blun-king-cli 9.1.429 → 9.1.430

package/bin/compaction-model-policy.cjs

@@ -0,0 +1,12 @@
+'use strict';
+
+function resolveCompactionModelAlias(configuredAlias, activeAlias) {
+  if (typeof configuredAlias === 'string' && configuredAlias.trim().length > 0) {
+    return configuredAlias.trim();
+  }
+  return activeAlias;
+}
+
+module.exports = {
+  resolveCompactionModelAlias,
+};

package/blun.mjs

@@ -9,6 +9,7 @@ import telegramConsoleStatusPolicy from "./bin/telegram-console-status-policy.cj
 import cognitiveGoalAutostartPolicy from "./bin/cognitive-goal-autostart-policy.cjs";
 import cognitiveGoalTimeTriggerController from "./bin/cognitive-goal-time-trigger-controller.cjs";
 import runtimeExitLedger from "./bin/runtime-exit-ledger.cjs";
+import compactionModelPolicy from "./bin/compaction-model-policy.cjs";
 import crypto$1, { createHash, randomBytes, randomInt, randomUUID, timingSafeEqual } from "node:crypto";
 import * as fs$16 from "node:fs";
 import Kt, { accessSync, appendFileSync, chmodSync, closeSync, constants, copyFileSync, createReadStream, createWriteStream, existsSync, fsyncSync, linkSync, lstatSync, mkdirSync, openSync, promises, readFileSync, readSync, readdirSync, realpathSync, renameSync, rmSync, statSync, unlinkSync, writeFileSync, writeSync } from "node:fs";
@@ -42,6 +43,7 @@ const { writeTelegramConsoleStatus } = telegramConsoleStatusPolicy;
 const { goalAutostartDecision, goalContinuationDecision } = cognitiveGoalAutostartPolicy;
 const { GoalTimeTriggerController } = cognitiveGoalTimeTriggerController;
 const { recordRuntimeExit } = runtimeExitLedger;
+const { resolveCompactionModelAlias } = compactionModelPolicy;
 import { EventEmitter as EventEmitter$1 } from "node:events";
 import { StringDecoder } from "node:string_decoder";
 import co from "node:assert";
@@ -10048,7 +10050,8 @@ var init_schema = __esmMin((() => {
 		maxRetriesPerStep: number$1().int().min(0).optional(),
 		maxRalphIterations: number$1().int().min(-1).optional(),
 		reservedContextSize: number$1().int().min(0).optional(),
-		compactionTriggerRatio: number$1().min(.5).max(.99).optional()
+		compactionTriggerRatio: number$1().min(.5).max(.99).optional(),
+		compactionModel: string().min(1).optional()
 	});
 	BackgroundConfigSchema = object({
 		maxRunningTasks: number$1().int().min(1).optional(),
@@ -75620,11 +75623,11 @@ var init_full = __esmMin((() => {
 			let retryCount = 0;
 			try {
 				await this.triggerPreCompactHook(data, tokensBefore, signal);
-				const model = this.agent.activeResponderModel ?? this.agent.config.model;
+				const model = resolveCompactionModelAlias(this.agent.blunConfig?.loopControl?.compactionModel, this.agent.activeResponderModel ?? this.agent.config.model);
 				const runtimeModel = this.agent.config.runtimeModel(model);
 				const capability = runtimeModel.modelCapabilities;
 				const buildCompactionProvider = (usedContextTokens) => {
-					const currentMaxContextTokens = this.getEffectiveMaxContextTokens();
+					const currentMaxContextTokens = this.getEffectiveMaxContextTokens(model);
 					const defaultCompactionCap = capCompactionCompletionTokens(currentMaxContextTokens);
 					const requestCapability = currentMaxContextTokens > 0 ? {
 						...capability,
@@ -75666,7 +75669,7 @@ var init_full = __esmMin((() => {
 					}), createUserMessage(requestInstruction)], capability));
 				};
 				while (true) {
-					const compactionRequestLimit = this.getEffectiveMaxContextTokens();
+					const compactionRequestLimit = this.getEffectiveMaxContextTokens(model);
 					const safeCompactionRequestLimit = compactionRequestLimit > 0 ? Math.max(1, Math.floor(compactionRequestLimit * (1 - COMPACTION_SUMMARY_RESERVE_RATIO))) : compactionRequestLimit;
 					const targetCompactionRequestLimit = capCompactionStageTarget(safeCompactionRequestLimit);
 					let messages = buildRequestMessages(historyForModel, instruction);
@@ -75764,7 +75767,7 @@ var init_full = __esmMin((() => {
 					};
 					armStallTimer();
 					try {
-						const response = await this.agent.generate(provider, COMPACTION_SYSTEM_PROMPT, compactionTools, messages, { onMessagePart: observeProgress }, { signal: attemptController.signal });
+						const response = await this.agent.generateForModel(model)(provider, COMPACTION_SYSTEM_PROMPT, compactionTools, messages, { onMessagePart: observeProgress }, { signal: attemptController.signal });
 						if (stalled && !signal.aborted && stallPolicy !== void 0) throw new CompactionStallError(stallPolicy.timeoutMs, stallPolicy.measuredIdleMs);
 						maxObservedIdleMs = Math.max(maxObservedIdleMs, Date.now() - lastProgressAt);
 						if (response.finishReason === "truncated") throw new CompactionTruncatedError();
@@ -75803,10 +75806,10 @@ var init_full = __esmMin((() => {
 						if (stalled && !signal.aborted && stallPolicy !== void 0) throw new CompactionStallError(stallPolicy.timeoutMs, stallPolicy.measuredIdleMs);
 						if (isModelFallbackStatus(error)) throw error;
 						const isContextOverflow = this.shouldRecoverFromContextOverflow(error);
-						if (isContextOverflow) this.observeContextOverflow(error);
+						if (isContextOverflow) this.observeContextOverflow(error, model);
 						if (isContextOverflow && historyForModel.length > 1) {
 							if (data.source === "auto") {
-								const learnedMaxContextTokens = this.getEffectiveMaxContextTokens();
+								const learnedMaxContextTokens = this.getEffectiveMaxContextTokens(model);
 								if (learnedMaxContextTokens > 0 && learnedMaxContextTokens < compactionRequestLimit) {
 									retryCount = 0;
 									continue;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "blun-king-cli",
-  "version": "9.1.429",
+  "version": "9.1.430",
   "description": "BLUN CLI - your own AI agent with a Telegram channel. Get it done. With BLUN.",
   "license": "MIT",
   "bin": {