bug(sandbox): arm64 remote cluster fails on x86_64 host without DOCKER_PLATFORM and missing default policy file #5

Summary

Creating a sandbox against an arm64 remote host from an x86_64 machine fails unless DOCKER_PLATFORM=aarch64 is set manually. After using that workaround, sandbox startup can still fail because dev-sandbox-policy.yaml is missing.

Actual behavior

• navigator sandbox create --remote spark.local -- claude prompts to create a cluster, then cluster startup fails with:
  • cluster container is not running while waiting for kubeconfig
  • exec /usr/local/bin/cluster-entrypoint.sh: exec format error
• With DOCKER_PLATFORM=aarch64, cluster creation succeeds, but sandbox startup later fails with:
  • failed to read sandbox policy from dev-sandbox-policy.yaml
  • No such file or directory (os error 2)

Expected behavior

• Navigator should detect/handle host vs target architecture automatically when creating remote clusters, without requiring users to set DOCKER_PLATFORM manually.
• If dev-sandbox-policy.yaml is required, it should be generated automatically (or a sensible default should be used) so sandbox startup does not fail.

Reproduction

1. Use an x86_64 machine.
2. Run navigator sandbox create --remote spark.local -- claude against a spark host (arm64).
3. Observe cluster creation failure with exec format error.
4. Retry with DOCKER_PLATFORM=aarch64 navigator sandbox create --remote spark.local -- claude.
5. Observe subsequent sandbox policy file error (dev-sandbox-policy.yaml missing).

Environment

• Local host arch: x86_64
• Remote host arch: arm64 (spark)
• Command: navigator sandbox create --remote spark.local -- claude

Notes

This may be easy to miss on macOS because architecture assumptions may differ there.

Originally by @drew on 2026-02-10T17:32:13.224-08:00