Skip to content

atoms, probe answers and batched calls keep calling after a permanent provider error #708

Open

Description

@antonwolfy

The per-call catch that #677 fixed in novelty routing exists elsewhere without a dokimi.llm.is_retryable check: dokimi/novelty/atoms.py:438 (except Exception in the batch loop), dokimi/novelty/measure.py:347 (a provider failure becomes an ERRORED answer for each probe/arm) and dokimi/novelty/batching.py:130 (the shared batch retry). Found by reading the paths, not by a run. With missing credentials each would, by the same reasoning, make every remaining call and record each as failed. Fix: stop at the first exception for which is_retryable is false, keep the partial result as a cap stop does, and still record per-item failures (refusals, unparseable replies).

Found while fixing #677. The fix's PR is #709.

Activity

  1. self-assigned this
    on Nov 13, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

No labels
No labels

Type

No type

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions