extension.js · part 23

60 text occurrences from desktop/Cursor.app/Contents/Resources/app/extensions/cursor-computer-use/dist/extension.js, part 23. Every entry preserves the shipped literal and its saved verdict or selection reason.

File contents and all parts · All files

no change after action

not run

${n+1}. not run: ${e.step}

${e.actions} action${1===e.actions?"":"s"}, ${e.calls} decision${1===e.calls?"":"s"}, ${e.ms} ms

${n+1}. done (${t}, goal_reached=${UA(e.goal_reached)}): ${e.step}

${n+1}. failed: ${e.reason} (${t}): ${e.step}

 Then call computer_attempt with the remaining steps: ${s.join("; ")}

Step ${a} acted but was not confirmed. Check now: if it already shows the result, step ${a} is done; otherwise do step ${a} yourself with the element tools.${o}

Step ${a} was not tried; nothing changed. Do step ${a} yourself with the element tools.${o}

ERR attempt: ${t}/${e.steps.length} done, step ${a+1} failed: ${s.reason}

OK attempt: ${t}/${e.steps.length} steps done

... ${i} earlier line${1===i?"":"s"}

now: ${e.presence.join("; ")}

To act by element_id, read computer_app_state first.

could not read the app state (${xA(e)})

not run

not run

${t.reason} at iteration ${t.at}

${e}: ${JSON.stringify(fA(a))}

in ${e.role} ${JSON.stringify(fA(e.attributes.name??e.attributes.description??""))}: 

... ${o.length-CA} more

Press a key or chord on the Windows keyboard, e.g. "Return", "Escape", "Tab", "ctrl+l", "win+r". Join modifiers with "+" (ctrl, alt, shift, win; cmd, super and meta also mean win). A modifier alone taps that key: "alt" taps Alt (Office KeyTips and the classic menu bar; follow with the tab letter, "alt h", or a second call); "shift" and "ctrl" tap those keys. Key names: space, backspace, delete (forward), insert, left, right, up, down, home, end, pageup, pagedown, f1–f12, apps (context menu), printscreen, capslock, numlock, scrolllock, pause, volumeup, volumedown, volumemute, mediaplaypause, medianext, mediaprev, mediastop, or any single character. Separate independent presses with a space ("win r" taps Win, then r).

call the same tool again, correcting any argument the message names; nothing was actuated

stop and tell the user what the message says; only the person at the machine can clear it

run the tool the reason names first (usually computer_screenshot), then repeat the action

end computer use for this turn; call no more input or control tools

Screenshots use a fixed ${ZA} canvas. Coordinates are pixels in the most recent computer_screenshot image, with (0, 0) at the top-left; x grows rightward and y grows downward. Valid coordinates are x=0..${XA.width-1} and y=0..${XA.height-1}.

 When something is too small to read or to aim at, call computer_zoom on that region; with the zoom_id it returns, an action's x and y are pixels of that zoom image, whose size the zoom result quotes (an edge can reach ${Bn} pixels, so a zoom y may exceed the canvas height).

${["A refused call returns isError=true with structuredContent {code, message, escalation: {recommended, reason}}; act on escalation.recommended and read the reason for the next step.",`Escalations, with the codes that carry them by default — ${Ta.map(e=>{const n=fa.filter(n=>Ia[n]===e);return`${e} (${pc[e]}): ${n.join(", ")}`}).join("; ")}.`].join(" ")}

A refused call returns isError=true with structuredContent {code, message, escalation: {recommended, reason}}; act on escalation.recommended and read the reason for the next step.

Escalations, with the codes that carry them by default — ${Ta.map(e=>{const n=fa.filter(n=>Ia[n]===e);return`${e} (${pc[e]}): ${n.join(", ")}`}).join("; ")}.

${e} (${pc[e]}): ${n.join(", ")}

These tools control the local Windows desktop: they read the real screen and send real mouse and keyboard input. No permission has to be checked or granted first; begin with computer_screenshot.

- Do not call Computer Use tools in parallel; they act on one machine and run one at a time.

- Look before you click: screenshot before your first coordinate action, rely on the fresh screenshot every input action returns, and take another before retrying anything unexpected.

- Windows cannot confirm an injected event was delivered: a changed screen is evidence something happened, not proof it was this action, and no visible change is not failure.

- Anything under about 15 canvas pixels (a close X, a menu item, a checkbox, a taskbar icon) needs computer_zoom first; aim with its zoom_id, and never retry a missed click by nudging the coordinates.

- App launches, page loads and dialogs outlast the screenshot an action returns. If little changed after a click or key meant to launch, navigate or open a dialog (a taskbar icon, a Start tile, Return in an address bar, a context-menu item such as Save As or Print), call computer_wait with 1000–3000 ms and read that screenshot before repeating the action or trying another route.

- "win" is the Windows key: "win" opens Start, "win+r" opens Run, and ctrl+esc opens Start too; cmd, super and meta also mean the Windows key here, so "cmd+l" is Win+L and locks the workstation.

- To read down a page, use one computer_scroll of 10–15 ticks (one wheel notch each) rather than several small scrolls.

- Save As dialogs open with the file name selected: type the new name, then press alt+s.

- Switch apps with the app's taskbar button (bottom of the canvas) or alt+tab; a script cannot bring a window to the front.

- Escape closes most consent and notification overlays; if one survives two attempts, read or scroll the page behind it instead.

- PowerPoint: review many slides with View → Slide Sorter rather than PageDown per slide; select an object and press Delete to remove it; to edit notes, click the notes pane, ctrl+a, then type.

Coordinates are pixels in the fixed ${ZA} screenshot canvas, origin (0, 0) at the top-left.

Returns a screenshot taken after ${e.returns}.

${s} is ${e[s]}, but with zoom_id coordinates are pixels of that zoom image, which is at most ${vc.width}×${vc.height} (${s}=0..${r}). The computer_zoom result quotes its actual size.

${s} is ${e[s]}, but without zoom_id coordinates are pixels of the ${ZA} screenshot canvas (${s}=0..${r}). To aim in a zoom image, pass its zoom_id.

(x1, y1) must be the top-left corner and (x2, y2) the bottom-right: x2 > x1 and y2 > y1.

The region is ${t}×${a}; with ${nc}% padding it does not fit the ${Bn} px output edge and the ${yn}-patch budget at the screenshot's own resolution, so it would come back no sharper than the screenshot. The long edge may be up to ${Sn} canvas pixels at any aspect when the other edge is modest; zoom into a smaller area, or read the screenshot directly.

Required with an app target; pass the coordinate_token from the latest computer_screenshot of this app.

element_id and snapshot_id go together; pass both from the same computer_app_state result, or neither.

Pass either from and to, or path — not both.

Required unless path is given: pass from and to, or a path of 2–64 points.

Typing into an element by element_id takes no coordinate_token.

${e.steps.length} steps; a batch may have at most ${Dn} when any step is computer_click, computer_drag, computer_scroll or computer_move (${Mn} when every step is computer_key or computer_type). Split it into several batches and look at the screen between them.

Pass exactly one of pid, bundle_id, app_path, or app_name.

 With zoom_id, a pixel of that zoom image instead (its size is quoted in the computer_zoom result; never more than ${Bn}).

${e} pixel coordinate in the ${ZA} screenshot canvas.

Drag ${e} point.