Problem

When running nemoclaw deploy <instance>, the workspace files that define the agent's identity and task context are not transferred to the remote sandbox. The deployed agent arrives blank — no SOUL.md, MEMORY.md, IDENTITY.md, USER.md, or AGENTS.md.

The expected lifecycle is:

1. Configure agent locally (edit workspace files, shape identity/memory/task)
2. nemoclaw deploy <instance> to a Brev GPU VM
3. Agent starts working on the remote machine with the same identity and context

Currently step 2 creates a fresh sandbox with default config. The user must manually:

openshell sandbox download <local-sandbox> /sandbox/.openclaw/workspace/ ~/backup/
scp ~/backup/* <instance>:/tmp/
ssh <instance> 'for f in IDENTITY.md SOUL.md USER.md AGENTS.md MEMORY.md; do openshell sandbox upload <remote-sandbox> /tmp/$f /sandbox/.openclaw/workspace/; done'

This is error-prone and breaks the "configure locally, deploy remotely" workflow.

Proposed solution

nemoclaw deploy should:

1. Before deploying, download workspace files from the local sandbox (if one exists)
2. After the remote sandbox is created, upload those workspace files to the remote sandbox
3. Optionally accept a --from <sandbox-name> flag to specify which local sandbox to clone from

Something like:

nemoclaw deploy my-gpu-box --from my-local-assistant

The workspace files are small (a few KB of markdown) so this adds negligible overhead to the deploy process.

Workaround

Manual download/upload via openshell sandbox download and openshell sandbox upload as described above.

Environment

• NemoClaw v0.1.0
• OpenShell 0.0.19+
• macOS (local) → Brev L40S (remote)