Skip to content

0.7.3 - honest failures, busy models

Choose a tag to compare

@unclecode unclecode released this 23 Aug 04:52
· 27 commits to main since this release

[0.7.3] - 2026-08-23

Fixed

  • A failed probe printed the first status number it could find, so a probe
    whose first call succeeded and whose second timed out reported
    probe-failed: 200. Failures now name the call that actually failed.
  • Failure text is now written for a human: model busy (rate limited),
    model overloaded (503), model too slow (no answer in time),
    key rejected (401).

Added

  • A busy provider is no longer treated as evidence. Cells that failed because
    the model was overloaded are marked, left out of the match colouring, and
    removed from both sides of the verdict score. The card says how many probes
    were skipped. Before this, two unrelated models that were both busy counted
    as a match.
  • A running lane reports a long wait every 3 seconds: waiting 33s, this model is busy. A lane stuck on one slow call used to look frozen.
  • SECURITY.md, CONTRIBUTING.md, CHANGELOG.md, and a VERSION file with
    bump.mjs so the page can no longer disagree with the release.