Use the k8s values file on Hawk by default, so sandbox pods get #72's requests

[bunny-baxter]

Problem

#72 added ephemeral-storage requests for the sandbox pods, but put them only in the k8s values file. Every task defaults to sandbox_backend="docker", and no Hawk config overrides it (bloom68 included). Hawk accepts the compose file by converting it to chart values, and compose has no way to express an ephemeral-storage request. So the requests never reached a pod.

Evidence from bloom68-alv3junt54xqrybd:

• The runner log shows Hawk loading /tmp/leanopenproblems_sandbox/0.1.11rc11/fc_488aade228ec/corpus/compose.yaml.
• Running Hawk's converter on that compose file gives both services limits/requests: memory: 32Gi and no ephemeral-storage request.
• PodReplacedError restarts continued through the run. Erdos23 alone restarted 4 times.

The comparator's runtimeClassName: CLUSTER_DEFAULT pin was dropped the same way. That run still landed on runc, but only because the deployment sets no sandbox_runtime_class_name.

Change

• sandbox_backend now defaults to None, which resolves to k8s when HAWK_JOB_ID is set and docker otherwise. An explicit value still wins.
  • Hawk's runner Job sets HAWK_JOB_ID unconditionally (api/helm_chart/templates/job.yaml), in the same env block as HAWK_RUNNER_PATCH_SANDBOX=true, the flag that turns on its docker→k8s conversion. hawk local sets neither and runs docker sandboxes as-is, so it keeps docker.
  • An earlier revision keyed on KUBERNETES_SERVICE_HOST. That would also have switched local-Docker runs from inside any other pod (a dev pod, a CI job), so it was changed after review.
• Applies to all six tasks, including apn_redteam_collatz and apn_math_software_test. Their helpers now branch on the resolved backend that get_sandbox_config returns.
• Tests and local runs are unaffected: they call get_compose_file, which asks for docker explicitly.
• New tests/test_sandbox_backend.py.

Testing

• uv run mypy: clean.
• The fast pytest suite: 202 passed.
• Built with the bloom68 args, apn_erdos resolves to ('docker', .../compose.yaml) by default, still docker with only KUBERNETES_SERVICE_HOST set, and ('k8s', .../values.yaml) with HAWK_JOB_ID set.
• Checked that Hawk's K8sSandboxEnvironmentValues model keeps the ephemeral-storage key when it parses and dumps the values file.

Before relying on it

• Watch for Pending sandbox pods on the first Hawk run. These requests (20Gi agent, 8Gi comparator) have never been scheduled on the cluster. Request ephemeral storage for the k8s sandbox pods #72 assumed a 200Gi data volume per node; if allocatable ephemeral storage is smaller, pods will sit in Pending.
• This makes Request ephemeral storage for the k8s sandbox pods #72's requests actually apply. It does not by itself prove disk pressure is what causes the pod replacements.
• Runs already in progress keep their current pods.

🤖 Generated with Claude Code

[edkins]

Good to merge if a test run comes out ok

[bunny-baxter]

Did some test runs using some easy OEIS problems, looks ok! https://viewer.hawk.hawkbench.com/eval-set/sandbox-values-smoke-gubh336kj77bspyc