@mastra/server 1.58.0-alpha.0 → 1.58.0-alpha.3

package/dist/{dist-DGqlUp0b.js → dist-BKRMiJcn.js}

@@ -5959,7 +5959,7 @@ function splitLines(text) {
 	return result;
 }
 //#endregion
-//#region ../memory/dist/src-Cwt9gefz.js
+//#region ../memory/dist/src-CdqJP57F.js
 var __defProp$3 = Object.defineProperty;
 var __exportAll = (all, no_symbols) => {
 	let target = {};
@@ -20722,6 +20722,7 @@ var Extractor = class Extractor {
 	includePreviousExtraction;
 	metadataKeyPath;
 	onExtracted;
+	retryStructuredExtractionOnEmptyObject;
 	/** @internal */
 	internal;
 	instructionsConfig;
@@ -20744,6 +20745,7 @@ var Extractor = class Extractor {
 		this.includePreviousExtraction = config.includePreviousExtraction ?? true;
 		this.metadataKeyPath = config.metadataKeyPath ?? `extracted.${slug}`;
 		this.onExtracted = config.onExtracted;
+		this.retryStructuredExtractionOnEmptyObject = config.retryStructuredExtractionOnEmptyObject ?? false;
 		this.internal = internal;
 	}
 	async resolve(context) {
@@ -20756,7 +20758,8 @@ var Extractor = class Extractor {
 			...schema ? { schema } : {},
 			includePreviousExtraction: this.includePreviousExtraction,
 			metadataKeyPath: this.metadataKeyPath,
-			onExtracted: this.onExtracted
+			onExtracted: this.onExtracted,
+			retryStructuredExtractionOnEmptyObject: this.retryStructuredExtractionOnEmptyObject
 		}, this.internal);
 	}
 };
@@ -21075,6 +21078,9 @@ if (OM_DEBUG_LOG) {
 function isAbortError$1(error, abortSignal) {
 	return abortSignal?.aborted === true || error instanceof DOMException && error.name === "AbortError" || error instanceof Error && error.name === "AbortError";
 }
+function shouldRetryEmptyStructuredObject(object, extractors) {
+	return Object.keys(object).length === 0 && extractors.some((extractor) => extractor.retryStructuredExtractionOnEmptyObject);
+}
 async function extractStructuredValues(opts) {
 	const structuredExtractors = (opts.extractors ?? []).filter((extractor) => extractor.mode === "structured");
 	if (structuredExtractors.length === 0) return {
@@ -21111,8 +21117,10 @@ ${extractorInstructions}${priorLines.length > 0 ? `\n\n## Prior Extracted Values
 		return output.object;
 	};
 	let object;
+	let retryEmptyObject = false;
 	try {
 		object = await generateWithStructuredOutput();
+		retryEmptyObject = shouldRetryEmptyStructuredObject(object, structuredExtractors);
 	} catch (error) {
 		if (isAbortError$1(error, opts.abortSignal)) throw error;
 		try {
@@ -21129,6 +21137,19 @@ ${extractorInstructions}${priorLines.length > 0 ? `\n\n## Prior Extracted Values
 			};
 		}
 	}
+	if (retryEmptyObject) try {
+		object = await generateWithStructuredOutput(coreFeatures.has("json-prompt-injection:inline") ? "inline" : true);
+	} catch (fallbackError) {
+		if (isAbortError$1(fallbackError, opts.abortSignal)) throw fallbackError;
+		const message = fallbackError instanceof Error ? fallbackError.message : String(fallbackError);
+		return {
+			values,
+			failures: structuredExtractors.map((extractor) => ({
+				slug: extractor.slug,
+				error: message
+			}))
+		};
+	}
 	for (const extractor of structuredExtractors) {
 		const value = object[extractor.slug];
 		if (value === void 0 || value === null || value === "") continue;
@@ -24412,6 +24433,17 @@ var TokenCounter = class TokenCounter {
 					toolResultDelta
 				};
 			}
+			if (invocation.state === "output-error") {
+				toolResultDelta++;
+				const errorText = invocation.errorText;
+				const errorMessage = typeof errorText === "string" ? errorText : "Tool execution failed";
+				tokens += this.readOrPersistPartEstimate(part, "tool-result-error", errorMessage);
+				return {
+					tokens,
+					overheadDelta,
+					toolResultDelta
+				};
+			}
 			throw new Error(`Unhandled tool-invocation state '${part.toolInvocation?.state}' in token counting for part type '${part.type}'`);
 		}
 		if (typeof part.type === "string" && part.type.startsWith("data-")) return {
@@ -24633,6 +24665,7 @@ var WorkingMemoryExtractor = class extends Extractor {
 			name: "Working Memory",
 			includePreviousExtraction: false,
 			metadataKeyPath: false,
+			retryStructuredExtractionOnEmptyObject: true,
 			instructions: async (context) => buildWorkingMemoryInstructions(await getWorkingMemoryDetails(context)),
 			schema: async (context) => {
 				return (await getWorkingMemoryDetails(context)).usesSchema ? z$2.union([z$2.record(z$2.string(), z$2.unknown()), z$2.null()]) : void 0;
@@ -24779,9 +24812,10 @@ async function listThreadsForResource({ memory, resourceId, currentThreadId, pag
 		hasMore
 	};
 }
+const SEARCH_NOT_CONFIGURED_MESSAGE = "Search is not configured. Enable it with `retrieval: { vector: true }` and configure a vector store and embedder on your Memory instance.";
 async function searchMessagesForResource({ memory, resourceId, currentThreadId, query, topK = 10, maxTokens = DEFAULT_MAX_RESULT_TOKENS, before, after, threadScope }) {
 	if (!memory.searchMessages) return {
-		results: "Search is not configured. Enable it with `retrieval: { vector: true }` and configure a vector store and embedder on your Memory instance.",
+		results: SEARCH_NOT_CONFIGURED_MESSAGE,
 		count: 0
 	};
 	const MAX_TOPK = 20;
@@ -25327,9 +25361,16 @@ async function recallThreadFromStart({ memory, threadId, resourceId, page = 1, l
 }
 const recallTool = (_memoryConfig, options) => {
 	const isResourceScope = (options?.retrievalScope ?? "thread") === "resource";
+	const searchEnabled = options?.searchEnabled ?? true;
+	const description = isResourceScope ? `Browse conversation history. Use mode="threads" to list all threads for the current user. Use mode="messages" (default) to browse messages in the current thread or pass threadId to browse another thread in the active resource. When mode="messages" has no cursor or threadId, it defaults to the current thread and says so at the top of the result. If you pass only a cursor, it must belong to the current thread.${searchEnabled ? " Use mode=\"search\" to find messages by content across all threads." : ""}` : `Browse conversation history in the current thread. Use mode="messages" (default) to page through messages near a cursor.${searchEnabled ? " Use mode=\"search\" to find messages by content in this thread." : ""} Use mode="threads" to get the current thread's ID and title.`;
+	const modeEnum = searchEnabled ? [
+		"messages",
+		"threads",
+		"search"
+	] : ["messages", "threads"];
 	return createTool({
 		id: "recall",
-		description: isResourceScope ? "Browse conversation history. Use mode=\"threads\" to list all threads for the current user. Use mode=\"messages\" (default) to browse messages in the current thread or pass threadId to browse another thread in the active resource. When mode=\"messages\" has no cursor or threadId, it defaults to the current thread and says so at the top of the result. If you pass only a cursor, it must belong to the current thread. Use mode=\"search\" to find messages by content across all threads." : "Browse conversation history in the current thread. Use mode=\"messages\" (default) to page through messages near a cursor. Use mode=\"search\" to find messages by content in this thread. Use mode=\"threads\" to get the current thread's ID and title.",
+		description,
 		inputSchema: {
 			$schema: "http://json-schema.org/draft-07/schema#",
 			type: "object",
@@ -25337,12 +25378,8 @@ const recallTool = (_memoryConfig, options) => {
 				...isResourceScope ? {
 					mode: {
 						type: "string",
-						enum: [
-							"messages",
-							"threads",
-							"search"
-						],
-						description: "What to retrieve. \"messages\" (default) pages through message history. \"threads\" lists all threads for the current user. \"search\" finds messages by semantic similarity across all threads."
+						enum: modeEnum,
+						description: `What to retrieve. "messages" (default) pages through message history. "threads" lists all threads for the current user.${searchEnabled ? " \"search\" finds messages by semantic similarity across all threads." : ""}`
 					},
 					threadId: {
 						type: "string",
@@ -25359,18 +25396,14 @@ const recallTool = (_memoryConfig, options) => {
 					}
 				} : { mode: {
 					type: "string",
-					enum: [
-						"messages",
-						"threads",
-						"search"
-					],
-					description: "What to retrieve. \"messages\" (default) pages through message history. \"threads\" returns info about the current thread. \"search\" finds messages by semantic similarity in this thread."
+					enum: modeEnum,
+					description: `What to retrieve. "messages" (default) pages through message history. "threads" returns info about the current thread.${searchEnabled ? " \"search\" finds messages by semantic similarity in this thread." : ""}`
 				} },
-				query: {
+				...searchEnabled ? { query: {
 					type: "string",
 					minLength: 1,
 					description: "Search query for mode=\"search\". Finds messages semantically similar to this text."
-				},
+				} } : {},
 				cursor: {
 					type: "string",
 					minLength: 1,
@@ -25431,6 +25464,10 @@ const recallTool = (_memoryConfig, options) => {
 			if (!memory) throw new Error("Memory instance is required for recall");
 			if (explicitThreadId === "current" && !currentThreadId) throw new Error("Could not resolve current thread.");
 			if (mode === "search") {
+				if (!searchEnabled) return {
+					results: SEARCH_NOT_CONFIGURED_MESSAGE,
+					count: 0
+				};
 				if (!query) throw new Error("query is required for mode=\"search\"");
 				if (!resourceId) throw new Error("Resource ID is required for recall");
 				return searchMessagesForResource({
@@ -26374,6 +26411,24 @@ function getUnobservedPartsPreservingToolCallPairs(message) {
 	return preservedCalls.length > 0 ? [...preservedCalls, ...unobservedParts] : unobservedParts;
 }
 /**
+* Get the messages Observational Memory is allowed to work with.
+*
+* Messages supplied through the `context` option are per-run ephemeral input. Core's
+* persistence contract already treats them as never-persist: `MessageStateManager` routes
+* them into `userContextMessages`, and `drainUnsavedMessages` only drains input/response.
+*
+* OM builds its windows from `get.all.db()`, which includes context messages, and then
+* seals and persists candidates directly — turning ephemeral context into durable user
+* messages. Excluding them here keeps OM's window, sealing, persistence and token
+* accounting consistent with that contract.
+*/
+function getObservableMessages(messageList) {
+	const allMessages = messageList.get.all.db();
+	const contextMessageIds = messageList.makeMessageSourceChecker().context;
+	if (contextMessageIds.size === 0) return allMessages;
+	return allMessages.filter((message) => !contextMessageIds.has(message.id));
+}
+/**
 * Safely extract buffered observation chunks from a record.
 * Handles both array and JSON-string formats, returning empty array if malformed.
 */
@@ -26386,7 +26441,7 @@ function getUnobservedPartsPreservingToolCallPairs(message) {
 */
 function filterObservedMessages(opts) {
 	const { messageList, record } = opts;
-	const allMessages = messageList.get.all.db();
+	const allMessages = getObservableMessages(messageList);
 	const useMarkerBoundaryPruning = opts.useMarkerBoundaryPruning ?? true;
 	const preserveMessageIds = opts.preserveMessageIds ?? /* @__PURE__ */ new Set();
 	const observedIds = new Set(Array.isArray(record?.observedMessageIds) ? record.observedMessageIds : []);
@@ -26798,7 +26853,7 @@ var ObservationStrategy = class ObservationStrategy {
 			transient: true
 		}).catch(() => {});
 		const markerThreadId = marker.data?.threadId ?? this.opts.threadId;
-		await this.persistMarkerToStorage(marker, markerThreadId, this.opts.resourceId);
+		if (!await this.persistMarkerToMessage(marker, this.opts.messageList, markerThreadId, this.opts.resourceId)) await this.persistMarkerToStorage(marker, markerThreadId, this.opts.resourceId);
 	}
 	getObservationMarkerConfig() {
 		return {
@@ -26926,10 +26981,14 @@ var ObservationStrategy = class ObservationStrategy {
 	/**
 	* Persist a marker part on the last assistant message in a MessageList
 	* AND save the updated message to the DB.
+	*
+	* @returns true when a marker was placed on an assistant message, false when
+	*   no list was provided or the list contains no assistant message (caller
+	*   should fall back to `persistMarkerToStorage`).
 	*/
 	async persistMarkerToMessage(marker, messageList, threadId, resourceId) {
-		if (!messageList) return;
-		const allMsgs = messageList.get.all.db();
+		if (!messageList) return false;
+		const allMsgs = getObservableMessages(messageList);
 		for (let i = allMsgs.length - 1; i >= 0; i--) {
 			const msg = allMsgs[i];
 			if (msg?.role === "assistant" && msg.content?.parts && Array.isArray(msg.content.parts)) {
@@ -26944,9 +27003,10 @@ var ObservationStrategy = class ObservationStrategy {
 				} catch (e) {
 					omDebug(`[OM:persistMarker] failed to save marker to DB: ${e}`);
 				}
-				return;
+				return true;
 			}
 		}
+		return false;
 	}
 };
 var SyncObservationStrategy = class extends ObservationStrategy {
@@ -27743,6 +27803,12 @@ var ObservationStep = class {
 	stepNumber;
 	_prepared = false;
 	_context;
+	/**
+	* True when this step seeded an empty assistant response message for a step-0
+	* observation. While set, the response-id rotation hook must NOT run — rotating
+	* would orphan the seed (markers would sit on a message the agent never streams into).
+	*/
+	seededResponseMessage = false;
 	constructor(turn, stepNumber) {
 		this.turn = turn;
 		this.stepNumber = stepNumber;
@@ -27789,7 +27855,7 @@ var ObservationStep = class {
 		let didThresholdCleanup = false;
 		let observerExchange;
 		if (this.stepNumber === 0) {
-			const step0Messages = messageList.get.all.db();
+			const step0Messages = getObservableMessages(messageList);
 			const activation = await om.activate({
 				threadId,
 				resourceId,
@@ -27821,7 +27887,7 @@ var ObservationStep = class {
 				currentModel: this.turn.actorModelContext,
 				requestContext: this.turn.requestContext,
 				observabilityContext: this.turn.observabilityContext,
-				lastActivityAt: getLastActivityFromMessages(messageList.get.all.db()),
+				lastActivityAt: getLastActivityFromMessages(getObservableMessages(messageList)),
 				reflectionHooks: om.composeHooks(void 0, {
 					threadId,
 					resourceId,
@@ -27831,7 +27897,7 @@ var ObservationStep = class {
 			await this.turn.refreshRecord();
 			if (this.turn.record.generationCount > preReflectGeneration) reflected = true;
 		}
-		const allMsgsForToolCheck = messageList.get.all.db();
+		const allMsgsForToolCheck = getObservableMessages(messageList);
 		const lastMessage = allMsgsForToolCheck[allMsgsForToolCheck.length - 1];
 		const pendingStepMessages = [...messageList.get.input.db(), ...messageList.get.response.db()];
 		const latestStepParts = [...getLatestStepParts(lastMessage?.content?.parts ?? []), ...pendingStepMessages.flatMap((msg) => getLatestStepParts(msg.content?.parts ?? []))];
@@ -27840,10 +27906,10 @@ var ObservationStep = class {
 		let statusSnapshot = await om.getStatus({
 			threadId,
 			resourceId,
-			messages: messageList.get.all.db()
+			messages: getObservableMessages(messageList)
 		});
 		if (statusSnapshot.shouldBuffer && !hasIncompleteToolCalls) {
-			const allMessages = messageList.get.all.db();
+			const allMessages = getObservableMessages(messageList);
 			const unobservedMessages = om.getUnobservedMessages(allMessages, statusSnapshot.record);
 			const candidates = om.getUnobservedMessages(unobservedMessages, statusSnapshot.record, { excludeBuffered: true });
 			if (candidates.length > 0) {
@@ -27871,15 +27937,41 @@ var ObservationStep = class {
 			});
 			buffered = true;
 		}
-		if (this.stepNumber > 0) {
-			const newInput = messageList.clear.input.db();
-			const newOutput = messageList.clear.response.db();
-			const messagesToSave = [...newInput, ...newOutput];
-			if (messagesToSave.length > 0) {
-				await om.persistMessages(messagesToSave, threadId, resourceId);
-				for (const msg of messagesToSave) messageList.add(msg, "memory");
+		const willObserveNow = statusSnapshot.shouldObserve && !hasIncompleteToolCalls;
+		/** In-flight message ids the step-0 cleanup must never remove from live context. */
+		let step0PreserveIds;
+		if (this.stepNumber > 0 || willObserveNow) {
+			if (this.stepNumber > 0) {
+				const newInput = messageList.clear.input.db();
+				const newOutput = messageList.clear.response.db();
+				const messagesToSave = [...newInput, ...newOutput];
+				if (messagesToSave.length > 0) {
+					await om.persistMessages(messagesToSave, threadId, resourceId);
+					for (const msg of messagesToSave) messageList.add(msg, "memory");
+				}
+			} else {
+				const pending = [...messageList.get.input.db(), ...messageList.get.response.db()];
+				if (pending.length > 0) await om.persistMessages(pending, threadId, resourceId);
+				step0PreserveIds = pending.map((msg) => msg.id);
+			}
+			if (this.stepNumber === 0 && willObserveNow && this.turn.responseMessageId) {
+				const seed = {
+					id: this.turn.responseMessageId,
+					role: "assistant",
+					content: {
+						format: 2,
+						parts: []
+					},
+					type: "text",
+					createdAt: /* @__PURE__ */ new Date(),
+					threadId,
+					resourceId
+				};
+				messageList.add(seed, "response");
+				this.seededResponseMessage = true;
+				omDebug(`[OM:step0] seeded response message ${seed.id} for step-0 observation markers`);
 			}
-			if (statusSnapshot.shouldObserve && !hasIncompleteToolCalls) {
+			if (willObserveNow) {
 				const preObsGeneration = this.turn.record.generationCount;
 				const obsResult = await this.runThresholdObservation();
 				observerExchange = obsResult.observerExchange;
@@ -27893,7 +27985,8 @@ var ObservationStep = class {
 						resourceId,
 						messages: messageList,
 						observedMessageIds: observedIds,
-						retentionFloor: minRemaining
+						retentionFloor: minRemaining,
+						preserveMessageIds: step0PreserveIds
 					});
 					if (statusSnapshot.asyncObservationEnabled) await om.resetBufferingState({
 						threadId,
@@ -27908,7 +28001,7 @@ var ObservationStep = class {
 			statusSnapshot = await om.getStatus({
 				threadId,
 				resourceId,
-				messages: messageList.get.all.db()
+				messages: getObservableMessages(messageList)
 			});
 		}
 		const otherThreadsContext = await this.turn.refreshOtherThreadsContext();
@@ -27957,10 +28050,11 @@ var ObservationStep = class {
 		const { threadId, resourceId, messageList } = this.turn;
 		const om = this.turn.om;
 		await om.waitForBuffering(threadId, resourceId);
+		const observableMessages = this.seededResponseMessage ? getObservableMessages(messageList).filter((msg) => msg.id !== this.turn.responseMessageId) : getObservableMessages(messageList);
 		const freshStatus = await om.getStatus({
 			threadId,
 			resourceId,
-			messages: messageList.get.all.db()
+			messages: observableMessages
 		});
 		if (!freshStatus.shouldObserve) return {
 			succeeded: false,
@@ -27970,7 +28064,7 @@ var ObservationStep = class {
 			const activation = await om.activate({
 				threadId,
 				resourceId,
-				messages: messageList.get.all.db(),
+				messages: observableMessages,
 				currentModel: this.turn.actorModelContext,
 				writer: this.turn.writer,
 				messageList
@@ -27986,7 +28080,7 @@ var ObservationStep = class {
 					currentModel: this.turn.actorModelContext,
 					requestContext: this.turn.requestContext,
 					observabilityContext: this.turn.observabilityContext,
-					lastActivityAt: getLastActivityFromMessages(messageList.get.all.db()),
+					lastActivityAt: getLastActivityFromMessages(getObservableMessages(messageList)),
 					reflectionHooks: om.composeHooks(void 0, {
 						threadId,
 						resourceId,
@@ -28003,7 +28097,8 @@ var ObservationStep = class {
 		const obsResult = await om.observe({
 			threadId,
 			resourceId,
-			messages: messageList.get.all.db(),
+			messages: observableMessages,
+			messageList,
 			trigger: "turn-sync",
 			requestContext: this.turn.requestContext,
 			writer: this.turn.writer,
@@ -28011,7 +28106,7 @@ var ObservationStep = class {
 		});
 		if (obsResult.observed) {
 			const observedMessageIds = new Set(obsResult.record.observedMessageIds ?? []);
-			const liveMessages = messageList.get.all.db();
+			const liveMessages = getObservableMessages(messageList);
 			let latestObservedIndex = -1;
 			for (let i = liveMessages.length - 1; i >= 0; i--) {
 				const message = liveMessages[i];
@@ -28020,10 +28115,12 @@ var ObservationStep = class {
 					break;
 				}
 			}
-			const messageToSeal = latestObservedIndex >= 0 ? liveMessages[latestObservedIndex] : void 0;
+			let messageToSeal = latestObservedIndex >= 0 ? liveMessages[latestObservedIndex] : void 0;
+			if (this.stepNumber === 0 && messageToSeal?.role !== "assistant") messageToSeal = void 0;
 			const messagesToSeal = messageToSeal ? [messageToSeal] : [];
 			om.sealMessagesForBuffering(messagesToSeal);
-			try {
+			if (this.seededResponseMessage) omDebug("[OM:observe] skipping response-id rotation — step-0 seeded response message holds the active id");
+			else try {
 				await this.turn.hooks?.onSyncObservationComplete?.();
 			} catch (error) {
 				omDebug(`[OM:observe] onSyncObservationComplete hook failed: ${error instanceof Error ? error.message : String(error)}`);
@@ -28082,6 +28179,8 @@ var ObservationTurn = class {
 	sendSignal;
 	/** Current actor model for this step. Updated by the processor before prepare(). */
 	actorModelContext;
+	/** The active assistant response message ID for this step. Updated by the processor before prepare(). */
+	responseMessageId;
 	/** Processor-provided hooks for turn/step lifecycle integration. */
 	hooks;
 	constructor(opts) {
@@ -28180,7 +28279,7 @@ var ObservationTurn = class {
 		const asyncObservationEnabled = this.om.buffering.isAsyncObservationEnabled();
 		const bufferOnIdle = this.om.getObservationConfig().bufferOnIdle;
 		if (asyncObservationEnabled && bufferOnIdle) {
-			const allMessages = this.messageList.get.all.db();
+			const allMessages = getObservableMessages(this.messageList);
 			const record = this._record;
 			const unobservedMessages = this.om.getUnobservedMessages(allMessages, record);
 			if (unobservedMessages.length > 0) this.om.buffer({
@@ -28546,7 +28645,7 @@ function getCurrentModel$1(model) {
 	return formatModelContext$1(model?.provider, model?.modelId);
 }
 function getLastModelFromMessageList(messageList) {
-	const messages = messageList?.get.all.db();
+	const messages = messageList ? getObservableMessages(messageList) : void 0;
 	if (!messages) return void 0;
 	for (let i = messages.length - 1; i >= 0; i--) {
 		const message = messages[i];
@@ -30126,7 +30225,7 @@ var ObservationalMemory = class ObservationalMemory {
 	*/
 	async persistMarkerToMessage(marker, messageList, threadId, resourceId) {
 		if (!messageList) return;
-		const allMsgs = messageList.get.all.db();
+		const allMsgs = getObservableMessages(messageList);
 		for (let i = allMsgs.length - 1; i >= 0; i--) {
 			const msg = allMsgs[i];
 			if (msg?.role === "assistant" && msg.content?.parts && Array.isArray(msg.content.parts)) {
@@ -30742,6 +30841,7 @@ ${formattedMessages}
 	*/
 	async getObservedMessageIdsForCleanup(opts) {
 		const { threadId, resourceId, messages, observedMessageIds, retentionFloor } = opts;
+		const preserveSet = opts.preserveMessageIds?.length ? new Set(opts.preserveMessageIds) : null;
 		const record = await this.getOrCreateRecord(threadId, resourceId);
 		const effectiveObservedIds = observedMessageIds && observedMessageIds.length > 0 ? observedMessageIds : Array.isArray(record.observedMessageIds) ? record.observedMessageIds : [];
 		if (effectiveObservedIds.length === 0) return [];
@@ -30753,6 +30853,10 @@ ${formattedMessages}
 		const retentionCounter = typeof retentionFloor === "number" ? new TokenCounter() : null;
 		for (const msg of messages) {
 			if (!msg?.id || msg.id === "om-continuation" || !observedSet.has(msg.id)) continue;
+			if (preserveSet?.has(msg.id)) {
+				skipped += 1;
+				continue;
+			}
 			const unobservedParts = getUnobservedParts(msg);
 			const totalParts = msg.content?.parts?.length ?? 0;
 			if (unobservedParts.length > 0 && unobservedParts.length < totalParts) {
@@ -30797,9 +30901,9 @@ ${formattedMessages}
 	*/
 	/** @internal Used by ObservationStep. */
 	async cleanupMessages(opts) {
-		const { threadId, resourceId, observedMessageIds, retentionFloor } = opts;
+		const { threadId, resourceId, observedMessageIds, retentionFloor, preserveMessageIds } = opts;
 		const messageList = this.isMessageList(opts.messages) ? opts.messages : void 0;
-		const allMsgs = messageList ? messageList.get.all.db() : opts.messages;
+		const allMsgs = messageList ? getObservableMessages(messageList) : opts.messages;
 		let markerIdx = -1;
 		let markerMsg = null;
 		for (let i = allMsgs.length - 1; i >= 0; i--) {
@@ -30818,7 +30922,8 @@ ${formattedMessages}
 				resourceId,
 				messages: allMsgs,
 				observedMessageIds,
-				retentionFloor
+				retentionFloor,
+				preserveMessageIds
 			});
 			if (messageList) {
 				if (idsToRemoveList.length > 0) messageList.removeByIds(idsToRemoveList);
@@ -30830,18 +30935,21 @@ ${formattedMessages}
 		if (markerMsg && markerIdx !== -1) {
 			const idsToRemove = [];
 			const messagesToSave = [];
+			const preserveSet = preserveMessageIds?.length ? new Set(preserveMessageIds) : null;
 			for (let i = 0; i < markerIdx; i++) {
 				const msg = allMsgs[i];
-				if (msg?.id && msg.id !== "om-continuation") {
+				if (msg?.id && msg.id !== "om-continuation" && !preserveSet?.has(msg.id)) {
 					idsToRemove.push(msg.id);
 					messagesToSave.push(msg);
 				}
 			}
 			messagesToSave.push(markerMsg);
-			const unobservedParts = getUnobservedParts(markerMsg);
-			if (unobservedParts.length === 0) {
-				if (markerMsg.id) idsToRemove.push(markerMsg.id);
-			} else if (unobservedParts.length < (markerMsg.content?.parts?.length ?? 0)) markerMsg.content.parts = unobservedParts;
+			if (!Boolean(markerMsg.id && preserveSet?.has(markerMsg.id))) {
+				const unobservedParts = getUnobservedParts(markerMsg);
+				if (unobservedParts.length === 0) {
+					if (markerMsg.id) idsToRemove.push(markerMsg.id);
+				} else if (unobservedParts.length < (markerMsg.content?.parts?.length ?? 0)) markerMsg.content.parts = unobservedParts;
+			}
 			if (messageList) {
 				if (idsToRemove.length > 0) messageList.removeByIds(idsToRemove);
 				if (messagesToSave.length > 0) await this.persistMessages(messagesToSave, threadId, resourceId);
@@ -31628,6 +31736,7 @@ ${formattedMessages}
 					threadId,
 					resourceId,
 					messages: unobservedMessages,
+					messageList: opts.messageList,
 					reflectionHooks,
 					agent: opts.agent,
 					requestContext,
@@ -31908,7 +32017,7 @@ function isTemporalGapMarkerForMessage(message, targetMessageId) {
 async function insertTemporalGapMarkers({ messageList, sendSignal }) {
 	const latestInputMessage = messageList.get.input.db().filter((message) => Boolean(message)).at(-1);
 	if (!latestInputMessage || isTemporalGapMarker(latestInputMessage)) return;
-	const allMessages = messageList.get.all.db().filter((message) => Boolean(message));
+	const allMessages = getObservableMessages(messageList).filter((message) => Boolean(message));
 	const latestInputIndex = allMessages.findIndex((message) => message.id === latestInputMessage.id);
 	if (latestInputIndex <= 0) return;
 	if (allMessages.some((message) => isTemporalGapMarkerForMessage(message, latestInputMessage.id))) return;
@@ -32013,7 +32122,7 @@ var ObservationalMemoryProcessor = class {
 		this.temporalMarkers = options?.temporalMarkers ?? false;
 	}
 	async processInputStep(args) {
-		const { messageList, requestContext, stepNumber, state: _state, writer, model, abortSignal, abort, rotateResponseMessageId } = args;
+		const { messageList, requestContext, stepNumber, state: _state, writer, model, abortSignal, abort, messageId, rotateResponseMessageId } = args;
 		const state = _state ?? {};
 		omDebug(`[OM:processInputStep:ENTER] step=${stepNumber}, hasMastraMemory=${!!requestContext?.get("MastraMemory")}, hasMemoryInfo=${!!messageList?.serialize()?.memoryInfo?.threadId}`);
 		const context = this.engine.getThreadContext(requestContext, messageList);
@@ -32099,6 +32208,7 @@ var ObservationalMemoryProcessor = class {
 			state.__omObservabilityContext = observabilityContext;
 			this.turn.observabilityContext = observabilityContext;
 			this.turn.actorModelContext = actorModelContext;
+			this.turn.responseMessageId = messageId;
 			{
 				const step = this.turn.step(stepNumber);
 				let ctx;
@@ -32129,7 +32239,7 @@ var ObservationalMemoryProcessor = class {
 					threadId,
 					resourceId
 				});
-				const allDbMsgs = messageList.get.all.db();
+				const allDbMsgs = getObservableMessages(messageList);
 				const tokenCounter = this.engine.getTokenCounter();
 				const contextTokens = await tokenCounter.countMessagesAsync(allDbMsgs);
 				const otherThreadsContext = this.turn.context.otherThreadsContext;
@@ -32476,6 +32586,18 @@ var Memory = class extends MastraMemory {
 	_omEngine;
 	_omEngineInstance;
 	_mastraInstance;
+	/**
+	* Every vector cleanup that deleteThread or deleteMessages started in the background.
+	* Callers do not wait for the cleanup, so this handle is the only join point.
+	*/
+	pendingVectorCleanup = Promise.resolve();
+	/**
+	* Adds a background vector cleanup to the join handle.
+	* The handle keeps the earlier cleanups, so it settles only after all of them end.
+	*/
+	trackVectorCleanup(cleanup) {
+		this.pendingVectorCleanup = Promise.allSettled([this.pendingVectorCleanup, cleanup]).then(() => void 0);
+	}
 	/** The shared ObservationalMemory engine. Lazily created on first access. */
 	get omEngine() {
 		if (!this._omEngine) this._omEngine = this._initOMEngine().then((engine) => {
@@ -32740,26 +32862,40 @@ var Memory = class extends MastraMemory {
 		const thread = await memoryStore.getThreadById({ threadId });
 		await memoryStore.deleteThread({ threadId });
 		if (thread?.resourceId && memoryStore.supportsObservationalMemory) await memoryStore.clearObservationalMemory(threadId, thread.resourceId);
-		if (this.vector) this.deleteThreadVectors(threadId);
+		if (this.vector) this.trackVectorCleanup(this.deleteThreadVectors(threadId));
 	}
 	/**
-	* Lists all vector indexes that match the memory messages prefix.
-	* Handles separator differences across vector store backends (e.g. '_' vs '-').
+	* Prefix shared by every message index. The index for the default embedding
+	* dimension is named with the bare prefix; other dimensions add a suffix.
 	*/
-	async getMemoryVectorIndexes() {
+	get messageIndexPrefix() {
+		return this.getEmbeddingIndexName();
+	}
+	/**
+	* Prefix shared by every observation index. Each observation index adds a dimension suffix.
+	*/
+	get observationIndexPrefix() {
+		return `memory${this.vector?.indexSeparator ?? "_"}observations`;
+	}
+	/**
+	* Lists the vector indexes whose name starts with one of the given prefixes.
+	* Index names can carry a dimension suffix, so discovery matches on the prefix.
+	*/
+	async getMemoryVectorIndexes(prefixes) {
 		if (!this.vector) return [];
-		const prefix = `memory${this.vector.indexSeparator ?? "_"}messages`;
-		return (await this.vector.listIndexes()).filter((name) => name.startsWith(prefix));
+		return (await this.vector.listIndexes()).filter((name) => prefixes.some((prefix) => name.startsWith(prefix)));
 	}
 	/**
 	* Deletes all vector embeddings associated with a thread.
 	* This is called internally by deleteThread to clean up orphaned vectors.
+	* Both message and observation vectors are removed, so no text of the deleted
+	* thread stays reachable through resource-scoped retrieval.
 	*
 	* @param threadId - The ID of the thread whose vectors should be deleted
 	*/
 	async deleteThreadVectors(threadId) {
 		try {
-			const memoryIndexes = await this.getMemoryVectorIndexes();
+			const memoryIndexes = await this.getMemoryVectorIndexes([this.messageIndexPrefix, this.observationIndexPrefix]);
 			await Promise.all(memoryIndexes.map(async (indexName) => {
 				try {
 					await this.vector.deleteVectors({
@@ -32767,14 +32903,14 @@ var Memory = class extends MastraMemory {
 						filter: { thread_id: threadId }
 					});
 				} catch {
-					this.logger.debug("Failed to delete vectors for thread, skipping", {
+					this.logger.warn("Failed to delete vectors of the deleted thread from index", {
 						threadId,
 						indexName
 					});
 				}
 			}));
 		} catch {
-			this.logger.debug("Failed to clean up vectors for thread", { threadId });
+			this.logger.warn("Failed to clean up vectors of the deleted thread", { threadId });
 		}
 	}
 	async updateWorkingMemory({ threadId, resourceId, workingMemory, memoryConfig, observabilityContext }) {
@@ -33447,7 +33583,7 @@ Notes:
 	getObservationEmbeddingIndexName(dimensions) {
 		const usedDimensions = dimensions ?? 384;
 		const separator = this.vector?.indexSeparator ?? "_";
-		return `memory${separator}observations${separator}${usedDimensions}`;
+		return `${this.observationIndexPrefix}${separator}${usedDimensions}`;
 	}
 	async createObservationEmbeddingIndex(dimensions) {
 		const usedDimensions = dimensions ?? 384;
@@ -33693,7 +33829,10 @@ Notes:
 			tools[name] = tool;
 		}
 		const omConfig = normalizeObservationalMemoryConfig(mergedConfig.observationalMemory);
-		if (omConfig?.retrieval) tools.recall = recallTool(mergedConfig, { retrievalScope: typeof omConfig.retrieval === "object" ? omConfig.retrieval.scope ?? "resource" : "resource" });
+		if (omConfig?.retrieval) tools.recall = recallTool(mergedConfig, {
+			retrievalScope: typeof omConfig.retrieval === "object" ? omConfig.retrieval.scope ?? "resource" : "resource",
+			searchEnabled: this.hasRetrievalSearch(omConfig.retrieval)
+		});
 		return tools;
 	}
 	/**
@@ -33749,7 +33888,7 @@ Notes:
 				}));
 				const messageIdsNeedingDeletion = /* @__PURE__ */ new Set([...messageIdsWithClearedContent, ...messageIdsWithNewEmbeddings]);
 				if (messageIdsNeedingDeletion.size > 0) try {
-					const memoryIndexes = await this.getMemoryVectorIndexes();
+					const memoryIndexes = await this.getMemoryVectorIndexes([this.messageIndexPrefix]);
 					const idsToDelete = [...messageIdsNeedingDeletion];
 					await Promise.all(memoryIndexes.map(async (indexName) => {
 						for (let i = 0; i < idsToDelete.length; i += VECTOR_DELETE_BATCH_SIZE) {
@@ -33808,7 +33947,7 @@ Notes:
 		const span = this.createMemorySpan("delete", observabilityContext, void 0, { messageCount: messageIds.length });
 		try {
 			await (await this.getMemoryStore()).deleteMessages(messageIds);
-			if (this.vector) this.deleteMessageVectors(messageIds);
+			if (this.vector) this.trackVectorCleanup(this.deleteMessageVectors(messageIds));
 			span?.end({
 				output: { success: true },
 				attributes: { messageCount: messageIds.length }
@@ -33824,12 +33963,14 @@ Notes:
 	/**
 	* Deletes vector embeddings for specific messages.
 	* This is called internally by deleteMessages to clean up orphaned vectors.
+	* Only the message indexes are touched, because observation vectors can hold
+	* text of other messages of the thread.
 	*
 	* @param messageIds - The IDs of the messages whose vectors should be deleted
 	*/
 	async deleteMessageVectors(messageIds) {
 		try {
-			const memoryIndexes = await this.getMemoryVectorIndexes();
+			const memoryIndexes = await this.getMemoryVectorIndexes([this.messageIndexPrefix]);
 			await Promise.all(memoryIndexes.map(async (indexName) => {
 				for (let i = 0; i < messageIds.length; i += VECTOR_DELETE_BATCH_SIZE) {
 					const batch = messageIds.slice(i, i + VECTOR_DELETE_BATCH_SIZE);
@@ -44602,4 +44743,4 @@ const agentBuilderWorkflows = {
 //#endregion
 export { agentBuilderWorkflows };
 
-//# sourceMappingURL=dist-DGqlUp0b.js.map
+//# sourceMappingURL=dist-BKRMiJcn.js.map