blun-king-cli 9.1.187 → 9.1.189

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

@@ -4,6 +4,7 @@ const MIN_RESPONSE_CHARS = 160;
 const MAX_RECENT_ASSISTANT_RESPONSES = 8;
 const REQUIRED_MATCHING_RESPONSES = 3;
 const RESPONSE_SIMILARITY_THRESHOLD = 0.78;
+const REPEATED_RESPONSE_PLACEHOLDER = '[Historical repeated assistant response omitted; newer user corrections and instructions supersede it.]';
 
 function isUserPrompt(message) {
   if (message?.role !== 'user') return false;
@@ -94,16 +95,33 @@ function detectRepeatedAssistantResponse(history) {
 
   return {
     count: matches.length,
+    indexes: matches.map((match) => match.index),
     minimumSimilarity: Math.min(...matches.slice(1).map((match) => match.similarity)),
   };
 }
 
+function projectRepeatedAssistantResponses(history) {
+  const detection = detectRepeatedAssistantResponse(history);
+  if (detection === null) return history;
+
+  const repeatedIndexes = new Set(detection.indexes);
+  return history.map((message, index) => {
+    if (!repeatedIndexes.has(index)) return message;
+    return {
+      ...message,
+      content: [{ type: 'text', text: REPEATED_RESPONSE_PLACEHOLDER }],
+    };
+  });
+}
+
 module.exports = {
   MAX_RECENT_ASSISTANT_RESPONSES,
   MIN_RESPONSE_CHARS,
   REQUIRED_MATCHING_RESPONSES,
+  REPEATED_RESPONSE_PLACEHOLDER,
   RESPONSE_SIMILARITY_THRESHOLD,
   assistantResponseText,
   detectRepeatedAssistantResponse,
+  projectRepeatedAssistantResponses,
   responseSimilarity,
 };

package/bin/turn-thinking-policy.cjs

@@ -3,10 +3,12 @@
 const EXPLICIT_THINKING_INTENT = /\b(?:analy[sz](?:e|ing|is)|root\s+cause|investigat(?:e|ion)|debug(?:ging)?|architect(?:ure|ural)|plan(?:ning)?|review|audit|threat\s+model|refactor|optim(?:ize|ise|ization)|migration|complex|thorough|deep(?:ly)?|untersuch(?:e|ung)|analysier(?:e|en)?|ursachenanalyse|architektur|plan(?:e|ung)|pruef(?:e|ung)|sicherheit|migration|komplex|gruendlich|analiz(?:a|iraj)|istrazi|arhitektur|planiraj|sigurnost|duboko)\b/iu;
 const ACTIONABLE_PROMPT = /\b(?:aender|analys|apply|bau|build|check|commit|cop(?:y|ier)|create|debug|delete|deploy|edit|entfern|erstelle|find|fix|implement|install|lad|loesch|merge|mess|move|napravi|oeffne|open|patch|pruef|publish|push|read|reparier|restart|run|schreib|search|starte?|stop|such|test|update|verschieb|write|zieh)\w*\b/iu;
 const STANDALONE_CHAT = /^(?:hi|hallo|hey|hello|hoi|moin|servus|yo|danke|thanks|thx|bye|tschau|ciao)[.!?]*$/iu;
-const IDENTITY_CHAT = /^(?:wer bist du|who are you|wie geht(?: es dir)?|how are you|na)[.!?]*$/iu;
+const BOSNIAN_STANDALONE_CHAT = /^(?:zdravo|pozdrav|hvala|ćao|cao)[.!?]*$/iu;
+const IDENTITY_CHAT = /^(?:wer bist du|who are you|wie geht(?: es dir)?|how are you|na|kako si|kako je king|ko si ti|ko stoji iza tebe)[.!?]*$/iu;
 const STATUS_ONLY_CHECK_IN = /(?:\bstandabfrage\b|\bstatusabfrage\b|\bkurzes?\s+update\b|\bwo\s+stehst\s+du\b|\bwie\s+weit\s+bist\s+du\b|\bwhere\s+are\s+you\b|\bquick\s+status\b)/iu;
 const STATUS_FOLLOW_UP_ACTION = /\b(?:aendere|arbeite|baue|build|change|delete|deploy|deploye|fahr|fix|fixe|implement|implementiere|install|installiere|loesche|mach|patch|pruefe|repariere|run|schreibe|setze|start|starte|stop|stoppe|test|teste|write)\b/iu;
 const LIGHTWEIGHT_STEER = /\b(?:cool|danke|dankeschoen|gute arbeit|passt|perfekt|stolz|super|freut mich)\b/iu;
+const STANDALONE_CHAT_COMPLETION_TOKENS = 1024;
 const ADAPTIVE_LOW_COMPLETION_TOKENS = 8192;
 const STANDARD_TURN_MAX_COMPLETION_TOKENS = 24_576;
 
@@ -17,7 +19,8 @@ function selectThinkingEffortForTurn(text, originKind) {
   if (!normalized) return "low";
 
   const lower = normalized.toLowerCase();
-  if (STANDALONE_CHAT.test(lower) || IDENTITY_CHAT.test(lower)) return "low";
+  if (IDENTITY_CHAT.test(lower) || BOSNIAN_STANDALONE_CHAT.test(lower)) return "off";
+  if (STANDALONE_CHAT.test(lower)) return "low";
   if (
     STATUS_ONLY_CHECK_IN.test(normalized)
     && !STATUS_FOLLOW_UP_ACTION.test(normalized)
@@ -42,13 +45,16 @@ function selectThinkingEffortForWorkStep(stepNumber, previousToolOutcome) {
 }
 
 function capCompletionBudgetForAdaptiveEffort(budget, thinkingEffort) {
-  if (thinkingEffort !== "low" || budget === undefined) return budget;
+  if ((thinkingEffort !== "off" && thinkingEffort !== "low") || budget === undefined) return budget;
   const configuredCap = budget.hardCap ?? budget.fallback;
+  const adaptiveCap = thinkingEffort === "off"
+    ? STANDALONE_CHAT_COMPLETION_TOKENS
+    : ADAPTIVE_LOW_COMPLETION_TOKENS;
   return {
     ...budget,
     hardCap: Math.min(
-      configuredCap ?? ADAPTIVE_LOW_COMPLETION_TOKENS,
-      ADAPTIVE_LOW_COMPLETION_TOKENS,
+      configuredCap ?? adaptiveCap,
+      adaptiveCap,
     ),
   };
 }
@@ -81,6 +87,7 @@ module.exports = {
   ACTIONABLE_PROMPT,
   ADAPTIVE_LOW_COMPLETION_TOKENS,
   EXPLICIT_THINKING_INTENT,
+  STANDALONE_CHAT_COMPLETION_TOKENS,
   STANDARD_TURN_MAX_COMPLETION_TOKENS,
   capCompletionBudgetForAdaptiveEffort,
   capTurnCompletionTokens,

package/blun.mjs

@@ -28236,9 +28236,10 @@ function personaSystemBlock(env = process.env) {
 		}
 	}
 	if (language !== void 0 && language.length > 0) {
+		const requestedLanguage = language === "Bosnian" ? "Bosnian (bosanski; not Slovenian or slovenski)" : language;
 		if (lines.length > 0) lines.push("");
-		lines.push(`## Language — HARD OVERRIDE (${language})`, "");
-		lines.push(`Write EVERYTHING you output in ${language}: your reasoning and thinking, your progress notes, and every reply to the user. This rule is ABSOLUTE and OVERRIDES the "# Language" section elsewhere in this prompt and any instinct to mirror the user's language. Even when the user writes to you in English or any other language, you still think and answer in ${language}. The only things that stay verbatim are code, shell commands, file paths, and identifiers. Do not switch away from ${language} unless the user explicitly changes the language with a command.`);
+		lines.push(`## Language — HARD OVERRIDE (${requestedLanguage})`, "");
+		lines.push(`Write EVERYTHING you output in ${requestedLanguage}: your reasoning and thinking, your progress notes, and every reply to the user. This rule is ABSOLUTE and OVERRIDES the "# Language" section elsewhere in this prompt and any instinct to mirror the user's language. Even when the user writes to you in English or any other language, you still think and answer in ${requestedLanguage}. The only things that stay verbatim are code, shell commands, file paths, and identifiers. Do not switch away from ${requestedLanguage} unless the user explicitly changes the language with a command.`);
 	}
 	return lines.join("\n");
 }
@@ -75985,12 +75986,13 @@ var init_full = __esmMin((() => {
 }));
 //#endregion
 //#region ../../packages/agent-core/src/agent/compaction/micro.ts
-var selectMicroCompactionCutoff, MICRO_COMPACTION_RECENT_MESSAGES, MICRO_COMPACTION_TOOL_ARGUMENTS_ENABLED, MICRO_COMPACTION_TOOL_ARGUMENT_MIN_TOKENS, isPersistedToolResultReference, projectHistoricalUnaddressedTelegramMessages, dedupeRepeatedUserMessages, compactHistoricalSkillActivations, dedupeRecurringCronWakeups, dedupeRepeatedInjections, DEFAULT_CONFIG, MicroCompaction;
+var selectMicroCompactionCutoff, MICRO_COMPACTION_RECENT_MESSAGES, MICRO_COMPACTION_TOOL_ARGUMENTS_ENABLED, MICRO_COMPACTION_TOOL_ARGUMENT_MIN_TOKENS, isPersistedToolResultReference, projectHistoricalUnaddressedTelegramMessages, dedupeRepeatedUserMessages, projectRepeatedAssistantResponses, compactHistoricalSkillActivations, dedupeRecurringCronWakeups, dedupeRepeatedInjections, DEFAULT_CONFIG, MicroCompaction;
 var init_micro = __esmMin((() => {
 	({ selectMicroCompactionCutoff, MICRO_COMPACTION_RECENT_MESSAGES, MICRO_COMPACTION_TOOL_ARGUMENTS_ENABLED, MICRO_COMPACTION_TOOL_ARGUMENT_MIN_TOKENS } = createRequire(import.meta.url)("./bin/micro-compaction-policy.cjs"));
 	({ isPersistedToolResultReference } = createRequire(import.meta.url)("./bin/tool-result-offload-policy.cjs"));
 	({ projectHistoricalUnaddressedTelegramMessages } = createRequire(import.meta.url)("./bin/telegram-context-projection-policy.cjs"));
 	({ dedupeRepeatedUserMessages } = createRequire(import.meta.url)("./bin/repeated-user-message-projection.cjs"));
+	({ projectRepeatedAssistantResponses } = createRequire(import.meta.url)("./bin/repeated-assistant-response-policy.cjs"));
 	({ compactHistoricalSkillActivations } = createRequire(import.meta.url)("./bin/skill-activation-performance-policy.cjs"));
 	({ dedupeRecurringCronWakeups, dedupeRepeatedInjections } = createRequire(import.meta.url)("./bin/repeated-injection-projection.cjs"));
 	init_tokens();
@@ -78970,7 +78972,8 @@ var init_context$2 = __esmMin((() => {
 			const historicalUserReferencesProjected = compactHistoricalPersistedUserMessageReferences(userOffloaded);
 			const duplicateUserMessagesProjected = dedupeRepeatedUserMessages(historicalUserReferencesProjected);
 			const historicalTelegramProjected = projectHistoricalUnaddressedTelegramMessages(duplicateUserMessagesProjected);
-			const result = project(this.agent.assistantMessageOffload.compact(this.agent.microCompaction.compact(compactHistoricalSuccessfulToolResults(this.agent.toolResultBatchOffload.compact(dedupeRecurringCronWakeups(dedupeRepeatedInjections(compactHistoricalSkillActivations(compactBaselineSkillInjections(historicalTelegramProjected)))))))), {
+			const repeatedAssistantProjected = projectRepeatedAssistantResponses(historicalTelegramProjected);
+			const result = project(this.agent.assistantMessageOffload.compact(this.agent.microCompaction.compact(compactHistoricalSuccessfulToolResults(this.agent.toolResultBatchOffload.compact(dedupeRecurringCronWakeups(dedupeRepeatedInjections(compactHistoricalSkillActivations(compactBaselineSkillInjections(repeatedAssistantProjected)))))))), {
 				...options,
 				onAnomaly: (anomaly) => {
 					anomalies.push(anomaly);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "blun-king-cli",
-  "version": "9.1.187",
+  "version": "9.1.189",
   "description": "BLUN CLI - your own AI agent with a Telegram channel. Get it done. With BLUN.",
   "license": "MIT",
   "bin": {