@schlenkr/ragents 0.1.18 → 0.1.19

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@schlenkr/ragents",
-  "version": "0.1.18",
+  "version": "0.1.19",
   "description": "RAgents host: server, engine, plugins and tools without a source checkout",
   "keywords": [
     "agents",
@@ -71,14 +71,14 @@
     "yaml": "2.9.0"
   },
   "optionalDependencies": {
-    "@schlenkr/ragents-tools-win32-x64": "0.1.18",
-    "@schlenkr/ragents-tools-win32-arm64": "0.1.18",
-    "@schlenkr/ragents-tools-darwin-x64": "0.1.18",
-    "@schlenkr/ragents-tools-darwin-arm64": "0.1.18",
-    "@schlenkr/ragents-tools-linux-x64": "0.1.18",
-    "@schlenkr/ragents-tools-linux-arm64": "0.1.18"
+    "@schlenkr/ragents-tools-win32-x64": "0.1.19",
+    "@schlenkr/ragents-tools-win32-arm64": "0.1.19",
+    "@schlenkr/ragents-tools-darwin-x64": "0.1.19",
+    "@schlenkr/ragents-tools-darwin-arm64": "0.1.19",
+    "@schlenkr/ragents-tools-linux-x64": "0.1.19",
+    "@schlenkr/ragents-tools-linux-arm64": "0.1.19"
   },
   "ragents": {
-    "hostVersion": "6f896323c8bdb2622d4dc9459a901949bf5fae33"
+    "hostVersion": "9fd1390d0bf180bd36be46f0409d318a223e6ef2"
   }
 }

package/packages/agent/src/core/agent-settings.ts

@@ -2,7 +2,7 @@
 export interface AgentSettings {
 	/** Retries after a retryable model error, with exponential backoff. */
 	retry: { enabled: boolean; maxRetries: number; baseDelayMs: number };
-	/** Timeout and retries of a single provider request; unset values leave the SDK defaults. */
+	/** Idle timeout (time without any received chunk) and retries of a single provider request; unset values leave the SDK defaults. */
 	providerRequest: { timeoutMs: number; maxRetries?: number; maxRetryDelayMs: number };
 }
 
@@ -13,5 +13,5 @@ export interface AgentSettingsInput {
 
 export const agentSettings = (input: AgentSettingsInput = {}): AgentSettings => ({
 	retry: { enabled: true, maxRetries: 3, baseDelayMs: 2000, ...input.retry },
-	providerRequest: { timeoutMs: 300_000, maxRetryDelayMs: 60_000, ...input.providerRequest },
+	providerRequest: { timeoutMs: 600_000, maxRetryDelayMs: 60_000, ...input.providerRequest },
 });

package/packages/ai/src/api/ai-sdk.ts

@@ -33,6 +33,16 @@ async function runStream(model: Model<"openai-completions">, context: Context, o
 	};
 	const controller = new AbortController();
 	const signal = options?.signal ? AbortSignal.any([options.signal, controller.signal]) : controller.signal;
+	let idleTimer: ReturnType<typeof setTimeout> | undefined;
+	let idleExpired = false;
+	const restartIdleTimer = () => {
+		clearTimeout(idleTimer);
+		if (options?.timeoutMs === undefined) return;
+		idleTimer = setTimeout(() => {
+			idleExpired = true;
+			controller.abort();
+		}, options.timeoutMs);
+	};
 	try {
 		const retention = options?.cacheRetention ?? (getProviderEnvValue("AGENT_CACHE_RETENTION", options?.env) === "long" ? "long" : "short");
 		const cacheControl = retention !== "none" && (model.compat?.cacheControlFormat === "anthropic" || model.id.startsWith("anthropic/"))
@@ -67,7 +77,6 @@ async function runStream(model: Model<"openai-completions">, context: Context, o
 			maxRetries: options?.maxRetries ?? 0,
 			streamRetries: 0,
 			abortSignal: signal,
-			timeout: options?.timeoutMs,
 			stopWhen: stepCountIs(1),
 			includeRawChunks: true,
 			onError: () => {},
@@ -83,7 +92,9 @@ async function runStream(model: Model<"openai-completions">, context: Context, o
 		let rawThinkingIndex: number | undefined;
 		let hasFinishReason = false;
 		let finished = false;
+		restartIdleTimer();
 		for await (const part of result.fullStream) {
+			restartIdleTimer();
 			switch (part.type) {
 				case "raw": {
 					const chunk = part.rawValue as { id?: string; model?: string; choices?: Array<{ finish_reason?: string | null; delta?: {
@@ -216,9 +227,11 @@ async function runStream(model: Model<"openai-completions">, context: Context, o
 	} catch (error) {
 		controller.abort();
 		output.stopReason = options?.signal?.aborted ? "aborted" : "error";
-		output.errorMessage = formatProviderError(normalizeProviderError(error), model.provider === "relay" ? `Relay ${model.baseUrl}` : undefined);
+		const failure = idleExpired ? new Error(`Timeout: no data from the provider for ${options?.timeoutMs} ms`) : error;
+		output.errorMessage = formatProviderError(normalizeProviderError(failure), model.provider === "relay" ? `Relay ${model.baseUrl}` : undefined);
 		events.push({ type: "error", reason: output.stopReason, error: output });
 	} finally {
+		clearTimeout(idleTimer);
 		events.end();
 	}
 }

package/packages/ai/src/types.ts

@@ -89,7 +89,8 @@ export interface StreamOptions {
 	 */
 	headers?: ProviderHeaders;
 	/**
-	 * HTTP request timeout in milliseconds for providers/SDKs that support it.
+	 * Timeout in milliseconds for providers/SDKs that support it. The openai-completions
+	 * stream counts it as idle time: any received chunk restarts it, so an active stream never expires.
 	 * For example, OpenAI and Anthropic SDK clients default to 10 minutes.
 	 */
 	timeoutMs?: number;