DynamicDevices · ajlennon · Aug 5, 2026 · Aug 5, 2026
@@ -127,7 +127,8 @@ Full table: [Dialog UX (humans)](docs/AGENTS.md#dialog-ux-humans).
 |-------|------|----|-------|--------|
 | Ubuntu 24.04 + GNOME + PipeWire | Cursor | Yes | Yes | **Verified** (2026-07) |
 | Same stack | Claude Code | Yes | Text-only | **Verified** (2026-07) — [@jackghx](https://github.com/jackghx) |
-| Windows 10/11 + tkinter | Cursor | Text | No | Phase 1 — **not yet reported** |
+| Windows 10/11 + WebView2 | Cursor | Text (Nebula UI) | No | Phase 2 UI — **Anthony laptop** |
+| Windows 10/11 + tkinter | Cursor | Text | No | Fallback if pywebview missing |
 | macOS / headless CI | — | No | No | Unsupported / N/A |
 
 More rows & how to report: [CONTRIBUTING.md](CONTRIBUTING.md).

@@ -1,24 +1,33 @@
-# Windows (Anthony / Cursor) — Phase 1 handoff
+# Windows (Anthony / Cursor) — WebView2 Nebula dialog
 
 Target user: **Anthony** ([@TheRealCheese](https://github.com/TheRealCheese)).
 Goal: text-only `ask_multiple_choice` on **Cursor for Windows** without WSL.
 
 Canonical install steps: this file (and a one-line pointer from the
 [README](../README.md#quick-start-linux)).
 
+## UI backend
+
+| Priority | Backend | Notes |
+| --- | --- | --- |
+| 1 (default) | **WebView2 + pywebview** | Nebula-styled floating dialog (`win_webview_ask.py`) |
+| 2 | tkinter | Fallback (`win_list_ask.py`) if pywebview missing |
+
+Force with env: `ASK_QUESTION_WIN_UI=webview` or `ASK_QUESTION_WIN_UI=tk`.
+
 ## Checklist for Anthony
 
-1. Python 3.12+ from python.org with **tcl/tk** → `python -c "import tkinter; print('ok')"`.
+1. Python 3.12+ (python.org). Edge WebView2 is already on Windows 11.
 2. Install uv → `where uv` (absolute path to `uv.exe`).
-3. `git clone https://github.com/DynamicDevices/ask-question-mcp.git` then:
+3. In the checkout:
    ```bat
    uv sync
    uv run ask-question-install --host cursor --skill
    ```
-   (or `git pull` + the same if you already have a clone).
 4. Cursor → **Developer: Reload Window**.
-5. Ask the agent: call **`check_setup`** (expect `ready.ui` / `ready.text_mcq` true; `audio_mode` text_only).
-6. Smoke **`ask_multiple_choice`** — dialog should appear on top; pick an option.
+5. Ask the agent: call **`check_setup`** (expect `ready.ui` / `ready.text_mcq` true;
+   `webview` ok; `audio_mode` text_only).
+6. Smoke **`ask_multiple_choice`** — Nebula WebView2 dialog on top; pick an option.
 7. Smoke **keyboard** — labels show `1 · …`; press **2**, wait for OK to arm,
    **Enter**. **Esc** cancels. Footer hint: `1–8 select · Enter OK · Esc cancel`.
 8. Smoke **Something else** — every MCQ should include a freeform row / entry; typing
@@ -37,30 +46,22 @@ Manual mcp.json edit is still fine if you skip the installer — use absolute
 
 ## Behaviour parity (vs Linux)
 
-Shared path (`zenity_ask` → `win_list_ask.py`):
-
-| Behaviour | Windows Phase 1 |
+| Behaviour | Windows |
 | --- | --- |
 | Something else always offered | Yes (same as Linux; `allow_other` ignored) |
 | Danger mark **⛔** + confirm arm | Yes (`danger_arm.py`) |
 | Danger banner wording | **⛔ Confirm** + lead ask (pink banner) |
-| Lead / detail (ask visible; tall referent scrolls) | Yes (`split_lead_detail`) |
+| Lead / detail (ask visible; tall referent scrolls) | Yes (`split_lead_detail` on tk; WebView body scrolls) |
 | Red OK on danger | Yes |
 | Voice / duck / STT | No (text-only) |
-| Image / images preview | No (ignored; text-only) |
-| Gtk footer / scroll layout fixes | N/A (tkinter layout) |
+| Image / images preview | No (ignored; text-only for now) |
 | 1–8 hotkeys + Enter / Esc | Yes |
 | Remember size/position | Yes (`prefs.window`) |
-| Scrollable options | Yes |
+| Aesthetic | Nebula glass WebView2 (frameless); tk fallback; optional Edge `--app` |
 
-## Out of scope (Phase 1)
+## Out of scope
 
 - Spoken questions / mic answers / media duck
 - WSL as the supported path
 - macOS
-
-## After Anthony verifies
-
-Maintainers: update README Tested platforms Windows row to **Verified**, and
-optionally add a `VERIFIED_PLATFORMS` entry with `"system": "windows"` in
-[`platform_info.py`](../src/ask_question_mcp/platform_info.py).
+- Hosting the dialog *inside* the Cursor chat chrome (MCP has no custom panel API yet)
@@ -11,6 +11,7 @@ authors = [
 ]
 dependencies = [
   "mcp[cli]>=1.9.0",
+  "pywebview>=5.0",
 ]
 
 [project.urls]

@@ -15,13 +15,17 @@
 from ask_question_mcp.zenity_ask import (
     _WIN_ENTRY_ASK,
     _WIN_LIST_ASK,
+    _WIN_WEBVIEW_ASK,
+    _WIN_WEBVIEW_ENTRY_ASK,
     ui_backend,
 )
 
 
 def main() -> None:
     assert _WIN_LIST_ASK.is_file(), _WIN_LIST_ASK
     assert _WIN_ENTRY_ASK.is_file(), _WIN_ENTRY_ASK
+    assert _WIN_WEBVIEW_ASK.is_file(), _WIN_WEBVIEW_ASK
+    assert _WIN_WEBVIEW_ENTRY_ASK.is_file(), _WIN_WEBVIEW_ENTRY_ASK
 
     if sys.platform == "win32":
         assert ui_backend() == "win"
@@ -75,7 +79,10 @@ def main() -> None:
         assert caps.audio_mode == "text_only"
         assert caps.speak_active is False
         assert caps.listen_active is False
-        assert any("Windows Phase 1" in n for n in caps.notes)
+        assert any(
+            "Windows" in n and ("WebView2" in n or "Phase 1" in n or "tkinter" in n)
+            for n in caps.notes
+        )
 
     # Ensure Linux UI-ready still requires DISPLAY when not Windows.
     if sys.platform != "win32":

@@ -1,19 +1,27 @@
 ---
 name: ask-multiple-choice
 description: >-
-  Prefer desktop MCP ask_multiple_choice for decision forks — not markdown
-  A/B/C. Use when choosing or confirming. Pass image=/images= when the human
-  must judge a still in the dialog.
+  Desktop MCP ask_multiple_choice for every decision fork — Anthony's opinion
+  when stuck, choosing paths, confirming work, or picking options. Prefer this
+  over markdown A/B/C or host AskQuestion. Supports multi-select and optional
+  image= / images= previews. Use whenever the ask-question MCP is available
+  and a human choice is needed.
 ---
 
 # Ask multiple choice (desktop MCP)
 
-**Need-band:** freedom — faster decisions without chat A/B/C noise.
+Cursor stand-in for Claude Code **AskUserQuestion**. **Need-band:** freedom —
+faster decisions without chat A/B/C noise.
 
-## When (auto)
+## When (auto — do not wait to be asked)
 
-Any **decision fork**: ship/wait, pick a path, confirm irreversible work,
-choose among options the human must decide.
+Any fork that needs **Anthony's opinion**, including when you are **stuck**:
+
+- ship / wait / pick a path
+- confirm irreversible or risky work
+- choose among concrete options
+- several answers may apply together → multi-select
+- preference unknown — ask; do not guess
 
 ## Do
 
@@ -27,63 +35,36 @@ choose among options the human must decide.
    (always fully visible); put Command/To+body/path **before** meta notes.
    Do **not** dump process templates, PATTERN blocks, or long meta into routine
    forks. No meta about dialogs/voice.
-4. **Permission / action asks (Alex 2026-07-30):** state **what** Briar will do
-   **and why** (one sentence each is enough). Opaque “run Task agents?” /
-   “proceed?” without purpose is not enough — Alex must understand the aim.
+4. **Permission / action asks:** state **what** will happen **and why** (one
+   sentence each is enough). Opaque “proceed?” without purpose is not enough.
    Pattern: `mcq-permission-what-and-why`.
-5. Mark preferred only as **`Label (recommended)`** + **`recommended_id`**.
-6. **`dangerous=true`** for irreversible / high-risk forks.
-7. **Images the human must judge** (Alex loves this — signed-off 2026-08-03):
-   pass **`image=`** (one path / `file://` URI) or **`images=`** (list, max 4).
-   Chat `Read` of a PNG does **not** put pixels in the MCQ — agents **must**
-   pass the path into the dialog. Linux Gtk: opens large on the **primary**
-   usable workarea (not the largest secondary); human can **click the preview**
-   (large ↔ compact ~320px) and **maximize** (header button or **F**).
-   **P0 — multi-image must never exceed primary usable resolution** (stack
-   shares one height budget / scrolls inside). Text-only MCQs stay compact.
-   Pattern: `mcq-with-image`.
-8. Wait for the JSON result. On cancel → stop. On freeform → use **`freeform_text`**.
-
-Humans use the dialog keyboard (**1–8**, Enter, Esc; **F** maximize when images);
-do not put hotkey instructions in `question`. Detail: repo `docs/AGENTS.md`
-(Dialog UX).
-
-## Voice — who is “I”? (Alex 2026-07-30)
-
-MCQs must not use ambiguous **I** / **you** for actions.
-
-| Role | How to refer |
-|------|----------------|
-| Human | **Alex** (or “Alex will…”) |
-| This assistant | **Briar** (or “Briar will…”) |
-
-**Do:** `Briar will restart the webhook` · `Alex approves the send` · option
-labels like `Briar sends now` / `Alex will edit first`.
-**Don't:** `I'll capture…` / `you refresh…` / `I mean the agent…` when either
-party could be “I”.
-
-Casual chat outside MCQs may still use normal I/you; **MCQ question + option
-labels** stay role-named. Pattern: `mcq-named-roles-alex-briar`.
+5. Mark preferred only as **`Label (recommended)`** + **`recommended_id`**
+   (or **`recommended_ids`** for multi).
+6. Set **`allow_multiple=true`** when more than one option can be correct together
+   (checklist). Default single-select otherwise.
+7. **`dangerous=true`** for irreversible / high-risk forks.
+8. **Images the human must judge:** pass **`image=`** (one path / `file://`) or
+   **`images=`** (list). Chat `Read` of a PNG does **not** put pixels in the MCQ.
+   Linux Gtk shows the preview in-dialog; Windows WebView path may ignore images
+   until wired.
+9. Wait for the JSON result.
+   - Cancel → **stop** (do not invent a choice).
+   - Freeform → honour **`freeform_text`**.
+   - Multi → use **`ids`** / **`labels`**.
+
+Humans use dialog keys (**1–8**, Enter, Esc); do not put hotkey text in
+`question`. Detail: repo `docs/AGENTS.md` (Dialog UX).
 
 ## Don't
 
 - Markdown A/B/C, numbered lists, or host AskQuestion when this MCP is available
-- Desktop MCQ for clarifications / next steps when the ask originated over
-  **WhatsApp admin inbound** — keep those on WhatsApp (Charlize voice or short
-  text) per `cursor-pa-whatsapp` SOUL; desktop MCQ is for Cursor-session forks
-  not originated on WA (unless Alex opts in). P0 send-gates to third parties
-  may still use desktop MCQ.
-- Asking Alex to judge a still that exists only in chat when the dialog can take
-  **`image=`** / **`images=`**
-- “Send now?” / “Ship it?” with no body/path when the human has not seen the draft
-- Permission MCQs that name a tool/action but omit **why** (Task agents, long
-  scans, enable MCP, restart services, etc.)
-- Stuffing PATTERN/PROPOSAL/OWNS walls into every MCQ
-- Soft MCQs before a send-gate (“draft OK?”, “ready?”, “shall Briar send?”) —
-  draft in chat, then **one** send-gate only (`email-one-send-gate`)
+- Asking the human to judge a still that exists only in chat when the dialog can
+  take **`image=`** / **`images=`**
+- Soft MCQs before a real send-gate — draft in chat, then one confirm
 - `check_setup` before routine MCQs (only first enable, dialog failure, or before voice)
 - Invent a choice after `cancelled: true`
-- Ambiguous **I/you** in MCQ question or option labels (use Alex / Briar)
+- Skip the dialog because the skill wasn't @-mentioned — the always-on user rule
+  still applies when the MCP is loaded
 
 ## Setup (humans)
 
@@ -92,4 +73,4 @@ cd /path/to/ask-question-mcp && uv sync
 uv run ask-question-install --host cursor --skill
 ```
 
-Then reload the host. Detail: repo `docs/AGENTS.md`.
+Then reload the host. Windows: see `docs/WINDOWS.md`. Detail: `docs/AGENTS.md`.