Adom Desktop - Puppeteer Bridge

Shipped all 5 in pup runtime 1.8.23. AD relays your verb args verbatim, so there is zero AD-side change — new behavior + new args only. Deployed + acceptance-tested live on a Windows box (open → navigate → eval-fill → screenshot with no alert calls flashed once, never stole focus, and list_tabs reported the touched tab). Arg shapes below so you can update the AD-side hints + the consumer skill.

1. AUTO-FLASH on agent updates ✅

Mutating verbs now auto-flash the window's taskbar (real FlashWindowEx, FLASHW_TRAY|FLASHW_TIMERNOFG) so the user always knows the agent touched a window — without stealing focus, and it self-suppresses while that window is foreground (GetForegroundWindow()==h → no-op) and clears on click (Windows' native flash semantics).

• Debounced: one flash per window per ~5s burst. A tight open→navigate→eval→screenshot sequence flashes once, then re-arms after 5s of quiet. No strobing.
• Which verbs count as mutating: browser_open_window, browser_open_tab, browser_navigate, browser_reload, browser_back, browser_forward, browser_switch_tab, browser_click, browser_type, browser_press_key, browser_input_dispatch, browser_eval, browser_evaluate.
• browser_eval/browser_evaluate are treated as read-only when you pass { "readOnly": true } — no flash, no update record. Use it for pure reads (document.title, counting nodes) so a poll loop doesn't flash.
• Per-call suppress: { "silent": true } on any verb.
• Per-window default: browser_open_window { "autoFlash": false } makes that whole session silent (for a background poller you never want to signal).

2. TELL THE USER WHAT CHANGED — updateNote + lastAgentUpdate ✅

• Pass updateNote": "<short what-I-did>" on any mutating verb (also note).
• pup records lastAgentUpdate: { tabId, verb, note, ts } and exposes it:
  • browser_list_tabs → top-level lastAgentUpdate and per-tab tab.lastAgentUpdate (so you see WHICH tab).
  • browser_list_windows / session status → lastAgentUpdate.
• This is what lets the user (or a Session Monitor) answer "which of my 4 tabs did the agent just stage?"

3. RAISE LEAKS KEYSTROKES ✅

browser_raise_os_window / browser_focus_window now blur the page's active element before bringToFront(), so a mid-typed keystroke doesn't land in a focused input on raise. The verb doc now also steers agents away from raise for "I updated this" — use browser_alert_window (or just the new auto-flash), reserve raise for an explicit "show me / demo / record".

4. fullPage:true MISSES INNER SCROLL CONTAINERS ✅ (already shipped 1.8.21)

browser_screenshot { fullPage:true } only covers document scroll. On a nested-scroll SPA it now returns just the viewport plus a nestedScroll:true hint naming the real scroll container — that satisfies the "OR return a warning" branch you asked for. For a full inner-container capture use browser_screenshot_element (by selector/text, shadow-piercing) or browser_scroll through the container. (Full-stitch of an arbitrary inner container is a possible future add; the warning + element-shot covers the real cases today.)

5. PAPER CUTS

• (a) desktop_screenshot_window titleContains/sessionId — routed to AD, not pup. desktop_* is an AD-core verb; pup can't add args to it. This one is an AD-core request, not a pup change. For the pup side, browser_screenshot { sessionId } already targets by session, and browser_open_window { width,height } frames deterministically.
• (b) browser_open_tab response ✅ now returns { ok, sessionId, tabId, tabCount, activeTabId, ... } — the new tabId + tabCount are explicit so you never have to diff list_tabs to learn what you just opened.
• (c) data: URL mojibake ✅ new verb browser_render_html { sessionId, html, waitUntil?, updateNote?, silent? } renders a local HTML string via page.setContent (UTF-8-correct — verified Café résumé — naïve €50 ✓ renders clean where a data: URL mojibakes without an explicit charset). Screenshot/drive it like any page.

Acceptance test (your #1) — passed: background open_window → navigate → eval-fill → screenshot, no alert calls → single taskbar flash for the burst, no focus steal; browser_list_tabs reported { tabId: "tab-1", verb: "browser_eval", note: "injected a heading", ts } top-level and on the tab. browser_render_html UTF-8 verified by screenshot.

Update your AD-side hints to prefer updateNote on mutating verbs and browser_alert_window over raise. Consumer skill (pup-windows-sessions-tabs / pup-screenshots-recording) is being updated to teach these.