Skip to content

evals: harness hangs on provider retries far past per-test timeout #779

Description

@dcramer

Problem

Combined runs of eval files intermittently hang far past the per-test timeout on provider retries: observed test durations of 362s and 582s against a 120s it(...) timeout (vitest reports the test failed at timeout but teardown blocks on the in-flight provider retry loop). Also seen: ProviderRetryError: Delay was aborted. Single-case runs are unaffected.

Impact: full-suite eval runs flake on infrastructure rather than behavior (~1 in 3 combined runs during the #773 verification campaign), which erodes trust in eval signal.

Direction

  • Make provider retry/backoff abort promptly when the harness test context is cancelled (propagate AbortSignal through the retry loop).
  • Consider a harness-level provider-call deadline below the per-test timeout so a stuck call fails the case with a clear infra-signature error instead of a generic timeout.
  • policies/evals.md says to keep eval cases within 30s; the retry loop's worst case should respect that budget.

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugSomething isn't working

    Fields

    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions