You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
[Bug]: Keyboard is taken over when using Orca browser skill #8139
When using orca browser skill, if the agent is trying to type something, it takes over my keyboard so I cant type and multi task. My keyboard ends up interfering with its task as well (ofc as its vice versa)
Reviewed computer-use / browser skill keyboard injection paths at a high level.
Could not run a full agent browser-skill session that synthesizes key events while measuring host keyboard steal without hijacking the interactive desktop of this machine.
Result
Needs an isolated computer-use automation that asserts keyboard focus ownership during agent typing. Not reproduced here.
Re: the cannot_repro above. I think the repro failed because the search was aimed at the typing path. In my measurements typing is not the trigger — synthetic mouse presses are.
Input.dispatchMouseEvent with mousePressed makes Chromium hand native focus to the guest view that received the press, so an agent clicking in a background tab pulls the caret out of whatever pane the user is typing in.
Method (no interactive-desktop hijack needed)
Drive a scratch tab with orca CLI on a loop while a human types continuously in a terminal pane in the same window. Each segment writes execution evidence into the page (a click counter and the final input value) so a silently-skipped command can't be mistaken for a clean run.
Measured on 1.4.158 / macOS, against a local file:// page — no app under test required:
Command
Commands issued
Caret in the user's pane
snapshot / get / is / network
32
held
eval (DOM .click() + value set)
224 @ 1.5/s
held
fill
37 + 38
held
click
39 + 38
lost
Order-reversed crossover
Running one command type per segment and then reversing the order rules out a warm-up or time-of-day confound:
Run
Order
click segment
fill segment
1
click → fill
caret lost
caret held
2
fill → click
caret lost
caret held
Same result per command regardless of position.
Why the other paths are fine
The two neighbouring input paths already compensate, which is why typing looked innocent:
Runtime.evaluate is deliberately excluded from auto-focus — the comment in cdp-ws-proxy.ts calls out this exact failure mode ("focusing on every eval steals the user's foreground window while background automation is running")
Input.dispatchMouseEvent had no equivalent.
Fix
Opened #10875, which captures the focused host WebContents around the synthetic mouse dispatch and hands focus back. Applied to click, drag and check. DOM focus still follows the click, so click → fill flows are unaffected.
Caveat on scope
I only have evidence for the mouse path, so I marked that PR Related to rather than Fixes this issue. The original report mentions typing; type / inserttext go through Input.insertText, which still calls webContents.focus() by design and which I did not measure. If the reporter's symptom persists after #10875 with typing alone, that path is worth a separate look.
Same on Orca 1.4.199, macOS 26.6, Claude Code agent.
While one worktree's session drives the embedded browser with the orca CLI, typing a prompt in a different session's terminal is unusable: the prompt input loses focus on every browser action, so you have to click back into the input and continue, then lose it again a moment later. Effectively nothing else can be done in Orca while any session is using the browser. Likely the same root cause as #19297 and the keystroke side of #12912.
Fixed by #28452, #28453 and #28454 (tracked in #20595), now on main: agent type, keypress, fill and click reach the targeted page and no longer land in, or take focus from, whatever you're typing in. Details and remaining limits are in #20595. Please reopen with your OS and the exact command if you still see it on the next release.
Activity
nwparker commentedon Jul 27, 2026
Attempted repro — cannot_repro
Tried
Result
Needs an isolated computer-use automation that asserts keyboard focus ownership during agent typing. Not reproduced here.
grboy6770 commentedon Jul 27, 2026
Reproduced — the trigger is
click, not typingRe: the
cannot_reproabove. I think the repro failed because the search was aimed at the typing path. In my measurements typing is not the trigger — synthetic mouse presses are.Input.dispatchMouseEventwithmousePressedmakes Chromium hand native focus to the guest view that received the press, so an agent clicking in a background tab pulls the caret out of whatever pane the user is typing in.Method (no interactive-desktop hijack needed)
Drive a scratch tab with
orcaCLI on a loop while a human types continuously in a terminal pane in the same window. Each segment writes execution evidence into the page (a click counter and the final input value) so a silently-skipped command can't be mistaken for a clean run.Measured on 1.4.158 / macOS, against a local
file://page — no app under test required:snapshot/get/is/networkeval(DOM.click()+ value set)fillclickOrder-reversed crossover
Running one command type per segment and then reversing the order rules out a warm-up or time-of-day confound:
Same result per command regardless of position.
Why the other paths are fine
The two neighbouring input paths already compensate, which is why typing looked innocent:
Input.insertTextreplays the pendingDOM.focus(fix(browser): restore programmatic focus for CDP text insertion #7041)Runtime.evaluateis deliberately excluded from auto-focus — the comment incdp-ws-proxy.tscalls out this exact failure mode ("focusing on every eval steals the user's foreground window while background automation is running")Input.dispatchMouseEventhad no equivalent.Fix
Opened #10875, which captures the focused host
WebContentsaround the synthetic mouse dispatch and hands focus back. Applied toclick,dragandcheck. DOM focus still follows the click, soclick→fillflows are unaffected.Caveat on scope
I only have evidence for the mouse path, so I marked that PR Related to rather than Fixes this issue. The original report mentions typing;
type/inserttextgo throughInput.insertText, which still callswebContents.focus()by design and which I did not measure. If the reporter's symptom persists after #10875 with typing alone, that path is worth a separate look.AmethystLiang commentedon Jul 27, 2026
Thanks for reporting @ryanmargono and @grboy6770 for the fix PR! Will try to land it soon
paul-vd commentedon Sep 10, 2026
Same on Orca 1.4.199, macOS 26.6, Claude Code agent.
While one worktree's session drives the embedded browser with the
orcaCLI, typing a prompt in a different session's terminal is unusable: the prompt input loses focus on every browser action, so you have to click back into the input and continue, then lose it again a moment later. Effectively nothing else can be done in Orca while any session is using the browser. Likely the same root cause as #19297 and the keystroke side of #12912.nwparker commentedon Oct 15, 2026
Fixed by #28452, #28453 and #28454 (tracked in #20595), now on
main: agenttype,keypress,fillandclickreach the targeted page and no longer land in, or take focus from, whatever you're typing in. Details and remaining limits are in #20595. Please reopen with your OS and the exact command if you still see it on the next release.