fix: TUI and GUI commands fail - workspace routing, UID mismatch, inference route

[LinoyBitan1]

Summary

Fix three bugs that break all four TUI/GUI commands after a fresh validated-pattern deploy.

Jira: APPENG-6158

What's broken on main

After ./pattern.sh make install and make openshell-saw-configure-gateway:

• make nemoclaw-tui SANDBOX_NAME=cuda-sandbox → sandbox not found
• make nemoclaw-gui SANDBOX_NAME=cuda-sandbox → Could not extract dashboard token
• make openclaw-tui SANDBOX_NAME=notebook → Unsafe fallback OpenClaw temp dir
• make openclaw-gui SANDBOX_NAME=notebook → Could not extract token

Fixes

1. Workspace routing — nemoclaw-tui, nemoclaw-gui, and openshell-saw-gui.sh pass --workspace to sandbox list but not to ssh-proxy or sandbox exec. The proxy defaults to workspace default and can't find cuda-sandbox in cuda-dev. Fixed by adding --workspace to all connection commands.

2. UID mismatch — The openclaw-openshell image owns /sandbox as UID 65532. The supervisor assigns a dynamic UID at runtime (1000, 998, …). OpenClaw fs-safe requires the temp dir to be owned by the running user — the mismatch crashes it. Fixed by running docker exec -u 0 chown -R sandbox:sandbox /sandbox after the sandbox reaches Ready.

3. Inference 503 — apply_bom.py calls openshell inference set without --workspace (so only default gets a route) and without --system (so inference.local has no backend). Fixed by passing --workspace on the user route and --system on the system route.

Changes

• Makefile-quickstart — --workspace on ssh-proxy in nemoclaw-tui; pass WORKSPACE to GUI script
• openshell-saw-gui.sh — --workspace on sandbox exec, ssh-proxy, and SSH hostname
• apply_bom.py — chown_sandbox_home() after sandbox Ready; --workspace and --system on inference set
• README.md — WORKSPACE=cuda-dev on NemoClaw commands; sandbox list --workspace cuda-dev

Verification (fresh deploy)

• make nemoclaw-tui SANDBOX_NAME=cuda-sandbox WORKSPACE=cuda-dev — connects, inference responds
• make openclaw-tui SANDBOX_NAME=notebook — starts without UID crash
• make nemoclaw-gui SANDBOX_NAME=cuda-sandbox WORKSPACE=cuda-dev GUI_PORT=18789 — works
• make openclaw-gui SANDBOX_NAME=notebook GUI_PORT=18790 — works
• Setup Job completes 10/10 ALL PASSED

[mtalvi]

LGTM!
Very nice work @LinoyBitan1 !
Yes the issue was testing on a fresh cluster vs an existing one.