feat: run dynamic apps with agentOS inline execution

Dockerfile

@@ -21,4 +21,4 @@ RUN pnpm build
 
 EXPOSE 3000
 
-CMD ["node", "--no-node-snapshot", "--import", "tsx", "benchmarks/dynamic-apps/src/server.ts", "--host", "0.0.0.0"]
+CMD ["node", "--import", "tsx", "benchmarks/dynamic-apps/src/server.ts", "--host", "0.0.0.0"]

benchmarks/dynamic-apps/RESULTS.md

@@ -1,4 +1,19 @@
-# Direct-isolate Dynamic Apps benchmark
+# agentOS inline Dynamic Apps benchmark
+
+Status: local correctness implemented; deployed performance qualification is
+pending.
+
+The active architecture keeps the durable per-app Rivet actor for deployment,
+then caches one agentOS VM per immutable release and executes the exported ESM
+dispatcher through headless JavaScript evaluation. `pooled` mode resets and
+reinitializes retained contexts; `ephemeral` mode requests a fresh context from
+the cached VM.
+
+No agentOS inline latency numbers are published yet. In particular, the old
+guest-server and native-isolate results below must not be reused as inline
+evaluation latency or as evidence that the new performance gates pass.
+
+## Historical comparison: direct-isolate runtime
 
 Qualified 2026-08-26 (US/Pacific). The rewritten architecture keeps the
 durable per-app Rivet actor for releases and invalidation, but removes scaler
@@ -26,7 +41,7 @@ direct isolate pool 2, actor-worker cache 4, and actor-worker heap limit 96 MiB.
 Every Cloud mutation targeted only
 `dynamic-apps-ben-562e-production-sqac`.
 
-## Runtime hardening qualification (2026-08-30)
+## Historical runtime hardening qualification (2026-08-30)
 
 The local stress suite now exercises multi-app cache churn, large payload
 bursts, release invalidation, cold-cache fan-out, queue overflow, oversized
@@ -75,7 +90,7 @@ from the same release. Its warm actor action cases were 100% successful at
 23.42 ms sequential p50 and 61.27 ms concurrency-16 p50; those numbers include
 the local Engine path, unlike the lower-level worker timing above.
 
-## Decision
+## Historical decision
 
 **Use the direct local-isolate path for ordinary request/response. Keep Rivet
 actors for durable state and app-defined actor semantics.**
@@ -92,7 +107,7 @@ Public Rivet Run ingress from this client is roughly 190–220 ms at low load, s
 outer latency hides the architectural difference. Server timing headers are
 the relevant comparison.
 
-## Final Rivet Compute direct results
+## Historical Rivet Compute direct results
 
 The workload is a zero-dependency JSON `fetch()` handler. Initialization and
 warm-up requests are excluded from the steady samples.
@@ -141,7 +156,7 @@ The small pool deliberately trades burst-tail CPU for bounded idle memory.
 Even with 21.23% overflow creation at concurrency 32, the stated server target
 (p50 <= 25 ms, p95 <= 50 ms) passed and memory returned to its steady bound.
 
-## App-defined actor results
+## Historical app-defined actor results
 
 The actor fixture uses ordinary RivetKit state, actions, an event subscription,
 and a direct HTTP handler in the same deployment. Correctness passed locally
@@ -164,7 +179,7 @@ about 5.5x the complete fresh-isolate server p50. Public ingress adds another
 roughly 225 ms at the median. This confirms that the earlier actor-heavy design
 was benchmarking network topology rather than a useful V8 warm-cache advantage.
 
-## Local direct baseline
+## Historical local direct baseline
 
 The same direct executor on a local Engine completed a 10,000-request pool-2,
 concurrency-2 run at 100% success:

benchmarks/dynamic-apps/package.json

@@ -20,10 +20,10 @@
 		"@rivet-dev/agentos-toolchain": "0.2.15",
 		"@rivet-dev/dynamic-apps": "workspace:*",
 		"hono": "^4.12.9",
-		"rivetkit": "0.0.0-fix-rivetkit-wasm-serve-config.e2b11f9"
+		"rivetkit": "2.3.11"
 	},
 	"devDependencies": {
-		"@rivetkit/engine-cli": "0.0.0-fix-rivetkit-wasm-serve-config.e2b11f9",
+		"@rivetkit/engine-cli": "2.3.11",
 		"@types/node": "^22.19.15",
 		"get-port": "^7.1.0",
 		"tsx": "^4.20.6",

benchmarks/dynamic-apps/src/cloud-stress.ts

@@ -9,7 +9,7 @@ import {
 type StressMode = "ramp" | "soak" | "both";
 
 interface StressCase {
-	name: "warm" | "snapshot" | "actor";
+	name: "pooled" | "ephemeral" | "actor";
 	path: string;
 	concurrency: number;
 	echoRequestId: boolean;
@@ -103,10 +103,10 @@ export async function runCloudStress(
 	try {
 		if (config.mode === "ramp" || config.mode === "both") {
 			result.ramps.push(
-				await runRamp(config, "warm", (concurrency) => [
+				await runRamp(config, "pooled", (concurrency) => [
 					{
-						name: "warm",
-						path: "/bench/warm",
+						name: "pooled",
+						path: "/bench/pooled",
 						concurrency,
 						echoRequestId: true,
 					},
@@ -271,20 +271,20 @@ async function runStage(
 }
 
 function mixedCases(concurrency: number): StressCase[] {
-	const warm = Math.max(1, Math.floor(concurrency * 0.8));
-	const snapshot = Math.max(1, Math.floor(concurrency * 0.1));
-	const actor = Math.max(1, concurrency - warm - snapshot);
+	const pooled = Math.max(1, Math.floor(concurrency * 0.8));
+	const ephemeral = Math.max(1, Math.floor(concurrency * 0.1));
+	const actor = Math.max(1, concurrency - pooled - ephemeral);
 	return [
 		{
-			name: "warm",
-			path: "/bench/warm",
-			concurrency: warm,
+			name: "pooled",
+			path: "/bench/pooled",
+			concurrency: pooled,
 			echoRequestId: true,
 		},
 		{
-			name: "snapshot",
-			path: "/bench/snapshot",
-			concurrency: snapshot,
+			name: "ephemeral",
+			path: "/bench/ephemeral",
+			concurrency: ephemeral,
 			echoRequestId: true,
 		},
 		{
@@ -357,7 +357,7 @@ async function setup(baseUrl: string): Promise<unknown> {
 	);
 	for (let index = 0; index < 10; index += 1) {
 		const response = await fetch(
-			`${baseUrl}/bench/warm?requestId=warmup-${index}`,
+			`${baseUrl}/bench/pooled?requestId=warmup-${index}`,
 			{
 				signal: AbortSignal.timeout(60_000),
 			},
@@ -372,7 +372,7 @@ async function setup(baseUrl: string): Promise<unknown> {
 			body.requestId !== `warmup-${index}`
 		) {
 			throw new Error(
-				`direct warmup ${index} failed with HTTP ${response.status}`,
+				`direct pooled warmup ${index} failed with HTTP ${response.status}`,
 			);
 		}
 	}

benchmarks/dynamic-apps/src/edge.ts

@@ -66,20 +66,15 @@ export function createBenchmarkApplication(): Hono {
 		...process.env,
 		DYNAMIC_APPS_TIMING_HEADERS: "1",
 	});
-	const warm = new DynamicAppsExecutor({
+	const pooled = new DynamicAppsExecutor({
 		...baseConfig,
-		isolateMode: "prewarm",
-		isolatePoolSize: integerEnv("BENCH_WARM_POOL_SIZE", 8, 1, 128),
+		executionMode: "pooled",
+		contextPoolSize: integerEnv("BENCH_CONTEXT_POOL_SIZE", 8, 1, 128),
 	});
-	const snapshot = new DynamicAppsExecutor({
+	const ephemeral = new DynamicAppsExecutor({
 		...baseConfig,
-		isolateMode: "snapshot",
-		isolatePoolSize: 0,
-	});
-	const fresh = new DynamicAppsExecutor({
-		...baseConfig,
-		isolateMode: "fresh",
-		isolatePoolSize: 0,
+		executionMode: "ephemeral",
+		contextPoolSize: 0,
 	});
 	const client = createClient() as unknown as StateClient &
 		BenchmarkDeploymentClient;
@@ -126,7 +121,7 @@ export function createBenchmarkApplication(): Hono {
 	const direct = (
 		executor: DynamicAppsExecutor,
 		prefix: string,
-		architecture: "warm" | "snapshot" | "fresh",
+		architecture: "pooled" | "ephemeral",
 		request: Request,
 	) => {
 		const url = new URL(request.url);
@@ -138,21 +133,17 @@ export function createBenchmarkApplication(): Hono {
 				return response;
 			});
 	};
-	app.all("/bench/warm", (c) => direct(warm, "/bench/warm", "warm", c.req.raw));
-	app.all("/bench/warm/*", (c) =>
-		direct(warm, "/bench/warm", "warm", c.req.raw),
-	);
-	app.all("/bench/fresh", (c) =>
-		direct(fresh, "/bench/fresh", "fresh", c.req.raw),
+	app.all("/bench/pooled", (c) =>
+		direct(pooled, "/bench/pooled", "pooled", c.req.raw),
 	);
-	app.all("/bench/fresh/*", (c) =>
-		direct(fresh, "/bench/fresh", "fresh", c.req.raw),
+	app.all("/bench/pooled/*", (c) =>
+		direct(pooled, "/bench/pooled", "pooled", c.req.raw),
 	);
-	app.all("/bench/snapshot", (c) =>
-		direct(snapshot, "/bench/snapshot", "snapshot", c.req.raw),
+	app.all("/bench/ephemeral", (c) =>
+		direct(ephemeral, "/bench/ephemeral", "ephemeral", c.req.raw),
 	);
-	app.all("/bench/snapshot/*", (c) =>
-		direct(snapshot, "/bench/snapshot", "snapshot", c.req.raw),
+	app.all("/bench/ephemeral/*", (c) =>
+		direct(ephemeral, "/bench/ephemeral", "ephemeral", c.req.raw),
 	);
 
 	app.all("/bench/actor/resolve", async () => {
@@ -229,7 +220,7 @@ export function createBenchmarkApplication(): Hono {
 			const second = await counter.add(3);
 			const current = await counter.inspect();
 			await counter.dispose();
-			const directResponse = await warm.request(
+			const directResponse = await pooled.request(
 				ACTOR_BENCHMARK_APP_ID,
 				new Request("http://dynamic-app.test/"),
 			);
@@ -280,16 +271,14 @@ export function createBenchmarkApplication(): Hono {
 				memory: process.memoryUsage(),
 			},
 			cpuParallelism: availableParallelism(),
-			warm: warm.diagnostics(),
-			snapshot: snapshot.diagnostics(),
-			fresh: fresh.diagnostics(),
+			pooled: pooled.diagnostics(),
+			ephemeral: ephemeral.diagnostics(),
 			paths: [
 				"/bench/noop",
 				"/bench/actor/resolve",
 				"/bench/actor/action",
-				"/bench/warm",
-				"/bench/snapshot",
-				"/bench/fresh",
+				"/bench/pooled",
+				"/bench/ephemeral",
 				"POST /bench/actor-app/setup",
 				"POST /bench/actor-app/verify",
 				"/bench/actor-app/action",
@@ -386,8 +375,7 @@ export function actorApplicationClientConfig(
 		poolName: deployment.pool,
 		...(deployment.token || endpointToken || process.env.RIVET_TOKEN
 			? {
-					token:
-						deployment.token ?? endpointToken ?? process.env.RIVET_TOKEN,
+					token: deployment.token ?? endpointToken ?? process.env.RIVET_TOKEN,
 				}
 			: {}),
 	};

benchmarks/dynamic-apps/src/fixture.ts

@@ -23,13 +23,20 @@ export async function deployBenchmarkFixture(
 					main: "index.js",
 				}),
 				"index.js": `
+import { basename } from "node:path";
+
 export default {
   fetch(request) {
     const url = new URL(request.url);
+		if (url.searchParams.get("logs") === "1") {
+			console.log("benchmark stdout");
+			console.error("benchmark stderr");
+		}
     return Response.json({
       ok: true,
       workload: "basic-request-response",
       requestId: url.searchParams.get("requestId"),
+			node: { platform: process.platform, file: basename(import.meta.filename) },
     });
   },
 };
@@ -61,7 +68,7 @@ export async function deployActorBenchmarkFixture(
 					type: "module",
 					main: "index.js",
 					dependencies: {
-						rivetkit: "0.0.0-fix-rivetkit-wasm-serve-config.e2b11f9",
+						rivetkit: "2.3.11",
 					},
 				}),
 				"index.js": `

benchmarks/dynamic-apps/src/load.ts

@@ -215,9 +215,6 @@ export async function runLoadTest(
 					}
 					for (const [phase, header] of Object.entries({
 						bundleLoad: "x-agentos-app-bundle-load-ms",
-						isolateStart: "x-agentos-app-isolate-start-ms",
-						processReady: "x-agentos-app-process-ready-ms",
-						dispatch: "x-agentos-app-dispatch-ms",
 						appReleaseLookup: "x-agentos-app-release-lookup-ms",
 						appRequestBody: "x-agentos-app-request-body-ms",
 						appScalerAcquire: "x-agentos-app-scaler-acquire-ms",