Nvidia fixes a startup path that bypassed disabled agent auto-approval Nvidia fixes a startup path that bypassed disabled agent auto-approval

What is the difference between policy and preference?

A saved startup preference describes how an application should open next time. A managed capability defines which behavior its owner permits. These settings can disagree. In this case, the owner disabled automatic approval while the agent retained a native preference for Auto. The distinction matters because an operator could inspect the management setting and reasonably expect the next session to follow it. The release makes that expectation part of startup resolution, instead of relying on the preference to match the policy.

How did the mismatch appear?

The October 27 issue describes a managed sandbox with automatic approval disabled and a native startup setting of auto. The reporter could not enable the same behavior through the command-line shortcut: that path refused the request. Opening an interactive session nevertheless displayed Auto. The report identified Ubuntu on x86 hardware and DGX Spark on Arm, using NemoClaw 0.0.138 and Deep Agents Code 0.1.71. Changing the saved preference required shell access inside the sandbox. The report isolates how two launch paths produced different outcomes under the same managed setting.

What does the new check change?

The merged fix applies the managed capability where Deep Agents Code resolves its startup mode. When automatic approval is disabled, that resolver returns Manual. When thread opt-in is allowed, native resolution remains available. This addresses the inherited setting rather than only blocking an explicit launch flag. The change also makes the managed image build fail if the expected resolver disappears upstream. That turns a changed integration point into a visible build problem instead of silently leaving the startup restriction unapplied.

What would a worked check look like?

Consider a disposable test sandbox, not a production incident. Its owner disables automatic approval, while its saved startup preference requests Auto. A new session should show Manual. Repeating the same check with the native Yolo preference tests a second inherited path. In both cases, the relevant observation is the running session, not just the stored file. Restoring the original configuration completes the exercise. This example illustrates why checking a rejected command-line option alone cannot establish that every route into a mode is restricted.

What does this establish about agent control?

The issue separates three observations: the owner setting, the launch request and the mode that actually starts. They are related, but none is a substitute for the others. A configuration screen can accurately display a saved preference even when policy prevents that preference from taking effect. For an agent integration, the useful boundary is where the effective behavior is selected. This release supplies a concrete example of moving the restriction to that boundary. It also shows why a small startup change can matter independently of model quality or task completion speed.