLDER = '[Exact repeated assistant paragraphs omitted from model projection; raw history preserved.]';
+const intraMessageProjectionCache = new WeakMap();
 
 function isUserPrompt(message) {
   if (message?.role !== 'user') return false;
@@ -132,8 +133,24 @@ function compactRepeatedAssistantParagraphs(message) {
   )) return null;
 
   const text = textParts[0].text;
+  const cached = intraMessageProjectionCache.get(message);
+  if (
+    cached?.content === message.content
+    && cached.textPart === textParts[0]
+    && cached.text === text
+  ) return cached.projected;
+
+  const cacheProjection = (projected) => {
+    intraMessageProjectionCache.set(message, {
+      content: message.content,
+      projected,
+      text,
+      textPart: textParts[0],
+    });
+    return projected;
+  };
   const blocks = text.split(/\n\s*\n/u).map((block) => block.trim()).filter(Boolean);
-  if (blocks.length < INTRA_MESSAGE_MIN_REPEATS + 1) return null;
+  if (blocks.length < INTRA_MESSAGE_MIN_REPEATS + 1) return cacheProjection(null);
   const protectedIndexes = fencedBlockIndexes(blocks);
 
   const clusters = [];
@@ -150,7 +167,7 @@ function compactRepeatedAssistantParagraphs(message) {
   const repeatedIndexes = new Set(clusters
     .filter((cluster) => cluster.indexes.length >= INTRA_MESSAGE_MIN_REPEATS)
     .flatMap((cluster) => cluster.indexes.slice(1)));
-  if (repeatedIndexes.size === 0) return null;
+  if (repeatedIndexes.size === 0) return cacheProjection(null);
 
   const projectedBlocks = [];
   let markerAdded = false;
@@ -166,13 +183,13 @@ function compactRepeatedAssistantParagraphs(message) {
   }
 
   const projectedText = projectedBlocks.join('\n\n');
-  if (text.length - projectedText.length < INTRA_MESSAGE_MIN_SAVED_CHARS) return null;
-  return {
+  if (text.length - projectedText.length < INTRA_MESSAGE_MIN_SAVED_CHARS) return cacheProjection(null);
+  return cacheProjection({
     ...message,
     content: message.content.map((part) => (
       part === textParts[0] ? { ...part, text: projectedText } : part
     )),
-  };
+  });
 }
 
 function projectRepeatedAssistantResponses(history) {

package/blun.mjs

@@ -78650,12 +78650,13 @@ function formatUndoUnavailableMessage(requestedCount, undoableCount, stoppedAtCo
 		return `${String(count)} ${count === 1 ? "prompt" : "prompts"}`;
 	}
 }
-var TOOL_ERROR_STATUS$2, TOOL_EMPTY_STATUS$2, TOOL_EMPTY_ERROR_STATUS$2, TOOL_OUTPUT_EMPTY_TEXT$2, TOOL_INTERRUPTED_ON_RESUME_OUTPUT$1, ContextMemory;
+var TOOL_ERROR_STATUS$2, TOOL_EMPTY_STATUS$2, TOOL_EMPTY_ERROR_STATUS$2, TOOL_OUTPUT_EMPTY_TEXT$2, TOOL_INTERRUPTED_ON_RESUME_OUTPUT$1, createPendingTokenEstimateCache, ContextMemory;
 var init_context$2 = __esmMin((() => {
-	init_src$4();
-	init_errors$8();
-	init_image_compress();
-	init_tokens();
+		init_src$4();
+		init_errors$8();
+		init_image_compress();
+		init_tokens();
+		({ createPendingTokenEstimateCache } = createRequire(import.meta.url)("./bin/pending-token-estimate-policy.cjs"));
 	init_xml_escape();
 	init_compaction();
 	init_projector();
@@ -78669,8 +78670,10 @@ var init_context$2 = __esmMin((() => {
 	ContextMemory = class {
 		agent;
 		_history = [];
-		_tokenCount = 0;
-		tokenCountCoveredMessageCount = 0;
+			_tokenCount = 0;
+			tokenCountCoveredMessageCount = 0;
+			pendingTokenRevision = 0;
+			pendingTokenEstimateCache = createPendingTokenEstimateCache(estimateTokensForMessages);
 		openSteps = /* @__PURE__ */ new Map();
 		pendingToolResultIds = /* @__PURE__ */ new Set();
 		deferredMessages = [];
@@ -78679,9 +78682,12 @@ var init_context$2 = __esmMin((() => {
 		constructor(agent) {
 			this.agent = agent;
 		}
-		get lastAssistantAt() {
-			return this._lastAssistantAt;
-		}
+			get lastAssistantAt() {
+				return this._lastAssistantAt;
+			}
+			markPendingTokenEstimateDirty() {
+				this.pendingTokenRevision = (this.pendingTokenRevision + 1) % Number.MAX_SAFE_INTEGER;
+			}
 		appendUserMessage(content, origin = USER_PROMPT_ORIGIN) {
 			if (content.length === 0) return;
 			const { captions, parts } = origin.kind === "user" ? splitImageCompressionCaptions(content) : {
@@ -78783,8 +78789,11 @@ var init_context$2 = __esmMin((() => {
 			const last = lastDeferred ?? this._history.at(-1);
 			if (last === void 0) return false;
 			if (!matcher(last.origin)) return false;
-			if (lastDeferred !== void 0) this.deferredMessages.pop();
-			else this._history.pop();
+				if (lastDeferred !== void 0) this.deferredMessages.pop();
+				else {
+					this._history.pop();
+					this.markPendingTokenEstimateDirty();
+				}
 			return true;
 		}
 		/** Remove transient system reminders before the next model projection. */
@@ -78806,7 +78815,8 @@ var init_context$2 = __esmMin((() => {
 				removedMessages.add(message);
 				return false;
 			});
-			if (removedMessages.size === 0) return 0;
+				if (removedMessages.size === 0) return 0;
+				this.markPendingTokenEstimateDirty();
 			this.agent.replayBuilder.removeLastMessages(removedMessages);
 			this.agent.microCompaction.reset(this._history.length);
 			this.agent.toolResultBatchOffload.reset(this._history);
@@ -78817,7 +78827,8 @@ var init_context$2 = __esmMin((() => {
 		}
 		clear() {
 			this.agent.records.logRecord({ type: "context.clear" });
-			this._history = [];
+				this._history = [];
+				this.markPendingTokenEstimateDirty();
 			this._tokenCount = 0;
 			this.tokenCountCoveredMessageCount = 0;
 			this.openSteps.clear();
@@ -78861,7 +78872,8 @@ var init_context$2 = __esmMin((() => {
 					if (removedUserCount >= count) break;
 				}
 			}
-			this.agent.replayBuilder.removeLastMessages(removedMessages);
+				this.agent.replayBuilder.removeLastMessages(removedMessages);
+				if (removedMessages.size > 0) this.markPendingTokenEstimateDirty();
 			this.openSteps.clear();
 			this.pendingToolResultIds.clear();
 			this.deferredMessages = [];
@@ -78940,7 +78952,8 @@ var init_context$2 = __esmMin((() => {
 				origin: { kind: "compaction_summary" }
 			};
 			const isLegacyRestore = this.agent.records.restoring !== null && input.keptUserMessageCount === void 0 && input.compactedCount < this._history.length;
-			this._history = isLegacyRestore ? [summaryMessage, ...this._history.slice(input.compactedCount)] : [...keptMessages, summaryMessage];
+				this._history = isLegacyRestore ? [summaryMessage, ...this._history.slice(input.compactedCount)] : [...keptMessages, summaryMessage];
+				this.markPendingTokenEstimateDirty();
 			this.openSteps.clear();
 			this.pendingToolResultIds.clear();
 			this.deferredMessages = [];
@@ -78963,9 +78976,12 @@ var init_context$2 = __esmMin((() => {
 		get tokenCount() {
 			return this._tokenCount;
 		}
-		get tokenCountWithPending() {
-			const pendingMessages = this._history.slice(this.tokenCountCoveredMessageCount);
-			return this._tokenCount + estimateTokensForMessages(pendingMessages);
+			get tokenCountWithPending() {
+				return this._tokenCount + this.pendingTokenEstimateCache.read(
+					this._history,
+					this.tokenCountCoveredMessageCount,
+					this.pendingTokenRevision
+				);
 		}
 		get history() {
 			return this._history;
@@ -79120,6 +79136,7 @@ var init_context$2 = __esmMin((() => {
 							this._tokenCount += estimateTokensForMessages(this._history.slice(previousCoveredCount, coveredCount));
 						}
 						this.tokenCountCoveredMessageCount = coveredCount;
+						this.markPendingTokenEstimateDirty();
 					}
 					this.flushDeferredMessagesIfToolExchangeClosed();
 					return;
@@ -79127,18 +79144,20 @@ var init_context$2 = __esmMin((() => {
 				case "content.part": {
 					const openStep = this.openSteps.get(event.stepUuid);
 					if (openStep === void 0) throw new Error(`Received content_part for unknown step_uuid '${event.stepUuid}' (no open step_begin)`);
-					openStep.content.push(event.part);
+						openStep.content.push(event.part);
+						this.markPendingTokenEstimateDirty();
 					return;
 				}
 				case "tool.call": {
 					const openStep = this.openSteps.get(event.stepUuid);
 					if (openStep === void 0) throw new Error(`Received tool_call for unknown step_uuid '${event.stepUuid}' (no open step_begin)`);
-					openStep.toolCalls.push({
+						openStep.toolCalls.push({
 						type: "function",
 						id: event.toolCallId,
 						name: event.name,
-						arguments: event.args === void 0 ? null : JSON.stringify(event.args)
-					});
+							arguments: event.args === void 0 ? null : JSON.stringify(event.args)
+						});
+						this.markPendingTokenEstimateDirty();
 					this.pendingToolResultIds.add(event.toolCallId);
 					return;
 				}
@@ -79175,8 +79194,9 @@ var init_context$2 = __esmMin((() => {
 		hasOpenToolExchange() {
 			return this.pendingToolResultIds.size > 0;
 		}
-		pushHistory(...messages) {
-			this._history.push(...messages);
+			pushHistory(...messages) {
+				this._history.push(...messages);
+				this.markPendingTokenEstimateDirty();
 			for (const message of messages) {
 				if (message.role === "assistant") this._lastAssistantAt = this.agent.records.restoring?.time ?? Date.now();
 				if (message.origin?.kind === "background_task") this.agent.background.markDeliveredNotification(message.origin);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "blun-king-cli",
-  "version": "9.1.219",
+  "version": "9.1.221",
   "description": "BLUN CLI - your own AI agent with a Telegram channel. Get it done. With BLUN.",
   "license": "MIT",
   "bin": {