Problem

nemoclaw sandbox connect (and other commands that use the last used sandbox as a default) can fail when the locally stored last-used sandbox points to a sandbox that no longer exists.

Example:

nemoclaw sandbox connect
→ Using sandbox 'commanding-sasquatch' (last used)
Error:   × status: NotFound, message: "sandbox not found", details: [], metadata: MetadataMap { headers: {"content-type": "application/grpc", "date": "Sun, 08 Mar 2026
  │ 02:56:56 GMT"} }

Context

This came up while running e2e tests against a cluster. The tests updated the local last used sandbox, then the sandbox was deleted, leaving the CLI pointing at a stale sandbox name.

Expected behavior

If the stored last-used sandbox no longer exists, the CLI should recover gracefully instead of failing with sandbox not found.

Possible approaches

• Delete or reset the stored last-used sandbox when a sandbox is deleted
• Have e2e tests clean up or restore the last_sandbox file after they run
• Keep a stack/list of recent sandboxes and fall back to the next valid one when the current one is gone

Acceptance criteria

• Reproduce the stale last used sandbox case with an automated test
• nemoclaw sandbox connect no longer hard-fails when the stored sandbox has been deleted
• The last-used sandbox state is cleaned up or recovered in a predictable way
• Documentation or behavior is clear enough for users and for e2e test workflows