Desktop: browser toolset next to the computer toolset #1

Merged
kubus merged 2 commits from t3code/browser-automation-use into main 2026-10-10 23:07:43 +00:00
Owner

Summary

  • Added browser_toolset_20260801 to desktop tasks alongside computer_toolset_20260801 for web page interactions
  • Browser toolset members include: navigate, read_page, find, form_input, left_click, get_page_text, screenshot, zoom, switch_tab, close_tab
  • Model reads pages as text via read_page and acts on elements by reference; takes screenshot only when needed for layout, images, or controls without refs
  • Added browserState() to report open tabs instead of always showing a screenshot at task start and after pauses
  • Browser element inspection checks policy before clicks/Enter, same as computer toolset
  • Halting: when a toolset call fails in a batch, remaining calls in that toolset don't execute
  • Removed Uptime Kuma confirmChecks waiting feature and kuma_waiting table; alerts start immediately
  • Updated PLAN.md (§5.5, §5.6, §6.1) with browser toolset flow and trust model

Testing

  • Browser navigation and page reading with filters (interactive, etc.)
  • Element location by reference and by viewport coordinate
  • Form input, tabs switching, and page scrolling
  • Browser screenshot capture to timeline
  • Policy check and approval for consequential browser clicks and Enter
  • Takeover and approval return to browser state instead of screenshot
  • Consequential click by element ref and coordinate
  • Failed toolset call halts remaining calls in that toolset this turn
  • Uptime Kuma alert without wait (confirmChecks removed)
  • Desktop task with browser and computer actions in one task
  • Not run: scheduling tests removed (confirm_checks references dropped)
## Summary - Added `browser_toolset_20260801` to desktop tasks alongside `computer_toolset_20260801` for web page interactions - Browser toolset members include: navigate, read_page, find, form_input, left_click, get_page_text, screenshot, zoom, switch_tab, close_tab - Model reads pages as text via `read_page` and acts on elements by reference; takes screenshot only when needed for layout, images, or controls without refs - Added `browserState()` to report open tabs instead of always showing a screenshot at task start and after pauses - Browser element inspection checks policy before clicks/Enter, same as computer toolset - Halting: when a toolset call fails in a batch, remaining calls in that toolset don't execute - Removed Uptime Kuma `confirmChecks` waiting feature and `kuma_waiting` table; alerts start immediately - Updated PLAN.md (§5.5, §5.6, §6.1) with browser toolset flow and trust model ## Testing - Browser navigation and page reading with filters (interactive, etc.) - Element location by reference and by viewport coordinate - Form input, tabs switching, and page scrolling - Browser screenshot capture to timeline - Policy check and approval for consequential browser clicks and Enter - Takeover and approval return to browser state instead of screenshot - Consequential click by element ref and coordinate - Failed toolset call halts remaining calls in that toolset this turn - Uptime Kuma alert without wait (`confirmChecks` removed) - Desktop task with browser and computer actions in one task - Not run: scheduling tests removed (confirm_checks references dropped)
The desktop-agent drives Chromium over CDP (/browser, /browser/state,
/browser/inspect): pages are read as text and acted on by element ref,
with screenshots only when needed. A task starts from the open tabs
instead of a screenshot. Browser clicks and Enter go through the same
consequential-step check. A task stops and asks when a consequential
step depends on a choice the task leaves open.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
kubus merged commit 162d6b043b into main 2026-10-10 23:07:43 +00:00
Sign in to join this conversation.
No reviewers
No labels
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
kubus/hatbot!1
No description provided.