nemoclaw onboard is not resumable — partial failures require full cleanup and restart
Summary
The nemoclaw onboard wizard runs 7 sequential steps (API key → preflight → gateway → sandbox → inference → OpenClaw → policies). If any step fails mid-way (due to OOM, network timeout, terminal disconnect, etc.), the wizard cannot be resumed from the point of failure. Worse, leftover state from the partial run can block subsequent attempts.
Environment
| Component |
Detail |
| Platform |
NVIDIA Brev (GCP, Ubuntu 24.04) |
| RAM |
Tested on 8 GiB and 16 GiB instances |
| NemoClaw |
Installed via official installer (March 18–19, 2026) |
Observed Behavior
Problem 1: No checkpoint/resume
When the onboard wizard fails at step 4 (sandbox image push — common on low-memory instances), rerunning nemoclaw onboard starts from step 1 again. There is no --resume or --from-step flag. On low-bandwidth or resource-constrained environments, this means repeating 5–10 minutes of work that already completed successfully.
Problem 2: Stale gateway blocks re-onboarding
If the wizard fails after the gateway has been started (step 3), the OpenShell gateway remains running on port 8080. On the next attempt:
Error: Gateway already running on port 8080
The user must manually clean up before retrying:
openshell gateway destroy -g nemoclaw
This is not documented in the quickstart or troubleshooting guides.
Problem 3: nemoclaw not in PATH after install
After a fresh install via the official script, nemoclaw is not available in the current shell session:
$ nemoclaw my-assistant connect
nemoclaw: command not found
Requires source ~/.bashrc to pick up the updated PATH. This is especially confusing when the install script's final output says to run nemoclaw my-assistant connect — the suggested command immediately fails.
Expected Behavior
- Resumable onboard: The wizard should detect which steps have already completed successfully and skip them. At minimum, provide a
--resume flag.
- Clean error on stale gateway: If a gateway is already running, the wizard should offer to restart it instead of failing with an opaque error.
- PATH updated in current session: The install script should
export PATH or instruct the user to source ~/.bashrc in its final output.
Suggested Improvements
# Ideal experience after a failure:
$ nemoclaw onboard --resume
[1/7] API Key ............ ✓ (cached)
[2/7] Preflight .......... ✓ (cached)
[3/7] Gateway ............ ✓ (running)
[4/7] Sandbox ............ resuming image push...
For the PATH issue, the install script could append:
echo ""
echo "[INFO] Run 'source ~/.bashrc' or open a new terminal to use 'nemoclaw'"
Workaround
Full cleanup before retrying:
openshell gateway destroy -g nemoclaw 2>/dev/null
# If sandbox was partially created:
openshell sandbox delete <name> 2>/dev/null
# Then re-run:
curl -fsSL https://nvidia.com/nemoclaw.sh | bash
nemoclaw onboardis not resumable — partial failures require full cleanup and restartSummary
The
nemoclaw onboardwizard runs 7 sequential steps (API key → preflight → gateway → sandbox → inference → OpenClaw → policies). If any step fails mid-way (due to OOM, network timeout, terminal disconnect, etc.), the wizard cannot be resumed from the point of failure. Worse, leftover state from the partial run can block subsequent attempts.Environment
Observed Behavior
Problem 1: No checkpoint/resume
When the onboard wizard fails at step 4 (sandbox image push — common on low-memory instances), rerunning
nemoclaw onboardstarts from step 1 again. There is no--resumeor--from-stepflag. On low-bandwidth or resource-constrained environments, this means repeating 5–10 minutes of work that already completed successfully.Problem 2: Stale gateway blocks re-onboarding
If the wizard fails after the gateway has been started (step 3), the OpenShell gateway remains running on port 8080. On the next attempt:
The user must manually clean up before retrying:
This is not documented in the quickstart or troubleshooting guides.
Problem 3:
nemoclawnot in PATH after installAfter a fresh install via the official script,
nemoclawis not available in the current shell session:Requires
source ~/.bashrcto pick up the updated PATH. This is especially confusing when the install script's final output says to runnemoclaw my-assistant connect— the suggested command immediately fails.Expected Behavior
--resumeflag.export PATHor instruct the user tosource ~/.bashrcin its final output.Suggested Improvements
# Ideal experience after a failure: $ nemoclaw onboard --resume [1/7] API Key ............ ✓ (cached) [2/7] Preflight .......... ✓ (cached) [3/7] Gateway ............ ✓ (running) [4/7] Sandbox ............ resuming image push...For the PATH issue, the install script could append:
Workaround
Full cleanup before retrying: