Skip to content

AI Integrator v0.1.4

Choose a tag to compare

@github-actions github-actions released this 17 Aug 20:20
· 17 commits to main since this release
1b1a08f

v0.1.4 focuses on upgrades and features in the browser, which still needs work, as well as a few other features.

The browser

  • Tabs remember the chat. Come back and the addresses are still there, asleep, until you or an agent actually open one. A sleeping tab is a row, not a page load.
  • A tab belongs to a task, and to whoever is driving it. One task cannot see another's tabs. A subagent gets the browser tools for the first time, on tabs it opened or was handed, and cannot pass them on. If you are typing in a tab, the agent is refused and the tab says it is held by you.
  • Pop-out is a real Integrator window. Theme, address field, history, annotate, capture. Several tabs can share it. Close it and they come home to the pane instead of vanishing.
  • Saved logins live in the OS keychain. Per project, bound to one origin. An agent can cause a login. It can never read a password, and from fill until the page navigates the tab will not snapshot or evaluate.
  • Dialogs, selects, drag, and wait_for. Dialogs no longer freeze the tab. Space types a space. A select picks the matching option. HTML5 drag moves something. wait_for waits. Same-origin iframes are in the snapshot. Enter submits. Hover opens listener menus, and CSS-only menus are refused with a reason instead of a fake success.

Agents and subagents

  • Subagents can drive a browser tab they own. Siblings do not trip over each other. The orchestrator can hand a tab over read-only or to drive.
  • Live tabs sit as a deck of cards. The raised one is the real webview. browser_focus brings one forward and says whether it landed.
  • A page an agent is working off-screen still measures at 1280×800 and says so, instead of reporting a 2×2 lie.

Auto-review

When an agent tries to cross a boundary (shell, network, a file outside the workspace, a tool call) a reviewer answers before it happens.

  • Codex uses its own reviewer. Claude, Cursor, Grok, Kimi, Antigravity and the rest we review ourselves, through the permission seam they already raise.
  • The reviewer is a cheap, fast model, not the one doing the work. It has about ten seconds. Silence is never an allow: it asks you, or it denies, whichever you set.
  • Every decision lands in the transcript as a line you can read back: auto-approved, auto-denied, or asked you.

Fixes

  • Closing a pop-out no longer orphans the tab.
  • The divider is a single hairline with a centered grip. The page stays put while you drag.
  • Offscreen and sleeping tabs no longer answer "not open" when they are right there in the strip.
  • Gestures report the page after they land, not before.
  • Cookie listing returns names and flags, never values. Consent dialogs are declined, not accepted.