Skip to content

v0.2.4

Choose a tag to compare

@charlesnutter charlesnutter released this 23 Sep 21:55
· 71 commits to main since this release

Changed

  • mtplx is read at the end of every step, not once per turn, so a turn
    that calls tools shows engine figures for the whole turn: tokens summed,
    the rate over every step's decode time, TTFT and prefill from the first
    step, MTP acceptance weighted across steps. Each step's receipt must match
    OpenCode's count for that step; if one doesn't, the turn shows the
    universal line with the overlapping-requests notice.
  • koboldcpp and mlxserve are read the same way: at the end of every
    step, combined across the turn, each step checked against OpenCode's
    count. KoboldCpp's N generations this turn (last shown only) no longer
    appears on turns read step by step, since nothing is dropped. Built from
    live captures; not yet run step by step against a live server.
  • llamacpp, llamafile, splash and omlx check that a window is this
    turn's before showing it: its tokens must equal OpenCode's count for the
    turn and, for Splash and oMLX, its requests must equal the turn's steps.
    A window that fails shows the universal line with the overlapping-requests
    notice. Splash's N requests this turn no longer appears once the
    requests are accounted for. On a multi-step turn oMLX shows OpenCode's
    generation rate instead of the server's all-time average. Built from live
    captures; not yet run against a live server.
  • Every engine line's total is OpenCode's -- what you waited, retries named
    -- rather than an engine's own request or phase timings.

Install: opencode plugin add @banburist/opencode-headsup