@mastra/observability 1.20.0-alpha.0 → 1.20.0

package/dist/index.js

@@ -1150,7 +1150,7 @@ const MODELS_DEV_API_URL = "https://models.dev/api.json";
 * not replace the pricing in use. The full catalog prices roughly 8,000 models.
 */
 const MIN_PRICED_MODELS = 1e3;
-/** models.dev `cost` keys (USD per 1M tokens) mapped to pricing meters. */
+/** models.dev `cost` keys (USD per 1M tokens) mapped to pricing meters. Keyed by core's `ModelPricingRates` so the two cannot drift. */
 const COST_KEY_TO_METER = {
 	input: "input_tokens",
 	output: "output_tokens",
@@ -1290,6 +1290,7 @@ const BEDROCK_GEOGRAPHY_PREFIXES = /* @__PURE__ */ new Set([
 	"au"
 ]);
 const AI_SDK_VERCEL_GATEWAY_PROVIDER_ID = "gateway";
+const AZURE_OPENAI_GATEWAY_ID = "azure-openai";
 const VERCEL_PRICING_PROVIDER_ID = "vercel";
 const AI_SDK_PROVIDER_NAMESPACE_ALIASES = /* @__PURE__ */ new Map([
 	["fireworks", "fireworks-ai"],
@@ -1300,6 +1301,11 @@ const AI_SDK_PROVIDER_NAMESPACE_ALIASES = /* @__PURE__ */ new Map([
 let cachedLoadError = null;
 let globalRegistry = null;
 let bundledRegistry = null;
+let gatewayRegistry = null;
+let gatewayCosts = {};
+let gatewaySource = null;
+const gatewayLoadAttempts = /* @__PURE__ */ new Map();
+let gatewayLoadInFlight = false;
 let refreshedAt = 0;
 let lastRefreshAttemptAt = 0;
 let refreshEtag;
@@ -1328,10 +1334,10 @@ var PricingRegistry = class PricingRegistry {
 		return globalRegistry;
 	}
 	get(args) {
-		const match = this.find(args);
-		if (match || this !== globalRegistry) return match;
-		const bundledMatch = this === bundledRegistry ? null : getBundledRegistry()?.find(args);
-		if (bundledMatch) return bundledMatch;
+		if (this !== globalRegistry) return this.find(args);
+		const match = this.find(args) ?? (this === bundledRegistry ? null : getBundledRegistry()?.find(args)) ?? gatewayRegistry?.find(args);
+		if (match) return match;
+		loadGatewayPricing();
 		refreshGlobalPricing();
 		return null;
 	}
@@ -1354,12 +1360,70 @@ function getBundledRegistry() {
 	return bundledRegistry;
 }
 /**
+* Where to find the model router gateways whose `pricing` fills the gaps in models.dev. Prices
+* load on the first lookup that has no price, so production does not call gateways at startup;
+* under the same switches as the models.dev refresh (`mastra dev`, or
+* `MASTRA_AUTO_REFRESH_PRICING=true`) they also load right away, off the hot path, so the first
+* model call is priced. Each gateway is read once; a failed read is retried at most once an hour.
+*/
+function setPricingGateways(source) {
+	gatewaySource = source;
+	if (isRefreshEnabled()) loadGatewayPricing();
+}
+/** Load `pricing` from every custom gateway not loaded yet. Off under `MASTRA_OFFLINE`. */
+async function loadGatewayPricing() {
+	if (!gatewaySource || gatewayLoadInFlight || isOffline()) return;
+	gatewayLoadInFlight = true;
+	try {
+		const builtInIds = await getBuiltInGatewayIds();
+		let changed = false;
+		for (const gateway of gatewaySource()) {
+			const gatewayId = gateway.getId?.() ?? gateway.id;
+			const lastAttempt = gatewayLoadAttempts.get(gateway) ?? 0;
+			if (builtInIds.has(gatewayId) || Date.now() - lastAttempt < MIN_REFRESH_ATTEMPT_INTERVAL_MS) continue;
+			try {
+				const providers = await gateway.fetchProviders();
+				for (const [providerId, config] of Object.entries(providers ?? {})) for (const [modelId, cost] of Object.entries(config?.pricing ?? {})) (gatewayCosts[providerId] ??= { models: {} }).models[modelId] = { cost };
+				gatewayLoadAttempts.set(gateway, Infinity);
+				changed = true;
+			} catch (error) {
+				gatewayLoadAttempts.set(gateway, Date.now());
+				console.warn(`Failed to load model prices from gateway ${gatewayId}:`, error);
+			}
+		}
+		if (changed) gatewayRegistry = new PricingRegistry(modelsDevToPricingModels(gatewayCosts));
+	} finally {
+		gatewayLoadInFlight = false;
+	}
+}
+/**
+* The gateways core ships: the ones Mastra registers on its own, plus Azure OpenAI, which users
+* construct themselves. Their `fetchProviders()` can hit the network and never carries prices.
+* Read from core when it exports the list (1.46+), else the known ids.
+*/
+async function getBuiltInGatewayIds() {
+	try {
+		const { defaultGateways, getGatewayId } = await import("@mastra/core/llm");
+		return /* @__PURE__ */ new Set([...defaultGateways.map(getGatewayId), AZURE_OPENAI_GATEWAY_ID]);
+	} catch {
+		return /* @__PURE__ */ new Set([
+			"models.dev",
+			"netlify",
+			"mastra",
+			AZURE_OPENAI_GATEWAY_ID
+		]);
+	}
+}
+function isOffline() {
+	const offline = process.env.MASTRA_OFFLINE;
+	return offline === "true" || offline === "1";
+}
+/**
 * Same switches as the model router's auto-refresh in `@mastra/core`: on by default only under
 * `mastra dev` (`MASTRA_DEV`), opt-in elsewhere, and always off with `MASTRA_OFFLINE`.
 */
 function isRefreshEnabled() {
-	const offline = process.env.MASTRA_OFFLINE;
-	if (offline === "true" || offline === "1") return false;
+	if (isOffline()) return false;
 	const isDev = process.env.MASTRA_DEV === "true" || process.env.MASTRA_DEV === "1";
 	const setting = process.env.MASTRA_AUTO_REFRESH_PRICING;
 	return setting === "true" || setting !== "false" && isDev;
@@ -4000,6 +4064,8 @@ function resolveTracingStorageStrategy$1(config, observabilityStorage, storageNa
 	return observabilityStrategy.preferred;
 }
 const FLUSH_SUCCEEDED$1 = { failed: false };
+/** Max join/write rounds a single flush() makes before it returns under continuous traffic */
+const MAX_FLUSH_ROUNDS$1 = 3;
 /**
 * @deprecated Use `MastraStorageExporter` from `@mastra/observability` instead.
 * This class is preserved unchanged so existing integrations (including code
@@ -4019,6 +4085,8 @@ var DefaultExporter = class extends BaseExporter {
 	#observabilityStorage;
 	#resolvedStrategy;
 	#flushTimer;
+	/** Storage writes (timer, size trigger, or flush()) that haven't finished yet */
+	#inFlightFlushes = /* @__PURE__ */ new Set();
 	#emitDropEvent;
 	#unsupportedSignals = /* @__PURE__ */ new Set();
 	constructor(config = {}) {
@@ -4093,7 +4161,7 @@ var DefaultExporter = class extends BaseExporter {
 		if (this.#flushTimer) clearTimeout(this.#flushTimer);
 		this.#flushTimer = setTimeout(() => {
 			this.#flushTimer = void 0;
-			this.flushBuffer().catch((error) => {
+			this.trackFlushBuffer().catch((error) => {
 				this.logger.error("Scheduled flush failed", { error: error instanceof Error ? error.message : String(error) });
 			});
 		}, delayMs);
@@ -4123,7 +4191,7 @@ var DefaultExporter = class extends BaseExporter {
 	* Returns the flush promise when flushing so callers can await it.
 	*/
 	async handleBatchedFlush() {
-		if (this.shouldFlush()) await this.flushBuffer();
+		if (this.shouldFlush()) await this.trackFlushBuffer();
 		else if (this.#eventBuffer.totalSize === 1) this.scheduleFlush();
 	}
 	sanitizeDropError(error) {
@@ -4227,6 +4295,15 @@ var DefaultExporter = class extends BaseExporter {
 		}
 	}
 	/**
+	* Runs flushBuffer() and tracks its promise so flush() can wait for writes
+	* that a timer or size trigger started on their own.
+	*/
+	trackFlushBuffer() {
+		const promise = this.flushBuffer().finally(() => this.#inFlightFlushes.delete(promise));
+		this.#inFlightFlushes.add(promise);
+		return promise;
+	}
+	/**
 	* Flushes the current buffer to storage.
 	*
 	* Creates are flushed first, then their span keys are added to allCreatedSpans.
@@ -4385,10 +4462,16 @@ var DefaultExporter = class extends BaseExporter {
 	* are exported before the runtime instance is terminated.
 	*/
 	async flush() {
-		if (this.#eventBuffer.totalSize > 0) {
-			this.logger.debug("Flushing buffered events", { bufferedEvents: this.#eventBuffer.totalSize });
-			await this.flushBuffer();
+		for (let round = 0; round < MAX_FLUSH_ROUNDS$1; round++) {
+			if (this.#inFlightFlushes.size > 0) await Promise.allSettled([...this.#inFlightFlushes]);
+			if (this.#eventBuffer.totalSize > 0) {
+				this.logger.debug("Flushing buffered events", { bufferedEvents: this.#eventBuffer.totalSize });
+				await this.trackFlushBuffer();
+			}
+			if (this.#inFlightFlushes.size === 0) return;
 		}
+		await Promise.allSettled([...this.#inFlightFlushes]);
+		this.logger.warn(`flush() stopped after ${MAX_FLUSH_ROUNDS$1} rounds; new writes kept starting`, { bufferedEventsLeftBehind: this.#eventBuffer.totalSize });
 	}
 	async shutdown() {
 		if (this.#flushTimer) {
@@ -4422,6 +4505,8 @@ function resolveTracingStorageStrategy(config, observabilityStorage, storageName
 	return observabilityStrategy.preferred;
 }
 const FLUSH_SUCCEEDED = { failed: false };
+/** Max join/write rounds a single flush() makes before it returns under continuous traffic */
+const MAX_FLUSH_ROUNDS = 3;
 /**
 * Storage-backed exporter. Buffers observability events and flushes them in
 * batches to the configured ObservabilityStorage backend with retry support.
@@ -4436,6 +4521,8 @@ var MastraStorageExporter = class extends BaseExporter {
 	#observabilityStorage;
 	#resolvedStrategy;
 	#flushTimer;
+	/** Storage writes (timer, size trigger, or flush()) that haven't finished yet */
+	#inFlightFlushes = /* @__PURE__ */ new Set();
 	#emitDropEvent;
 	#unsupportedSignals = /* @__PURE__ */ new Set();
 	constructor(config = {}) {
@@ -4525,7 +4612,7 @@ var MastraStorageExporter = class extends BaseExporter {
 		if (this.#flushTimer) clearTimeout(this.#flushTimer);
 		this.#flushTimer = setTimeout(() => {
 			this.#flushTimer = void 0;
-			this.flushBuffer().catch((error) => {
+			this.trackFlushBuffer().catch((error) => {
 				this.logger.error("Scheduled flush failed", { error: error instanceof Error ? error.message : String(error) });
 			});
 		}, delayMs);
@@ -4555,7 +4642,7 @@ var MastraStorageExporter = class extends BaseExporter {
 	* Returns the flush promise when flushing so callers can await it.
 	*/
 	async handleBatchedFlush() {
-		if (this.shouldFlush()) await this.flushBuffer();
+		if (this.shouldFlush()) await this.trackFlushBuffer();
 		else if (this.#eventBuffer.totalSize === 1) this.scheduleFlush();
 	}
 	sanitizeDropError(error) {
@@ -4659,6 +4746,15 @@ var MastraStorageExporter = class extends BaseExporter {
 		}
 	}
 	/**
+	* Runs flushBuffer() and tracks its promise so flush() can wait for writes
+	* that a timer or size trigger started on their own.
+	*/
+	trackFlushBuffer() {
+		const promise = this.flushBuffer().finally(() => this.#inFlightFlushes.delete(promise));
+		this.#inFlightFlushes.add(promise);
+		return promise;
+	}
+	/**
 	* Flushes the current buffer to storage.
 	*
 	* Creates are flushed first, then their span keys are added to allCreatedSpans.
@@ -4806,10 +4902,16 @@ var MastraStorageExporter = class extends BaseExporter {
 	* are exported before the runtime instance is terminated.
 	*/
 	async flush() {
-		if (this.#eventBuffer.totalSize > 0) {
-			this.logger.debug("Flushing buffered events", { bufferedEvents: this.#eventBuffer.totalSize });
-			await this.flushBuffer();
+		for (let round = 0; round < MAX_FLUSH_ROUNDS; round++) {
+			if (this.#inFlightFlushes.size > 0) await Promise.allSettled([...this.#inFlightFlushes]);
+			if (this.#eventBuffer.totalSize > 0) {
+				this.logger.debug("Flushing buffered events", { bufferedEvents: this.#eventBuffer.totalSize });
+				await this.trackFlushBuffer();
+			}
+			if (this.#inFlightFlushes.size === 0) return;
 		}
+		await Promise.allSettled([...this.#inFlightFlushes]);
+		this.logger.warn(`flush() stopped after ${MAX_FLUSH_ROUNDS} rounds; new writes kept starting`, { bufferedEventsLeftBehind: this.#eventBuffer.totalSize });
 	}
 	async shutdown() {
 		if (this.#flushTimer) {
@@ -9931,6 +10033,7 @@ var Observability = class extends MastraBase {
 		const instances = this.listInstances();
 		const { mastra } = options;
 		this.#mastra = mastra;
+		setPricingGateways(() => Object.values(mastra.listGateways?.() ?? {}));
 		const mastraEnvironment = mastra.getEnvironment?.();
 		instances.forEach((instance) => {
 			instance.__setMastraEnvironment?.(mastraEnvironment);