Skip to content

AskKit 0.7.2

Latest

Choose a tag to compare

@gregyuzik gregyuzik released this 10 Nov 01:41
· 2 commits to main since this release
Immutable release. Only release title and notes can be modified.
v0.7.2
f6796f7

Eleven fixes, and no change to the API. A chat keeps saving when one of its rows won't encode, a yes
given twice is carried out once, a title the clamp rewrote or a tool cut short still names its row,
and a reader comes back to the thread where they left it.

Fixed

  • A yes given again after one that failed off screen is carried out once, and its card says so.
    When a yes came to nothing after its chat had left the screen (New Chat in the same moment, or while
    the perform ran and then failed), AskConversationStore kept that failure and wrote it over the turn
    whenever the chat was read back. So a reader who opened the chat, said yes again and left while that
    yes ran found the card, on opening the chat once more, saying it hadn't worked and offering the
    action; the second yes's success found no turn waiting on it, the chat was saved as failed, and a tap
    carried the action out a second time. A new yes now retires what came of the one before it, and the
    chat read back shows the new yes under way, then what came of it.
  • A row that won't encode costs its answer that row, not every save of its chat. A host's row
    can decode and still not encode, such as one holding a Double that is NaN or infinite, which a
    default JSONEncoder refuses, or one whose own encode(to:) throws. The thread, a chat's filing and
    a proposal's outcome were each encoded whole, so from the turn that read such a row on, every save of
    the thread failed and a relaunch brought back only the turns before it, every yes to a proposal
    failed with its save, New Chat couldn't file the chat, and no earlier chat opened while it was on
    screen. Reads already salvaged row by row. Now a write that won't encode whole tries each row alone,
    writes the turns without the ones that fail, in matches and tiles alike, and logs how many: the
    turn keeps its question, prose, outcome and other rows, and an answer whose only tile went reads back
    with none. A thread whose rows all encode is written as before, and one that still won't encode fails
    as before.
  • A title names its row as the clamp showed it to the model. AskPrompt.clamp swaps a title's
    angle brackets for ‹ ›, makes its tabs and line breaks spaces and drops what a reader can't see, and
    the model names the row as it read it, but tile and mention matching compared the title as stored: a
    row titled <3 Love Songs>, shown as ‹3 Love Songs›, got no tile when the model chose it, and one
    whose title held a tab, a zero-width space, a soft hyphen or a left-to-right mark wasn't found in
    prose that named it, so a follow-up naming it could fail as ungrounded. Matching now reads each title
    both as stored and as the clamp shows it, short of a cut, through the clamp's own steps, in prose and
    in the model's tile titles alike, and AskEval counts a title it names that way as named.
  • A title the tool cut short names its row. A tool clamps each title, as the README's sample and
    the demo do with AskPrompt.clamp(title, AskPrompt.shortClamp), so the model sees a title past 60
    characters as its first 60 and an ellipsis, and names it so, as AskReply's tile guide asks it to.
    Matching compared that with the whole title: a pick or a superlative on such a row showed no tile,
    the turn's referent was empty, and a follow-up that named the row as it was shown failed as
    ungrounded. Now a tile title or a mention in prose that ends in an ellipsis, the clamp's "…" or three
    periods, names the row whose title opens with what comes before it and goes on past it, at whatever
    limit the tool clamped to, once that is 16 characters or more: fewer can be the model trailing off,
    as in "You could play…". An exact title still comes first, so a title that ends in an ellipsis is
    still that title. Next comes a row whose whole title is what comes before the ellipsis, so
    "Mission: Impossible…" names Mission: Impossible and not Mission: Impossible – Fallout, as a tile
    and in prose that trails off after it: a model does that far oftener than a tool's limit lands on
    the end of another row's title. Then a cut title, then a near one. Titles that open alike were
    shown cut alike: a tile title is the first of them read, and prose names them all.
  • A promptPrefix without a trailing newline ends its line. The engine appended a domain's prefix
    as it stood, so "Answer about my records." ran straight into the first of the domain's lines, or into
    the question's <user_request> fence, where every line already ended in a newline. A non-empty
    prefix that doesn't end in one now gets one in the prompt, after it, so what prewarm caches is
    still how the prompt opens.
  • A chat set aside beside a filing this build reads only in part reaches History. A chat opened
    from such a filing, asked on, then set aside because New Chat couldn't file it, was put back at the
    next launch only into its earlier filing, which is never written over while some of its turns don't
    decode. So it was never filed, stayed out of History, and was tried again at every launch. It is now
    filed beside that filing, as New Chat files a chat beside one, and the filing stays as it was. An
    original kept for a build that reads more still waits for its chat's filing to read whole.
  • A turn draws "Earlier in this chat is summarized" only once a model was sent the recap. The
    engine reported AskTurnEvent.summarized as it built a fresh session's recap, before the session,
    the overview and the domain's prompt lines were in hand, so a turn that failed on any of them, such
    as a session the seam couldn't build, drew the hairline and kept it on disk, and one handed down
    from such a failure reported it twice. It is now reported just before the prompt that carries the
    recap is sent, still ahead of the answer's first words.
  • A bank of tiles with no column maximum lays out under an unbounded width. An AskTileGrid given
    columns: .adaptive(minimum: 76, maximum: .infinity), the maximum GridItem.adaptive defaults to,
    put its first tile at an x of NaN and handed its parent a width that wasn't a number wherever it was
    asked how wide it could be: in a .fixedSize(), a horizontal ScrollView or a ViewThatFits, the
    non-finite geometry behind LayoutSubview.place's assertion. AskKit's own columns are finite, so only
    a host's could. Asked with an unbounded width, such a bank now answers with every tile in one row at
    the minimum, the one finite width its columns name, and lays out that row when placed at it; a bank
    with a maximum answers at the maximum, as before.
  • A scope with no words reads as no scope. A domain whose evidenceScope(from:) returned "" or
    whitespace got the evidence line "From 3 items in your ", and an AskActivity with a scope like it
    got the wait row "Searching your …" or "Reading …". Each now reads as it does with no scope: "From 3
    items", and "Thinking…".
  • The Mac menu's Send and Stop follow a disabled composer. .disabled reaches a composer's controls
    and not what it hands the menu bar, so on the can't-serve first screen, whose composer is disabled,
    File > Send (⌘Return) stayed enabled whenever the field held a draft, such as a host's
    prefill(_:send: false) or words typed before Apple Intelligence went off, and sending put a failed
    thread where that screen had been. The composer now offers the menu neither Send nor Stop while it is
    disabled, as its circle offers neither.
  • Coming back to the thread leaves the reader where they were. A reader who scrolled up to an
    earlier answer and opened one of its tiles in a pushed row, or went to another tab, came back to the
    thread gliding to its foot, the answer they had been reading gone: the scroll that takes a new
    question into view ran again each time the thread reappeared, and took the reader back down. It now
    runs only for a last turn the screen hasn't seen arrive, the thread's first appearance and a new
    question, so a restored conversation and a screen built mid-answer still open at the end, and for a
    reader the thread was following down a streaming answer as it went, who comes back to the end of it
    whether or not it finished meanwhile. Anyone else comes back to where they left the thread, with the
    pill where it was, a card they had opened in the last answer included.