@mastra/memory 1.39.1-alpha.2 → 1.39.1-alpha.3

package/dist/{src-stSVckwY.js → src-kIpEG6XZ.js}

@@ -4,7 +4,7 @@ import * as z4 from "zod/v4";
 import { any, array, boolean, custom, discriminatedUnion, enum as enum$1, instanceof as instanceof$1, json, lazy, literal, looseObject, never, null as null$1, number, object, record, safeParseAsync, string, union, unknown, z as z$1 } from "zod/v4";
 import { toJSONSchema } from "zod/v4/core";
 import { Agent, MessageList } from "@mastra/core/agent";
-import { ErrorCategory, ErrorDomain, MastraError } from "@mastra/core/error";
+import { ErrorCategory, ErrorDomain, MastraError, classifyProviderError } from "@mastra/core/error";
 import { coreFeatures } from "@mastra/core/features";
 import { MastraMemory, getMemoryTokenBoundary, getThreadOMMetadata, isAfterMemoryTokenBoundary, isThreadTitlePinned, loadMessageHistory, normalizeMessageHistoryConfig, parseMemoryRequestContext, setThreadOMMetadata } from "@mastra/core/memory";
 import { EntityType, SpanType, createObservabilityContext, getOrCreateSpan } from "@mastra/core/observability";
@@ -27105,15 +27105,16 @@ var OmIncompleteResponseError = class extends Error {
 	}
 };
 /**
-* A model reply that finished but contains no text, or only whitespace.
+* A model reply that finished but carries nothing usable: no text, only whitespace,
+* or text with no recognised Observer sections (e.g. a refusal).
 * Retryable, so an empty reply is retried like a transport error.
 *
 * @internal
 */
 var OmEmptyResponseError = class extends Error {
 	isRetryable = true;
-	constructor(label) {
-		super(`${label} returned empty output`);
+	constructor(label, what = "empty output") {
+		super(`${label} returned ${what}`);
 		this.name = "OmEmptyResponseError";
 	}
 };
@@ -29644,13 +29645,15 @@ function parseMemorySectionXml(content) {
 		suggestedResponse: "",
 		threadTitle: ""
 	};
-	const observationsMatches = [...content.matchAll(/^[ \t]*<observations>([\s\S]*?)^[ \t]*<\/observations>/gim)];
-	if (observationsMatches.length > 0) result.observations = observationsMatches.map((m) => m[1]?.trim() ?? "").filter(Boolean).join("\n");
+	const observationsMatches = [...content.matchAll(/^[ \t]*<observations>(?:([^\r\n]*?)<\/observations>[ \t]*$|([\s\S]*?)^[ \t]*<\/observations>)/gim)];
+	if (observationsMatches.length > 0) result.observations = observationsMatches.map((m) => (m[1] ?? m[2])?.trim() ?? "").filter(Boolean).join("\n");
 	else result.observations = extractListItemsOnly(content);
-	const currentTaskMatch = content.match(/^[ \t]*<current-task>([\s\S]*?)^[ \t]*<\/current-task>/im);
-	if (currentTaskMatch?.[1]) result.currentTask = currentTaskMatch[1].trim();
-	const suggestedResponseMatch = content.match(/^[ \t]*<suggested-response>([\s\S]*?)^[ \t]*<\/suggested-response>/im);
-	if (suggestedResponseMatch?.[1]) result.suggestedResponse = suggestedResponseMatch[1].trim();
+	const currentTaskMatch = content.match(/^[ \t]*<current-task>(?:([^\r\n]*?)<\/current-task>[ \t]*$|([\s\S]*?)^[ \t]*<\/current-task>)/im);
+	const currentTask = currentTaskMatch?.[1] ?? currentTaskMatch?.[2];
+	if (currentTask) result.currentTask = currentTask.trim();
+	const suggestedResponseMatch = content.match(/^[ \t]*<suggested-response>(?:([^\r\n]*?)<\/suggested-response>[ \t]*$|([\s\S]*?)^[ \t]*<\/suggested-response>)/im);
+	const suggestedResponse = suggestedResponseMatch?.[1] ?? suggestedResponseMatch?.[2];
+	if (suggestedResponse) result.suggestedResponse = suggestedResponse.trim();
 	const threadTitleMatch = content.match(/^[ \t]*<thread-title>([\s\S]*?)<\/thread-title>/im);
 	if (threadTitleMatch?.[1]) result.threadTitle = threadTitleMatch[1].trim();
 	return result;
@@ -30223,7 +30226,7 @@ var ObserverRunner = class {
 				callback: (childObservabilityContext) => this.withAbortCheck(async () => {
 					try {
 						attemptMemory = temporaryMemory?.newThread();
-						return assertNonEmptyObserverOutput(assertCompleteModelResponse(await (await agent.stream(observerMessages, {
+						return assertRecognizedObserverOutput(assertNonEmptyObserverOutput(assertCompleteModelResponse(await (await agent.stream(observerMessages, {
 							maxSteps: 1,
 							modelSettings: { ...this.observationConfig.modelSettings },
 							providerOptions: this.observationConfig.providerOptions,
@@ -30232,7 +30235,7 @@ var ObserverRunner = class {
 							onChunk: onActivity,
 							...internalRequestContext ? { requestContext: internalRequestContext } : {},
 							...childObservabilityContext
-						})).getFullOutput(), "OM observer"), "OM observer");
+						})).getFullOutput(), "OM observer"), "OM observer"), "OM observer", (text) => hasSingleThreadObserverContent(text, activeExtractors));
 					} catch (error) {
 						this.mastra?.getLogger?.().error("OM observer provider call failed", {
 							diagnostic: formatOmError(error),
@@ -30335,7 +30338,7 @@ var ObserverRunner = class {
 		for (const threadId of allThreadOrder) {
 			const threadResult = output.results.get(threadId);
 			if (!threadResult) {
-				output.results.set(threadId, { observations: "" });
+				if (!messagesByThread.has(threadId)) output.results.set(threadId, { observations: "" });
 				continue;
 			}
 			threadResult.observations = await applyTextTransform(this.hooks, "afterObservation", threadResult.observations, contextFor(threadId));
@@ -30437,7 +30440,7 @@ var ObserverRunner = class {
 				},
 				callback: (childObservabilityContext) => this.withAbortCheck(async () => {
 					try {
-						return assertNonEmptyObserverOutput(assertCompleteModelResponse(await (await agent.stream(observerMessages, {
+						return assertRecognizedObserverOutput(assertNonEmptyObserverOutput(assertCompleteModelResponse(await (await agent.stream(observerMessages, {
 							maxSteps: 1,
 							modelSettings: { ...this.observationConfig.modelSettings },
 							providerOptions: this.observationConfig.providerOptions,
@@ -30445,7 +30448,7 @@ var ObserverRunner = class {
 							onChunk: onActivity,
 							...internalRequestContext ? { requestContext: internalRequestContext } : {},
 							...childObservabilityContext
-						})).getFullOutput(), "OM multi-thread observer"), "OM multi-thread observer");
+						})).getFullOutput(), "OM multi-thread observer"), "OM multi-thread observer"), "OM multi-thread observer", (text) => hasMultiThreadObserverContent(text, activeExtractors, threadOrder));
 					} catch (error) {
 						this.mastra?.getLogger?.().error("OM multi-thread observer provider call failed", {
 							diagnostic: formatOmError(error),
@@ -30522,8 +30525,7 @@ var ObserverRunner = class {
 				extractors: activeExtractors
 			});
 		}
-		for (const threadId of threadOrder) if (!results.has(threadId)) results.set(threadId, { observations: "" });
-		for (const msgs of messagesByThread.values()) for (const msg of msgs) this.observedMessageIds.add(msg.id);
+		for (const threadId of results.keys()) for (const msg of messagesByThread.get(threadId) ?? []) this.observedMessageIds.add(msg.id);
 		const usage = result.totalUsage ?? result.usage;
 		return {
 			results,
@@ -30546,6 +30548,29 @@ function assertNonEmptyObserverOutput(output, label) {
 	if (!output.text?.trim()) throw new OmEmptyResponseError(label);
 	return output;
 }
+/**
+* A reply with text but nothing the parser recognises (e.g. a refusal like
+* "I can't help with that") parses to zero observations. Committing it would drop
+* the observed messages the same way an empty reply does, so it gets the same retry.
+* Degenerate replies pass through; the degenerate retry handles them.
+*/
+function assertRecognizedObserverOutput(output, label, isRecognized) {
+	if (!isRecognized(output.text)) throw new OmEmptyResponseError(label, "no observations");
+	return output;
+}
+function hasSingleThreadObserverContent(text, extractors) {
+	if (/^[ \t]*<observations>/im.test(text)) return true;
+	const parsed = parseObserverOutput(text, extractors);
+	return Boolean(parsed.degenerate || parsed.observations || parsed.currentTask || parsed.suggestedContinuation || parsed.threadTitle || Object.keys(parsed.extractedValues ?? {}).length || parsed.extractionFailures?.length);
+}
+/**
+* Multi-thread replies carry every recognised section inside `<thread>` blocks, so a reply
+* with no block for any requested thread has nothing to record.
+*/
+function hasMultiThreadObserverContent(text, extractors, threadOrder) {
+	const parsed = parseMultiThreadObserverOutput(text, extractors);
+	return Boolean(parsed.degenerate || threadOrder.some((threadId) => parsed.threads.has(threadId)));
+}
 // src/processors/observational-memory/measure-image-buffer.ts
 /**
 * Synchronous image dimension lookup for in-memory image buffers.
@@ -42080,54 +42105,6 @@ ${formattedMessages}
 		});
 	}
 };
-// src/processors/observational-memory/context-overflow.ts
-/**
-* Detects provider errors that reject a request for exceeding the model's context window.
-*
-* Providers report this with different codes and wording, so this matches the known shapes
-* (OpenAI and OpenAI-compatible APIs, Anthropic, Google, Bedrock, xAI, Mistral, local servers)
-* against the error's message, code, response body and causes.
-*/
-const CONTEXT_OVERFLOW_PATTERNS = [
-	/context[_ ]length[_ ]exceeded/i,
-	/maximum context length/i,
-	/maximum prompt length/i,
-	/prompt is too long/i,
-	/input is too long/i,
-	/exceed\w*\b[^.]{0,40}\bcontext (?:window|length|limit|size)/i,
-	/input token count[^.]{0,40}exceeds the maximum/i,
-	/reduce the length of the (?:messages|prompt|input)/i
-];
-const MAX_CAUSE_DEPTH = 4;
-function collectErrorText(error, depth = 0) {
-	if (depth > MAX_CAUSE_DEPTH || error == null) return [];
-	if (typeof error === "string") return [error];
-	if (typeof error !== "object") return [];
-	const record = error;
-	const texts = [];
-	for (const key of [
-		"message",
-		"code",
-		"type",
-		"responseBody"
-	]) {
-		const value = record[key];
-		if (typeof value === "string") texts.push(value);
-	}
-	if (record.data && typeof record.data === "object") texts.push(...collectErrorText(record.data.error ?? record.data, depth + 1));
-	if (record.error && typeof record.error === "object") texts.push(...collectErrorText(record.error, depth + 1));
-	texts.push(...collectErrorText(record.cause, depth + 1));
-	return texts;
-}
-function isRateLimited(error, depth = 0) {
-	if (depth > MAX_CAUSE_DEPTH || error == null || typeof error !== "object") return false;
-	const record = error;
-	return record.statusCode === 429 || isRateLimited(record.cause, depth + 1);
-}
-function isContextOverflowError(error) {
-	if (isRateLimited(error)) return false;
-	return collectErrorText(error).some((text) => CONTEXT_OVERFLOW_PATTERNS.some((pattern) => pattern.test(text)));
-}
 // src/processors/observational-memory/temporal-markers.ts
 const TEMPORAL_GAP_REMINDER_TYPE = "temporal-gap";
 function getTemporalGapReminderText(gapText, timestamp) {
@@ -42529,7 +42506,7 @@ var ObservationalMemoryProcessor = class {
 		if (!this.engine.getObservationConfig().observeOnContextOverflow) return;
 		if (state[GATEWAY_STATE_KEY]) return;
 		if (state[CONTEXT_OVERFLOW_STEP_KEY] === stepNumber) return;
-		if (!isContextOverflowError(error)) return;
+		if (classifyProviderError(error) !== "context_overflow") return;
 		const context = this.engine.getThreadContext(requestContext, messageList);
 		if (!context) return;
 		if (parseMemoryRequestContext(requestContext)?.memoryConfig?.readOnly) return;
@@ -45341,4 +45318,4 @@ Notes:
 };
 export { renderSubconsciousActivity as $, buildObserverSystemPrompt as A, OBSERVATION_CONTEXT_INSTRUCTIONS as B, WorkingMemoryExtractor as C, TokenCounter as D, summarizeConversation as E, parseObserverOutput as F, Subconscious as G, OBSERVATION_CONTEXT_PROMPT_THREAD as H, injectAnchorIds as I, SubconsciousRemindExtractor as J, KnowledgeSemanticIndexCoordinator as K, parseAnchorId as L, formatMessagesForObserver as M, hasCurrentTaskSection as N, OBSERVER_SYSTEM_PROMPT as O, optimizeObservationsForContext as P, publishSubconsciousActivity as Q, stripEphemeralAnchorIds as R, wrapInObservationGroup as S, SUMMARIZE_THREAD_DEFAULTS as T, OBSERVATION_CONTINUATION_HINT as U, OBSERVATION_CONTEXT_PROMPT as V, getObservationContextPrompt as W, SUBCONSCIOUS_ACTIVITY_STATE_ID as X, ModelByInputTokens as Y, buildSubconsciousActivitySnapshot as Z, deriveObservationGroupProvenance as _, extractWorkingMemoryContent as a, renderObservationGroupsForReflection as b, getObservationsAsOf as c, AUTO_MODEL_BY_PROVIDER as d, Extractor as et, resolveAutoModelId as f, combineObservationGroupRanges as g, WorkingMemoryStateProcessor as h, WorkingMemory as i, extractCurrentTask as j, buildObserverPrompt as k, ObservationalMemoryProcessor as l, WORKING_MEMORY_STATE_PROCESSOR_ID as m, MessageHistory$1 as n, extractWorkingMemoryTags as o, WORKING_MEMORY_STATE_ID as p, StaleKnowledgeSemanticIndexError as q, SemanticRecall as r, removeWorkingMemoryTags as s, Memory as t, ObservationalMemory as u, parseObservationGroups as v, deepMergeWorkingMemory as w, stripObservationGroups as x, reconcileObservationGroupsFromReflection as y, OBSERVATIONAL_MEMORY_DEFAULTS as z };
 
-//# sourceMappingURL=src-stSVckwY.js.map
+//# sourceMappingURL=src-kIpEG6XZ.js.map