fix(agent): keep Deep Agents Code sandbox alive with a stable entrypoint

agents/langchain-deepagents-code/start.sh

@@ -82,8 +82,16 @@ prepare_runtime_env() {
 
 prepare_runtime_env
 
+# With no command, this invocation IS the sandbox's long-running entrypoint.
+# Deep Agents Code is a terminal-runtime agent invoked on demand via
+# `openshell sandbox exec`, so the entrypoint has no daemon to run and must
+# stay alive as a stable foreground process. A bare `/bin/bash` exits
+# immediately in a non-interactive sandbox (no TTY, EOF on stdin), leaving the
+# sandbox with no persistent process: OpenShell then flaps it into the Error
+# phase, which breaks the Docker GPU-patch supervisor reconnect and leaves GPU
+# posture unreliable (#5717). Idle forever instead so the sandbox stays Ready.
 if [ "$#" -eq 0 ]; then
-  set -- /bin/bash
+  exec sleep infinity
 fi
 
 exec "$@"

test/dcode-start-keepalive.test.ts

@@ -0,0 +1,58 @@
+// @ts-nocheck
+// SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+// SPDX-License-Identifier: Apache-2.0
+
+import { spawnSync } from "node:child_process";
+import fs from "node:fs";
+import path from "node:path";
+import { afterEach, describe, expect, it } from "vitest";
+
+const START_SCRIPT = path.join(
+  import.meta.dirname,
+  "..",
+  "agents",
+  "langchain-deepagents-code",
+  "start.sh",
+);
+
+// start.sh hardcodes this runtime-env path; clean it up so the test is hermetic.
+const RUNTIME_ENV_FILE = "/tmp/nemoclaw-proxy-env.sh";
+
+afterEach(() => {
+  fs.rmSync(RUNTIME_ENV_FILE, { force: true });
+});
+
+describe("Deep Agents Code sandbox entrypoint keep-alive (#5717)", () => {
+  it("stays alive as a long-running process when invoked with no command", () => {
+    // The terminal-runtime sandbox runs this entrypoint with no args as its
+    // sole foreground process. It must NOT exit on its own — a self-exiting
+    // entrypoint (e.g. a bare non-interactive /bin/bash) leaves the sandbox
+    // with no persistent process, flapping it into OpenShell's Error phase and
+    // breaking the Docker GPU-patch supervisor reconnect. Run with stdin closed
+    // and a short timeout: a correct keep-alive is still running at the
+    // deadline (killed by the timeout signal), not exited cleanly. Execute the
+    // script directly (not via `bash`) so this also exercises the real ENTRYPOINT
+    // contract — the image runs /usr/local/bin/nemoclaw-start directly, so a
+    // broken shebang or execute bit would also be caught here.
+    const result = spawnSync(START_SCRIPT, [], {
+      input: "",
+      timeout: 3000,
+      encoding: "utf-8",
+    });
+
+    // Killed by the timeout (still running) => signal set, status null.
+    // A self-exiting entrypoint would return status 0 with no signal.
+    expect(result.signal).toBe("SIGTERM");
+    expect(result.status).toBeNull();
+  });
+
+  it("execs an explicitly supplied command instead of idling", () => {
+    const result = spawnSync(START_SCRIPT, ["printf", "RAN_CMD"], {
+      input: "",
+      timeout: 3000,
+      encoding: "utf-8",
+    });
+    expect(result.status).toBe(0);
+    expect(result.stdout).toContain("RAN_CMD");
+  });
+});