@boardwalk-labs/runner 0.1.16 → 0.1.17

package/dist/runtime/http_pool_reset.d.ts

@@ -0,0 +1,4 @@
+/** Discard the global HTTP connection pool so post-restore requests open fresh sockets. Never
+ *  throws (a wake must not fail on pool hygiene) and never blocks (the dead pool drains in the
+ *  background). Returns true if the dispatcher was swapped. */
+export declare function resetHttpConnectionPool(): boolean;

package/dist/runtime/http_pool_reset.js

@@ -0,0 +1,49 @@
+// Reset the global HTTP connection pool after a snapshot restore.
+//
+// The North Star substrate freezes the whole VM (memory + every socket) and restores it later,
+// possibly minutes or days on. Keep-alive TCP sockets do NOT survive that: undici pools them, but
+// on restore the frozen socket state is stale and the peer — the per-host egress proxy (a CONNECT
+// tunnel) and, through it, the broker — long since closed its end. undici would REUSE such a socket
+// for the next request, which then hangs (no bytes flow, no RST arrives on a half-dead tunnel) until
+// the control client's 30s AbortSignal.timeout fires and aborts it. On a woken run's FIRST broker
+// call (finalize) that abort crashes the run — the "socket is dead but never reset" gap the control
+// client's timeout only papered over (see runner_control_client.ts).
+//
+// The fix is to discard the dead pool on resume. Swapping in a fresh EnvHttpProxyAgent — which
+// re-reads the unchanged HTTP(S)_PROXY env, so proxying is preserved — means every post-wake request
+// opens a NEW socket instead of reusing a frozen one. Node's global `fetch` reads the same global
+// dispatcher, so this covers the control client, inference, artifacts, and any author fetch alike.
+import { EnvHttpProxyAgent, getGlobalDispatcher, setGlobalDispatcher } from "undici";
+import { createLogger } from "./support/index.js";
+const log = createLogger("http_pool_reset");
+/** Discard the global HTTP connection pool so post-restore requests open fresh sockets. Never
+ *  throws (a wake must not fail on pool hygiene) and never blocks (the dead pool drains in the
+ *  background). Returns true if the dispatcher was swapped. */
+export function resetHttpConnectionPool() {
+    try {
+        const prev = getGlobalDispatcher();
+        setGlobalDispatcher(new EnvHttpProxyAgent());
+        // Drain the old pool's dead sockets in the background; never block the wake on it.
+        const drain = async () => {
+            try {
+                await prev?.close?.();
+            }
+            catch {
+                try {
+                    await prev?.destroy?.(new Error("http pool reset on wake"));
+                }
+                catch {
+                    /* already gone */
+                }
+            }
+        };
+        void drain();
+        return true;
+    }
+    catch (err) {
+        log.warn("http_pool_reset_failed", {
+            error: err instanceof Error ? err.message : String(err),
+        });
+        return false;
+    }
+}

package/dist/runtime/index.js

@@ -38,6 +38,7 @@ import { parseByoProviders } from "./direct_inference.js";
 import { applyIdentityToEnv, connectIdentityRelayFd, relayFdFromEnv, workerDiagnostics, } from "./identity_relay.js";
 import { FreezeCoordinator } from "./freeze_coordinator.js";
 import { reseedUserspaceCsprng } from "./uniqueness_reseed.js";
+import { resetHttpConnectionPool } from "./http_pool_reset.js";
 import { WorkerWorkflowHost } from "./workflow_host.js";
 import { BrowserSessionManager } from "./browser_session.js";
 import { loadGuestBrowserConfig, makeGuestBrowserBackend, connectSessionMcp, } from "./browser_session_backend.js";
@@ -393,6 +394,10 @@ export function assembleWorkerDeps(runtime) {
                 // BEFORE the seam resolves, so no woken author code draws from the stale (pre-suspend) DRBG.
                 onAfterWake: (wake) => {
                     reseedUserspaceCsprng();
+                    // Keep-alive sockets (to the egress proxy + through it to the broker) do NOT survive the
+                    // freeze; discard the connection pool so the first post-wake broker call (finalize) opens a
+                    // fresh socket instead of hanging on a stale one until its 30s timeout crashes the run.
+                    resetHttpConnectionPool();
                     broker.swapRunToken(wake.run_token);
                     redactor.record(wake.run_token);
                     if (wake.api_token !== undefined && wake.api_token !== "") {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@boardwalk-labs/runner",
-  "version": "0.1.16",
+  "version": "0.1.17",
   "description": "Boardwalk self-hosted runner: execute cloud-scheduled runs on your own machines. Currently: the canonical runner contract types.",
   "license": "Apache-2.0",
   "repository": {
@@ -58,6 +58,7 @@
     "@modelcontextprotocol/sdk": "^1.29.0",
     "node-gyp-build": "^4.8.4",
     "tar": "^7.5.16",
+    "undici": "^6.27.0",
     "zod": "^4.0.0"
   },
   "devDependencies": {