[hemildesai]

Summary

• inspect OpenSandbox lifecycle status after an exec backend returns HTTP 502
• surface a typed OOM error with lifecycle context while preserving the original behavior for non-OOM failures
• add race-aware unit coverage for status reconciliation and non-OOM behavior

Example OOM message

SandboxBackendUnreachableError: Sandbox was OOM-killed. OpenSandbox status: state='Failed', reason='FAILED', message='1/1 observed pods failed; primary reason=OOMKilled; sample pod=<pod-name>; container sandbox terminated with OOMKilled (exit code 137)'; sandbox_id='<sandbox-id>'

Test plan

• pytest -q tests/unit_tests/test_opensandbox_provider.py (45 passed)
• scoped pre-commit hooks on both changed files
• local live E2E: in-flight exec returned 502, SDK reconciled to Failed / OOMKilled, Gym surfaced the OOM diagnosis, and cleanup succeeded

🤖 Generated with Claude Code

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[hemildesai]

/claude review

[hemildesai]

/ok to test 9e20c69

[hemildesai]

/claude review

[hemildesai]

/ok to test 42f732d

[claude]

SHIP — LGTM, no reliability concerns.

Focused, defensive change to OpenSandbox exec error classification. On a submission/polling 502 (or an existing SandboxBackendUnreachableError), it polls read-only get_info for up to 5s and, if the status reports an OOM-kill, raises a more specific SandboxBackendUnreachableError; otherwise it re-raises the original error unchanged.

Verified:

• No exec double-run risk. The added path only calls get_info (read-only status), never re-submits the command, so retry semantics are untouched.
• Error routing is correct. Non-502 / non-typed exceptions re-raise immediately (provider.py:1244-1245); the hard-cap TimeoutError (no status code) propagates as before. get_info absent or raising both fall through to re-raise the original error.
• Async is clean. Loop uses asyncio.get_running_loop().time(), await self._await_sdk_call(...), bounded per-call timeout_s=min(2.0, remaining_s), and await asyncio.sleep(...). No httpx, no ray.get(), no missing await.
• Tests assert real behavior — OOM classification, message truncation (<800 / [:500]), __cause__ chaining, poll count, and the non-OOM re-raise path.

Worst case is a bounded ~5s extra latency on a 502 when the status stays non-terminal — acceptable for the improved diagnosability.