fix(ai-gateway): keep long response streams alive

services/ai-gateway/src/config.ts

@@ -20,6 +20,9 @@ const RawConfig = Config.all({
   port: Config.int("AI_GATEWAY_LISTEN_PORT").pipe(
     Config.withDefault(8787),
   ),
+  idleTimeoutSeconds: Config.int("AI_GATEWAY_IDLE_TIMEOUT_SECONDS").pipe(
+    Config.withDefault(255),
+  ),
   gracefulShutdownMillis: Config.int(
     "AI_GATEWAY_GRACEFUL_SHUTDOWN_MILLIS",
   ).pipe(Config.withDefault(20_000)),
@@ -84,6 +87,7 @@ export interface AIGatewayConfig {
   readonly stateDirectory: string;
   readonly hostname: string;
   readonly port: number;
+  readonly idleTimeoutSeconds: number;
   readonly gracefulShutdownMillis: number;
   readonly clientAuthenticationMode: "shared_token" | "workload_identity";
   readonly clientToken: Redacted.Redacted<string>;
@@ -134,6 +138,7 @@ export const loadAIGatewayConfig = Effect.fn(
     stateDirectory,
     hostname: raw.hostname,
     port: raw.port,
+    idleTimeoutSeconds: raw.idleTimeoutSeconds,
     gracefulShutdownMillis: raw.gracefulShutdownMillis,
     clientAuthenticationMode,
     clientToken: raw.clientToken,
@@ -194,6 +199,7 @@ function validConfiguration(
     validSecret(operatorToken) &&
     validSecret(openAIApiKey) &&
     validPositiveInteger(raw.port, 65_535) &&
+    validNonNegativeInteger(raw.idleTimeoutSeconds, 255) &&
     validPositiveInteger(raw.gracefulShutdownMillis) &&
     validPositiveInteger(raw.heartbeatMillis) &&
     validPositiveInteger(raw.maximumUsageEventBytes) &&
@@ -215,6 +221,10 @@ function validPositiveInteger(value: number, maximum?: number): boolean {
     (maximum === undefined || value <= maximum);
 }
 
+function validNonNegativeInteger(value: number, maximum: number): boolean {
+  return Number.isSafeInteger(value) && value >= 0 && value <= maximum;
+}
+
 function validString(value: string, maximum: number): boolean {
   return value.length > 0 && value.length <= maximum && value === value.trim();
 }

services/ai-gateway/src/main.ts

@@ -248,6 +248,7 @@ function makeAIGatewayRuntimeLive(
             Layer.provide(BunHttpServer.layer({
               hostname: serveConfig.hostname,
               port: serveConfig.port,
+              idleTimeout: serveConfig.idleTimeoutSeconds,
               gracefulShutdownTimeout: serveConfig.gracefulShutdownMillis,
             })),
           );

services/ai-gateway/tests/cli.effect.test.ts

@@ -39,6 +39,7 @@ function configuration(
     stateDirectory: "/state",
     hostname: "0.0.0.0",
     port: 8787,
+    idleTimeoutSeconds: 255,
     gracefulShutdownMillis: 20_000,
     clientAuthenticationMode: "shared_token",
     clientToken: Redacted.make("fleet-secret"),
@@ -71,6 +72,7 @@ const makeHarness = Effect.fn("test.aiGateway.makeCliHarness")(function*() {
   const served = yield* Ref.make<ReadonlyArray<{
     readonly hostname: string;
     readonly port: number;
+    readonly idleTimeoutSeconds: number;
     readonly authentication: string;
   }>>([]);
   const statusTokens = yield* Ref.make<ReadonlyArray<string>>([]);
@@ -117,6 +119,7 @@ const makeHarness = Effect.fn("test.aiGateway.makeCliHarness")(function*() {
           {
             hostname: config.hostname,
             port: config.port,
+            idleTimeoutSeconds: config.idleTimeoutSeconds,
             authentication: config.authentication.kind,
           },
         ]),
@@ -172,17 +175,22 @@ describe("AI Gateway Effect CLI", () => {
       ]);
     }));
 
-  it.effect("passes the explicit native bind address to the Effect runtime", () =>
+  it.effect("passes HTTP server settings to the Effect runtime", () =>
     Effect.gen(function*() {
       const harness = yield* makeHarness();
       const exitCode = yield* runAIGatewayCli(
         ["serve"],
-        configuration({ hostname: "127.0.0.2", port: 9876 }),
+        configuration({
+          hostname: "127.0.0.2",
+          port: 9876,
+          idleTimeoutSeconds: 120,
+        }),
       ).pipe(Effect.provide(harness.layer));
       assert.strictEqual(exitCode, 0);
       assert.deepStrictEqual(yield* Ref.get(harness.served), [{
         hostname: "127.0.0.2",
         port: 9876,
+        idleTimeoutSeconds: 120,
         authentication: "shared_token",
       }]);
     }));
@@ -202,6 +210,7 @@ describe("AI Gateway Effect CLI", () => {
       assert.deepStrictEqual(yield* Ref.get(harness.served), [{
         hostname: "0.0.0.0",
         port: 8787,
+        idleTimeoutSeconds: 255,
         authentication: "workload_identity",
       }]);
     }));

services/ai-gateway/tests/config.effect.test.ts

@@ -27,6 +27,7 @@ describe("AI Gateway Effect configuration", () => {
       );
       assert.strictEqual(config.hostname, "0.0.0.0");
       assert.strictEqual(config.port, 8787);
+      assert.strictEqual(config.idleTimeoutSeconds, 255);
       assert.strictEqual(String(config.clientToken), "<redacted>");
       assert.strictEqual(String(config.openAIApiKey), "<redacted>");
       assert.strictEqual(Redacted.value(config.clientToken), "fleet-secret");
@@ -52,6 +53,25 @@ describe("AI Gateway Effect configuration", () => {
       );
     }));
 
+  it.effect("loads an explicit Bun HTTP idle timeout", () =>
+    Effect.gen(function*() {
+      const configured = yield* loadAIGatewayConfig().pipe(
+        Effect.provide(environment({
+          HOME: "/home/agentos",
+          AI_GATEWAY_IDLE_TIMEOUT_SECONDS: "120",
+        })),
+      );
+      const disabled = yield* loadAIGatewayConfig().pipe(
+        Effect.provide(environment({
+          HOME: "/home/agentos",
+          AI_GATEWAY_IDLE_TIMEOUT_SECONDS: "0",
+        })),
+      );
+
+      assert.strictEqual(configured.idleTimeoutSeconds, 120);
+      assert.strictEqual(disabled.idleTimeoutSeconds, 0);
+    }));
+
   it.effect("rejects missing shared auth and malformed runtime bounds", () =>
     Effect.gen(function*() {
       const missingToken = yield* loadAIGatewayConfig().pipe(
@@ -64,6 +84,8 @@ describe("AI Gateway Effect configuration", () => {
       const invalidEnvironments: ReadonlyArray<Readonly<Record<string, string>>> = [
         { AI_GATEWAY_LISTEN_PORT: "0" },
         { AI_GATEWAY_LISTEN_PORT: "65536" },
+        { AI_GATEWAY_IDLE_TIMEOUT_SECONDS: "-1" },
+        { AI_GATEWAY_IDLE_TIMEOUT_SECONDS: "256" },
         { AI_GATEWAY_HEARTBEAT_MILLIS: "0" },
         { AGENTOS_PROVIDER_BUDGET_SETTLEMENT_TIMEOUT_MILLIS: "0" },
       ];