Description
PR #3925 (reopen/pr-3832-upgrade-deps, head fa28360e) post-rebuild leaves the in-sandbox OpenClaw gateway unable to start on aarch64. The gateway-recovery warning fires (#2478), nothing more is logged, and the container stays Docker-unhealthy. Reproduces 3/3 sandboxes across DGX Spark + DGX Station; same Spark box on main HEAD (cfa817b) passes rebuild's own deployment check.
Environment
Device: DGX Spark (and reproduced on DGX Station GB300)
OS: Ubuntu 24.04.4 LTS (Noble Numbat)
Architecture: aarch64
Node.js: v22.22.3
npm: 10.9.8
Docker: 29.2.1, build a5c7197
OpenShell CLI: 0.0.39 (host) — sandboxes.json keeps openshellVersion "0.0.39" even though PR's package.json pins 0.0.44
NemoClaw: v0.1.0 (PR branch reopen/pr-3832-upgrade-deps fa28360e); pre-upgrade v0.0.46
OpenClaw: 2026.5.18 (50a2481) — bundled by PR; pre-upgrade and on main is 2026.4.24 (cbcfdf6)
Steps to Reproduce
- On DGX Spark (aarch64): start from NemoClaw v0.0.46 with sandbox
my-assistant onboarded and healthy.
- Clone the PR branch:
git clone -b reopen/pr-3832-upgrade-deps --depth 1 \
https://github.com/NVIDIA/NemoClaw.git ~/NemoClaw-pr3925
- Run the PR-branch installer:
cd ~/NemoClaw-pr3925 && \
NEMOCLAW_ACCEPT_THIRD_PARTY_SOFTWARE=1 NEMOCLAW_NONINTERACTIVE=1 \
yes y | bash install.sh
- Trigger a sandbox rebuild:
nemoclaw my-assistant rebuild --yes
- Observe rebuild's own deployment-verify step and
docker ps after.
Reproduced identically on DGX Station GB300 with two openclaw sandboxes (discord-sb, gemini-sb).
Expected Result
Rebuild's deployment-verify reports Deployment verified — gateway and dashboard are healthy. Container shows (health: starting) briefly then becomes healthy. Gateway log contains [gateway] ready (N plugins: ...) within a few seconds. Dashboard port (18789/18790) is reachable.
This is exactly what happens on main HEAD (cfa817b), which still bundles OpenClaw 2026.4.24.
Actual Result
Rebuild's own deployment-verify on PR branch reports:
✗ gateway: HTTP 0 (gateway not responding)
✗ dashboard: port forward not working (connection refused)
⚠ Deployment verification found issues
The sandbox was created successfully but may not be fully functional.
docker ps shows new container Up X minutes (unhealthy) — does not transition out of unhealthy. docker exec <ctr> cat /tmp/gateway.log contains only:
[gateway-recovery] WARNING: /tmp/nemoclaw-proxy-env.sh missing - gateway launching without library guards (#2478)
No further gateway init lines ever appear (no [gateway] ready, no plugin register, no HTTP server start). Reproduced 3/3 sandboxes on 2/2 ARM64 boxes (DGX Spark + DGX Station GB300).
Logs
=== Spark PR-branch rebuild ===
✓ Sandbox 'my-assistant' rebuilt successfully
Now running: OpenClaw v2026.5.18
✗ gateway: HTTP 0 (gateway not responding)
✗ dashboard: port forward not working (connection refused)
Container: openshell-my-assistant-18b489dd-... Up 15 minutes (unhealthy)
$ docker exec <ctr> cat /tmp/gateway.log
[gateway-recovery] WARNING: /tmp/nemoclaw-proxy-env.sh missing - gateway launching without library guards (#2478)
(EOF -- no further lines)
=== Negative control on same Spark box: main HEAD (cfa817b) rebuild ===
✓ Deployment verified - gateway and dashboard are healthy.
OpenClaw version: 2026.4.24
Now running: OpenClaw v2026.4.24
Container: openshell-my-assistant-37b15ded-... Up 1 minute (health: starting)
$ docker exec <ctr> cat /tmp/gateway.log
[sandbox-safety-net] loaded (launcher)
[guard] ciao-network-guard loaded (launcher)
2026-05-21T06:04:06 [gateway] loading configuration...
2026-05-21T06:04:06 [gateway] starting...
[guard] os.networkInterfaces() failed: A system error occurred: uv_interface_addresses returned Unknown system error 1 - returning empty (mDNS disabled)
2026-05-21T06:04:07 [plugins] plugins.allow is empty; discovered non-bundled plugins may auto-load: nemoclaw
2026-05-21T06:04:07 [plugins] nemoclaw failed during register from /sandbox/.openclaw/extensions/nemoclaw/dist/index.js: SyntaxError: Unexpected end of JSON input
2026-05-21T06:04:07 [gateway] starting HTTP server...
2026-05-21T06:04:07 [gateway] ready (4 plugins: browser, device-pair, phone-control, talk-voice; 1.6s)
DevTest evidence:
- 597724 (Spark
my-assistant)
- 597725 (Station
discord-sb)
- 597726 (Station
gemini-sb)
All under \DevTest\NemoClaw\NemoClaw Test\v0.0.47\manual\{spark,station,windows-arm-reference-host}.
Related
GH #2478 (closed) — same ARM64 networkInterfaces EPERM context; main has the workaround active and the gateway still reaches [gateway] ready. PR #3925 appears to regress the workaround or remove the proxy-env.sh setup path that triggers the gateway-recovery WARNING.
NVB#6198894
Description
PR #3925 (
reopen/pr-3832-upgrade-deps, headfa28360e) post-rebuild leaves the in-sandbox OpenClaw gateway unable to start on aarch64. The gateway-recovery warning fires (#2478), nothing more is logged, and the container stays Docker-unhealthy. Reproduces 3/3 sandboxes across DGX Spark + DGX Station; same Spark box on main HEAD (cfa817b) passes rebuild's own deployment check.Environment
Steps to Reproduce
my-assistantonboarded and healthy.git clone -b reopen/pr-3832-upgrade-deps --depth 1 \ https://github.com/NVIDIA/NemoClaw.git ~/NemoClaw-pr3925docker psafter.Reproduced identically on DGX Station GB300 with two openclaw sandboxes (
discord-sb,gemini-sb).Expected Result
Rebuild's deployment-verify reports
Deployment verified — gateway and dashboard are healthy. Container shows(health: starting)briefly then becomes healthy. Gateway log contains[gateway] ready (N plugins: ...)within a few seconds. Dashboard port (18789/18790) is reachable.This is exactly what happens on main HEAD (
cfa817b), which still bundles OpenClaw 2026.4.24.Actual Result
Rebuild's own deployment-verify on PR branch reports:
docker psshows new containerUp X minutes (unhealthy)— does not transition out of unhealthy.docker exec <ctr> cat /tmp/gateway.logcontains only:No further gateway init lines ever appear (no
[gateway] ready, no plugin register, no HTTP server start). Reproduced 3/3 sandboxes on 2/2 ARM64 boxes (DGX Spark + DGX Station GB300).Logs
DevTest evidence:
my-assistant)discord-sb)gemini-sb)All under
\DevTest\NemoClaw\NemoClaw Test\v0.0.47\manual\{spark,station,windows-arm-reference-host}.Related
GH #2478 (closed) — same ARM64
networkInterfacesEPERM context; main has the workaround active and the gateway still reaches[gateway] ready. PR #3925 appears to regress the workaround or remove the proxy-env.sh setup path that triggers the gateway-recovery WARNING.NVB#6198894