@mastra/memory 1.28.0-alpha.2 → 1.28.0-alpha.4

package/dist/{src-CIBCjg2A.cjs → src-C7Ozl6Pf.cjs}

@@ -57,8 +57,8 @@ probe_image_size_sync_js = __toESM$1(probe_image_size_sync_js, 1);
 let _mastra_core_schema = require("@mastra/core/schema");
 let _mastra_core_error = require("@mastra/core/error");
 let _mastra_core_processors = require("@mastra/core/processors");
-let util = require("util");
 let diff = require("diff");
+let util = require("util");
 //#region ../_vendored/ai_v4/dist/dist-Cm5DZgxA.js
 var marker$6 = "vercel.ai.error";
 var symbol$6 = Symbol.for(marker$6);
@@ -3734,7 +3734,7 @@ function trimStartOfStream() {
 	};
 }
 //#endregion
-//#region ../_vendored/ai_v5/dist/dist-BuEMdYEn.js
+//#region ../_vendored/ai_v5/dist/dist-nVfjvSbw.js
 var marker$1$4 = "vercel.ai.error";
 var symbol$1$3 = Symbol.for(marker$1$4);
 var _a$1$3;
@@ -4287,6 +4287,7 @@ async function cancelResponseBody$1(response) {
 var initialGlobalFetch$1 = globalThis.fetch;
 isNodeDefaultFetch$1(initialGlobalFetch$1);
 function isNodeDefaultFetch$1(fetch) {
+	if (typeof fetch !== "function") return false;
 	const source = Function.prototype.toString.call(fetch);
 	return source.includes("internal/deps/undici") || source.includes("lazy loading of undici");
 }
@@ -8622,7 +8623,7 @@ var object$1 = ({ schema: inputSchema }) => {
 	};
 };
 //#endregion
-//#region ../_vendored/ai_v6/dist/dist-BTWHzT8H.js
+//#region ../_vendored/ai_v6/dist/dist-Gd4EIxPM.js
 var marker$1 = "vercel.ai.error";
 var symbol$1$1 = Symbol.for(marker$1);
 var _a$1$1;
@@ -9196,6 +9197,7 @@ var DownloadError = class extends (_b$2 = AISDKError, _a$2 = symbol$2, _b$2) {
 var initialGlobalFetch = globalThis.fetch;
 isNodeDefaultFetch(initialGlobalFetch);
 function isNodeDefaultFetch(fetch) {
+	if (typeof fetch !== "function") return false;
 	const source = Function.prototype.toString.call(fetch);
 	return source.includes("internal/deps/undici") || source.includes("lazy loading of undici");
 }
@@ -10774,7 +10776,7 @@ function withoutTrailingSlash(url) {
 	return url == null ? void 0 : url.replace(/\/$/, "");
 }
 //#endregion
-//#region ../_vendored/ai_v6/dist/dist-3jq9sPhA.js
+//#region ../_vendored/ai_v6/dist/dist-BoeKQsDe.js
 function getContext() {
 	return { headers: {} };
 }
@@ -22434,6 +22436,151 @@ function createWorkingMemoryTool(config, options = {}) {
 	};
 }
 //#endregion
+//#region src/processors/working-memory-state/processor.ts
+/**
+* WorkingMemoryStateProcessor
+*
+* Experimental: delivers working memory to the model as a state signal instead
+* of folding it into the system message. Storage and the `setWorkingMemory`
+* tool are unchanged — this processor only changes the delivery path.
+*
+* Pattern matches `BrowserContextProcessor` in `@mastra/core/browser`:
+* - `stateId` namespaces the state lane on the thread.
+* - `cacheKey` is derived from the rendered payload so dedup is automatic.
+* - `contextWindow.hasSnapshot` re-injection ensures the model still sees the
+*   current snapshot after older messages drop out of the window.
+*
+* Delta emission (markdown mode only): when a prior snapshot exists in the
+* context window, the processor emits a unified-diff delta against that
+* snapshot's contents. Schema mode and the snapshot fallback always emit a
+* full snapshot.
+*
+* @example
+* ```ts
+* new Memory({
+*   options: {
+*     workingMemory: {
+*       enabled: true,
+*       template: '...',
+*       useStateSignals: true, // auto-attaches this processor
+*     },
+*   },
+* });
+* ```
+*/
+const WORKING_MEMORY_STATE_ID = "working-memory";
+const WORKING_MEMORY_STATE_PROCESSOR_ID = "working-memory-state";
+var WorkingMemoryStateProcessor = class {
+	memory;
+	memoryConfig;
+	id = WORKING_MEMORY_STATE_PROCESSOR_ID;
+	stateId = WORKING_MEMORY_STATE_ID;
+	constructor(memory, memoryConfig) {
+		this.memory = memory;
+		this.memoryConfig = memoryConfig;
+	}
+	async computeStateSignal(args) {
+		const template = await this.memory.getWorkingMemoryTemplate({ memoryConfig: this.memoryConfig });
+		if (!template) return;
+		const contents = (await this.memory.getWorkingMemory({
+			threadId: args.threadId,
+			resourceId: args.resourceId,
+			memoryConfig: this.memoryConfig
+		}))?.trim();
+		if (!contents) return;
+		const cacheKey = stableWorkingMemoryCacheKey({
+			format: template.format,
+			data: contents
+		});
+		const shouldMakeSnapshot = !args.contextWindow.hasSnapshot;
+		if (args.tracking?.currentCacheKey === cacheKey && !shouldMakeSnapshot) return;
+		const scope = this.memory.getMergedThreadConfig(this.memoryConfig).workingMemory?.scope ?? "resource";
+		const deltaCandidate = template.format === "markdown" && !shouldMakeSnapshot ? buildMarkdownDelta({
+			lastSnapshot: args.lastSnapshot,
+			deltasSinceSnapshot: args.deltasSinceSnapshot,
+			nextContents: contents
+		}) : void 0;
+		if (deltaCandidate) return {
+			id: WORKING_MEMORY_STATE_ID,
+			mode: "delta",
+			cacheKey,
+			tagName: "working-memory",
+			contents: deltaCandidate.contents,
+			delta: deltaCandidate.contents,
+			value: contents,
+			attributes: {
+				format: template.format,
+				scope,
+				patch: "unified-diff"
+			}
+		};
+		return {
+			id: WORKING_MEMORY_STATE_ID,
+			mode: "snapshot",
+			cacheKey,
+			tagName: "working-memory",
+			contents,
+			value: contents,
+			attributes: {
+				format: template.format,
+				scope
+			}
+		};
+	}
+};
+/**
+* Stable cache key for the rendered working memory payload. Returns a SHA-256
+* digest so dedup metadata stays compact regardless of payload size (working
+* memory blobs can grow arbitrarily long).
+*/
+function stableWorkingMemoryCacheKey(input) {
+	const hash = (0, crypto$1.createHash)("sha256");
+	hash.update(input.format);
+	hash.update("\0");
+	hash.update(input.data ?? "");
+	return `sha256:${hash.digest("hex")}`;
+}
+/**
+* Build a unified-diff delta against the most recently emitted state. Prefers
+* the latest delta's `value` (the post-edit full text) when available, falling
+* back to the snapshot's `value` and finally the snapshot's `contents`. This
+* keeps deltas incremental (B→C) instead of cumulative against a stale
+* snapshot (A→C), which matters when many small edits land between snapshots.
+*
+* Returns undefined when:
+* - there's no prior state to diff against
+* - the prior state isn't a plain string (multimodal signal)
+*
+* In either case the caller falls back to emitting a full snapshot.
+*/
+function buildMarkdownDelta(args) {
+	const { lastSnapshot, deltasSinceSnapshot, nextContents } = args;
+	const prior = pickStringValue(readSignalValue(deltasSinceSnapshot.at(-1))) ?? pickStringValue(readSignalValue(lastSnapshot)) ?? (typeof lastSnapshot?.contents === "string" ? lastSnapshot.contents : void 0);
+	if (!prior) return;
+	return { contents: renderHunksOnly(prior, nextContents) };
+}
+function pickStringValue(value) {
+	return typeof value === "string" ? value : void 0;
+}
+function readSignalValue(signal) {
+	return (signal?.metadata)?.value;
+}
+/**
+* Render a unified-diff-style patch body containing only `@@` hunks and their
+* lines — dropping the filename preamble (`Index:` / `===` / `---` / `+++`)
+* that `createPatch` emits and the `` trailer.
+* The preamble exists for tooling like `patch -p1` to know which file to
+* apply to; we only ever diff a single working-memory blob. The newline
+* trailer is semantically meaningless to the model and adds noise to the
+* state signal.
+*/
+function renderHunksOnly(prior, next) {
+	const { hunks } = (0, diff.structuredPatch)("", "", prior, next, "", "", { context: 0 });
+	return hunks.map((hunk) => {
+		return [`@@ -${hunk.oldStart},${hunk.oldLines} +${hunk.newStart},${hunk.newLines} @@`, ...hunk.lines.filter((line) => !line.startsWith("\"))].join("\n");
+	}).join("\n");
+}
+//#endregion
 //#region src/processors/observational-memory/activation-ttl.ts
 const MINUTE = 6e4;
 const HOUR = 60 * MINUTE;
@@ -25412,6 +25559,21 @@ var ReflectorRunner = class {
 	memory;
 	onReflectionCommitted;
 	mastra;
+	/**
+	* Per-lock-key observation-token count of the last unproductive synchronous
+	* reflection (failed, or committed while still over the reflection
+	* threshold). `call()` already runs the full bounded compression ladder for
+	* a given input, so re-running it against unchanged observations cannot
+	* succeed — threshold-triggered sync reflection is skipped while the count
+	* is unchanged and retried as soon as the input differs.
+	* Keyed by lock key (thread/resource), which is stable across generations —
+	* record ids change every time a generation is created.
+	*
+	* Lifecycle: an entry is deleted when a sync reflection finishes under
+	* threshold and replaced on each unproductive attempt. Entries are one
+	* number per active thread/resource key and live for the runner's lifetime.
+	*/
+	syncReflectionSuppression = /* @__PURE__ */ new Map();
 	constructor(opts) {
 		this.reflectionConfig = opts.reflectionConfig;
 		this.observationConfig = opts.observationConfig;
@@ -25494,6 +25656,8 @@ var ReflectorRunner = class {
 		};
 		let reflectedTokens = 0;
 		let attemptNumber = 0;
+		/** True when the latest attempt returned an empty block for non-empty input. */
+		let emptyOutput = false;
 		while (currentLevel <= maxLevel) {
 			attemptNumber++;
 			const isRetry = attemptNumber > 1;
@@ -25557,14 +25721,18 @@ var ReflectorRunner = class {
 			}
 			finalProviderMetadata = result.providerMetadata ?? finalProviderMetadata;
 			parsed = parseReflectorOutput(result.text, observations, activeExtractors);
+			emptyOutput = !parsed.degenerate && observations.trim().length > 0 && parsed.observations.trim().length === 0;
 			if (parsed.degenerate) {
 				omDebug(`[OM:callReflector] attempt #${attemptNumber}: degenerate repetition detected, treating as compression failure. ${describeDegenerateOutput(result.text, 2e3)}`);
 				reflectedTokens = originalTokens;
+			} else if (emptyOutput) {
+				omDebug(`[OM:callReflector] attempt #${attemptNumber}: empty observations block for non-empty input, treating as compression failure`);
+				reflectedTokens = originalTokens;
 			} else reflectedTokens = this.tokenCounter.countObservations(parsed.observations);
 			omDebug(`[OM:callReflector] attempt #${attemptNumber} parsed: reflectedTokens=${reflectedTokens}, targetThreshold=${targetThreshold}, compressionValid=${validateCompression(reflectedTokens, targetThreshold)}, parsedObsLen=${parsed.observations?.length}, degenerate=${parsed.degenerate ?? false}`);
-			if (!parsed.degenerate && (validateCompression(reflectedTokens, targetThreshold) || currentLevel >= maxLevel)) break;
-			if (parsed.degenerate && currentLevel >= maxLevel) {
-				omDebug(`[OM:callReflector] degenerate output persists at maxLevel=${maxLevel}, breaking`);
+			if (!parsed.degenerate && !emptyOutput && (validateCompression(reflectedTokens, targetThreshold) || currentLevel >= maxLevel)) break;
+			if ((parsed.degenerate || emptyOutput) && currentLevel >= maxLevel) {
+				omDebug(`[OM:callReflector] ${parsed.degenerate ? "degenerate" : "empty"} output persists at maxLevel=${maxLevel}, breaking`);
 				break;
 			}
 			if (streamContext?.writer) {
@@ -25602,6 +25770,7 @@ var ReflectorRunner = class {
 			}
 			currentLevel = Math.min(currentLevel + 1, maxLevel);
 		}
+		if (observations.trim().length > 0 && parsed.observations.trim().length === 0) throw new Error(`Reflector produced empty output after ${attemptNumber} attempt(s)${parsed.degenerate ? " (degenerate repetition)" : ""} — refusing to commit an empty reflection over ${originalTokens} observation tokens`);
 		const structuredExtraction = await extractStructuredValues({
 			agent,
 			source: "reflector",
@@ -25955,6 +26124,13 @@ var ReflectorRunner = class {
 				return;
 			}
 		}
+		if (activationTriggeredBy === "threshold") {
+			const suppressedAtTokens = this.syncReflectionSuppression.get(lockKey);
+			if (suppressedAtTokens !== void 0 && observationTokens === suppressedAtTokens) {
+				omDebug(`[OM:reflect] skipping sync reflection — observations unchanged at ${observationTokens} tokens since the last unproductive attempt; retrying once they change`);
+				return;
+			}
+		}
 		await this.storage.setReflectingFlag(record.id, true);
 		registerOp(record.id, "reflecting");
 		const cycleId = crypto.randomUUID();
@@ -26015,6 +26191,8 @@ var ReflectorRunner = class {
 				reflection: reflectResult.observations,
 				tokenCount: reflectionTokenCount
 			});
+			if (reflectionTokenCount >= reflectThreshold) this.syncReflectionSuppression.set(lockKey, reflectionTokenCount);
+			else this.syncReflectionSuppression.delete(lockKey);
 			await this.notifyReflectionCommitted({
 				parentThreadId: requestedThreadId ?? record.threadId ?? "",
 				resourceId: record.resourceId ?? "",
@@ -26074,6 +26252,7 @@ var ReflectorRunner = class {
 			}
 			reflectionError = error instanceof Error ? error : new Error(String(error));
 			if (lifecycleError !== void 0 || abortSignal?.aborted) throw error;
+			this.syncReflectionSuppression.set(lockKey, observationTokens);
 			omError("[OM] Reflection failed", error);
 		} finally {
 			try {
@@ -26420,10 +26599,24 @@ function getLatestStepParts(parts) {
 	return parts;
 }
 /**
-* Returns true when a message contains at least one part with visible user/assistant
-* content (text, tool-invocation, reasoning, image, file).  Messages that only carry
-* internal `data-*` parts (buffering markers, observation markers, etc.) return false.
+* Returns true for persisted Working Memory state signals (`role: 'signal'`,
+* `signal.type: 'state'`, `state.id: 'working-memory'`). OM must not observe these:
+* they mirror memory OM itself manages, and re-observing them creates a self-feedback
+* loop that can overwrite stored working memory (#21961). Only the working-memory lane
+* is filtered — other state lanes are outside OM's write path, so re-observing them
+* cannot corrupt OM-managed state; excluding them broadly is a separate decision.
 */
+function isWorkingMemoryStateSignal(message) {
+	if (message.role !== "signal") return false;
+	const signal = message.content.metadata?.signal;
+	if (!signal || typeof signal !== "object" || Array.isArray(signal)) return false;
+	const signalRecord = signal;
+	if ((signalRecord.type ?? message.type) !== "state") return false;
+	const metadata = signalRecord.metadata;
+	if (!metadata || typeof metadata !== "object" || Array.isArray(metadata)) return false;
+	const state = metadata.state;
+	return !!state && typeof state === "object" && !Array.isArray(state) && state.id === "working-memory";
+}
 function messageHasVisibleContent(msg) {
 	const content = msg.content;
 	if (content?.parts && Array.isArray(content.parts)) return content.parts.some((p) => {
@@ -27194,7 +27387,7 @@ var ObservationalMemory = class ObservationalMemory {
 		}
 		const result = [];
 		for (const msg of allMessages) {
-			if (msg.role === "system") continue;
+			if (msg.role === "system" || isWorkingMemoryStateSignal(msg)) continue;
 			if (observedMessageIds?.has(msg.id)) continue;
 			const endMarkerIndex = findLastCompletedObservationBoundary(msg);
 			if (this.hasInProgressObservation(msg)) result.push(msg);
@@ -27416,7 +27609,7 @@ var ObservationalMemory = class ObservationalMemory {
 			},
 			filter: startDate ? { dateRange: { start: startDate } } : void 0
 		});
-		return result.messages.filter((msg) => msg.role !== "system");
+		return result.messages.filter((msg) => msg.role !== "system" && !isWorkingMemoryStateSignal(msg));
 	}
 	/**
 	* Format unobserved messages from other threads as <unobserved-context> blocks.
@@ -27875,7 +28068,7 @@ ${formattedMessages}
 					direction: "ASC"
 				},
 				filter: startDate ? { dateRange: { start: startDate } } : void 0
-			})).messages.filter((m) => !this.observedMessageIds.has(m.id));
+			})).messages.filter((message) => !this.observedMessageIds.has(message.id) && !isWorkingMemoryStateSignal(message));
 			if (filtered.length > 0) messagesByThread.set(thread.id, filtered);
 		}
 		if (messagesByThread.size === 0) return void 0;
@@ -29301,151 +29494,6 @@ var observational_memory_exports = /* @__PURE__ */ __exportAll({
 	wrapInObservationGroup: () => wrapInObservationGroup
 });
 //#endregion
-//#region src/processors/working-memory-state/processor.ts
-/**
-* WorkingMemoryStateProcessor
-*
-* Experimental: delivers working memory to the model as a state signal instead
-* of folding it into the system message. Storage and the `setWorkingMemory`
-* tool are unchanged — this processor only changes the delivery path.
-*
-* Pattern matches `BrowserContextProcessor` in `@mastra/core/browser`:
-* - `stateId` namespaces the state lane on the thread.
-* - `cacheKey` is derived from the rendered payload so dedup is automatic.
-* - `contextWindow.hasSnapshot` re-injection ensures the model still sees the
-*   current snapshot after older messages drop out of the window.
-*
-* Delta emission (markdown mode only): when a prior snapshot exists in the
-* context window, the processor emits a unified-diff delta against that
-* snapshot's contents. Schema mode and the snapshot fallback always emit a
-* full snapshot.
-*
-* @example
-* ```ts
-* new Memory({
-*   options: {
-*     workingMemory: {
-*       enabled: true,
-*       template: '...',
-*       useStateSignals: true, // auto-attaches this processor
-*     },
-*   },
-* });
-* ```
-*/
-const WORKING_MEMORY_STATE_ID = "working-memory";
-const WORKING_MEMORY_STATE_PROCESSOR_ID = "working-memory-state";
-var WorkingMemoryStateProcessor = class {
-	memory;
-	memoryConfig;
-	id = WORKING_MEMORY_STATE_PROCESSOR_ID;
-	stateId = WORKING_MEMORY_STATE_ID;
-	constructor(memory, memoryConfig) {
-		this.memory = memory;
-		this.memoryConfig = memoryConfig;
-	}
-	async computeStateSignal(args) {
-		const template = await this.memory.getWorkingMemoryTemplate({ memoryConfig: this.memoryConfig });
-		if (!template) return;
-		const contents = (await this.memory.getWorkingMemory({
-			threadId: args.threadId,
-			resourceId: args.resourceId,
-			memoryConfig: this.memoryConfig
-		}))?.trim();
-		if (!contents) return;
-		const cacheKey = stableWorkingMemoryCacheKey({
-			format: template.format,
-			data: contents
-		});
-		const shouldMakeSnapshot = !args.contextWindow.hasSnapshot;
-		if (args.tracking?.currentCacheKey === cacheKey && !shouldMakeSnapshot) return;
-		const scope = this.memory.getMergedThreadConfig(this.memoryConfig).workingMemory?.scope ?? "resource";
-		const deltaCandidate = template.format === "markdown" && !shouldMakeSnapshot ? buildMarkdownDelta({
-			lastSnapshot: args.lastSnapshot,
-			deltasSinceSnapshot: args.deltasSinceSnapshot,
-			nextContents: contents
-		}) : void 0;
-		if (deltaCandidate) return {
-			id: WORKING_MEMORY_STATE_ID,
-			mode: "delta",
-			cacheKey,
-			tagName: "working-memory",
-			contents: deltaCandidate.contents,
-			delta: deltaCandidate.contents,
-			value: contents,
-			attributes: {
-				format: template.format,
-				scope,
-				patch: "unified-diff"
-			}
-		};
-		return {
-			id: WORKING_MEMORY_STATE_ID,
-			mode: "snapshot",
-			cacheKey,
-			tagName: "working-memory",
-			contents,
-			value: contents,
-			attributes: {
-				format: template.format,
-				scope
-			}
-		};
-	}
-};
-/**
-* Stable cache key for the rendered working memory payload. Returns a SHA-256
-* digest so dedup metadata stays compact regardless of payload size (working
-* memory blobs can grow arbitrarily long).
-*/
-function stableWorkingMemoryCacheKey(input) {
-	const hash = (0, crypto$1.createHash)("sha256");
-	hash.update(input.format);
-	hash.update("\0");
-	hash.update(input.data ?? "");
-	return `sha256:${hash.digest("hex")}`;
-}
-/**
-* Build a unified-diff delta against the most recently emitted state. Prefers
-* the latest delta's `value` (the post-edit full text) when available, falling
-* back to the snapshot's `value` and finally the snapshot's `contents`. This
-* keeps deltas incremental (B→C) instead of cumulative against a stale
-* snapshot (A→C), which matters when many small edits land between snapshots.
-*
-* Returns undefined when:
-* - there's no prior state to diff against
-* - the prior state isn't a plain string (multimodal signal)
-*
-* In either case the caller falls back to emitting a full snapshot.
-*/
-function buildMarkdownDelta(args) {
-	const { lastSnapshot, deltasSinceSnapshot, nextContents } = args;
-	const prior = pickStringValue(readSignalValue(deltasSinceSnapshot.at(-1))) ?? pickStringValue(readSignalValue(lastSnapshot)) ?? (typeof lastSnapshot?.contents === "string" ? lastSnapshot.contents : void 0);
-	if (!prior) return;
-	return { contents: renderHunksOnly(prior, nextContents) };
-}
-function pickStringValue(value) {
-	return typeof value === "string" ? value : void 0;
-}
-function readSignalValue(signal) {
-	return (signal?.metadata)?.value;
-}
-/**
-* Render a unified-diff-style patch body containing only `@@` hunks and their
-* lines — dropping the filename preamble (`Index:` / `===` / `---` / `+++`)
-* that `createPatch` emits and the `` trailer.
-* The preamble exists for tooling like `patch -p1` to know which file to
-* apply to; we only ever diff a single working-memory blob. The newline
-* trailer is semantically meaningless to the model and adds noise to the
-* state signal.
-*/
-function renderHunksOnly(prior, next) {
-	const { hunks } = (0, diff.structuredPatch)("", "", prior, next, "", "", { context: 0 });
-	return hunks.map((hunk) => {
-		return [`@@ -${hunk.oldStart},${hunk.oldLines} +${hunk.newStart},${hunk.newLines} @@`, ...hunk.lines.filter((line) => !line.startsWith("\"))].join("\n");
-	}).join("\n");
-}
-//#endregion
 //#region src/processors/working-memory-state/index.ts
 var working_memory_state_exports = /* @__PURE__ */ __exportAll({
 	WORKING_MEMORY_STATE_ID: () => WORKING_MEMORY_STATE_ID,
@@ -30542,6 +30590,7 @@ ${workingMemory}`;
 				instruction: omConfig.observation.instruction,
 				threadTitle: omConfig.observation.threadTitle,
 				observeAttachments: omConfig.observation.observeAttachments,
+				continuationHints: omConfig.observation.continuationHints,
 				extract: omConfig.observation.extract
 			} : void 0,
 			reflection: omConfig.reflection ? {
@@ -30552,6 +30601,7 @@ ${workingMemory}`;
 				bufferActivation: omConfig.reflection.bufferActivation,
 				blockAfter: omConfig.reflection.blockAfter,
 				instruction: omConfig.reflection.instruction,
+				continuationHints: omConfig.reflection.continuationHints,
 				extract: omConfig.reflection.extract
 			} : void 0
 		});
@@ -31759,4 +31809,4 @@ Object.defineProperty(exports, "wrapInObservationGroup", {
 	}
 });
 
-//# sourceMappingURL=src-CIBCjg2A.cjs.map
+//# sourceMappingURL=src-C7Ozl6Pf.cjs.map