Skip to content

0.3.0

Choose a tag to compare

@charlesnutter charlesnutter released this 24 Sep 04:28
· 52 commits to main since this release

Added

  • A Session section below the per-turn figures, collapsed by default:
    click its heading to open it. Collapsed, it still shows the session's
    generation speed (▸ Session · 14 turns 48.2 tok/s). Open, it
    shows generation tok/s with a trend of recent turns, TTFT median and
    worst, cache hit rate, how the time split between generating, waiting
    for the first token and everything else, the engine's own averages
    (MTP or draft acceptance, prefill rate) where the engine provides them,
    and retries. Only the current model's turns count; the heading says so
    when the model changed partway through. The session's tokens, context
    and cost are left to OpenCode's own sidebar.
  • Sub-agent roll-ups. A turn that started sub-agents adds a line to the
    per-turn figures (sub-agents 2 · 4,210 tok, then 38.10s and
    $0.012 on the rows below): their tokens and cost summed, and the time from the first starting to the last
    finishing. Rates are never combined across them. The Session section
    totals them in a sub-agents row, and its time split gives the real time
    sub-agents were running its own share instead of folding it into other.
  • On vllm, sglang, vllmmlx, aphrodite, lmdeploy, llamacpp,
    llamafile, splash and omlx, a turn whose sub-agent used the same
    engine is no longer declined: the engine's window is expected to hold the
    turn's requests and tokens plus the sub-agents'. Those figures then cover
    both, and the line says incl. sub-agents. Rate and TTFT stay the turn's
    own. Built and tested against captures; not yet run live.
  • background option (default on): shades each box one step above the
    sidebar's own background.

Changed

  • The sidebar is two boxes, the last turn and the session, each opened and
    closed independently by clicking its heading. Figures are laid out one
    per line as label and value; a figure with parts continues on the next
    line. Collapsed, a heading keeps one figure. The heading names the
    engine, not the model, which OpenCode already shows under the prompt.
  • mlx-serve no longer shows a non-streamed request's whole-request rate:
    it includes prefill, so it is not generation speed.
  • mtplx, koboldcpp and mlxserve are read when each step finishes
    streaming, not when OpenCode marks the step ended. For a step that calls
    tools, "ended" comes only after the tools have run, and a sub-agent on the
    same engine had replaced the step's figures by then, so the turn was
    declined with engine data skipped: overlapping requests. Measured on
    MTPLX: 194 tokens (the step's own) at the end of streaming, 163 (the
    sub-agent's) at "ended".
  • The history panel's headline rate is generation tok/s on the newest
    turn's model (tokens over streaming time, as in the Session box), instead
    of a mean of per-turn rates across every model.

Install: opencode plugin add @banburist/opencode-headsup