Skip to content

v0.2.3

Choose a tag to compare

@charlesnutter charlesnutter released this 23 Sep 17:40
· 76 commits to main since this release

Changed

  • A turn that calls tools is now measured as the whole turn, not its last
    step. Tokens, cost and cache reuse are summed over every step; measured, a
    316-token, 37s turn had shown 140 tok 10.20s. Applies to the sidebar,
    the collapsed line and history. The engine line from mtplx, koboldcpp
    and mlxserve, which report the latest request, still describes a tool
    turn's last step; reading them per step follows in a later release.
  • Turn time is what you waited, from the request to the last step's end,
    and names any retries OpenCode made: 60.00s (6 retries).
  • tok/s is generation speed only: tokens over the time spent streaming
    after each step's first token, so tool execution and prefill never count.
    A turn that cannot be timed from its stream shows no rate, where it used to
    show a whole-turn figure labelled overall.

Fixed

  • Prometheus engines declined every tool-using turn: its window holds one
    request per step. One per step is now expected, with the engine's tokens
    checked against the turn's total.
  • On a single-step Prometheus turn the total was the engine's request
    duration, which excludes retries; it is now the total you waited.

Install: opencode plugin add @banburist/opencode-headsup