Surface remote eval errors instead of silently failing

[shfunc]

Note

Medium Risk
Behavior changes from best-effort to raising on remote API failures, which may break existing callers that relied on silent fallback but improves correctness and debuggability.

Overview
Remote hud eval --remote now treats remote submission as fail-fast: it captures the trace_ids returned by submit_rollouts() and raises an error if none are accepted, rather than always reporting success.

_send_job_enter() in hud/eval/manager.py was tightened to surface backend failures by removing the broad try/except fallback and instead calling raise_for_status() and raising a ValueError on unexpected responses (still returning None only when telemetry is disabled/no API key).

^{Written by Cursor Bugbot for commit 9980d62. This will update automatically on new commits. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{Bugbot Autofix is OFF. To automatically fix reported issues with Cloud Agents, enable Autofix in the Cursor dashboard.}

[cursor]

Raises on valid responses missing task_version_ids

High Severity

_send_job_enter now unconditionally raises ValueError when the API response doesn't contain task_version_ids as a list of strings (line 114). Previously it returned None in that case, which callers handle gracefully. When called from run_eval's parallel path without a taskset, the server likely returns a success response without task_version_ids, and the new code will crash with "Job registration failed: unexpected response" even though registration succeeded. This breaks all parallel evals that don't use tasksets.

 

[jdchawla29]

ignore