v0.1.8 — No more white page

AI Integrator v0.1.8 — No more white page

Small follow-up to 0.1.7, from the first round of hands-on testing.

The page never goes white

A native browser page paints above every piece of the app's own HTML. So when something has to be drawn over its slot — the compact deck being dragged, the address field's suggestions, a dialog — the page has to stand down for a moment and its still shows in its place. Until now that still was taken once, at load; if that capture missed (it did on some Macs), the slot went white.

• The surface now takes a fresh still the moment the address field takes focus, so typing shows the page as it is under the suggestions.
• While the page stands down it is re-captured every couple of seconds (browser_tab_poster_refresh, from the tab's own camera, parked or not), so a page that keeps moving reads as near-live rather than frozen.

The deck goes anywhere

The compact deck no longer glides away from the page when you drop it there. It rests wherever you put it, the browser pane included; while it sits on the page, the page stands down under it (still refreshing) and comes back live the moment you move the deck off.

Notes

• Everything from the v0.1.6 and v0.1.7 notes still applies.

v0.1.7 — Fixes for 0.1.6, plus the work that was in flight

AI Integrator v0.1.7 — Fixes for 0.1.6, plus the work that was in flight

v0.1.6 shipped from a clean branch while several other sessions had unfinished work in the main checkout. This release merges that work in, and fixes the first things found on macOS with 0.1.6.

Fixes to the browser windows (0.1.6 regressions)

• Clicking a tab in a browser window switches to it again. The new drag machine captured the pointer on press, which retargeted the release and ate the tab button's click. Capture now starts only once a drag has actually begun.
• X always closes. The "recent agent work" guard no longer turns X into a minimize — in the browser window and in the work pane. If you want a page to keep running, use – (to the pane, or to the compact browser).
• The group pill sits on the tab row, centred on the tabs' height, instead of floating mid-strip.
• CI on macOS was red on a Windows-only path assertion in a group-name test; it is now cfg(windows).

Merged in-flight work

• Tabs capture themselves. New webview-capture crate: WebView2 CapturePreview on Windows, WKWebView takeSnapshot on macOS. A covered, parked or off-screen tab now screenshots its own page rather than whatever is on top of it. Falls back to the old crop where unsupported. (macOS path is compiled by CI for the first time in this release; treat as new.)
• The embedded browser reads as a desktop browser. User agent derived from the installed engine major (Safari-style on macOS), client-hint brands aligned, webdriver=false, AutomationControlled off — sites like Google sign-in stop refusing the "embedded" token.
• Guest runtime robustness. Actions that finish later reply through a ticket the host polls (no more "malformed reply" / null for promise-shaped answers); a navigation mid-gesture answers page-navigated; a tab between pages answers guest-not-ready and browser_wait_for waits through it. Refs are stamped with the layout they were measured at and refused as stale after a resize; verbs describe their target before the gesture; bare-coordinate targets and role-only targets work; scrolling walks up to a box that scrolls. Agent replies drop favicon bytes.
• Schedules move between chats from the edit sheet; cancelled/completed schedules can be edited and re-run; only a running one is locked; dispatch no longer requires the main window.
• Claude Chat isolation without --safe-mode: --setting-sources "" plus settings/env that disable connectors and auto-memory, so Chat keeps its MCP tools.
• Layout-shift tracking for native pages: a page follows its slot while the sidebar slides, not just after.
• Composer route dropdown shows icon + caret at narrow widths.

Notes

• If favicons or browser_snapshot were failing for you on 0.1.6, both went through the same eval path the guest-runtime work rewrote; please retest on 0.1.7.
• Everything from the v0.1.6 notes still applies: grouped strips, per-project identity, shared popped-out tabs, draggable windows, restart persistence, Recently closed.

v0.1.6 — Browser windows, grouped

Popped-out tabs now live in real browser windows: grouped by project, restored after a restart, and draggable between windows and the pane.

Groups
Each popped-out tab sits under a coloured pill for its project (standalone chats share a Chat group). Click the pill to collapse the group; hover a tab to see the host and, if an agent is driving it, which one. Collapse is remembered per window.

One login per project
Cookies and saved logins follow the project, not the task. Sign in once and the next task in that project is already signed in. Old per-task identities are still listed under Settings → Browser if you need to clear them.

Popped out = shared with the project
Any agent on a task in that project can drive the project's popped-out tabs. Tabs still in the pane stay private to their task. Two agents cannot drive the same tab — the second is told who is holding it. Live-tab caps are 8 per group, 24 overall.

Windows
A window is not tied to a task. Mix groups, open as many as you want. Popping a tab out drops it in a window that already has that group when it can. Drag along the strip to reorder, off the strip to tear off a new window, onto another window to join it, or back onto the app to dock. Resize from the edges; double-click empty strip to maximize.

X closes a tab (recent agent work is still kept). Minus sends it back to the pane. Ctrl+T / Ctrl+W / Ctrl+Shift+T / Ctrl+Tab / Ctrl+1–9 work in the window; collapsed groups are skipped.

After a restart
Windows come back where you left them. Tabs stay asleep until you click them — nothing loads at boot. Tabs untouched for 7 days (or past 100 popped-out tabs) go to Recently closed. Right-click a group pill to reopen one.

Agents
browser_open can pop a tab out. New browser_pop_out and browser_dock. browser_list now includes group, window, who holds the tab, and (with includeRecent) that group's recently closed tabs.

Also
A scoped search could come back empty on macOS and some CI paths where the folder spelling did not match the index. That is fixed.

AI Integrator v0.1.5

The first of many security passes, and a lot more browser — with a lot more still to come.

Security — pass one, with Daybreak Blue from Codex

• Browser identity and profiles are properly scoped: chats share one bucket, every other task is isolated, and an unknown task gets a one-way identifier rather than falling into someone else's session.
• Hardening across the credential store and the saved-login vault. A password moves between the page and the OS store without ever passing through the app window.
• Tighter boundaries on what an agent may reach — which tabs it can address, which window may speak for which task, and which files it may hand back to you.

Popped-out browsing is a real browser window now

• Pop a page out of a chat and it joins one shared chat window; pop one out of a task and that task gets its own. No more one window per conversation.
• Tabs shrink to share the strip instead of scrolling off it, animate in and out, and answer the shortcuts you expect: Ctrl+T, Ctrl+W, Ctrl+Tab, Ctrl+1–9, middle-click to close.
• The window strip, its controls and its empty state were rebuilt in the app's own design language.

The new tab

• Browser chrome is there before you have anywhere to go: back, forward, reload and the address field, with the same action cluster a live page wears, so nothing shifts once the page arrives.
• The page itself was redesigned — brand lockup, a proper search field, bookmarks as tiles, recently opened and local servers as clean rows, all on a soft accent wash.
• Suggestions stay quiet until you actually type.
• A file or review opened in the work pane now wears the same colour as its tab, the way the browser surface already did.

Favicons

• Real sites get their real icons. The resolver now ranks candidates by nearness to the size a tab draws them at, walks a fallback chain when one fails, tolerates the mislabelled content types servers actually send, and falls back to /favicon.ico.
• Icons appear while the page is still loading, survive a restart with your remembered tabs, and fill in for bookmarks made before a site had shown one.

Agents and pictures

• An agent can photograph a popped-out tab: the window it lives in is raised first, so the picture is the page rather than whatever was covering it.
• Agents can show you what they saw — a screenshot can be dropped straight into a reply and it renders inline. Only this app's own captures are readable that way.
• The overflow menu is an OS menu now, so opening it no longer hides the page underneath.

Elsewhere

• Antigravity improved, and every fallback model list refreshed.
• Fewer main-thread stalls in the tab deck: the z-order fix no longer reparents every card on every placement.
• The tree is green again — formatting, the bridge-contract snapshot, two long-broken resume tests, and the maintainability ceilings.

Full Changelog: v0.1.41...v0.1.5

AI Integrator v0.1.4

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.

AI Integrator v0.1.3

A browser inside Integrator

This release is about one surface: a real browser that lives in the work pane,
that you and the agent drive together.

The browser

• A tab is a real webview, not a screenshot or an iframe. Pages behave like
  pages: sign-ins work, cookies persist between runs, and video and canvas
  render at full speed.
• Chrome that stays out of the way — history, an address field, annotate,
  screenshot, record, pop out, and open in your system browser. A popped-out
  tab keeps running and stays available to the agent.
• A blank tab starts where you work. It lists the local dev servers that
  are actually listening, probes each one once over HTTP, and offers the ones
  that answer first with everything else behind a single "show all".
• Annotate an element to send it to the chat: pick it on the page, write a
  note beside it, and it arrives in the composer as a chip rather than a block
  of markup in the middle of your draft.

For agents

Ten browser tools, task-scoped, refused politely when Settings says no:
browser_open, browser_list, browser_close, browser_navigate,
browser_snapshot, browser_click, browser_type, browser_press,
browser_scroll, browser_wait_for, browser_evaluate.

• Tabs belong to a task. One task cannot reach another's tabs, and delegated
  children cannot reach any.
• The agent has a cursor you can watch. It travels to each target, taps a
  ring where it lands, and captions what it did, then fades. Driving a page no
  longer looks like the page moving on its own.
• A tab an agent opens gets a pane tab of its own without stealing focus.
• Harness policy asks an agent to check page and local-server work in a tab you
  can see, acting on snapshot refs instead of guessed selectors.

Signing in

One browser profile on this computer, so sites stay signed in between runs and
an automation does not have to log in again. Settings lists the origins holding
state and can sign out everywhere. Integrator never sees or stores a
password — sign-in stays between you and the site.

The work pane

Files, review, subagents and browser tabs are peer tabs in one pane between the
transcript and the right rail. Closing a tab never stops the process behind it,
and the empty pane offers what can open there.

Fixes

• The work pane could animate out and stay mounted, holding its width while
  every control reported it closed.
• The browser no longer blinks: a tab follows its slot instead of being taken
  off screen and put back on every scroll or re-render, and it is positioned in
  one call rather than a move followed by a resize.
• Dragging the pane edge works with a page showing. The tab steps aside for the
  gesture instead of swallowing the pointer halfway through it.
• No more console window flashing over the app when a blank tab looks for dev
  servers.
• The window moves from the right side of the titlebar, not only the left.
• Project folder icons draw themselves instead of writing "undefined" into the
  path and rendering nothing.

AI Integrator v0.1.2

Full Changelog: v0.1.1...v0.1.2

AI Integrator v0.1.1

Full Changelog: v0.1.0...v0.1.1

AI Integrator v0.1.0

Full Changelog: https://github.com/lukekabbash/integrator-3/commits/v0.1.0