From 58b18fc85e9c165e75d8fdbeaeeb03be6117f56c Mon Sep 17 00:00:00 2001
From: JJ Eaton <58495321+jayleaton@users.noreply.github.com>
Date: Sun, 25 Oct 2026 10:07:51 +0700
Subject: [PATCH] fix(mcp-gateway): stop killing slow-starting owners during
 session handover

The launcher gave a starting shared owner 15 s to report ready and then
killed it. Owner startup waits on no peer, so it always ends in ready, an
error, or an exit; the deadline only killed healthy owners on a loaded
machine and ended the MCP session being moved to them. CI hit this in the
stdio handover test, which spawns the TypeScript source (5-8 s to start).

The handover test now fails with the launcher's stderr when the session
exits, instead of waiting for the 60 s test timeout.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---
 packages/mcp-gateway/src/sharedGateway.test.ts | 7 +++++++
 packages/mcp-gateway/src/sharedLauncher.ts     | 6 ++----
 2 files changed, 9 insertions(+), 4 deletions(-)

diff --git a/packages/mcp-gateway/src/sharedGateway.test.ts b/packages/mcp-gateway/src/sharedGateway.test.ts
index a55eddab4c5bc..79f2d7ffd8ebc 100644
--- a/packages/mcp-gateway/src/sharedGateway.test.ts
+++ b/packages/mcp-gateway/src/sharedGateway.test.ts
@@ -672,10 +672,17 @@ describe("shared MCP gateway", () => {
         T3_MCP_GRANTS: "{}",
       },
     });
+    let stderr = "";
+    transport.stderr?.on("data", (chunk: Buffer) => {
+      stderr += chunk.toString();
+    });
     cleanup.push(() => client.close());
     await client.connect(transport);
     const changed = Promise.withResolvers<void>();
     client.setNotificationHandler(ToolListChangedNotificationSchema, () => changed.resolve());
+    // A launcher that cannot move the session exits; fail with its reason instead of waiting.
+    // oxlint-disable-next-line unicorn/prefer-add-event-listener -- MCP Client exposes an onclose callback rather than a DOM event API.
+    client.onclose = () => changed.reject(new Error(`Launcher exited: ${stderr.trim()}`));
     expect((await client.listTools()).tools.length).toBeGreaterThan(0);
 
     await first!.close();
diff --git a/packages/mcp-gateway/src/sharedLauncher.ts b/packages/mcp-gateway/src/sharedLauncher.ts
index 8e2fe558b498a..532ff1d209a72 100644
--- a/packages/mcp-gateway/src/sharedLauncher.ts
+++ b/packages/mcp-gateway/src/sharedLauncher.ts
@@ -39,9 +39,9 @@ export function launchSharedOwner(
         },
       },
     );
-    const timeout = AbortSignal.timeout(15_000);
+    // Startup waits on no peer, so it always ends in ready, an error, or an exit. A deadline would
+    // only kill a healthy owner a loaded machine starts slowly; a stuck launcher is its host's to end.
     const cleanup = () => {
-      timeout.removeEventListener("abort", onTimeout);
       child.off("error", fail);
       child.off("exit", onExit);
       child.off("message", onMessage);
@@ -52,7 +52,6 @@ export function launchSharedOwner(
       if (child.connected) child.disconnect();
       reject(error);
     };
-    const onTimeout = () => fail(new Error("Timed out starting the shared MCP gateway owner."));
     const onExit = (code: number | null) =>
       fail(new Error(`Shared MCP gateway owner exited before startup (code ${code}).`));
     const onMessage = (value: unknown) => {
@@ -66,7 +65,6 @@ export function launchSharedOwner(
         resolve(child);
       }
     };
-    timeout.addEventListener("abort", onTimeout, { once: true });
     child.once("error", fail);
     child.once("exit", onExit);
     child.on("message", onMessage);