1.0.8
published 9 November 2026
superseded by 1.0.9 · current release 1.0.12
Computer use acts on what the page hides: a target scrolled out of view is scrolled into view and acted on in one act, after the session's claim and the challenge check, on macOS and on Windows, where a row hidden by its pane and a window taller than the screen are now seen; the sequence state line tells the truth about the whole sequence; the side jobs' effort floor follows the words a provider serves; the word sample reaches the Samples feature; /usage shows the TensorFold engine a session runs on; and four text scanners that walked a long digit run quadratically are linear.
Added
- Added a one-line reminder to the model when the first message you type in a session says "sample" or "samples": what a sample is, how it is made in an Eval cell, and whether Samples is on in the session (again after a compaction); the Eval tool's roster words name sample() when Samples is on, and the Boot Menu row's off words carry the one exception to "identical to a build without samples".
- Added a TensorFold slot to /usage's Local models block when the session's model is one TensorFold serves: the loaded model, its address, served window and memory in use, serving, draining, loading or not served, an idle mark until the first turn, and the ledger lines under it; a rented TensorFold pod keeps its ledger lines; a session on another local server shows no slot.
Changed
- Changed the side jobs' effort floor: low wherever the provider serves it, otherwise the served word nearest to low, the cheaper of two equally near, so a Mistral or Cohere list that serves only none and high runs titles, summaries and judges at none; the Cohere wire now sends that word (it sent high whenever thinking was on).
- Changed the Computer tool's edge rule: a target a browser clamps to a strip at a pane's edge (an element scrolled out of view) is scrolled into view and acted on in one act, after the session's claim and the challenge check, and the result says scrolled into view first; an element still hidden after the scroll, or a pane that will not scroll, is refused with the next move named; a human-verification checkbox scrolled below the page is a challenge like a visible one; identical siblings resolve apart by their order in the window and keep separate refs; a find that matches nothing keeps your refs; a ref asks for a fresh find once its list has changed.
- Changed the Computer tool's consent to one ask per session: the first act of the session shows the card, naming the application in front and the act, and Yes now covers every application for the rest of the session (a resume keeps it, a new session asks again) instead of that one application; the grant is recorded at the Yes, so the allowed act never asks twice; and when answering the card brought your own terminal forward (macOS), the act or read brings the application back in front and runs, saying so in its receipt, instead of refusing as a changed window.
- Changed wait with until to wait 5 s by default (a plain pause keeps 1 s) and to say how long the condition took.
- Changed the Computer tool's words: fill and fields are taught; a role shown with its subrole answers to either word; an element without a title is matched by its value; focusApp takes capture, settleMs and label and returns a screenshot after the switch.
- Changed four text scanners that walked a long digit run quadratically (the terminal's mouse-fragment and orphan-probe regexes, the capacity wait sentence, the TensorFold refusal reader) to linear forms with the same answers on every input, removed three casts that erased a type the module already had, and merged two repeated bodies into the helper they repeated; nothing a user or a model sees changes.
Fixed
- Fixed the crew view's slash commands: with a crewmate opened in its view (or pinned as the main chat), /model, /effort, /compact, /context and /usage and the model-picker key act on that crewmate (its model and effort change at its next turn, or at its next resume once it has finished, and the status row shows the change waiting beside the running model; its transcript compacts the same way; its own numbers show), the main chat's model, effort, thinking and transcript stay as they were, and every other command runs for the whole session and says so (typing /model at a selected crewmate changed the main chat's model).
- Fixed window screenshots on Retina displays: the capture of a window was taken from twice the window's offset and twice its size (a window at y 122 was pictured from y 244, so a click mapped from its screenshot landed 122 pt above the target); full-display screenshots were right, and a window at the top of the screen hid the fault.
- Fixed two misses of that rule on Windows: a row scrolled out of a page's inner pane, which Windows lists with its full layout rectangle and no pane between it and the page, now counts as out of view by the window's own off-screen report (find and tree mark it scrolled out of view), is scrolled into view and clicked where it lands, and a row cut by its pane's edge is clicked at the point Windows reports inside the part on view; and a target whose frame lies inside the window but past the screen's working area (a window taller than the screen, on any platform; the taskbar's strip included, unless the window covers its whole display) is refused with the display's edge named instead of a click the taskbar takes; a window spanning two displays is on view on both where windows span displays; a refusal after a scroll says whether the target moved; a scroll pointed past the screen wheels at the window's nearest point on it.
- Fixed the Windows tree's values and the switch role: text fields, checkboxes, sliders and selections read their value (every control read an empty one), and a switch reads as AXCheckBox (AXSwitch) as it does on macOS.
- Fixed the sequence state line: it compares the screen before the first step with the screen after the last and carries any step's change (a sequence ending in a screenshot said "not compared · focus not read"; one ending in a wait said "unchanged" after earlier steps had changed the window); a step that scrolled its target into view no longer hides a focus the click then moved).
- Fixed the step after a screenshot step in a sequence acting through the previous screenshot's coordinates while the screenshot's line said they now referred to the new window: a pixel step after an in-sequence screenshot is now refused with the reason (an image the model has not seen; take the screenshot alone, then act); a failed sequence returns no image (an error answer travels as text alone), so its screenshot line says the image was not returned and the previous screenshot's coordinates still apply.
- Fixed a drag landing one step short of its receipt: a window drag anchored at the first drag event that moved and lost that step (103 of 110 points), so a drag now nudges a twentieth of a point before it steps and the window lands where the receipt says, on Retina displays too.