Remote Rollout Processor, Raise Exception from its Logs, and new Vercel default server

[xzrderek]

Note

Introduce EP exception types and mapping, use them to raise errors from remote logs, expand Status error helpers, update retry logic, and replace the Vercel server with an async Flask app; add tests and config updates.

• Core:
  • Exceptions: Add eval_protocol/exceptions.py with EP exception classes mapped to gRPC/status codes and exception_for_status_code().
  • Status API: Expand Status with many rollout_*_error helpers and _build_details_with_extra_info; route rollout_error() to internal_error().
• Remote Execution:
  • RemoteRolloutProcessor: Use exception_for_status_code to raise mapped exceptions from Fireworks logs; increase /init timeout to 300s; poll searches via asyncio.to_thread; use rollout_deadline_exceeded_error on timeout.
  • Retry Config: Include EP exceptions in DEFAULT_RETRYABLE_EXCEPTIONS.
  • Tracing Utils: Set rollout_not_found_error/rollout_internal_error for missing/invalid remote data.
  • Eval Utils: Use str(e) for non-retryable error messages.
• Quickstart Server:
  • Replace BaseHTTP server with async Flask app (/api/init) that returns immediately and runs rollout in background, logging Status codes; map OpenAI errors to specific Status helpers; add CORS.
  • Update requirements.txt (add Flask[async], bump eval_protocol) and vercel.json (rewrite all to /api/init).
• Tests:
  • Add tests/test_exceptions.py for mapping/behavior.
  • Update remote server test to auto-start server and assert status propagation.

^{Written by Cursor Bugbot for commit d63e109. This will update automatically on new commits. Configure here.}