Skip to content

Releases

0.5.0

Choose a tag to compare

@charlesnutter charlesnutter released this 25 Sep 23:55

Changed

  • The details dialog is one column with three tabs -- Turn, Session and
    History -- switched with tab, instead of two side-by-side columns. It is
    as tall as its content (up to most of the screen), fills the width of an
    xlarge dialog, and sits on a lighter panel inside a darker ring. Each
    section has a title and rule with a blank row under it; shares are drawn
    as square bars; and figures within a section sit on shared columns
    (Tokens' values, qualifiers and bars; the engine's figures in a
    three-column grid).
  • Tools by time on the Session tab always shows: the tools used, that none
    were, or that tool use was not recorded for older turns.
  • The Turn tab adds a timeline: each step's wait, generation and tools on
    one time scale. Its steps table shows the engine's own per-step rate
    beside OpenCode's where the engine is read per step.
  • ctrl+shift+h opens the dialog's History tab instead of the history
    panel. The tab lays each turn out in fixed columns, so a long row no
    longer wraps; s switches between this session and every session.
  • /headsup session and /headsup history open the dialog on those tabs.

Fixed

  • The History tab's headline says 1 turn, not 1 turns.

Install: opencode plugin add @banburist/opencode-headsup

0.4.0

Choose a tag to compare

@charlesnutter charlesnutter released this 25 Sep 19:38

Added

  • A details dialog, opened by details › under the sidebar boxes,
    ctrl+shift+d or /headsup, and closed the same way or with esc. Last
    turn on the left, session on the right; one at a time below 110 columns,
    switched with tab. The sidebar is unchanged.
    • Last turn: where the time went, in seconds and shares that add up to the
      total; every step with its tool calls and their times; tokens in all
      five kinds and context used; the engine's own figures only, marked ◆,
      including MTPLX acceptance at every depth and per-step engine rates;
      retry reasons in full.
    • Session: speed and time-to-first-token spread (min, median, p90, max),
      time split in seconds, tools by time, retries by reason, and coverage:
      how many turns had engine figures and why the others did not.
  • Compaction is named. OpenCode compacting the conversation mid-turn gets
    its own share of the turn's time, and the step it delayed says so. On an
    engine that publishes cumulative counters (Splash, llama.cpp, llamafile,
    the vLLM family), the compaction's own request is read before and after
    and taken out of the turn, so the turn keeps its engine figures; where it
    can't be, the reason reads compaction ran this turn.

Fixed

  • A sub-agent on the same counter engine as its parent no longer makes the
    parent's turn lose its engine figures. The sub-agent's own report moved
    the engine-wide starting reading, so the parent's window held only its
    last step and was declined as overlapping requests. Each turn now takes
    its own starting reading when it starts.
  • When engine figures were declined, the dialog says the engine's figures
    were left out and why, rather than reading as though the engine reported
    nothing.

Install: opencode plugin add @banburist/opencode-headsup

0.3.3

Choose a tag to compare

@charlesnutter charlesnutter released this 25 Sep 04:55

Fixed

  • Generation speed counts the time a model spends writing a tool call's
    arguments. Those tokens were counted but their streaming time was not, so
    a step that wrote a file showed several times its real speed (3,672 tokens
    "at 162.9 tok/s" against the engine's 36.4), and the Session average was
    inflated with it. Affects OpenCode's figures, not an engine's own. On
    MTPLX, a turn that wrote a file now gives 36.4 tok/s from OpenCode's
    figures against the engine's 36.1.

Install: opencode plugin add @banburist/opencode-headsup

0.3.2

Choose a tag to compare

@charlesnutter charlesnutter released this 25 Sep 03:41

Fixed

  • A reply is recorded when it ends, not when OpenCode's execution does. A
    message sent while a reply is running is queued into the same execution,
    so the first reply was never shown: a 19m 53s reply left the previous
    turn in the sidebar.
  • A reply that was interrupted or failed is shown, marked interrupted or
    failed, with OpenCode's figures up to that point, instead of leaving the
    previous turn in place. The history panel marks it too.

Install: opencode plugin add @banburist/opencode-headsup

0.3.1

Choose a tag to compare

@charlesnutter charlesnutter released this 25 Sep 00:13

Changed

  • The first turn after OpenCode starts can show engine figures on
    vllm, sglang, vllmmlx, aphrodite, lmdeploy, llamacpp,
    llamafile, splash and omlx, instead of engine telemetry from the next turn. When a turn starts on one of these engines with no reading
    yet, that one engine is read once. Nothing is read at startup, and no
    other engine is read. A new session that has not used or selected a
    model yet is not primed, as before.
  • A turn shown with OpenCode's own figures is headed by the engine's name,
    as its engine figures are (vllm-mlx, not vllmmlx).

Install: opencode plugin add @banburist/opencode-headsup

0.3.0

Choose a tag to compare

@charlesnutter charlesnutter released this 24 Sep 04:28

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

v0.2.4

Choose a tag to compare

@charlesnutter charlesnutter released this 23 Sep 21:55

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

v0.2.3

Choose a tag to compare

@charlesnutter charlesnutter released this 23 Sep 17:40

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

v0.2.2

Choose a tag to compare

@charlesnutter charlesnutter released this 23 Sep 03:54

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

v0.2.1

Choose a tag to compare

@charlesnutter charlesnutter released this 22 Sep 21:52

Fixed

  • The plugin crashed on load with ctx.storage.memory is not a function when
    installed from npm. package.json carried a main field, which OpenCode's
    loader falls back to for the server process, so a TUI-only plugin was
    being started there. It now exposes ./tui and nothing else, and the server
    skips it.

Install: opencode plugin add @banburist/opencode-headsup