blun-king-cli 9.1.110 → 9.1.112

package/bin/user-message-offload-policy.cjs

@@ -1,6 +1,8 @@
 'use strict';
 
 const USER_MESSAGE_MAX_CHARS = 200_000;
+const HISTORICAL_USER_MESSAGE_MAX_CHARS = 4_000;
+const USER_MESSAGE_KEEP_RECENT_MESSAGES = 20;
 const USER_MESSAGE_PREVIEW_LINES = 5;
 const USER_MESSAGE_PREVIEW_LINE_CHARS = 1_000;
 const USER_MESSAGE_OFFLOAD_MARKER = '[User message offloaded]';
@@ -9,6 +11,22 @@ function shouldOffloadUserMessage(textChars) {
   return Number.isSafeInteger(textChars) && textChars > USER_MESSAGE_MAX_CHARS;
 }
 
+function shouldOffloadHistoricalUserMessage(options = {}) {
+  const historyIndex = Number(options.historyIndex);
+  const historyLength = Number(options.historyLength);
+  const textChars = Number(options.textChars);
+  if (
+    !Number.isSafeInteger(historyIndex)
+    || !Number.isSafeInteger(historyLength)
+    || !Number.isSafeInteger(textChars)
+    || historyIndex < 0
+    || historyLength < 0
+    || historyIndex >= historyLength
+  ) return false;
+  const cutoff = Math.max(0, historyLength - USER_MESSAGE_KEEP_RECENT_MESSAGES);
+  return historyIndex < cutoff && textChars > HISTORICAL_USER_MESSAGE_MAX_CHARS;
+}
+
 function createUserMessagePreview(text) {
   const lines = String(text).split(/\r?\n/u);
   const render = (line, index) => `${index}: ${line.slice(0, USER_MESSAGE_PREVIEW_LINE_CHARS)}`;
@@ -28,10 +46,13 @@ function createUserMessagePreview(text) {
 }
 
 module.exports = {
+  HISTORICAL_USER_MESSAGE_MAX_CHARS,
   USER_MESSAGE_MAX_CHARS,
+  USER_MESSAGE_KEEP_RECENT_MESSAGES,
   USER_MESSAGE_OFFLOAD_MARKER,
   USER_MESSAGE_PREVIEW_LINES,
   USER_MESSAGE_PREVIEW_LINE_CHARS,
   createUserMessagePreview,
+  shouldOffloadHistoricalUserMessage,
   shouldOffloadUserMessage,
 };

package/blun.mjs

@@ -260422,7 +260422,7 @@ async function saveUserMessage(homedir, text) {
 function renderPersistedUserMessage(text, outputPath) {
 	return [
 		USER_MESSAGE_OFFLOAD_MARKER,
-		`Message text exceeded ${String(USER_MESSAGE_MAX_CHARS)} characters; the complete original remains in conversation history.`,
+		"The complete original was archived to keep the active context fast and remains available on disk.",
 		`message_size_chars: ${String(text.length)}`,
 		`message_size_bytes: ${String(Buffer.byteLength(text, "utf8"))}`,
 		`output_path: ${outputPath}`,
@@ -260432,12 +260432,13 @@ function renderPersistedUserMessage(text, outputPath) {
 		createUserMessagePreview(text)
 	].join("\n");
 }
-var USER_MESSAGE_MAX_CHARS, USER_MESSAGE_OFFLOAD_MARKER, shouldOffloadUserMessage, createUserMessagePreview;
+var USER_MESSAGE_MAX_CHARS, USER_MESSAGE_OFFLOAD_MARKER, shouldOffloadUserMessage, shouldOffloadHistoricalUserMessage, createUserMessagePreview;
 var init_user_message_offload = __esmMin((() => {
 	const policy = createRequire(import.meta.url)("./bin/user-message-offload-policy.cjs");
 	USER_MESSAGE_MAX_CHARS = policy.USER_MESSAGE_MAX_CHARS;
 	USER_MESSAGE_OFFLOAD_MARKER = policy.USER_MESSAGE_OFFLOAD_MARKER;
 	shouldOffloadUserMessage = policy.shouldOffloadUserMessage;
+	shouldOffloadHistoricalUserMessage = policy.shouldOffloadHistoricalUserMessage;
 	createUserMessagePreview = policy.createUserMessagePreview;
 }));
 var UserMessageOffload = class {
@@ -260447,30 +260448,44 @@ var UserMessageOffload = class {
 		this.agent = agent;
 	}
 	async detect() {
-		const message = this.agent.context.history.at(-1);
-		const text = persistableUserMessageText(message);
-		if (text === void 0 || !shouldOffloadUserMessage(text.length)) return 0;
-		const messageHash = userMessageOffloadHash(message, text);
-		if (this.replacements.has(messageHash) || this.agent.homedir === void 0) return 0;
-		const outputPath = await saveUserMessage(this.agent.homedir, text);
-		if (outputPath === void 0) return 0;
-		const replacementText = renderPersistedUserMessage(text, outputPath);
-		const content = [{
-			type: "text",
-			text: replacementText
-		}, ...message.content.filter((part) => part.type !== "text")];
-		this.apply({
-			messageHash,
-			content,
-			outputPath
-		});
-		this.agent.telemetry.track("user_message_offloaded", {
-			chars_before: text.length,
-			chars_after: replacementText.length,
-			chars_saved: text.length - replacementText.length,
-			media_part_count: content.length - 1
-		});
-		return 1;
+		const history = this.agent.context.history;
+		if (this.agent.homedir === void 0) return 0;
+		let offloaded = 0;
+		for (let historyIndex = 0; historyIndex < history.length; historyIndex++) {
+			const message = history[historyIndex];
+			const text = persistableUserMessageText(message);
+			if (text === void 0) continue;
+			const immediateOversize = historyIndex === history.length - 1 && shouldOffloadUserMessage(text.length);
+			const historicalOversize = shouldOffloadHistoricalUserMessage({
+				historyIndex,
+				historyLength: history.length,
+				textChars: text.length
+			});
+			if (!immediateOversize && !historicalOversize) continue;
+			const messageHash = userMessageOffloadHash(message, text);
+			if (this.replacements.has(messageHash)) continue;
+			const outputPath = await saveUserMessage(this.agent.homedir, text);
+			if (outputPath === void 0) continue;
+			const replacementText = renderPersistedUserMessage(text, outputPath);
+			const content = [{
+				type: "text",
+				text: replacementText
+			}, ...message.content.filter((part) => part.type !== "text")];
+			this.apply({
+				messageHash,
+				content,
+				outputPath
+			});
+			this.agent.telemetry.track("user_message_offloaded", {
+				chars_before: text.length,
+				chars_after: replacementText.length,
+				chars_saved: text.length - replacementText.length,
+				media_part_count: content.length - 1,
+				reason: immediateOversize ? "immediate_oversize" : "historical_oversize"
+			});
+			offloaded++;
+		}
+		return offloaded;
 	}
 	apply(replacement) {
 		this.replacements.set(replacement.messageHash, replacement);
@@ -260584,12 +260599,12 @@ function blunFastConversationTools(tools, input, pendingMediaJobIds = []) {
 * name through ToolSearch. The absolute cap prevents large context windows
 * from turning every deferred schema into permanent request overhead.
 */
-function blunSelectTurnTools(tools, maxContextTokens, loadedToolNames) {
+function blunSelectTurnTools(tools, maxContextTokens, loadedToolNames, requiredToolNames = /* @__PURE__ */ new Set()) {
 	if (estimateTokensForTools(tools) <= toolSchemaBudgetTokens(maxContextTokens)) return {
 		tools,
 		deferredToolCount: 0
 	};
-	const selected = tools.filter((tool) => BLUN_LEAN_TOOL_NAMES.has(tool.name) || loadedToolNames.has(tool.name));
+	const selected = tools.filter((tool) => BLUN_LEAN_TOOL_NAMES.has(tool.name) || loadedToolNames.has(tool.name) || requiredToolNames.has(tool.name));
 	const selectedNames = new Set(selected.map((tool) => tool.name));
 	const deferred = tools.filter((tool) => !selectedNames.has(tool.name));
 	if (deferred.length > 0) selected.push(createDeferredToolLoader(selected, deferred, loadedToolNames));
@@ -261430,7 +261445,7 @@ var init_turn = __esmMin((() => {
 					const toolSelection = fastConversation ? {
 						tools: [...blunFastConversationTools(eligibleTools, input, pendingMediaJobIds)],
 						deferredToolCount: 0
-					} : turnNeedsTools ? blunSelectTurnTools(eligibleTools, this.agent.config.modelCapabilities?.max_context_tokens, this.loadedToolNames) : {
+					} : turnNeedsTools ? blunSelectTurnTools(eligibleTools, this.agent.config.modelCapabilities?.max_context_tokens, this.loadedToolNames, turnHasAttachment ? new Set(["mcp__plugin-telegram_telegram__download_attachment"]) : void 0) : {
 						tools: [],
 						deferredToolCount: 0
 					};

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "blun-king-cli",
-  "version": "9.1.110",
+  "version": "9.1.112",
   "description": "BLUN CLI - your own AI agent with a Telegram channel. Get it done. With BLUN.",
   "license": "MIT",
   "bin": {