bug: Claude Code crashes on GB200 (ARM64 64K page kernel)

[klueska]

Agent Diagnostic

Investigated by: Claude Code (debug-openshell-cluster skill)
Cluster: openshell · https://127.0.0.1:8080 · v0.0.7
Sandbox: strong-treefrog · image ghcr.io/nvidia/openshell-community/sandboxes/base@sha256:97a6a668...

Cluster is fully healthy — this is not a cluster startup issue.

openshell status        → Connected (v0.0.7)
node status             → Ready (k3s v1.35.2+k3s1)
openshell-0             → Running, 0 restarts

Root cause: the Claude Code binary aborts on ARM64 kernels with 64K page size.

Property	Value
Kernel	6.14.0-1013-nvidia-64k
Architecture	aarch64
getconf PAGE_SIZE	65536 (64K)
/usr/local/bin/claude	ELF ARM64 standalone binary, 222MB
Bundled runtime	Bun v1.3.1 (confirmed via grep -boa "Bun v")
Crash signal	SIGABRT (exit 134, not SIGSEGV)
Seccomp	Disabled — not the cause
Node.js v22.22.1	Works fine in same sandbox

Bun's JavaScript engine (JavaScriptCore) performs memory layout operations assuming ≤16K page alignment. On a 64K page kernel, these assumptions are violated and JSC calls abort() before any user code executes. node itself is unaffected — the NPM-based install path uses Node.js and avoids this entirely.

Commands run:

openshell status
openshell doctor exec -- kubectl get pods -A -o wide
openshell doctor exec -- kubectl -n openshell exec strong-treefrog -- uname -a
openshell doctor exec -- kubectl -n openshell exec strong-treefrog -- getconf PAGE_SIZE
openshell doctor exec -- kubectl -n openshell exec strong-treefrog -- /usr/local/bin/claude --version
openshell doctor exec -- kubectl -n openshell exec strong-treefrog -- sh -c 'cat /proc/self/status | grep -E "Seccomp|NoNewPrivs|CapEff"'
openshell doctor exec -- kubectl -n openshell exec strong-treefrog -- ldd /usr/local/bin/claude
openshell doctor exec -- kubectl -n openshell exec strong-treefrog -- sh -c 'ls -lh /usr/local/bin/claude'
openshell doctor exec -- kubectl -n openshell exec strong-treefrog -- sh -c 'grep -boa "Bun v" /usr/local/bin/claude'
openshell doctor exec -- kubectl -n openshell exec strong-treefrog -- node -e "process.exit(0)"  # ✓ works

Description

The default Claude Code installer produces a binary that does not run correctly on GB200 nodes.

$ openshell sandbox create -- claude
...
Created sandbox: strong-treefrog

$ openshell sandbox connect strong-treefrog
sandbox@strong-treefrog:~$ claude
Aborted (core dumped)

This is a known issue with the Claude Code installer on these machines. The workaround is to install via NPM instead:

npm install -g @anthropic-ai/claude-code

OpenShell should consider using the NPM-based install path for Claude on
affected systems. It is unclear whether this is specific to Grace-based
systems (GB200) or affects ARM64 in general — worth testing on other
aarch64 machines to narrow it down.

Reproduction Steps

1. On a GB200 node, create a Claude sandbox: openshell sandbox create -- claude
2. Connect to the sandbox: openshell sandbox connect <name>
3. Run claude
4. Observe: Aborted (core dumped)

Environment

• Host: GB200 (Grace Blackwell), aarch64
• OpenShell sandbox with Claude provider

Logs

Agent-First Checklist

• I pointed my agent at the repo and had it investigate this issue
• I loaded relevant skills (e.g., debug-openshell-cluster, debug-inference, openshell-cli)
• My agent could not resolve this — the diagnostic above explains why

[github-actions]

This issue appears to have been opened without an agent investigation.

OpenShell is an agent-first project - please point your coding agent at the repo and have it diagnose this before we triage. Your agent can load skills like debug-openshell-cluster, debug-inference, openshell-cli, and generate-sandbox-policy.

See CONTRIBUTING.md for the full workflow.

[klueska]

I've updated the description after pointing (native) claude to the affected sandbox (trying to run claude)

[jgehrcke]

This is a known issue with the Claude Code installer on these machines.

anthropics/claude-code#25713

[pimlock]

Thanks for reporting this!

Overall, we would like for Claude to fix this, reverting back to npm install for claude (which is also marked as deprecated) would have disadvantages, e.g. the binary that would run claude would be set to node and we use that information to enforce policies.

The workaround for now would be to create a custom image with claude installed with the npm, see: https://docs.nvidia.com/openshell/latest/sandboxes/manage-sandboxes.html#custom-containers

It could be achieved by something like this (by changing the existing base image)

# Minimal OpenShell sandbox that installs Claude Code with npm instead of the
# native installer shipped in the base image.
ARG BASE_IMAGE=ghcr.io/nvidia/openshell-community/sandboxes/base:latest
FROM ${BASE_IMAGE}
USER root
RUN rm -f /usr/local/bin/claude \
    && npm install -g @anthropic-ai/claude-code \
    && ln -sf "$(npm config get prefix)/bin/claude" /usr/local/bin/claude \
    && chmod 755 /usr/local/bin/claude
USER sandbox
ENTRYPOINT ["/bin/bash"]

Or it could be a completely fresh image.

And then the sandbox can be started with

openshell sandbox create --from Dockerfile.claude-npm