Bug type

Regression (worked before, now fails)

Beta release blocker

No

Summary

Environment

• NemoClaw v0.0.55, bundled OpenClaw runtime 2026.5.22
• Platform: NVIDIA DGX Spark, Linux aarch64 (kernel 6.17), Node v22.22.3
• Sandbox agent: OpenClaw, model Qwen/Qwen3.6-27B-FP8 (local vLLM)
• Selected messaging channel: Discord (Telegram alone builds fine)

Summary

Running nemoclaw onboard with Discord enabled aborts during the sandbox image build at
step 66/80, when openclaw-build-messaging-plugins.py returns a non-zero exit code. With
Discord deselected the build completes. This makes Discord effectively uninstallable via
the normal onboarding path.

Analysis

For Discord the script runs (subprocess.run(check=True)):
openclaw plugins install @openclaw/discord@${OPENCLAW_VERSION} # = @openclaw/discord@2026.5.22
openclaw doctor --fix --non-interactive
so a non-zero exit from either crashes it (uncaught CalledProcessError → exit 1).

Notably:

• @openclaw/discord@2026.5.22 DOES exist on npm and installs fine at runtime inside a
  running sandbox - so this is not a missing-version problem.
• The build fails in ~1.4s - too fast for an npm fetch - pointing to a build-stage
  environment issue (npm registry / egress proxy not reachable, or openclaw not yet
  usable during docker build), or to openclaw doctor --fix failing in that stage.
  The orchestrator only surfaces the non-zero exit code; the script's stderr is not
  propagated, so the precise cause can't be confirmed from the user side.

Asks

1. Surface openclaw-build-messaging-plugins.py stdout/stderr in the build output (or run
   it with set -x) so the failing sub-command is visible.
2. Confirm whether the npm registry / egress proxy is available during this build stage on
   DGX Spark (aarch64), and whether openclaw doctor --fix is the failing step.
3. Consider not hard-failing the whole image build when a single messaging-plugin install
   fails (degrade gracefully / retry).

Related

After bypassing this step (Discord plugin @openclaw/discord@2026.5.22 correctly installed),
the Discord gateway WebSocket never reaches READY in the proxy-only sandbox
(getaddrinfo EAI_AGAIN discord.com, restart loop) - the symptom of #409, which the release
notes say was fixed in v0.0.48 with "a loopback proxy for gateway traffic." On v0.0.55 that
loopback proxy is not active (no loopback proxy listener running; generated openclaw.json
sets no proxy for the Discord account). Looks like a regression of #409.

Steps to reproduce

1. nemoclaw onboard on a DGX Spark.
2. Select Discord as a messaging channel and provide bot token / server id / user id.
3. Let the sandbox image build run.

Expected behavior

The agent can use discord

Actual behavior

Actual result

Error:   × Docker build stream error
  ╰-> Docker stream error: The command '/bin/bash -o pipefail -c python3 /usr/
      local/lib/nemoclaw/openclaw-build-messaging-plugins.py' returned a non-
      zero code: 1
  Step 66/80 stopped after 1.4s (RUN python3 /usr/local/lib/nemoclaw/openclaw-build-messaging-plugins.py)
  Sandbox image build stopped after 206.8s

OpenClaw version

NemoClaw v0.0.55, bundled OpenClaw runtime 2026.5.22

Operating system

Ubuntu 24.04

Install method

No response

Model

qwen3.6 27B FP8

Provider / routing chain

vllm-local -> Qwen/Qwen3.6-27B-FP8

Additional provider/model setup details

No response

Logs, screenshots, and evidence

Impact and severity

No response

Additional information

No response