Skip to content

v0.2.2

Choose a tag to compare

@charlesnutter charlesnutter released this 23 Sep 03:54
· 85 commits to main since this release

Fixed

  • vllm, sglang, vllmmlx, aphrodite and lmdeploy rendered figures
    for several requests as one turn's whenever other requests reached the
    engine in the same window — a new session's title, an interrupted turn, a
    second tab. Measured: a 46-token answer showed 116135.1 tok/s over
    8594 tok. Such a turn now shows the universal line with
    engine data skipped: overlapping requests.
  • On engines that publish a request-duration histogram (sglang,
    vllmmlx, lmdeploy), the turn's total and engine-derived decode rate
    were averaged with any other request in the window. On vllm-mlx every
    OpenCode turn is joined by a rejected ~0s title request, which halved the
    total (0.32s against OpenCode's own 649ms). The engine's duration is
    now used only when exactly one was recorded; otherwise OpenCode's timing.
  • With several tabs open, a turn completing in one tab replaced every other
    tab's sidebar line with inference · —. The sidebar kept one line for the
    whole TUI; it now keeps one per session. The collapsed line had the same
    fault and now uses its own session's latest turn. A turn in one tab can no
    longer suppress another tab's line as out of order.
  • The history panel's legend said * rows were "engine-measured". Every row
    holds OpenCode's own figures; * only means the sidebar line for that turn
    came from the engine. The legend now says so.

Install: opencode plugin add @banburist/opencode-headsup