blun-king-cli 9.1.610 → 9.1.612

package/worker-host.mjs

@@ -1,11 +1,11 @@
 #!/usr/bin/env node
-// BLUN_BUILD_INPUT_SHA256:8c6e16ea727bc28d31e8a784ec7ab7cc656e35f70f996bb1e64df97776875360
+// BLUN_BUILD_INPUT_SHA256:f9f2527561c5e5ec1afc53eef27313667c4d323a444532439e73a66865340f69
 import { fileURLToPath as __cjsShimFileURLToPath } from 'node:url';
 import { dirname as __cjsShimDirname } from 'node:path';
 const __filename = __cjsShimFileURLToPath(import.meta.url);
 const __dirname = __cjsShimDirname(__filename);
 import { createRequire } from "node:module";
-import crypto$1, { createHash, randomBytes, randomInt, randomUUID } from "node:crypto";
+import crypto$1, { createHash, createHmac, randomBytes, randomInt, randomUUID } from "node:crypto";
 import * as fs$12 from "node:fs";
 import Kt, { appendFileSync, chmodSync, closeSync, constants, createReadStream, createWriteStream, existsSync, fstatSync, fsyncSync, linkSync, lstatSync, mkdirSync, openSync, opendirSync, promises, readFileSync, readSync, readdirSync, realpathSync, renameSync, statSync, unlinkSync, writeFileSync, writeSync } from "node:fs";
 import * as path$11 from "node:path";
@@ -1888,6 +1888,9 @@ var FetchTransport = class {
 		const maxRetries = normalizeMaxRetries(requestOptions?.maxRetries);
 		const requestBytes = exactBodyBytes(init.body);
 		const requestBreakdown = pathname === "/chat/completions" && requestOptions?.onTransportAttempt !== void 0 && typeof init.body === "string" && requestBytes !== void 0 ? measureSerializedRequest(init.body, requestBytes) : void 0;
+		if (pathname === "/chat/completions" && typeof init.body === "string") try {
+			requestOptions?.onSerializedRequest?.(init.body);
+		} catch {}
 		let attempt = 0;
 		while (true) {
 			const headers = new Headers(init.headers);
@@ -1896,6 +1899,9 @@ var FetchTransport = class {
 			const startedAt = performance.now();
 			const startedAtMs = Date.now();
 			try {
+				try {
+					requestOptions?.onRequestSent?.();
+				} catch {}
 				const response = await this.fetch(`${this.baseUrl}${pathname}`, {
 					...init,
 					headers
@@ -2648,6 +2654,13 @@ function collectBlunVisionAttachments(messages) {
 	}
 	return attachments;
 }
+function estimateBlunRequestExtraBytes(tools, messages) {
+	const attachments = collectBlunVisionAttachments(messages);
+	return {
+		toolsBytes: tools.length === 0 ? 0 : Buffer.byteLength(JSON.stringify({ tools: tools.map(convertTool) }), "utf8"),
+		visionBytes: attachments.length === 0 ? 0 : Buffer.byteLength(JSON.stringify({ visionAttachments: attachments }), "utf8")
+	};
+}
 function convertMessage(message) {
 	let reasoningContent = "";
 	const nonThinkParts = [];
@@ -2957,12 +2970,13 @@ var BlunChatProvider = class {
 		if (this._stream) createParams["stream_options"] = { include_usage: true };
 		try {
 			const client = this._createClient(options?.auth);
-			const requestOptions = options?.signal !== void 0 || options?.transportMaxRetries !== void 0 || options?.onTransportAttempt !== void 0 ? {
+			const requestOptions = options?.signal !== void 0 || options?.transportMaxRetries !== void 0 || options?.onRequestSent !== void 0 || options?.onSerializedRequest !== void 0 || options?.onTransportAttempt !== void 0 ? {
 				...options.signal !== void 0 ? { signal: options.signal } : {},
 				...options.transportMaxRetries !== void 0 ? { maxRetries: options.transportMaxRetries } : {},
+				...options.onRequestSent !== void 0 ? { onRequestSent: options.onRequestSent } : {},
+				...options.onSerializedRequest !== void 0 ? { onSerializedRequest: options.onSerializedRequest } : {},
 				...options.onTransportAttempt !== void 0 ? { onTransportAttempt: options.onTransportAttempt } : {}
 			} : void 0;
-			options?.onRequestSent?.();
 			let response;
 			try {
 				response = await client.chat.completions.create(createParams, requestOptions);
@@ -2971,7 +2985,6 @@ var BlunChatProvider = class {
 				const cap = rejection instanceof APIContextOverflowError ? contextOverflowOutputCap(rejection, createParams["max_completion_tokens"]) : void 0;
 				if (cap === void 0) throw rejection;
 				options?.signal?.throwIfAborted();
-				options?.onRequestSent?.();
 				response = await client.chat.completions.create({
 					...createParams,
 					max_completion_tokens: cap
@@ -21757,10 +21770,15 @@ async function loadAgentsMdForRoots(kaos, brandHome, workDirs) {
 	const discovered = [];
 	const emptySources = /* @__PURE__ */ new Map();
 	const seen = /* @__PURE__ */ new Set();
+	const notLoadedCandidates = /* @__PURE__ */ new Set();
 	const collect = async (path) => {
 		const file = await readAgentFile(kaos, path);
-		if (file === void 0) return false;
+		if (file === void 0) {
+			notLoadedCandidates.add(kaos.normpath(path));
+			return false;
+		}
 		const key = kaos.normpath(file.path);
+		notLoadedCandidates.delete(key);
 		if (file.content.length === 0) {
 			emptySources.set(key, InstructionText.file(file.path, "", file.sourceTextBytes));
 			return false;
@@ -21796,7 +21814,10 @@ async function loadAgentsMdForRoots(kaos, brandHome, workDirs) {
 	return {
 		content,
 		warning: totalBytes > AGENTS_MD_RECOMMENDED_MAX_BYTES ? `AGENTS.md total ${formatKB(totalBytes)} KB exceeds the recommended ${formatKB(AGENTS_MD_RECOMMENDED_MAX_BYTES)} KB. Large instruction files increase cost and may impact performance; consider trimming.` : void 0,
-		sources: rendered.measure()
+		sources: {
+			...rendered.measure(),
+			...notLoadedCandidates.size === 0 ? {} : { notLoadedCandidates: [...notLoadedCandidates] }
+		}
 	};
 }
 async function loadAdditionalDirsInfo(kaos, additionalDirs) {
@@ -28778,17 +28799,20 @@ function createSystemPromptRenderer(merged) {
 				const counts = renderPromptContributions(merged.systemPromptTemplate, vars, rendered, true) ?? renderPromptContributions(merged.systemPromptTemplate, vars, rendered);
 				const sources = context.agentsMdSources;
 				const copies = counts?.instructionCopies ?? 0;
-				const sourceMatches = sources !== void 0 && sources.version === 1 && Number.isSafeInteger(sources.generatedBytes) && sources.generatedBytes >= 0 && sources.files.every((file) => typeof file.path === "string" && Number.isSafeInteger(file.bytes) && file.bytes >= 0) && sources.totalBytes === Buffer.byteLength(vars["BLUN_AGENTS_MD"] ?? "", "utf8") && sources.generatedBytes + sources.files.reduce((sum, file) => sum + file.bytes, 0) === sources.totalBytes;
+				const sourceMatches = sources !== void 0 && sources.version === 1 && Number.isSafeInteger(sources.generatedBytes) && sources.generatedBytes >= 0 && sources.files.every((file) => typeof file.path === "string" && Number.isSafeInteger(file.bytes) && file.bytes >= 0) && (sources.notLoadedCandidates?.every((path) => typeof path === "string") ?? true) && sources.totalBytes === Buffer.byteLength(vars["BLUN_AGENTS_MD"] ?? "", "utf8") && sources.generatedBytes + sources.files.reduce((sum, file) => sum + file.bytes, 0) === sources.totalBytes;
+				const notLoadedCandidates = sourceMatches && sources.notLoadedCandidates !== void 0 ? { notLoadedCandidates: [...sources.notLoadedCandidates] } : {};
 				const instructionSources = copies === 0 ? {
 					version: 1,
 					totalBytes: 0,
 					generatedBytes: 0,
-					files: sourceMatches ? sources.files.map((file) => copyInstructionFileBytes(file, 0)) : []
+					files: sourceMatches ? sources.files.map((file) => copyInstructionFileBytes(file, 0)) : [],
+					...notLoadedCandidates
 				} : !sourceMatches ? void 0 : {
 					version: 1,
 					totalBytes: sources.totalBytes * copies,
 					generatedBytes: sources.generatedBytes * copies,
-					files: sources.files.map((file) => copyInstructionFileBytes(file, copies))
+					files: sources.files.map((file) => copyInstructionFileBytes(file, copies)),
+					...notLoadedCandidates
 				};
 				const skillListingSources = counts === void 0 ? void 0 : scaleSkillListingSources(skillSources, vars["BLUN_SKILLS"] ?? "", counts.skillListingBytes);
 				context.onContributions(counts === void 0 ? void 0 : {
@@ -230115,6 +230139,7 @@ function buildCompactionSummaryText(summary) {
 const MAX_SESSION_COMPACTION_FAILURES = 2;
 const DEFAULT_COMPACTION_MAX_COMPLETION_TOKENS = 128 * 1024;
 const COMPACTION_SUMMARY_RESERVE_RATIO = .1;
+const BLUN_COMPACTION_REQUEST_BUDGET_BYTES = 75e4;
 const MAX_HIERARCHICAL_COMPACTION_PASSES = 64;
 const HIERARCHICAL_COMPACTION_PREFIX = "This is a complete summary of an earlier chronological segment. Preserve it as source material when merging it with the following conversation:";
 var CompactionTruncatedError = class extends Error {
@@ -230234,16 +230259,19 @@ var FullCompaction = class {
 		const projected = context.additionalMessages === void 0 ? messages : [...messages, ...context.additionalMessages];
 		const historyBudget = this.requestHistoryByteBudget(projected);
 		const historyThresholdReached = historyBudget !== void 0 && historyBudget.usedHistoryBytes >= historyBudget.thresholdHistoryBytes;
+		const requestBudget = this.requestBodyByteBudget(projected, context);
+		const requestThresholdReached = requestBudget !== void 0 && requestBudget.usedRequestBytesEstimate >= requestBudget.thresholdRequestBytes;
 		if (this.sessionCompactionFailures >= MAX_SESSION_COMPACTION_FAILURES) reason = "session_stopped";
 		else if (this.isCompacting) reason = "already_compacting";
-		else if (thresholdTokens === void 0 && historyBudget === void 0) reason = "no_budget";
-		else if (!historyThresholdReached && (thresholdTokens === void 0 || !Number.isFinite(usedTokens) || usedTokens < thresholdTokens)) reason = "below_threshold";
+		else if (thresholdTokens === void 0 && historyBudget === void 0 && requestBudget === void 0) reason = "no_budget";
+		else if (!historyThresholdReached && !requestThresholdReached && (thresholdTokens === void 0 || !Number.isFinite(usedTokens) || usedTokens < thresholdTokens)) reason = "below_threshold";
 		else if (!this.hasShrinkableHistory(projected)) reason = "nothing_to_compact";
 		return {
 			eligible: reason === "eligible",
 			usedTokens,
 			thresholdTokens,
 			...historyBudget,
+			...requestBudget,
 			reason
 		};
 	}
@@ -230265,6 +230293,35 @@ var FullCompaction = class {
 			thresholdHistoryBytes: limit
 		};
 	}
+	requestBodyByteBudget(messages, context) {
+		const limit = this.agent.blunConfig?.loopControl?.maxRequestHistoryBytes;
+		if (!this.agent.config.hasProvider || this.strategy.compactionBudgetTokens === void 0 || limit === void 0 || !Number.isSafeInteger(limit) || limit <= 0) return void 0;
+		const model = this.agent.activeResponderModel ?? this.agent.config.modelAlias;
+		if (model === void 0 || this.agent.config.runtimeModel(model).provider.name !== "blun") return void 0;
+		const systemPrompt = context.systemPrompt ?? this.agent.effectiveSystemPrompt;
+		const tools = (context.tools ?? this.agent.tools.loopTools).map((tool) => ({
+			name: tool.name,
+			description: tool.description,
+			parameters: tool.parameters
+		}));
+		let extraBytes;
+		try {
+			extraBytes = estimateBlunRequestExtraBytes(tools, messages);
+		} catch {
+			return;
+		}
+		if (Buffer.byteLength(JSON.stringify({
+			systemPrompt,
+			messages: []
+		}), "utf8") + extraBytes.toolsBytes >= BLUN_COMPACTION_REQUEST_BUDGET_BYTES) return void 0;
+		return {
+			usedRequestBytesEstimate: Buffer.byteLength(JSON.stringify({
+				systemPrompt,
+				messages
+			}), "utf8") + extraBytes.toolsBytes + extraBytes.visionBytes,
+			thresholdRequestBytes: BLUN_COMPACTION_REQUEST_BUDGET_BYTES
+		};
+	}
 	requestProactiveCompaction(context = {}) {
 		const status = this.proactiveCompactionStatus(context);
 		if (status.eligible && !this.proactiveCompactionRequested) {
@@ -230273,7 +230330,9 @@ var FullCompaction = class {
 				used_tokens: status.usedTokens,
 				threshold_tokens: status.thresholdTokens,
 				used_history_bytes: status.usedHistoryBytes,
-				threshold_history_bytes: status.thresholdHistoryBytes
+				threshold_history_bytes: status.thresholdHistoryBytes,
+				used_request_bytes_estimate: status.usedRequestBytesEstimate,
+				threshold_request_bytes: status.thresholdRequestBytes
 			});
 		}
 		return status;
@@ -230402,7 +230461,9 @@ var FullCompaction = class {
 		const requestTokens = this.estimateCurrentRequestTokens(requestContext);
 		const budget = this.getCompactionBudgetTokens();
 		const historyBudget = this.requestHistoryByteBudget(projectedMessages);
-		const proactive = (historyBudget !== void 0 && historyBudget.usedHistoryBytes >= historyBudget.thresholdHistoryBytes || !this.isCompacting && budget !== void 0 && Number.isFinite(requestTokens) && requestTokens >= Math.max(1, Math.floor(budget * .9))) && this.hasShrinkableHistory(projectedMessages);
+		const historyThresholdReached = historyBudget !== void 0 && historyBudget.usedHistoryBytes >= historyBudget.thresholdHistoryBytes;
+		const requestBudget = this.requestBodyByteBudget(projectedMessages, requestContext);
+		const proactive = (requestBudget !== void 0 && requestBudget.usedRequestBytesEstimate >= requestBudget.thresholdRequestBytes || historyThresholdReached || !this.isCompacting && budget !== void 0 && Number.isFinite(requestTokens) && requestTokens >= Math.max(1, Math.floor(budget * .9))) && this.hasShrinkableHistory(projectedMessages);
 		if (this.checkAutoCompaction(true, requestTokens, projectedMessages, proactive) && (proactive || this.strategy.shouldBlock(requestTokens))) await this.block(signal);
 		signal.throwIfAborted();
 		const compacted = this.agent.context.history !== historyBefore;
@@ -230604,6 +230665,7 @@ var FullCompaction = class {
 			let charsReceived = 0;
 			let lastProgressEmitAt = 0;
 			let initialCompactionRequestTokens;
+			let initialCompactionRequestBytes;
 			const compactionTools = [];
 			const buildRequestMessages = (history, requestInstruction) => {
 				return stripMediaForCompaction(downgradeUnsupportedMedia([...this.agent.context.project(history, {
@@ -230611,6 +230673,7 @@ var FullCompaction = class {
 					dropOrphanResults: true
 				}), createUserMessage(requestInstruction)], capability));
 			};
+			const estimateRequestBytes = (messages) => Buffer.byteLength(JSON.stringify(messages), "utf8") + Buffer.byteLength(this.agent.config.systemPrompt, "utf8");
 			while (true) {
 				signal.throwIfAborted();
 				if (hierarchicalPassCount >= MAX_HIERARCHICAL_COMPACTION_PASSES) throw new BlunError(ErrorCodes.COMPACTION_FAILED, `Compaction stopped after ${String(MAX_HIERARCHICAL_COMPACTION_PASSES)} staged requests without reaching a final summary. The conversation history was not changed.`, { details: {
@@ -230620,30 +230683,39 @@ var FullCompaction = class {
 				const compactionRequestLimit = this.getEffectiveMaxContextTokens(model);
 				const safeCompactionRequestLimit = compactionRequestLimit > 0 ? Math.max(1, Math.floor(compactionRequestLimit * (1 - COMPACTION_SUMMARY_RESERVE_RATIO))) : compactionRequestLimit;
 				let messages = buildRequestMessages(historyForModel, instruction);
+				const requestByteBudget = provider.name === "blun" ? BLUN_COMPACTION_REQUEST_BUDGET_BYTES : void 0;
 				let estimatedCompactionRequestTokens = this.estimateRequestTokens(messages, this.agent.effectiveSystemPrompt, compactionTools);
+				let estimatedCompactionRequestBytes = requestByteBudget === void 0 ? void 0 : estimateRequestBytes(messages);
 				initialCompactionRequestTokens ??= estimatedCompactionRequestTokens;
+				initialCompactionRequestBytes ??= estimatedCompactionRequestBytes;
 				let hierarchicalChunkEnd;
-				if (safeCompactionRequestLimit > 0 && estimatedCompactionRequestTokens >= safeCompactionRequestLimit) {
+				const tokenBudgetExceeded = safeCompactionRequestLimit > 0 && estimatedCompactionRequestTokens >= safeCompactionRequestLimit;
+				const byteBudgetExceeded = requestByteBudget !== void 0 && estimatedCompactionRequestBytes !== void 0 && estimatedCompactionRequestBytes >= requestByteBudget;
+				if (tokenBudgetExceeded || byteBudgetExceeded) {
 					if (historyForModel.length > 1) {
 						const chunkInstruction = `${instruction}\n\nThis request contains only the oldest chronological segment. Summarize every fact needed by a later merge pass. Do not assume later messages are visible.`;
-						const chunk = selectHierarchicalCompactionChunk(historyForModel, safeCompactionRequestLimit, safeCompactionRequestLimit, (candidate) => {
+						const chunk = selectHierarchicalCompactionChunk(historyForModel, safeCompactionRequestLimit, safeCompactionRequestLimit, requestByteBudget, (candidate) => {
 							const candidateMessages = buildRequestMessages(candidate, chunkInstruction);
 							return {
 								messages: candidateMessages,
-								estimatedTokens: this.estimateRequestTokens(candidateMessages, this.agent.effectiveSystemPrompt, compactionTools)
+								estimatedTokens: this.estimateRequestTokens(candidateMessages, this.agent.effectiveSystemPrompt, compactionTools),
+								estimatedBytes: requestByteBudget === void 0 ? void 0 : estimateRequestBytes(candidateMessages)
 							};
 						});
 						if (chunk !== void 0) {
 							hierarchicalChunkEnd = chunk.end;
 							messages = chunk.messages;
 							estimatedCompactionRequestTokens = chunk.estimatedTokens;
+							estimatedCompactionRequestBytes = chunk.estimatedBytes;
 						}
 					}
-					if (hierarchicalChunkEnd === void 0) throw new BlunError(ErrorCodes.CONTEXT_OVERFLOW, `Compaction stopped before upload: the conversation (${String(tokensBefore)} tokens) could not be divided into a request below the active model window (${String(compactionRequestLimit)} tokens).`, { details: {
+					if (hierarchicalChunkEnd === void 0) throw new BlunError(ErrorCodes.CONTEXT_OVERFLOW, byteBudgetExceeded ? "Compaction stopped before upload: the conversation could not be divided into a request below the active byte limit." : `Compaction stopped before upload: the conversation (${String(tokensBefore)} tokens) could not be divided into a request below the active model window (${String(compactionRequestLimit)} tokens).`, { details: {
 						contextBrakeBlocked: true,
 						estimatedRequestTokens: estimatedCompactionRequestTokens,
 						maxContextTokens: compactionRequestLimit,
 						requestBudgetTokens: safeCompactionRequestLimit,
+						estimatedRequestBytes: estimatedCompactionRequestBytes,
+						requestBudgetBytes: requestByteBudget,
 						summaryReserveTokens: compactionRequestLimit - safeCompactionRequestLimit,
 						contextUnchanged: true
 					} });
@@ -230652,7 +230724,7 @@ var FullCompaction = class {
 				attemptCount += 1;
 				charsReceived = 0;
 				const stage = hierarchicalPassCount + 1;
-				const estimatedStageCount = estimateCompactionStageCount(initialCompactionRequestTokens, safeCompactionRequestLimit, stage);
+				const estimatedStageCount = estimateCompactionStageCount(initialCompactionRequestTokens, safeCompactionRequestLimit, stage, initialCompactionRequestBytes, requestByteBudget);
 				const estimatedProgressPercent = estimateCompactionProgressPercent(stage, estimatedStageCount);
 				const windowUsagePercent = estimateCompactionWindowUsagePercent(estimatedCompactionRequestTokens, compactionRequestLimit);
 				this.agent.log.info("compaction stage request", {
@@ -230661,6 +230733,8 @@ var FullCompaction = class {
 					estimatedStageCount,
 					activeContextTokens: this.estimateProjectedRequestTokens(),
 					estimatedInputTokens: estimatedCompactionRequestTokens,
+					estimatedInputBytes: estimatedCompactionRequestBytes,
+					requestBudgetBytes: requestByteBudget,
 					windowUsagePercent,
 					safeInputLimitTokens: safeCompactionRequestLimit,
 					maxContextTokens: compactionRequestLimit
@@ -230732,13 +230806,17 @@ var FullCompaction = class {
 					if (hierarchicalChunkEnd !== void 0) {
 						hierarchicalPassCount += 1;
 						const previousTokens = estimateTokensForMessages(historyForModel);
+						const previousBytes = requestByteBudget === void 0 ? void 0 : Buffer.byteLength(JSON.stringify(historyForModel), "utf8");
 						const nextHistory = [createUserMessage(`${HIERARCHICAL_COMPACTION_PREFIX}\n${extractedSummary}`), ...historyForModel.slice(hierarchicalChunkEnd)];
 						const nextTokens = estimateTokensForMessages(nextHistory);
-						if (nextTokens >= previousTokens) throw new BlunError(ErrorCodes.COMPACTION_FAILED, "A staged compaction response did not reduce the pending history. The conversation history was not changed.", { details: {
+						const nextBytes = requestByteBudget === void 0 ? void 0 : Buffer.byteLength(JSON.stringify(nextHistory), "utf8");
+						if (nextTokens >= previousTokens || previousBytes !== void 0 && nextBytes !== void 0 && nextBytes >= previousBytes) throw new BlunError(ErrorCodes.COMPACTION_FAILED, "A staged compaction response did not reduce the pending history. The conversation history was not changed.", { details: {
 							contextUnchanged: true,
 							hierarchicalPassCount,
 							previousTokens,
-							nextTokens
+							nextTokens,
+							previousBytes,
+							nextBytes
 						} });
 						historyForModel = nextHistory;
 						retryCount = 0;
@@ -230934,9 +231012,10 @@ function isModelFallbackStatus(error) {
 function compactionTimingKey(provider) {
 	return `${provider.name}\u0000${provider.modelName}`;
 }
-function estimateCompactionStageCount(initialInputTokens, safeRequestLimitTokens, currentStage) {
-	if (safeRequestLimitTokens <= 0) return currentStage;
-	return Math.max(currentStage, Math.ceil(initialInputTokens / safeRequestLimitTokens));
+function estimateCompactionStageCount(initialInputTokens, safeRequestLimitTokens, currentStage, initialInputBytes, safeRequestLimitBytes) {
+	const tokenStages = safeRequestLimitTokens > 0 ? Math.ceil(initialInputTokens / safeRequestLimitTokens) : 0;
+	const byteStages = initialInputBytes !== void 0 && safeRequestLimitBytes !== void 0 && safeRequestLimitBytes > 0 ? Math.ceil(initialInputBytes / safeRequestLimitBytes) : 0;
+	return Math.max(currentStage, tokenStages, byteStages);
 }
 function estimateCompactionProgressPercent(stage, estimatedStageCount) {
 	if (estimatedStageCount <= 0) return 0;
@@ -230952,7 +231031,7 @@ const COMPACTION_OVERFLOW_SHRINK_RATIOS = [
 	.5,
 	.35
 ];
-function selectHierarchicalCompactionChunk(history, targetRequestTokens, hardRequestLimit, build) {
+function selectHierarchicalCompactionChunk(history, targetRequestTokens, hardRequestLimit, requestByteBudget, build) {
 	const safeEnds = [];
 	for (let end = 1; end < history.length; end++) if (history[end]?.role !== "tool") safeEnds.push(end);
 	if (safeEnds.length === 0) return void 0;
@@ -230964,7 +231043,7 @@ function selectHierarchicalCompactionChunk(history, targetRequestTokens, hardReq
 			const middle = Math.floor((low + high) / 2);
 			const end = safeEnds[middle];
 			const request = build(history.slice(0, end));
-			if (request.estimatedTokens < limit) {
+			if (request.estimatedTokens < limit && (requestByteBudget === void 0 || request.estimatedBytes !== void 0 && request.estimatedBytes < requestByteBudget)) {
 				best = {
 					end,
 					...request
@@ -236771,8 +236850,14 @@ var PersonalMemoryRecallInjector = class {
 	* to read, write, list, or change any personal-memory state.
 	*/
 	filterToolsForTurn(turnId, input, origin, tools) {
-		if (this.toolsAllowedForTurn(turnId, input, origin)) return tools.filter((tool) => !isPersonalMemoryRecallToolName(tool.name) && (!isPersonalMemoryToolName(tool.name) || this.isTrustedHostTool(tool.name)));
-		return tools.filter((tool) => !isPersonalMemoryToolName(tool.name));
+		const brokerAllowed = this.toolsAllowedForTurn(turnId, input, origin);
+		const localSearchAllowed = origin.kind === "user" && this.agent.type === "main" && !isTelegramGroupInput(input) && this.agent.localWorkHistorySearchEligibility?.() === true;
+		return tools.filter((tool) => {
+			if (!isPersonalMemoryToolName(tool.name)) return true;
+			if (isPersonalMemoryRecallToolName(tool.name)) return false;
+			if (tool.name === "mcp__personal-memory__memory_search" && localSearchAllowed) return true;
+			return brokerAllowed && this.isTrustedHostTool(tool.name);
+		});
 	}
 	isTrustedHostTool(name) {
 		const isHostTool = this.agent.isPersonalMemoryHostTool;
@@ -238951,6 +239036,32 @@ var import_turn_tool_performance_policy = (/* @__PURE__ */ __commonJSMin(((expor
 		"suchen"
 	]);
 	const TOOL_SEARCH_RELATED_TERM_GROUPS = Object.freeze([
+		Object.freeze([
+			"persist",
+			"save",
+			"speichere",
+			"speichern",
+			"store"
+		]),
+		Object.freeze([
+			"erinnerung",
+			"erinnerungen",
+			"fact",
+			"facts",
+			"fakt",
+			"fakten",
+			"memory",
+			"memories",
+			"merk",
+			"merken",
+			"remember"
+		]),
+		Object.freeze([
+			"note",
+			"notes",
+			"notiz",
+			"notizen"
+		]),
 		Object.freeze([
 			"checklist",
 			"list",
@@ -239209,7 +239320,7 @@ var import_turn_tool_performance_policy = (/* @__PURE__ */ __commonJSMin(((expor
 				if (document.nameTokens.has(token)) exactNameMatches += 1;
 				else if (relatedSearchTerms(token).some((term) => document.nameTokens.has(term))) relatedNameMatches += 1;
 			}
-			return matchedTokens >= 2 || exactNameMatches >= 1 && score >= 10 || relatedNameMatches >= 1 && score >= 5 ? {
+			return matchedTokens >= 2 || exactNameMatches >= 1 && score >= 10 || queryTokens.length === 1 && relatedNameMatches >= 1 && score >= 5 ? {
 				name: document.name,
 				score,
 				matchedTokens,
@@ -239230,11 +239341,11 @@ var import_turn_tool_performance_policy = (/* @__PURE__ */ __commonJSMin(((expor
 	function searchRelatedDeferredTools(tools, query, options = {}) {
 		const normalized = normalizeDeferredToolQuery(query).toLowerCase();
 		if (!normalized || normalized.startsWith("select:")) return [];
-		const tokens = normalized.split(/\s+/).filter(Boolean);
+		const tokens = normalized.match(/\+?[\p{L}\p{N}_-]+/gu) || [];
 		const requestedNameTokens = tokens.filter((token) => token.startsWith("+")).map((token) => token.slice(1)).filter(Boolean);
 		const relaxRequiredNameTokens = options.relaxRequiredNameTokens === true && requestedNameTokens.length > 0;
 		const requiredNameTokens = relaxRequiredNameTokens ? [] : requestedNameTokens;
-		const searchTokens = [...tokens.filter((token) => !token.startsWith("+")), ...relaxRequiredNameTokens ? requestedNameTokens : []].filter((token) => !TOOL_SEARCH_INTENT_WORDS.has(token));
+		const searchTokens = [...tokens.filter((token) => !token.startsWith("+")), ...relaxRequiredNameTokens ? requestedNameTokens : []].filter((token) => !TOOL_SEARCH_INTENT_WORDS.has(token) && !TOOL_RANK_STOP_WORDS.has(token));
 		if (searchTokens.length < 2 && !TOOL_SEARCH_SINGLE_TERM_RELATED_FALLBACKS.has(searchTokens[0])) return [];
 		const minimumMatchedTokens = searchTokens.length === 1 ? 1 : Math.max(2, Math.ceil(searchTokens.length * .75));
 		const ranked = tools.map((tool) => {
@@ -239295,9 +239406,12 @@ var import_turn_tool_performance_policy = (/* @__PURE__ */ __commonJSMin(((expor
 			const name = String(tool?.name || "").trim();
 			if (name && !available.has(name)) available.set(name, tool);
 		}
+		const usedNames = new Set([...selectedTools, ...available.values()].map((tool) => String(tool?.name ?? "")));
+		let loaderName = DEFERRED_TOOL_LOADER_NAME;
+		for (let suffix = 1; usedNames.has(loaderName); suffix += 1) loaderName = `${DEFERRED_TOOL_LOADER_NAME}_${suffix}`;
 		const deferredCount = available.size;
 		const loader = {
-			name: DEFERRED_TOOL_LOADER_NAME,
+			name: loaderName,
 			description: [
 				`Search ${deferredCount} deferred ${deferredCount === 1 ? "tool schema" : "tool schemas"} without loading all definitions.`,
 				"Use select:leaf_name for one known tool, plain keywords to search, or +word to require that word in the tool name.",
@@ -239317,7 +239431,7 @@ var import_turn_tool_performance_policy = (/* @__PURE__ */ __commonJSMin(((expor
 				const query = normalizeDeferredToolQuery(args?.query);
 				return {
 					description: query ? `Searching tool schemas for ${query}` : "Searching tool schemas",
-					approvalRule: DEFERRED_TOOL_LOADER_NAME,
+					approvalRule: loaderName,
 					execute: async () => {
 						let matches = searchDeferredTools([...available.values()], query);
 						let relatedFallback = false;
@@ -239330,7 +239444,7 @@ var import_turn_tool_performance_policy = (/* @__PURE__ */ __commonJSMin(((expor
 							relatedFallback = matches.length > 0;
 						}
 						if (matches.length === 0) {
-							const alreadyLoaded = query.toLowerCase().startsWith("select:") ? searchDeferredTools(selectedTools.filter((tool) => tool?.name !== DEFERRED_TOOL_LOADER_NAME), query) : [];
+							const alreadyLoaded = query.toLowerCase().startsWith("select:") ? searchDeferredTools(selectedTools.filter((tool) => tool?.name !== loaderName), query) : [];
 							if (alreadyLoaded.length === 1) return {
 								isError: false,
 								output: `Tool schema already loaded: ${alreadyLoaded[0].name}. Invoke it now.`
@@ -239340,7 +239454,7 @@ var import_turn_tool_performance_policy = (/* @__PURE__ */ __commonJSMin(((expor
 								output: [
 									`No eligible deferred tool is available for: ${query || "[empty query]"}.`,
 									"Use a loaded alternative or report this integration as unavailable.",
-									"ToolSearch remains available for a different tool needed later."
+									`${loaderName} remains available for a different tool needed later.`
 								].join(" ")
 							};
 						}
@@ -250032,6 +250146,9 @@ function logStepTiming(log, turnId, step, response) {
 		streamDurationMs: timing.streamDurationMs,
 		...timing.serverDecodeMs !== void 0 ? { serverDecodeMs: timing.serverDecodeMs } : {},
 		...timing.clientConsumeMs !== void 0 ? { clientConsumeMs: timing.clientConsumeMs } : {},
+		inputOtherTokens: response.usage.inputOther,
+		cacheReadTokens: response.usage.inputCacheRead,
+		cacheCreationTokens: response.usage.inputCacheCreation,
 		outputTokens: response.usage.output
 	});
 }
@@ -266492,7 +266609,7 @@ var HistoricalToolProjection = class {
 		for (const { index, message, outputPath, text, reference, shellOutput } of candidates) try {
 			context.signal.throwIfAborted();
 			if (!await this.canRead(read, outputPath, context)) continue;
-			const save = this.pending.then(() => this.persist(outputPath, text, context.signal));
+			const save = this.pending.then(() => this.persist(outputPath, text, context.signal, true));
 			this.pending = save.catch(() => void 0);
 			await save;
 			const stat = await this.agent.kaos.stat(outputPath);
@@ -266584,7 +266701,7 @@ var HistoricalToolProjection = class {
 			revision: revision(after)
 		});
 	}
-	async persist(path, text, signal) {
+	async persist(path, text, signal, finalVerificationFollows = false) {
 		signal.throwIfAborted();
 		await checkedAncestors(dirname(dirname(dirname(path))));
 		signal.throwIfAborted();
@@ -266593,7 +266710,7 @@ var HistoricalToolProjection = class {
 		await this.ensureDirectory(dirname(path));
 		signal.throwIfAborted();
 		if (this.files.has(path)) {
-			await this.verifyFile(path, text);
+			if (!finalVerificationFollows) await this.verifyFile(path, text);
 			return;
 		}
 		try {
@@ -267062,7 +267179,7 @@ function blunToolsForOrigin(tools, origin) {
 * the ordinary permission pipeline.
 */
 function blunSelectTurnTools(tools, maxContextTokens, loadedToolNames = /* @__PURE__ */ new Set(), requiredToolNames = /* @__PURE__ */ new Set(), selected = []) {
-	if (estimateTokensForTools(tools) <= (0, import_turn_tool_performance_policy.toolSchemaBudgetTokens)(maxContextTokens) || tools.some((tool) => tool.name === "ToolSearch")) {
+	if (estimateTokensForTools(tools) <= (0, import_turn_tool_performance_policy.toolSchemaBudgetTokens)(maxContextTokens) || tools.some(import_turn_tool_performance_policy.isDeferredToolLoader)) {
 		selected.splice(0, selected.length, ...tools);
 		return selected;
 	}
@@ -267559,6 +267676,7 @@ var TurnFlow = class {
 		const appendedMessage = this.agent.context.history.at(-1);
 		const turnAnchor = appendedMessage !== previousLastMessage && appendedMessage?.role === "user" && appendedMessage.origin === origin ? appendedMessage : void 0;
 		const startedAt = Date.now();
+		const preRequestHookStartedAt = performance.now();
 		let ended;
 		let blockedByUserPromptHook = false;
 		let completedStopReason;
@@ -267584,6 +267702,13 @@ var TurnFlow = class {
 					this.agent.log.warn("Turn-start observation failed");
 				}
 				signal.throwIfAborted();
+				if (this.agent.type === "main" && process.env["BLUN_DIAG_PRE_REQUEST_PHASES"] === "1") try {
+					this.agent.log.info("turn pre-request phase", {
+						turnId,
+						phase: "turn-hooks",
+						durationMs: Math.max(0, Math.round(performance.now() - preRequestHookStartedAt))
+					});
+				} catch {}
 				const stopReason = await this.runStepLoop(turnId, signal, input, origin, turnAnchor, (response) => {
 					completedResponse = response;
 				});
@@ -267753,9 +267878,26 @@ var TurnFlow = class {
 		let stopHookContinuationUsed = false;
 		let goalOutcomeMessageContinuationUsed = false;
 		const deduper = new ToolCallDeduplicator({ telemetry: this.agent.telemetry });
+		const diagnosePreRequest = this.agent.type === "main" && process.env["BLUN_DIAG_PRE_REQUEST_PHASES"] === "1";
+		const recordPreRequestPhase = (phase, startedAt) => {
+			if (!diagnosePreRequest) return;
+			try {
+				this.agent.log.info("turn pre-request phase", {
+					turnId,
+					phase,
+					durationMs: Math.max(0, Math.round(performance.now() - startedAt))
+				});
+			} catch {}
+		};
+		let phaseStartedAt = performance.now();
 		await this.agent.mcp?.waitForInitialLoad(signal);
+		recordPreRequestPhase("mcp-initial-load", phaseStartedAt);
+		phaseStartedAt = performance.now();
 		const personalMemoryRecall = await this.agent.injection.injectPersonalMemoryForTurn(turnId, input, origin, signal);
+		recordPreRequestPhase("personal-memory-recall", phaseStartedAt);
+		phaseStartedAt = performance.now();
 		await this.agent.injection.injectGoal();
+		recordPreRequestPhase("goal-injection", phaseStartedAt);
 		this.setActiveSteerAcceptance(turnId, true);
 		const turnNeedsThinking = blunTurnNeedsTools(blunThinkingIntentInput(input), origin);
 		let toolInput = input;
@@ -267766,12 +267908,14 @@ var TurnFlow = class {
 			const raw = [...this.agent.context.history];
 			const history = personalMemoryRecall === void 0 ? raw : [...raw, personalMemoryRecall];
 			const selected = taskToolsNeeded ? history : greetingHistory(history, greetingAnchor);
+			const projectionStartedAt = performance.now();
 			const projected = await this.historicalToolProjection.project(selected, turnAnchor, tools, {
 				turnId: String(turnId),
 				stepNumber: this.currentStep,
 				signal,
 				llm: turnLLM
 			});
+			recordPreRequestPhase("historical-tool-projection", projectionStartedAt);
 			const current = this.agent.context.history;
 			if (current.length !== raw.length || current.some((message, index) => message !== raw[index])) {
 				const latest = personalMemoryRecall === void 0 ? current : [...current, personalMemoryRecall];
@@ -267782,7 +267926,7 @@ var TurnFlow = class {
 			partialContext = selected.length < history.length || projected.changed;
 			return projected.history;
 		};
-		const turnLLM = this.agent.llmForTurn(turnNeedsThinking ? void 0 : "off");
+		const turnLLM = this.agent.llmForTurn(turnNeedsThinking ? void 0 : "off", turnId);
 		const turnTools = this.agent.tools.captureTurnTools();
 		while (true) {
 			signal.throwIfAborted();
@@ -267792,6 +267936,7 @@ var TurnFlow = class {
 			try {
 				const selectedTools = [];
 				this.refreshSelectedTools = (snapshot) => {
+					const selectionStartedAt = performance.now();
 					if (snapshot.messages.length > 0) toolInput = snapshot.messages.flatMap((message) => message.content);
 					const eligibleTools = blunToolsForOrigin(this.agent.injection.filterPersonalMemoryToolsForTurn(turnId, input, origin, turnTools()), origin);
 					const intent = blunThinkingIntentInput(toolInput);
@@ -267804,6 +267949,7 @@ var TurnFlow = class {
 							if (tool !== void 0 && !selectedTools.some((candidate) => candidate.name === name) && estimateTokensForTools([...selectedTools, tool]) <= (0, import_turn_tool_performance_policy.toolSchemaBudgetTokens)(this.agent.config.modelCapabilities?.max_context_tokens)) selectedTools.push(tool);
 						}
 					} else selectedTools.splice(0, selectedTools.length, ...eligibleTools.filter((tool) => BLUN_LANGUAGE_REVIEW_TOOL_NAMES.has(tool.name) || BLUN_TELEGRAM_OUTBOUND_TOOL_RE.test(tool.name)));
+					recordPreRequestPhase("tool-selection", selectionStartedAt);
 					return selectedTools;
 				};
 				return (await runTurn({
@@ -267816,14 +267962,24 @@ var TurnFlow = class {
 						}, completion);
 					},
 					llm: turnLLM,
-					buildMessages: async () => this.agent.context.project(await turnHistory(selectedTools), { dropOrphanResults: true }),
-					buildMessagesStrict: async () => this.agent.context.project(await turnHistory(selectedTools), {
-						synthesizeMissing: true,
-						dropOrphanResults: true,
-						dedupeDuplicateToolCalls: true,
-						dropLeadingNonUser: true,
-						mergeConsecutiveAssistants: true
-					}),
+					buildMessages: async () => {
+						const startedAt = performance.now();
+						const messages = this.agent.context.project(await turnHistory(selectedTools), { dropOrphanResults: true });
+						recordPreRequestPhase("message-build", startedAt);
+						return messages;
+					},
+					buildMessagesStrict: async () => {
+						const startedAt = performance.now();
+						const messages = this.agent.context.project(await turnHistory(selectedTools), {
+							synthesizeMissing: true,
+							dropOrphanResults: true,
+							dedupeDuplicateToolCalls: true,
+							dropLeadingNonUser: true,
+							mergeConsecutiveAssistants: true
+						});
+						recordPreRequestPhase("message-build-strict", startedAt);
+						return messages;
+					},
 					dispatchEvent: this.buildDispatchEvent(turnId, signal, onFinalResponse, () => partialContext),
 					tools: selectedTools,
 					log: this.agent.log,
@@ -267840,8 +267996,10 @@ var TurnFlow = class {
 					},
 					hooks: {
 						beforeStep: async ({ signal: stepSignal, llm }) => {
+							const injectionStartedAt = performance.now();
 							this.agent.microCompaction.detect();
 							await this.agent.injection.inject();
+							recordPreRequestPhase("step-injection", injectionStartedAt);
 							stepSignal.throwIfAborted();
 							deduper.beginStep();
 						},
@@ -271507,8 +271665,13 @@ var LlmRequestLogger = class {
 	toolsEstimatedTokens = 0;
 	requestSequence = 0;
 	lastTurnRequest;
-	constructor(log) {
+	serializedPrefixKey;
+	requestSignatures = /* @__PURE__ */ new Map();
+	prefixComparisons = /* @__PURE__ */ new Map();
+	lastSerializedRequest;
+	constructor(log, serializedPrefixDiagnostic = process.env["BLUN_DIAG_SERIALIZED_PREFIX"] === "1") {
 		this.log = log;
+		this.serializedPrefixKey = serializedPrefixDiagnostic ? randomBytes(32) : void 0;
 	}
 	logRequest(input) {
 		const sequence = ++this.requestSequence;
@@ -271525,6 +271688,10 @@ var LlmRequestLogger = class {
 			toolsHash: fingerprint(JSON.stringify(toolSignature(tools)))
 		};
 		const signature = JSON.stringify(config);
+		if (this.serializedPrefixKey !== void 0) {
+			this.requestSignatures.set(sequence, signature);
+			if (this.requestSignatures.size > 32) this.requestSignatures.delete(this.requestSignatures.keys().next().value);
+		}
 		if (signature !== this.lastConfigLogSignature) {
 			this.lastConfigLogSignature = signature;
 			this.configRevision++;
@@ -271554,6 +271721,20 @@ var LlmRequestLogger = class {
 			}
 		};
 		if (partialMessageCount > 0) requestFields.partialMessageCount = partialMessageCount;
+		const source = input.systemPromptContributions;
+		if (source?.version === 1 && [
+			source.totalBytes,
+			source.instructionBytes,
+			source.skillListingBytes,
+			source.runtimeMemoryBytes,
+			source.otherBytes
+		].every((value) => Number.isSafeInteger(value) && value >= 0) && source.instructionBytes + source.skillListingBytes + source.runtimeMemoryBytes + source.otherBytes === source.totalBytes && source.totalBytes === Buffer.byteLength(systemPrompt, "utf8")) requestFields.systemPromptSourceBytes = {
+			totalBytes: source.totalBytes,
+			instructionBytes: source.instructionBytes,
+			skillListingBytes: source.skillListingBytes,
+			runtimeMemoryBytes: source.runtimeMemoryBytes,
+			otherBytes: source.otherBytes
+		};
 		this.log.info("llm request", requestFields);
 		if (fields?.turnStep !== void 0 && /^\d+\.\d+$/.test(fields.turnStep)) {
 			const historyTextOrigins = summarizeRequestTextOrigins(messages);
@@ -271579,10 +271760,51 @@ var LlmRequestLogger = class {
 	getLastTurnRequest() {
 		return this.lastTurnRequest === void 0 ? void 0 : structuredClone(this.lastTurnRequest);
 	}
+	get serializedPrefixDiagnosticEnabled() {
+		return this.serializedPrefixKey !== void 0;
+	}
+	observeSerializedRequest(body, fields, sequence) {
+		const key = this.serializedPrefixKey;
+		if (key === void 0 || !/^\d+\.\d+$/.test(fields?.turnStep ?? "")) return;
+		const signature = this.requestSignatures.get(sequence);
+		if (signature === void 0) return;
+		try {
+			const bytes = Buffer.from(body, "utf8");
+			const granularityBytes = 1024;
+			const chunks = [];
+			for (let start = 0; start + granularityBytes <= bytes.length; start += granularityBytes) chunks.push(createHmac("sha256", key).update(bytes.subarray(start, start + granularityBytes)).digest("hex"));
+			const previous = this.lastSerializedRequest;
+			if (previous !== void 0 && previous.signature === signature) {
+				let common = 0;
+				while (common < previous.chunks.length && common < chunks.length && previous.chunks[common] === chunks[common]) common++;
+				this.prefixComparisons.set(sequence, {
+					version: 1,
+					granularityBytes,
+					previousBytes: previous.bytes,
+					currentBytes: bytes.length,
+					commonPrefixBytesLowerBound: common * granularityBytes,
+					previousSequence: previous.sequence,
+					currentSequence: sequence
+				});
+				if (this.prefixComparisons.size > 32) this.prefixComparisons.delete(this.prefixComparisons.keys().next().value);
+			}
+			this.lastSerializedRequest = {
+				sequence,
+				signature,
+				bytes: bytes.length,
+				chunks
+			};
+		} catch {
+			this.lastSerializedRequest = void 0;
+		}
+	}
 	logTransportAttempt(stats, fields, sequence) {
+		const serializedPrefix = sequence === void 0 ? void 0 : this.prefixComparisons.get(sequence);
 		this.log.info("llm transport attempt", {
 			...fields,
-			...stats
+			...stats,
+			...this.serializedPrefixKey === void 0 || sequence === void 0 ? {} : { requestSequence: sequence },
+			...serializedPrefix === void 0 ? {} : { serializedPrefix }
 		});
 		if (sequence === void 0 || this.lastTurnRequest?.sequence !== sequence) return;
 		const breakdown = stats.requestBreakdown;
@@ -271665,7 +271887,8 @@ function copyContributions(value) {
 			version: source.version,
 			totalBytes: source.totalBytes,
 			generatedBytes: source.generatedBytes,
-			files: source.files.map((file) => copyInstructionFileBytes(file))
+			files: source.files.map((file) => copyInstructionFileBytes(file)),
+			...source.notLoadedCandidates === void 0 ? {} : { notLoadedCandidates: [...source.notLoadedCandidates] }
 		} }
 	};
 }
@@ -271761,6 +271984,7 @@ var Agent$1 = class {
 	errorMemory;
 	replayBuilder;
 	additionalDirs;
+	systemPromptStartedAt = /* @__PURE__ */ new Date();
 	activeProfile;
 	brandHome;
 	_activeResponderModel;
@@ -271831,6 +272055,7 @@ var Agent$1 = class {
 	}
 	personalMemoryRecallEligibility = () => this.type === "main" && this.personalMemoryHostCapabilityEnabled;
 	isPersonalMemoryHostTool = (name) => this.personalMemoryHostCapabilityEnabled && this.tools.hasUserTool(name);
+	localWorkHistorySearchEligibility = () => this.type === "main" && this.tools.hasUserTool("mcp__personal-memory__memory_search");
 	setPersonalMemoryHostCapability(enabled) {
 		this.personalMemoryHostCapabilityEnabled = enabled === true;
 	}
@@ -271847,6 +272072,7 @@ var Agent$1 = class {
 			const retryScope = providerRetryScope(modelAlias, provider);
 			await this.providerRetryWait.wait(retryScope, generateOptions?.signal);
 			const callerTransportAttempt = generateOptions?.onTransportAttempt;
+			const callerSerializedRequest = generateOptions?.onSerializedRequest;
 			const ownedGenerateOptions = {
 				...generateOptions,
 				transportMaxRetries: 0
@@ -271869,6 +272095,12 @@ var Agent$1 = class {
 							callerTransportAttempt?.(stats);
 							this.llmRequestLogger.logTransportAttempt(stats, requestLogFields, requestSequence);
 						},
+						...callerSerializedRequest !== void 0 || this.llmRequestLogger.serializedPrefixDiagnosticEnabled ? { onSerializedRequest: (body) => {
+							try {
+								callerSerializedRequest?.(body);
+							} catch {}
+							this.llmRequestLogger.observeSerializedRequest(body, requestLogFields, requestSequence);
+						} } : {},
 						onRequestStart: () => {
 							requestOptions?.onRequestStart?.();
 							requestStarted = true;
@@ -271894,7 +272126,7 @@ var Agent$1 = class {
 	get llm() {
 		return this.llmForTurn();
 	}
-	llmForTurn(thinkingEffort) {
+	llmForTurn(thinkingEffort, diagnosticTurnId) {
 		const selectedAlias = this.config.model;
 		const modelProvider = this.modelProvider;
 		if (modelProvider === void 0) throw new BlunError(ErrorCodes.MODEL_NOT_CONFIGURED, "Provider not set");
@@ -271909,6 +272141,7 @@ var Agent$1 = class {
 				messages: await rebuildMessages()
 			};
 			let systemPrompt = this.effectiveSystemPrompt;
+			const preflightStartedAt = performance.now();
 			const compacted = await this.fullCompaction.beforeStep(params.signal, {
 				messages: params.messages,
 				additionalMessages: steers.messages,
@@ -271917,6 +272150,13 @@ var Agent$1 = class {
 				rebuildSystemPrompt: () => systemPrompt = this.effectiveSystemPrompt,
 				tools: params.tools
 			});
+			if (this.type === "main" && diagnosticTurnId !== void 0 && process.env["BLUN_DIAG_PRE_REQUEST_PHASES"] === "1") try {
+				this.log.info("turn pre-request phase", {
+					turnId: diagnosticTurnId,
+					phase: "compaction-preflight",
+					durationMs: Math.max(0, Math.round(performance.now() - preflightStartedAt))
+				});
+			} catch {}
 			const tools = params.tools.some((tool) => tool.name === "CompactConversation") && !this.fullCompaction.proactiveCompactionStatus({
 				messages: compacted ? await rebuildMessages() : params.messages,
 				additionalMessages: steers.messages,
@@ -272021,6 +272261,7 @@ var Agent$1 = class {
 		let contributions;
 		const systemPrompt = profile.systemPrompt({
 			osEnv: this.kaos.osEnv,
+			now: this.systemPromptStartedAt,
 			cwd: this.config.cwd,
 			skills: this.skills?.registry,
 			cwdListing: context?.cwdListing,