feat(vom): fixture test

fix(recording): support user action recording in iframes and OOPIFs

.github/workflows/release-dsh-plugin.yml

@@ -0,0 +1,105 @@
+name: Release dsh plugin
+
+on:
+  push:
+    tags:
+      - dsh-plugin-v*
+  workflow_dispatch:
+
+permissions:
+  contents: read
+
+concurrency:
+  group: release-dsh-plugin-${{ github.ref }}
+  cancel-in-progress: false
+
+env:
+  PACKAGE_NAME: "@wxg-prc-cpg/browser-skill-dsh-plugin"
+  PACKAGE_DIR: packages/dsh-plugin-browserskill
+
+jobs:
+  resolve:
+    runs-on: ubuntu-latest
+    outputs:
+      version: ${{ steps.version.outputs.version }}
+      tag: ${{ steps.version.outputs.tag }}
+    steps:
+      - uses: actions/checkout@v6
+
+      - name: Resolve plugin version
+        id: version
+        shell: bash
+        run: |
+          set -euo pipefail
+          if [ "${{ github.event_name }}" = "workflow_dispatch" ]; then
+            VERSION="$(node -p "require('./${{ env.PACKAGE_DIR }}/package.json').version")"
+          else
+            VERSION="${GITHUB_REF_NAME#dsh-plugin-v}"
+          fi
+          VERSION="${VERSION#v}"
+          echo "version=${VERSION}" >> "$GITHUB_OUTPUT"
+          echo "tag=dsh-plugin-v${VERSION}" >> "$GITHUB_OUTPUT"
+
+  guard:
+    needs: resolve
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v6
+
+      - name: Verify tag matches package.json name and version
+        shell: bash
+        run: |
+          set -euo pipefail
+          PKG_NAME="$(node -p "require('./${{ env.PACKAGE_DIR }}/package.json').name")"
+          PKG_VERSION="$(node -p "require('./${{ env.PACKAGE_DIR }}/package.json').version")"
+          RELEASE_VERSION="${{ needs.resolve.outputs.version }}"
+          if [ "$PKG_NAME" != "$PACKAGE_NAME" ]; then
+            echo "package.json name (${PKG_NAME}) does not match expected ${PACKAGE_NAME}"
+            exit 1
+          fi
+          if [ "$PKG_VERSION" != "$RELEASE_VERSION" ]; then
+            echo "${PACKAGE_DIR}/package.json version (${PKG_VERSION}) does not match release version (${RELEASE_VERSION})"
+            exit 1
+          fi
+          echo "Version guard passed: ${PKG_NAME}@${RELEASE_VERSION}"
+
+  publish:
+    needs: [resolve, guard]
+    runs-on: ubuntu-latest
+    # Secret NPM_TOKEN is stored on the "npm-publish" GitHub Environment.
+    environment: npm-publish
+    permissions:
+      contents: read
+    steps:
+      - uses: actions/checkout@v6
+
+      - uses: pnpm/action-setup@v6
+
+      - uses: actions/setup-node@v6
+        with:
+          node-version: 22
+          cache: pnpm
+          registry-url: https://registry.npmjs.org
+          scope: "@wxg-prc-cpg"
+
+      - name: Install dependencies
+        run: pnpm install --frozen-lockfile
+
+      - name: Typecheck plugin
+        run: pnpm --filter "${{ env.PACKAGE_NAME }}" typecheck
+
+      - name: Test plugin
+        run: pnpm --filter "${{ env.PACKAGE_NAME }}" test
+
+      - name: Publish to npm
+        working-directory: ${{ env.PACKAGE_DIR }}
+        env:
+          NODE_AUTH_TOKEN: ${{ secrets.NPM_TOKEN }}
+        run: |
+          set -euo pipefail
+          if [ -z "${NODE_AUTH_TOKEN:-}" ]; then
+            echo "::error::Environment secret NPM_TOKEN is not configured"
+            exit 1
+          fi
+          pnpm publish --access public --no-git-checks
+          echo "Published ${PACKAGE_NAME}@${{ needs.resolve.outputs.version }}"

.stylelintignore

@@ -3,3 +3,6 @@
 **/.output/**
 **/target/**
 **/node_modules/**
+# Generated client CSS (tailwind compile output); built by
+# packages/dsh-plugin-browserskill/scripts/build-client-css.mjs.
+**/dsh-plugin-browserskill/src/client/bsk-ui.nomodule.css

AGENT_INSTALL.md

@@ -34,13 +34,17 @@ where only `extension connected` fails is expected; go to Step 3.
 ## 3. Open the extension install page
 
 If `extension connected` is `FAIL` (`0 browsers connected`), the user likely
-has not installed or enabled the browser extension yet. Open the
-[Chrome Web Store install page](https://chromewebstore.google.com/detail/hhcmgoofomhgciiibhipgmgkgnoenaoi)
-in the user's browser, then tell the user:
+has not installed or enabled the browser extension yet. Open the install page
+matching the user's browser —
+[Chrome Web Store](https://chromewebstore.google.com/detail/hhcmgoofomhgciiibhipgmgkgnoenaoi)
+for Chrome and other Chromium browsers,
+[Edge Add-ons](https://microsoftedge.microsoft.com/addons/detail/browserskill/emacgiaaaiojkkpkddmmdfhmokgmnikg)
+for Microsoft Edge — then tell the user:
 
 > Install **BrowserSkill** from the
-> [Chrome Web Store](https://chromewebstore.google.com/detail/hhcmgoofomhgciiibhipgmgkgnoenaoi),
-> open the popup, and wait until it turns green. Reply when done.
+> [Chrome Web Store](https://chromewebstore.google.com/detail/hhcmgoofomhgciiibhipgmgkgnoenaoi)
+> (or [Edge Add-ons](https://microsoftedge.microsoft.com/addons/detail/browserskill/emacgiaaaiojkkpkddmmdfhmokgmnikg)
+> on Edge), open the popup, and wait until it turns green. Reply when done.
 
 If opening the page fails, give the user the same link. Then run `bsk doctor`
 once more. All `ok`/`na` → tell the user it's ready.

README.md

@@ -13,7 +13,7 @@
 </p>
 
 **BrowserSkill** connects Cursor, Claude Code, Codex, OpenClaw, CodeBuddy,
-WorkBuddy, Pi, Hermes Agent, and other shell-capable AI agents to your already logged-in
+WorkBuddy, Pi, Hermes Agent, DeepSeek Harness, and other AI agents to your already logged-in
 browser.
 
 Need the agent to touch a tab you already have open? It must borrow that tab
@@ -67,7 +67,8 @@ Set up browser-skill on this machine by following https://raw.githubusercontent.
 
 <br>
 
-Install the CLI, then install the extension from the [Chrome Web Store](https://chromewebstore.google.com/detail/hhcmgoofomhgciiibhipgmgkgnoenaoi).
+Install the CLI, then install the extension from the [Chrome Web Store](https://chromewebstore.google.com/detail/hhcmgoofomhgciiibhipgmgkgnoenaoi)
+or [Edge Add-ons](https://microsoftedge.microsoft.com/addons/detail/browserskill/emacgiaaaiojkkpkddmmdfhmokgmnikg).
 
 #### 1. Install the `bsk` CLI
 
@@ -91,7 +92,14 @@ bsk --version
 
 #### 2. Install the browser extension
 
-Install BrowserSkill from the [Chrome Web Store](https://chromewebstore.google.com/detail/hhcmgoofomhgciiibhipgmgkgnoenaoi).
+Install BrowserSkill from your browser's store:
+
+| Browser | Store listing |
+| --- | --- |
+| Chrome | [Chrome Web Store](https://chromewebstore.google.com/detail/hhcmgoofomhgciiibhipgmgkgnoenaoi) |
+| Microsoft Edge | [Edge Add-ons](https://microsoftedge.microsoft.com/addons/detail/browserskill/emacgiaaaiojkkpkddmmdfhmokgmnikg) |
+
+On other Chromium-based browsers, install the Chrome Web Store build.
 
 #### 3. Install the skill
 
@@ -123,7 +131,8 @@ internal variants and install paths.
 
 Other shell-capable agent harnesses are supported too. Copy
 [`skill/SKILL.md`](skill/SKILL.md) into your harness's skills directory as
-`browser-skill/SKILL.md` to install the skill manually.
+`browser-skill/SKILL.md` to install the skill manually. DeepSeek Harness uses a
+dedicated plugin instead — see [DeepSeek Harness plugin](#deepseek-harness-plugin).
 
 </details>
 
@@ -133,6 +142,26 @@ Start a new Agent session and write a prompt that needs the browser, for example
 /browser-skill open example.com and summarize what is on the page.
 ```
 
+## DeepSeek Harness plugin
+
+Using [DeepSeek Harness](https://github.com/deepseek-ai/deepseek-harness) (`dsh`)?
+BrowserSkill ships a first-class dsh plugin on npm as
+[`@wxg-prc-cpg/browser-skill-dsh-plugin`](https://www.npmjs.com/package/@wxg-prc-cpg/browser-skill-dsh-plugin).
+It injects native `browser_*` tools (no shelling out to `bsk`) and a live Web UI
+overlay of each Agent Window.
+
+Add it to a dsh profile, then start that profile:
+
+```sh
+dsh plugin --profile web add @wxg-prc-cpg/browser-skill-dsh-plugin
+dsh --profile web
+```
+
+The plugin carries its own copy of the skill, so `bsk install-skill` is not needed
+for dsh — but the `bsk` CLI and the browser extension are still prerequisites. See
+the [plugin README](packages/dsh-plugin-browserskill/README.md) for the tool list,
+configuration, and the observation overlay.
+
 ## How It Works
 
 BrowserSkill is a local bridge between your agent harness and your browser.
@@ -166,7 +195,9 @@ flowchart TB
 
 The agent never talks to the browser directly. It asks the `bsk` CLI to perform a
 browser task; the local daemon routes that request to the extension; the
-extension runs it in an Agent Window.
+extension runs it in an Agent Window. DeepSeek Harness takes the same path
+through the [plugin](#deepseek-harness-plugin): the agent calls injected
+`browser_*` tools, and the plugin invokes `bsk` on its behalf.
 
 ## For Developers
 
@@ -176,6 +207,7 @@ The repository is a Cargo + pnpm workspace:
 - `crates/bsk-protocol` — shared wire types and JSON schemas
 - `apps/extension` — browser extension
 - `packages/ui` and `packages/i18n` — shared extension UI support
+- `packages/dsh-plugin-browserskill` — DeepSeek Harness plugin (`@wxg-prc-cpg/browser-skill-dsh-plugin`)
 
 ## License
 

README.zh-CN.md

@@ -12,7 +12,7 @@
   <a href="README.md">English</a> · 中文
 </p>
 
-**BrowserSkill** 把 Cursor、Claude Code、Codex、OpenClaw、CodeBuddy、WorkBuddy、Pi、Hermes Agent 等支持 Shell 的 AI Agent 连接到你已登录的浏览器。
+**BrowserSkill** 把 Cursor、Claude Code、Codex、OpenClaw、CodeBuddy、WorkBuddy、Pi、Hermes Agent、DeepSeek Harness 等 AI Agent 连接到你已登录的浏览器。
 
 需要 Agent 操作你已打开的标签页？必须显式借用该标签，任务结束后归还，其余浏览器窗口不受影响。
 
@@ -54,7 +54,8 @@ BrowserSkill 由两个本地运行组件组成：`bsk` CLI/daemon 和浏览器
 
 <br>
 
-先安装 CLI，再从 [Chrome Web Store](https://chromewebstore.google.com/detail/hhcmgoofomhgciiibhipgmgkgnoenaoi) 安装浏览器扩展。
+先安装 CLI，再从 [Chrome Web Store](https://chromewebstore.google.com/detail/hhcmgoofomhgciiibhipgmgkgnoenaoi)
+或 [Edge 加载项商店](https://microsoftedge.microsoft.com/addons/detail/browserskill/emacgiaaaiojkkpkddmmdfhmokgmnikg) 安装浏览器扩展。
 
 #### 1. 安装 `bsk` CLI
 
@@ -64,8 +65,11 @@ BrowserSkill 由两个本地运行组件组成：`bsk` CLI/daemon 和浏览器
 curl -fsSL https://raw.githubusercontent.com/Tencent/BrowserSkill/main/install.sh | sh
 ```
 
-**Windows**：从 [最新 CLI release](https://github.com/Tencent/BrowserSkill/releases/latest)
-下载 `bsk-v<version>-x86_64-pc-windows-msvc.zip`，解压后将 `bsk.exe` 加入 `PATH`。
+**Windows**（PowerShell，安装到 `~/.local/bin`）：
+
+```powershell
+irm https://raw.githubusercontent.com/Tencent/BrowserSkill/main/install.ps1 | iex
+```
 
 验证二进制：
 
@@ -75,7 +79,14 @@ bsk --version
 
 #### 2. 安装浏览器扩展
 
-从 [Chrome Web Store](https://chromewebstore.google.com/detail/hhcmgoofomhgciiibhipgmgkgnoenaoi) 安装 BrowserSkill。
+在对应浏览器的商店安装 BrowserSkill：
+
+| 浏览器 | 商店页面 |
+| --- | --- |
+| Chrome | [Chrome Web Store](https://chromewebstore.google.com/detail/hhcmgoofomhgciiibhipgmgkgnoenaoi) |
+| Microsoft Edge | [Edge 加载项商店](https://microsoftedge.microsoft.com/addons/detail/browserskill/emacgiaaaiojkkpkddmmdfhmokgmnikg) |
+
+其他基于 Chromium 的浏览器，安装 Chrome Web Store 版本即可。
 
 #### 3. 安装 skill
 
@@ -102,7 +113,7 @@ bsk install-skill
 
 用 <kbd>Space</kbd> 选择需要安装的 Agent harness，然后按 <kbd>Enter</kbd> 安装 skill。运行 `bsk install-skill --list` 可查看 internal 变体及安装路径。
 
-其他支持 Shell 的 Agent harness 也可使用 BrowserSkill，但需手动将 [`skill/SKILL.md`](skill/SKILL.md) 复制到对应 skills 目录下的 `browser-skill/SKILL.md`。
+其他支持 Shell 的 Agent harness 也可使用 BrowserSkill，但需手动将 [`skill/SKILL.md`](skill/SKILL.md) 复制到对应 skills 目录下的 `browser-skill/SKILL.md`。DeepSeek Harness 走独立插件，见 [DeepSeek Harness 插件](#deepseek-harness-插件)。
 
 </details>
 
@@ -112,6 +123,19 @@ bsk install-skill
 /browser-skill open example.com and summarize what is on the page.
 ```
 
+## DeepSeek Harness 插件
+
+在用 [DeepSeek Harness](https://github.com/deepseek-ai/deepseek-harness)（`dsh`）？BrowserSkill 提供了官方 dsh 插件，已发布到 npm：[`@wxg-prc-cpg/browser-skill-dsh-plugin`](https://www.npmjs.com/package/@wxg-prc-cpg/browser-skill-dsh-plugin)。它会注入原生 `browser_*` 工具（无需再通过 Shell 调用 `bsk`），并在 Web UI 中实时观察每个 Agent Window。
+
+把它装进某个 dsh profile，然后启动该 profile：
+
+```sh
+dsh plugin --profile web add @wxg-prc-cpg/browser-skill-dsh-plugin
+dsh --profile web
+```
+
+插件自带 skill，所以在 dsh 下无需执行 `bsk install-skill`；但 `bsk` CLI 和浏览器扩展仍是前置条件。工具清单、配置项与观察浮层见[插件 README](packages/dsh-plugin-browserskill/README.md)。
+
 ## 工作原理
 
 BrowserSkill 是 Agent 运行时与浏览器之间的本地桥接层。
@@ -143,7 +167,7 @@ flowchart TB
   style UserWindows fill:#f8fafc,stroke:#cbd5e1,color:#334155
 ```
 
-Agent 不直接与浏览器通信。它通过 `bsk` CLI 下发浏览器任务；本地 daemon 把请求路由到扩展；扩展在 Agent Window 中执行。
+Agent 不直接与浏览器通信。它通过 `bsk` CLI 下发浏览器任务；本地 daemon 把请求路由到扩展；扩展在 Agent Window 中执行。DeepSeek Harness 走同一条链路，只是经由 [插件](#deepseek-harness-插件)：Agent 调用注入的 `browser_*` 工具，由插件代为执行 `bsk`。
 
 ## 面向开发者
 
@@ -153,6 +177,7 @@ Agent 不直接与浏览器通信。它通过 `bsk` CLI 下发浏览器任务；
 - `crates/bsk-protocol` — 共享协议类型与 JSON Schema
 - `apps/extension` — 浏览器扩展
 - `packages/ui` 和 `packages/i18n` — 扩展 UI 共享支持
+- `packages/dsh-plugin-browserskill` — DeepSeek Harness 插件（`@wxg-prc-cpg/browser-skill-dsh-plugin`）
 
 ## 许可证
 

apps/extension/src/browser-driver/__tests__/chromium-cdp.test.ts

@@ -1,5 +1,5 @@
 import { describe, expect, it, vi } from "vitest";
-import { type CdpDebuggerApi, ChromiumCdp } from "../chromium-cdp";
+import { type CdpDebuggee, type CdpDebuggerApi, ChromiumCdp } from "../chromium-cdp";
 
 function fakeChromeEvent<TArgs extends unknown[]>() {
   const listeners = new Set<(...args: TArgs) => void>();
@@ -14,7 +14,7 @@ function fakeChromeEvent<TArgs extends unknown[]>() {
 }
 
 function fakeApi() {
-  const onEvent = fakeChromeEvent<[chrome.debugger.Debuggee, string, unknown]>();
+  const onEvent = fakeChromeEvent<[CdpDebuggee, string, unknown]>();
   const onDetach = fakeChromeEvent<[chrome.debugger.Debuggee, string]>();
   const api: CdpDebuggerApi = {
     attach: vi.fn(async () => {}),
@@ -29,6 +29,135 @@ function fakeApi() {
 }
 
 describe("ChromiumCdp", () => {
+  it("discovers multiple iframe targets and recursively routes nested OOPIF commands", async () => {
+    const { api, onEvent } = fakeApi();
+    (api.sendCommand as ReturnType<typeof vi.fn>).mockImplementation(
+      async (
+        target: chrome.debugger.Debuggee & { sessionId?: string },
+        method: string,
+        params?: { frameId?: string },
+      ) => {
+        if (method === "DOM.getFrameOwner") {
+          return { backendNodeId: params?.frameId === "nested" ? 300 : 200 };
+        }
+        if (method !== "Page.getFrameTree") return {};
+        if (target.sessionId === "right-session") {
+          return {
+            frameTree: {
+              frame: { id: "right", url: "https://right.test" },
+              childFrames: [
+                { frame: { id: "nested", parentId: "right", url: "https://nested.test" } },
+              ],
+            },
+          };
+        }
+        if (target.sessionId === "nested-session") {
+          return { frameTree: { frame: { id: "nested", url: "https://nested.test" } } };
+        }
+        return {
+          frameTree: {
+            frame: { id: "main", url: "https://app.test" },
+            childFrames: [
+              { frame: { id: "left", parentId: "main", url: "https://left.test" } },
+              {
+                frame: { id: "right", parentId: "main", url: "https://right.test" },
+                childFrames: [
+                  { frame: { id: "nested", parentId: "right", url: "https://nested.test" } },
+                ],
+              },
+            ],
+          },
+        };
+      },
+    );
+    const cdp = new ChromiumCdp(api);
+    await cdp.ensureAttached(4);
+    onEvent.fire({ tabId: 4 }, "Target.attachedToTarget", { sessionId: "right-session" });
+    onEvent.fire({ tabId: 4 } as chrome.debugger.Debuggee, "Target.attachedToTarget", {
+      sessionId: "nested-session",
+    });
+
+    const graph = await cdp.getFrameGraph(4);
+
+    expect(graph.frames).toHaveLength(4);
+    expect(graph.frames.find((frame) => frame.frameId === "left")?.target).toEqual({ tabId: 4 });
+    expect(graph.frames.find((frame) => frame.frameId === "right")?.target.sessionId).toBe(
+      "right-session",
+    );
+    expect(graph.frames.find((frame) => frame.frameId === "nested")?.target.sessionId).toBe(
+      "nested-session",
+    );
+    expect(graph.frames.find((frame) => frame.frameId === "right")?.ownerBackendNodeId).toBe(200);
+    expect(graph.frames.find((frame) => frame.frameId === "nested")?.ownerBackendNodeId).toBe(300);
+    expect(api.sendCommand).toHaveBeenCalledWith(
+      { tabId: 4, sessionId: "right-session" },
+      "DOM.getFrameOwner",
+      { frameId: "nested" },
+    );
+    expect(api.sendCommand).toHaveBeenCalledWith(
+      { tabId: 4, sessionId: "nested-session" },
+      "Target.setAutoAttach",
+      expect.objectContaining({ flatten: true }),
+    );
+  });
+
+  it("waits for recursively attached iframe targets without a depth limit", async () => {
+    const { api, onEvent } = fakeApi();
+    const depth = 6;
+    const scheduled = new Set<string>();
+    const nestedTree = (index: number): Record<string, unknown> => ({
+      frame: {
+        id: index === 0 ? "main" : `frame-${index}`,
+        ...(index > 0 ? { parentId: index === 1 ? "main" : `frame-${index - 1}` } : {}),
+      },
+      ...(index < depth ? { childFrames: [nestedTree(index + 1)] } : {}),
+    });
+    (api.sendCommand as ReturnType<typeof vi.fn>).mockImplementation(
+      async (target: CdpDebuggee, method: string, params?: { frameId?: string }) => {
+        if (method === "Target.setAutoAttach") {
+          const parentIndex = target.sessionId
+            ? Number(target.sessionId.replace("session-", ""))
+            : 0;
+          const nextIndex = parentIndex + 1;
+          const key = `${target.sessionId ?? "root"}:${nextIndex}`;
+          if (nextIndex <= depth && !scheduled.has(key)) {
+            scheduled.add(key);
+            setTimeout(() => {
+              onEvent.fire(
+                { tabId: 4, ...(target.sessionId ? { sessionId: target.sessionId } : {}) },
+                "Target.attachedToTarget",
+                {
+                  sessionId: `session-${nextIndex}`,
+                  targetInfo: { type: "iframe" },
+                },
+              );
+            }, 5);
+          }
+          return {};
+        }
+        if (method === "Page.getFrameTree") {
+          if (!target.sessionId) return { frameTree: nestedTree(0) };
+          const index = Number(target.sessionId.replace("session-", ""));
+          return { frameTree: nestedTree(index) };
+        }
+        if (method === "DOM.getFrameOwner") {
+          return { backendNodeId: Number(params?.frameId?.replace("frame-", "")) + 100 };
+        }
+        return {};
+      },
+    );
+    const cdp = new ChromiumCdp(api);
+
+    const graph = await cdp.getFrameGraph(4);
+
+    expect(graph.frames).toHaveLength(depth + 1);
+    for (let index = 1; index <= depth; index += 1) {
+      expect(
+        graph.frames.find((frame) => frame.frameId === `frame-${index}`)?.target.sessionId,
+      ).toBe(`session-${index}`);
+    }
+  });
+
   it("coalesces concurrent attach calls for the same tab", async () => {
     const { api } = fakeApi();
     let releaseAttach!: () => void;

apps/extension/src/browser-driver/__tests__/frame-graph.test.ts

@@ -0,0 +1,136 @@
+import { describe, expect, it } from "vitest";
+import { buildFrameGraph, type CdpFrameTreeNode, type CdpFrameTreeSource } from "../frame-graph";
+
+describe("buildFrameGraph", () => {
+  it("keeps sibling frames distinct and routes nested OOPIFs to their child sessions", () => {
+    const graph = buildFrameGraph([
+      {
+        target: { tabId: 4 },
+        tree: {
+          frame: { id: "main", url: "https://app.test" },
+          childFrames: [
+            { frame: { id: "left", parentId: "main", url: "https://left.test" } },
+            {
+              frame: { id: "right", parentId: "main", url: "https://right.test" },
+              childFrames: [
+                { frame: { id: "nested", parentId: "right", url: "https://nested.test" } },
+              ],
+            },
+          ],
+        },
+      },
+      {
+        target: { tabId: 4, sessionId: "right-session" },
+        tree: {
+          frame: { id: "right", url: "https://right.test" },
+          childFrames: [{ frame: { id: "nested", parentId: "right", url: "https://nested.test" } }],
+        },
+      },
+      {
+        target: { tabId: 4, sessionId: "nested-session" },
+        tree: { frame: { id: "nested", url: "https://nested.test" } },
+      },
+    ]);
+
+    expect(graph?.frames).toHaveLength(4);
+    expect(graph?.frames.find((frame) => frame.frameId === "left")?.target).toEqual({ tabId: 4 });
+    expect(graph?.frames.find((frame) => frame.frameId === "right")).toMatchObject({
+      parentFrameId: "main",
+      target: { tabId: 4, sessionId: "right-session" },
+    });
+    expect(graph?.frames.find((frame) => frame.frameId === "nested")).toMatchObject({
+      parentFrameId: "right",
+      target: { tabId: 4, sessionId: "nested-session" },
+    });
+  });
+
+  it("propagates each target boundary to its same-process descendants", () => {
+    const sources: CdpFrameTreeSource[] = [
+      {
+        target: { tabId: 4 },
+        tree: {
+          frame: { id: "main" },
+          childFrames: [
+            {
+              frame: { id: "oopif-a", parentId: "main" },
+              childFrames: [
+                {
+                  frame: { id: "same-process-b", parentId: "oopif-a" },
+                  childFrames: [
+                    {
+                      frame: { id: "oopif-c", parentId: "same-process-b" },
+                      childFrames: [{ frame: { id: "same-process-d", parentId: "oopif-c" } }],
+                    },
+                  ],
+                },
+              ],
+            },
+          ],
+        },
+      },
+      {
+        target: { tabId: 4, sessionId: "session-a" },
+        tree: {
+          frame: { id: "oopif-a" },
+          childFrames: [
+            {
+              frame: { id: "same-process-b", parentId: "oopif-a" },
+              childFrames: [{ frame: { id: "oopif-c", parentId: "same-process-b" } }],
+            },
+          ],
+        },
+      },
+      {
+        target: { tabId: 4, sessionId: "session-c" },
+        tree: {
+          frame: { id: "oopif-c" },
+          childFrames: [{ frame: { id: "same-process-d", parentId: "oopif-c" } }],
+        },
+      },
+    ];
+
+    const graph = buildFrameGraph(sources);
+
+    expect(graph?.frames.map((frame) => [frame.frameId, frame.target.sessionId])).toEqual([
+      ["main", undefined],
+      ["oopif-a", "session-a"],
+      ["same-process-b", "session-a"],
+      ["oopif-c", "session-c"],
+      ["same-process-d", "session-c"],
+    ]);
+    expect(graph?.frames.find((frame) => frame.frameId === "oopif-a")?.parentFrameId).toBe("main");
+    expect(graph?.frames.find((frame) => frame.frameId === "oopif-c")?.parentFrameId).toBe(
+      "same-process-b",
+    );
+
+    const reordered = buildFrameGraph([sources[2], sources[1], sources[0]]);
+    expect(reordered).toEqual(graph);
+  });
+
+  it("walks deeply nested frame trees without consuming the JavaScript call stack", () => {
+    const depth = 5_000;
+    const root: CdpFrameTreeNode = { frame: { id: "frame-0" } };
+    let parent = root;
+    for (let index = 1; index <= depth; index += 1) {
+      const child: CdpFrameTreeNode = {
+        frame: { id: `frame-${index}`, parentId: `frame-${index - 1}` },
+      };
+      parent.childFrames = [child];
+      parent = child;
+    }
+
+    const graph = buildFrameGraph([{ target: { tabId: 4 }, tree: root }]);
+
+    expect(graph?.frames).toHaveLength(depth + 1);
+    expect(graph?.frames.at(-1)?.frameId).toBe(`frame-${depth}`);
+  });
+
+  it("does not loop when an in-memory frame tree contains an object cycle", () => {
+    const root: CdpFrameTreeNode = { frame: { id: "main" } };
+    root.childFrames = [root];
+
+    expect(buildFrameGraph([{ target: { tabId: 4 }, tree: root }])?.frames).toEqual([
+      { frameId: "main", target: { tabId: 4 } },
+    ]);
+  });
+});

apps/extension/src/browser-driver/chromium-cdp.ts

@@ -31,6 +31,15 @@ import type {
   NetworkEntryKind,
   NetworkResult,
 } from "@/transport/types";
+import {
+  buildFrameGraph,
+  type CdpFrameGraph,
+  type CdpFrameTreeNode,
+  type CdpFrameTreeSource,
+  type CdpTarget,
+} from "./frame-graph";
+
+export type CdpDebuggee = chrome.debugger.Debuggee & { sessionId?: string };
 
 /**
  * Minimal slice of `chrome.debugger` the rest of the extension
@@ -38,21 +47,15 @@ import type {
  * fake without monkey-patching the real `chrome` global.
  */
 export interface CdpDebuggerApi {
-  attach(target: chrome.debugger.Debuggee, requiredVersion: string): Promise<void>;
-  detach(target: chrome.debugger.Debuggee): Promise<void>;
-  sendCommand(
-    target: chrome.debugger.Debuggee,
-    method: string,
-    commandParams?: object,
-  ): Promise<unknown>;
+  attach(target: CdpDebuggee, requiredVersion: string): Promise<void>;
+  detach(target: CdpDebuggee): Promise<void>;
+  sendCommand(target: CdpDebuggee, method: string, commandParams?: object): Promise<unknown>;
   /**
    * Fires for every CDP event (`Page.lifecycleEvent`, `DOM.documentUpdated`,
    * …). The first callback argument is the source debuggee; the second
    * is the CDP method name; the third is the payload.
    */
-  onEvent: chrome.events.Event<
-    (source: chrome.debugger.Debuggee, method: string, params: unknown) => void
-  >;
+  onEvent: chrome.events.Event<(source: CdpDebuggee, method: string, params: unknown) => void>;
   /**
    * Fires when Chrome unilaterally detaches us — most commonly because
    * the tab navigated to a chrome:// URL or the user clicked
@@ -110,6 +113,8 @@ const MAX_CONSOLE_STACK_FRAMES = 20;
 const MAX_NETWORK_BUFFER = 200;
 const MAX_NETWORK_FIELD_LENGTH = 4096;
 const MAX_NETWORK_REQUEST_META = 1024;
+const FRAME_DISCOVERY_TIMEOUT_MS = 1000;
+const FRAME_DISCOVERY_QUIET_MS = 20;
 
 interface ParsedDialogOpening {
   type: JavaScriptDialogType;
@@ -131,6 +136,27 @@ interface NetworkRequestMeta {
   truncated: boolean;
 }
 
+interface FrameDiscoveryState {
+  sessions: Set<string>;
+  pending: Set<Promise<void>>;
+  generation: number;
+}
+
+async function settleBeforeDeadline(promises: Promise<void>[], deadline: number): Promise<boolean> {
+  if (promises.length === 0) return true;
+  const remaining = deadline - Date.now();
+  if (remaining <= 0) return false;
+  let timer: ReturnType<typeof setTimeout> | undefined;
+  const timedOut = await Promise.race([
+    Promise.allSettled(promises).then(() => false),
+    new Promise<boolean>((resolve) => {
+      timer = setTimeout(() => resolve(true), remaining);
+    }),
+  ]);
+  if (timer) clearTimeout(timer);
+  return !timedOut;
+}
+
 /**
  * Wrapper around `chrome.debugger` that owns the "attach once per
  * tabId" cache and exposes typed `send<T>()`.
@@ -149,10 +175,12 @@ export class ChromiumCdp {
   private readonly networkSequences = new Map<number, number>();
   private readonly networkDomainsEnabledTabs = new Set<number>();
   private readonly networkRequestMeta = new Map<number, Map<string, NetworkRequestMeta>>();
+  private readonly frameDiscovery = new Map<number, FrameDiscoveryState>();
   private detachSubscription: { dispose(): void } | null = null;
   private dialogSubscription: { dispose(): void } | null = null;
   private consoleSubscription: { dispose(): void } | null = null;
   private networkSubscription: { dispose(): void } | null = null;
+  private frameTargetSubscription: { dispose(): void } | null = null;
 
   constructor(api: CdpDebuggerApi = chromeDebuggerApi) {
     this.api = api;
@@ -160,6 +188,7 @@ export class ChromiumCdp {
     this.bindDialogHandler();
     this.bindConsoleHandler();
     this.bindNetworkHandler();
+    this.bindFrameTargetHandler();
   }
 
   /** Attach to `tabId` if we haven't already in this driver. */
@@ -177,6 +206,9 @@ export class ChromiumCdp {
         await this.enableConsoleDomains(tabId);
         await this.enableNetworkDomainBestEffort(tabId);
         this.attachedTabs.add(tabId);
+        await this.enableFrameDiscovery({ tabId }).catch((err) => {
+          console.debug("[bsk cdp] frame discovery unavailable", { tabId, err });
+        });
       } catch (err) {
         // A CDP domain enable failed after the raw attach succeeded
         // (e.g. `Page.enable` rejects because the tab just navigated to
@@ -222,6 +254,73 @@ export class ChromiumCdp {
     }
   }
 
+  async sendToTarget<T = unknown>(target: CdpTarget, method: string, params?: object): Promise<T> {
+    if (!this.attachedTabs.has(target.tabId)) {
+      await this.ensureAttached(target.tabId);
+    }
+    try {
+      return (await this.api.sendCommand(target, method, params ?? {})) as T;
+    } catch (err) {
+      throw normalizeError(err);
+    }
+  }
+
+  async getFrameGraph(tabId: number): Promise<CdpFrameGraph> {
+    await this.ensureAttached(tabId);
+    await this.enableFrameDiscovery({ tabId }).catch(() => {});
+    await this.drainFrameAttachTasks(tabId);
+
+    const sources: CdpFrameTreeSource[] = [];
+    const root = await this.sendToTarget<{ frameTree?: CdpFrameTreeNode }>(
+      { tabId },
+      "Page.getFrameTree",
+      {},
+    );
+    if (root.frameTree) sources.push({ target: { tabId }, tree: root.frameTree });
+
+    const sessions = [...(this.frameDiscovery.get(tabId)?.sessions ?? [])];
+    const childTrees = await Promise.all(
+      sessions.map(async (sessionId): Promise<CdpFrameTreeSource | null> => {
+        const target = { tabId, sessionId };
+        try {
+          const reply = await this.sendToTarget<{ frameTree?: CdpFrameTreeNode }>(
+            target,
+            "Page.getFrameTree",
+            {},
+          );
+          return reply.frameTree ? { target, tree: reply.frameTree } : null;
+        } catch {
+          return null;
+        }
+      }),
+    );
+    for (const source of childTrees) {
+      if (source) sources.push(source);
+    }
+
+    const graph = buildFrameGraph(sources);
+    if (!graph) throw new Error("Page.getFrameTree returned no root frame");
+    const frameById = new Map(graph.frames.map((frame) => [frame.frameId, frame]));
+    await Promise.all(
+      graph.frames.map(async (frame) => {
+        if (!frame.parentFrameId) return;
+        const parent = frameById.get(frame.parentFrameId);
+        if (!parent) return;
+        try {
+          const owner = await this.sendToTarget<{ backendNodeId?: number }>(
+            parent.target,
+            "DOM.getFrameOwner",
+            { frameId: frame.frameId },
+          );
+          if (owner.backendNodeId !== undefined) frame.ownerBackendNodeId = owner.backendNodeId;
+        } catch {
+          // The frame may have navigated between tree capture and owner lookup.
+        }
+      }),
+    );
+    return graph;
+  }
+
   /** Return a cursor marking the current dialog sequence for `tabId`. */
   dialogCursor(tabId: number): DialogCursor {
     return this.dialogSequences.get(tabId) ?? 0;
@@ -332,6 +431,7 @@ export class ChromiumCdp {
     this.clearDialogState(tabId);
     this.clearConsoleState(tabId);
     this.clearNetworkState(tabId);
+    this.clearFrameState(tabId);
     try {
       await this.api.detach({ tabId });
     } catch (err) {
@@ -354,7 +454,7 @@ export class ChromiumCdp {
   }
 
   /** Subscribe to all CDP events. Returned disposable removes the listener. */
-  onEvent(handler: (source: chrome.debugger.Debuggee, method: string, params: unknown) => void): {
+  onEvent(handler: (source: CdpDebuggee, method: string, params: unknown) => void): {
     dispose(): void;
   } {
     this.api.onEvent.addListener(handler);
@@ -378,6 +478,7 @@ export class ChromiumCdp {
     this.networkSequences.clear();
     this.networkDomainsEnabledTabs.clear();
     this.networkRequestMeta.clear();
+    this.frameDiscovery.clear();
     await Promise.all(
       tabs.map(async (tabId) => {
         try {
@@ -428,9 +529,104 @@ export class ChromiumCdp {
     }
   }
 
+  private async enableFrameDiscovery(target: CdpTarget): Promise<void> {
+    await this.api.sendCommand(target, "Target.setAutoAttach", {
+      autoAttach: true,
+      waitForDebuggerOnStart: false,
+      flatten: true,
+      filter: [{ type: "iframe", exclude: false }],
+    });
+  }
+
+  private bindFrameTargetHandler(): void {
+    if (this.frameTargetSubscription) return;
+    const listener = (source: CdpDebuggee, method: string, params: unknown) => {
+      const tabId = source.tabId;
+      if (typeof tabId !== "number") return;
+      const raw = (params ?? {}) as Record<string, unknown>;
+      const sessionId = typeof raw.sessionId === "string" ? raw.sessionId : undefined;
+      if (!sessionId) return;
+
+      if (method === "Target.detachedFromTarget") {
+        const state = this.frameDiscovery.get(tabId);
+        if (state?.sessions.delete(sessionId)) state.generation += 1;
+        return;
+      }
+      if (method !== "Target.attachedToTarget") return;
+      const targetInfo = raw.targetInfo as { type?: string } | undefined;
+      if (targetInfo?.type && targetInfo.type !== "iframe") return;
+
+      const state = this.frameDiscoveryState(tabId);
+      if (state.sessions.has(sessionId)) return;
+      state.sessions.add(sessionId);
+      state.generation += 1;
+
+      const task = this.initializeFrameTarget({ tabId, sessionId });
+      state.pending.add(task);
+      void task.finally(() => {
+        state.pending.delete(task);
+      });
+    };
+    this.api.onEvent.addListener(listener);
+    this.frameTargetSubscription = {
+      dispose: () => this.api.onEvent.removeListener(listener),
+    };
+  }
+
+  private async initializeFrameTarget(target: CdpTarget): Promise<void> {
+    try {
+      await this.enableFrameDiscovery(target);
+    } catch (err) {
+      const state = this.frameDiscovery.get(target.tabId);
+      if (state?.sessions.delete(target.sessionId as string)) state.generation += 1;
+      console.debug("[bsk cdp] child frame target initialization failed", { target, err });
+    }
+  }
+
+  private async drainFrameAttachTasks(tabId: number): Promise<void> {
+    const deadline = Date.now() + FRAME_DISCOVERY_TIMEOUT_MS;
+    while (Date.now() < deadline) {
+      const state = this.frameDiscovery.get(tabId);
+      const generation = state?.generation ?? 0;
+      if (state?.pending.size && !(await settleBeforeDeadline([...state.pending], deadline))) break;
+
+      // Target.setAutoAttach may enqueue the next attachedToTarget event after
+      // its command promise settles. Wait for a short quiet window, then finish
+      // only if neither the state object nor its generation changed.
+      const quietTime = Math.min(FRAME_DISCOVERY_QUIET_MS, deadline - Date.now());
+      if (quietTime <= 0) break;
+      await new Promise<void>((resolve) => setTimeout(resolve, quietTime));
+      const current = this.frameDiscovery.get(tabId);
+      if (
+        current === state &&
+        (current?.generation ?? 0) === generation &&
+        !current?.pending.size
+      ) {
+        return;
+      }
+    }
+    console.debug("[bsk cdp] frame discovery did not reach quiescence before timeout", { tabId });
+  }
+
+  private frameDiscoveryState(tabId: number): FrameDiscoveryState {
+    const existing = this.frameDiscovery.get(tabId);
+    if (existing) return existing;
+    const created: FrameDiscoveryState = {
+      sessions: new Set(),
+      pending: new Set(),
+      generation: 0,
+    };
+    this.frameDiscovery.set(tabId, created);
+    return created;
+  }
+
+  private clearFrameState(tabId: number): void {
+    this.frameDiscovery.delete(tabId);
+  }
+
   private bindDialogHandler(): void {
     if (this.dialogSubscription) return;
-    const listener = (source: chrome.debugger.Debuggee, method: string, params: unknown) => {
+    const listener = (source: CdpDebuggee, method: string, params: unknown) => {
       if (method !== "Page.javascriptDialogOpening") return;
       const tabId = source.tabId;
       if (typeof tabId !== "number") return;
@@ -614,6 +810,7 @@ export class ChromiumCdp {
         this.clearDialogState(source.tabId);
         this.clearConsoleState(source.tabId);
         this.clearNetworkState(source.tabId);
+        this.clearFrameState(source.tabId);
       }
     };
     this.api.onDetach.addListener(listener);
@@ -632,6 +829,8 @@ export class ChromiumCdp {
     this.consoleSubscription = null;
     this.networkSubscription?.dispose();
     this.networkSubscription = null;
+    this.frameTargetSubscription?.dispose();
+    this.frameTargetSubscription = null;
   }
 
   /** Detach tabs only when no other live session has claimed them. */

apps/extension/src/browser-driver/frame-graph.ts

@@ -0,0 +1,107 @@
+export interface CdpTarget {
+  tabId: number;
+  sessionId?: string;
+}
+
+export interface CdpFrame {
+  frameId: string;
+  parentFrameId?: string;
+  ownerBackendNodeId?: number;
+  url?: string;
+  target: CdpTarget;
+}
+
+export interface CdpFrameGraph {
+  rootFrameId: string;
+  frames: CdpFrame[];
+}
+
+export interface CdpFrameTreeNode {
+  frame: {
+    id: string;
+    parentId?: string;
+    url?: string;
+  };
+  childFrames?: CdpFrameTreeNode[];
+}
+
+export interface CdpFrameTreeSource {
+  target: CdpTarget;
+  tree: CdpFrameTreeNode;
+}
+
+export function cdpTargetKey(target: CdpTarget): string {
+  return `${target.tabId}:${target.sessionId ?? "root"}`;
+}
+
+function mergeFrame(
+  frames: Map<string, CdpFrame>,
+  order: string[],
+  node: CdpFrameTreeNode,
+  target: CdpTarget,
+): void {
+  const existing = frames.get(node.frame.id);
+  if (!existing) order.push(node.frame.id);
+  const parentFrameId = node.frame.parentId || existing?.parentFrameId;
+  const url = node.frame.url || existing?.url;
+  frames.set(node.frame.id, {
+    frameId: node.frame.id,
+    ...(parentFrameId ? { parentFrameId } : {}),
+    ...(url ? { url } : {}),
+    target,
+  });
+}
+
+function walkFrameTree(
+  source: CdpFrameTreeSource,
+  targetByRootFrameId: ReadonlyMap<string, CdpTarget>,
+  frames: Map<string, CdpFrame>,
+  order: string[],
+): void {
+  const stack: Array<{ node: CdpFrameTreeNode; inheritedTarget: CdpTarget }> = [
+    { node: source.tree, inheritedTarget: source.target },
+  ];
+  const expanded = new WeakSet<CdpFrameTreeNode>();
+
+  while (stack.length > 0) {
+    const current = stack.pop();
+    if (!current || expanded.has(current.node)) continue;
+    expanded.add(current.node);
+
+    const target = targetByRootFrameId.get(current.node.frame.id) ?? current.inheritedTarget;
+    mergeFrame(frames, order, current.node, target);
+
+    const children = current.node.childFrames ?? [];
+    for (let index = children.length - 1; index >= 0; index -= 1) {
+      stack.push({ node: children[index], inheritedTarget: target });
+    }
+  }
+}
+
+export function buildFrameGraph(sources: CdpFrameTreeSource[]): CdpFrameGraph | null {
+  const rootSource = sources.find((source) => source.target.sessionId === undefined) ?? sources[0];
+  if (!rootSource) return null;
+
+  const frames = new Map<string, CdpFrame>();
+  const order: string[] = [];
+  const targetByRootFrameId = new Map(
+    sources.map((source) => [source.tree.frame.id, source.target] as const),
+  );
+  // The top-level source is authoritative if malformed input reports the same
+  // frame as both a root target and a child target.
+  targetByRootFrameId.set(rootSource.tree.frame.id, rootSource.target);
+
+  walkFrameTree(rootSource, targetByRootFrameId, frames, order);
+  for (const source of sources) {
+    if (source === rootSource) continue;
+    walkFrameTree(source, targetByRootFrameId, frames, order);
+  }
+
+  return {
+    rootFrameId: rootSource.tree.frame.id,
+    frames: order.flatMap((frameId) => {
+      const frame = frames.get(frameId);
+      return frame ? [frame] : [];
+    }),
+  };
+}

apps/extension/src/content/HelpRequestOverlay.tsx

@@ -8,6 +8,7 @@ import {
   useRef,
   useState,
 } from "react";
+import type { HelpHighlightRect } from "@/lib/help-bridge";
 import logoUrl from "../../assets/logo.png";
 
 export interface HelpRequestData {
@@ -19,6 +20,10 @@ export interface HelpRequestData {
   displayMode?: "full" | "compact";
   /** CSS selectors to scroll to + flash-highlight. */
   selectors: string[];
+  /** Pre-resolved top-viewport rectangles, used for cross-frame targets. */
+  rects?: HelpHighlightRect[];
+  /** Re-resolve cross-frame rectangles after scroll, resize, or layout changes. */
+  refreshRects?: () => Promise<HelpHighlightRect[] | undefined>;
   onContinue: (note: string) => void;
   onCancel: () => void;
 }
@@ -49,6 +54,7 @@ const PANEL_WIDTH = 420;
 const FALLBACK_PANEL_H = 180;
 const FALLBACK_PANEL_H_COLLAPSED = 44;
 const EMPTY_SELECTORS: string[] = [];
+const EMPTY_RECTS: HelpHighlightRect[] = [];
 
 function clamp(value: number, min: number, max: number): number {
   return Math.max(min, Math.min(max, value));
@@ -179,6 +185,8 @@ export function HelpRequestOverlay({ request }: Props) {
   const { t } = useTranslation("extension");
   const isCompact = request?.displayMode === "compact";
   const effectiveSelectors = isCompact ? EMPTY_SELECTORS : (request?.selectors ?? EMPTY_SELECTORS);
+  const [liveRects, setLiveRects] = useState<HelpHighlightRect[] | null>(null);
+  const effectiveRects = isCompact ? EMPTY_RECTS : (liveRects ?? request?.rects ?? EMPTY_RECTS);
   const [note, setNote] = useState("");
   const [collapsed, setCollapsed] = useState(false);
   const [boxes, setBoxes] = useState<Box[]>([]);
@@ -201,6 +209,7 @@ export function HelpRequestOverlay({ request }: Props) {
   // Reset the note and collapse state whenever a new request appears.
   useEffect(() => {
     setNote("");
+    setLiveRects(null);
     setCollapsed(false);
     placementRef.current = null;
     setDragPos(null);
@@ -209,6 +218,32 @@ export function HelpRequestOverlay({ request }: Props) {
     dragStartRef.current = null;
   }, [request?.id]);
 
+  useEffect(() => {
+    if (!request || isCompact || !request.refreshRects) return;
+    let disposed = false;
+    let resolving = false;
+    const refresh = async () => {
+      if (resolving) return;
+      resolving = true;
+      try {
+        const rects = await request.refreshRects?.();
+        if (!disposed && rects) setLiveRects(rects);
+      } finally {
+        resolving = false;
+      }
+    };
+    const onViewportChange = () => void refresh();
+    window.addEventListener("scroll", onViewportChange, true);
+    window.addEventListener("resize", onViewportChange);
+    const interval = window.setInterval(onViewportChange, 500);
+    return () => {
+      disposed = true;
+      window.removeEventListener("scroll", onViewportChange, true);
+      window.removeEventListener("resize", onViewportChange);
+      window.clearInterval(interval);
+    };
+  }, [request, isCompact]);
+
   // Scroll the first matched target into view once per distinct request id.
   useLayoutEffect(() => {
     if (!request) return;
@@ -231,7 +266,7 @@ export function HelpRequestOverlay({ request }: Props) {
       return;
     }
     const update = () => {
-      const nextBoxes = measure(effectiveSelectors);
+      const nextBoxes = [...measure(effectiveSelectors), ...effectiveRects];
       setBoxes((prev) => (boxesEqual(prev, nextBoxes) ? prev : nextBoxes));
     };
     update();
@@ -243,7 +278,7 @@ export function HelpRequestOverlay({ request }: Props) {
       window.removeEventListener("resize", update);
       window.clearInterval(interval);
     };
-  }, [request, effectiveSelectors]);
+  }, [request, effectiveSelectors, effectiveRects]);
 
   const onHeaderPointerDown = useCallback(
     (e: ReactPointerEvent<HTMLDivElement>) => {
@@ -305,7 +340,7 @@ export function HelpRequestOverlay({ request }: Props) {
       return;
     }
     if (userMovedRef.current) return;
-    const measured = boxes.length > 0 ? boxes : measure(effectiveSelectors);
+    const measured = boxes.length > 0 ? boxes : [...measure(effectiveSelectors), ...effectiveRects];
     const anchor = unionBox(measured);
     const banner = bannerRef.current;
     if (!anchor || !banner) {
@@ -329,7 +364,7 @@ export function HelpRequestOverlay({ request }: Props) {
     }
     const nextPos = placePanel(anchor, panelW, panelH, placementRef.current.placement);
     setPanelPos((prev) => (panelPosEqual(prev, nextPos) ? prev : nextPos));
-  }, [request, boxes, collapsed, isCompact, effectiveSelectors]);
+  }, [request, boxes, collapsed, isCompact, effectiveSelectors, effectiveRects]);
 
   if (!request) return null;
 

apps/extension/src/content/__tests__/HelpRequestOverlay.test.tsx

@@ -1,6 +1,6 @@
 import { i18n } from "@browser-skill/i18n";
 import { I18nextProvider } from "@browser-skill/i18n/react";
-import { cleanup, fireEvent, render, screen } from "@testing-library/react";
+import { cleanup, fireEvent, render, screen, waitFor } from "@testing-library/react";
 import { createElement } from "react";
 import { afterEach, describe, expect, it, vi } from "vitest";
 import { type HelpRequestData, HelpRequestOverlay } from "../HelpRequestOverlay";
@@ -38,6 +38,36 @@ describe("HelpRequestOverlay", () => {
     expect(screen.getByText("Please complete the captcha")).toBeTruthy();
   });
 
+  it("renders explicit viewport rectangles for cross-frame targets", () => {
+    const { container } = renderOverlay(
+      baseRequest({ rects: [{ top: 30, left: 40, width: 120, height: 50 }] }),
+    );
+    const highlight = container.querySelector<HTMLElement>("[data-slot='help-highlight']");
+    expect(highlight?.style.top).toBe("30px");
+    expect(highlight?.style.left).toBe("40px");
+    expect(highlight?.style.width).toBe("120px");
+    expect(highlight?.style.height).toBe("50px");
+  });
+
+  it("re-resolves cross-frame rectangles after the viewport changes", async () => {
+    const refreshRects = vi.fn(async () => [{ top: 80, left: 90, width: 140, height: 60 }]);
+    const { container } = renderOverlay(
+      baseRequest({
+        rects: [{ top: 30, left: 40, width: 120, height: 50 }],
+        refreshRects,
+      }),
+    );
+
+    window.dispatchEvent(new Event("scroll"));
+
+    await waitFor(() => {
+      const highlight = container.querySelector<HTMLElement>("[data-slot='help-highlight']");
+      expect(highlight?.style.top).toBe("80px");
+      expect(highlight?.style.left).toBe("90px");
+    });
+    expect(refreshRects).toHaveBeenCalledTimes(1);
+  });
+
   it("keeps the inactive render stable", () => {
     const { container, rerender } = renderOverlay(null);
 

apps/extension/src/content/__tests__/help-request.test.ts

@@ -0,0 +1,64 @@
+import { describe, expect, it, vi } from "vitest";
+import { createHelpRequestData } from "../help-request";
+
+describe("createHelpRequestData", () => {
+  it("gives live delivery and recovery the same rectangle refresh mechanism", async () => {
+    const finish = vi.fn();
+    const query = vi.fn(async () => ({
+      active: true,
+      request: {
+        requestId: "help-1",
+        prompt: "complete the challenge",
+        selectors: [],
+        rects: [{ top: 80, left: 90, width: 140, height: 60 }],
+        timeoutMs: 1_000,
+      },
+    }));
+    const request = createHelpRequestData(
+      {
+        requestId: "help-1",
+        prompt: "complete the challenge",
+        selectors: [],
+        rects: [{ top: 10, left: 20, width: 30, height: 40 }],
+        timeoutMs: 1_000,
+      },
+      { finish, query },
+    );
+
+    await expect(request.refreshRects?.()).resolves.toEqual([
+      { top: 80, left: 90, width: 140, height: 60 },
+    ]);
+    request.onContinue(" done ");
+    request.onCancel();
+
+    expect(query).toHaveBeenCalledOnce();
+    expect(finish).toHaveBeenNthCalledWith(1, "help-1", "continued", " done ");
+    expect(finish).toHaveBeenNthCalledWith(2, "help-1", "cancelled");
+  });
+
+  it("ignores rectangle replies for a replaced request", async () => {
+    const request = createHelpRequestData(
+      {
+        requestId: "help-1",
+        prompt: "complete the challenge",
+        selectors: [],
+        timeoutMs: 1_000,
+      },
+      {
+        finish: vi.fn(),
+        query: vi.fn(async () => ({
+          active: true,
+          request: {
+            requestId: "help-2",
+            prompt: "new challenge",
+            selectors: [],
+            rects: [{ top: 1, left: 2, width: 3, height: 4 }],
+            timeoutMs: 1_000,
+          },
+        })),
+      },
+    );
+
+    await expect(request.refreshRects?.()).resolves.toBeUndefined();
+  });
+});

apps/extension/src/content/__tests__/record-capture.test.ts

@@ -1,10 +1,12 @@
 import { beforeEach, describe, expect, it, vi } from "vitest";
-import { RECORD_START, RECORD_STOP, type RecordStepPayload } from "@/lib/record-bridge";
-import { handleRecordContentMessage, startRecordCapture } from "../record-capture";
+import type { RecordStepPayload } from "@/lib/record-bridge";
+import { startRecordCapture } from "../record-capture";
 
 vi.stubGlobal("chrome", {
   runtime: {
-    sendMessage: vi.fn(() => Promise.resolve()),
+    sendMessage: vi.fn((message: { sequence?: number }) =>
+      Promise.resolve({ ok: true, sequence: message.sequence }),
+    ),
   },
 });
 
@@ -46,117 +48,6 @@ function click(el: Element): void {
   el.dispatchEvent(new MouseEvent("click", { bubbles: true, button: 0, detail: 1 }));
 }
 
-describe("handleRecordContentMessage stop/cancel", () => {
-  it("ignores STOP when no recording is active", () => {
-    const dispose = vi.fn();
-    const onStop = vi.fn();
-    const setActiveRequestId = vi.fn();
-    const setCapture = vi.fn();
-    const sendResponse = vi.fn();
-
-    const needsAsync = handleRecordContentMessage(
-      { type: RECORD_STOP, requestId: "rec-stale" },
-      {
-        activeRequestId: null,
-        capture: { dispose },
-        setActiveRequestId,
-        setCapture,
-        onStart: vi.fn(),
-        onStop,
-      },
-      sendResponse,
-    );
-
-    expect(needsAsync).toBe(false);
-    expect(dispose).not.toHaveBeenCalled();
-    expect(onStop).not.toHaveBeenCalled();
-    expect(setActiveRequestId).not.toHaveBeenCalled();
-    expect(setCapture).not.toHaveBeenCalled();
-    expect(sendResponse).not.toHaveBeenCalled();
-  });
-
-  it("ignores STOP for a mismatched requestId", () => {
-    const dispose = vi.fn();
-    const onStop = vi.fn();
-
-    const needsAsync = handleRecordContentMessage(
-      { type: RECORD_STOP, requestId: "rec-other" },
-      {
-        activeRequestId: "rec-1",
-        capture: { dispose },
-        setActiveRequestId: vi.fn(),
-        setCapture: vi.fn(),
-        onStart: vi.fn(),
-        onStop,
-      },
-      vi.fn(),
-    );
-
-    expect(needsAsync).toBe(false);
-    expect(dispose).not.toHaveBeenCalled();
-    expect(onStop).not.toHaveBeenCalled();
-  });
-
-  it("keeps a failed STOP retryable and redelivers its recorded step", async () => {
-    document.body.innerHTML = `<label for="retry">Draft</label><input id="retry" />`;
-    const sendMessage = vi.mocked(chrome.runtime.sendMessage);
-    sendMessage.mockReset();
-    sendMessage.mockRejectedValueOnce(new Error("service worker unavailable"));
-    sendMessage.mockResolvedValueOnce(undefined);
-
-    let activeRequestId: string | null = null;
-    let capture: ReturnType<typeof startRecordCapture> | null = null;
-    const onStop = vi.fn();
-    const dispatch = (
-      message:
-        | { type: typeof RECORD_START; requestId: string; startedAtMs?: number }
-        | { type: typeof RECORD_STOP; requestId: string },
-      sendResponse?: (response: unknown) => void,
-    ) =>
-      handleRecordContentMessage(
-        message,
-        {
-          activeRequestId,
-          capture,
-          setActiveRequestId: (id) => {
-            activeRequestId = id;
-          },
-          setCapture: (next) => {
-            capture = next;
-          },
-          onStart: vi.fn(),
-          onStop,
-        },
-        sendResponse,
-      );
-
-    dispatch({ type: RECORD_START, requestId: "rec-retry" });
-    const input = document.querySelector("input")!;
-    input.dispatchEvent(new FocusEvent("focusin", { bubbles: true }));
-    input.value = "dirty final value";
-    input.dispatchEvent(new Event("input", { bubbles: true }));
-
-    const firstResponse = vi.fn();
-    expect(dispatch({ type: RECORD_STOP, requestId: "rec-retry" }, firstResponse)).toBe(true);
-    await vi.waitFor(() =>
-      expect(firstResponse).toHaveBeenCalledWith({
-        ok: false,
-        error: "failed to deliver one or more recorded steps",
-      }),
-    );
-    expect(activeRequestId).toBe("rec-retry");
-    expect(onStop).not.toHaveBeenCalled();
-
-    const retryResponse = vi.fn();
-    expect(dispatch({ type: RECORD_STOP, requestId: "rec-retry" }, retryResponse)).toBe(true);
-    await vi.waitFor(() => expect(retryResponse).toHaveBeenCalledWith({ ok: true }));
-
-    expect(sendMessage).toHaveBeenCalledTimes(2);
-    expect(activeRequestId).toBeNull();
-    expect(onStop).toHaveBeenCalledTimes(1);
-  });
-});
-
 describe("record-capture semantic", () => {
   let steps: RecordStepPayload[];
 
@@ -233,6 +124,19 @@ describe("record-capture semantic", () => {
     ]);
   });
 
+  it("does not emit page navigation from a child Document capture", () => {
+    const originalUrl = location.href;
+    const capture = startRecordCapture("rec-child", (step) => steps.push(step), {
+      captureNavigation: false,
+    });
+
+    history.pushState({}, "", "#inside-frame");
+    expect(steps).toEqual([]);
+
+    capture.dispose();
+    history.replaceState({}, "", originalUrl);
+  });
+
   it("does not record clicks on anonymous layout divs", () => {
     document.body.innerHTML = `
       <div id="chrome">page chrome</div>

apps/extension/src/content/__tests__/record-frame-agent.test.ts

@@ -0,0 +1,102 @@
+import { afterEach, describe, expect, it, vi } from "vitest";
+import type { RecordFramePortMessage } from "@/lib/recording/frame-bridge";
+import { RECORD_FRAME_START } from "@/lib/recording/frame-bridge";
+import { RECORD_DOCUMENT_ATTRIBUTE } from "@/shared/recording-document-identity";
+import { RecordFrameAgent } from "../recording/frame-agent";
+
+class PortListeners<T extends (...args: never[]) => unknown> {
+  readonly values = new Set<T>();
+  addListener = (listener: T) => this.values.add(listener);
+  removeListener = (listener: T) => this.values.delete(listener);
+}
+
+function portHarness() {
+  const onMessage = new PortListeners<(message: unknown) => void>();
+  const onDisconnect = new PortListeners<() => void>();
+  const outbound: RecordFramePortMessage[] = [];
+  const port = {
+    name: "bsk-record-frame",
+    onMessage,
+    onDisconnect,
+    postMessage(message: RecordFramePortMessage) {
+      outbound.push(message);
+      if (message.type === "ready") {
+        queueMicrotask(() => {
+          for (const listener of onMessage.values) {
+            listener({
+              type: "ready_ack",
+              requestId: message.requestId,
+              producerId: message.producerId,
+            });
+          }
+        });
+      }
+    },
+    disconnect: vi.fn(),
+  } as unknown as chrome.runtime.Port;
+  return {
+    port,
+    outbound,
+    receive(message: RecordFramePortMessage) {
+      for (const listener of onMessage.values) listener(message);
+    },
+  };
+}
+
+describe("RecordFrameAgent", () => {
+  afterEach(() => {
+    document.documentElement.removeAttribute(RECORD_DOCUMENT_ATTRIBUTE);
+    document.body.replaceChildren();
+  });
+
+  it("keeps a failed stop retryable and flushes the final dirty fill", async () => {
+    const harness = portHarness();
+    const sendMessage = vi
+      .fn<(message: { sequence: number }) => Promise<unknown>>()
+      .mockRejectedValueOnce(new Error("offline"))
+      .mockRejectedValueOnce(new Error("still offline"))
+      .mockImplementation(async (message) => ({ ok: true, sequence: message.sequence }));
+    vi.stubGlobal("chrome", {
+      runtime: {
+        connect: vi.fn(() => harness.port),
+        sendMessage,
+      },
+    });
+    document.body.innerHTML = `<label for="draft">Draft</label><input id="draft" />`;
+    const agent = new RecordFrameAgent();
+
+    await expect(
+      agent.start({ type: RECORD_FRAME_START, requestId: "rec-1", startedAtMs: 10 }),
+    ).resolves.toEqual({ ok: true });
+    expect(document.documentElement.hasAttribute(RECORD_DOCUMENT_ATTRIBUTE)).toBe(true);
+
+    const input = document.querySelector("input")!;
+    input.dispatchEvent(new FocusEvent("focusin", { bubbles: true }));
+    input.value = "final value";
+    input.dispatchEvent(new Event("input", { bubbles: true }));
+
+    harness.receive({ type: "stop", requestId: "rec-1", commandId: "stop-1" });
+    await vi.waitFor(() =>
+      expect(harness.outbound).toContainEqual({
+        type: "stopped",
+        requestId: "rec-1",
+        commandId: "stop-1",
+        ok: false,
+        error: "failed to deliver one or more recorded steps",
+      }),
+    );
+    expect(document.documentElement.hasAttribute(RECORD_DOCUMENT_ATTRIBUTE)).toBe(true);
+
+    harness.receive({ type: "stop", requestId: "rec-1", commandId: "stop-2" });
+    await vi.waitFor(() =>
+      expect(harness.outbound).toContainEqual({
+        type: "stopped",
+        requestId: "rec-1",
+        commandId: "stop-2",
+        ok: true,
+      }),
+    );
+    expect(sendMessage).toHaveBeenCalledTimes(3);
+    expect(document.documentElement.hasAttribute(RECORD_DOCUMENT_ATTRIBUTE)).toBe(false);
+  });
+});

apps/extension/src/content/__tests__/record-step-delivery.test.ts

@@ -0,0 +1,36 @@
+import { describe, expect, it, vi } from "vitest";
+import type { RecordStepMessage } from "@/lib/record-bridge";
+import { RecordStepDelivery } from "../record-step-delivery";
+
+describe("RecordStepDelivery", () => {
+  it("retries from the first unacknowledged sequence before sending later steps", async () => {
+    const sent: RecordStepMessage[] = [];
+    const send = vi.fn(async (message: RecordStepMessage) => {
+      sent.push(message);
+      if (sent.length === 1) throw new Error("service worker unavailable");
+      return { ok: true, sequence: message.sequence };
+    });
+    const delivery = new RecordStepDelivery("rec-ordered", send, "document-1");
+
+    delivery.enqueue({ op: "click", target: { tag: "button", name: "First" } });
+    delivery.enqueue({ op: "click", target: { tag: "button", name: "Second" } });
+
+    await expect(delivery.flush()).resolves.toBe(true);
+    expect(sent.map((message) => message.sequence)).toEqual([1, 1, 2]);
+    expect(sent.map((message) => message.step.target?.name)).toEqual(["First", "First", "Second"]);
+  });
+
+  it("keeps unacknowledged steps pending for a later flush", async () => {
+    const send = vi
+      .fn<(message: RecordStepMessage) => Promise<unknown>>()
+      .mockRejectedValueOnce(new Error("offline"))
+      .mockRejectedValueOnce(new Error("still offline"))
+      .mockImplementation(async (message) => ({ ok: true, sequence: message.sequence }));
+    const delivery = new RecordStepDelivery("rec-retry", send, "document-1");
+    delivery.enqueue({ op: "click", target: { tag: "button", name: "Save" } });
+
+    await expect(delivery.flush()).resolves.toBe(false);
+    await expect(delivery.flush()).resolves.toBe(true);
+    expect(send).toHaveBeenCalledTimes(3);
+  });
+});

apps/extension/src/content/help-request.ts

@@ -0,0 +1,31 @@
+import type { HelpQueryResponse, HelpRequestMessage } from "@/lib/help-bridge";
+import type { HelpRequestData } from "./HelpRequestOverlay";
+
+export interface HelpRequestDataDeps {
+  finish(requestId: string, outcome: "continued" | "cancelled", note?: string): void;
+  query(): Promise<HelpQueryResponse | undefined>;
+}
+
+/** Build the content-side representation shared by live delivery and recovery. */
+export function createHelpRequestData(
+  message: Omit<HelpRequestMessage, "type">,
+  deps: HelpRequestDataDeps,
+): HelpRequestData {
+  return {
+    id: message.requestId,
+    prompt: message.prompt,
+    ...(message.title ? { title: message.title } : {}),
+    ...(message.displayMode ? { displayMode: message.displayMode } : {}),
+    selectors: message.selectors,
+    rects: message.rects,
+    refreshRects: async () => {
+      const response = await deps.query();
+      return response?.active && response.request?.requestId === message.requestId
+        ? response.request.rects
+        : undefined;
+    },
+    onContinue: (note: string) =>
+      deps.finish(message.requestId, "continued", note.trim() ? note : undefined),
+    onCancel: () => deps.finish(message.requestId, "cancelled"),
+  };
+}

apps/extension/src/content/record-capture.ts

@@ -12,22 +12,8 @@ import {
   hasDirectHoverInteractiveSignal,
   hasStrongHoverExpansionSignal,
 } from "@/lib/hover-trigger-policy";
-import {
-  isRecordCancelMessage,
-  isRecordStartMessage,
-  isRecordStopMessage,
-  RECORD_CANCEL,
-  RECORD_START,
-  RECORD_STEP,
-  RECORD_STOP,
-  type RecordCancelMessage,
-  type RecordStartAck,
-  type RecordStartMessage,
-  type RecordStepPayload,
-  type RecordStopAck,
-  type RecordStopMessage,
-} from "@/lib/record-bridge";
-import { shouldRecordPress } from "@/lib/trace-reducer";
+import type { RecordStepPayload } from "@/lib/record-bridge";
+import { shouldRecordPress } from "@/lib/recording/draft-policy";
 import {
   closestHoverSurfaceCandidate,
   collectHoverSurfaceStates,
@@ -37,37 +23,6 @@ import {
   isLikelyHoverSurfaceOwner,
 } from "./record-hover-surface";
 
-const pendingStepSends = new Map<string, Set<Promise<boolean>>>();
-const failedStepDeliveries = new Set<string>();
-const failedStepPayloads = new Map<string, RecordStepPayload[]>();
-const pendingStopFlushes = new Map<string, Promise<RecordStopAck>>();
-const knownRecordRequests = new Set<string>();
-
-function deliverRecordStep(requestId: string, step: RecordStepPayload): Promise<boolean> {
-  const payload = { type: RECORD_STEP, requestId, step };
-  return Promise.resolve(chrome.runtime.sendMessage(payload)).then(
-    () => true,
-    () => {
-      failedStepDeliveries.add(requestId);
-      const failed = failedStepPayloads.get(requestId) ?? [];
-      failed.push(step);
-      failedStepPayloads.set(requestId, failed);
-      return false;
-    },
-  );
-}
-
-function sendRecordStep(requestId: string, step: RecordStepPayload): void {
-  const pending = deliverRecordStep(requestId, step);
-  const sends = pendingStepSends.get(requestId) ?? new Set<Promise<boolean>>();
-  sends.add(pending);
-  pendingStepSends.set(requestId, sends);
-  void pending.then(() => {
-    sends.delete(pending);
-    if (sends.size === 0) pendingStepSends.delete(requestId);
-  });
-}
-
 export interface RecordCaptureController {
   dispose(): void;
 }
@@ -169,12 +124,8 @@ function nearbyFillableFromSearchChrome(target: Element): FillableElement | null
 
 function captureGeometry(el: Element): RecordStepPayload["geometry"] {
   const rect = el.getBoundingClientRect();
-  const style = window.getComputedStyle(el);
   return {
     rect: { x: rect.x, y: rect.y, w: rect.width, h: rect.height },
-    scrollX: window.scrollX,
-    scrollY: window.scrollY,
-    position: style.position,
     tag: el.tagName.toLowerCase(),
   };
 }
@@ -408,6 +359,7 @@ function scheduleInputCompletionCommit(
 export function startRecordCapture(
   _requestId: string,
   sendStep: (step: RecordStepPayload) => void,
+  options: { captureNavigation?: boolean } = {},
 ): RecordCaptureController {
   const emitStep = (step: RecordStepPayload) => {
     sendStep({ page_url: location.href, ...step });
@@ -944,26 +896,33 @@ export function startRecordCapture(
   document.addEventListener("change", onChange, true);
   document.addEventListener("keydown", onKeyDown, true);
 
-  const urlObserver = new MutationObserver(() => emitNavigateIfChanged());
-  urlObserver.observe(document, { subtree: true, childList: true });
+  const captureNavigation = options.captureNavigation ?? true;
+  const urlObserver = captureNavigation
+    ? new MutationObserver(() => emitNavigateIfChanged())
+    : null;
+  urlObserver?.observe(document, { subtree: true, childList: true });
   const onUrlEvent = () => emitNavigateIfChanged();
   // Wrap: passing commitFillSession directly would forward the DOM event as
   // the `commit` argument and stamp it onto the recorded fill step.
   const onPageHide = () => commitFillSession();
-  window.addEventListener("hashchange", onUrlEvent);
-  window.addEventListener("popstate", onUrlEvent);
+  if (captureNavigation) {
+    window.addEventListener("hashchange", onUrlEvent);
+    window.addEventListener("popstate", onUrlEvent);
+  }
   window.addEventListener("pagehide", onPageHide);
 
   const originalPushState = history.pushState;
   const originalReplaceState = history.replaceState;
-  history.pushState = function (...args: Parameters<History["pushState"]>) {
-    originalPushState.apply(this, args);
-    emitNavigateIfChanged(navigationActionPending ? true : undefined);
-  };
-  history.replaceState = function (...args: Parameters<History["replaceState"]>) {
-    originalReplaceState.apply(this, args);
-    emitNavigateIfChanged(navigationActionPending ? true : undefined);
-  };
+  if (captureNavigation) {
+    history.pushState = function (...args: Parameters<History["pushState"]>) {
+      originalPushState.apply(this, args);
+      emitNavigateIfChanged(navigationActionPending ? true : undefined);
+    };
+    history.replaceState = function (...args: Parameters<History["replaceState"]>) {
+      originalReplaceState.apply(this, args);
+      emitNavigateIfChanged(navigationActionPending ? true : undefined);
+    };
+  }
 
   return {
     dispose() {
@@ -977,118 +936,16 @@ export function startRecordCapture(
       document.removeEventListener("compositionend", onCompositionEnd, true);
       document.removeEventListener("change", onChange, true);
       document.removeEventListener("keydown", onKeyDown, true);
-      urlObserver.disconnect();
-      window.removeEventListener("hashchange", onUrlEvent);
-      window.removeEventListener("popstate", onUrlEvent);
+      urlObserver?.disconnect();
+      if (captureNavigation) {
+        window.removeEventListener("hashchange", onUrlEvent);
+        window.removeEventListener("popstate", onUrlEvent);
+      }
       window.removeEventListener("pagehide", onPageHide);
-      history.pushState = originalPushState;
-      history.replaceState = originalReplaceState;
+      if (captureNavigation) {
+        history.pushState = originalPushState;
+        history.replaceState = originalReplaceState;
+      }
     },
   };
 }
-
-export type RecordContentMessage = RecordStartMessage | RecordStopMessage | RecordCancelMessage;
-
-export function isRecordContentMessage(msg: unknown): msg is RecordContentMessage {
-  return isRecordStartMessage(msg) || isRecordStopMessage(msg) || isRecordCancelMessage(msg);
-}
-
-export function handleRecordContentMessage(
-  message: RecordContentMessage,
-  state: {
-    activeRequestId: string | null;
-    capture: RecordCaptureController | null;
-    setActiveRequestId(id: string | null): void;
-    setCapture(capture: RecordCaptureController | null): void;
-    onStart(requestId: string, startedAtMs?: number): void;
-    onStop(): void;
-  },
-  sendResponse?: (response: RecordStartAck | RecordStopAck) => void,
-): boolean {
-  if (isRecordStartMessage(message)) {
-    if (!knownRecordRequests.has(message.requestId)) {
-      knownRecordRequests.add(message.requestId);
-      failedStepDeliveries.delete(message.requestId);
-      failedStepPayloads.delete(message.requestId);
-    }
-    state.capture?.dispose();
-    state.setCapture(
-      startRecordCapture(message.requestId, (step) => {
-        sendRecordStep(message.requestId, step);
-      }),
-    );
-    state.setActiveRequestId(message.requestId);
-    state.onStart(message.requestId, message.startedAtMs);
-    sendResponse?.({ ok: true });
-    return sendResponse !== undefined;
-  }
-
-  if (isRecordStopMessage(message) || isRecordCancelMessage(message)) {
-    // Require an active recording that matches this requestId — otherwise a
-    // stray STOP/CANCEL (e.g. after teardown) would still run finishStop and
-    // clear overlay state even though nothing was capturing.
-    if (!state.activeRequestId || state.activeRequestId !== message.requestId) {
-      return false;
-    }
-    state.capture?.dispose();
-    state.setCapture(null);
-    const finishStop = () => {
-      state.onStop();
-      state.setActiveRequestId(null);
-    };
-    if (isRecordStopMessage(message) && sendResponse) {
-      const existingFlush = pendingStopFlushes.get(message.requestId);
-      if (existingFlush) {
-        void existingFlush.then(sendResponse);
-        return true;
-      }
-
-      const failedBeforeStop = [...(failedStepPayloads.get(message.requestId) ?? [])];
-      if (failedBeforeStop.length > 0) {
-        failedStepPayloads.delete(message.requestId);
-        failedStepDeliveries.delete(message.requestId);
-      }
-      const pending = [...(pendingStepSends.get(message.requestId) ?? [])];
-      const flush = Promise.all(pending).then(async (delivered): Promise<RecordStopAck> => {
-        const retried = await Promise.all(
-          failedBeforeStop.map((step) => deliverRecordStep(message.requestId, step)),
-        );
-        const succeeded =
-          delivered.every(Boolean) &&
-          retried.every(Boolean) &&
-          !failedStepDeliveries.has(message.requestId);
-        if (succeeded) {
-          failedStepDeliveries.delete(message.requestId);
-          failedStepPayloads.delete(message.requestId);
-          knownRecordRequests.delete(message.requestId);
-          finishStop();
-        }
-        return succeeded
-          ? { ok: true }
-          : {
-              ok: false,
-              error: "failed to deliver one or more recorded steps",
-            };
-      });
-      pendingStopFlushes.set(message.requestId, flush);
-      void flush.then(sendResponse).finally(() => {
-        if (pendingStopFlushes.get(message.requestId) === flush) {
-          pendingStopFlushes.delete(message.requestId);
-        }
-      });
-      return true;
-    }
-    if (isRecordCancelMessage(message)) {
-      failedStepDeliveries.delete(message.requestId);
-      failedStepPayloads.delete(message.requestId);
-      pendingStopFlushes.delete(message.requestId);
-      knownRecordRequests.delete(message.requestId);
-    }
-    finishStop();
-    return false;
-  }
-
-  return false;
-}
-
-export { RECORD_CANCEL, RECORD_START, RECORD_STEP, RECORD_STOP };

apps/extension/src/content/record-step-delivery.ts

@@ -0,0 +1,65 @@
+import {
+  isAcceptedRecordStepAck,
+  RECORD_STEP,
+  type RecordStepMessage,
+  type RecordStepPayload,
+} from "@/lib/record-bridge";
+
+export type SendRecordStepMessage = (message: RecordStepMessage) => Promise<unknown>;
+
+export class RecordStepDelivery {
+  readonly #requestId: string;
+  readonly #producerId: string;
+  readonly #send: SendRecordStepMessage;
+  readonly #pending = new Map<number, RecordStepPayload>();
+  #nextSequence = 1;
+  #tail = Promise.resolve();
+
+  constructor(
+    requestId: string,
+    send: SendRecordStepMessage = (message) => chrome.runtime.sendMessage(message),
+    producerId: string = crypto.randomUUID(),
+  ) {
+    this.#requestId = requestId;
+    this.#producerId = producerId;
+    this.#send = send;
+  }
+
+  enqueue(step: RecordStepPayload): void {
+    const sequence = this.#nextSequence;
+    this.#nextSequence += 1;
+    this.#pending.set(sequence, step);
+    void this.#schedule();
+  }
+
+  async flush(): Promise<boolean> {
+    await this.#tail;
+    await this.#schedule();
+    return this.#pending.size === 0;
+  }
+
+  #schedule(): Promise<void> {
+    const drain = () => this.#drain();
+    this.#tail = this.#tail.then(drain, drain);
+    return this.#tail;
+  }
+
+  async #drain(): Promise<void> {
+    for (const [sequence, step] of this.#pending) {
+      const message: RecordStepMessage = {
+        type: RECORD_STEP,
+        requestId: this.#requestId,
+        producerId: this.#producerId,
+        sequence,
+        step,
+      };
+      try {
+        const response = await this.#send(message);
+        if (!isAcceptedRecordStepAck(response, sequence)) return;
+        this.#pending.delete(sequence);
+      } catch {
+        return;
+      }
+    }
+  }
+}

apps/extension/src/content/recording/frame-agent.ts

@@ -0,0 +1,209 @@
+import {
+  markRecordingDocument,
+  type RecordingDocumentMarker,
+  waitForDocumentElement,
+} from "@/lib/recording/document-marker";
+import {
+  isRecordFrameStartMessage,
+  RECORD_FRAME_PORT,
+  RECORD_FRAME_QUERY,
+  RECORD_FRAME_START,
+  type RecordFramePortMessage,
+  type RecordFrameQueryResponse,
+  type RecordFrameStartMessage,
+} from "@/lib/recording/frame-bridge";
+import { type RecordCaptureController, startRecordCapture } from "../record-capture";
+import { RecordStepDelivery } from "../record-step-delivery";
+
+interface ActiveFrameRecording {
+  requestId: string;
+  producerId: string;
+  port: chrome.runtime.Port;
+  delivery: RecordStepDelivery;
+  capture: RecordCaptureController;
+  marker: RecordingDocumentMarker;
+}
+
+const RECORD_FRAME_REGISTER_TIMEOUT_MS = 5_000;
+
+export class RecordFrameAgent {
+  #active: ActiveFrameRecording | null = null;
+  #startPromise: Promise<{ ok: true } | { ok: false; error: string }> | null = null;
+  #disposed = false;
+
+  async start(
+    message: RecordFrameStartMessage,
+  ): Promise<{ ok: true } | { ok: false; error: string }> {
+    if (this.#active?.requestId === message.requestId) return { ok: true };
+    if (this.#disposed) return { ok: false, error: "recording document is disposed" };
+    if (this.#startPromise) {
+      await this.#startPromise;
+      return this.start(message);
+    }
+    if (this.#active) this.cancel(this.#active.requestId);
+    const starting = this.#start(message);
+    this.#startPromise = starting;
+    try {
+      return await starting;
+    } finally {
+      if (this.#startPromise === starting) this.#startPromise = null;
+    }
+  }
+
+  cancel(requestId: string): void {
+    const active = this.#active;
+    if (!active || active.requestId !== requestId) return;
+    active.capture.dispose();
+    active.marker.restore();
+    active.port.disconnect();
+    this.#active = null;
+  }
+
+  dispose(): void {
+    this.#disposed = true;
+    if (this.#active) this.cancel(this.#active.requestId);
+  }
+
+  async #start(
+    message: RecordFrameStartMessage,
+  ): Promise<{ ok: true } | { ok: false; error: string }> {
+    const producerId = crypto.randomUUID();
+    const root = await waitForDocumentElement();
+    if (this.#disposed) return { ok: false, error: "recording document is disposed" };
+    const marker = markRecordingDocument(producerId, root);
+    const port = chrome.runtime.connect({ name: RECORD_FRAME_PORT });
+    try {
+      await registerPort(port, message.requestId, producerId);
+    } catch {
+      marker.restore();
+      port.disconnect();
+      return { ok: false, error: "failed to register recording document" };
+    }
+    if (this.#disposed) {
+      marker.restore();
+      port.disconnect();
+      return { ok: false, error: "recording document is disposed" };
+    }
+
+    const delivery = new RecordStepDelivery(message.requestId, undefined, producerId);
+    const capture = startRecordCapture(
+      message.requestId,
+      (step) => {
+        marker.ensure();
+        delivery.enqueue(step);
+      },
+      { captureNavigation: window.top === window },
+    );
+    const active: ActiveFrameRecording = {
+      requestId: message.requestId,
+      producerId,
+      port,
+      delivery,
+      capture,
+      marker,
+    };
+    this.#active = active;
+    port.onMessage.addListener((raw: unknown) => {
+      const command = raw as Partial<RecordFramePortMessage>;
+      if (command.type === "stop" && command.requestId === active.requestId) {
+        void this.#stop(active, command.commandId);
+      } else if (command.type === "cancel" && command.requestId === active.requestId) {
+        this.cancel(active.requestId);
+      }
+    });
+    port.onDisconnect.addListener(() => {
+      if (this.#active !== active) return;
+      active.capture.dispose();
+      active.marker.restore();
+      this.#active = null;
+    });
+    return { ok: true };
+  }
+
+  async #stop(active: ActiveFrameRecording, commandId: string | undefined): Promise<void> {
+    if (!commandId || this.#active !== active) return;
+    active.capture.dispose();
+    const succeeded = await active.delivery.flush();
+    try {
+      active.port.postMessage({
+        type: "stopped",
+        requestId: active.requestId,
+        commandId,
+        ok: succeeded,
+        ...(succeeded ? {} : { error: "failed to deliver one or more recorded steps" }),
+      } satisfies RecordFramePortMessage);
+    } catch {
+      return;
+    }
+    if (!succeeded || this.#active !== active) return;
+    active.marker.restore();
+    this.#active = null;
+    active.port.disconnect();
+  }
+}
+
+function registerPort(
+  port: chrome.runtime.Port,
+  requestId: string,
+  producerId: string,
+): Promise<void> {
+  return new Promise((resolve, reject) => {
+    const timer = setTimeout(() => {
+      cleanup();
+      reject(new Error("recording document registration timed out"));
+    }, RECORD_FRAME_REGISTER_TIMEOUT_MS);
+    const cleanup = () => {
+      clearTimeout(timer);
+      port.onMessage.removeListener(onMessage);
+      port.onDisconnect.removeListener(onDisconnect);
+    };
+    const onMessage = (raw: unknown) => {
+      const message = raw as Partial<RecordFramePortMessage>;
+      if (
+        message.type !== "ready_ack" ||
+        message.requestId !== requestId ||
+        message.producerId !== producerId
+      ) {
+        return;
+      }
+      cleanup();
+      resolve();
+    };
+    const onDisconnect = () => {
+      cleanup();
+      reject(new Error("recording document port disconnected"));
+    };
+    port.onMessage.addListener(onMessage);
+    port.onDisconnect.addListener(onDisconnect);
+    port.postMessage({ type: "ready", requestId, producerId } satisfies RecordFramePortMessage);
+  });
+}
+
+export function attachRecordFrameAgent(): () => void {
+  const agent = new RecordFrameAgent();
+  const onMessage = (
+    message: unknown,
+    _sender: chrome.runtime.MessageSender,
+    sendResponse: (response: { ok: true } | { ok: false; error: string }) => void,
+  ) => {
+    if (!isRecordFrameStartMessage(message)) return false;
+    void agent.start(message).then(sendResponse);
+    return true;
+  };
+  chrome.runtime.onMessage.addListener(onMessage);
+  void chrome.runtime
+    .sendMessage({ type: RECORD_FRAME_QUERY })
+    .then((response: RecordFrameQueryResponse | undefined) => {
+      if (!response?.active || !response.requestId || response.startedAtMs === undefined) return;
+      return agent.start({
+        type: RECORD_FRAME_START,
+        requestId: response.requestId,
+        startedAtMs: response.startedAtMs,
+      });
+    })
+    .catch(() => {});
+  return () => {
+    chrome.runtime.onMessage.removeListener(onMessage);
+    agent.dispose();
+  };
+}

apps/extension/src/entrypoints/background.ts

@@ -21,6 +21,7 @@ import {
   type OverlayMode,
 } from "@/lib/overlay-bridge";
 import { POPUP_PORT_NAME, type PopupInbound, type PopupOutbound } from "@/lib/popup-bridge";
+import { recordFrameCoordinator } from "@/lib/recording/frame-coordinator";
 import { attachSessionsLiveFlag } from "@/lib/sessions-live-flag";
 import { createDisconnectCleanup } from "@/session-manager/disconnect-cleanup";
 import { attachSessionEventHandler } from "@/session-manager/event-handler";
@@ -38,6 +39,7 @@ import {
   attachRecordFinishListener,
   attachRecordQueryListener,
   attachRecordStepListener,
+  type RecordRuntimeDeps,
 } from "@/tools/record";
 import { detectBrowserMeta } from "@/transport/handshake";
 import type { Transport } from "@/transport/transport";
@@ -145,10 +147,33 @@ export default defineBackground(() => {
       void sessionsLive.syncFromManager();
     },
   });
+  const recordDeps = {
+    tabsApi: chrome.tabs,
+    cdp,
+    frameCoordinator: recordFrameCoordinator,
+    sendToTab: (tabId: number, msg: Parameters<typeof chrome.tabs.sendMessage>[1]) =>
+      chrome.tabs.sendMessage(tabId, msg),
+    bypassOverlay: async (tabId: number, enabled: boolean) => {
+      try {
+        await chrome.tabs.sendMessage(tabId, {
+          type: OVERLAY_AUTOMATION_BYPASS,
+          enabled,
+        });
+      } catch {
+        // Content script may be unavailable on restricted pages.
+      }
+    },
+  } satisfies RecordRuntimeDeps;
+  recordFrameCoordinator.attach();
+  attachRecordStepListener(recordDeps);
+  attachRecordFinishListener(recordDeps);
+  attachRecordQueryListener(recordDeps);
+
   const dispatcher = new ToolDispatcher({
     transport,
     sessions,
     cdp,
+    recording: recordDeps,
     onSessionsChanged: onOverlaySessionStateChanged,
     onBrowserControlResumed,
     approveBorrow: (ctx) =>
@@ -170,28 +195,6 @@ export default defineBackground(() => {
     }),
   });
   dispatcher.start();
-  const recordDeps = {
-    tabsApi: chrome.tabs,
-    cdp,
-    sendToTab: (tabId: number, msg: Parameters<typeof chrome.tabs.sendMessage>[1]) =>
-      chrome.tabs.sendMessage(tabId, msg),
-    bypassOverlay: async (tabId: number, enabled: boolean) => {
-      try {
-        await chrome.tabs.sendMessage(tabId, {
-          type: OVERLAY_AUTOMATION_BYPASS,
-          enabled,
-        });
-      } catch {
-        // Content script may be unavailable on restricted pages.
-      }
-    },
-  };
-  // Message listeners stay up (cheap; fire only for record message types).
-  // Tab / webNavigation observation attaches lazily while a recording is
-  // active — see ensureBrowserObservationListeners in tools/record.ts.
-  attachRecordStepListener(recordDeps);
-  attachRecordFinishListener(recordDeps);
-  attachRecordQueryListener(recordDeps);
   if (typeof chrome.notifications?.onClicked?.addListener === "function") {
     attachBorrowNotificationClickHandler({
       onClicked: chrome.notifications.onClicked,

apps/extension/src/entrypoints/content.ts

@@ -7,14 +7,10 @@ import { BorrowConfirmationOverlay } from "@/content/BorrowConfirmationOverlay";
 import { ControlOverlay } from "@/content/ControlOverlay";
 import { createCaptureSuppressController } from "@/content/capture-suppress";
 import { HelpRequestOverlay } from "@/content/HelpRequestOverlay";
+import { createHelpRequestData } from "@/content/help-request";
 import overlayCss from "@/content/overlay.css?inline";
 import { OverlayController, shouldShowAgentControlOverlay } from "@/content/overlay-controller";
 import { RecordOverlay } from "@/content/RecordOverlay";
-import {
-  handleRecordContentMessage,
-  isRecordContentMessage,
-  type RecordCaptureController,
-} from "@/content/record-capture";
 import {
   type CaptureSuppressAck,
   type CaptureSuppressMessage,
@@ -44,11 +40,17 @@ import {
 } from "@/lib/overlay-bridge";
 import { sendInterrupt } from "@/lib/overlay-interrupt-client";
 import {
+  isRecordCancelMessage,
+  isRecordStartMessage,
+  isRecordStopMessage,
   RECORD_FINISH,
   RECORD_QUERY,
+  type RecordCancelMessage,
   type RecordQueryResponse,
   type RecordStartAck,
+  type RecordStartMessage,
   type RecordStopAck,
+  type RecordStopMessage,
 } from "@/lib/record-bridge";
 import type {
   BorrowCancelMessage,
@@ -68,7 +70,6 @@ export default defineContentScript({
     if (window.top !== window) return;
 
     const overlays = new OverlayController();
-    let recordCapture: RecordCaptureController | null = null;
     let activeRecordRequestId: string | null = null;
     let reactRoot: ReactDOM.Root | null = null;
     let overlayHost: HTMLElement | null = null;
@@ -227,8 +228,6 @@ export default defineContentScript({
       if (previousHelp) {
         void sendHelpFinish(previousHelp.id, "cancelled");
       }
-      recordCapture?.dispose();
-      recordCapture = null;
       activeRecordRequestId = null;
       renderAll();
     }
@@ -258,6 +257,9 @@ export default defineContentScript({
         | HelpRequestMessage
         | HelpCancelMessage
         | CaptureSuppressMessage
+        | RecordStartMessage
+        | RecordStopMessage
+        | RecordCancelMessage
         | OverlayAgentOverlayResetMessage
         | OverlayAgentStateMessage
         | OverlayAutomationBypassMessage,
@@ -268,41 +270,32 @@ export default defineContentScript({
         return captureSuppress.handleMessage(message, sendResponse);
       }
 
-      if (isRecordContentMessage(message)) {
-        const needsAsync = handleRecordContentMessage(
-          message,
-          {
-            activeRequestId: activeRecordRequestId,
-            capture: recordCapture,
-            setActiveRequestId: (id) => {
-              activeRecordRequestId = id;
-            },
-            setCapture: (capture) => {
-              recordCapture = capture;
-            },
-            onStart: (requestId, startedAtMs) => {
-              overlays.setAgentRecordRequest({
-                id: requestId,
-                ...(typeof startedAtMs === "number" ? { startedAtMs } : {}),
-                onFinish: () => {
-                  void chrome.runtime.sendMessage({
-                    type: RECORD_FINISH,
-                    requestId,
-                  });
-                },
-              });
-              renderAll();
-            },
-            onStop: () => {
-              overlays.clearAgentRecordRequest(activeRecordRequestId ?? undefined);
-              renderAll();
-            },
+      if (isRecordStartMessage(message)) {
+        activeRecordRequestId = message.requestId;
+        overlays.setAgentRecordRequest({
+          id: message.requestId,
+          ...(typeof message.startedAtMs === "number" ? { startedAtMs: message.startedAtMs } : {}),
+          onFinish: () => {
+            void chrome.runtime.sendMessage({
+              type: RECORD_FINISH,
+              requestId: message.requestId,
+            });
           },
-          sendResponse as unknown as
-            | ((response: RecordStartAck | RecordStopAck) => void)
-            | undefined,
-        );
-        return needsAsync;
+        });
+        renderAll();
+        (sendResponse as unknown as (response: RecordStartAck) => void)({ ok: true });
+        return false;
+      }
+
+      if (isRecordStopMessage(message) || isRecordCancelMessage(message)) {
+        if (activeRecordRequestId !== message.requestId) return false;
+        overlays.clearAgentRecordRequest(message.requestId);
+        activeRecordRequestId = null;
+        renderAll();
+        if (isRecordStopMessage(message)) {
+          (sendResponse as unknown as (response: RecordStopAck) => void)({ ok: true });
+        }
+        return false;
       }
 
       if (
@@ -344,16 +337,7 @@ export default defineContentScript({
 
       if (isHelpRequestMessage(message)) {
         const helpMsg = message as HelpRequestMessage;
-        const previousHelp = overlays.setAgentHelpRequest({
-          id: helpMsg.requestId,
-          prompt: helpMsg.prompt,
-          ...(helpMsg.title ? { title: helpMsg.title } : {}),
-          ...(helpMsg.displayMode ? { displayMode: helpMsg.displayMode } : {}),
-          selectors: helpMsg.selectors,
-          onContinue: (note: string) =>
-            void sendHelpFinish(helpMsg.requestId, "continued", note.trim() ? note : undefined),
-          onCancel: () => void sendHelpFinish(helpMsg.requestId, "cancelled"),
-        });
+        const previousHelp = mountHelpRequest(helpMsg);
         if (previousHelp && previousHelp.id !== helpMsg.requestId) {
           void sendHelpFinish(previousHelp.id, "cancelled");
         }
@@ -406,17 +390,18 @@ export default defineContentScript({
       });
     }
 
-    function mountHelpRequest(helpMsg: Omit<HelpRequestMessage, "type">): void {
-      overlays.setAgentHelpRequest({
-        id: helpMsg.requestId,
-        prompt: helpMsg.prompt,
-        ...(helpMsg.title ? { title: helpMsg.title } : {}),
-        ...(helpMsg.displayMode ? { displayMode: helpMsg.displayMode } : {}),
-        selectors: helpMsg.selectors,
-        onContinue: (note: string) =>
-          void sendHelpFinish(helpMsg.requestId, "continued", note.trim() ? note : undefined),
-        onCancel: () => void sendHelpFinish(helpMsg.requestId, "cancelled"),
-      });
+    function mountHelpRequest(helpMsg: Omit<HelpRequestMessage, "type">) {
+      return overlays.setAgentHelpRequest(
+        createHelpRequestData(helpMsg, {
+          finish: (requestId, outcome, note) => {
+            void sendHelpFinish(requestId, outcome, note);
+          },
+          query: () =>
+            chrome.runtime.sendMessage({ type: HELP_QUERY }) as Promise<
+              HelpQueryResponse | undefined
+            >,
+        }),
+      );
     }
 
     async function queryActiveHelpWithRetry(): Promise<void> {
@@ -533,9 +518,6 @@ export default defineContentScript({
       chrome.runtime.onMessage.removeListener(onMessage);
       chrome.storage.onChanged.removeListener(onStorageChange);
       window.removeEventListener("pageshow", onPageShow);
-      // Restore history hooks / remove capture listeners before the CS unloads.
-      recordCapture?.dispose();
-      recordCapture = null;
       activeRecordRequestId = null;
     });
   },

apps/extension/src/entrypoints/popup/App.test.tsx

@@ -100,24 +100,6 @@ describe("App", () => {
     expect(screen.queryByText("录制你的操作，供 Agent 参考")).toBeNull();
   });
 
-  it("renders localized copy for a version failure instead of protocol details", () => {
-    mockUseConnectionState.mockReturnValue({
-      snapshot: {
-        ...baseSnapshot,
-        lastError:
-          "[handshake] daemon rejected handshake: version_too_old — peer protocol 1.0 is below local min_compatible_protocol 1.0",
-      },
-      statusState: "disconnected",
-      setLabel,
-      setConnectionEnabled,
-    });
-
-    render(<App />);
-
-    const error = screen.getByText(/bsk 与 BrowserSkill 扩展版本不兼容/);
-    expect(error.textContent).not.toContain("min_compatible_protocol");
-  });
-
   it("shows single-line compact metadata and copies the instance id", async () => {
     mockUseConnectionState.mockReturnValue({
       snapshot: {
@@ -220,11 +202,7 @@ describe("App", () => {
     expect(copied).not.toMatch(/bsk record start[^\n]*--url/);
     expect(copied).toContain("./trace");
     expect(copied).toContain("trace.json");
-    expect(copied).toContain("pages/");
-    expect(copied).toContain("总结我做了什么");
-    expect(copied).not.toContain("不要重跑");
-    expect(copied).not.toContain("pages + steps");
-    expect(copied).not.toContain("@eN");
+    expect(copied).toContain("states/");
     // Button label stays static; a transient toast confirms the copy.
     expect(copyButton.textContent).toContain("复制录制指令");
     await waitFor(() => expect(screen.getByRole("status").textContent).toContain("已复制"));

apps/extension/src/entrypoints/popup/App.tsx

@@ -83,9 +83,6 @@ export function App() {
   const daemonProtocol = snapshot.handshake?.protocol_version ?? "—";
   const extensionVersion = snapshot.extensionVersion || "—";
   const instanceId = snapshot.instanceId || "—";
-  const errorMessage = snapshot.lastError?.includes("version_too_old")
-    ? t("popup.versionIncompatibleError")
-    : snapshot.lastError;
 
   const copyInstanceId = async () => {
     if (!snapshot.instanceId || !navigator.clipboard?.writeText) return;
@@ -255,12 +252,12 @@ export function App() {
             </div>
           </section>
 
-          {errorMessage && (
+          {snapshot.lastError && (
             <div
               className="rounded-lg border border-destructive/25 bg-destructive/10 px-3 py-2 text-xs leading-snug text-destructive"
               data-slot="popup-error"
             >
-              {errorMessage}
+              {snapshot.lastError}
             </div>
           )}
 

apps/extension/src/entrypoints/record-frame.content.ts

@@ -0,0 +1,14 @@
+import { attachRecordFrameAgent } from "@/content/recording/frame-agent";
+
+export default defineContentScript({
+  matches: ["<all_urls>"],
+  runAt: "document_start",
+  allFrames: true,
+  matchAboutBlank: true,
+  matchOriginAsFallback: true,
+
+  main(ctx) {
+    const dispose = attachRecordFrameAgent();
+    ctx.onInvalidated(dispose);
+  },
+});

apps/extension/src/lib/__tests__/connection-controller.test.ts

@@ -26,20 +26,20 @@ function handshake(
 }
 
 describe("computeConnectedState (protocol-based compat)", () => {
-  it("returns connected when protocol strings match", () => {
-    expect(computeConnectedState(handshake("1.1", "1.1"), MIN_COMPATIBLE_PROTOCOL)).toEqual({
+  it("returns connected when daemon protocol equals extension protocol", () => {
+    expect(computeConnectedState(handshake("1.1", "1.0"), MIN_COMPATIBLE_PROTOCOL)).toEqual({
       kind: "connected",
     });
   });
 
   it("returns version_skew when daemon protocol minor is newer", () => {
-    expect(computeConnectedState(handshake("1.2", "1.1"))).toEqual({
+    expect(computeConnectedState(handshake("1.2", "1.0"))).toEqual({
       kind: "version_skew",
     });
   });
 
   it("returns version_skew when daemon protocol string differs but floor is satisfied", () => {
-    expect(computeConnectedState(handshake("1.1.0", "1.1"))).toEqual({
+    expect(computeConnectedState(handshake("1.1.0", "1.0"))).toEqual({
       kind: "version_skew",
     });
   });
@@ -48,9 +48,7 @@ describe("computeConnectedState (protocol-based compat)", () => {
     const result = computeConnectedState(handshake("2.0", "1.0"));
     expect(result.kind).toBe("rejected");
     if (result.kind === "rejected") {
-      expect(result.reason).toContain("bsk CLI");
-      expect(result.reason).toContain("BrowserSkill extension");
-      expect(result.reason).not.toContain("protocol");
+      expect(result.reason).toContain("protocol-major mismatch");
     }
   });
 
@@ -58,9 +56,8 @@ describe("computeConnectedState (protocol-based compat)", () => {
     const result = computeConnectedState(handshake("1.1", "1.5"));
     expect(result.kind).toBe("rejected");
     if (result.kind === "rejected") {
-      expect(result.reason).toContain("BrowserSkill extension is out of date");
-      expect(result.reason).toContain("Update and reload the extension");
-      expect(result.reason).not.toContain("min_compatible_protocol");
+      expect(result.reason).toContain("min_compatible_protocol");
+      expect(result.reason).toContain("1.5");
     }
   });
 
@@ -78,24 +75,16 @@ describe("computeConnectedState (protocol-based compat)", () => {
     const result = computeConnectedState(handshake("1.0", "1.0"), "1.1");
     expect(result.kind).toBe("rejected");
     if (result.kind === "rejected") {
-      expect(result.reason).toContain("bsk CLI is out of date");
-      expect(result.reason).toContain("bsk daemon restart");
-      expect(result.reason).not.toContain("min_compatible_protocol");
+      expect(result.reason).toContain("below extension min_compatible_protocol");
     }
   });
 
-  it("returns version_skew when daemon protocol is 1.0 and floor is satisfied", () => {
-    expect(computeConnectedState(handshake("1.0", "1.0"))).toEqual({
-      kind: "version_skew",
-    });
-  });
-
   it("rejects malformed daemon min_compatible_protocol with a daemon-floor reason", () => {
     const result = computeConnectedState(handshake("1.1", "not-a-protocol"));
     expect(result.kind).toBe("rejected");
     if (result.kind === "rejected") {
-      expect(result.reason).toContain("could not verify version compatibility");
-      expect(result.reason).not.toContain("min_compatible_protocol");
+      expect(result.reason).toContain("daemon min_compatible_protocol");
+      expect(result.reason).toContain("not-a-protocol");
     }
   });
 
@@ -253,11 +242,11 @@ describe("ConnectionController connectionEnabled", () => {
     const second = transport.send.mock.calls[1]?.[0] as { id: string };
     expect(second.id).not.toBe(first.id);
 
-    transport.emitMessage({ id: first.id, result: handshake("1.1", "1.1") });
+    transport.emitMessage({ id: first.id, result: handshake("1.1", "1.0") });
     await Promise.resolve();
     expect(controller.snapshot().state).not.toBe("connected");
 
-    transport.emitMessage({ id: second.id, result: handshake("1.1", "1.1") });
+    transport.emitMessage({ id: second.id, result: handshake("1.1", "1.0") });
     await vi.waitFor(() => expect(controller.snapshot().state).toBe("connected"));
   });
 });

apps/extension/src/lib/__tests__/document-settle.test.ts

@@ -0,0 +1,101 @@
+import { afterEach, describe, expect, it, vi } from "vitest";
+import type { CdpRunner } from "@/tools/shared";
+import { waitForDocumentSettled } from "../recording/document-settle";
+
+function quietCdp(): { cdp: CdpRunner; send: ReturnType<typeof vi.fn> } {
+  const send = vi.fn(async () => ({
+    result: { value: { idleMs: 1_000, readyState: "complete" } },
+  }));
+  return {
+    cdp: { send: send as unknown as CdpRunner["send"] },
+    send,
+  };
+}
+
+describe("waitForDocumentSettled", () => {
+  afterEach(() => {
+    vi.useRealTimers();
+  });
+
+  it("waits through the minimum observation floor before accepting a quiet page", async () => {
+    vi.useFakeTimers();
+    const { cdp, send } = quietCdp();
+
+    const settled = waitForDocumentSettled(cdp, { target: { tabId: 7 } });
+    await vi.advanceTimersByTimeAsync(180);
+
+    await expect(settled).resolves.toBe("quiet");
+    expect(send).toHaveBeenCalledTimes(3);
+  });
+
+  it("cancels before probing a superseded observation", async () => {
+    vi.useFakeTimers();
+    const { cdp, send } = quietCdp();
+
+    const controller = new AbortController();
+    const settled = waitForDocumentSettled(
+      cdp,
+      { target: { tabId: 7 } },
+      {
+        signal: controller.signal,
+      },
+    );
+    controller.abort();
+    await vi.advanceTimersByTimeAsync(60);
+
+    await expect(settled).resolves.toBe("cancelled");
+    expect(send).not.toHaveBeenCalled();
+  });
+
+  it("probes a same-process iframe through its isolated execution context", async () => {
+    vi.useFakeTimers();
+    const send = vi.fn(async (_tabId: number, method: string) => {
+      if (method === "Page.createIsolatedWorld") return { executionContextId: 91 };
+      return { result: { value: { idleMs: 1_000, readyState: "complete" } } };
+    });
+    const cdp: CdpRunner = { send: send as unknown as CdpRunner["send"] };
+
+    const settled = waitForDocumentSettled(cdp, {
+      target: { tabId: 7 },
+      frameId: "child-frame",
+    });
+    await vi.advanceTimersByTimeAsync(180);
+
+    await expect(settled).resolves.toBe("quiet");
+    expect(send).toHaveBeenCalledWith(
+      7,
+      "Page.createIsolatedWorld",
+      expect.objectContaining({ frameId: "child-frame" }),
+    );
+    expect(send).toHaveBeenCalledWith(
+      7,
+      "Runtime.evaluate",
+      expect.objectContaining({ contextId: 91 }),
+    );
+  });
+
+  it("routes an OOPIF probe through its CDP target session", async () => {
+    vi.useFakeTimers();
+    const send = vi.fn();
+    const sendToTarget = vi.fn(async (_target, method: string) => {
+      if (method === "Page.createIsolatedWorld") return { executionContextId: 27 };
+      return { result: { value: { idleMs: 1_000, readyState: "complete" } } };
+    });
+    const cdp: CdpRunner = {
+      send: send as unknown as CdpRunner["send"],
+      sendToTarget: sendToTarget as unknown as NonNullable<CdpRunner["sendToTarget"]>,
+    };
+
+    const target = { tabId: 7, sessionId: "oopif-session" };
+    const settled = waitForDocumentSettled(cdp, { target, frameId: "oopif-frame" });
+    await vi.advanceTimersByTimeAsync(180);
+
+    await expect(settled).resolves.toBe("quiet");
+    expect(send).not.toHaveBeenCalled();
+    expect(sendToTarget).toHaveBeenCalledWith(
+      target,
+      "Runtime.evaluate",
+      expect.objectContaining({ contextId: 27 }),
+    );
+  });
+});

apps/extension/src/lib/__tests__/help-bridge.test.ts

@@ -32,6 +32,7 @@ describe("help-bridge", () => {
         prompt: "log in",
         displayMode: "compact",
         selectors: ["#login"],
+        rects: [{ top: 10, left: 20, width: 100, height: 40 }],
         timeoutMs: 1000,
       }),
     ).toBe(true);

apps/extension/src/lib/__tests__/match-target.test.ts

@@ -1,116 +0,0 @@
-import type { RenderedRef } from "@browser-skill/vom";
-import { describe, expect, it } from "vitest";
-import type { CapturedNode } from "@/tools/vom/capture";
-import { matchTarget } from "../match-target";
-
-function node(backendNodeId: number, overrides: Partial<CapturedNode> = {}): CapturedNode {
-  return {
-    backendNodeId,
-    parentBackendNodeId: null,
-    tag: "button",
-    attrs: {},
-    rect: { x: 10, y: 20, w: 100, h: 30 },
-    paintOrder: 1,
-    position: "static",
-    pointerEvents: "auto",
-    ...overrides,
-  };
-}
-
-describe("matchTarget", () => {
-  const refs: RenderedRef[] = [
-    { ref: "e1", backendNodeId: 42, role: "button", name: "发布", line: 3 },
-  ];
-
-  it("matches a unique node by viewport coordinates", () => {
-    const target = matchTarget({
-      geometry: {
-        rect: { x: 10, y: 20, w: 100, h: 30 },
-        scrollX: 0,
-        scrollY: 0,
-        position: "static",
-        tag: "button",
-      },
-      captured: [node(42)],
-      refs,
-      fallback: { tag: "button", name: "发布" },
-    });
-    expect(target).toEqual({ ref: "e1", role: "button", name: "发布" });
-  });
-
-  it("matches a non-fixed node after the top frame scrolls", () => {
-    const target = matchTarget({
-      geometry: {
-        rect: { x: 10, y: 20, w: 100, h: 30 },
-        scrollX: 40,
-        scrollY: 300,
-        position: "static",
-        tag: "button",
-      },
-      captured: [
-        node(42, {
-          documentRect: { x: 50, y: 320, w: 100, h: 30 },
-          localRect: { x: 10, y: 20, w: 100, h: 30 },
-          rect: { x: 10, y: 20, w: 100, h: 30 },
-        }),
-      ],
-      refs,
-      fallback: { tag: "button", name: "发布" },
-    });
-
-    expect(target).toEqual({ ref: "e1", role: "button", name: "发布" });
-  });
-
-  it("matches a static node by document coordinates when scroll changed after observation", () => {
-    const target = matchTarget({
-      geometry: {
-        rect: { x: 10, y: 50, w: 100, h: 30 },
-        scrollX: 0,
-        scrollY: 250,
-        position: "static",
-        tag: "button",
-      },
-      captured: [
-        node(42, {
-          documentRect: { x: 10, y: 300, w: 100, h: 30 },
-          localRect: { x: 10, y: 200, w: 100, h: 30 },
-          rect: { x: 10, y: 200, w: 100, h: 30 },
-        }),
-      ],
-      refs,
-      fallback: { tag: "button", name: "发布" },
-    });
-
-    expect(target).toEqual({ ref: "e1", role: "button", name: "发布" });
-  });
-
-  it("returns unmatched when multiple candidates match", () => {
-    const target = matchTarget({
-      geometry: {
-        rect: { x: 10, y: 20, w: 100, h: 30 },
-        scrollX: 0,
-        scrollY: 0,
-        position: "static",
-        tag: "button",
-      },
-      captured: [node(42), node(43)],
-      refs,
-    });
-    expect(target.unmatched).toBe(true);
-  });
-
-  it("uses viewport coordinates for fixed elements", () => {
-    const target = matchTarget({
-      geometry: {
-        rect: { x: 5, y: 5, w: 80, h: 24 },
-        scrollX: 100,
-        scrollY: 200,
-        position: "fixed",
-        tag: "button",
-      },
-      captured: [node(42, { position: "fixed", rect: { x: 5, y: 5, w: 80, h: 24 } })],
-      refs,
-    });
-    expect(target.ref).toBe("e1");
-  });
-});

apps/extension/src/lib/__tests__/record-bridge.test.ts

@@ -0,0 +1,32 @@
+import { describe, expect, it } from "vitest";
+import {
+  isAcceptedRecordStepAck,
+  isRecordStepMessage,
+  RECORD_STEP,
+  type RecordStepMessage,
+} from "../record-bridge";
+
+const message: RecordStepMessage = {
+  type: RECORD_STEP,
+  requestId: "rec-1",
+  producerId: "document-1",
+  sequence: 1,
+  step: { op: "click", target: { tag: "button", name: "Submit" } },
+};
+
+describe("record step delivery protocol", () => {
+  it("requires a producer and a positive integer sequence", () => {
+    expect(isRecordStepMessage(message)).toBe(true);
+    expect(isRecordStepMessage({ ...message, producerId: "" })).toBe(false);
+    expect(isRecordStepMessage({ ...message, sequence: 0 })).toBe(false);
+    expect(isRecordStepMessage({ ...message, sequence: 1.5 })).toBe(false);
+  });
+
+  it("accepts only the acknowledgement for the sent sequence", () => {
+    expect(isAcceptedRecordStepAck({ ok: true, sequence: 2 }, 2)).toBe(true);
+    expect(isAcceptedRecordStepAck({ ok: true, sequence: 1 }, 2)).toBe(false);
+    expect(isAcceptedRecordStepAck({ ok: false, expectedSequence: 2, error: "gap" }, 2)).toBe(
+      false,
+    );
+  });
+});

apps/extension/src/lib/__tests__/record-frame-coordinator.test.ts

@@ -0,0 +1,213 @@
+import { beforeEach, describe, expect, it, vi } from "vitest";
+import { RECORD_FRAME_PORT, RECORD_FRAME_QUERY } from "../recording/frame-bridge";
+import { RecordFrameCoordinator } from "../recording/frame-coordinator";
+
+class ListenerSet<T extends (...args: never[]) => unknown> {
+  readonly listeners = new Set<T>();
+  addListener = (listener: T) => this.listeners.add(listener);
+  removeListener = (listener: T) => this.listeners.delete(listener);
+}
+
+function fakePort(sender: chrome.runtime.MessageSender) {
+  const messages: unknown[] = [];
+  const onMessage = new ListenerSet<(message: unknown) => void>();
+  const onDisconnect = new ListenerSet<() => void>();
+  return {
+    port: {
+      name: RECORD_FRAME_PORT,
+      sender,
+      onMessage,
+      onDisconnect,
+      postMessage: vi.fn((message: unknown) => messages.push(message)),
+      disconnect: vi.fn(),
+    } as unknown as chrome.runtime.Port,
+    messages,
+    receive(message: unknown) {
+      for (const listener of onMessage.listeners) listener(message);
+    },
+  };
+}
+
+describe("RecordFrameCoordinator", () => {
+  const onConnect = new ListenerSet<(port: chrome.runtime.Port) => void>();
+  const onMessage = new ListenerSet<
+    (
+      message: unknown,
+      sender: chrome.runtime.MessageSender,
+      sendResponse: (response: unknown) => void,
+    ) => boolean
+  >();
+
+  beforeEach(() => {
+    onConnect.listeners.clear();
+    onMessage.listeners.clear();
+    vi.stubGlobal("chrome", {
+      runtime: { onConnect, onMessage },
+    });
+  });
+
+  it("answers child-document queries for the initial tab before it is armed", () => {
+    const coordinator = new RecordFrameCoordinator({
+      getAllFrames: async () => [],
+      sendToDocument: async () => ({ ok: true }),
+    });
+    coordinator.attach();
+    coordinator.begin("rec-1", 10, 3);
+
+    const query = (tabId: number) => {
+      let response: unknown;
+      for (const listener of onMessage.listeners) {
+        listener(
+          { type: RECORD_FRAME_QUERY },
+          {
+            tab: { id: tabId },
+            frameId: 7,
+            documentId: "child-document",
+          } as chrome.runtime.MessageSender,
+          (value) => {
+            response = value;
+          },
+        );
+      }
+      return response;
+    };
+
+    expect(query(3)).toEqual({ active: true, requestId: "rec-1", startedAtMs: 10 });
+    expect(query(4)).toEqual({ active: false });
+  });
+
+  it("starts a child Document that appears after the initial frame snapshot", async () => {
+    let notifyFrameNavigation:
+      | ((frame: {
+          tabId: number;
+          frameId: number;
+          documentId?: string;
+          lifecycle: "committed" | "completed";
+        }) => void)
+      | undefined;
+    const sendToDocument = vi.fn(async () => ({ ok: true }));
+    const coordinator = new RecordFrameCoordinator({
+      getAllFrames: async () => [{ frameId: 0, documentId: "top-document" }],
+      sendToDocument,
+      subscribeFrameNavigation(listener) {
+        notifyFrameNavigation = listener;
+        return () => {};
+      },
+    });
+    coordinator.attach();
+    const onDocumentReady = vi.fn();
+    coordinator.begin("rec-1", 10, 3, onDocumentReady);
+    await coordinator.armTab("rec-1", 3);
+    sendToDocument.mockClear();
+
+    notifyFrameNavigation?.({
+      tabId: 3,
+      frameId: 7,
+      documentId: "late-child",
+      lifecycle: "completed",
+    });
+    await vi.waitFor(() => {
+      expect(sendToDocument).toHaveBeenCalledWith(
+        3,
+        { type: "bsk-record-frame-start", requestId: "rec-1", startedAtMs: 10 },
+        { documentId: "late-child" },
+      );
+    });
+    expect(onDocumentReady).not.toHaveBeenCalled();
+
+    const child = fakePort({
+      tab: { id: 3 },
+      frameId: 7,
+      documentId: "late-child",
+    } as chrome.runtime.MessageSender);
+    for (const listener of onConnect.listeners) listener(child.port);
+    child.receive({ type: "ready", requestId: "rec-1", producerId: "producer-1" });
+    notifyFrameNavigation?.({
+      tabId: 3,
+      frameId: 7,
+      documentId: "late-child",
+      lifecycle: "completed",
+    });
+    await vi.waitFor(() => {
+      expect(onDocumentReady).toHaveBeenCalledWith({
+        tabId: 3,
+        documentId: "late-child",
+        browserFrameId: 7,
+        producerId: "producer-1",
+      });
+    });
+  });
+
+  it("binds a producer to its sender Document and keeps final steps valid while stopping", async () => {
+    const sendToDocument = vi.fn(async () => ({ ok: true }));
+    const coordinator = new RecordFrameCoordinator({
+      getAllFrames: async () => [
+        { frameId: 0, documentId: "top-document" },
+        { frameId: 7, documentId: "child-document" },
+      ],
+      sendToDocument,
+    });
+    coordinator.attach();
+    coordinator.begin("rec-1", 10, 3);
+    await expect(coordinator.armTab("rec-1", 3)).resolves.toBe(true);
+
+    const sender = {
+      tab: { id: 3 },
+      frameId: 7,
+      documentId: "child-document",
+    } as chrome.runtime.MessageSender;
+    const frame = fakePort(sender);
+    for (const listener of onConnect.listeners) listener(frame.port);
+    frame.receive({ type: "ready", requestId: "rec-1", producerId: "producer-1" });
+
+    expect(coordinator.sourceFor("rec-1", "producer-1", sender)).toEqual({
+      tabId: 3,
+      documentId: "child-document",
+      browserFrameId: 7,
+      producerId: "producer-1",
+    });
+
+    const stopping = coordinator.stop("rec-1");
+    expect(coordinator.sourceFor("rec-1", "producer-1", sender)).not.toBeNull();
+    const stop = frame.messages.find(
+      (message): message is { type: "stop"; commandId: string } =>
+        typeof message === "object" &&
+        message !== null &&
+        (message as { type?: string }).type === "stop",
+    );
+    expect(stop).toBeDefined();
+    frame.receive({
+      type: "stopped",
+      requestId: "rec-1",
+      commandId: stop!.commandId,
+      ok: true,
+    });
+    await expect(stopping).resolves.toBe(true);
+    expect(coordinator.sourceFor("rec-1", "producer-1", sender)).toBeNull();
+  });
+
+  it("rejects a producer used from another Document", async () => {
+    const coordinator = new RecordFrameCoordinator({
+      getAllFrames: async () => [{ frameId: 0, documentId: "top-document" }],
+      sendToDocument: async () => ({ ok: true }),
+    });
+    coordinator.attach();
+    coordinator.begin("rec-1", 10, 3);
+    await coordinator.armTab("rec-1", 3);
+    const sender = {
+      tab: { id: 3 },
+      frameId: 0,
+      documentId: "top-document",
+    } as chrome.runtime.MessageSender;
+    const frame = fakePort(sender);
+    for (const listener of onConnect.listeners) listener(frame.port);
+    frame.receive({ type: "ready", requestId: "rec-1", producerId: "producer-1" });
+
+    expect(
+      coordinator.sourceFor("rec-1", "producer-1", {
+        ...sender,
+        documentId: "other-document",
+      }),
+    ).toBeNull();
+  });
+});

apps/extension/src/lib/__tests__/record-observation.test.ts

@@ -1,114 +0,0 @@
-import { beforeEach, describe, expect, it } from "vitest";
-import type { CapturedNode } from "@/tools/vom/capture";
-import type { DraftTraceStep } from "@/transport/types";
-import { resetStateIdCounterForTests } from "../record-constants";
-import {
-  applyTargetMatching,
-  buildTraceV3,
-  createObservationState,
-  rememberStepOnPage,
-} from "../record-observation";
-import { registerObservation } from "../trace-reducer";
-
-const URL = "https://example.com/login";
-
-function capturedInput(): CapturedNode {
-  return {
-    backendNodeId: 42,
-    parentBackendNodeId: null,
-    tag: "input",
-    attrs: {},
-    rect: { x: 20, y: 40, w: 200, h: 30 },
-    localRect: { x: 20, y: 40, w: 200, h: 30 },
-    paintOrder: 1,
-    position: "static",
-    pointerEvents: "auto",
-  };
-}
-
-function finalizedFillBody(value: string, redactValues: boolean): string {
-  const obs = createObservationState({ redactValues });
-  const rawVomText = '@vom 1\ntextbox "Password" value="••••••" [ref=e1]';
-  const stateId = registerObservation(obs.stateRegistry, { url: URL, rawVomText });
-  obs.lastSettled = {
-    stateId,
-    captured: [capturedInput()],
-    refs: [{ ref: "e1", backendNodeId: 42, role: "textbox", name: "Password", line: 1 }],
-    url: URL,
-    vomText: rawVomText,
-  };
-
-  const draft: DraftTraceStep = {
-    op: "fill",
-    target: { unmatched: true },
-    value,
-    geometry: {
-      rect: { x: 20, y: 40, w: 200, h: 30 },
-      scrollX: 0,
-      scrollY: 0,
-      position: "static",
-      tag: "input",
-    },
-  };
-  applyTargetMatching(obs, draft, 1);
-  draft.postStateId = stateId;
-  rememberStepOnPage(obs, draft, 1);
-
-  return (
-    buildTraceV3({
-      obs,
-      steps: [draft],
-      startedAt: "2026-08-12T00:00:00.000Z",
-      stoppedBy: "user_finish",
-      bskVersion: "test",
-    }).states[0]?.body ?? ""
-  );
-}
-
-describe("record observation annotations", () => {
-  beforeEach(() => {
-    resetStateIdCounterForTests();
-  });
-
-  it("omits a fill literal from finalized state bodies when values are redacted", () => {
-    const secret = "hunter2-private";
-    const body = finalizedFillBody(secret, true);
-
-    expect(body).toContain("step 1: fill");
-    expect(body).not.toContain(secret);
-  });
-
-  it("keeps fill details in ordinary recording annotations", () => {
-    const value = "ordinary text";
-    const body = finalizedFillBody(value, false);
-
-    expect(body).toContain(`step 1: fill: ${JSON.stringify(value)}`);
-  });
-});
-
-describe("applyTargetMatching without a settled observation", () => {
-  it("keeps the capture role/name instead of a bare unmatched target", () => {
-    const obs = createObservationState();
-    const draft: DraftTraceStep = {
-      op: "hover",
-      target: { unmatched: true },
-      captureTarget: { tag: "button", role: "button", name: "新建" },
-      geometry: {
-        rect: { x: 900, y: 8, w: 60, h: 32 },
-        scrollX: 0,
-        scrollY: 0,
-        position: "static",
-        tag: "button",
-      },
-    };
-
-    applyTargetMatching(obs, draft, 1);
-
-    expect(draft.preStateId).toBeUndefined();
-    expect(draft.target).toEqual({
-      role: "button",
-      name: "新建",
-      unmatched: true,
-    });
-  });
-});

apps/extension/src/lib/__tests__/recording-document-marker.test.ts

@@ -0,0 +1,70 @@
+import { describe, expect, it } from "vitest";
+import {
+  RECORD_DOCUMENT_ATTRIBUTE,
+  recordingDocumentMarkerValue,
+} from "@/shared/recording-document-identity";
+import { markRecordingDocument } from "../recording/document-marker";
+import { ObservationNodeIndex } from "../recording/observation-capture";
+
+describe("recording document marker", () => {
+  it("rejects a non-random Document identity", () => {
+    expect(() => markRecordingDocument("user@example.com")).toThrow(
+      "recording Document identity must be a random UUID",
+    );
+  });
+
+  it("restores the Document attribute exactly", () => {
+    document.documentElement.setAttribute(RECORD_DOCUMENT_ATTRIBUTE, "page-value");
+    const producerId = "123e4567-e89b-42d3-a456-426614174000";
+    const marker = markRecordingDocument(producerId);
+
+    expect(document.documentElement.getAttribute(RECORD_DOCUMENT_ATTRIBUTE)).toBe(
+      recordingDocumentMarkerValue(producerId),
+    );
+    marker.restore();
+    expect(document.documentElement.getAttribute(RECORD_DOCUMENT_ATTRIBUTE)).toBe("page-value");
+    document.documentElement.removeAttribute(RECORD_DOCUMENT_ATTRIBUTE);
+  });
+
+  it("indexes the CDP scope from the safe frame identity", () => {
+    const index = new ObservationNodeIndex({
+      rootFrameId: "root",
+      frames: [
+        {
+          frameId: "child",
+          target: { tabId: 3, sessionId: "oopif-session" },
+          recordingDocumentId: "producer-1",
+        },
+      ],
+      matchNodes: [],
+      refs: [],
+    });
+
+    expect(index.documentScope("producer-1")).toEqual({
+      frameId: "child",
+      target: { tabId: 3, sessionId: "oopif-session" },
+    });
+  });
+
+  it("fails closed when a recording identity maps to multiple Documents", () => {
+    const index = new ObservationNodeIndex({
+      rootFrameId: "root",
+      frames: [
+        {
+          frameId: "left",
+          target: { tabId: 3, sessionId: "left-session" },
+          recordingDocumentId: "duplicate",
+        },
+        {
+          frameId: "right",
+          target: { tabId: 3, sessionId: "right-session" },
+          recordingDocumentId: "duplicate",
+        },
+      ],
+      matchNodes: [],
+      refs: [],
+    });
+
+    expect(index.documentScope("duplicate")).toBeUndefined();
+  });
+});

apps/extension/src/lib/__tests__/recording-observation.test.ts

@@ -0,0 +1,146 @@
+import { describe, expect, it } from "vitest";
+import { ObservationNodeIndex } from "../recording/observation-capture";
+import { RecordingObservationSession } from "../recording/observation-session";
+import { RecordingStateRegistry } from "../recording/state-registry";
+import { buildTraceV3 } from "../recording/trace-builder-v3";
+import type { RecordingDraftStep } from "../recording/types";
+
+const URL = "https://example.com/login";
+
+function sessionWithInput(redactValues = false): RecordingObservationSession {
+  const session = new RecordingObservationSession({ redactValues });
+  const state = session.registry.register({
+    url: URL,
+    vomText: '@vom 1\ntextbox "Password" value="••••••" [ref=e1]',
+  });
+  session.cursor.lastSettled = {
+    stateId: state.id,
+    rootFrameId: "root",
+    index: new ObservationNodeIndex({
+      rootFrameId: "root",
+      matchNodes: [
+        {
+          frameId: "root",
+          backendNodeId: 42,
+          tag: "input",
+          rect: { x: 20, y: 40, w: 200, h: 30 },
+          localRect: { x: 20, y: 40, w: 200, h: 30 },
+        },
+      ],
+      refs: [{ ref: "e1", backendNodeId: 42, role: "textbox", name: "Password", line: 1 }],
+    }),
+    url: URL,
+  };
+  return session;
+}
+
+function finalizedFillTrace(value: string, redactValues: boolean) {
+  const session = sessionWithInput(redactValues);
+  const draft: RecordingDraftStep = {
+    op: "fill",
+    captureTarget: { tag: "input", role: "textbox", name: "Password" },
+    value,
+    targetHint: {
+      geometry: { rect: { x: 20, y: 40, w: 200, h: 30 }, tag: "input" },
+    },
+  };
+  session.bindDraft(draft, 1);
+  draft.postStateId = draft.preStateId;
+  return buildTraceV3({
+    registry: session.registry,
+    drafts: [draft],
+    annotations: session.annotations,
+    startedAt: "2026-08-12T00:00:00.000Z",
+    stoppedBy: "user_finish",
+    bskVersion: "test",
+    redactValues,
+  });
+}
+
+function finalizedFillBody(value: string, redactValues: boolean): string {
+  return finalizedFillTrace(value, redactValues).states[0]!.body;
+}
+
+describe("record observation annotations", () => {
+  it("omits a fill literal when values are redacted", () => {
+    const secret = "hunter2-private";
+    const dumped = JSON.stringify(finalizedFillTrace(secret, true));
+    expect(dumped).toContain("step 1: fill");
+    expect(dumped).not.toContain(secret);
+  });
+
+  it("keeps ordinary fill details", () => {
+    expect(finalizedFillBody("ordinary text", false)).toContain('step 1: fill: "ordinary text"');
+  });
+
+  it("encodes title and URL so line breaks cannot corrupt state metadata", () => {
+    const registry = new RecordingStateRegistry();
+    const state = registry.register({
+      url: "https://example.com/a\nb",
+      title: "hello\nworld",
+      vomText: "@vom 1",
+    });
+    const trace = buildTraceV3({
+      registry,
+      drafts: [{ op: "scroll", preStateId: state.id, postStateId: state.id }],
+      startedAt: "2026-08-12T00:00:00.000Z",
+      stoppedBy: "user_finish",
+      bskVersion: "test",
+    });
+    expect(trace.states[0]?.body).toContain('url: "https://example.com/a\\nb"');
+    expect(trace.states[0]?.body).toContain('title: "hello\\nworld"');
+  });
+});
+
+describe("recording state ownership", () => {
+  it("deduplicates within one recording and isolates ids between recordings", () => {
+    const first = new RecordingStateRegistry();
+    const second = new RecordingStateRegistry();
+    expect(first.register({ url: URL, vomText: "same" }).id).toBe("s1");
+    expect(first.register({ url: URL, vomText: "same" }).id).toBe("s1");
+    expect(second.register({ url: URL, vomText: "other" }).id).toBe("s1");
+  });
+
+  it("enriches metadata when a deduplicated observation becomes more complete", () => {
+    const registry = new RecordingStateRegistry();
+    registry.register({ url: URL, vomText: "same" });
+    const state = registry.register({
+      url: URL,
+      title: "Login",
+      vomText: "same",
+      truncated: true,
+    });
+    expect(state).toMatchObject({ id: "s1", title: "Login", truncated: true });
+  });
+});
+
+describe("draft binding", () => {
+  it("keeps capture semantics when no observation exists", () => {
+    const session = new RecordingObservationSession();
+    const draft: RecordingDraftStep = {
+      op: "hover",
+      captureTarget: { tag: "button", role: "button", name: "新建" },
+    };
+    session.bindDraft(draft, 1);
+    expect(draft.matchedTarget).toEqual({ role: "button", name: "新建", unmatched: true });
+    expect(draft.preStateId).toBeUndefined();
+  });
+
+  it("does not bind an unmatched new action to a stale observation", () => {
+    const session = sessionWithInput();
+    const draft: RecordingDraftStep = {
+      op: "click",
+      captureTarget: { tag: "button", role: "button", name: "Confirm" },
+      targetHint: {
+        geometry: { rect: { x: 400, y: 300, w: 80, h: 30 }, tag: "button" },
+      },
+    };
+    session.bindDraft(draft, 2, true);
+    expect(draft.preStateId).toBeUndefined();
+    expect("matchedTarget" in draft ? draft.matchedTarget : undefined).toEqual({
+      role: "button",
+      name: "Confirm",
+      unmatched: true,
+    });
+  });
+});

apps/extension/src/lib/__tests__/recording-runtime.test.ts

@@ -0,0 +1,289 @@
+import { afterEach, describe, expect, it, vi } from "vitest";
+import type { CdpRunner, ChromeTabsApi } from "@/tools/shared";
+
+const captureRecordingObservation = vi.hoisted(() => vi.fn());
+const waitForDocumentSettled = vi.hoisted(() => vi.fn(async () => "quiet" as const));
+
+vi.mock("../recording/observation-capture", async (importOriginal) => {
+  const actual = await importOriginal<typeof import("../recording/observation-capture")>();
+  return { ...actual, captureRecordingObservation };
+});
+
+vi.mock("../recording/document-settle", async (importOriginal) => {
+  const actual = await importOriginal<typeof import("../recording/document-settle")>();
+  return { ...actual, waitForDocumentSettled };
+});
+
+import { ObservationNodeIndex } from "../recording/observation-capture";
+import { RecordingObservationRuntime } from "../recording/recording-runtime";
+import type { RecordingDraftStep } from "../recording/types";
+
+function observation(url = "https://example.com/") {
+  return {
+    rootFrameId: "root",
+    index: new ObservationNodeIndex({ rootFrameId: "root", matchNodes: [], refs: [] }),
+    url,
+    title: "Example",
+    vomText: '@vom 1\nRootWebArea "Example"',
+    truncated: false,
+  };
+}
+
+function runtime(): RecordingObservationRuntime {
+  return new RecordingObservationRuntime({
+    cdp: { send: vi.fn() as unknown as CdpRunner["send"] },
+    tabsApi: { get: vi.fn(), query: vi.fn() } as unknown as ChromeTabsApi,
+  });
+}
+
+describe("RecordingObservationRuntime", () => {
+  afterEach(() => {
+    captureRecordingObservation.mockReset();
+    waitForDocumentSettled.mockClear();
+  });
+
+  it("shares one initial capture and includes it in flush", async () => {
+    let release!: (value: ReturnType<typeof observation>) => void;
+    captureRecordingObservation.mockReturnValueOnce(
+      new Promise((resolve) => {
+        release = resolve;
+      }),
+    );
+    const recording = runtime();
+
+    const first = recording.captureInitial(7);
+    const second = recording.captureInitial(7);
+    const flushed = recording.flush();
+    let flushCompleted = false;
+    void flushed.then(() => {
+      flushCompleted = true;
+    });
+
+    await Promise.resolve();
+    expect(captureRecordingObservation).toHaveBeenCalledTimes(1);
+    expect(flushCompleted).toBe(false);
+
+    release(observation());
+    await Promise.all([first, second, flushed]);
+    expect(flushCompleted).toBe(true);
+  });
+
+  it("cancels an in-flight initial capture", async () => {
+    captureRecordingObservation.mockImplementationOnce(
+      (input: { signal?: AbortSignal }) =>
+        new Promise((_, reject) => {
+          input.signal?.addEventListener(
+            "abort",
+            () => reject(new DOMException("observation aborted", "AbortError")),
+            { once: true },
+          );
+        }),
+    );
+    const recording = runtime();
+
+    const pending = recording.captureInitial(7);
+    recording.cancel();
+
+    await expect(pending).rejects.toMatchObject({ name: "AbortError" });
+  });
+
+  it("retries a failed initial capture when the recording settles at stop", async () => {
+    captureRecordingObservation
+      .mockRejectedValueOnce(new Error("document swapped"))
+      .mockResolvedValueOnce(observation());
+    const recording = runtime();
+
+    await expect(recording.captureInitial(7)).rejects.toThrow("document swapped");
+    await recording.settleTrailing(7, []);
+    const trace = recording.buildTrace({
+      drafts: [],
+      startedAt: "2026-08-19T00:00:00.000Z",
+      stoppedBy: "user_finish",
+      bskVersion: "0.1.6",
+    });
+
+    expect(captureRecordingObservation).toHaveBeenCalledTimes(2);
+    expect(trace.states).toHaveLength(1);
+  });
+
+  it("keeps runtime matching data out of the serialized trace", async () => {
+    const captured = observation();
+    captured.index = new ObservationNodeIndex({
+      rootFrameId: "root",
+      matchNodes: [
+        {
+          frameId: "root",
+          backendNodeId: 42,
+          tag: "button",
+          rect: { x: 10, y: 20, w: 100, h: 30 },
+          localRect: { x: 10, y: 20, w: 100, h: 30 },
+        },
+      ],
+      refs: [{ ref: "e1", backendNodeId: 42, role: "button", name: "Submit", line: 1 }],
+    });
+    captureRecordingObservation.mockResolvedValueOnce(captured);
+    const recording = runtime();
+
+    await recording.captureInitial(7);
+    const dumped = JSON.stringify(
+      recording.buildTrace({
+        drafts: [],
+        startedAt: "2026-08-19T00:00:00.000Z",
+        stoppedBy: "user_finish",
+        bskVersion: "0.1.6",
+      }),
+    );
+
+    expect(dumped).toContain("RootWebArea");
+    expect(dumped).not.toContain("matchNodes");
+    expect(dumped).not.toContain("localRect");
+    expect(dumped).not.toContain("backendNodeId");
+    expect(dumped).not.toContain("ObservationNodeIndex");
+  });
+
+  it("captures a tab transition without sharing observation cursors", async () => {
+    captureRecordingObservation
+      .mockResolvedValueOnce(observation("https://example.com/first"))
+      .mockResolvedValueOnce(observation("https://example.com/second"));
+    const recording = runtime();
+
+    await recording.captureInitial(4);
+    const transition = await recording.captureTabTransition(4, 5);
+    await recording.captureInitial(4);
+    await recording.captureInitial(5);
+
+    expect(transition).toEqual({
+      preStateId: "s1",
+      postStateId: "s2",
+      targetUrl: "https://example.com/second",
+    });
+    expect(captureRecordingObservation.mock.calls.map(([input]) => input.tabId)).toEqual([4, 5]);
+  });
+
+  it("binds an iframe draft to its marked CDP Document scope", async () => {
+    captureRecordingObservation.mockResolvedValueOnce({
+      ...observation(),
+      index: new ObservationNodeIndex({
+        rootFrameId: "root",
+        frames: [
+          {
+            frameId: "child",
+            target: { tabId: 7, sessionId: "oopif-session" },
+            recordingDocumentId: "producer-1",
+          },
+        ],
+        matchNodes: [
+          {
+            backendNodeId: 42,
+            frameId: "child",
+            tag: "button",
+            rect: { x: 410, y: 20, w: 100, h: 30 },
+            localRect: { x: 10, y: 20, w: 100, h: 30 },
+          },
+        ],
+        refs: [{ ref: "e1", backendNodeId: 42, frameId: "child", line: 1 }],
+      }),
+    });
+    const recording = runtime();
+    const drafts: RecordingDraftStep[] = [
+      {
+        op: "click" as const,
+        captureTarget: { tag: "button", role: "button", name: "Save" },
+        targetHint: {
+          geometry: { rect: { x: 10, y: 20, w: 100, h: 30 }, tag: "button" },
+        },
+      },
+    ];
+
+    await recording.captureInitial(7);
+    await recording.processDraft(7, drafts, 0, "producer-1");
+    recording.cancel();
+
+    const draft = drafts[0];
+    expect(draft?.op).toBe("click");
+    if (!draft || draft.op !== "click") throw new Error("expected click draft");
+    expect(draft.targetHint).toMatchObject({
+      frameId: "child",
+      geometrySpace: "local",
+    });
+    expect(draft.matchedTarget?.ref).toBe("e1");
+  });
+
+  it("refreshes a late iframe before matching its first action", async () => {
+    const iframeObservation = {
+      ...observation(),
+      index: new ObservationNodeIndex({
+        rootFrameId: "root",
+        frames: [
+          {
+            frameId: "child",
+            target: { tabId: 7, sessionId: "oopif-session" },
+            recordingDocumentId: "producer-1",
+          },
+        ],
+        matchNodes: [
+          {
+            backendNodeId: 42,
+            frameId: "child",
+            tag: "button",
+            rect: { x: 410, y: 20, w: 100, h: 30 },
+            localRect: { x: 10, y: 20, w: 100, h: 30 },
+          },
+        ],
+        refs: [
+          {
+            ref: "e1",
+            backendNodeId: 42,
+            frameId: "child",
+            role: "button",
+            name: "表格视图",
+            line: 1,
+          },
+        ],
+      }),
+    };
+    captureRecordingObservation
+      .mockResolvedValueOnce(observation())
+      .mockResolvedValueOnce(iframeObservation)
+      .mockResolvedValueOnce(iframeObservation);
+    const recording = runtime();
+
+    await recording.captureInitial(7);
+    await recording.refreshDocument(7, "producer-1");
+    const drafts: RecordingDraftStep[] = [
+      {
+        op: "click",
+        captureTarget: { tag: "button", role: "button", name: "表格视图" },
+        targetHint: {
+          geometry: { rect: { x: 10, y: 20, w: 100, h: 30 }, tag: "button" },
+        },
+      },
+    ];
+    await recording.processDraft(7, drafts, 0, "producer-1");
+    recording.cancel();
+
+    expect(captureRecordingObservation).toHaveBeenCalledTimes(3);
+    expect(waitForDocumentSettled).toHaveBeenCalledWith(
+      expect.anything(),
+      { frameId: "child", target: { tabId: 7, sessionId: "oopif-session" } },
+      { signal: expect.any(AbortSignal) },
+    );
+    const draft = drafts[0];
+    expect(draft?.op).toBe("click");
+    if (!draft || draft.op !== "click") throw new Error("expected click draft");
+    expect(draft.matchedTarget).toMatchObject({ ref: "e1", name: "表格视图" });
+  });
+
+  it("does not poison final flush when a frame readiness refresh fails", async () => {
+    captureRecordingObservation
+      .mockResolvedValueOnce(observation())
+      .mockRejectedValueOnce(new Error("child Document replaced"));
+    const recording = runtime();
+
+    await recording.captureInitial(7);
+    await expect(recording.refreshDocument(7, "producer-1")).rejects.toThrow(
+      "child Document replaced",
+    );
+    await expect(recording.flush()).resolves.toBeUndefined();
+  });
+});

apps/extension/src/lib/__tests__/settle-controller.test.ts

@@ -0,0 +1,136 @@
+import { afterEach, describe, expect, it, vi } from "vitest";
+import type { CdpRunner, ChromeTabsApi } from "@/tools/shared";
+import { ObservationNodeIndex, type RegisteredObservation } from "../recording/observation-capture";
+import { RecordingObservationSession } from "../recording/observation-session";
+import { SettleController } from "../recording/settle-controller";
+import type { RecordingDraftStep } from "../recording/types";
+
+const OBSERVATION: RegisteredObservation = {
+  stateId: "s-next",
+  rootFrameId: "root",
+  index: new ObservationNodeIndex({ rootFrameId: "root", matchNodes: [], refs: [] }),
+  url: "https://example.com/next",
+};
+
+describe("SettleController", () => {
+  afterEach(() => vi.useRealTimers());
+
+  it("aborts superseded capture work and lands the prior action on the next origin", async () => {
+    vi.useFakeTimers();
+    const cdp: CdpRunner = {
+      send: vi.fn(async () => ({
+        result: { value: { idleMs: 1_000, readyState: "complete" } },
+      })) as unknown as CdpRunner["send"],
+    };
+    const tabsApi: ChromeTabsApi = {
+      get: vi.fn(async () => ({ id: 7, url: OBSERVATION.url }) as chrome.tabs.Tab),
+      query: vi.fn(async () => []),
+    };
+    const session = new RecordingObservationSession();
+    let firstSignal: AbortSignal | undefined;
+    vi.spyOn(session, "capture")
+      .mockImplementationOnce(async (_cdp, _tabs, _tabId, signal) => {
+        firstSignal = signal;
+        return new Promise<RegisteredObservation>((_resolve, reject) => {
+          signal?.addEventListener(
+            "abort",
+            () => reject(new DOMException("observation aborted", "AbortError")),
+            { once: true },
+          );
+        });
+      })
+      .mockResolvedValueOnce(OBSERVATION);
+
+    const drafts: RecordingDraftStep[] = [
+      { op: "click", captureTarget: { tag: "button", name: "First" }, preStateId: "s1" },
+    ];
+    const controller = new SettleController({ session, cdp, tabsApi, tabId: 7 });
+    controller.schedule(drafts, 0);
+    await vi.advanceTimersByTimeAsync(180);
+
+    drafts.push({
+      op: "click",
+      captureTarget: { tag: "button", name: "Second" },
+      preStateId: "s-next",
+    });
+    controller.schedule(drafts, 1);
+    expect(firstSignal?.aborted).toBe(true);
+    expect(drafts[0]?.postStateId).toBe("s-next");
+
+    await vi.advanceTimersByTimeAsync(500);
+    await controller.flush();
+    expect(drafts[1]?.postStateId).toBe("s-next");
+  });
+
+  it("does not consume a newer redirect while reading the prior landing URL", async () => {
+    vi.useFakeTimers();
+    const cdp: CdpRunner = {
+      send: vi.fn(async () => ({
+        result: { value: { idleMs: 1_000, readyState: "complete" } },
+      })) as unknown as CdpRunner["send"],
+    };
+    let resolveFirstTab!: (tab: chrome.tabs.Tab) => void;
+    const firstTab = new Promise<chrome.tabs.Tab>((resolve) => {
+      resolveFirstTab = resolve;
+    });
+    const tabsApi: ChromeTabsApi = {
+      get: vi
+        .fn()
+        .mockImplementationOnce(() => firstTab)
+        .mockResolvedValue({ id: 7, url: OBSERVATION.url }),
+      query: vi.fn(async () => []),
+    };
+    const session = new RecordingObservationSession();
+    session.cursor.lastSettled = OBSERVATION;
+    const drafts: RecordingDraftStep[] = [];
+    const controller = new SettleController({ session, cdp, tabsApi, tabId: 7 });
+
+    controller.scheduleRedirect(drafts, "https://example.com/intermediate");
+    await vi.advanceTimersByTimeAsync(180);
+    expect(tabsApi.get).toHaveBeenCalledTimes(1);
+
+    controller.scheduleRedirect(drafts, OBSERVATION.url);
+    resolveFirstTab({ id: 7, url: "https://example.com/intermediate" } as chrome.tabs.Tab);
+    await vi.advanceTimersByTimeAsync(500);
+    await controller.flushRedirects();
+
+    expect(drafts).toEqual([]);
+    expect(controller.hasPending).toBe(false);
+  });
+
+  it("settles an action in its own OOPIF document scope", async () => {
+    vi.useFakeTimers();
+    const send = vi.fn();
+    const sendToTarget = vi.fn(async (_target, method: string) => {
+      if (method === "Page.createIsolatedWorld") return { executionContextId: 39 };
+      return { result: { value: { idleMs: 1_000, readyState: "complete" } } };
+    });
+    const cdp: CdpRunner = {
+      send: send as unknown as CdpRunner["send"],
+      sendToTarget: sendToTarget as unknown as NonNullable<CdpRunner["sendToTarget"]>,
+    };
+    const tabsApi: ChromeTabsApi = {
+      get: vi.fn(async () => ({ id: 7, url: OBSERVATION.url }) as chrome.tabs.Tab),
+      query: vi.fn(async () => []),
+    };
+    const session = new RecordingObservationSession();
+    vi.spyOn(session, "capture").mockResolvedValue(OBSERVATION);
+    const drafts: RecordingDraftStep[] = [
+      { op: "click", captureTarget: { tag: "button", name: "Inside frame" } },
+    ];
+    const controller = new SettleController({ session, cdp, tabsApi, tabId: 7 });
+    const target = { tabId: 7, sessionId: "oopif-session" };
+
+    controller.schedule(drafts, 0, { target, frameId: "oopif-frame" });
+    await vi.advanceTimersByTimeAsync(180);
+    await controller.flush();
+
+    expect(send).not.toHaveBeenCalled();
+    expect(sendToTarget).toHaveBeenCalledWith(
+      target,
+      "Runtime.evaluate",
+      expect.objectContaining({ contextId: 39 }),
+    );
+    expect(drafts[0]?.postStateId).toBe("s-next");
+  });
+});

apps/extension/src/lib/__tests__/step-buffer.test.ts

@@ -1,9 +1,9 @@
 import { describe, expect, it } from "vitest";
-import { appendRecordedPayload, observeRecordedNavigation } from "../recording-step-buffer";
+import { appendRecordedPayload, observeRecordedNavigation } from "../recording/step-buffer";
 
 describe("recording-step-buffer", () => {
   it("stores semantic click without summary", () => {
-    const buffer = { steps: [], pendingNavigation: false };
+    const buffer = { steps: [], navigation: { pendingNavigation: false } };
     appendRecordedPayload(buffer, {
       op: "click",
       target: { tag: "button", role: "button", name: "发布" },
@@ -12,37 +12,40 @@ describe("recording-step-buffer", () => {
     expect(buffer.steps).toEqual([
       {
         op: "click",
-        target: { unmatched: true },
         captureTarget: { tag: "button", role: "button", name: "发布" },
       },
     ]);
-    expect(buffer.pendingNavigation).toBe(true);
+    expect(buffer.navigation.pendingNavigation).toBe(true);
   });
 
   it("keeps the hovered element description as a capture fallback", () => {
-    const buffer = { steps: [], pendingNavigation: false };
-    appendRecordedPayload(buffer, {
-      op: "hover",
-      target: { tag: "button", role: "button", name: "新建" },
-      geometry: {
-        rect: { x: 900, y: 8, w: 60, h: 32 },
-        scrollX: 0,
-        scrollY: 0,
-        position: "static",
-        tag: "button",
+    const buffer = { steps: [], navigation: { pendingNavigation: false } };
+    appendRecordedPayload(
+      buffer,
+      {
+        op: "hover",
+        target: { tag: "button", role: "button", name: "新建" },
+        geometry: {
+          rect: { x: 900, y: 8, w: 60, h: 32 },
+          tag: "button",
+        },
       },
-    });
+      {
+        geometry: {
+          rect: { x: 900, y: 8, w: 60, h: 32 },
+          tag: "button",
+        },
+      },
+    );
     expect(buffer.steps).toEqual([
       {
         op: "hover",
-        target: { unmatched: true },
         captureTarget: { tag: "button", role: "button", name: "新建" },
-        geometry: {
-          rect: { x: 900, y: 8, w: 60, h: 32 },
-          scrollX: 0,
-          scrollY: 0,
-          position: "static",
-          tag: "button",
+        targetHint: {
+          geometry: {
+            rect: { x: 900, y: 8, w: 60, h: 32 },
+            tag: "button",
+          },
         },
       },
     ]);
@@ -53,19 +56,21 @@ describe("recording-step-buffer", () => {
       steps: [
         {
           op: "click" as const,
-          target: { tag: "button", role: "button", name: "发布" },
+          captureTarget: { tag: "button", role: "button", name: "发布" },
         },
       ],
-      currentUrl: "https://example.com/a",
-      pendingNavigation: true,
-      pendingNavigationDeadline: Date.now() + 5_000,
+      navigation: {
+        currentUrl: "https://example.com/a",
+        pendingNavigation: true,
+        pendingNavigationDeadline: Date.now() + 5_000,
+      },
     };
     observeRecordedNavigation(buffer, "https://example.com/b", true);
     expect(buffer.steps).toEqual([
       {
         op: "click",
-        target: { tag: "button", role: "button", name: "发布" },
-        navigated_to: "https://example.com/b",
+        captureTarget: { tag: "button", role: "button", name: "发布" },
+        navigatedTo: "https://example.com/b",
       },
     ]);
     expect(JSON.stringify(buffer.steps)).not.toContain("wait_for_navigation");
@@ -76,21 +81,23 @@ describe("recording-step-buffer", () => {
       steps: [
         {
           op: "select" as const,
-          target: { tag: "select", role: "combobox", name: "分类" },
+          captureTarget: { tag: "select", role: "combobox", name: "分类" },
           values: ["tech"],
         },
       ],
-      currentUrl: "https://example.com/list",
-      pendingNavigation: true,
-      pendingNavigationDeadline: Date.now() + 5_000,
+      navigation: {
+        currentUrl: "https://example.com/list",
+        pendingNavigation: true,
+        pendingNavigationDeadline: Date.now() + 5_000,
+      },
     };
     observeRecordedNavigation(buffer, "https://example.com/list?cat=tech", true);
     expect(buffer.steps).toEqual([
       {
         op: "select",
-        target: { tag: "select", role: "combobox", name: "分类" },
+        captureTarget: { tag: "select", role: "combobox", name: "分类" },
         values: ["tech"],
-        navigated_to: "https://example.com/list?cat=tech",
+        navigatedTo: "https://example.com/list?cat=tech",
       },
     ]);
   });
@@ -98,8 +105,7 @@ describe("recording-step-buffer", () => {
   it("emits navigate for uncaused URL changes", () => {
     const buffer = {
       steps: [],
-      currentUrl: "https://example.com/a",
-      pendingNavigation: false,
+      navigation: { currentUrl: "https://example.com/a", pendingNavigation: false },
     };
     const result = observeRecordedNavigation(buffer, "https://example.com/b", false);
     expect(result).toEqual({ kind: "appended", index: 0 });
@@ -112,8 +118,10 @@ describe("recording-step-buffer", () => {
   it("asks the recorder to coalesce redirect hops instead of emitting each one", () => {
     const buffer = {
       steps: [],
-      currentUrl: "https://passport.example/login",
-      pendingNavigation: false,
+      navigation: {
+        currentUrl: "https://passport.example/login",
+        pendingNavigation: false,
+      },
     };
     const hop1 = observeRecordedNavigation(
       buffer,
@@ -136,6 +144,21 @@ describe("recording-step-buffer", () => {
       url: "https://app.example/dashboard",
     });
     expect(buffer.steps).toEqual([]);
-    expect(buffer.currentUrl).toBe("https://app.example/dashboard");
+    expect(buffer.navigation.currentUrl).toBe("https://app.example/dashboard");
+  });
+
+  it("does not reuse redirect metadata for a later content-observed URL change", () => {
+    const buffer = {
+      steps: [],
+      navigation: {
+        currentUrl: "https://example.com/redirect",
+        pendingNavigation: false,
+      },
+    };
+
+    const result = observeRecordedNavigation(buffer, "https://example.com/spa");
+
+    expect(result).toEqual({ kind: "appended", index: 0 });
+    expect(buffer.steps[0]).toMatchObject({ op: "navigate", url: "https://example.com/spa" });
   });
 });

apps/extension/src/lib/__tests__/tab-coordinator.test.ts

@@ -0,0 +1,39 @@
+import { describe, expect, it } from "vitest";
+import { RecordingTabCoordinator } from "../recording/tab-coordinator";
+
+describe("RecordingTabCoordinator", () => {
+  it("keeps navigation state isolated per tab", () => {
+    const tabs = new RecordingTabCoordinator(4, "https://example.com/first");
+    const first = tabs.navigation(4);
+    first.pendingNavigation = true;
+    const second = tabs.navigation(5, "https://example.com/second");
+
+    expect(second).toEqual({
+      currentUrl: "https://example.com/second",
+      pendingNavigation: false,
+    });
+    expect(tabs.navigation(4)).toBe(first);
+    expect(tabs.navigation(4).pendingNavigation).toBe(true);
+  });
+
+  it("invalidates an earlier activation when a newer tab becomes active", () => {
+    const tabs = new RecordingTabCoordinator(4);
+    const second = tabs.noteActivation(5);
+    const third = tabs.noteActivation(6);
+
+    expect(tabs.isLatest(second)).toBe(false);
+    expect(tabs.isLatest(third)).toBe(true);
+    tabs.commit(6);
+    expect(tabs.currentTabId).toBe(6);
+  });
+
+  it("commits the freshly observed URL when returning to an existing tab", () => {
+    const tabs = new RecordingTabCoordinator(4);
+    tabs.navigation(5, "https://example.com/old");
+
+    tabs.commit(5, "https://example.com/new");
+
+    expect(tabs.currentTabId).toBe(5);
+    expect(tabs.navigation(5).currentUrl).toBe("https://example.com/new");
+  });
+});

apps/extension/src/lib/__tests__/target-matcher.test.ts

@@ -0,0 +1,159 @@
+import type { RenderedRef } from "@browser-skill/vom";
+import { describe, expect, it } from "vitest";
+import type { CaptureVomMatchNode } from "@/tools/capture-vom-observation";
+import { ObservationNodeIndex, type RegisteredObservation } from "../recording/observation-capture";
+import { matchObservationTarget } from "../recording/target-matcher";
+
+function node(backendNodeId: number, frameId: string, x = 10): CaptureVomMatchNode {
+  return {
+    backendNodeId,
+    frameId,
+    tag: "button",
+    rect: { x, y: 20, w: 100, h: 30 },
+    localRect: { x, y: 20, w: 100, h: 30 },
+  };
+}
+
+function observation(
+  matchNodes: CaptureVomMatchNode[],
+  refs: RenderedRef[],
+): RegisteredObservation {
+  return {
+    stateId: "s1",
+    rootFrameId: "root",
+    index: new ObservationNodeIndex({ rootFrameId: "root", matchNodes, refs }),
+    url: "https://example.com",
+  };
+}
+
+describe("matchObservationTarget", () => {
+  it("indexes the safe geometry DTO without discarding frame-local geometry", () => {
+    const geometry = node(42, "child");
+    geometry.localRect = { x: 5, y: 6, w: 100, h: 30 };
+    const ref: RenderedRef = {
+      ref: "e1",
+      backendNodeId: 42,
+      frameId: "child",
+      line: 1,
+    };
+    const candidate = observation([geometry], [ref]).index.candidates("child", "button")[0];
+
+    expect(candidate).toEqual({ frameId: "child", geometry, ref });
+  });
+
+  it("matches a unique node using canonical top-level viewport geometry", () => {
+    const target = matchObservationTarget({
+      observation: observation(
+        [node(42, "root")],
+        [{ ref: "e1", backendNodeId: 42, role: "button", name: "发布", line: 1 }],
+      ),
+      hint: { geometry: { rect: { x: 10, y: 20, w: 100, h: 30 }, tag: "button" } },
+    });
+    expect(target).toEqual({ ref: "e1", role: "button", name: "发布" });
+  });
+
+  it("uses frame id with backend node id so sibling frames cannot collide", () => {
+    const target = matchObservationTarget({
+      observation: observation(
+        [node(42, "left"), node(42, "right")],
+        [
+          { ref: "e1", backendNodeId: 42, frameId: "left", line: 1 },
+          { ref: "e2", backendNodeId: 42, frameId: "right", line: 2 },
+        ],
+      ),
+      hint: {
+        frameId: "right",
+        geometry: { rect: { x: 10, y: 20, w: 100, h: 30 }, tag: "button" },
+      },
+    });
+    expect(target.ref).toBe("e2");
+  });
+
+  it("matches iframe capture geometry in the child viewport coordinate space", () => {
+    const child = node(42, "child", 410);
+    child.localRect = { x: 10, y: 20, w: 100, h: 30 };
+    const target = matchObservationTarget({
+      observation: observation(
+        [child],
+        [{ ref: "e1", backendNodeId: 42, frameId: "child", line: 1 }],
+      ),
+      hint: {
+        frameId: "child",
+        geometrySpace: "local",
+        geometry: { rect: { x: 10, y: 20, w: 100, h: 30 }, tag: "button" },
+      },
+    });
+    expect(target.ref).toBe("e1");
+  });
+
+  it("restricts a missing frame hint to the root frame", () => {
+    const target = matchObservationTarget({
+      observation: observation(
+        [node(42, "child")],
+        [{ ref: "e1", backendNodeId: 42, frameId: "child", line: 1 }],
+      ),
+      hint: { geometry: { rect: { x: 10, y: 20, w: 100, h: 30 }, tag: "button" } },
+      fallback: { tag: "button", name: "发布" },
+    });
+    expect(target).toEqual({ name: "发布", unmatched: true });
+  });
+
+  it("fails closed when the source Document has no VOM frame mapping", () => {
+    const target = matchObservationTarget({
+      observation: observation(
+        [node(42, "root")],
+        [{ ref: "e1", backendNodeId: 42, role: "button", name: "发布", line: 1 }],
+      ),
+      hint: {
+        frameId: null,
+        geometry: { rect: { x: 10, y: 20, w: 100, h: 30 }, tag: "button" },
+      },
+      fallback: { tag: "button", role: "button", name: "发布" },
+    });
+    expect(target).toEqual({ role: "button", name: "发布", unmatched: true });
+  });
+
+  it("returns unmatched for ambiguous geometry", () => {
+    const target = matchObservationTarget({
+      observation: observation(
+        [node(42, "root"), node(43, "root")],
+        [
+          { ref: "e1", backendNodeId: 42, line: 1 },
+          { ref: "e2", backendNodeId: 43, line: 2 },
+        ],
+      ),
+      hint: { geometry: { rect: { x: 10, y: 20, w: 100, h: 30 }, tag: "button" } },
+    });
+    expect(target.unmatched).toBe(true);
+  });
+
+  it("uses semantics when geometry is unavailable without crossing frame boundaries", () => {
+    const target = matchObservationTarget({
+      observation: observation(
+        [node(41, "root"), node(42, "child")],
+        [
+          { ref: "e1", backendNodeId: 41, role: "button", name: "保存", line: 1 },
+          { ref: "e2", backendNodeId: 42, frameId: "child", role: "button", name: "保存", line: 2 },
+        ],
+      ),
+      hint: { frameId: "child" },
+      fallback: { tag: "button", role: "button", name: "保存" },
+    });
+    expect(target.ref).toBe("e2");
+  });
+
+  it("uses semantics to disambiguate equal geometry in one frame", () => {
+    const target = matchObservationTarget({
+      observation: observation(
+        [node(42, "root"), node(43, "root")],
+        [
+          { ref: "e1", backendNodeId: 42, role: "button", name: "保存", line: 1 },
+          { ref: "e2", backendNodeId: 43, role: "button", name: "取消", line: 2 },
+        ],
+      ),
+      hint: { geometry: { rect: { x: 10, y: 20, w: 100, h: 30 }, tag: "button" } },
+      fallback: { tag: "button", role: "button", name: "取消" },
+    });
+    expect(target.ref).toBe("e2");
+  });
+});

apps/extension/src/lib/__tests__/trace-reducer-v2.test.ts

@@ -1,70 +1,202 @@
 import { describe, expect, it } from "vitest";
-import { buildTraceV2, shouldRecordPress } from "@/lib/trace-reducer-v2";
-import type { DraftTraceStep } from "@/transport/types";
+import type { RecordingDraftStep } from "@/lib/recording/types";
+import { shouldRecordPress } from "../recording/draft-policy";
+import { buildTraceV2 } from "../recording/trace-reducer-v2";
 
-describe("trace-reducer-v2", () => {
-  it("drops scroll and bare character press steps", () => {
-    const steps: DraftTraceStep[] = [
+function reduceTraceSteps(steps: RecordingDraftStep[], startUrl?: string) {
+  return buildTraceV2({
+    steps,
+    startedAt: "2026-01-01T00:00:00.000Z",
+    ...(startUrl ? { startUrl } : {}),
+  });
+}
+
+describe("shouldRecordPress", () => {
+  it("keeps Enter and Escape", () => {
+    expect(shouldRecordPress("Enter")).toBe(true);
+    expect(shouldRecordPress("Escape")).toBe(true);
+  });
+
+  it("drops modifiers, clipboard shortcuts, and bare typing", () => {
+    expect(shouldRecordPress("Meta")).toBe(false);
+    expect(shouldRecordPress("c", ["meta"])).toBe(false);
+    expect(shouldRecordPress("a", ["ctrl"])).toBe(false);
+    expect(shouldRecordPress("x")).toBe(false);
+    expect(shouldRecordPress("中")).toBe(false);
+  });
+});
+
+describe("reduceTraceSteps", () => {
+  it("does not expose internal tab transitions in trace v2", () => {
+    const trace = reduceTraceSteps([
       {
-        op: "scroll",
-        page_url: "https://example.com/",
+        op: "switch_tab",
+        preStateId: "s1",
+        postStateId: "s2",
+      },
+    ]);
+    expect(trace.steps).toEqual([]);
+  });
+
+  it("builds steps with pages dictionary and page id references", () => {
+    const drafts: RecordingDraftStep[] = [
+      {
+        op: "navigate",
+        url: "https://example.com/search?q=hello&utm_source=x",
+        pageUrl: "https://example.com/search?q=hello&utm_source=x",
+      },
+      {
+        op: "fill",
+        captureTarget: { tag: "input", role: "textbox", name: "搜索", name_attr: "q" },
+        value: "browser skill",
+        pageUrl: "https://example.com/search?q=hello&utm_source=x",
+      },
+      {
+        op: "press",
+        key: "Enter",
+        captureTarget: { tag: "input", role: "textbox", name: "搜索", name_attr: "q" },
+        navigatedTo: "https://example.com/results/42",
+        pageUrl: "https://example.com/search?q=hello&utm_source=x",
+      },
+      {
+        op: "click",
+        captureTarget: { tag: "button", role: "button", name: "发布" },
+        navigatedTo: "https://example.com/p/99",
+        pageUrl: "https://example.com/results/42",
       },
       {
         op: "press",
         key: "a",
-        page_url: "https://example.com/",
+        pageUrl: "https://example.com/p/99",
       },
     ];
-    const trace = buildTraceV2({
-      steps,
-      startedAt: "2026-01-01T00:00:00.000Z",
-      startUrl: "https://example.com/",
+
+    const { pages, steps } = reduceTraceSteps(
+      drafts,
+      "https://example.com/search?q=hello&utm_source=x",
+    );
+    expect(JSON.stringify(steps)).not.toContain("parameters");
+    expect(JSON.stringify(steps)).not.toContain("intent");
+    expect(JSON.stringify(steps)).not.toContain("summary");
+    expect(steps.map((s) => s.op)).toEqual(["navigate", "fill", "press", "click"]);
+    expect(pages.map((p) => p.url)).toEqual([
+      "https://example.com/search?q=hello&utm_source=x",
+      "https://example.com/results/42",
+      "https://example.com/p/99",
+    ]);
+
+    expect(steps[0]).toMatchObject({
+      id: 1,
+      op: "navigate",
+      page: "p1",
+      to: "https://example.com/search?q=hello&utm_source=x",
+    });
+
+    expect(steps[1]).toMatchObject({
+      op: "fill",
+      page: "p1",
+      value: "browser skill",
+    });
+
+    expect(steps[2]).toMatchObject({
+      op: "press",
+      key: "Enter",
+      page: "p1",
+      effect: { navigated_to: "p2" },
+    });
+
+    expect(steps[3]).toMatchObject({
+      op: "click",
+      page: "p2",
+      effect: { navigated_to: "p3" },
     });
-    expect(trace.steps).toHaveLength(0);
-    expect(trace.pages).toHaveLength(1);
   });
 
-  it("drops hover steps unsupported by historical v2 clients", () => {
-    const trace = buildTraceV2({
-      steps: [
+  it("collapses consecutive navigations", () => {
+    const { steps } = reduceTraceSteps([
+      { op: "navigate", url: "https://a.example/redirect1" },
+      { op: "navigate", url: "https://a.example/final" },
+    ]);
+    expect(steps).toHaveLength(1);
+    expect(steps[0]).toMatchObject({
+      op: "navigate",
+      to: "https://a.example/final",
+    });
+  });
+
+  it("keeps committed empty fill steps", () => {
+    const { steps } = reduceTraceSteps(
+      [
         {
-          op: "hover",
-          target: { unmatched: true },
-          captureTarget: { tag: "button", role: "button", name: "结束" },
-          page_url: "https://example.com/",
+          op: "fill",
+          captureTarget: { tag: "input", role: "textbox", name: "Search query" },
+          value: "",
+          pageUrl: "https://example.com/search",
         },
       ],
-      startedAt: "2026-01-01T00:00:00.000Z",
-      startUrl: "https://example.com/",
-    });
+      "https://example.com/search",
+    );
 
-    expect(trace.steps).toHaveLength(0);
+    expect(steps).toEqual([
+      expect.objectContaining({
+        op: "fill",
+        value: "",
+      }),
+    ]);
   });
 
-  it("maps captureTarget to v2 target with required tag", () => {
-    const steps: DraftTraceStep[] = [
-      {
-        op: "click",
-        target: { unmatched: true },
-        captureTarget: { tag: "button", role: "button", name: "Submit" },
-        page_url: "https://example.com/",
-      },
-    ];
-    const trace = buildTraceV2({
-      steps,
-      startedAt: "2026-01-01T00:00:00.000Z",
-      startUrl: "https://example.com/",
+  it("maps select navigatedTo onto effect.navigatedTo (page id)", () => {
+    const { pages, steps } = reduceTraceSteps(
+      [
+        {
+          op: "select",
+          captureTarget: { tag: "select", role: "combobox", name: "分类" },
+          values: ["tech"],
+          labels: ["技术"],
+          navigatedTo: "https://example.com/list?cat=tech",
+          pageUrl: "https://example.com/list",
+        },
+      ],
+      "https://example.com/list",
+    );
+    expect(pages.map((p) => p.url)).toEqual([
+      "https://example.com/list",
+      "https://example.com/list?cat=tech",
+    ]);
+    expect(steps[0]).toMatchObject({
+      op: "select",
+      page: "p1",
+      selection: [{ value: "tech", label: "技术" }],
+      effect: { navigated_to: "p2" },
     });
-    expect(trace.steps).toHaveLength(1);
-    expect(trace.steps[0]?.op).toBe("click");
-    if (trace.steps[0]?.op === "click") {
-      expect(trace.steps[0].target.tag).toBe("button");
-      expect(trace.steps[0].target.name).toBe("Submit");
-    }
-    expect("version" in trace).toBe(false);
   });
 
-  it("records Enter presses", () => {
-    expect(shouldRecordPress("Enter")).toBe(true);
+  it("drops hover steps, which only exist in trace v3", () => {
+    const { steps } = reduceTraceSteps(
+      [
+        {
+          op: "hover",
+          captureTarget: { tag: "span", role: "button", name: "Account" },
+          pageUrl: "https://example.com/app",
+        },
+        {
+          op: "click",
+          captureTarget: { tag: "a", role: "link", name: "Profile" },
+          pageUrl: "https://example.com/app",
+        },
+      ],
+      "https://example.com/app",
+    );
+    expect(steps.map((s) => s.op)).toEqual(["click"]);
+  });
+
+  it("resolveTraceStartUrl prefers explicit start URL", () => {
+    expect(
+      buildTraceV2({
+        steps: [{ op: "navigate", url: "https://example.com/other" }],
+        startedAt: "2026-01-01T00:00:00.000Z",
+        startUrl: "https://example.com/start",
+      }).entry.start_url,
+    ).toBe("https://example.com/start");
   });
 });

apps/extension/src/lib/__tests__/trace-reducer-v3.test.ts

@@ -0,0 +1,132 @@
+import { describe, expect, it } from "vitest";
+import { TRACE_VERSION_V3, VOM_FORMAT_VERSION } from "@/transport/types";
+import { RecordingStateRegistry } from "../recording/state-registry";
+import { buildTraceV3 } from "../recording/trace-builder-v3";
+import { reduceTraceStepsV3 } from "../recording/trace-reducer-v3";
+import type { RecordingDraftStep } from "../recording/types";
+
+describe("trace reducer v3", () => {
+  it("removes form literals when value redaction is requested", () => {
+    const reduced = reduceTraceStepsV3(
+      [
+        {
+          op: "fill",
+          value: "private@example.com",
+          preStateId: "s1",
+          postStateId: "s1",
+        },
+        {
+          op: "select",
+          values: ["private-account-id"],
+          labels: ["Private account"],
+          preStateId: "s1",
+          postStateId: "s1",
+        },
+      ],
+      { redactValues: true },
+    );
+
+    expect(reduced.steps[0]).toMatchObject({ op: "fill", value: "***", redacted: true });
+    expect(reduced.steps[1]).toMatchObject({ op: "select" });
+    expect(reduced.steps[1]).not.toHaveProperty("selection");
+    expect(JSON.stringify(reduced.steps)).not.toContain("private@example.com");
+    expect(JSON.stringify(reduced.steps)).not.toContain("private-account-id");
+  });
+
+  it("emits tab transitions only for callers that advertised support", () => {
+    const drafts: RecordingDraftStep[] = [
+      {
+        op: "switch_tab",
+        preStateId: "s1",
+        postStateId: "s2",
+      },
+    ];
+
+    expect(reduceTraceStepsV3(drafts).steps).toEqual([]);
+    expect(reduceTraceStepsV3(drafts, { includeTabSwitches: true }).steps).toEqual([
+      { op: "switch_tab", id: 1, state: "s1", result: { state: "s2" } },
+    ]);
+  });
+
+  it("keeps hover steps before menu clicks", () => {
+    const drafts: RecordingDraftStep[] = [
+      {
+        op: "hover",
+        captureTarget: { tag: "span", role: "button", name: "Account" },
+        preStateId: "s1",
+        postStateId: "s1",
+      },
+      {
+        op: "click",
+        captureTarget: { tag: "a", role: "link", name: "Profile" },
+        preStateId: "s1",
+        postStateId: "s1",
+      },
+    ];
+
+    expect(reduceTraceStepsV3(drafts).steps.map((s) => s.op)).toEqual(["hover", "click"]);
+  });
+
+  it("does not export a tab transition without both observation endpoints", () => {
+    const sourceOnly: RecordingDraftStep = { op: "switch_tab", preStateId: "s1" };
+    const targetOnly: RecordingDraftStep = { op: "switch_tab", postStateId: "s2" };
+
+    expect(
+      reduceTraceStepsV3([sourceOnly, targetOnly], { includeTabSwitches: true }).steps,
+    ).toEqual([]);
+  });
+
+  it("collapses redirect hops while retaining draft-to-step identity", () => {
+    const drafts: RecordingDraftStep[] = [
+      { op: "navigate", url: "https://example.com/start", preStateId: "s1", postStateId: "s2" },
+      {
+        op: "navigate",
+        url: "https://example.com/final",
+        transitionQualifiers: ["server_redirect"],
+        preStateId: "s2",
+        postStateId: "s3",
+      },
+    ];
+    const before = structuredClone(drafts);
+    const reduced = reduceTraceStepsV3(drafts);
+
+    expect(drafts).toEqual(before);
+    expect(reduced.steps).toEqual([
+      expect.objectContaining({
+        id: 1,
+        op: "navigate",
+        state: "s1",
+        to: "https://example.com/final",
+        result: { state: "s3" },
+      }),
+    ]);
+    expect(reduced.stepIdByDraftId.get(1)).toBe(1);
+    expect(reduced.stepIdByDraftId.get(2)).toBe(1);
+  });
+
+  it("builds the wire model from protocol constants", () => {
+    const registry = new RecordingStateRegistry();
+    const state = registry.register({ url: "https://example.com", vomText: "@vom 1" });
+    const trace = buildTraceV3({
+      registry,
+      drafts: [
+        {
+          op: "click",
+          captureTarget: { tag: "button", role: "button", name: "Save" },
+          preStateId: state.id,
+          postStateId: state.id,
+        },
+      ],
+      startedAt: "2026-08-12T00:00:00.000Z",
+      stoppedBy: "user_finish",
+      bskVersion: "test",
+    });
+
+    expect(trace.version).toBe(TRACE_VERSION_V3);
+    expect(trace.recorder.vom).toBe(VOM_FORMAT_VERSION);
+    expect(trace.steps[0]).toMatchObject({
+      op: "click",
+      target: { role: "button", name: "Save", unmatched: true },
+    });
+  });
+});

apps/extension/src/lib/__tests__/trace-reducer.test.ts

@@ -1,252 +0,0 @@
-import { describe, expect, it } from "vitest";
-import { resetStateIdCounterForTests } from "@/lib/record-constants";
-import {
-  reduceTraceSteps,
-  registerObservation,
-  resolveTraceStartUrl,
-  shouldRecordPress,
-} from "../trace-reducer";
-
-describe("shouldRecordPress", () => {
-  it("keeps Enter and Escape", () => {
-    expect(shouldRecordPress("Enter")).toBe(true);
-    expect(shouldRecordPress("Escape")).toBe(true);
-  });
-
-  it("drops modifiers, clipboard shortcuts, and bare typing", () => {
-    expect(shouldRecordPress("Meta")).toBe(false);
-    expect(shouldRecordPress("c", ["meta"])).toBe(false);
-    expect(shouldRecordPress("a", ["ctrl"])).toBe(false);
-    expect(shouldRecordPress("x")).toBe(false);
-    expect(shouldRecordPress("中")).toBe(false);
-  });
-});
-
-describe("reduceTraceSteps v3", () => {
-  it("builds steps with state/result references and deduped states", () => {
-    resetStateIdCounterForTests();
-    const registry = new Map();
-    const body = "@vom 1\nL1 page";
-    const s1 = registerObservation(registry, {
-      url: "https://example.com/search",
-      rawVomText: body,
-    });
-    const s2 = registerObservation(registry, {
-      url: "https://example.com/results/42",
-      rawVomText: "@vom 1\nL1 results",
-    });
-
-    const { states, steps } = reduceTraceSteps(
-      [
-        {
-          op: "navigate",
-          url: "https://example.com/search",
-          preStateId: s1,
-          postStateId: s1,
-        },
-        {
-          op: "fill",
-          target: { ref: "e1", role: "textbox", name: "搜索" },
-          value: "browser skill",
-          commit: "enter",
-          preStateId: s1,
-          postStateId: s1,
-        },
-        {
-          op: "press",
-          key: "Enter",
-          target: { ref: "e1", role: "textbox", name: "搜索" },
-          preStateId: s1,
-          postStateId: s2,
-          navigated_to: "https://example.com/results/42",
-        },
-        {
-          op: "click",
-          target: { ref: "e2", role: "button", name: "发布" },
-          preStateId: s2,
-          postStateId: s2,
-        },
-      ],
-      registry,
-    );
-
-    expect(states).toHaveLength(2);
-    expect(steps.map((s) => s.op)).toEqual(["navigate", "fill", "press", "click"]);
-    expect(steps[1]).toMatchObject({
-      op: "fill",
-      state: s1,
-      result: { state: s1 },
-      commit: "enter",
-    });
-    expect(steps[2]).toMatchObject({
-      op: "press",
-      state: s1,
-      result: { state: s2 },
-    });
-    expect(JSON.stringify(steps)).not.toContain("page");
-    expect(JSON.stringify(steps)).not.toContain("effect");
-  });
-
-  it("collapses consecutive navigations", () => {
-    resetStateIdCounterForTests();
-    const registry = new Map();
-    const s1 = registerObservation(registry, {
-      url: "https://a.example/redirect1",
-      rawVomText: "a",
-    });
-    const { steps } = reduceTraceSteps(
-      [
-        { op: "navigate", url: "https://a.example/redirect1", preStateId: s1, postStateId: s1 },
-        { op: "navigate", url: "https://a.example/final", preStateId: s1, postStateId: s1 },
-      ],
-      registry,
-    );
-    expect(steps).toHaveLength(1);
-    expect(steps[0]).toMatchObject({
-      op: "navigate",
-      to: "https://a.example/final",
-    });
-  });
-
-  it("keeps the origin state and cause of the hop that started a redirect chain", () => {
-    resetStateIdCounterForTests();
-    const registry = new Map();
-    const start = registerObservation(registry, {
-      url: "https://example.com/",
-      rawVomText: "start",
-    });
-    const hop = registerObservation(registry, {
-      url: "https://iwiki.woa.com/",
-      rawVomText: "hop",
-    });
-    const landing = registerObservation(registry, {
-      url: "https://iwiki.woa.com/dashboard",
-      rawVomText: "landing",
-    });
-
-    const { steps, stepIdByDraftId } = reduceTraceSteps(
-      [
-        {
-          op: "navigate",
-          url: "https://iwiki.woa.com/",
-          preStateId: start,
-          postStateId: hop,
-          transitionType: "typed",
-        },
-        {
-          op: "navigate",
-          url: "https://iwiki.woa.com/dashboard",
-          preStateId: hop,
-          postStateId: landing,
-          transitionType: "link",
-        },
-      ],
-      registry,
-    );
-
-    expect(steps).toHaveLength(1);
-    expect(steps[0]).toMatchObject({
-      op: "navigate",
-      id: 1,
-      state: start,
-      result: { state: landing },
-      to: "https://iwiki.woa.com/dashboard",
-      cause: "user_typed",
-    });
-    // Both drafts fold into step 1, so anything tagged with either draft id
-    // must resolve to it.
-    expect(stepIdByDraftId.get(1)).toBe(1);
-    expect(stepIdByDraftId.get(2)).toBe(1);
-  });
-
-  it("maps draft ids to published step ids across dropped drafts", () => {
-    resetStateIdCounterForTests();
-    const registry = new Map();
-    const s1 = registerObservation(registry, { url: "https://a.example/", rawVomText: "a" });
-
-    const { steps, stepIdByDraftId } = reduceTraceSteps(
-      [
-        // Dropped: bare typing press.
-        { op: "press", key: "a", preStateId: s1, postStateId: s1 },
-        { op: "click", target: { ref: "e2" }, preStateId: s1, postStateId: s1 },
-      ],
-      registry,
-    );
-
-    expect(steps.map((step) => step.op)).toEqual(["click"]);
-    expect(stepIdByDraftId.get(1)).toBeUndefined();
-    expect(stepIdByDraftId.get(2)).toBe(1);
-  });
-
-  it("keeps committed empty fills", () => {
-    resetStateIdCounterForTests();
-    const registry = new Map();
-    const s1 = registerObservation(registry, { url: "https://a.example/", rawVomText: "a" });
-
-    const { steps } = reduceTraceSteps(
-      [
-        {
-          op: "fill",
-          target: { ref: "e1", role: "textbox", name: "Search" },
-          value: "",
-          preStateId: s1,
-          postStateId: s1,
-        },
-        {
-          op: "click",
-          target: { ref: "e2", role: "button", name: "Apply" },
-          preStateId: s1,
-          postStateId: s1,
-        },
-      ],
-      registry,
-    );
-
-    expect(steps.map((step) => step.op)).toEqual(["fill", "click"]);
-    expect(steps[0]).toMatchObject({ op: "fill", value: "" });
-  });
-
-  it("keeps hover steps before menu clicks", () => {
-    resetStateIdCounterForTests();
-    const registry = new Map();
-    const state = registerObservation(registry, {
-      url: "https://example.com/app",
-      rawVomText: '@e1 button "Account"\n@e2 link "Profile"',
-    });
-    const { steps } = reduceTraceSteps(
-      [
-        {
-          op: "hover",
-          target: { ref: "e1", role: "button", name: "Account" },
-          page_url: "https://example.com/app",
-          preStateId: state,
-          postStateId: state,
-        },
-        {
-          op: "click",
-          target: { ref: "e2", role: "link", name: "Profile" },
-          page_url: "https://example.com/app",
-          preStateId: state,
-          postStateId: state,
-        },
-      ],
-      registry,
-    );
-    expect(steps.map((s) => s.op)).toEqual(["hover", "click"]);
-    expect(steps[0]).toMatchObject({
-      op: "hover",
-      state,
-      result: { state },
-      target: { name: "Account" },
-    });
-  });
-
-  it("resolveTraceStartUrl prefers explicit start URL", () => {
-    expect(
-      resolveTraceStartUrl(
-        [{ op: "navigate", url: "https://example.com/other", preStateId: "s1", postStateId: "s1" }],
-        "https://example.com/start",
-      ),
-    ).toBe("https://example.com/start");
-  });
-});

apps/extension/src/lib/connection-controller.ts

@@ -313,15 +313,6 @@ export type HandshakeVerdict =
   | { kind: "version_skew" }
   | { kind: "rejected"; reason: string };
 
-const UPDATE_BOTH_VERSION_MESSAGE =
-  "The installed bsk CLI and BrowserSkill extension are incompatible. Update both, reload the extension, then run `bsk daemon restart`.";
-const UPDATE_BSK_VERSION_MESSAGE =
-  "The installed bsk CLI is out of date for this BrowserSkill extension. Update bsk, then run `bsk daemon restart`.";
-const UPDATE_EXTENSION_VERSION_MESSAGE =
-  "This BrowserSkill extension is out of date for the installed bsk CLI. Update and reload the extension, then try again.";
-const INVALID_VERSION_MESSAGE =
-  "BrowserSkill could not verify version compatibility. Update bsk and the BrowserSkill extension, reload the extension, then run `bsk daemon restart`.";
-
 /**
  * Symmetric extension-side **protocol** compat check.
  *
@@ -348,19 +339,19 @@ export function computeConnectedState(
   if (daemonMajor === null || ourMajor === null || daemonMajor !== ourMajor) {
     return {
       kind: "rejected",
-      reason: UPDATE_BOTH_VERSION_MESSAGE,
+      reason: `protocol-major mismatch (daemon protocol v${handshake.protocol_version}, extension protocol v${PROTOCOL_VERSION})`,
     };
   }
   if (compareProtocol(handshake.protocol_version, localMinCompatibleProtocol) === null) {
     return {
       kind: "rejected",
-      reason: INVALID_VERSION_MESSAGE,
+      reason: `daemon protocol v${handshake.protocol_version} is unparseable`,
     };
   }
   if (compareProtocol(handshake.protocol_version, localMinCompatibleProtocol)! < 0) {
     return {
       kind: "rejected",
-      reason: UPDATE_BSK_VERSION_MESSAGE,
+      reason: `daemon protocol v${handshake.protocol_version} below extension min_compatible_protocol ${localMinCompatibleProtocol}`,
     };
   }
   const peerFloor = handshake.min_compatible_protocol;
@@ -368,13 +359,13 @@ export function computeConnectedState(
     if (compareProtocol(PROTOCOL_VERSION, peerFloor) === null) {
       return {
         kind: "rejected",
-        reason: INVALID_VERSION_MESSAGE,
+        reason: `daemon min_compatible_protocol ${peerFloor} is unparseable`,
       };
     }
     if (compareProtocol(PROTOCOL_VERSION, peerFloor)! < 0) {
       return {
         kind: "rejected",
-        reason: UPDATE_EXTENSION_VERSION_MESSAGE,
+        reason: `extension protocol v${PROTOCOL_VERSION} below daemon min_compatible_protocol ${peerFloor}`,
       };
     }
   }

apps/extension/src/lib/describe-target.ts

@@ -1,5 +1,5 @@
 /**
- * Build a semantic TargetDescriptor for an interacted element.
+ * Build a semantic capture descriptor for an interacted element.
  *
  * Trace steps are an LLM *textbook*: each click must say what to look for
  * on screen (usually a short visible name). Tag-only noise like

apps/extension/src/lib/format-observation-file.ts

@@ -1,62 +0,0 @@
-import type { Step } from "@/transport/types";
-import { OBSERVATION_FILE_VERSION } from "./record-constants";
-
-export interface ObservationAnnotation {
-  stepId: number;
-  op: Step["op"];
-  line: number;
-  stateId: string;
-  detail?: string;
-}
-
-function formatAnnotation({ stepId, op, detail }: ObservationAnnotation): string {
-  const suffix = detail ? `: ${detail}` : "";
-  return ` ⟵ step ${stepId}: ${op}${suffix}`;
-}
-
-/** Serialize a page observation file (front matter + VOM body + step annotations). */
-export function formatObservationFile(input: {
-  stateId: string;
-  url: string;
-  title?: string;
-  stepsHere: number[];
-  body: string;
-  annotations?: ObservationAnnotation[];
-}): string {
-  const lines: string[] = [
-    `# bsk-observation ${OBSERVATION_FILE_VERSION}`,
-    `state: ${input.stateId}`,
-    `url: ${input.url}`,
-  ];
-  if (input.title) lines.push(`title: ${input.title}`);
-  if (input.stepsHere.length > 0) {
-    lines.push(`steps_here: [${input.stepsHere.join(", ")}]`);
-  }
-  lines.push("---");
-
-  const bodyLines = input.body.split("\n");
-  const annotationMap = new Map<number, ObservationAnnotation[]>();
-  for (const ann of input.annotations ?? []) {
-    const bucket = annotationMap.get(ann.line) ?? [];
-    bucket.push(ann);
-    annotationMap.set(ann.line, bucket);
-  }
-
-  for (let i = 0; i < bodyLines.length; i += 1) {
-    let line = bodyLines[i] ?? "";
-    const anns = annotationMap.get(i);
-    if (anns) {
-      for (const ann of anns) {
-        line += formatAnnotation(ann);
-      }
-    }
-    lines.push(line);
-  }
-
-  return `${lines.join("\n")}\n`;
-}
-
-/** Hash input: VOM body **before** annotations are inserted. */
-export function observationBodyForHash(vomText: string): string {
-  return vomText;
-}

apps/extension/src/lib/help-bridge.ts

@@ -16,6 +16,13 @@ export const HELP_ACK = "bsk-help-ack";
 export const HELP_QUERY = "bsk-help-query";
 export const HELP_FINISH = "bsk-help-finish";
 
+export interface HelpHighlightRect {
+  top: number;
+  left: number;
+  width: number;
+  height: number;
+}
+
 export interface HelpRequestMessage {
   type: typeof HELP_REQUEST;
   requestId: string;
@@ -26,6 +33,8 @@ export interface HelpRequestMessage {
   displayMode?: "full" | "compact";
   /** CSS selectors to scroll to + flash-highlight (may be empty). */
   selectors: string[];
+  /** Top-viewport rectangles for targets that cannot be represented by a root selector. */
+  rects?: HelpHighlightRect[];
   timeoutMs: number;
 }
 
@@ -74,6 +83,16 @@ export function isHelpRequestMessage(msg: unknown): msg is HelpRequestMessage {
     (m.displayMode === undefined || m.displayMode === "full" || m.displayMode === "compact") &&
     Array.isArray(m.selectors) &&
     m.selectors.every((selector) => typeof selector === "string") &&
+    (m.rects === undefined ||
+      (Array.isArray(m.rects) &&
+        m.rects.every(
+          (rect) =>
+            typeof rect === "object" &&
+            rect !== null &&
+            ["top", "left", "width", "height"].every((key) =>
+              Number.isFinite((rect as Record<string, unknown>)[key]),
+            ),
+        ))) &&
     typeof m.timeoutMs === "number"
   );
 }

apps/extension/src/lib/match-target.ts

@@ -1,103 +0,0 @@
-import type { RenderedRef } from "@browser-skill/vom";
-import type { CapturedNode } from "@/tools/vom/capture";
-import type { CaptureGeometry, TargetDescriptor } from "@/transport/types";
-import type { CaptureTargetDescriptor } from "./describe-target";
-import { GEOM_MATCH_TOLERANCE_PX } from "./record-constants";
-
-export interface MatchTargetInput {
-  geometry: CaptureGeometry;
-  captured: CapturedNode[];
-  refs: RenderedRef[];
-  fallback?: CaptureTargetDescriptor;
-}
-
-function withinTolerance(a: number, b: number): boolean {
-  return Math.abs(a - b) <= GEOM_MATCH_TOLERANCE_PX;
-}
-
-function rectsMatch(
-  a: { x: number; y: number; w: number; h: number },
-  b: { x: number; y: number; w: number; h: number },
-): boolean {
-  return (
-    withinTolerance(a.x, b.x) &&
-    withinTolerance(a.y, b.y) &&
-    withinTolerance(a.w, b.w) &&
-    withinTolerance(a.h, b.h)
-  );
-}
-
-function nodeViewportRect(
-  node: CapturedNode,
-): { x: number; y: number; w: number; h: number } | null {
-  return node.localRect ?? node.rect;
-}
-
-function matchingRects(
-  geometry: CaptureGeometry,
-  node: CapturedNode,
-): {
-  target: { x: number; y: number; w: number; h: number };
-  node: { x: number; y: number; w: number; h: number } | null;
-} {
-  const topFrame = (geometry.ownerFrameBackendNodeId ?? null) === null;
-  const viewportPosition = geometry.position === "fixed" || geometry.position === "sticky";
-  if (!topFrame || viewportPosition || !node.documentRect) {
-    return { target: geometry.rect, node: nodeViewportRect(node) };
-  }
-  return {
-    target: {
-      x: geometry.rect.x + geometry.scrollX,
-      y: geometry.rect.y + geometry.scrollY,
-      w: geometry.rect.w,
-      h: geometry.rect.h,
-    },
-    node: node.documentRect,
-  };
-}
-
-function tagMatches(geometryTag: string, nodeTag: string): boolean {
-  return geometryTag.toLowerCase() === nodeTag.toLowerCase();
-}
-
-/** Best description available when the element cannot be located in the VOM. */
-export function fallbackDescriptor(fallback?: CaptureTargetDescriptor): TargetDescriptor {
-  if (!fallback) {
-    return { unmatched: true };
-  }
-  return {
-    ...(fallback.role ? { role: fallback.role } : {}),
-    ...(fallback.name ? { name: fallback.name } : {}),
-    unmatched: true,
-  };
-}
-
-/** Locate the interacted element in the last settled observation by geometry. */
-export function matchTarget(input: MatchTargetInput): TargetDescriptor {
-  const ownerFrame = input.geometry.ownerFrameBackendNodeId ?? null;
-
-  const candidates = input.captured.filter((node) => {
-    if (ownerFrame !== (node.ownerFrameBackendNodeId ?? null)) return false;
-    if (!tagMatches(input.geometry.tag, node.tag)) return false;
-    const rects = matchingRects(input.geometry, node);
-    if (!rects.node) return false;
-    return rectsMatch(rects.target, rects.node);
-  });
-
-  if (candidates.length !== 1) {
-    return fallbackDescriptor(input.fallback);
-  }
-
-  const backendNodeId = candidates[0]!.backendNodeId;
-  const refEntry = input.refs.find((r) => r.backendNodeId === backendNodeId);
-  if (!refEntry) {
-    return fallbackDescriptor(input.fallback);
-  }
-
-  return {
-    ref: refEntry.ref,
-    ...(refEntry.role ? { role: refEntry.role } : {}),
-    ...(refEntry.name ? { name: refEntry.name } : {}),
-    ...(refEntry.ctx ? { ctx: refEntry.ctx } : {}),
-  };
-}

apps/extension/src/lib/page-settled.ts

@@ -1,92 +0,0 @@
-// Deciding *when* a page is worth observing.
-//
-// A fixed delay cannot serve both a modal that toggles in one frame and a
-// route change that renders for a second. Instead of guessing, ask the page:
-// a document that has stopped mutating and is no longer loading is done
-// reacting to whatever the user just did.
-
-import type { CdpRunner } from "@/tools/shared";
-import { SETTLE_MAX_MS, SETTLE_MIN_MS, SETTLE_POLL_MS, SETTLE_QUIET_MS } from "./record-constants";
-
-/**
- * Installed once per document. Records only the timestamp of the last DOM
- * change so each poll stays O(1) no matter how large the page is.
- */
-const QUIET_PROBE = `(() => {
-  const scope = window;
-  let probe = scope.__bskRecordQuiet;
-  if (!probe) {
-    probe = { changedAt: Date.now() };
-    const observer = new MutationObserver(() => {
-      probe.changedAt = Date.now();
-    });
-    observer.observe(document.documentElement, {
-      subtree: true,
-      childList: true,
-      attributes: true,
-      characterData: true,
-    });
-    scope.__bskRecordQuiet = probe;
-  }
-  return { idleMs: Date.now() - probe.changedAt, readyState: document.readyState };
-})()`;
-
-export type SettleOutcome =
-  /** The page stopped changing on its own. */
-  | "quiet"
-  /** Still changing when the budget ran out; observe it as it is. */
-  | "timeout"
-  /** A newer action took over; this observation is no longer wanted. */
-  | "cancelled";
-
-interface QuietProbe {
-  idleMs: number;
-  readyState: string;
-}
-
-function sleep(ms: number): Promise<void> {
-  return new Promise((resolve) => setTimeout(resolve, ms));
-}
-
-async function readQuietProbe(cdp: CdpRunner, tabId: number): Promise<QuietProbe | null> {
-  try {
-    const reply = await cdp.send<{ result?: { value?: unknown } }>(tabId, "Runtime.evaluate", {
-      expression: QUIET_PROBE,
-      returnByValue: true,
-    });
-    const value = reply.result?.value;
-    if (!value || typeof value !== "object") return null;
-    const { idleMs, readyState } = value as { idleMs?: unknown; readyState?: unknown };
-    if (typeof idleMs !== "number" || typeof readyState !== "string") return null;
-    return { idleMs, readyState };
-  } catch {
-    // An unreadable page is a page mid-swap; that is a reason to keep waiting,
-    // not a reason to give up.
-    return null;
-  }
-}
-
-/** Wait until the page has finished reacting, or until the budget runs out. */
-export async function waitForPageSettled(
-  cdp: CdpRunner,
-  tabId: number,
-  options: { cancelled?: () => boolean } = {},
-): Promise<SettleOutcome> {
-  const startedAt = Date.now();
-  const floor = startedAt + SETTLE_MIN_MS;
-  const deadline = startedAt + SETTLE_MAX_MS;
-
-  for (;;) {
-    if (options.cancelled?.()) return "cancelled";
-    await sleep(SETTLE_POLL_MS);
-    if (options.cancelled?.()) return "cancelled";
-
-    const probe = await readQuietProbe(cdp, tabId);
-    const now = Date.now();
-    if (now < floor) continue;
-    if (now >= deadline) return "timeout";
-    if (!probe) continue;
-    if (probe.readyState === "loading") continue;
-    if (probe.idleMs >= SETTLE_QUIET_MS) return "quiet";
-  }
-}

apps/extension/src/lib/record-bridge.ts

@@ -16,6 +16,10 @@ export interface RecordStartAck {
   ok: true;
 }
 
+export type RecordStepAck =
+  | { ok: true; sequence: number }
+  | { ok: false; expectedSequence: number; error: string };
+
 export type RecordStopAck = { ok: true } | { ok: false; error: string };
 
 export interface RecordQueryMessage {
@@ -41,7 +45,7 @@ export interface RecordStartMessage {
 }
 
 export interface RecordStepPayload {
-  op: "click" | "hover" | "fill" | "press" | "select" | "navigate" | "scroll";
+  op: "click" | "hover" | "fill" | "press" | "select" | "navigate";
   target?: CaptureTargetDescriptor;
   value?: string;
   key?: string;
@@ -53,14 +57,10 @@ export interface RecordStepPayload {
   commit?: "enter" | "suggestion" | "blur";
   /** Page URL when the step was captured. */
   page_url?: string;
-  /** Document geometry at capture time for background geometric matching. */
+  /** Event-target bounds in the source frame's viewport coordinate space. */
   geometry?: {
     rect: { x: number; y: number; w: number; h: number };
-    scrollX: number;
-    scrollY: number;
-    position: string;
     tag: string;
-    ownerFrameBackendNodeId?: number | null;
   };
   /** Capture-only hint; never persisted unless converted to navigated_to. */
   expects_navigation?: boolean;
@@ -74,6 +74,10 @@ export interface RecordStepPayload {
 export interface RecordStepMessage {
   type: typeof RECORD_STEP;
   requestId: string;
+  /** Stable for the lifetime of one content-script document. */
+  producerId: string;
+  /** Monotonic within one content-script producer and recording request. */
+  sequence: number;
   step: RecordStepPayload;
 }
 
@@ -125,8 +129,27 @@ export function isRecordQueryMessage(msg: unknown): msg is RecordQueryMessage {
 export function isRecordStepMessage(msg: unknown): msg is RecordStepMessage {
   if (typeof msg !== "object" || msg === null) return false;
   const m = msg as Record<string, unknown>;
-  if (m.type !== RECORD_STEP || typeof m.requestId !== "string") return false;
+  if (
+    m.type !== RECORD_STEP ||
+    typeof m.requestId !== "string" ||
+    typeof m.producerId !== "string" ||
+    m.producerId.length === 0 ||
+    typeof m.sequence !== "number" ||
+    !Number.isSafeInteger(m.sequence) ||
+    m.sequence < 1
+  ) {
+    return false;
+  }
   const step = m.step;
   if (typeof step !== "object" || step === null) return false;
   return typeof (step as RecordStepPayload).op === "string";
 }
+
+export function isAcceptedRecordStepAck(
+  value: unknown,
+  sequence: number,
+): value is Extract<RecordStepAck, { ok: true }> {
+  if (typeof value !== "object" || value === null) return false;
+  const ack = value as Partial<RecordStepAck>;
+  return ack.ok === true && ack.sequence === sequence;
+}

apps/extension/src/lib/record-constants.ts

@@ -1,97 +0,0 @@
-import type { FillCommit, NavigationCause } from "@/transport/types";
-
-/**
- * Floor before a post-action observation may be taken. A page usually has not
- * started reacting in the first frames after an action, and capturing then
- * would record the page as it was, not as the action left it.
- */
-export const SETTLE_MIN_MS = 150;
-
-/** How long the DOM must stop changing before a page counts as settled. */
-export const SETTLE_QUIET_MS = 250;
-
-/** Upper bound on waiting for a page to settle; animations never stop. */
-export const SETTLE_MAX_MS = 2_000;
-
-/** How often to ask the page whether it has gone quiet. */
-export const SETTLE_POLL_MS = 60;
-
-/** Minimum interval between consecutive observations on the same tab. */
-export const OBSERVATION_MIN_INTERVAL_MS = 200;
-
-/** Delay before retrying a capture that lost its execution context. */
-export const CAPTURE_RETRY_DELAY_MS = 250;
-
-/** Document-coordinate matching tolerance in CSS pixels. */
-export const GEOM_MATCH_TOLERANCE_PX = 2;
-
-/** Default max tokens per page observation when CLI omits `--max-page-tokens`. */
-export const DEFAULT_MAX_PAGE_TOKENS = 3000;
-
-/** VOM observation file format version (front matter header). */
-export const OBSERVATION_FILE_VERSION = 1;
-
-/** FNV-1a 64-bit offset basis. */
-const FNV_OFFSET = 0xcbf29ce484222325n;
-/** FNV-1a 64-bit prime. */
-const FNV_PRIME = 0x100000001b3n;
-
-/** Content-hash for state deduplication (sync, non-cryptographic). */
-export function fnv1a64(text: string): string {
-  let hash = FNV_OFFSET;
-  for (let i = 0; i < text.length; i += 1) {
-    hash ^= BigInt(text.charCodeAt(i));
-    hash = (hash * FNV_PRIME) & 0xffffffffffffffffn;
-  }
-  return hash.toString(16).padStart(16, "0");
-}
-
-let nextStateSerial = 0;
-
-/** Monotonic state id generator (`s1`, `s2`, …). */
-export function nextStateId(): string {
-  nextStateSerial += 1;
-  return `s${nextStateSerial}`;
-}
-
-/** Test seam: reset the state id counter. */
-export function resetStateIdCounterForTests(): void {
-  nextStateSerial = 0;
-}
-
-const REDIRECT_QUALIFIERS = new Set(["client_redirect", "server_redirect"]);
-
-const TRANSITION_TO_CAUSE: Record<string, NavigationCause> = {
-  typed: "user_typed",
-  generated: "user_typed",
-  keyword: "user_typed",
-  keyword_generated: "user_typed",
-  link: "link",
-  form_submit: "form_submit",
-  reload: "reload",
-  auto_bookmark: "browser",
-  start_page: "browser",
-};
-
-export interface NavigationTransitionMeta {
-  transitionType?: string;
-  transitionQualifiers?: string[];
-  navigationActionPending?: boolean;
-}
-
-/** Map webNavigation metadata to protocol `NavigationCause`. Returns null for redirects. */
-export function mapNavigationCause(meta: NavigationTransitionMeta): NavigationCause | null {
-  const qualifiers = meta.transitionQualifiers ?? [];
-  if (qualifiers.includes("forward_back")) return "history";
-  if (qualifiers.some((q) => REDIRECT_QUALIFIERS.has(q))) return null;
-  if (qualifiers.includes("from_address_bar")) return "user_typed";
-
-  const type = meta.transitionType ?? "";
-  const mapped = TRANSITION_TO_CAUSE[type];
-  if (mapped) return mapped;
-
-  if (!type && meta.navigationActionPending === false) return "script";
-  return "browser";
-}
-
-export const DEFAULT_FILL_COMMIT: FillCommit = "blur";

apps/extension/src/lib/record-observation.ts

@@ -1,663 +0,0 @@
-import type { RenderedRef } from "@browser-skill/vom";
-import { formatObservationFile, type ObservationAnnotation } from "@/lib/format-observation-file";
-import { fallbackDescriptor, matchTarget } from "@/lib/match-target";
-import { waitForPageSettled } from "@/lib/page-settled";
-import {
-  CAPTURE_RETRY_DELAY_MS,
-  DEFAULT_MAX_PAGE_TOKENS,
-  OBSERVATION_MIN_INTERVAL_MS,
-} from "@/lib/record-constants";
-import { registerObservation, type StateRegistryEntry } from "@/lib/trace-reducer";
-import { captureVomObservation } from "@/tools/capture-vom-observation";
-import type { CdpRunner, ChromeTabsApi } from "@/tools/shared";
-import type { CapturedNode } from "@/tools/vom/capture";
-import type { DraftTraceStep, Step, StopReason, Trace, TraceState } from "@/transport/types";
-import { reduceTraceSteps, resolveTraceStartUrl } from "./trace-reducer";
-
-export interface LastSettledObservation {
-  stateId: string;
-  captured: CapturedNode[];
-  refs: RenderedRef[];
-  url: string;
-  title?: string;
-  vomText: string;
-}
-
-/** A post-action observation that has been queued but not yet written down. */
-interface PendingSettle {
-  draftIndex: number;
-  /** Set when a newer action has already decided where this step landed. */
-  cancelled: boolean;
-}
-
-/** Latest URL in an in-flight redirect chain awaiting coalesce + settle. */
-export interface PendingRedirectLanding {
-  url: string;
-  /** Bumped on every hop so an in-flight `waitForPageSettled` can cancel. */
-  generation: number;
-}
-
-export interface RecordingObservationState {
-  stateRegistry: Map<string, StateRegistryEntry>;
-  lastSettled: LastSettledObservation | null;
-  maxPageTokens: number;
-  redactValues: boolean;
-  lastCaptureAtMs: number;
-  /**
-   * Tail of the settle chain. Observations run one at a time so that a slow
-   * capture can never land after a faster one and report the pages out of the
-   * order the user visited them.
-   */
-  settleQueue: Promise<void>;
-  /** Queued and in-flight settles, by draft index, so newer actions can supersede them. */
-  settles: Map<number, PendingSettle>;
-  stepAnnotations: Map<number, ObservationAnnotation[]>;
-  /**
-   * OAuth / server redirects are coalesced here: intermediate hops only update
-   * `url`+`generation`, and one `navigate` is emitted after the page settles.
-   */
-  pendingRedirect: PendingRedirectLanding | null;
-  redirectFlushQueue: Promise<void>;
-}
-
-export function createObservationState(options?: {
-  maxPageTokens?: number;
-  redactValues?: boolean;
-}): RecordingObservationState {
-  return {
-    stateRegistry: new Map(),
-    lastSettled: null,
-    maxPageTokens: options?.maxPageTokens ?? DEFAULT_MAX_PAGE_TOKENS,
-    redactValues: options?.redactValues ?? false,
-    lastCaptureAtMs: 0,
-    settleQueue: Promise.resolve(),
-    settles: new Map(),
-    stepAnnotations: new Map(),
-    pendingRedirect: null,
-    redirectFlushQueue: Promise.resolve(),
-  };
-}
-
-function sleep(ms: number): Promise<void> {
-  return new Promise((resolve) => setTimeout(resolve, ms));
-}
-
-async function readTabMeta(
-  tabsApi: ChromeTabsApi,
-  tabId: number,
-): Promise<{ url: string; title?: string }> {
-  try {
-    const tab = await tabsApi.get(tabId);
-    return { url: tab.url ?? "about:blank", title: tab.title };
-  } catch {
-    return { url: "about:blank" };
-  }
-}
-
-export async function captureAndRegisterObservation(
-  obs: RecordingObservationState,
-  cdp: CdpRunner,
-  tabId: number,
-  tabsApi: ChromeTabsApi,
-  urlOverride?: string,
-): Promise<LastSettledObservation> {
-  const now = Date.now();
-  const waitMs = Math.max(0, OBSERVATION_MIN_INTERVAL_MS - (now - obs.lastCaptureAtMs));
-  if (waitMs > 0) await sleep(waitMs);
-
-  const { url, title } = urlOverride
-    ? { url: urlOverride, title: undefined }
-    : await readTabMeta(tabsApi, tabId);
-  const meta = urlOverride ? await readTabMeta(tabsApi, tabId) : { url, title };
-  const resolvedTitle = title ?? meta.title;
-
-  const rendered = await captureVomObservation(cdp, tabId, url, {
-    maxTokens: obs.maxPageTokens,
-    redactValues: obs.redactValues,
-    conditionalSurfaceProbe: false,
-  });
-
-  const stateId = registerObservation(obs.stateRegistry, {
-    url,
-    title: resolvedTitle,
-    rawVomText: rendered.text,
-    truncated: rendered.truncated,
-  });
-
-  obs.lastCaptureAtMs = Date.now();
-  const settled: LastSettledObservation = {
-    stateId,
-    captured: rendered.captured,
-    refs: rendered.refs,
-    url,
-    title: resolvedTitle,
-    vomText: rendered.text,
-  };
-  obs.lastSettled = settled;
-  return settled;
-}
-
-/**
- * Bind a draft to the page it was performed on: origin state, geometric target
- * match, and the inline annotation. All three read the *current* observation,
- * which is only the page the user acted on while the draft is still fresh —
- * once the action settles, `lastSettled` is the destination page instead.
- */
-export function applyTargetMatching(
-  obs: RecordingObservationState,
-  draft: DraftTraceStep,
-  stepId: number,
-): void {
-  if (!obs.lastSettled) {
-    // Initial observation can fail or race the first action. Keep whatever
-    // the content script captured so the exported step is still teachable.
-    if ("target" in draft && "captureTarget" in draft) {
-      draft.target = fallbackDescriptor(draft.captureTarget);
-    }
-    rememberStepAnnotation(obs, stepId, draft);
-    return;
-  }
-  draft.preStateId = obs.lastSettled.stateId;
-
-  if ("target" in draft) {
-    draft.target =
-      "geometry" in draft && draft.geometry
-        ? matchTarget({
-            geometry: draft.geometry,
-            captured: obs.lastSettled.captured,
-            refs: obs.lastSettled.refs,
-            fallback: draft.captureTarget,
-          })
-        : // No geometry to match on, but the capture still knew what the user
-          // touched — keep that instead of an anonymous unmatched target.
-          fallbackDescriptor(draft.captureTarget);
-  }
-  rememberStepAnnotation(obs, stepId, draft);
-}
-
-function fillDetailForDraft(
-  obs: RecordingObservationState,
-  draft: DraftTraceStep,
-): string | undefined {
-  if (obs.redactValues) return undefined;
-  if (draft.op === "fill") return JSON.stringify(draft.value);
-  return undefined;
-}
-
-function refLineForDraft(
-  obs: RecordingObservationState,
-  draft: DraftTraceStep,
-): number | undefined {
-  if (!("target" in draft) || !obs.lastSettled) return undefined;
-  const targetRef = draft.target?.ref;
-  if (!targetRef) return undefined;
-  return obs.lastSettled.refs.find((r) => r.ref === targetRef)?.line;
-}
-
-function rememberStepAnnotation(
-  obs: RecordingObservationState,
-  stepId: number,
-  draft: DraftTraceStep,
-): void {
-  const line = refLineForDraft(obs, draft);
-  if (line === undefined || draft.op === "navigate" || draft.op === "scroll") return;
-  if (!draft.preStateId) return;
-  const ann: ObservationAnnotation = {
-    stepId,
-    op: draft.op,
-    line,
-    stateId: draft.preStateId,
-    detail: fillDetailForDraft(obs, draft),
-  };
-  const bucket = obs.stepAnnotations.get(line) ?? [];
-  bucket.push(ann);
-  obs.stepAnnotations.set(line, bucket);
-}
-
-/** Record the step under the page it was performed on, not the one it led to. */
-export function rememberStepOnPage(
-  obs: RecordingObservationState,
-  draft: DraftTraceStep,
-  stepId: number,
-): void {
-  const stateId = draft.preStateId;
-  if (!stateId) return;
-  const entry = obs.stateRegistry.get(stateId);
-  if (entry && !entry.stepsHere.includes(stepId)) {
-    entry.stepsHere.push(stepId);
-  }
-}
-
-/** Drop an in-flight redirect coalesce (e.g. a real navigate superseded it). */
-export function clearPendingRedirectLanding(obs: RecordingObservationState): void {
-  obs.pendingRedirect = null;
-}
-
-function enqueueRedirectFlush(obs: RecordingObservationState, task: () => Promise<void>): void {
-  obs.redirectFlushQueue = obs.redirectFlushQueue.then(task, task).catch(() => {});
-}
-
-/**
- * Remember a redirect hop and schedule a settle-then-emit of a single navigate
- * to the final URL. Later hops only bump `generation` / replace `url`.
- */
-export function scheduleRedirectLandingFlush(
-  obs: RecordingObservationState,
-  steps: DraftTraceStep[],
-  cdp: CdpRunner,
-  tabId: number,
-  tabsApi: ChromeTabsApi,
-  url: string,
-  options: { cancelled?: () => boolean } = {},
-): void {
-  const generation = (obs.pendingRedirect?.generation ?? 0) + 1;
-  obs.pendingRedirect = { url, generation };
-  enqueueRedirectFlush(obs, () =>
-    coalesceRedirectLanding(obs, steps, cdp, tabId, tabsApi, options),
-  );
-}
-
-/**
- * Drain any coalesced redirect so the next action matches against the real
- * landing page. Safe to call when nothing is pending.
- */
-export async function flushPendingRedirectLanding(
-  obs: RecordingObservationState,
-  steps: DraftTraceStep[],
-  cdp: CdpRunner | undefined,
-  tabId: number,
-  tabsApi: ChromeTabsApi,
-  options: { cancelled?: () => boolean } = {},
-): Promise<void> {
-  if (obs.pendingRedirect && cdp) {
-    enqueueRedirectFlush(obs, () =>
-      coalesceRedirectLanding(obs, steps, cdp, tabId, tabsApi, options),
-    );
-  }
-  await obs.redirectFlushQueue;
-}
-
-/**
- * Wait until the redirect chain stops changing the document, then emit one
- * `navigate` (cause `browser`) to the tab's final URL and settle its observation.
- */
-async function coalesceRedirectLanding(
-  obs: RecordingObservationState,
-  steps: DraftTraceStep[],
-  cdp: CdpRunner,
-  tabId: number,
-  tabsApi: ChromeTabsApi,
-  options: { cancelled?: () => boolean },
-): Promise<void> {
-  while (obs.pendingRedirect) {
-    if (options.cancelled?.()) {
-      obs.pendingRedirect = null;
-      return;
-    }
-
-    const snap = obs.pendingRedirect;
-    const outcome = await waitForPageSettled(cdp, tabId, {
-      cancelled: () =>
-        !!options.cancelled?.() ||
-        obs.pendingRedirect === null ||
-        obs.pendingRedirect.generation !== snap.generation,
-    });
-
-    if (options.cancelled?.()) {
-      obs.pendingRedirect = null;
-      return;
-    }
-    // A newer hop cancelled this wait — loop and settle against the latest URL.
-    if (outcome === "cancelled") continue;
-    if (!obs.pendingRedirect || obs.pendingRedirect.generation !== snap.generation) {
-      continue;
-    }
-
-    const finalUrl = (await readTabMeta(tabsApi, tabId)).url || snap.url;
-    obs.pendingRedirect = null;
-
-    if (!finalUrl || finalUrl === "about:blank") return;
-    if (obs.lastSettled?.url === finalUrl) return;
-
-    const last = steps[steps.length - 1];
-    if (last?.op === "navigate" && last.url === finalUrl) {
-      if (!last.postStateId) {
-        const draftIndex = steps.length - 1;
-        scheduleDraftSettle(obs, draftIndex, draftIndex + 1, cdp, tabId, tabsApi, steps);
-        await obs.settleQueue;
-      }
-      return;
-    }
-
-    if (outcome === "timeout") {
-      console.debug(
-        `[bsk record] redirect landing still changing after settle budget; ` +
-          `recording navigate to ${finalUrl}`,
-      );
-    }
-
-    const draft: DraftTraceStep = {
-      op: "navigate",
-      url: finalUrl,
-      page_url: finalUrl,
-      cause: "browser",
-      preStateId: obs.lastSettled?.stateId,
-    };
-    steps.push(draft);
-    const draftIndex = steps.length - 1;
-    const stepId = draftIndex + 1;
-    rememberStepOnPage(obs, draft, stepId);
-    scheduleDraftSettle(obs, draftIndex, stepId, cdp, tabId, tabsApi, steps);
-    // Callers that flush before the next action need `lastSettled` to already
-    // be the landing page so target matching does not use the pre-redirect view.
-    await obs.settleQueue;
-    return;
-  }
-}
-
-/**
- * A capture that lands mid-navigation fails on a destroyed execution context.
- * That is transient, so give it one more chance before the step has to fall
- * back to a stale state.
- */
-async function captureWithRetry(
-  obs: RecordingObservationState,
-  cdp: CdpRunner,
-  tabId: number,
-  tabsApi: ChromeTabsApi,
-  urlOverride?: string,
-): Promise<LastSettledObservation> {
-  try {
-    return await captureAndRegisterObservation(obs, cdp, tabId, tabsApi, urlOverride);
-  } catch (err) {
-    console.debug("[bsk record] observation failed, retrying once", err);
-    await sleep(CAPTURE_RETRY_DELAY_MS);
-    return captureAndRegisterObservation(obs, cdp, tabId, tabsApi, urlOverride);
-  }
-}
-
-export async function settleDraftObservation(
-  obs: RecordingObservationState,
-  cdp: CdpRunner,
-  tabId: number,
-  tabsApi: ChromeTabsApi,
-  steps: DraftTraceStep[],
-  draftIndex: number,
-  pending?: PendingSettle,
-): Promise<void> {
-  const cancelled = () => pending?.cancelled === true;
-  const outcome = await waitForPageSettled(cdp, tabId, { cancelled });
-  if (outcome === "cancelled") return;
-  if (outcome === "timeout") {
-    console.debug(
-      `[bsk record] step ${draftIndex + 1} was still changing the page after the settle ` +
-        "budget; observing it as it stands",
-    );
-  }
-
-  const started = steps[draftIndex];
-  if (!started) return;
-
-  // The URL attached to the draft may only be an intermediate redirect hop.
-  // Once the page is settled, register the capture against live tab metadata.
-  const settled = await captureWithRetry(obs, cdp, tabId, tabsApi);
-  // The observation still counts as the latest view of the page even when this
-  // step no longer wants it — the next action will start from it.
-  if (cancelled()) return;
-  // Re-read the slot: a navigation observed while the capture ran may have
-  // rewritten this draft, and the observation belongs to whatever occupies the
-  // slot now — writing to the object we started with would strand it.
-  const draft = steps[draftIndex] ?? started;
-  draft.postStateId = settled.stateId;
-}
-
-/**
- * Recover landing states from the shape of the recording itself: wherever the
- * next action was performed is, by definition, where the previous one landed.
- * Only the immediately following draft counts — a later one would claim a page
- * that several unobserved actions away, which is worse than admitting we saw
- * no change.
- */
-export function inferMissingPostStates(steps: DraftTraceStep[]): void {
-  for (let i = 0; i < steps.length - 1; i += 1) {
-    const draft = steps[i];
-    const next = steps[i + 1];
-    if (!draft || !next || draft.postStateId || !next.preStateId) continue;
-    draft.postStateId = next.preStateId;
-    console.debug(
-      `[bsk record] step ${i + 1} (${draft.op}) had no post-action observation; ` +
-        `using where step ${i + 2} started (${next.preStateId})`,
-    );
-  }
-}
-
-/**
- * Last chance for the closing actions of a recording. Stopping right after the
- * final action is normal, and so is that action navigating away, so take one
- * final look at the page. Only trailing drafts may claim it: an earlier step
- * did not land on whatever the user happens to be looking at when they stop.
- */
-export async function settleUnsettledDrafts(
-  obs: RecordingObservationState,
-  cdp: CdpRunner,
-  tabId: number,
-  tabsApi: ChromeTabsApi,
-  steps: DraftTraceStep[],
-): Promise<void> {
-  const trailing: DraftTraceStep[] = [];
-  for (let i = steps.length - 1; i >= 0; i -= 1) {
-    const draft = steps[i];
-    if (!draft || draft.postStateId) break;
-    trailing.push(draft);
-  }
-  if (trailing.length === 0) return;
-
-  let settled: LastSettledObservation;
-  try {
-    settled = await captureWithRetry(obs, cdp, tabId, tabsApi);
-  } catch (err) {
-    console.warn("[bsk record] final observation at stop failed", err);
-    return;
-  }
-  for (const draft of trailing) draft.postStateId = settled.stateId;
-  console.debug(
-    `[bsk record] settled ${trailing.length} trailing step(s) against the page at stop ` +
-      `(${settled.stateId} ${settled.url})`,
-  );
-}
-
-/**
- * Performing a new action ends the previous one: the page the user reached for
- * is, by definition, where the earlier step left them. Taking the landing from
- * the newer step also keeps the trace monotonic — a capture that finished late
- * would otherwise credit an earlier step with a page that only exists because
- * of the newer action.
- */
-function supersedeEarlierSettles(
-  obs: RecordingObservationState,
-  draftIndex: number,
-  steps: DraftTraceStep[],
-): void {
-  const landing = steps[draftIndex]?.preStateId;
-  for (const [index, pending] of obs.settles) {
-    if (index >= draftIndex) continue;
-    pending.cancelled = true;
-    obs.settles.delete(index);
-
-    const draft = steps[index];
-    if (!draft || draft.postStateId) continue;
-    if (!landing) continue;
-    draft.postStateId = landing;
-    console.debug(
-      `[bsk record] step ${index + 1} (${draft.op}) was still settling when step ` +
-        `${draftIndex + 1} started; landing it on ${landing}`,
-    );
-  }
-}
-
-export function scheduleDraftSettle(
-  obs: RecordingObservationState,
-  draftIndex: number,
-  stepId: number,
-  cdp: CdpRunner,
-  tabId: number,
-  tabsApi: ChromeTabsApi,
-  steps: DraftTraceStep[],
-): void {
-  supersedeEarlierSettles(obs, draftIndex, steps);
-
-  // Rescheduling the same step (a navigation showed up after the action)
-  // replaces the pending observation rather than racing it.
-  const superseded = obs.settles.get(draftIndex);
-  if (superseded) superseded.cancelled = true;
-
-  const pending: PendingSettle = { draftIndex, cancelled: false };
-  obs.settles.set(draftIndex, pending);
-
-  enqueueSettle(obs, async () => {
-    if (pending.cancelled) return;
-    try {
-      await settleDraftObservation(obs, cdp, tabId, tabsApi, steps, draftIndex, pending);
-    } catch (err) {
-      // Recoverable: stop-time repair still gives the step a landing page.
-      const op = steps[draftIndex]?.op ?? "?";
-      console.warn(`[bsk record] post-action observation failed for step ${stepId} (${op})`, err);
-    } finally {
-      if (obs.settles.get(draftIndex) === pending) obs.settles.delete(draftIndex);
-    }
-  });
-}
-
-function enqueueSettle(obs: RecordingObservationState, task: () => Promise<void>): void {
-  obs.settleQueue = obs.settleQueue.then(task, task).catch(() => {});
-}
-
-export function cancelPendingSettles(obs: RecordingObservationState): void {
-  for (const pending of obs.settles.values()) pending.cancelled = true;
-  obs.settles.clear();
-}
-
-/** Bound the drain loop so a self-rescheduling settle cannot block stop. */
-const MAX_SETTLE_FLUSH_ROUNDS = 10;
-
-/**
- * Drain settle work before exporting, including work queued while draining —
- * a navigation observed during the last capture schedules another one, and
- * the trace would drop it if stop did not wait.
- */
-export async function flushPendingSettles(obs: RecordingObservationState): Promise<void> {
-  for (let round = 0; round < MAX_SETTLE_FLUSH_ROUNDS; round += 1) {
-    const drained = obs.settleQueue;
-    await drained;
-    if (obs.settleQueue === drained) return;
-  }
-  console.warn("[bsk record] settle queue kept growing at stop; exporting what has been observed");
-}
-
-function finalizeStateBodies(
-  entries: StateRegistryEntry[],
-  annotationsByState: Map<string, ObservationAnnotation[]>,
-  stepIdByDraftId: Map<number, number>,
-  idByOldId: Map<string, string>,
-): TraceState[] {
-  return entries.map((entry) => {
-    const id = idByOldId.get(entry.id) ?? entry.id;
-    const body = formatObservationFile({
-      stateId: id,
-      url: entry.url,
-      title: entry.title,
-      stepsHere: remapStepIds(entry.stepsHere, stepIdByDraftId),
-      body: entry.rawVomText,
-      annotations: annotationsByState.get(entry.id) ?? [],
-    });
-    return {
-      id,
-      url: entry.url,
-      ...(entry.title ? { title: entry.title } : {}),
-      body,
-      ...(entry.truncated ? { truncated: true } : {}),
-    };
-  });
-}
-
-/** Draft ids only become step ids after collapsing and filtering. */
-function remapStepIds(draftIds: number[], stepIdByDraftId: Map<number, number>): number[] {
-  const mapped = new Set<number>();
-  for (const draftId of draftIds) {
-    const stepId = stepIdByDraftId.get(draftId);
-    if (stepId !== undefined) mapped.add(stepId);
-  }
-  return [...mapped].sort((a, b) => a - b);
-}
-
-function collectAnnotationsByState(
-  obs: RecordingObservationState,
-  stepIdByDraftId: Map<number, number>,
-): Map<string, ObservationAnnotation[]> {
-  const byState = new Map<string, ObservationAnnotation[]>();
-  for (const anns of obs.stepAnnotations.values()) {
-    for (const ann of anns) {
-      const stepId = stepIdByDraftId.get(ann.stepId);
-      if (stepId === undefined) continue;
-      const bucket = byState.get(ann.stateId) ?? [];
-      bucket.push({ ...ann, stepId });
-      byState.set(ann.stateId, bucket);
-    }
-  }
-  return byState;
-}
-
-/**
- * Keep only the pages the published steps point at. Redirect hops and
- * mid-load captures land in the registry too, and shipping them would invite
- * a reader to treat a page the flow merely passed through as a real stop.
- * With no steps at all the first observation is the whole artifact, so it
- * stays.
- */
-function selectPublishedStates(
-  registry: Map<string, StateRegistryEntry>,
-  steps: Step[],
-): StateRegistryEntry[] {
-  const entries = [...registry.values()];
-  if (steps.length === 0) return entries.slice(0, 1);
-  const referenced = new Set<string>();
-  for (const step of steps) {
-    referenced.add(step.state);
-    referenced.add(step.result.state);
-  }
-  return entries.filter((entry) => referenced.has(entry.id));
-}
-
-export function buildTraceV3(input: {
-  obs: RecordingObservationState;
-  steps: DraftTraceStep[];
-  startedAt: string;
-  purpose?: string;
-  startUrl?: string;
-  stoppedBy: StopReason;
-  bskVersion: string;
-}): Trace {
-  const { steps, stepIdByDraftId } = reduceTraceSteps(input.steps, input.obs.stateRegistry);
-  const published = selectPublishedStates(input.obs.stateRegistry, steps);
-  // Renumber so the shipped dictionary reads s1..sN without holes where the
-  // dropped captures used to be.
-  const idByOldId = new Map(published.map((entry, index) => [entry.id, `s${index + 1}`]));
-  for (const step of steps) {
-    step.state = idByOldId.get(step.state) ?? step.state;
-    step.result.state = idByOldId.get(step.result.state) ?? step.result.state;
-  }
-  const annotationsByState = collectAnnotationsByState(input.obs, stepIdByDraftId);
-  const states = finalizeStateBodies(published, annotationsByState, stepIdByDraftId, idByOldId);
-  const startUrl = resolveTraceStartUrl(input.steps, input.startUrl, states);
-  return {
-    version: 3,
-    ...(input.purpose ? { purpose: input.purpose } : {}),
-    recorded_at: new Date().toISOString(),
-    started_at: input.startedAt,
-    stopped_by: input.stoppedBy,
-    entry: { start_url: startUrl },
-    recorder: { bsk: input.bskVersion, vom: 1 },
-    states,
-    steps,
-  };
-}

apps/extension/src/lib/recording-step-buffer.ts

@@ -1,212 +0,0 @@
-import type { DraftTraceStep, NavigationCause } from "@/transport/types";
-import type { RecordStepPayload } from "./record-bridge";
-import { mapNavigationCause } from "./record-constants";
-
-export interface RecordingStepBuffer {
-  steps: DraftTraceStep[];
-  currentUrl?: string;
-  pendingNavigation: boolean;
-  pendingNavigationDeadline?: number;
-  lastTransitionType?: string;
-  lastTransitionQualifiers?: string[];
-}
-
-const NAVIGATION_TRIGGER_WINDOW_MS = 3_000;
-
-function toDraftStep(payload: RecordStepPayload): DraftTraceStep | null {
-  const pageUrl = payload.page_url;
-  const geometry = payload.geometry;
-  switch (payload.op) {
-    case "click":
-      return payload.target
-        ? {
-            op: "click",
-            target: { unmatched: true },
-            captureTarget: payload.target,
-            ...(geometry ? { geometry } : {}),
-            ...(pageUrl ? { page_url: pageUrl } : {}),
-          }
-        : null;
-    case "hover":
-      return payload.target
-        ? {
-            op: "hover",
-            target: { unmatched: true },
-            captureTarget: payload.target,
-            ...(geometry ? { geometry } : {}),
-            ...(pageUrl ? { page_url: pageUrl } : {}),
-          }
-        : null;
-    case "fill":
-      return payload.target
-        ? {
-            op: "fill",
-            target: { unmatched: true },
-            captureTarget: payload.target,
-            ...(geometry ? { geometry } : {}),
-            value: payload.value ?? "",
-            ...(payload.commit ? { commit: payload.commit } : {}),
-            ...(payload.redacted ? { redacted: true } : {}),
-            ...(pageUrl ? { page_url: pageUrl } : {}),
-          }
-        : null;
-    case "press":
-      return payload.key
-        ? {
-            op: "press",
-            key: payload.key,
-            ...(payload.target
-              ? { target: { unmatched: true }, captureTarget: payload.target }
-              : {}),
-            ...(geometry ? { geometry } : {}),
-            ...(payload.modifiers?.length ? { modifiers: payload.modifiers } : {}),
-            ...(pageUrl ? { page_url: pageUrl } : {}),
-          }
-        : null;
-    case "select":
-      return payload.target && payload.values
-        ? {
-            op: "select",
-            target: { unmatched: true },
-            captureTarget: payload.target,
-            ...(geometry ? { geometry } : {}),
-            values: payload.values,
-            ...(payload.labels?.length ? { labels: payload.labels } : {}),
-            ...(pageUrl ? { page_url: pageUrl } : {}),
-          }
-        : null;
-    case "scroll":
-      return { op: "scroll", ...(pageUrl ? { page_url: pageUrl } : {}) };
-    case "navigate":
-      return null;
-  }
-}
-
-/** Annotate the latest action with `navigated_to`. Returns its index, or -1. */
-function annotateLastStepNavigation(buffer: RecordingStepBuffer, url: string): number {
-  for (let i = buffer.steps.length - 1; i >= 0; i -= 1) {
-    const step = buffer.steps[i];
-    if (!step) continue;
-    if (step.op === "click" || step.op === "press" || step.op === "select" || step.op === "fill") {
-      // Mutate in place: an in-flight settle holds this object, and replacing
-      // it would strand that settle's observation on the discarded copy.
-      step.navigated_to = url;
-      return i;
-    }
-    break;
-  }
-  return -1;
-}
-
-function causeFromBuffer(buffer: RecordingStepBuffer): NavigationCause | null {
-  return mapNavigationCause({
-    transitionType: buffer.lastTransitionType,
-    transitionQualifiers: buffer.lastTransitionQualifiers,
-  });
-}
-
-/** Outcome of folding a main-frame URL change into the draft buffer. */
-export type NavigationObserveResult =
-  | { kind: "noop" }
-  | { kind: "annotated"; index: number }
-  | { kind: "appended"; index: number }
-  /**
-   * Redirect hop: do not emit a step yet. The recorder coalesces the chain and
-   * emits one `navigate` after `waitForPageSettled` on the final URL.
-   */
-  | { kind: "coalesce_redirect"; url: string };
-
-export function observeRecordedNavigation(
-  buffer: RecordingStepBuffer,
-  url: string,
-  causedByAction?: boolean,
-  transitionType?: string,
-  transitionQualifiers?: string[],
-): NavigationObserveResult {
-  if (!url || url === buffer.currentUrl) return { kind: "noop" };
-  buffer.currentUrl = url;
-  if (transitionType !== undefined) buffer.lastTransitionType = transitionType;
-  if (transitionQualifiers !== undefined) {
-    buffer.lastTransitionQualifiers = transitionQualifiers;
-  }
-
-  const pendingIsCurrent =
-    buffer.pendingNavigation &&
-    (buffer.pendingNavigationDeadline === undefined ||
-      buffer.pendingNavigationDeadline >= Date.now());
-
-  if (causedByAction === true || (causedByAction === undefined && pendingIsCurrent)) {
-    buffer.pendingNavigation = false;
-    buffer.pendingNavigationDeadline = undefined;
-    const annotatedIndex = annotateLastStepNavigation(buffer, url);
-    if (annotatedIndex >= 0) {
-      return { kind: "annotated", index: annotatedIndex };
-    }
-    const cause = causeFromBuffer(buffer);
-    if (cause === null) {
-      return { kind: "coalesce_redirect", url };
-    }
-    buffer.steps.push({
-      op: "navigate",
-      url,
-      page_url: url,
-      cause,
-      transitionType: buffer.lastTransitionType,
-      transitionQualifiers: buffer.lastTransitionQualifiers,
-    });
-    return { kind: "appended", index: buffer.steps.length - 1 };
-  }
-
-  buffer.pendingNavigation = false;
-  buffer.pendingNavigationDeadline = undefined;
-  const cause = causeFromBuffer(buffer);
-  if (cause === null) {
-    return { kind: "coalesce_redirect", url };
-  }
-  buffer.steps.push({
-    op: "navigate",
-    url,
-    page_url: url,
-    cause,
-    transitionType: buffer.lastTransitionType,
-    transitionQualifiers: buffer.lastTransitionQualifiers,
-  });
-  return { kind: "appended", index: buffer.steps.length - 1 };
-}
-
-/**
- * Buffer one captured payload. Returns the index of the draft it appended, or
- * `null` when the payload produced none (an unnamed click target, a navigation
- * that only annotated the previous draft, …). Callers must not fall back to
- * "the last draft" on `null`: that would re-match an already recorded step
- * against a newer observation and overwrite its origin state.
- */
-export function appendRecordedPayload(
-  buffer: RecordingStepBuffer,
-  payload: RecordStepPayload,
-): number | null {
-  if (payload.op === "navigate") {
-    if (!payload.url) return null;
-    const result = observeRecordedNavigation(
-      buffer,
-      payload.url,
-      payload.navigation_caused_by_action,
-      payload.transitionType,
-      payload.transitionQualifiers,
-    );
-    return result.kind === "appended" ? result.index : null;
-  }
-  const step = toDraftStep({
-    ...payload,
-    page_url: payload.page_url ?? buffer.currentUrl,
-  });
-  if (!step) return null;
-  buffer.steps.push(step);
-  if (step.op === "click" || step.op === "press" || step.op === "select" || step.op === "fill") {
-    buffer.pendingNavigation = payload.expects_navigation === true;
-    buffer.pendingNavigationDeadline = buffer.pendingNavigation
-      ? Date.now() + NAVIGATION_TRIGGER_WINDOW_MS
-      : undefined;
-  }
-  return buffer.steps.length - 1;
-}

apps/extension/src/lib/recording/document-marker.ts

@@ -0,0 +1,39 @@
+import {
+  RECORD_DOCUMENT_ATTRIBUTE,
+  recordingDocumentMarkerValue,
+} from "@/shared/recording-document-identity";
+
+export interface RecordingDocumentMarker {
+  restore(): void;
+  ensure(): void;
+}
+
+export function waitForDocumentElement(): Promise<HTMLElement> {
+  if (document.documentElement) return Promise.resolve(document.documentElement);
+  return new Promise((resolve) => {
+    const observer = new MutationObserver(() => {
+      if (!document.documentElement) return;
+      observer.disconnect();
+      resolve(document.documentElement);
+    });
+    observer.observe(document, { childList: true });
+  });
+}
+
+export function markRecordingDocument(
+  producerId: string,
+  root: HTMLElement = document.documentElement,
+): RecordingDocumentMarker {
+  const previous = root.getAttribute(RECORD_DOCUMENT_ATTRIBUTE);
+  const markerValue = recordingDocumentMarkerValue(producerId);
+  const ensure = () => root.setAttribute(RECORD_DOCUMENT_ATTRIBUTE, markerValue);
+  ensure();
+  return {
+    ensure,
+    restore() {
+      if (root.getAttribute(RECORD_DOCUMENT_ATTRIBUTE) !== markerValue) return;
+      if (previous === null) root.removeAttribute(RECORD_DOCUMENT_ATTRIBUTE);
+      else root.setAttribute(RECORD_DOCUMENT_ATTRIBUTE, previous);
+    },
+  };
+}

apps/extension/src/lib/recording/document-settle.ts

@@ -0,0 +1,134 @@
+import type { CdpTarget } from "@/browser-driver/frame-graph";
+import { type CdpRunner, sendToCdpTarget } from "@/tools/shared";
+
+const SETTLE_MIN_MS = 150;
+const SETTLE_QUIET_MS = 250;
+const SETTLE_MAX_MS = 2_000;
+const SETTLE_POLL_MS = 60;
+const SETTLE_WORLD_NAME = "__bsk_record_settle__";
+
+const QUIET_PROBE = `(() => {
+  const scope = window;
+  let probe = scope.__bskRecordQuiet;
+  if (!probe) {
+    probe = { changedAt: Date.now() };
+    const observer = new MutationObserver(() => {
+      probe.changedAt = Date.now();
+    });
+    observer.observe(document.documentElement, {
+      subtree: true,
+      childList: true,
+      attributes: true,
+      characterData: true,
+    });
+    scope.__bskRecordQuiet = probe;
+  }
+  return { idleMs: Date.now() - probe.changedAt, readyState: document.readyState };
+})()`;
+
+export interface DocumentSettleScope {
+  target: CdpTarget;
+  /** Omit only for the target's root document. */
+  frameId?: string;
+}
+
+type SettleOutcome = "quiet" | "timeout" | "cancelled";
+
+interface QuietProbe {
+  idleMs: number;
+  readyState: string;
+}
+
+interface ProbeContext {
+  executionContextId?: number;
+}
+
+function sleep(ms: number, signal?: AbortSignal): Promise<void> {
+  return new Promise((resolve) => {
+    if (signal?.aborted) return resolve();
+    const timer = setTimeout(resolve, ms);
+    signal?.addEventListener(
+      "abort",
+      () => {
+        clearTimeout(timer);
+        resolve();
+      },
+      { once: true },
+    );
+  });
+}
+
+async function executionContextForScope(
+  cdp: CdpRunner,
+  scope: DocumentSettleScope,
+): Promise<number | undefined> {
+  if (!scope.frameId) return undefined;
+  const result = await sendToCdpTarget<{ executionContextId?: number }>(
+    cdp,
+    scope.target,
+    "Page.createIsolatedWorld",
+    {
+      frameId: scope.frameId,
+      worldName: SETTLE_WORLD_NAME,
+      grantUniveralAccess: false,
+    },
+  );
+  return result.executionContextId;
+}
+
+async function readQuietProbe(
+  cdp: CdpRunner,
+  scope: DocumentSettleScope,
+  context: ProbeContext,
+): Promise<QuietProbe | null> {
+  try {
+    if (scope.frameId && context.executionContextId === undefined) {
+      context.executionContextId = await executionContextForScope(cdp, scope);
+      if (context.executionContextId === undefined) return null;
+    }
+    const reply = await sendToCdpTarget<{ result?: { value?: unknown } }>(
+      cdp,
+      scope.target,
+      "Runtime.evaluate",
+      {
+        expression: QUIET_PROBE,
+        returnByValue: true,
+        ...(context.executionContextId !== undefined
+          ? { contextId: context.executionContextId }
+          : {}),
+      },
+    );
+    const value = reply.result?.value;
+    if (!value || typeof value !== "object") return null;
+    const { idleMs, readyState } = value as { idleMs?: unknown; readyState?: unknown };
+    if (typeof idleMs !== "number" || typeof readyState !== "string") return null;
+    return { idleMs, readyState };
+  } catch {
+    context.executionContextId = undefined;
+    return null;
+  }
+}
+
+export async function waitForDocumentSettled(
+  cdp: CdpRunner,
+  scope: DocumentSettleScope,
+  options: { signal?: AbortSignal } = {},
+): Promise<SettleOutcome> {
+  const startedAt = Date.now();
+  const floor = startedAt + SETTLE_MIN_MS;
+  const deadline = startedAt + SETTLE_MAX_MS;
+  const context: ProbeContext = {};
+
+  for (;;) {
+    if (options.signal?.aborted) return "cancelled";
+    await sleep(SETTLE_POLL_MS, options.signal);
+    if (options.signal?.aborted) return "cancelled";
+
+    const probe = await readQuietProbe(cdp, scope, context);
+    const now = Date.now();
+    if (now < floor) continue;
+    if (now >= deadline) return "timeout";
+    if (!probe || probe.readyState === "loading") continue;
+    if (probe.idleMs >= SETTLE_QUIET_MS) return "quiet";
+  }
+}

apps/extension/src/lib/recording/draft-policy.ts

@@ -0,0 +1,34 @@
+import type { RecordingDraftStep } from "./types";
+
+const CLIPBOARD_KEYS = new Set(["a", "c", "v", "x", "A", "C", "V", "X"]);
+const MODIFIER_ONLY_KEYS = new Set(["Meta", "Control", "Alt", "Shift", "OS", "Hyper", "Super"]);
+
+export function shouldRecordPress(
+  key: string,
+  modifiers?: Array<"alt" | "ctrl" | "meta" | "shift">,
+): boolean {
+  if (MODIFIER_ONLY_KEYS.has(key)) return false;
+  const mods = modifiers ?? [];
+  const hasCtrlOrMeta = mods.includes("ctrl") || mods.includes("meta");
+  if (hasCtrlOrMeta && CLIPBOARD_KEYS.has(key)) return false;
+  if (key === "Enter" || key === "Escape") return true;
+  if (key.length === 1 && !hasCtrlOrMeta && !mods.includes("alt")) return false;
+  return false;
+}
+
+export function shouldIncludeDraft(step: RecordingDraftStep): boolean {
+  return step.op !== "press" || shouldRecordPress(step.key, step.modifiers);
+}
+
+export function resolveDraftStartUrl(
+  drafts: RecordingDraftStep[],
+  explicitStartUrl?: string,
+  fallbackUrl?: string,
+): string {
+  if (explicitStartUrl) return explicitStartUrl;
+  const navigation = drafts.find(
+    (step): step is Extract<RecordingDraftStep, { op: "navigate" }> => step.op === "navigate",
+  );
+  if (navigation) return navigation.url;
+  return drafts.find((step) => step.pageUrl)?.pageUrl ?? fallbackUrl ?? "about:blank";
+}

apps/extension/src/lib/recording/frame-bridge.ts

@@ -0,0 +1,65 @@
+export const RECORD_FRAME_PORT = "bsk-record-frame";
+export const RECORD_FRAME_QUERY = "bsk-record-frame-query";
+export const RECORD_FRAME_START = "bsk-record-frame-start";
+
+export interface RecordFrameQueryMessage {
+  type: typeof RECORD_FRAME_QUERY;
+}
+
+export interface RecordFrameQueryResponse {
+  active: boolean;
+  requestId?: string;
+  startedAtMs?: number;
+}
+
+export interface RecordFrameStartMessage {
+  type: typeof RECORD_FRAME_START;
+  requestId: string;
+  startedAtMs: number;
+}
+
+export type RecordFramePortMessage =
+  | {
+      type: "ready";
+      requestId: string;
+      producerId: string;
+    }
+  | {
+      type: "ready_ack";
+      requestId: string;
+      producerId: string;
+    }
+  | {
+      type: "stop";
+      requestId: string;
+      commandId: string;
+    }
+  | {
+      type: "cancel";
+      requestId: string;
+    }
+  | {
+      type: "stopped";
+      requestId: string;
+      commandId: string;
+      ok: boolean;
+      error?: string;
+    };
+
+export function isRecordFrameQueryMessage(value: unknown): value is RecordFrameQueryMessage {
+  return (
+    typeof value === "object" &&
+    value !== null &&
+    (value as { type?: unknown }).type === RECORD_FRAME_QUERY
+  );
+}
+
+export function isRecordFrameStartMessage(value: unknown): value is RecordFrameStartMessage {
+  if (typeof value !== "object" || value === null) return false;
+  const message = value as Partial<RecordFrameStartMessage>;
+  return (
+    message.type === RECORD_FRAME_START &&
+    typeof message.requestId === "string" &&
+    typeof message.startedAtMs === "number"
+  );
+}

apps/extension/src/lib/recording/frame-coordinator.ts

@@ -0,0 +1,399 @@
+import {
+  isRecordFrameQueryMessage,
+  RECORD_FRAME_PORT,
+  RECORD_FRAME_START,
+  type RecordFramePortMessage,
+  type RecordFrameQueryResponse,
+  type RecordFrameStartMessage,
+} from "./frame-bridge";
+
+export interface RecordingCaptureScope {
+  tabId: number;
+  documentId: string;
+  browserFrameId: number;
+  producerId: string;
+}
+
+interface ArmedRecording {
+  requestId: string;
+  startedAtMs: number;
+  tabIds: Set<number>;
+  agents: Map<string, FrameAgent>;
+  onDocumentReady?: (scope: RecordingCaptureScope) => void | Promise<void>;
+  finishing: boolean;
+}
+
+interface FrameAgent extends RecordingCaptureScope {
+  requestId: string;
+  port: chrome.runtime.Port;
+  stopWaiters: Map<string, (ok: boolean) => void>;
+}
+
+interface BrowserFrame {
+  frameId: number;
+  documentId?: string;
+}
+
+interface BrowserFrameNavigation extends BrowserFrame {
+  tabId: number;
+  lifecycle: "committed" | "completed";
+}
+
+const RECORD_FRAME_STOP_TIMEOUT_MS = 5_000;
+
+export interface RecordFrameCoordinatorDeps {
+  getAllFrames(tabId: number): Promise<BrowserFrame[]>;
+  sendToDocument(
+    tabId: number,
+    message: RecordFrameStartMessage,
+    target: { documentId?: string; frameId?: number },
+  ): Promise<unknown>;
+  subscribeFrameNavigation?(listener: (frame: BrowserFrameNavigation) => void): () => void;
+}
+
+function documentKey(tabId: number, documentId: string): string {
+  return `${tabId}:${documentId}`;
+}
+
+function senderAddress(
+  sender: chrome.runtime.MessageSender,
+): { tabId: number; documentId: string; browserFrameId: number } | null {
+  const tabId = sender.tab?.id;
+  const documentId = sender.documentId;
+  const browserFrameId = sender.frameId;
+  if (
+    typeof tabId !== "number" ||
+    typeof documentId !== "string" ||
+    typeof browserFrameId !== "number"
+  ) {
+    return null;
+  }
+  return { tabId, documentId, browserFrameId };
+}
+
+function defaultDeps(): RecordFrameCoordinatorDeps {
+  return {
+    async getAllFrames(tabId) {
+      return (await chrome.webNavigation.getAllFrames({ tabId })) ?? [];
+    },
+    sendToDocument(tabId, message, target) {
+      return chrome.tabs.sendMessage(tabId, message, target);
+    },
+    subscribeFrameNavigation(listener) {
+      const onCommitted = (details: chrome.webNavigation.WebNavigationTransitionCallbackDetails) =>
+        listener({ ...details, lifecycle: "committed" });
+      const onCompleted = (details: chrome.webNavigation.WebNavigationFramedCallbackDetails) =>
+        listener({ ...details, lifecycle: "completed" });
+      chrome.webNavigation.onCommitted.addListener(onCommitted);
+      chrome.webNavigation.onCompleted.addListener(onCompleted);
+      return () => {
+        chrome.webNavigation.onCommitted.removeListener(onCommitted);
+        chrome.webNavigation.onCompleted.removeListener(onCompleted);
+      };
+    },
+  };
+}
+
+export class RecordFrameCoordinator {
+  readonly #deps: RecordFrameCoordinatorDeps;
+  readonly #recordings = new Map<string, ArmedRecording>();
+  #attached = false;
+  #detachFrameNavigation: (() => void) | null = null;
+
+  constructor(deps: RecordFrameCoordinatorDeps = defaultDeps()) {
+    this.#deps = deps;
+  }
+
+  attach(): () => void {
+    if (this.#attached) return () => {};
+    this.#attached = true;
+    chrome.runtime.onConnect.addListener(this.#onConnect);
+    chrome.runtime.onMessage.addListener(this.#onMessage);
+    this.#detachFrameNavigation =
+      this.#deps.subscribeFrameNavigation?.(this.#onFrameNavigation) ?? null;
+    return () => {
+      if (!this.#attached) return;
+      this.#attached = false;
+      chrome.runtime.onConnect.removeListener(this.#onConnect);
+      chrome.runtime.onMessage.removeListener(this.#onMessage);
+      this.#detachFrameNavigation?.();
+      this.#detachFrameNavigation = null;
+      for (const recording of this.#recordings.values()) this.#cancelAgents(recording);
+      this.#recordings.clear();
+    };
+  }
+
+  begin(
+    requestId: string,
+    startedAtMs: number,
+    initialTabId: number,
+    onDocumentReady?: (scope: RecordingCaptureScope) => void | Promise<void>,
+  ): void {
+    const previous = this.#recordings.get(requestId);
+    if (previous) this.#cancelAgents(previous);
+    this.#recordings.set(requestId, {
+      requestId,
+      startedAtMs,
+      // Register the initial tab before navigation begins. A newly-created
+      // child document can query at document_start before armTab's frame
+      // snapshot sees it, especially when Chromium promotes it to an OOPIF.
+      tabIds: new Set([initialTabId]),
+      agents: new Map(),
+      ...(onDocumentReady ? { onDocumentReady } : {}),
+      finishing: false,
+    });
+  }
+
+  async armTab(requestId: string, tabId: number): Promise<boolean> {
+    const recording = this.#recordings.get(requestId);
+    if (!recording || recording.finishing) return false;
+    recording.tabIds.add(tabId);
+
+    let frames: BrowserFrame[];
+    try {
+      frames = await this.#deps.getAllFrames(tabId);
+    } catch {
+      frames = [{ frameId: 0 }];
+    }
+    if (!frames.some((frame) => frame.frameId === 0)) frames.unshift({ frameId: 0 });
+
+    const results = await Promise.all(
+      frames.map(async (frame) => {
+        const started = await this.#startFrame(recording, tabId, frame);
+        return { frameId: frame.frameId, started };
+      }),
+    );
+    return results.some((result) => result.frameId === 0 && result.started);
+  }
+
+  sourceFor(
+    requestId: string,
+    producerId: string,
+    sender: chrome.runtime.MessageSender,
+  ): RecordingCaptureScope | null {
+    const address = senderAddress(sender);
+    const recording = this.#recordings.get(requestId);
+    if (!address || !recording) return null;
+    const agent = recording.agents.get(documentKey(address.tabId, address.documentId));
+    if (!agent || agent.producerId !== producerId) return null;
+    return {
+      tabId: agent.tabId,
+      documentId: agent.documentId,
+      browserFrameId: agent.browserFrameId,
+      producerId: agent.producerId,
+    };
+  }
+
+  async stop(requestId: string): Promise<boolean> {
+    const recording = this.#recordings.get(requestId);
+    if (!recording) return true;
+    recording.finishing = true;
+    const commandId = crypto.randomUUID();
+    const results = await Promise.all(
+      [...recording.agents.values()].map((agent) => this.#stopAgent(agent, requestId, commandId)),
+    );
+    const succeeded = results.every(Boolean);
+    if (succeeded) this.#recordings.delete(requestId);
+    else recording.finishing = false;
+    return succeeded;
+  }
+
+  cancel(requestId: string): void {
+    const recording = this.#recordings.get(requestId);
+    if (!recording) return;
+    this.#cancelAgents(recording);
+    this.#recordings.delete(requestId);
+  }
+
+  readonly #onMessage = (
+    message: unknown,
+    sender: chrome.runtime.MessageSender,
+    sendResponse: (response: RecordFrameQueryResponse) => void,
+  ): boolean => {
+    if (!isRecordFrameQueryMessage(message)) return false;
+    const tabId = sender.tab?.id;
+    const recording =
+      typeof tabId === "number"
+        ? [...this.#recordings.values()].find(
+            (candidate) => !candidate.finishing && candidate.tabIds.has(tabId),
+          )
+        : undefined;
+    sendResponse(
+      recording
+        ? {
+            active: true,
+            requestId: recording.requestId,
+            startedAtMs: recording.startedAtMs,
+          }
+        : { active: false },
+    );
+    return false;
+  };
+
+  readonly #onFrameNavigation = (frame: BrowserFrameNavigation): void => {
+    if (frame.frameId === 0) return;
+    for (const recording of this.#recordings.values()) {
+      if (recording.finishing || !recording.tabIds.has(frame.tabId)) continue;
+      // onCommitted can race document_start, while onCompleted runs after the
+      // content script is available. Starting twice is safe and closes both
+      // the early-query and late-SPA-frame gaps.
+      void (async () => {
+        const started = await this.#startFrame(recording, frame.tabId, frame);
+        if (!started || frame.lifecycle !== "completed" || recording.finishing) return;
+        const agent = frame.documentId
+          ? recording.agents.get(documentKey(frame.tabId, frame.documentId))
+          : [...recording.agents.values()].find(
+              (candidate) =>
+                candidate.tabId === frame.tabId && candidate.browserFrameId === frame.frameId,
+            );
+        if (!agent) return;
+        await recording.onDocumentReady?.({
+          tabId: agent.tabId,
+          documentId: agent.documentId,
+          browserFrameId: agent.browserFrameId,
+          producerId: agent.producerId,
+        });
+      })();
+    }
+  };
+
+  readonly #onConnect = (port: chrome.runtime.Port): void => {
+    if (port.name !== RECORD_FRAME_PORT) return;
+    const address = senderAddress(port.sender ?? {});
+    if (!address) {
+      port.disconnect();
+      return;
+    }
+    let registered: FrameAgent | null = null;
+    port.onMessage.addListener((raw: unknown) => {
+      const message = raw as Partial<RecordFramePortMessage>;
+      if (message.type === "ready") {
+        if (typeof message.requestId !== "string" || typeof message.producerId !== "string") {
+          port.disconnect();
+          return;
+        }
+        const requestId = message.requestId;
+        const producerId = message.producerId;
+        if (registered) {
+          if (registered.requestId === requestId && registered.producerId === producerId) {
+            port.postMessage({
+              type: "ready_ack",
+              requestId,
+              producerId,
+            } satisfies RecordFramePortMessage);
+          } else {
+            port.disconnect();
+          }
+          return;
+        }
+        const recording = this.#recordings.get(requestId);
+        if (
+          !recording ||
+          recording.finishing ||
+          !recording.tabIds.has(address.tabId) ||
+          producerId.length === 0
+        ) {
+          port.disconnect();
+          return;
+        }
+        const agent: FrameAgent = {
+          ...address,
+          requestId,
+          producerId,
+          port,
+          stopWaiters: new Map(),
+        };
+        registered = agent;
+        const key = documentKey(address.tabId, address.documentId);
+        const previous = recording.agents.get(key);
+        if (previous && previous !== agent) previous.port.disconnect();
+        recording.agents.set(key, agent);
+        port.postMessage({
+          type: "ready_ack",
+          requestId,
+          producerId,
+        } satisfies RecordFramePortMessage);
+        return;
+      }
+      if (
+        message.type === "stopped" &&
+        registered &&
+        message.requestId === registered.requestId &&
+        typeof message.commandId === "string"
+      ) {
+        registered.stopWaiters.get(message.commandId)?.(message.ok === true);
+        registered.stopWaiters.delete(message.commandId);
+      }
+    });
+    port.onDisconnect.addListener(() => {
+      if (!registered) return;
+      for (const resolve of registered.stopWaiters.values()) resolve(true);
+      const recording = this.#recordings.get(registered.requestId);
+      const key = documentKey(registered.tabId, registered.documentId);
+      if (recording?.agents.get(key) === registered) recording.agents.delete(key);
+    });
+  };
+
+  #stopAgent(agent: FrameAgent, requestId: string, commandId: string): Promise<boolean> {
+    return new Promise((resolve) => {
+      const timer = setTimeout(() => {
+        agent.stopWaiters.delete(commandId);
+        resolve(false);
+      }, RECORD_FRAME_STOP_TIMEOUT_MS);
+      agent.stopWaiters.set(commandId, (ok) => {
+        clearTimeout(timer);
+        resolve(ok);
+      });
+      try {
+        agent.port.postMessage({
+          type: "stop",
+          requestId,
+          commandId,
+        } satisfies RecordFramePortMessage);
+      } catch {
+        clearTimeout(timer);
+        agent.stopWaiters.delete(commandId);
+        resolve(true);
+      }
+    });
+  }
+
+  async #startFrame(
+    recording: ArmedRecording,
+    tabId: number,
+    frame: BrowserFrame,
+  ): Promise<boolean> {
+    const message: RecordFrameStartMessage = {
+      type: RECORD_FRAME_START,
+      requestId: recording.requestId,
+      startedAtMs: recording.startedAtMs,
+    };
+    try {
+      const response = await this.#deps.sendToDocument(tabId, message, {
+        ...(frame.documentId ? { documentId: frame.documentId } : { frameId: frame.frameId }),
+      });
+      return isStarted(response);
+    } catch {
+      return false;
+    }
+  }
+
+  #cancelAgents(recording: ArmedRecording): void {
+    for (const agent of recording.agents.values()) {
+      try {
+        agent.port.postMessage({
+          type: "cancel",
+          requestId: recording.requestId,
+        } satisfies RecordFramePortMessage);
+      } catch {
+        // A destroyed Document has no remaining capture state to cancel.
+      }
+    }
+  }
+}
+
+function isStarted(value: unknown): boolean {
+  return typeof value === "object" && value !== null && (value as { ok?: unknown }).ok === true;
+}
+
+export const recordFrameCoordinator = new RecordFrameCoordinator();

apps/extension/src/lib/recording/navigation-policy.ts

@@ -0,0 +1,5 @@
+const REDIRECT_QUALIFIERS = new Set(["client_redirect", "server_redirect"]);
+
+export function hasRedirectQualifier(qualifiers?: readonly string[]): boolean {
+  return (qualifiers ?? []).some((qualifier) => REDIRECT_QUALIFIERS.has(qualifier));
+}

apps/extension/src/lib/recording/observation-capture.ts

@@ -0,0 +1,140 @@
+import type { RenderedRef } from "@browser-skill/vom";
+import type { CdpTarget } from "@/browser-driver/frame-graph";
+import {
+  type CaptureVomMatchNode,
+  type CaptureVomObservationResult,
+  captureVomObservation,
+} from "@/tools/capture-vom-observation";
+import type { CdpRunner, ChromeTabsApi } from "@/tools/shared";
+
+export interface IndexedObservationNode {
+  frameId: string;
+  geometry: CaptureVomMatchNode;
+  ref?: RenderedRef;
+}
+
+export interface CapturedRecordingObservation {
+  rootFrameId: string;
+  index: ObservationNodeIndex;
+  url: string;
+  title?: string;
+  vomText: string;
+  truncated: boolean;
+}
+
+export interface RegisteredObservation {
+  stateId: string;
+  rootFrameId: string;
+  index: ObservationNodeIndex;
+  url: string;
+}
+
+export interface RecordingDocumentScope {
+  frameId: string;
+  target: CdpTarget;
+}
+
+function nodeKey(frameId: string, backendNodeId: number): string {
+  return `${frameId}:${backendNodeId}`;
+}
+
+function frameTagKey(frameId: string, tag: string): string {
+  return `${frameId}:${tag.toLowerCase()}`;
+}
+
+export class ObservationNodeIndex {
+  readonly #nodesByFrameTag = new Map<string, IndexedObservationNode[]>();
+  readonly #refById = new Map<string, RenderedRef>();
+  readonly #refsByFrame = new Map<string, RenderedRef[]>();
+  readonly #scopeByProducer = new Map<string, RecordingDocumentScope | null>();
+
+  constructor(
+    input: Pick<CaptureVomObservationResult, "rootFrameId" | "matchNodes" | "refs"> &
+      Partial<Pick<CaptureVomObservationResult, "frames">>,
+  ) {
+    const refByNode = new Map<string, RenderedRef>();
+    for (const ref of input.refs) {
+      const frameId = ref.frameId ?? input.rootFrameId;
+      refByNode.set(nodeKey(frameId, ref.backendNodeId), ref);
+      this.#refById.set(ref.ref, ref);
+      const frameRefs = this.#refsByFrame.get(frameId) ?? [];
+      frameRefs.push(ref);
+      this.#refsByFrame.set(frameId, frameRefs);
+    }
+    for (const frame of input.frames ?? []) {
+      const documentId = frame.recordingDocumentId;
+      if (!documentId) continue;
+      this.#scopeByProducer.set(
+        documentId,
+        this.#scopeByProducer.has(documentId)
+          ? null
+          : { frameId: frame.frameId, target: frame.target },
+      );
+    }
+    for (const geometry of input.matchNodes) {
+      const { frameId } = geometry;
+      const entry = {
+        frameId,
+        geometry,
+        ref: refByNode.get(nodeKey(frameId, geometry.backendNodeId)),
+      };
+      const key = frameTagKey(frameId, geometry.tag);
+      const bucket = this.#nodesByFrameTag.get(key) ?? [];
+      bucket.push(entry);
+      this.#nodesByFrameTag.set(key, bucket);
+    }
+  }
+
+  candidates(frameId: string, tag: string): readonly IndexedObservationNode[] {
+    return this.#nodesByFrameTag.get(frameTagKey(frameId, tag)) ?? [];
+  }
+
+  ref(refId: string): RenderedRef | undefined {
+    return this.#refById.get(refId);
+  }
+
+  refs(frameId: string): readonly RenderedRef[] {
+    return this.#refsByFrame.get(frameId) ?? [];
+  }
+
+  documentScope(producerId: string): RecordingDocumentScope | undefined {
+    return this.#scopeByProducer.get(producerId) ?? undefined;
+  }
+}
+
+async function readTabMeta(
+  tabsApi: ChromeTabsApi,
+  tabId: number,
+): Promise<{ url: string; title?: string }> {
+  try {
+    const tab = await tabsApi.get(tabId);
+    return { url: tab.url ?? "about:blank", title: tab.title };
+  } catch {
+    return { url: "about:blank" };
+  }
+}
+
+export async function captureRecordingObservation(input: {
+  cdp: CdpRunner;
+  tabsApi: ChromeTabsApi;
+  tabId: number;
+  maxTokens: number;
+  redactValues: boolean;
+  signal?: AbortSignal;
+}): Promise<CapturedRecordingObservation> {
+  const { url, title } = await readTabMeta(input.tabsApi, input.tabId);
+  const captured = await captureVomObservation(input.cdp, input.tabId, url, {
+    maxTokens: input.maxTokens,
+    redactValues: input.redactValues,
+    conditionalSurfaceProbe: false,
+    signal: input.signal,
+  });
+  return {
+    rootFrameId: captured.rootFrameId,
+    index: new ObservationNodeIndex(captured),
+    url,
+    title,
+    vomText: captured.text,
+    truncated: captured.truncated,
+  };
+}

apps/extension/src/lib/recording/observation-session.ts

@@ -0,0 +1,128 @@
+import type { CdpRunner, ChromeTabsApi } from "@/tools/shared";
+import { captureRecordingObservation, type RegisteredObservation } from "./observation-capture";
+import { RecordingStateRegistry } from "./state-registry";
+import { matchObservationTarget, unmatchedTarget } from "./target-matcher";
+import type { RecordingDraftStep, StepAnnotation, TargetedRecordingDraft } from "./types";
+
+const DEFAULT_MAX_PAGE_TOKENS = 3_000;
+const MIN_CAPTURE_INTERVAL_MS = 200;
+
+export interface TabObservationCursor {
+  lastSettled: RegisteredObservation | null;
+  lastCaptureAt: number;
+}
+
+function abortableDelay(ms: number, signal?: AbortSignal): Promise<void> {
+  return new Promise((resolve) => {
+    if (signal?.aborted) return resolve();
+    const timer = setTimeout(resolve, ms);
+    signal?.addEventListener(
+      "abort",
+      () => {
+        clearTimeout(timer);
+        resolve();
+      },
+      { once: true },
+    );
+  });
+}
+
+function isTargeted(draft: RecordingDraftStep): draft is TargetedRecordingDraft {
+  return (
+    draft.op === "click" ||
+    draft.op === "hover" ||
+    draft.op === "fill" ||
+    draft.op === "press" ||
+    draft.op === "select"
+  );
+}
+
+export class RecordingObservationSession {
+  readonly registry: RecordingStateRegistry;
+  readonly cursor: TabObservationCursor;
+  readonly annotations: StepAnnotation[];
+  readonly #maxTokens: number;
+  readonly #redactValues: boolean;
+
+  constructor(
+    options: {
+      registry?: RecordingStateRegistry;
+      cursor?: TabObservationCursor;
+      annotations?: StepAnnotation[];
+      maxTokens?: number;
+      redactValues?: boolean;
+    } = {},
+  ) {
+    this.registry = options.registry ?? new RecordingStateRegistry();
+    this.cursor = options.cursor ?? { lastSettled: null, lastCaptureAt: 0 };
+    this.annotations = options.annotations ?? [];
+    this.#maxTokens = options.maxTokens ?? DEFAULT_MAX_PAGE_TOKENS;
+    this.#redactValues = options.redactValues ?? false;
+  }
+
+  async capture(
+    cdp: CdpRunner,
+    tabsApi: ChromeTabsApi,
+    tabId: number,
+    signal?: AbortSignal,
+  ): Promise<RegisteredObservation> {
+    const waitMs = Math.max(0, MIN_CAPTURE_INTERVAL_MS - (Date.now() - this.cursor.lastCaptureAt));
+    if (waitMs > 0) await abortableDelay(waitMs, signal);
+    if (signal?.aborted) throw new DOMException("observation aborted", "AbortError");
+    const captured = await captureRecordingObservation({
+      cdp,
+      tabsApi,
+      tabId,
+      maxTokens: this.#maxTokens,
+      redactValues: this.#redactValues,
+      signal,
+    });
+    const state = this.registry.register({
+      url: captured.url,
+      title: captured.title,
+      vomText: captured.vomText,
+      truncated: captured.truncated,
+    });
+    const observation: RegisteredObservation = {
+      stateId: state.id,
+      rootFrameId: captured.rootFrameId,
+      index: captured.index,
+      url: captured.url,
+    };
+    this.cursor.lastSettled = observation;
+    this.cursor.lastCaptureAt = Date.now();
+    return observation;
+  }
+
+  bindDraft(draft: RecordingDraftStep, draftId: number, previousActionPending = false): void {
+    const observation = this.cursor.lastSettled;
+    if (isTargeted(draft)) {
+      draft.matchedTarget = observation
+        ? matchObservationTarget({
+            observation,
+            hint: draft.targetHint,
+            fallback: draft.captureTarget,
+          })
+        : unmatchedTarget(draft.captureTarget);
+    }
+    if (!observation) return;
+
+    const unmatched = isTargeted(draft) && draft.matchedTarget?.unmatched === true;
+    if (previousActionPending && unmatched) return;
+    draft.preStateId = observation.stateId;
+    this.registry.markStep(observation.stateId, draftId);
+
+    if (!isTargeted(draft) || !draft.matchedTarget?.ref) return;
+    const ref = observation.index.ref(draft.matchedTarget.ref);
+    if (!ref) return;
+    this.annotations.push({
+      draftId,
+      op: draft.op,
+      line: ref.line,
+      stateId: observation.stateId,
+      ...(draft.op === "fill" && !this.#redactValues
+        ? { detail: JSON.stringify(draft.value) }
+        : {}),
+    });
+  }
+}

apps/extension/src/lib/recording/recording-runtime.ts

@@ -0,0 +1,277 @@
+import type { CdpRunner, ChromeTabsApi } from "@/tools/shared";
+import type { StopReason, TraceV3 } from "@/transport/types";
+import { type DocumentSettleScope, waitForDocumentSettled } from "./document-settle";
+import type { RegisteredObservation } from "./observation-capture";
+import { RecordingObservationSession } from "./observation-session";
+import { inferMissingPostStates, SettleController } from "./settle-controller";
+import { RecordingStateRegistry } from "./state-registry";
+import { buildTraceV3 } from "./trace-builder-v3";
+import type { RecordingDraftStep, StepAnnotation } from "./types";
+
+interface TabRecordingContext {
+  session: RecordingObservationSession;
+  settle: SettleController;
+  pendingCapture: PendingCapture | null;
+  frameRefresh: Promise<void>;
+  frameRefreshAbort: AbortController | null;
+}
+
+interface PendingCapture {
+  promise: Promise<RegisteredObservation>;
+  abort: AbortController;
+}
+
+export class RecordingObservationRuntime {
+  readonly #cdp: CdpRunner;
+  readonly #tabsApi: ChromeTabsApi;
+  readonly #registry = new RecordingStateRegistry();
+  readonly #annotations: StepAnnotation[] = [];
+  readonly #contexts = new Map<number, TabRecordingContext>();
+  readonly #maxTokens?: number;
+  readonly #redactValues: boolean;
+
+  constructor(input: {
+    cdp: CdpRunner;
+    tabsApi: ChromeTabsApi;
+    maxTokens?: number;
+    redactValues?: boolean;
+  }) {
+    this.#cdp = input.cdp;
+    this.#tabsApi = input.tabsApi;
+    this.#maxTokens = input.maxTokens;
+    this.#redactValues = input.redactValues ?? false;
+  }
+
+  #context(tabId: number): TabRecordingContext {
+    const existing = this.#contexts.get(tabId);
+    if (existing) return existing;
+    const session = new RecordingObservationSession({
+      registry: this.#registry,
+      annotations: this.#annotations,
+      maxTokens: this.#maxTokens,
+      redactValues: this.#redactValues,
+    });
+    const context: TabRecordingContext = {
+      session,
+      settle: new SettleController({
+        session,
+        cdp: this.#cdp,
+        tabsApi: this.#tabsApi,
+        tabId,
+      }),
+      pendingCapture: null,
+      frameRefresh: Promise.resolve(),
+      frameRefreshAbort: null,
+    };
+    this.#contexts.set(tabId, context);
+    return context;
+  }
+
+  async #capture(tabId: number): Promise<RegisteredObservation> {
+    const context = this.#context(tabId);
+    if (context.pendingCapture) return context.pendingCapture.promise;
+
+    const abort = new AbortController();
+    let pending!: PendingCapture;
+    const promise = context.session.capture(this.#cdp, this.#tabsApi, tabId, abort.signal);
+    pending = {
+      abort,
+      promise: promise.finally(() => {
+        if (context.pendingCapture === pending) context.pendingCapture = null;
+      }),
+    };
+    context.pendingCapture = pending;
+    return pending.promise;
+  }
+
+  async captureInitial(tabId: number): Promise<void> {
+    const context = this.#context(tabId);
+    if (context.session.cursor.lastSettled) return;
+    await this.#capture(tabId);
+  }
+
+  /** Refresh the safe observation after a late child Document becomes usable. */
+  async refreshDocument(tabId: number, producerId: string): Promise<void> {
+    const context = this.#context(tabId);
+    context.frameRefreshAbort?.abort();
+    const abort = new AbortController();
+    context.frameRefreshAbort = abort;
+    const previous = context.frameRefresh;
+    const refresh = (async () => {
+      await previous;
+      if (abort.signal.aborted) return;
+      if (context.pendingCapture) await Promise.allSettled([context.pendingCapture.promise]);
+      if (abort.signal.aborted) return;
+
+      // The first capture discovers the safe producer -> CDP Document scope.
+      await this.#capture(tabId);
+      if (abort.signal.aborted) return;
+      const scope = context.session.cursor.lastSettled?.index.documentScope(producerId);
+      if (scope) {
+        await waitForDocumentSettled(this.#cdp, scope, { signal: abort.signal });
+      }
+      if (abort.signal.aborted) return;
+
+      // Capture again after the child SPA is quiet so the next action can be
+      // matched against refs that actually exist in its pre-action state.
+      await this.#capture(tabId);
+    })();
+    const tracked = refresh.catch(() => {});
+    context.frameRefresh = tracked;
+    try {
+      await refresh;
+    } finally {
+      if (context.frameRefresh === tracked) context.frameRefreshAbort = null;
+    }
+  }
+
+  async captureTabTransition(
+    fromTabId: number,
+    toTabId: number,
+  ): Promise<{ preStateId?: string; postStateId?: string; targetUrl?: string }> {
+    const from = this.#context(fromTabId);
+    await this.#flushContext(from);
+    if (!from.session.cursor.lastSettled) {
+      try {
+        await this.captureInitial(fromTabId);
+      } catch {
+        // A target-tab observation can still preserve the forward transition.
+      }
+    }
+
+    const to = this.#context(toTabId);
+    let target = to.session.cursor.lastSettled;
+    try {
+      target = await this.#capture(toTabId);
+    } catch {
+      // Reuse the last settled target state when a refresh races navigation.
+    }
+
+    const source = from.session.cursor.lastSettled;
+    return {
+      ...(source ? { preStateId: source.stateId } : {}),
+      ...(target ? { postStateId: target.stateId, targetUrl: target.url } : {}),
+    };
+  }
+
+  bindTabTransition(
+    draft: Extract<RecordingDraftStep, { op: "switch_tab" }>,
+    draftId: number,
+  ): void {
+    if (draft.preStateId) this.#registry.markStep(draft.preStateId, draftId);
+  }
+
+  async processDraft(
+    tabId: number,
+    drafts: RecordingDraftStep[],
+    draftIndex: number,
+    producerId?: string,
+  ): Promise<void> {
+    const draft = drafts[draftIndex];
+    if (!draft) return;
+    const context = this.#context(tabId);
+    if (!context.session.cursor.lastSettled) {
+      try {
+        await this.captureInitial(tabId);
+      } catch {
+        // Post-action settle can still provide a usable state.
+      }
+    }
+    const scope = producerId
+      ? context.session.cursor.lastSettled?.index.documentScope(producerId)
+      : undefined;
+    if (
+      producerId &&
+      (draft.op === "click" ||
+        draft.op === "hover" ||
+        draft.op === "fill" ||
+        draft.op === "press" ||
+        draft.op === "select")
+    ) {
+      draft.targetHint = {
+        ...(draft.targetHint ?? {}),
+        frameId: scope?.frameId ?? null,
+        geometrySpace: "local",
+      };
+    }
+    context.session.bindDraft(draft, draftIndex + 1, context.settle.hasPending);
+    context.settle.schedule(drafts, draftIndex, scope);
+  }
+
+  scheduleSettle(
+    tabId: number,
+    drafts: RecordingDraftStep[],
+    draftIndex: number,
+    scope?: DocumentSettleScope,
+  ): void {
+    this.#context(tabId).settle.schedule(drafts, draftIndex, scope);
+  }
+
+  clearRedirect(tabId: number): void {
+    this.#contexts.get(tabId)?.settle.clearRedirect();
+  }
+
+  scheduleRedirect(tabId: number, drafts: RecordingDraftStep[], url: string): void {
+    this.#context(tabId).settle.scheduleRedirect(drafts, url);
+  }
+
+  async #flushContext(context: TabRecordingContext): Promise<void> {
+    await context.frameRefresh;
+    if (context.pendingCapture) await Promise.allSettled([context.pendingCapture.promise]);
+    await context.settle.flushRedirects();
+    await context.settle.flush();
+  }
+
+  async flushRedirects(tabId: number): Promise<void> {
+    await this.#contexts.get(tabId)?.settle.flushRedirects();
+  }
+
+  async flush(): Promise<void> {
+    for (const context of this.#contexts.values()) await this.#flushContext(context);
+  }
+
+  async settleTrailing(tabId: number, drafts: RecordingDraftStep[]): Promise<void> {
+    const context = this.#context(tabId);
+    if (context.pendingCapture) await Promise.allSettled([context.pendingCapture.promise]);
+    if (!context.session.cursor.lastSettled) {
+      try {
+        await this.captureInitial(tabId);
+      } catch {
+        // The trace may remain observation-free when CDP is unavailable.
+      }
+    }
+    await context.settle.settleTrailing(drafts);
+    inferMissingPostStates(drafts);
+  }
+
+  cancel(): void {
+    for (const context of this.#contexts.values()) {
+      context.frameRefreshAbort?.abort();
+      context.pendingCapture?.abort.abort();
+      context.settle.cancel();
+    }
+  }
+
+  buildTrace(input: {
+    drafts: RecordingDraftStep[];
+    startedAt: string;
+    purpose?: string;
+    startUrl?: string;
+    stoppedBy: StopReason;
+    bskVersion: string;
+    includeTabSwitches?: boolean;
+  }): TraceV3 {
+    return buildTraceV3({
+      registry: this.#registry,
+      drafts: input.drafts,
+      annotations: this.#annotations,
+      startedAt: input.startedAt,
+      purpose: input.purpose,
+      startUrl: input.startUrl,
+      stoppedBy: input.stoppedBy,
+      bskVersion: input.bskVersion,
+      redactValues: this.#redactValues,
+      includeTabSwitches: input.includeTabSwitches,
+    });
+  }
+}

apps/extension/src/lib/recording/settle-controller.ts

@@ -0,0 +1,233 @@
+import type { CdpRunner, ChromeTabsApi } from "@/tools/shared";
+import { type DocumentSettleScope, waitForDocumentSettled } from "./document-settle";
+import { RecordingObservationSession } from "./observation-session";
+import type { RecordingDraftStep } from "./types";
+
+interface PendingSettle {
+  abort: AbortController;
+  scope: DocumentSettleScope;
+}
+
+interface PendingRedirect {
+  url: string;
+  abort: AbortController;
+}
+
+const CAPTURE_RETRY_DELAY_MS = 250;
+
+function delay(ms: number, signal?: AbortSignal): Promise<void> {
+  return new Promise((resolve) => {
+    if (signal?.aborted) return resolve();
+    const timer = setTimeout(resolve, ms);
+    signal?.addEventListener(
+      "abort",
+      () => {
+        clearTimeout(timer);
+        resolve();
+      },
+      { once: true },
+    );
+  });
+}
+
+function isAbortError(error: unknown): boolean {
+  return (
+    typeof error === "object" &&
+    error !== null &&
+    (error as { name?: string }).name === "AbortError"
+  );
+}
+
+export function inferMissingPostStates(drafts: RecordingDraftStep[]): void {
+  for (let index = 0; index < drafts.length - 1; index += 1) {
+    const draft = drafts[index];
+    const next = drafts[index + 1];
+    if (draft && next && !draft.postStateId && next.preStateId) {
+      draft.postStateId = next.preStateId;
+    }
+  }
+}
+
+export class SettleController {
+  readonly #session: RecordingObservationSession;
+  readonly #cdp: CdpRunner;
+  readonly #tabsApi: ChromeTabsApi;
+  readonly #tabId: number;
+  readonly #rootScope: DocumentSettleScope;
+  readonly #pending = new Map<number, PendingSettle>();
+  #queue = Promise.resolve();
+  #pendingRedirect: PendingRedirect | null = null;
+  #redirectQueue = Promise.resolve();
+
+  constructor(input: {
+    session: RecordingObservationSession;
+    cdp: CdpRunner;
+    tabsApi: ChromeTabsApi;
+    tabId: number;
+  }) {
+    this.#session = input.session;
+    this.#cdp = input.cdp;
+    this.#tabsApi = input.tabsApi;
+    this.#tabId = input.tabId;
+    this.#rootScope = { target: { tabId: input.tabId } };
+  }
+
+  get hasPending(): boolean {
+    return this.#pending.size > 0 || this.#pendingRedirect !== null;
+  }
+
+  async #captureWithRetry(signal?: AbortSignal) {
+    try {
+      return await this.#session.capture(this.#cdp, this.#tabsApi, this.#tabId, signal);
+    } catch (error) {
+      if (signal?.aborted) throw error;
+      await delay(CAPTURE_RETRY_DELAY_MS, signal);
+      return this.#session.capture(this.#cdp, this.#tabsApi, this.#tabId, signal);
+    }
+  }
+
+  schedule(
+    drafts: RecordingDraftStep[],
+    draftIndex: number,
+    scope: DocumentSettleScope = this.#rootScope,
+  ): void {
+    if (scope.target.tabId !== this.#tabId) {
+      throw new Error(
+        `settle scope tab ${scope.target.tabId} does not belong to tab ${this.#tabId}`,
+      );
+    }
+    const landing = drafts[draftIndex]?.preStateId;
+    for (const [index, pending] of this.#pending) {
+      if (index >= draftIndex) continue;
+      pending.abort.abort();
+      this.#pending.delete(index);
+      if (landing && drafts[index] && !drafts[index].postStateId) {
+        drafts[index].postStateId = landing;
+      }
+    }
+
+    this.#pending.get(draftIndex)?.abort.abort();
+    const pending = { abort: new AbortController(), scope };
+    this.#pending.set(draftIndex, pending);
+    const task = async () => {
+      if (pending.abort.signal.aborted) return;
+      try {
+        const outcome = await waitForDocumentSettled(this.#cdp, pending.scope, {
+          signal: pending.abort.signal,
+        });
+        if (outcome === "cancelled") return;
+        const observation = await this.#captureWithRetry(pending.abort.signal);
+        if (!pending.abort.signal.aborted && drafts[draftIndex]) {
+          drafts[draftIndex].postStateId = observation.stateId;
+        }
+      } catch (error) {
+        if (!isAbortError(error)) {
+          console.warn(
+            `[bsk record] post-action observation failed for step ${draftIndex + 1}`,
+            error,
+          );
+        }
+      } finally {
+        if (this.#pending.get(draftIndex) === pending) this.#pending.delete(draftIndex);
+      }
+    };
+    this.#queue = this.#queue.then(task, task).catch(() => {});
+  }
+
+  cancel(): void {
+    for (const pending of this.#pending.values()) pending.abort.abort();
+    this.#pending.clear();
+    this.clearRedirect();
+  }
+
+  async flush(): Promise<void> {
+    for (let round = 0; round < 10; round += 1) {
+      const current = this.#queue;
+      await current;
+      if (current === this.#queue) return;
+    }
+    console.warn("[bsk record] settle queue kept growing; using completed observations");
+  }
+
+  async settleTrailing(drafts: RecordingDraftStep[]): Promise<void> {
+    const trailing: RecordingDraftStep[] = [];
+    for (let index = drafts.length - 1; index >= 0; index -= 1) {
+      const draft = drafts[index];
+      if (!draft || draft.postStateId) break;
+      trailing.push(draft);
+    }
+    if (trailing.length === 0) return;
+    try {
+      const observation = await this.#captureWithRetry();
+      for (const draft of trailing) draft.postStateId = observation.stateId;
+    } catch (error) {
+      console.warn("[bsk record] final observation at stop failed", error);
+    }
+  }
+
+  clearRedirect(): void {
+    this.#pendingRedirect?.abort.abort();
+    this.#pendingRedirect = null;
+  }
+
+  scheduleRedirect(drafts: RecordingDraftStep[], url: string): void {
+    this.#pendingRedirect?.abort.abort();
+    const pending: PendingRedirect = { url, abort: new AbortController() };
+    this.#pendingRedirect = pending;
+    const task = () => this.#settleRedirect(drafts, pending);
+    this.#redirectQueue = this.#redirectQueue.then(task, task).catch(() => {});
+  }
+
+  async #settleRedirect(drafts: RecordingDraftStep[], pending: PendingRedirect): Promise<void> {
+    const outcome = await waitForDocumentSettled(this.#cdp, this.#rootScope, {
+      signal: pending.abort.signal,
+    });
+    if (outcome === "cancelled" || this.#pendingRedirect !== pending) return;
+
+    let finalUrl = pending.url;
+    try {
+      finalUrl = (await this.#tabsApi.get(this.#tabId)).url || finalUrl;
+    } catch {
+      // The navigation event URL remains the best available destination.
+    }
+    if (this.#pendingRedirect !== pending) return;
+    this.#pendingRedirect = null;
+    if (
+      !finalUrl ||
+      finalUrl === "about:blank" ||
+      this.#session.cursor.lastSettled?.url === finalUrl
+    ) {
+      return;
+    }
+
+    const last = drafts[drafts.length - 1];
+    if (last?.op === "navigate" && last.url === finalUrl) {
+      if (!last.postStateId) this.schedule(drafts, drafts.length - 1);
+      await this.flush();
+      return;
+    }
+
+    const draft: RecordingDraftStep = {
+      op: "navigate",
+      url: finalUrl,
+      pageUrl: finalUrl,
+      cause: "browser",
+      transitionQualifiers: ["server_redirect"],
+      preStateId: this.#session.cursor.lastSettled?.stateId,
+    };
+    drafts.push(draft);
+    const draftIndex = drafts.length - 1;
+    if (draft.preStateId) this.#session.registry.markStep(draft.preStateId, draftIndex + 1);
+    this.schedule(drafts, draftIndex);
+    await this.flush();
+  }
+
+  async flushRedirects(): Promise<void> {
+    for (let round = 0; round < 10; round += 1) {
+      const current = this.#redirectQueue;
+      await current;
+      if (current === this.#redirectQueue) return;
+    }
+    console.warn("[bsk record] redirect queue kept growing; using the final completed landing");
+  }
+}

apps/extension/src/lib/recording/state-registry.ts

@@ -0,0 +1,57 @@
+export interface RecordedStateEntry {
+  id: string;
+  url: string;
+  title?: string;
+  /** Rendered VOM text; capture indexes and source DOM/AX data never enter the registry. */
+  vomText: string;
+  truncated: boolean;
+  stepsHere: number[];
+}
+
+function stateIdentity(url: string, body: string): string {
+  return `${url}\0${body}`;
+}
+
+export class RecordingStateRegistry {
+  readonly #entriesById = new Map<string, RecordedStateEntry>();
+  readonly #idByIdentity = new Map<string, string>();
+  #nextId = 1;
+
+  register(input: {
+    url: string;
+    title?: string;
+    vomText: string;
+    truncated?: boolean;
+  }): RecordedStateEntry {
+    const identity = stateIdentity(input.url, input.vomText);
+    const existingId = this.#idByIdentity.get(identity);
+    if (existingId) {
+      const existing = this.#entriesById.get(existingId)!;
+      if (!existing.title && input.title) existing.title = input.title;
+      if (input.truncated) existing.truncated = true;
+      return existing;
+    }
+
+    const entry: RecordedStateEntry = {
+      id: `s${this.#nextId}`,
+      url: input.url,
+      ...(input.title ? { title: input.title } : {}),
+      vomText: input.vomText,
+      truncated: input.truncated ?? false,
+      stepsHere: [],
+    };
+    this.#nextId += 1;
+    this.#entriesById.set(entry.id, entry);
+    this.#idByIdentity.set(identity, entry.id);
+    return entry;
+  }
+
+  values(): RecordedStateEntry[] {
+    return [...this.#entriesById.values()];
+  }
+
+  markStep(stateId: string, draftId: number): void {
+    const entry = this.#entriesById.get(stateId);
+    if (entry && !entry.stepsHere.includes(draftId)) entry.stepsHere.push(draftId);
+  }
+}

apps/extension/src/lib/recording/step-buffer.ts

@@ -0,0 +1,155 @@
+import type { RecordStepPayload } from "../record-bridge";
+import { hasRedirectQualifier } from "./navigation-policy";
+import type { RecordingDraftStep, TargetMatchHint } from "./types";
+
+export interface RecordingStepBuffer {
+  steps: RecordingDraftStep[];
+  navigation: RecordingNavigationCursor;
+}
+
+export interface RecordingNavigationCursor {
+  currentUrl?: string;
+  pendingNavigation: boolean;
+  pendingNavigationDeadline?: number;
+}
+
+const NAVIGATION_TRIGGER_WINDOW_MS = 3_000;
+
+function toDraftStep(
+  payload: RecordStepPayload,
+  targetHint?: TargetMatchHint,
+): RecordingDraftStep | null {
+  const pageUrl = payload.page_url;
+  const common = {
+    ...(pageUrl ? { pageUrl } : {}),
+    ...(targetHint ? { targetHint } : {}),
+  };
+  switch (payload.op) {
+    case "click":
+      return payload.target ? { op: "click", captureTarget: payload.target, ...common } : null;
+    case "hover":
+      return payload.target ? { op: "hover", captureTarget: payload.target, ...common } : null;
+    case "fill":
+      return payload.target
+        ? {
+            op: "fill",
+            captureTarget: payload.target,
+            value: payload.value ?? "",
+            ...(payload.commit ? { commit: payload.commit } : {}),
+            ...(payload.redacted ? { redacted: true } : {}),
+            ...common,
+          }
+        : null;
+    case "press":
+      return payload.key
+        ? {
+            op: "press",
+            key: payload.key,
+            ...(payload.target ? { captureTarget: payload.target } : {}),
+            ...(payload.modifiers?.length ? { modifiers: payload.modifiers } : {}),
+            ...common,
+          }
+        : null;
+    case "select":
+      return payload.target && payload.values
+        ? {
+            op: "select",
+            captureTarget: payload.target,
+            values: payload.values,
+            ...(payload.labels?.length ? { labels: payload.labels } : {}),
+            ...common,
+          }
+        : null;
+    case "navigate":
+      return null;
+  }
+}
+
+function annotateLastStepNavigation(buffer: RecordingStepBuffer, url: string): number {
+  for (let index = buffer.steps.length - 1; index >= 0; index -= 1) {
+    const step = buffer.steps[index];
+    if (!step) continue;
+    if (step.op === "click" || step.op === "press" || step.op === "select" || step.op === "fill") {
+      step.navigatedTo = url;
+      return index;
+    }
+    break;
+  }
+  return -1;
+}
+
+export type NavigationObserveResult =
+  | { kind: "noop" }
+  | { kind: "annotated"; index: number }
+  | { kind: "appended"; index: number }
+  | { kind: "coalesce_redirect"; url: string };
+
+export function observeRecordedNavigation(
+  buffer: RecordingStepBuffer,
+  url: string,
+  causedByAction?: boolean,
+  transitionType?: string,
+  transitionQualifiers?: string[],
+): NavigationObserveResult {
+  const navigation = buffer.navigation;
+  if (!url || url === navigation.currentUrl) return { kind: "noop" };
+  navigation.currentUrl = url;
+
+  const pendingIsCurrent =
+    navigation.pendingNavigation &&
+    (navigation.pendingNavigationDeadline === undefined ||
+      navigation.pendingNavigationDeadline >= Date.now());
+
+  if (causedByAction === true || (causedByAction === undefined && pendingIsCurrent)) {
+    navigation.pendingNavigation = false;
+    navigation.pendingNavigationDeadline = undefined;
+    const annotatedIndex = annotateLastStepNavigation(buffer, url);
+    if (annotatedIndex >= 0) return { kind: "annotated", index: annotatedIndex };
+  } else {
+    navigation.pendingNavigation = false;
+    navigation.pendingNavigationDeadline = undefined;
+  }
+
+  if (hasRedirectQualifier(transitionQualifiers)) {
+    return { kind: "coalesce_redirect", url };
+  }
+  buffer.steps.push({
+    op: "navigate",
+    url,
+    pageUrl: url,
+    transitionType,
+    transitionQualifiers,
+  });
+  return { kind: "appended", index: buffer.steps.length - 1 };
+}
+
+export function appendRecordedPayload(
+  buffer: RecordingStepBuffer,
+  payload: RecordStepPayload,
+  targetHint?: TargetMatchHint,
+): number | null {
+  if (payload.op === "navigate") {
+    if (!payload.url) return null;
+    const result = observeRecordedNavigation(
+      buffer,
+      payload.url,
+      payload.navigation_caused_by_action,
+      payload.transitionType,
+      payload.transitionQualifiers,
+    );
+    return result.kind === "appended" ? result.index : null;
+  }
+  const step = toDraftStep(
+    { ...payload, page_url: payload.page_url ?? buffer.navigation.currentUrl },
+    targetHint,
+  );
+  if (!step) return null;
+  buffer.steps.push(step);
+  if (step.op === "click" || step.op === "press" || step.op === "select" || step.op === "fill") {
+    buffer.navigation.pendingNavigation = payload.expects_navigation === true;
+    buffer.navigation.pendingNavigationDeadline = buffer.navigation.pendingNavigation
+      ? Date.now() + NAVIGATION_TRIGGER_WINDOW_MS
+      : undefined;
+  }
+  return buffer.steps.length - 1;
+}

apps/extension/src/lib/recording/tab-coordinator.ts

@@ -0,0 +1,61 @@
+import type { RecordingNavigationCursor } from "./step-buffer";
+
+export interface TabActivation {
+  tabId: number;
+  revision: number;
+}
+
+export class RecordingTabCoordinator {
+  readonly #navigationByTab = new Map<number, RecordingNavigationCursor>();
+  #activeTabId: number;
+  #currentTabId: number;
+  #activationRevision = 0;
+
+  constructor(initialTabId: number, initialUrl?: string) {
+    this.#activeTabId = initialTabId;
+    this.#currentTabId = initialTabId;
+    this.#navigationByTab.set(initialTabId, {
+      currentUrl: initialUrl,
+      pendingNavigation: false,
+    });
+  }
+
+  get activeTabId(): number {
+    return this.#activeTabId;
+  }
+
+  get currentTabId(): number {
+    return this.#currentTabId;
+  }
+
+  noteActivation(tabId: number): TabActivation {
+    this.#activeTabId = tabId;
+    this.#activationRevision += 1;
+    return { tabId, revision: this.#activationRevision };
+  }
+
+  isLatest(activation: TabActivation): boolean {
+    return (
+      activation.tabId === this.#activeTabId && activation.revision === this.#activationRevision
+    );
+  }
+
+  commit(tabId: number, currentUrl?: string): void {
+    if (currentUrl !== undefined) this.navigation(tabId).currentUrl = currentUrl;
+    this.#currentTabId = tabId;
+  }
+
+  navigation(tabId: number, fallbackUrl?: string): RecordingNavigationCursor {
+    const existing = this.#navigationByTab.get(tabId);
+    if (existing) {
+      if (!existing.currentUrl && fallbackUrl) existing.currentUrl = fallbackUrl;
+      return existing;
+    }
+    const created: RecordingNavigationCursor = {
+      currentUrl: fallbackUrl,
+      pendingNavigation: false,
+    };
+    this.#navigationByTab.set(tabId, created);
+    return created;
+  }
+}

apps/extension/src/lib/recording/target-matcher.ts

@@ -0,0 +1,86 @@
+import type { RenderedRef } from "@browser-skill/vom";
+import type { TargetDescriptorV3 } from "@/transport/types";
+import type { CaptureTargetDescriptor } from "../describe-target";
+import type { IndexedObservationNode, RegisteredObservation } from "./observation-capture";
+import type { TargetGeometry, TargetMatchHint } from "./types";
+
+const MATCH_TOLERANCE_PX = 2;
+
+function close(a: number, b: number): boolean {
+  return Math.abs(a - b) <= MATCH_TOLERANCE_PX;
+}
+
+function rectMatches(a: TargetGeometry["rect"], b: TargetGeometry["rect"]): boolean {
+  return close(a.x, b.x) && close(a.y, b.y) && close(a.w, b.w) && close(a.h, b.h);
+}
+
+function candidateMatches(
+  candidate: IndexedObservationNode,
+  geometry: TargetGeometry,
+  geometrySpace: "top" | "local",
+): boolean {
+  const rect = geometrySpace === "local" ? candidate.geometry.localRect : candidate.geometry.rect;
+  return rect != null && rectMatches(geometry.rect, rect);
+}
+
+export function unmatchedTarget(fallback?: CaptureTargetDescriptor): TargetDescriptorV3 {
+  return {
+    ...(fallback?.role ? { role: fallback.role } : {}),
+    ...(fallback?.name ? { name: fallback.name } : {}),
+    unmatched: true,
+  };
+}
+
+function normalized(value: string | undefined): string {
+  return value?.replace(/\s+/g, " ").trim().toLowerCase() ?? "";
+}
+
+function matchesSemantics(ref: RenderedRef, fallback?: CaptureTargetDescriptor): boolean {
+  const role = normalized(fallback?.role);
+  const name = normalized(fallback?.name);
+  if (!role && !name) return false;
+  if (role && normalized(ref.role) !== role) return false;
+  if (name && normalized(ref.name) !== name) return false;
+  return true;
+}
+
+function descriptor(ref: RenderedRef): TargetDescriptorV3 {
+  return {
+    ref: ref.ref,
+    ...(ref.role ? { role: ref.role } : {}),
+    ...(ref.name ? { name: ref.name } : {}),
+    ...(ref.ctx ? { ctx: ref.ctx } : {}),
+  };
+}
+
+export function matchObservationTarget(input: {
+  observation: RegisteredObservation;
+  hint?: TargetMatchHint;
+  fallback?: CaptureTargetDescriptor;
+}): TargetDescriptorV3 {
+  if (input.hint?.frameId === null) return unmatchedTarget(input.fallback);
+  const frameId = input.hint?.frameId ?? input.observation.rootFrameId;
+  const geometry = input.hint?.geometry;
+  if (geometry) {
+    const matches = input.observation.index
+      .candidates(frameId, geometry.tag)
+      .filter(
+        (candidate) =>
+          candidate.ref &&
+          candidateMatches(candidate, geometry, input.hint?.geometrySpace ?? "top"),
+      );
+    if (matches.length === 1) return descriptor(matches[0]!.ref!);
+    const semanticMatches = matches.filter(
+      (candidate) => candidate.ref && matchesSemantics(candidate.ref, input.fallback),
+    );
+    if (semanticMatches.length === 1) return descriptor(semanticMatches[0]!.ref!);
+    return unmatchedTarget(input.fallback);
+  }
+
+  const semanticMatches = input.observation.index
+    .refs(frameId)
+    .filter((ref) => matchesSemantics(ref, input.fallback));
+  return semanticMatches.length === 1
+    ? descriptor(semanticMatches[0]!)
+    : unmatchedTarget(input.fallback);
+}

apps/extension/src/lib/recording/trace-builder-v3.ts

@@ -0,0 +1,92 @@
+import {
+  type StepV3,
+  type StopReason,
+  TRACE_VERSION_V3,
+  type TraceStateV3,
+  type TraceV3,
+  VOM_FORMAT_VERSION,
+} from "@/transport/types";
+import { resolveDraftStartUrl } from "./draft-policy";
+import type { RecordedStateEntry, RecordingStateRegistry } from "./state-registry";
+import { reduceTraceStepsV3 } from "./trace-reducer-v3";
+import { formatTraceStateBody } from "./trace-state-body";
+import type { RecordingDraftStep, StepAnnotation } from "./types";
+
+function publishedEntries(registry: RecordingStateRegistry, steps: StepV3[]): RecordedStateEntry[] {
+  const entries = registry.values();
+  if (steps.length === 0) return entries.slice(0, 1);
+  const referenced = new Set(steps.flatMap((step) => [step.state, step.result.state]));
+  return entries.filter((entry) => referenced.has(entry.id));
+}
+
+function remapDraftIds(draftIds: number[], stepIdByDraftId: Map<number, number>): number[] {
+  return [
+    ...new Set(
+      draftIds.flatMap((id) => {
+        const stepId = stepIdByDraftId.get(id);
+        return stepId === undefined ? [] : [stepId];
+      }),
+    ),
+  ].sort((a, b) => a - b);
+}
+
+export function buildTraceV3(input: {
+  registry: RecordingStateRegistry;
+  drafts: RecordingDraftStep[];
+  annotations?: StepAnnotation[];
+  startedAt: string;
+  purpose?: string;
+  startUrl?: string;
+  stoppedBy: StopReason;
+  bskVersion: string;
+  redactValues?: boolean;
+  includeTabSwitches?: boolean;
+}): TraceV3 {
+  const reduced = reduceTraceStepsV3(input.drafts, {
+    includeTabSwitches: input.includeTabSwitches,
+    redactValues: input.redactValues,
+  });
+  const entries = publishedEntries(input.registry, reduced.steps);
+  const publishedId = new Map(entries.map((entry, index) => [entry.id, `s${index + 1}`]));
+  const annotationsByState = new Map<string, StepAnnotation[]>();
+  for (const annotation of input.annotations ?? []) {
+    const bucket = annotationsByState.get(annotation.stateId) ?? [];
+    bucket.push(annotation);
+    annotationsByState.set(annotation.stateId, bucket);
+  }
+  const steps = reduced.steps.map((step) => ({
+    ...step,
+    state: publishedId.get(step.state) ?? step.state,
+    result: { state: publishedId.get(step.result.state) ?? step.result.state },
+  }));
+  const states: TraceStateV3[] = entries.map((entry) => {
+    const id = publishedId.get(entry.id) ?? entry.id;
+    return {
+      id,
+      url: entry.url,
+      ...(entry.title ? { title: entry.title } : {}),
+      body: formatTraceStateBody({
+        stateId: id,
+        url: entry.url,
+        title: entry.title,
+        stepIds: remapDraftIds(entry.stepsHere, reduced.stepIdByDraftId),
+        vomText: entry.vomText,
+        annotations: annotationsByState.get(entry.id) ?? [],
+        stepIdByDraftId: reduced.stepIdByDraftId,
+      }),
+      ...(entry.truncated ? { truncated: true } : {}),
+    };
+  });
+
+  return {
+    version: TRACE_VERSION_V3,
+    ...(input.purpose ? { purpose: input.purpose } : {}),
+    recorded_at: new Date().toISOString(),
+    started_at: input.startedAt,
+    stopped_by: input.stoppedBy,
+    entry: { start_url: resolveDraftStartUrl(input.drafts, input.startUrl, states[0]?.url) },
+    recorder: { bsk: input.bskVersion, vom: VOM_FORMAT_VERSION },
+    states,
+    steps,
+  };
+}

apps/extension/src/lib/recording/trace-reducer-v2.ts

@@ -0,0 +1,215 @@
+import type { PageRefV2, SelectedOptionV2, StepV2, TraceV2 } from "@/transport/types";
+import { resolveDraftStartUrl, shouldIncludeDraft } from "./draft-policy";
+import type { RecordingDraftStep } from "./types";
+
+/** Collapse consecutive navigations to the last hop. */
+function collapseNavigations(steps: RecordingDraftStep[]): RecordingDraftStep[] {
+  const out: RecordingDraftStep[] = [];
+  for (const step of steps) {
+    const prev = out[out.length - 1];
+    if (step.op === "navigate" && prev?.op === "navigate") {
+      out[out.length - 1] = step;
+      continue;
+    }
+    out.push(step);
+  }
+  return out;
+}
+
+function collectUrls(steps: RecordingDraftStep[], startUrl?: string): string[] {
+  const urls: string[] = [];
+  if (startUrl) urls.push(startUrl);
+  for (const step of steps) {
+    if (step.op === "navigate") {
+      urls.push(step.url);
+      continue;
+    }
+    if (step.pageUrl) urls.push(step.pageUrl);
+    if ("navigatedTo" in step && step.navigatedTo) urls.push(step.navigatedTo);
+  }
+  const seen = new Set<string>();
+  const unique: string[] = [];
+  for (const url of urls) {
+    if (!url || seen.has(url)) continue;
+    seen.add(url);
+    unique.push(url);
+  }
+  return unique;
+}
+
+function buildPageRegistry(
+  steps: RecordingDraftStep[],
+  startUrl?: string,
+): { pages: PageRefV2[]; urlToId: Map<string, string> } {
+  const urls = collectUrls(steps, startUrl);
+  const urlToId = new Map<string, string>();
+  const pages = urls.map((url, index) => {
+    const id = `p${index + 1}`;
+    urlToId.set(url, id);
+    return { id, url };
+  });
+  return { pages, urlToId };
+}
+
+function pageIdFor(
+  url: string | undefined,
+  urlToId: Map<string, string>,
+  fallbackUrl?: string,
+): string {
+  if (url && urlToId.has(url)) return urlToId.get(url)!;
+  if (fallbackUrl && urlToId.has(fallbackUrl)) return urlToId.get(fallbackUrl)!;
+  return urlToId.values().next().value ?? "p1";
+}
+
+function pageUrlForDraft(step: RecordingDraftStep, fallbackUrl?: string): string | undefined {
+  if (step.op === "navigate") return step.pageUrl ?? step.url;
+  if (step.pageUrl) return step.pageUrl;
+  return fallbackUrl;
+}
+
+function effectForNavigation(
+  navigatedTo: string | undefined,
+  urlToId: Map<string, string>,
+): StepV2["effect"] {
+  if (!navigatedTo) return undefined;
+  const pageId = urlToId.get(navigatedTo);
+  if (!pageId) return undefined;
+  return { navigated_to: pageId };
+}
+
+function withEffect(step: StepV2, effect: StepV2["effect"]): StepV2 {
+  if (!effect) return step;
+  return { ...step, effect };
+}
+
+function toSelection(values: string[], labels?: string[]): SelectedOptionV2[] {
+  return values.map((value, index) => ({
+    value,
+    ...(labels?.[index] ? { label: labels[index] } : {}),
+  }));
+}
+
+function toV2Step(
+  step: RecordingDraftStep,
+  id: number,
+  urlToId: Map<string, string>,
+  fallbackUrl?: string,
+): StepV2 | null {
+  if (!shouldIncludeDraft(step)) return null;
+
+  const pageUrl = pageUrlForDraft(step, fallbackUrl);
+  const page = pageIdFor(pageUrl, urlToId, fallbackUrl);
+
+  switch (step.op) {
+    case "navigate":
+      return {
+        op: "navigate",
+        id,
+        page: pageIdFor(step.url, urlToId, fallbackUrl),
+        to: step.url,
+      };
+    case "click":
+      if (!step.captureTarget) return null;
+      return withEffect(
+        {
+          op: "click",
+          id,
+          page,
+          target: step.captureTarget,
+        },
+        effectForNavigation(step.navigatedTo, urlToId),
+      );
+    case "fill":
+      if (!step.captureTarget) return null;
+      return {
+        op: "fill",
+        id,
+        page,
+        target: step.captureTarget,
+        value: step.value,
+        ...(step.redacted ? { redacted: true } : {}),
+      };
+    case "press":
+      return withEffect(
+        {
+          op: "press",
+          id,
+          page,
+          key: step.key,
+          ...(step.captureTarget ? { target: step.captureTarget } : {}),
+          ...(step.modifiers?.length ? { modifiers: step.modifiers } : {}),
+        },
+        effectForNavigation(step.navigatedTo, urlToId),
+      );
+    case "select":
+      if (!step.captureTarget) return null;
+      return withEffect(
+        {
+          op: "select",
+          id,
+          page,
+          target: step.captureTarget,
+          selection: toSelection(step.values, step.labels),
+        },
+        effectForNavigation(step.navigatedTo, urlToId),
+      );
+    // `hover` only exists in Trace v3. Peers that negotiate v2 predate the
+    // step variant and fail to decode the whole result if we emit it.
+    case "hover":
+    case "scroll":
+    case "switch_tab":
+      return null;
+  }
+}
+
+interface ReducedTrace {
+  pages: PageRefV2[];
+  steps: StepV2[];
+}
+
+/**
+ * Compile capture drafts into record-only trace v2 steps.
+ * Variable inputs are NOT classified here — executing agents infer that at run time.
+ */
+function reduceTraceSteps(steps: RecordingDraftStep[], startUrl?: string): ReducedTrace {
+  const collapsed = collapseNavigations(steps);
+  const { pages, urlToId } = buildPageRegistry(collapsed, startUrl);
+  const out: StepV2[] = [];
+  let id = 1;
+  let lastUrl = startUrl;
+  for (const draft of collapsed) {
+    if (draft.op === "navigate") lastUrl = draft.url;
+    else if ("navigatedTo" in draft && draft.navigatedTo) lastUrl = draft.navigatedTo;
+    else if (draft.pageUrl) lastUrl = draft.pageUrl;
+    const step = toV2Step(draft, id, urlToId, lastUrl);
+    if (!step) continue;
+    out.push(step);
+    id += 1;
+  }
+  return { pages, steps: out };
+}
+
+function resolveTraceStartUrl(
+  drafts: RecordingDraftStep[],
+  startUrl?: string,
+  pages?: PageRefV2[],
+): string {
+  return resolveDraftStartUrl(drafts, startUrl, pages?.[0]?.url);
+}
+
+export function buildTraceV2(input: {
+  steps: RecordingDraftStep[];
+  startedAt: string;
+  startUrl?: string;
+  purpose?: string;
+}): TraceV2 {
+  const { pages, steps } = reduceTraceSteps(input.steps, input.startUrl);
+  return {
+    recorded_at: new Date().toISOString(),
+    started_at: input.startedAt,
+    ...(input.purpose ? { purpose: input.purpose } : {}),
+    entry: { start_url: resolveTraceStartUrl(input.steps, input.startUrl, pages) },
+    pages,
+    steps,
+  };
+}

apps/extension/src/lib/recording/trace-reducer-v3.ts

@@ -0,0 +1,154 @@
+import type { NavigationCause, StepV3 } from "@/transport/types";
+import { shouldIncludeDraft } from "./draft-policy";
+import { hasRedirectQualifier } from "./navigation-policy";
+import { unmatchedTarget } from "./target-matcher";
+import type { RecordingDraftStep } from "./types";
+
+interface CollapsedDraft {
+  draft: RecordingDraftStep;
+  draftIds: number[];
+}
+
+const TRANSITION_CAUSES: Record<string, NavigationCause> = {
+  typed: "user_typed",
+  generated: "user_typed",
+  keyword: "user_typed",
+  keyword_generated: "user_typed",
+  link: "link",
+  form_submit: "form_submit",
+  reload: "reload",
+  auto_bookmark: "browser",
+  start_page: "browser",
+};
+
+function isRedirect(step: Extract<RecordingDraftStep, { op: "navigate" }>): boolean {
+  return hasRedirectQualifier(step.transitionQualifiers);
+}
+
+function collapseRedirects(steps: RecordingDraftStep[]): CollapsedDraft[] {
+  const output: CollapsedDraft[] = [];
+  steps.forEach((step, index) => {
+    const previous = output[output.length - 1];
+    if (step.op === "navigate" && previous?.draft.op === "navigate" && isRedirect(step)) {
+      previous.draft = {
+        ...previous.draft,
+        url: step.url,
+        postStateId: step.postStateId ?? previous.draft.postStateId,
+      };
+      previous.draftIds.push(index + 1);
+      return;
+    }
+    output.push({ draft: { ...step }, draftIds: [index + 1] });
+  });
+  return output;
+}
+
+function navigationCause(step: Extract<RecordingDraftStep, { op: "navigate" }>): NavigationCause {
+  if (step.cause) return step.cause;
+  const qualifiers = step.transitionQualifiers ?? [];
+  if (qualifiers.includes("forward_back")) return "history";
+  if (qualifiers.includes("from_address_bar")) return "user_typed";
+  return TRANSITION_CAUSES[step.transitionType ?? ""] ?? "browser";
+}
+
+function selection(values: string[], labels?: string[]): Array<{ value: string; label?: string }> {
+  return values.map((value, index) => ({
+    value,
+    ...(labels?.[index] ? { label: labels[index] } : {}),
+  }));
+}
+
+interface ReduceDraftOptions {
+  includeTabSwitches: boolean;
+  redactValues: boolean;
+}
+
+function reduceDraft(
+  draft: RecordingDraftStep,
+  id: number,
+  options: ReduceDraftOptions,
+): StepV3 | null {
+  if (!shouldIncludeDraft(draft)) return null;
+  if (draft.op === "switch_tab") {
+    if (!options.includeTabSwitches || !draft.preStateId || !draft.postStateId) return null;
+    return {
+      op: "switch_tab",
+      id,
+      state: draft.preStateId,
+      result: { state: draft.postStateId },
+    };
+  }
+  const state = draft.preStateId ?? draft.postStateId;
+  const resultState = draft.postStateId ?? draft.preStateId;
+  if (!state || !resultState) return null;
+  const common = { id, state, result: { state: resultState } };
+
+  switch (draft.op) {
+    case "navigate":
+      return { op: "navigate", ...common, to: draft.url, cause: navigationCause(draft) };
+    case "click":
+      return {
+        op: "click",
+        ...common,
+        target: draft.matchedTarget ?? unmatchedTarget(draft.captureTarget),
+      };
+    case "hover":
+      return {
+        op: "hover",
+        ...common,
+        target: draft.matchedTarget ?? unmatchedTarget(draft.captureTarget),
+      };
+    case "fill":
+      const fillIsRedacted = options.redactValues || draft.redacted === true;
+      return {
+        op: "fill",
+        ...common,
+        target: draft.matchedTarget ?? unmatchedTarget(draft.captureTarget),
+        value: fillIsRedacted ? "***" : draft.value,
+        commit: draft.commit ?? "blur",
+        ...(fillIsRedacted ? { redacted: true } : {}),
+      };
+    case "press":
+      return {
+        op: "press",
+        ...common,
+        key: draft.key,
+        ...(draft.captureTarget || draft.matchedTarget
+          ? { target: draft.matchedTarget ?? unmatchedTarget(draft.captureTarget) }
+          : {}),
+        ...(draft.modifiers?.length ? { modifiers: draft.modifiers } : {}),
+      };
+    case "select":
+      return {
+        op: "select",
+        ...common,
+        target: draft.matchedTarget ?? unmatchedTarget(draft.captureTarget),
+        ...(!options.redactValues ? { selection: selection(draft.values, draft.labels) } : {}),
+      };
+    case "scroll":
+      return { op: "scroll", ...common };
+  }
+}
+
+export interface ReducedTraceV3 {
+  steps: StepV3[];
+  stepIdByDraftId: Map<number, number>;
+}
+
+export function reduceTraceStepsV3(
+  steps: RecordingDraftStep[],
+  options: { includeTabSwitches?: boolean; redactValues?: boolean } = {},
+): ReducedTraceV3 {
+  const output: StepV3[] = [];
+  const stepIdByDraftId = new Map<number, number>();
+  for (const { draft, draftIds } of collapseRedirects(steps)) {
+    const step = reduceDraft(draft, output.length + 1, {
+      includeTabSwitches: options.includeTabSwitches === true,
+      redactValues: options.redactValues === true,
+    });
+    if (!step) continue;
+    output.push(step);
+    for (const draftId of draftIds) stepIdByDraftId.set(draftId, step.id);
+  }
+  return { steps: output, stepIdByDraftId };
+}

apps/extension/src/lib/recording/trace-state-body.ts

@@ -0,0 +1,39 @@
+import type { StepAnnotation } from "./types";
+
+function annotationText(annotation: StepAnnotation, stepId: number): string {
+  return ` ⟵ step ${stepId}: ${annotation.op}${annotation.detail ? `: ${annotation.detail}` : ""}`;
+}
+
+export function formatTraceStateBody(input: {
+  stateId: string;
+  url: string;
+  title?: string;
+  stepIds: number[];
+  vomText: string;
+  annotations: StepAnnotation[];
+  stepIdByDraftId: Map<number, number>;
+}): string {
+  const lines = ["# bsk-observation 1", `state: ${JSON.stringify(input.stateId)}`];
+  lines.push(`url: ${JSON.stringify(input.url)}`);
+  if (input.title) lines.push(`title: ${JSON.stringify(input.title)}`);
+  if (input.stepIds.length > 0) lines.push(`steps_here: [${input.stepIds.join(", ")}]`);
+  lines.push("---");
+
+  const byLine = new Map<number, Array<{ annotation: StepAnnotation; stepId: number }>>();
+  for (const annotation of input.annotations) {
+    const stepId = input.stepIdByDraftId.get(annotation.draftId);
+    if (stepId === undefined) continue;
+    const bucket = byLine.get(annotation.line) ?? [];
+    bucket.push({ annotation, stepId });
+    byLine.set(annotation.line, bucket);
+  }
+
+  input.vomText.split("\n").forEach((bodyLine, lineIndex) => {
+    let line = bodyLine;
+    for (const item of byLine.get(lineIndex) ?? []) {
+      line += annotationText(item.annotation, item.stepId);
+    }
+    lines.push(line);
+  });
+  return `${lines.join("\n")}\n`;
+}

apps/extension/src/lib/recording/types.ts

@@ -0,0 +1,85 @@
+import type { CaptureTargetDescriptor } from "@/lib/describe-target";
+import type {
+  FillCommit,
+  KeyModifier,
+  NavigationCause,
+  StepV3,
+  TargetDescriptorV3,
+} from "@/transport/types";
+
+export interface TargetGeometry {
+  /** Top-level viewport-relative CSS pixels, as defined by the geometry module. */
+  rect: { x: number; y: number; w: number; h: number };
+  tag: string;
+}
+
+export interface TargetMatchHint {
+  geometry?: TargetGeometry;
+  /** Missing means top frame; null means the source Document could not be resolved. */
+  frameId?: string | null;
+  geometrySpace?: "top" | "local";
+}
+
+export interface StepAnnotation {
+  draftId: number;
+  op: StepV3["op"];
+  line: number;
+  stateId: string;
+  detail?: string;
+}
+
+interface DraftStateLink {
+  pageUrl?: string;
+  preStateId?: string;
+  postStateId?: string;
+}
+
+interface DraftTarget {
+  captureTarget?: CaptureTargetDescriptor;
+  targetHint?: TargetMatchHint;
+  matchedTarget?: TargetDescriptorV3;
+}
+
+interface DraftNavigationEffect {
+  navigatedTo?: string;
+}
+
+export type RecordingDraftStep =
+  | ({ op: "click" } & DraftStateLink & DraftTarget & DraftNavigationEffect)
+  | ({ op: "hover" } & DraftStateLink & DraftTarget)
+  | ({
+      op: "fill";
+      value: string;
+      commit?: FillCommit;
+      redacted?: boolean;
+    } & DraftStateLink &
+      DraftTarget &
+      DraftNavigationEffect)
+  | ({
+      op: "press";
+      key: string;
+      modifiers?: KeyModifier[];
+    } & DraftStateLink &
+      DraftTarget &
+      DraftNavigationEffect)
+  | ({
+      op: "select";
+      values: string[];
+      labels?: string[];
+    } & DraftStateLink &
+      DraftTarget &
+      DraftNavigationEffect)
+  | ({ op: "scroll" } & DraftStateLink)
+  | ({ op: "switch_tab" } & DraftStateLink)
+  | ({
+      op: "navigate";
+      url: string;
+      cause?: NavigationCause;
+      transitionType?: string;
+      transitionQualifiers?: string[];
+    } & DraftStateLink);
+
+export type TargetedRecordingDraft = Extract<
+  RecordingDraftStep,
+  { op: "click" | "hover" | "fill" | "press" | "select" }
+>;

apps/extension/src/lib/trace-reducer-v2.ts

@@ -1,297 +0,0 @@
-import type { CaptureTargetDescriptor } from "@/lib/describe-target";
-import type { DraftTraceStep, KeyModifier } from "@/transport/types";
-
-const CLIPBOARD_KEYS = new Set(["a", "c", "v", "x", "A", "C", "V", "X"]);
-const MODIFIER_ONLY_KEYS = new Set(["Meta", "Control", "Alt", "Shift", "OS", "Hyper", "Super"]);
-
-export interface TargetDescriptorV2 {
-  role?: string;
-  name?: string;
-  tag: string;
-  name_attr?: string;
-  placeholder?: string;
-  nearby_label?: string;
-}
-
-export interface PageRef {
-  id: string;
-  url: string;
-  title?: string;
-}
-
-export interface SelectedOptionV2 {
-  value: string;
-  label?: string;
-}
-
-export interface StepEffectV2 {
-  navigated_to: string;
-}
-
-export type StepV2 =
-  | { op: "navigate"; id: number; page: string; to: string; effect?: StepEffectV2 }
-  | { op: "click"; id: number; page: string; target: TargetDescriptorV2; effect?: StepEffectV2 }
-  | {
-      op: "fill";
-      id: number;
-      page: string;
-      target: TargetDescriptorV2;
-      value: string;
-      redacted?: boolean;
-      effect?: StepEffectV2;
-    }
-  | {
-      op: "select";
-      id: number;
-      page: string;
-      target: TargetDescriptorV2;
-      selection: SelectedOptionV2[];
-      effect?: StepEffectV2;
-    }
-  | {
-      op: "press";
-      id: number;
-      page: string;
-      key: string;
-      modifiers?: KeyModifier[];
-      target?: TargetDescriptorV2;
-      effect?: StepEffectV2;
-    };
-
-export interface TraceV2 {
-  recorded_at: string;
-  started_at?: string;
-  purpose?: string;
-  entry: { start_url: string };
-  pages: PageRef[];
-  steps: StepV2[];
-}
-
-export function shouldRecordPress(
-  key: string,
-  modifiers?: Array<"alt" | "ctrl" | "meta" | "shift">,
-): boolean {
-  if (MODIFIER_ONLY_KEYS.has(key)) return false;
-  const mods = modifiers ?? [];
-  const hasCtrlOrMeta = mods.includes("ctrl") || mods.includes("meta");
-  if (hasCtrlOrMeta && CLIPBOARD_KEYS.has(key)) return false;
-  if (key === "Enter" || key === "Escape") return true;
-  if (key.length === 1 && !hasCtrlOrMeta && !mods.includes("alt")) return false;
-  return false;
-}
-
-function shouldIncludeDraft(step: DraftTraceStep): boolean {
-  if (step.op === "scroll" || step.op === "hover") return false;
-  if (step.op === "fill" && !(step.value ?? "").trim() && !step.redacted) return false;
-  if (step.op === "press" && !shouldRecordPress(step.key, step.modifiers)) return false;
-  return true;
-}
-
-function collapseNavigations(steps: DraftTraceStep[]): DraftTraceStep[] {
-  const out: DraftTraceStep[] = [];
-  for (const step of steps) {
-    const prev = out[out.length - 1];
-    if (step.op === "navigate" && prev?.op === "navigate") {
-      out[out.length - 1] = step;
-      continue;
-    }
-    out.push(step);
-  }
-  return out;
-}
-
-function collectUrls(steps: DraftTraceStep[], startUrl?: string): string[] {
-  const urls: string[] = [];
-  if (startUrl) urls.push(startUrl);
-  for (const step of steps) {
-    if (step.op === "navigate") {
-      urls.push(step.url);
-      continue;
-    }
-    if ("page_url" in step && step.page_url) urls.push(step.page_url);
-    if ("navigated_to" in step && step.navigated_to) urls.push(step.navigated_to);
-  }
-  const seen = new Set<string>();
-  const unique: string[] = [];
-  for (const url of urls) {
-    if (!url || seen.has(url)) continue;
-    seen.add(url);
-    unique.push(url);
-  }
-  return unique;
-}
-
-function buildPageRegistry(
-  steps: DraftTraceStep[],
-  startUrl?: string,
-): { pages: PageRef[]; urlToId: Map<string, string> } {
-  const urls = collectUrls(steps, startUrl);
-  const urlToId = new Map<string, string>();
-  const pages = urls.map((url, index) => {
-    const id = `p${index + 1}`;
-    urlToId.set(url, id);
-    return { id, url };
-  });
-  return { pages, urlToId };
-}
-
-function pageIdFor(
-  url: string | undefined,
-  urlToId: Map<string, string>,
-  fallbackUrl?: string,
-): string {
-  if (url && urlToId.has(url)) return urlToId.get(url)!;
-  if (fallbackUrl && urlToId.has(fallbackUrl)) return urlToId.get(fallbackUrl)!;
-  return urlToId.values().next().value ?? "p1";
-}
-
-function pageUrlForDraft(step: DraftTraceStep, fallbackUrl?: string): string | undefined {
-  if (step.op === "navigate") return step.page_url ?? step.url;
-  if ("page_url" in step && step.page_url) return step.page_url;
-  return fallbackUrl;
-}
-
-function effectForNavigation(
-  navigatedTo: string | undefined,
-  urlToId: Map<string, string>,
-): StepEffectV2 | undefined {
-  if (!navigatedTo) return undefined;
-  const pageId = urlToId.get(navigatedTo);
-  if (!pageId) return undefined;
-  return { navigated_to: pageId };
-}
-
-function captureTargetToV2Target(capture?: CaptureTargetDescriptor): TargetDescriptorV2 {
-  return {
-    tag: capture?.tag ?? "unknown",
-    ...(capture?.role ? { role: capture.role } : {}),
-    ...(capture?.name ? { name: capture.name } : {}),
-    ...(capture?.name_attr ? { name_attr: capture.name_attr } : {}),
-    ...(capture?.placeholder ? { placeholder: capture.placeholder } : {}),
-    ...(capture?.nearby_label ? { nearby_label: capture.nearby_label } : {}),
-  };
-}
-
-function targetForDraft(step: DraftTraceStep): TargetDescriptorV2 | undefined {
-  if (!("target" in step) && !("captureTarget" in step)) return undefined;
-  const capture = "captureTarget" in step ? step.captureTarget : undefined;
-  if (capture) return captureTargetToV2Target(capture);
-  if ("target" in step && step.target && "tag" in (step.target as object)) {
-    const legacy = step.target as TargetDescriptorV2 & { tag?: string };
-    if (legacy.tag) return legacy;
-  }
-  return undefined;
-}
-
-function toSelection(values: string[], labels?: string[]): SelectedOptionV2[] {
-  return values.map((value, index) => ({
-    value,
-    ...(labels?.[index] ? { label: labels[index] } : {}),
-  }));
-}
-
-function toV2Step(
-  step: DraftTraceStep,
-  id: number,
-  urlToId: Map<string, string>,
-  fallbackUrl?: string,
-): StepV2 | null {
-  if (!shouldIncludeDraft(step)) return null;
-
-  const pageUrl = pageUrlForDraft(step, fallbackUrl);
-  const page = pageIdFor(pageUrl, urlToId, fallbackUrl);
-  const effect =
-    "navigated_to" in step ? effectForNavigation(step.navigated_to, urlToId) : undefined;
-
-  switch (step.op) {
-    case "navigate":
-      return {
-        op: "navigate",
-        id,
-        page: pageIdFor(step.url, urlToId, fallbackUrl),
-        to: step.url,
-      };
-    case "click": {
-      const target = targetForDraft(step);
-      if (!target) return null;
-      return { op: "click", id, page, target, ...(effect ? { effect } : {}) };
-    }
-    case "hover":
-      return null;
-    case "fill": {
-      const target = targetForDraft(step);
-      if (!target) return null;
-      return {
-        op: "fill",
-        id,
-        page,
-        target,
-        value: step.value,
-        ...(step.redacted ? { redacted: true } : {}),
-      };
-    }
-    case "press":
-      return {
-        op: "press",
-        id,
-        page,
-        key: step.key,
-        ...(step.target ? { target: targetForDraft(step) } : {}),
-        ...(step.modifiers?.length ? { modifiers: step.modifiers } : {}),
-        ...(effect ? { effect } : {}),
-      };
-    case "select": {
-      const target = targetForDraft(step);
-      if (!target) return null;
-      return {
-        op: "select",
-        id,
-        page,
-        target,
-        selection: toSelection(step.values, step.labels),
-        ...(effect ? { effect } : {}),
-      };
-    }
-    case "scroll":
-      return null;
-  }
-}
-
-export interface BuildTraceV2Input {
-  steps: DraftTraceStep[];
-  startedAt: string;
-  startUrl?: string;
-  purpose?: string;
-}
-
-export function buildTraceV2(input: BuildTraceV2Input): TraceV2 {
-  const collapsed = collapseNavigations(input.steps);
-  const startUrl =
-    input.startUrl ??
-    collapsed.find(
-      (step): step is Extract<DraftTraceStep, { op: "navigate" }> => step.op === "navigate",
-    )?.url ??
-    collapsed.find((step) => "page_url" in step && step.page_url)?.page_url ??
-    "about:blank";
-  const { pages, urlToId } = buildPageRegistry(collapsed, startUrl);
-  const out: StepV2[] = [];
-  let id = 1;
-  let lastUrl = startUrl;
-  for (const draft of collapsed) {
-    if (draft.op === "navigate") lastUrl = draft.url;
-    else if ("navigated_to" in draft && draft.navigated_to) lastUrl = draft.navigated_to;
-    else if ("page_url" in draft && draft.page_url) lastUrl = draft.page_url;
-    const step = toV2Step(draft, id, urlToId, lastUrl);
-    if (!step) continue;
-    out.push(step);
-    id += 1;
-  }
-  return {
-    recorded_at: new Date().toISOString(),
-    started_at: input.startedAt,
-    ...(input.purpose ? { purpose: input.purpose } : {}),
-    entry: { start_url: startUrl },
-    pages,
-    steps: out,
-  };
-}

apps/extension/src/lib/trace-reducer.ts

@@ -1,236 +0,0 @@
-import type {
-  DraftTraceStep,
-  FillCommit,
-  NavigationCause,
-  Step,
-  TraceState,
-} from "@/transport/types";
-import {
-  DEFAULT_FILL_COMMIT,
-  fnv1a64,
-  mapNavigationCause,
-  nextStateId,
-  resetStateIdCounterForTests,
-} from "./record-constants";
-
-const CLIPBOARD_KEYS = new Set(["a", "c", "v", "x", "A", "C", "V", "X"]);
-const MODIFIER_ONLY_KEYS = new Set(["Meta", "Control", "Alt", "Shift", "OS", "Hyper", "Super"]);
-
-export function shouldRecordPress(
-  key: string,
-  modifiers?: Array<"alt" | "ctrl" | "meta" | "shift">,
-): boolean {
-  if (MODIFIER_ONLY_KEYS.has(key)) return false;
-  const mods = modifiers ?? [];
-  const hasCtrlOrMeta = mods.includes("ctrl") || mods.includes("meta");
-  if (hasCtrlOrMeta && CLIPBOARD_KEYS.has(key)) return false;
-  if (key === "Enter" || key === "Escape") return true;
-  if (key.length === 1 && !hasCtrlOrMeta && !mods.includes("alt")) return false;
-  return false;
-}
-
-function shouldIncludeDraft(step: DraftTraceStep): boolean {
-  if (step.op === "press" && !shouldRecordPress(step.key, step.modifiers)) return false;
-  return true;
-}
-
-/** A reduced draft plus the capture-order ids of every draft folded into it. */
-interface CollapsedDraft {
-  draft: DraftTraceStep;
-  draftIds: number[];
-}
-
-/**
- * Collapse a redirect chain into a single hop.
- *
- * The surviving step keeps the *first* hop's origin state and cause — that is
- * where the user was and why they left — and the *last* hop's destination, so
- * the intermediate URL the browser only passed through never becomes the
- * step's `state` or `to`.
- */
-function collapseNavigations(steps: DraftTraceStep[]): CollapsedDraft[] {
-  const out: CollapsedDraft[] = [];
-  steps.forEach((step, index) => {
-    const draftId = index + 1;
-    const prev = out[out.length - 1];
-    if (step.op === "navigate" && prev?.draft.op === "navigate") {
-      prev.draft = {
-        ...prev.draft,
-        url: step.url,
-        postStateId: step.postStateId ?? prev.draft.postStateId,
-      };
-      prev.draftIds.push(draftId);
-      return;
-    }
-    out.push({ draft: step, draftIds: [draftId] });
-  });
-  return out;
-}
-
-export interface StateRegistryEntry extends TraceState {
-  contentHash: string;
-  rawVomText: string;
-  stepsHere: number[];
-}
-
-export function stateDedupKey(body: string, url: string): string {
-  return `${fnv1a64(body)}:${url}`;
-}
-
-/** Register or reuse a page observation in the states dictionary. */
-export function registerObservation(
-  registry: Map<string, StateRegistryEntry>,
-  input: {
-    url: string;
-    title?: string;
-    rawVomText: string;
-    truncated?: boolean;
-  },
-): string {
-  const key = stateDedupKey(input.rawVomText, input.url);
-  for (const entry of registry.values()) {
-    if (entry.contentHash === key) return entry.id;
-  }
-  const id = nextStateId();
-  registry.set(id, {
-    id,
-    url: input.url,
-    ...(input.title ? { title: input.title } : {}),
-    truncated: input.truncated ?? false,
-    contentHash: key,
-    rawVomText: input.rawVomText,
-    stepsHere: [],
-  });
-  return id;
-}
-
-function toSelection(
-  values: string[],
-  labels?: string[],
-): Array<{ value: string; label?: string }> {
-  return values.map((value, index) => ({
-    value,
-    ...(labels?.[index] ? { label: labels[index] } : {}),
-  }));
-}
-
-function navigateCauseForDraft(
-  draft: Extract<DraftTraceStep, { op: "navigate" }>,
-): NavigationCause {
-  if (draft.cause) return draft.cause;
-  const mapped = mapNavigationCause({
-    transitionType: draft.transitionType,
-    transitionQualifiers: draft.transitionQualifiers,
-  });
-  return mapped ?? "browser";
-}
-
-function toV3Step(draft: DraftTraceStep, id: number): Step | null {
-  if (!shouldIncludeDraft(draft)) return null;
-  // An action the user really performed must not disappear because one of its
-  // observations is missing — a capture racing a document swap is routine, and
-  // the last step of a recording is the most exposed to it. Fall back to the
-  // endpoint we do know; only a step with no observation at all has nothing to
-  // point at.
-  const preState = draft.preStateId ?? draft.postStateId;
-  const postState = draft.postStateId ?? draft.preStateId;
-  if (!preState || !postState) return null;
-
-  const common = { id, state: preState, result: { state: postState } };
-
-  switch (draft.op) {
-    case "navigate":
-      return {
-        op: "navigate",
-        ...common,
-        to: draft.url,
-        cause: navigateCauseForDraft(draft),
-      };
-    case "click":
-      return { op: "click", ...common, target: draft.target };
-    case "hover":
-      return { op: "hover", ...common, target: draft.target };
-    case "fill":
-      return {
-        op: "fill",
-        ...common,
-        target: draft.target,
-        value: draft.value,
-        commit: draft.commit ?? DEFAULT_FILL_COMMIT,
-        ...(draft.redacted ? { redacted: true } : {}),
-      };
-    case "press":
-      return {
-        op: "press",
-        ...common,
-        key: draft.key,
-        ...(draft.target ? { target: draft.target } : {}),
-        ...(draft.modifiers?.length ? { modifiers: draft.modifiers } : {}),
-      };
-    case "select":
-      return {
-        op: "select",
-        ...common,
-        target: draft.target,
-        selection: toSelection(draft.values, draft.labels),
-      };
-    case "scroll":
-      return { op: "scroll", ...common };
-  }
-}
-
-export interface ReducedTrace {
-  states: TraceState[];
-  steps: Step[];
-  /**
-   * Capture-order draft id (`draftIndex + 1`) → published step id. Collapsed
-   * and dropped drafts make the two diverge, so anything the recorder tagged
-   * with a draft id (`steps_here`, inline annotations) must be remapped.
-   */
-  stepIdByDraftId: Map<number, number>;
-}
-
-/**
- * Compile capture drafts into record-only trace v3 steps.
- * Drafts must already carry `preStateId` / `postStateId` from the recorder.
- */
-export function reduceTraceSteps(
-  steps: DraftTraceStep[],
-  registry: Map<string, StateRegistryEntry>,
-): ReducedTrace {
-  const collapsed = collapseNavigations(steps);
-  const out: Step[] = [];
-  const stepIdByDraftId = new Map<number, number>();
-  let id = 1;
-  for (const { draft, draftIds } of collapsed) {
-    const step = toV3Step(draft, id);
-    if (!step) continue;
-    out.push(step);
-    for (const draftId of draftIds) stepIdByDraftId.set(draftId, id);
-    id += 1;
-  }
-  const states = [...registry.values()].map(
-    ({ contentHash: _hash, rawVomText: _body, stepsHere: _steps, ...state }) => state,
-  );
-  return { states, steps: out, stepIdByDraftId };
-}
-
-export function resolveTraceStartUrl(
-  drafts: DraftTraceStep[],
-  startUrl?: string,
-  states?: TraceState[],
-): string {
-  if (startUrl) return startUrl;
-  const navigate = drafts.find((step): step is Extract<DraftTraceStep, { op: "navigate" }> => {
-    return step.op === "navigate";
-  });
-  if (navigate) return navigate.url;
-  for (const draft of drafts) {
-    if ("page_url" in draft && draft.page_url) return draft.page_url;
-  }
-  return states?.[0]?.url ?? "about:blank";
-}
-
-export type { FillCommit, NavigationCause };
-/** Test seam */
-export { resetStateIdCounterForTests };

apps/extension/src/session-manager/__tests__/ref-store.test.ts

@@ -13,6 +13,8 @@ describe("RefStore", () => {
       backendNodeId: 42,
       tabId: 7,
       generation: 0,
+      kind: "dom",
+      capabilities: ["interact", "screenshot"],
     });
     expect(s.size()).toBe(1);
     expect(s.isEmpty()).toBe(false);
@@ -33,4 +35,46 @@ describe("RefStore", () => {
     expect(s.resolve("@e11")).toBe(11);
     expect(s.size()).toBe(2);
   });
+
+  it("preserves the child CDP session as part of ref identity", () => {
+    const s = new RefStore();
+    s.set("e1", 42, {
+      tabId: 7,
+      frameId: "child-frame",
+      cdpSessionId: "child-session",
+    });
+    expect(s.resolveEntry("e1")).toMatchObject({
+      backendNodeId: 42,
+      tabId: 7,
+      frameId: "child-frame",
+      cdpSessionId: "child-session",
+    });
+  });
+
+  it("preserves explicit capabilities for visual surface refs", () => {
+    const s = new RefStore();
+    s.set("e1", 42, {
+      tabId: 7,
+      kind: "surface",
+      capabilities: ["screenshot"],
+    });
+
+    expect(s.resolveEntry("e1")).toMatchObject({
+      kind: "surface",
+      capabilities: ["screenshot"],
+    });
+  });
+
+  it("never grants ordinary interaction to a surface through defaults or caller input", () => {
+    const s = new RefStore();
+    s.set("e1", 42, { tabId: 7, kind: "surface" });
+    s.set("e2", 43, {
+      tabId: 7,
+      kind: "surface",
+      capabilities: ["interact", "screenshot"],
+    });
+
+    expect(s.resolveEntry("e1")?.capabilities).toEqual(["screenshot"]);
+    expect(s.resolveEntry("e2")?.capabilities).toEqual(["screenshot"]);
+  });
 });

apps/extension/src/session-manager/ref-store.ts

@@ -1,9 +1,12 @@
+import type { Rect } from "@browser-skill/vom";
+
 /**
- * Per-session map from `@e<N>` snapshot refs to CDP `backendNodeId`.
+ * Per-session map from `@e<N>` snapshot refs to a CDP node address.
  *
  * Each fresh `tool.snapshot` resets the store: M6 will call
- * `replace(...)` with the new ref → backendNodeId pairs. Tools like
- * `tool.click` consume the store via `resolve("@e1")`.
+ * `replace(...)` with the new ref → node address pairs. A node address
+ * includes the owning flat CDP session/frame when the element lives in
+ * an OOPIF; tools resolve live geometry from that identity at call time.
  *
  * Refs are session-scoped (§7): looking up a ref in the wrong session
  * returns `null`, never silently leaks. Storing values in different
@@ -11,14 +14,39 @@
  * inside the `SessionContext`.
  */
 export type BackendNodeId = number;
+export type RefCapability = "interact" | "screenshot";
+export type RefTargetKind = "dom" | "surface";
 
 export interface RefEntry {
   backendNodeId: BackendNodeId;
   tabId: number | null;
+  frameId?: string;
+  cdpSessionId?: string;
+  visibleRect?: Rect;
+  kind: RefTargetKind;
+  capabilities: RefCapability[];
   generation: number;
 }
 
-type RefInput = BackendNodeId | { backendNodeId: BackendNodeId; tabId: number };
+export type RefInput =
+  | BackendNodeId
+  | {
+      backendNodeId: BackendNodeId;
+      tabId: number;
+      frameId?: string;
+      cdpSessionId?: string;
+      visibleRect?: Rect;
+      kind?: RefTargetKind;
+      capabilities?: RefCapability[];
+    };
+
+function capabilitiesFor(
+  kind: RefTargetKind,
+  capabilities: RefCapability[] | undefined,
+): RefCapability[] {
+  if (kind === "surface") return ["screenshot"];
+  return capabilities ?? ["interact", "screenshot"];
+}
 
 export class RefStore {
   private readonly map = new Map<string, RefEntry>();
@@ -44,7 +72,7 @@ export class RefStore {
   }
 
   /**
-   * Replace the entire store with a new ref → backendNodeId mapping.
+   * Replace the entire store with a new ref → CDP node identity mapping.
    * Used after every fresh `tool.snapshot`.
    */
   replace(entries: Iterable<readonly [string, RefInput]>): void {
@@ -53,10 +81,27 @@ export class RefStore {
     for (const [ref, input] of entries) this.map.set(normaliseRef(ref), this.entry(input));
   }
 
-  set(ref: string, id: BackendNodeId, opts: { tabId?: number } = {}): void {
+  set(
+    ref: string,
+    id: BackendNodeId,
+    opts: {
+      tabId?: number;
+      frameId?: string;
+      cdpSessionId?: string;
+      visibleRect?: Rect;
+      kind?: RefTargetKind;
+      capabilities?: RefCapability[];
+    } = {},
+  ): void {
+    const kind = opts.kind ?? "dom";
     this.map.set(normaliseRef(ref), {
       backendNodeId: id,
       tabId: opts.tabId ?? null,
+      ...(opts.frameId ? { frameId: opts.frameId } : {}),
+      ...(opts.cdpSessionId ? { cdpSessionId: opts.cdpSessionId } : {}),
+      ...(opts.visibleRect ? { visibleRect: opts.visibleRect } : {}),
+      kind,
+      capabilities: capabilitiesFor(kind, opts.capabilities),
       generation: this.generation,
     });
   }
@@ -74,12 +119,20 @@ export class RefStore {
       return {
         backendNodeId: input,
         tabId: null,
+        kind: "dom",
+        capabilities: ["interact", "screenshot"],
         generation: this.generation,
       };
     }
+    const kind = input.kind ?? "dom";
     return {
       backendNodeId: input.backendNodeId,
       tabId: input.tabId,
+      ...(input.frameId ? { frameId: input.frameId } : {}),
+      ...(input.cdpSessionId ? { cdpSessionId: input.cdpSessionId } : {}),
+      ...(input.visibleRect ? { visibleRect: input.visibleRect } : {}),
+      kind,
+      capabilities: capabilitiesFor(kind, input.capabilities),
       generation: this.generation,
     };
   }

apps/extension/src/shared/recording-document-identity.ts

@@ -0,0 +1,22 @@
+export const RECORD_DOCUMENT_ATTRIBUTE = "data-bsk-record-document";
+
+const RECORDING_DOCUMENT_ID_PATTERN =
+  /^[0-9a-f]{8}-[0-9a-f]{4}-4[0-9a-f]{3}-[89ab][0-9a-f]{3}-[0-9a-f]{12}$/i;
+const RECORDING_DOCUMENT_MARKER_PREFIX = "bsk:";
+
+export function recordingDocumentMarkerValue(identity: string): string {
+  if (!RECORDING_DOCUMENT_ID_PATTERN.test(identity)) {
+    throw new TypeError("recording Document identity must be a random UUID");
+  }
+  return `${RECORDING_DOCUMENT_MARKER_PREFIX}${identity}`;
+}
+
+/** Read BrowserSkill's opaque per-Document recording identity from captured attributes. */
+export function readRecordingDocumentIdentity(
+  attrs: Readonly<Record<string, string>>,
+): string | undefined {
+  const marker = attrs[RECORD_DOCUMENT_ATTRIBUTE];
+  if (!marker?.startsWith(RECORDING_DOCUMENT_MARKER_PREFIX)) return undefined;
+  const identity = marker.slice(RECORDING_DOCUMENT_MARKER_PREFIX.length);
+  return identity && RECORDING_DOCUMENT_ID_PATTERN.test(identity) ? identity : undefined;
+}

apps/extension/src/tools/__tests__/borrow-confirmation.test.ts

@@ -612,4 +612,14 @@ describe("isInjectableContentScriptUrl", () => {
     ).toBe(false);
     expect(isInjectableContentScriptUrl("https://chromewebstore.google.com/")).toBe(false);
   });
+
+  it("rejects the Edge Add-ons store", () => {
+    expect(
+      isInjectableContentScriptUrl(
+        "https://microsoftedge.microsoft.com/addons/detail/browserskill/emacgiaaaiojkkpkddmmdfhmokgmnikg",
+      ),
+    ).toBe(false);
+    // Non-store paths on the same host stay injectable.
+    expect(isInjectableContentScriptUrl("https://microsoftedge.microsoft.com/")).toBe(true);
+  });
 });

apps/extension/src/tools/__tests__/dispatcher.test.ts

@@ -8,6 +8,7 @@ import type {
   RequestFrame,
 } from "@/transport/types";
 import { ToolDispatcher } from "../dispatcher";
+import { resetBrowserObservationForTests, setBrowserObservationAttachForTests } from "../record";
 
 type TestDispatcherCdp = NonNullable<ConstructorParameters<typeof ToolDispatcher>[0]["cdp"]>;
 
@@ -44,9 +45,72 @@ function makeRequest(method: string, params: unknown): RequestFrame {
 
 describe("ToolDispatcher", () => {
   afterEach(() => {
+    resetBrowserObservationForTests();
     vi.unstubAllGlobals();
   });
 
+  it("uses the configured recording runtime for start and stop", async () => {
+    const { transport, sent, deliver } = fakeTransport();
+    const sessions = new SessionManager({
+      agentWindow: {
+        create: vi.fn(async () => 4242),
+        remove: vi.fn(),
+        ensureActiveTab: vi.fn(async () => {}),
+      },
+    });
+    await sessions.start("aa11");
+    const tab = {
+      id: 7,
+      windowId: 4242,
+      active: true,
+      status: "complete",
+      url: "https://example.com/start",
+    } as chrome.tabs.Tab;
+    const frameCoordinator = {
+      begin: vi.fn(),
+      armTab: vi.fn(async () => true),
+      sourceFor: vi.fn(() => null),
+      stop: vi.fn(async () => true),
+      cancel: vi.fn(),
+    };
+    const sendToTab = vi.fn(async () => ({ ok: true }));
+    setBrowserObservationAttachForTests(
+      () => () => {},
+      () => () => {},
+    );
+    const dispatcher = new ToolDispatcher({
+      transport,
+      sessions,
+      recording: {
+        tabsApi: {
+          get: vi.fn(async () => tab),
+          query: vi.fn(async () => [tab]),
+        },
+        frameCoordinator,
+        sendToTab,
+      },
+    });
+    dispatcher.start();
+
+    deliver(
+      makeRequest("tool.record_start", {
+        session_id: "aa11",
+        url: "https://example.com/start",
+      }),
+    );
+    await vi.waitFor(() => expect(sent).toHaveLength(1));
+
+    expect(sent[0]).toMatchObject({ result: { tab_id: 7, recording: true } });
+    expect(frameCoordinator.begin).toHaveBeenCalledOnce();
+    expect(frameCoordinator.armTab).toHaveBeenCalledWith(expect.any(String), 7);
+
+    deliver({ id: "r-2", method: "tool.record_stop", params: { session_id: "aa11" } });
+    await vi.waitFor(() => expect(sent).toHaveLength(2));
+
+    expect(frameCoordinator.stop).toHaveBeenCalledOnce();
+    expect(sent[1]).toMatchObject({ id: "r-2", result: { trace: { steps: [] } } });
+  });
+
   it("routes tool.session_start to the SessionManager and replies with the window id", async () => {
     const { transport, sent, deliver } = fakeTransport();
     const sessions = new SessionManager({
@@ -203,6 +267,55 @@ describe("ToolDispatcher", () => {
     expect(sent[0]).toEqual({ id: "r-1", result: {} });
   });
 
+  it("forwards daemon cancellation into an in-flight session stop", async () => {
+    const { transport, sent, deliver } = fakeTransport();
+    const remove = vi.fn(async () => {});
+    const sessions = new SessionManager({
+      agentWindow: {
+        create: vi.fn(async () => 4242),
+        remove,
+        ensureActiveTab: vi.fn(async () => {}),
+      },
+    });
+    await sessions.start("aa11");
+    let finishDetach: () => void = () => {};
+    const detach = new Promise<void>((resolve) => {
+      finishDetach = resolve;
+    });
+    const cdp = {
+      send: vi.fn(),
+      detachSession: vi.fn(() => detach),
+      ensureNetworkCapture: vi.fn(async () => {}),
+      networkEntriesSince: vi.fn(() => ({
+        tab_id: 7,
+        entries: [],
+        next_since: 0,
+        truncated: false,
+      })),
+      setDeviceMetricsOverride: vi.fn(async () => {}),
+      clearDeviceMetricsOverride: vi.fn(async () => {}),
+      setUserAgentOverride: vi.fn(async () => {}),
+      setTouchEmulationEnabled: vi.fn(async () => {}),
+    };
+    const dispatcher = new ToolDispatcher({ transport, sessions, cdp: cdp as TestDispatcherCdp });
+    dispatcher.start();
+
+    deliver(makeRequest("tool.session_stop", { session_id: "aa11" }));
+    await vi.waitFor(() => expect(cdp.detachSession).toHaveBeenCalledWith("aa11"));
+    deliver({ id: "cancel-stop", method: "cancel", params: { rpc_id: "r-1" } });
+    await flushMicrotasks();
+    finishDetach();
+    await flushMicrotasks();
+
+    expect(sent).toContainEqual({ id: "cancel-stop", result: { cancelled: true } });
+    expect(sent).toContainEqual({
+      id: "r-1",
+      error: expect.objectContaining({ code: "cancelled" }),
+    });
+    expect(sessions.has("aa11")).toBe(true);
+    expect(remove).not.toHaveBeenCalled();
+  });
+
   it("returns not_found when stopping an unknown session", async () => {
     const { transport, sent, deliver } = fakeTransport();
     const sessions = new SessionManager({

apps/extension/src/tools/__tests__/element-geometry.test.ts

@@ -1,24 +1,54 @@
 import { describe, expect, it } from "vitest";
-import { nodeBoundingRect, quadBoundingRect } from "../element-geometry";
+import { nodeContentRegion } from "../element-geometry";
 import type { CdpRunner } from "../shared";
 
-describe("quadBoundingRect", () => {
-  it("computes axis-aligned bounds from an 8-double quad", () => {
-    expect(quadBoundingRect([10, 20, 110, 20, 110, 60, 10, 60])).toEqual({
-      x: 10,
-      y: 20,
-      width: 100,
-      height: 40,
+describe("nodeContentRegion", () => {
+  it("rejects invalid content quads before using the box model fallback", async () => {
+    const send = async (_tabId: number, method: string) => {
+      if (method === "DOM.getContentQuads") return { quads: [[0, 0, 0, 0, 0, 0, 0, 0]] };
+      if (method === "DOM.getBoxModel") {
+        return { model: { content: [5, 10, 25, 10, 25, 30, 5, 30] } };
+      }
+      throw new Error(`unexpected CDP call ${method}`);
+    };
+
+    await expect(nodeContentRegion({ send: send as CdpRunner["send"] }, 7, 555)).resolves.toEqual([
+      [
+        { x: 5, y: 10 },
+        { x: 25, y: 10 },
+        { x: 25, y: 30 },
+        { x: 5, y: 30 },
+      ],
+    ]);
+  });
+
+  it("preserves every visible content fragment", async () => {
+    const send = async () => ({
+      quads: [
+        [10, 20, 60, 20, 60, 40, 10, 40],
+        [15, 50, 90, 50, 90, 80, 15, 80],
+      ],
     });
+    const cdp: CdpRunner = {
+      send: send as CdpRunner["send"],
+    };
+
+    await expect(nodeContentRegion(cdp, 7, 555)).resolves.toEqual([
+      [
+        { x: 10, y: 20 },
+        { x: 60, y: 20 },
+        { x: 60, y: 40 },
+        { x: 10, y: 40 },
+      ],
+      [
+        { x: 15, y: 50 },
+        { x: 90, y: 50 },
+        { x: 90, y: 80 },
+        { x: 15, y: 80 },
+      ],
+    ]);
   });
 
-  it("returns null for degenerate quads", () => {
-    expect(quadBoundingRect([0, 0, 0, 0, 0, 0, 0, 0])).toBeNull();
-    expect(quadBoundingRect([1, 2, 3])).toBeNull();
-  });
-});
-
-describe("nodeBoundingRect", () => {
   it("falls back to visible descendant bounds for zero-size containers", async () => {
     const calls: Array<{ method: string; params?: object }> = [];
     const send = async (_tabId: number, method: string, params?: object) => {
@@ -38,12 +68,14 @@ describe("nodeBoundingRect", () => {
     };
     const cdp = { send: send as CdpRunner["send"] };
 
-    await expect(nodeBoundingRect(cdp, 7, 555)).resolves.toEqual({
-      x: 242,
-      y: 468,
-      width: 769,
-      height: 180,
-    });
+    await expect(nodeContentRegion(cdp, 7, 555)).resolves.toEqual([
+      [
+        { x: 242, y: 468 },
+        { x: 1011, y: 468 },
+        { x: 1011, y: 648 },
+        { x: 242, y: 648 },
+      ],
+    ]);
     expect(calls.map((c) => c.method)).toEqual([
       "DOM.getContentQuads",
       "DOM.getBoxModel",

apps/extension/src/tools/__tests__/frame-geometry.test.ts

@@ -0,0 +1,286 @@
+import { describe, expect, it, vi } from "vitest";
+import { resolveFrameProjection, resolveNodeGeometry } from "../frame-geometry";
+import {
+  clipPolygon,
+  polygonArea,
+  polygonBounds,
+  projectAndClipRegion,
+  projectUnitPoint,
+  type Quad,
+  rectPolygon,
+  regionBounds,
+} from "../geometry";
+import type { CdpRunner } from "../shared";
+
+describe("frame geometry projection", () => {
+  it("keeps region bounds separate from polygon area", () => {
+    const region = [
+      rectPolygon({ x: 0, y: 0, w: 10, h: 10 }),
+      rectPolygon({ x: 20, y: 5, w: 5, h: 5 }),
+    ];
+    expect(regionBounds(region)).toEqual({ x: 0, y: 0, width: 25, height: 10 });
+    expect(polygonArea(region[0])).toBe(100);
+  });
+
+  it("maps through the iframe content box instead of its border box", () => {
+    const contentQuad: Quad = [
+      { x: 204, y: 306 },
+      { x: 604, y: 306 },
+      { x: 604, y: 506 },
+      { x: 204, y: 506 },
+    ];
+    const projected = projectAndClipRegion(
+      [rectPolygon({ x: 10, y: 20, w: 100, h: 40 })],
+      [{ sourceViewport: { width: 200, height: 100 }, destinationQuad: contentQuad }],
+      { width: 1000, height: 800 },
+    );
+    expect(polygonBounds(projected.flat())).toEqual({ x: 224, y: 346, width: 200, height: 80 });
+  });
+
+  it("uses a projective mapping for perspective-transformed iframe quads", () => {
+    const quad: Quad = [
+      { x: 0, y: 0 },
+      { x: 200, y: 20 },
+      { x: 180, y: 140 },
+      { x: 20, y: 100 },
+    ];
+    expect(projectUnitPoint({ x: 0, y: 0 }, quad)).toEqual(quad[0]);
+    expect(projectUnitPoint({ x: 1, y: 0 }, quad)).toEqual(quad[1]);
+    expect(projectUnitPoint({ x: 1, y: 1 }, quad)).toEqual(quad[2]);
+    expect(projectUnitPoint({ x: 0, y: 1 }, quad)).toEqual(quad[3]);
+  });
+
+  it("clips against the iframe polygon rather than its bounding box", () => {
+    const frameQuad: Quad = [
+      { x: 50, y: 0 },
+      { x: 100, y: 50 },
+      { x: 50, y: 100 },
+      { x: 0, y: 50 },
+    ];
+    expect(clipPolygon(rectPolygon({ x: 0, y: 0, w: 10, h: 10 }), frameQuad)).toEqual([]);
+  });
+
+  it("does not reapply same-target frame transforms around an OOPIF", async () => {
+    const send = vi.fn(async (_tabId, method, params) => {
+      if (method === "Page.getLayoutMetrics") {
+        return { cssLayoutViewport: { clientWidth: 800, clientHeight: 600 } };
+      }
+      if (method === "DOM.getBoxModel") {
+        const backendNodeId = (params as { backendNodeId?: number }).backendNodeId;
+        return backendNodeId === 20
+          ? { model: { content: [110, 70, 310, 70, 310, 170, 110, 170] } }
+          : { model: { content: [100, 50, 400, 50, 400, 250, 100, 250] } };
+      }
+      throw new Error(`unexpected root command ${method}`);
+    });
+    const cdp: CdpRunner = {
+      send: send as CdpRunner["send"],
+      sendToTarget: vi.fn(async (_target, method) => {
+        if (method === "Page.getLayoutMetrics") {
+          return { cssLayoutViewport: { clientWidth: 200, clientHeight: 100 } };
+        }
+        throw new Error(`unexpected child command ${method}`);
+      }) as CdpRunner["sendToTarget"],
+    };
+    const projection = await resolveFrameProjection(
+      cdp,
+      {
+        rootFrameId: "main",
+        frames: [
+          { frameId: "main", target: { tabId: 4 } },
+          {
+            frameId: "same-process-parent",
+            parentFrameId: "main",
+            ownerBackendNodeId: 10,
+            target: { tabId: 4 },
+          },
+          {
+            frameId: "oopif",
+            parentFrameId: "same-process-parent",
+            ownerBackendNodeId: 20,
+            target: { tabId: 4, sessionId: "oopif-session" },
+          },
+        ],
+      },
+      "oopif",
+    );
+
+    expect(projection?.edges).toEqual([
+      {
+        sourceViewport: { width: 200, height: 100 },
+        destinationQuad: [
+          { x: 110, y: 70 },
+          { x: 310, y: 70 },
+          { x: 310, y: 170 },
+          { x: 110, y: 170 },
+        ],
+        destinationClips: [
+          [
+            { x: 100, y: 50 },
+            { x: 400, y: 50 },
+            { x: 400, y: 250 },
+            { x: 100, y: 250 },
+          ],
+        ],
+      },
+    ]);
+  });
+
+  it("clips live OOPIF geometry at every frame boundary", async () => {
+    const cdp: CdpRunner = {
+      send: vi.fn(async (_tabId, method) => {
+        if (method === "Page.getLayoutMetrics") {
+          return { cssLayoutViewport: { clientWidth: 800, clientHeight: 600 } };
+        }
+        if (method === "DOM.getBoxModel") {
+          return { model: { content: [100, 100, 300, 100, 300, 200, 100, 200] } };
+        }
+        throw new Error(`unexpected root command ${method}`);
+      }) as CdpRunner["send"],
+      sendToTarget: vi.fn(async (_target, method) => {
+        if (method === "Page.getLayoutMetrics") {
+          return { cssLayoutViewport: { clientWidth: 200, clientHeight: 100 } };
+        }
+        if (method === "DOM.getContentQuads") {
+          return { quads: [[-50, -20, 250, -20, 250, 120, -50, 120]] };
+        }
+        throw new Error(`unexpected child command ${method}`);
+      }) as CdpRunner["sendToTarget"],
+      getFrameGraph: vi.fn(async () => ({
+        rootFrameId: "main",
+        frames: [
+          { frameId: "main", target: { tabId: 4 } },
+          {
+            frameId: "child",
+            parentFrameId: "main",
+            ownerBackendNodeId: 10,
+            target: { tabId: 4, sessionId: "child-session" },
+          },
+        ],
+      })),
+    };
+
+    const geometry = await resolveNodeGeometry(cdp, 4, {
+      target: { tabId: 4, sessionId: "child-session" },
+      frameId: "child",
+      backendNodeId: 101,
+    });
+
+    expect(geometry).toMatchObject({
+      topBounds: { x: 100, y: 100, width: 200, height: 100 },
+      actionPoint: { x: 200, y: 150 },
+    });
+  });
+
+  it("reuses one frame graph while scrolling and projecting live geometry", async () => {
+    const getFrameGraph = vi.fn(async () => ({
+      rootFrameId: "main",
+      frames: [
+        { frameId: "main", target: { tabId: 4 } },
+        {
+          frameId: "child",
+          parentFrameId: "main",
+          ownerBackendNodeId: 10,
+          target: { tabId: 4, sessionId: "child-session" },
+        },
+      ],
+    }));
+    const cdp: CdpRunner = {
+      send: vi.fn(async (_tabId, method) => {
+        if (method === "DOM.scrollIntoViewIfNeeded") return {};
+        if (method === "Page.getLayoutMetrics") {
+          return { cssLayoutViewport: { clientWidth: 800, clientHeight: 600 } };
+        }
+        if (method === "DOM.getBoxModel") {
+          return { model: { content: [100, 100, 300, 100, 300, 200, 100, 200] } };
+        }
+        throw new Error(`unexpected root command ${method}`);
+      }) as CdpRunner["send"],
+      sendToTarget: vi.fn(async (_target, method) => {
+        if (method === "DOM.scrollIntoViewIfNeeded") return {};
+        if (method === "Page.getLayoutMetrics") {
+          return { cssLayoutViewport: { clientWidth: 200, clientHeight: 100 } };
+        }
+        if (method === "DOM.getContentQuads") {
+          return { quads: [[0, 0, 20, 0, 20, 20, 0, 20]] };
+        }
+        throw new Error(`unexpected child command ${method}`);
+      }) as CdpRunner["sendToTarget"],
+      getFrameGraph,
+    };
+
+    await expect(
+      resolveNodeGeometry(
+        cdp,
+        4,
+        {
+          target: { tabId: 4, sessionId: "child-session" },
+          frameId: "child",
+          backendNodeId: 101,
+        },
+        { scrollIntoView: true },
+      ),
+    ).resolves.not.toMatchObject({ code: expect.any(String) });
+    expect(getFrameGraph).toHaveBeenCalledOnce();
+  });
+
+  it("clips live geometry to same-target iframe ancestors", async () => {
+    const cdp: CdpRunner = {
+      send: vi.fn(async (_tabId, method) => {
+        if (method === "Page.getLayoutMetrics") {
+          return { cssLayoutViewport: { clientWidth: 800, clientHeight: 600 } };
+        }
+        if (method === "DOM.getBoxModel") {
+          return { model: { content: [100, 100, 300, 100, 300, 200, 100, 200] } };
+        }
+        if (method === "DOM.getContentQuads") {
+          return { quads: [[50, 50, 350, 50, 350, 250, 50, 250]] };
+        }
+        throw new Error(`unexpected ${method}`);
+      }) as CdpRunner["send"],
+      getFrameGraph: vi.fn(async () => ({
+        rootFrameId: "main",
+        frames: [
+          { frameId: "main", target: { tabId: 4 } },
+          {
+            frameId: "child",
+            parentFrameId: "main",
+            ownerBackendNodeId: 10,
+            target: { tabId: 4 },
+          },
+        ],
+      })),
+    };
+
+    const geometry = await resolveNodeGeometry(cdp, 4, {
+      target: { tabId: 4 },
+      frameId: "child",
+      backendNodeId: 101,
+    });
+
+    expect(geometry).toMatchObject({
+      topBounds: { x: 100, y: 100, width: 200, height: 100 },
+      actionPoint: { x: 200, y: 150 },
+    });
+  });
+
+  it("fails closed when an OOPIF frame graph is unavailable", async () => {
+    const cdp: CdpRunner = {
+      send: vi.fn(),
+      sendToTarget: vi.fn(async (_target, method) => {
+        if (method === "DOM.getContentQuads") {
+          return { quads: [[0, 0, 20, 0, 20, 20, 0, 20]] };
+        }
+        throw new Error(`unexpected ${method}`);
+      }) as CdpRunner["sendToTarget"],
+    };
+
+    await expect(
+      resolveNodeGeometry(cdp, 4, {
+        target: { tabId: 4, sessionId: "child-session" },
+        frameId: "child",
+        backendNodeId: 101,
+      }),
+    ).resolves.toMatchObject({ code: "cdp_failed" });
+  });
+});

apps/extension/src/tools/__tests__/human-loop.test.ts

@@ -1,5 +1,6 @@
 import { afterEach, describe, expect, it, vi } from "vitest";
 import { SessionManager } from "@/session-manager/manager";
+import { RefStore } from "@/session-manager/ref-store";
 import type { RequestHelpParams } from "@/transport/types";
 import { handleRequestHelp, type RequestHelpDeps, resetHelpLifecycleForTests } from "../human-loop";
 
@@ -50,11 +51,10 @@ function installHelpLifecycleChrome() {
 }
 
 function fakeManager(sessionId: string, agentWindowId: number, tabId: number) {
+  const refStore = new RefStore();
   const mgr = {
     get: (id: string) =>
-      id === sessionId
-        ? { sessionId, agentWindowId, refStore: { resolve: () => null }, borrowedTabs: new Map() }
-        : null,
+      id === sessionId ? { sessionId, agentWindowId, refStore, borrowedTabs: new Map() } : null,
     findByWindowId: (wid: number) => (wid === agentWindowId ? { sessionId } : null),
   } as unknown as SessionManager;
   return mgr;
@@ -339,26 +339,23 @@ describe("handleRequestHelp", () => {
 
     ac.abort();
     await expect(pending).resolves.toMatchObject({ code: "cancelled" });
-    expect(cdpSend).toHaveBeenCalledWith(
-      5,
-      "DOM.querySelectorAll",
-      expect.objectContaining({ selector: "[data-bsk-help]" }),
-    );
-    expect(cdpSend).toHaveBeenCalledWith(
-      6,
-      "DOM.querySelectorAll",
-      expect.objectContaining({ selector: "[data-bsk-help]" }),
+    expect(cdpSend).not.toHaveBeenCalledWith(
+      expect.anything(),
+      "DOM.removeAttribute",
+      expect.anything(),
     );
   });
 
-  it("tags ref targets via CDP and reports them matched", async () => {
+  it("resolves ref targets to explicit viewport rectangles without mutating the page", async () => {
+    const refStore = new RefStore();
+    refStore.set("e1", 42, { tabId: 5 });
     const mgr = {
       get: (id: string) =>
         id === "abcd"
           ? {
               sessionId: "abcd",
               agentWindowId: 99,
-              refStore: { resolve: () => 42 },
+              refStore,
               borrowedTabs: new Map(),
             }
           : null,
@@ -366,7 +363,17 @@ describe("handleRequestHelp", () => {
     } as unknown as SessionManager;
     const deps = baseDeps({
       cdp: {
-        send: vi.fn(async () => ({ object: { objectId: "obj-1" } })),
+        send: vi.fn(async (_tabId, method) => {
+          if (method === "DOM.getDocument") return { root: { nodeId: 1 } };
+          if (method === "DOM.scrollIntoViewIfNeeded") return {};
+          if (method === "DOM.getContentQuads") {
+            return { quads: [[10, 20, 110, 20, 110, 60, 10, 60]] };
+          }
+          if (method === "Page.getLayoutMetrics") {
+            return { cssLayoutViewport: { clientWidth: 1280, clientHeight: 720 } };
+          }
+          throw new Error(`unexpected ${method}`);
+        }),
       } as unknown as RequestHelpDeps["cdp"],
     });
     const res = await handleRequestHelp(
@@ -375,14 +382,97 @@ describe("handleRequestHelp", () => {
       deps,
     );
     const sentMsg = (deps.sendToTab as ReturnType<typeof vi.fn>).mock.calls[0][1];
-    expect(sentMsg.selectors).toContain('[data-bsk-help="0"]');
+    expect(sentMsg.selectors).toEqual([]);
+    expect(sentMsg.rects).toEqual([{ top: 20, left: 10, width: 100, height: 40 }]);
     expect(res).toMatchObject({
       outcome: "continued",
       tab_id: 5,
-      resolved_targets: [{ matched: true, ref: "@e1" }],
+      resolved_targets: [
+        {
+          matched: true,
+          ref: "@e1",
+        },
+      ],
     });
   });
 
+  it("projects an OOPIF ref through its live content quad for the help overlay", async () => {
+    const mgr = {
+      get: (id: string) =>
+        id === "abcd"
+          ? {
+              sessionId: "abcd",
+              agentWindowId: 99,
+              refStore: {
+                resolveEntry: () => ({
+                  backendNodeId: 42,
+                  tabId: 5,
+                  frameId: "child",
+                  cdpSessionId: "child-session",
+                  generation: 1,
+                }),
+              },
+              borrowedTabs: new Map(),
+            }
+          : null,
+      findByWindowId: (wid: number) => (wid === 99 ? { sessionId: "abcd" } : null),
+    } as unknown as SessionManager;
+    const send = vi.fn(async (_tabId, method) => {
+      if (method === "DOM.getDocument") return { root: { nodeId: 1 } };
+      if (method === "DOM.scrollIntoViewIfNeeded") return {};
+      if (method === "DOM.getBoxModel") {
+        return { model: { content: [204, 306, 604, 306, 604, 506, 204, 506] } };
+      }
+      if (method === "Page.getLayoutMetrics") {
+        return { cssLayoutViewport: { clientWidth: 1280, clientHeight: 720 } };
+      }
+      throw new Error(`unexpected root ${method}`);
+    });
+    const sendToTarget = vi.fn(async (_target, method) => {
+      if (method === "DOM.scrollIntoViewIfNeeded") return {};
+      if (method === "DOM.getContentQuads") {
+        return { quads: [[10, 20, 110, 20, 110, 60, 10, 60]] };
+      }
+      if (method === "Page.getLayoutMetrics") {
+        return { cssLayoutViewport: { clientWidth: 200, clientHeight: 100 } };
+      }
+      throw new Error(`unexpected child ${method}`);
+    });
+    const deps = baseDeps({
+      cdp: {
+        send,
+        sendToTarget,
+        getFrameGraph: vi.fn(async () => ({
+          rootFrameId: "main",
+          frames: [
+            { frameId: "main", target: { tabId: 5 } },
+            {
+              frameId: "child",
+              parentFrameId: "main",
+              ownerBackendNodeId: 99,
+              target: { tabId: 5, sessionId: "child-session" },
+            },
+          ],
+        })),
+      } as unknown as NonNullable<RequestHelpDeps["cdp"]>,
+    });
+
+    const res = await handleRequestHelp(
+      mgr,
+      baseParams({ tab_id: 5, targets: [{ ref: "@e1" }] }),
+      deps,
+    );
+
+    const sentMsg = (deps.sendToTab as ReturnType<typeof vi.fn>).mock.calls[0][1];
+    expect(sentMsg.rects).toEqual([{ top: 346, left: 224, width: 200, height: 80 }]);
+    expect(res).toMatchObject({ resolved_targets: [{ matched: true, ref: "@e1" }] });
+    expect(sendToTarget).toHaveBeenCalledWith(
+      { tabId: 5, sessionId: "child-session" },
+      "DOM.getContentQuads",
+      { backendNodeId: 42 },
+    );
+  });
+
   it("reports ref target unmatched when ref does not resolve", async () => {
     const deps = baseDeps();
     const res = await handleRequestHelp(
@@ -398,16 +488,15 @@ describe("handleRequestHelp", () => {
   });
 
   it("reports ref target unmatched when ref is for another tab", async () => {
+    const refStore = new RefStore();
+    refStore.set("e1", 42, { tabId: 4 });
     const mgr = {
       get: (id: string) =>
         id === "abcd"
           ? {
               sessionId: "abcd",
               agentWindowId: 99,
-              refStore: {
-                resolve: (ref: string, opts: { tabId?: number }) =>
-                  ref === "e1" && opts.tabId === 4 ? 42 : null,
-              },
+              refStore,
               borrowedTabs: new Map(),
             }
           : null,

apps/extension/src/tools/__tests__/interaction.test.ts

@@ -30,6 +30,9 @@ function makeFakeCdp(handlers: Record<string, (params: unknown) => unknown>) {
   const sendImpl = async (tabId: number, method: string, params?: object) => {
     sent.push({ tabId, method, params });
     const h = handlers[method];
+    if (!h && method === "Page.getLayoutMetrics") {
+      return { cssLayoutViewport: { clientWidth: 1280, clientHeight: 720 } };
+    }
     if (!h) throw new Error(`unexpected CDP call ${method}`);
     return h(params);
   };
@@ -120,6 +123,29 @@ describe("handleClick", () => {
     expect(fake.cdp.send).not.toHaveBeenCalled();
   });
 
+  it("rejects screenshot-only visual surface refs before issuing CDP calls", async () => {
+    const sm = new SessionManager({ agentWindow: fakeAgentWindow([100]) });
+    const ctx = await sm.start("aa11");
+    ctx.refStore.set("e3", 1234, {
+      tabId: 4,
+      kind: "surface",
+      capabilities: ["screenshot"],
+    });
+    const fake = makeFakeCdp({});
+
+    const res = await handleClick(
+      sm,
+      { session_id: "aa11", ref: "@e3" },
+      { cdp: fake.cdp, tabsApi: fake.tabsApi },
+    );
+
+    expect(res).toMatchObject({
+      code: "permission_denied",
+      data: { reason: "ref_capability_denied", required_capability: "interact" },
+    });
+    expect(fake.cdp.send).not.toHaveBeenCalled();
+  });
+
   it("clicks by ref, computes the quad centre, dispatches three mouse events", async () => {
     const sm = new SessionManager({ agentWindow: fakeAgentWindow([100]) });
     const ctx = await sm.start("aa11");
@@ -164,6 +190,63 @@ describe("handleClick", () => {
     });
   });
 
+  it("resolves frame refs in their CDP session and dispatches input in top coordinates", async () => {
+    const sm = new SessionManager({ agentWindow: fakeAgentWindow([100]) });
+    const ctx = await sm.start("aa11");
+    ctx.refStore.set("e3", 1234, {
+      tabId: 4,
+      frameId: "child-frame",
+      cdpSessionId: "child-session",
+    });
+    const fake = makeFakeCdp({
+      "Input.dispatchMouseEvent": () => ({}),
+      "DOM.scrollIntoViewIfNeeded": () => ({}),
+      "DOM.getBoxModel": () => ({
+        model: { content: [204, 306, 604, 306, 604, 506, 204, 506] },
+      }),
+    });
+    fake.cdp.getFrameGraph = vi.fn(async () => ({
+      rootFrameId: "main",
+      frames: [
+        { frameId: "main", target: { tabId: 4 } },
+        {
+          frameId: "child-frame",
+          parentFrameId: "main",
+          ownerBackendNodeId: 99,
+          target: { tabId: 4, sessionId: "child-session" },
+        },
+      ],
+    }));
+    const targetCalls: Array<{ sessionId?: string; method: string }> = [];
+    fake.cdp.sendToTarget = vi.fn(async (target, method) => {
+      targetCalls.push({ sessionId: target.sessionId, method });
+      if (method === "DOM.scrollIntoViewIfNeeded") return {};
+      if (method === "DOM.getContentQuads") {
+        return { quads: [[10, 20, 110, 20, 110, 60, 10, 60]] };
+      }
+      if (method === "Page.getLayoutMetrics") {
+        return { cssLayoutViewport: { clientWidth: 200, clientHeight: 100 } };
+      }
+      throw new Error(`unexpected child CDP call ${method}`);
+    }) as CdpRunner["sendToTarget"];
+
+    const res = await handleClick(
+      sm,
+      { session_id: "aa11", ref: "@e3" },
+      { cdp: fake.cdp, tabsApi: fake.tabsApi },
+    );
+
+    if ("code" in res) throw new Error(`unexpected error: ${JSON.stringify(res)}`);
+    // The iframe content box starts at (204, 306) after its border and is scaled 2x.
+    expect(res).toMatchObject({ x: 324, y: 386 });
+    expect(targetCalls).toEqual([
+      { sessionId: "child-session", method: "DOM.scrollIntoViewIfNeeded" },
+      { sessionId: "child-session", method: "DOM.getContentQuads" },
+      { sessionId: "child-session", method: "Page.getLayoutMetrics" },
+    ]);
+    expect(fake.sent.filter((call) => call.method === "Input.dispatchMouseEvent")).toHaveLength(3);
+  });
+
   it("enables overlay bypass before mouse events when overlay blocks the click point", async () => {
     const order: string[] = [];
     const bypassOverlay = vi.fn(async (_tabId: number, enabled: boolean) => {

apps/extension/src/tools/__tests__/record-steps.test.ts

@@ -1,19 +1,11 @@
 import { afterEach, describe, expect, it, vi } from "vitest";
-import {
-  RECORD_FINISH,
-  RECORD_QUERY,
-  RECORD_START,
-  RECORD_STEP,
-  RECORD_STOP,
-} from "@/lib/record-bridge";
-import { resetStateIdCounterForTests } from "@/lib/trace-reducer";
+import { RECORD_FINISH, RECORD_START, RECORD_STEP, RECORD_STOP } from "@/lib/record-bridge";
 import type { SessionManager } from "@/session-manager/manager";
 import type { CdpRunner } from "@/tools/shared";
 import { EXTENSION_VERSION } from "@/transport/handshake";
 import type { RecordedTrace, RecordStopResult, TraceV3 } from "@/transport/types";
 import {
   attachRecordFinishListener,
-  attachRecordQueryListener,
   attachRecordStepListener,
   handleRecordStart,
   handleRecordStop,
@@ -23,7 +15,13 @@ import {
 const AGENT_WINDOW_ID = 100;
 const TAB_ID = 4;
 const START_URL = "https://example.com/";
-const RECORD_START_V3 = { session_id: "abcd", url: START_URL, trace_version: 3 as const };
+const RECORD_START_V3 = {
+  session_id: "abcd",
+  url: START_URL,
+  trace_version: 3 as const,
+  supports_tab_switch_steps: true,
+};
+const stepSequenceByProducer = new Map<string, number>();
 
 function asTraceV3(trace: RecordedTrace): TraceV3 {
   if (!("version" in trace)) {
@@ -55,6 +53,7 @@ function chromeEvent<T extends (...args: never[]) => unknown>() {
 
 function installChrome() {
   const runtimeOnMessage = chromeEvent<RuntimeListener>();
+  const tabsOnActivated = chromeEvent<(activeInfo: chrome.tabs.TabActiveInfo) => unknown>();
   const webNavigationOnCompleted =
     chromeEvent<(details: chrome.webNavigation.WebNavigationFramedCallbackDetails) => unknown>();
   const webNavigationOnCommitted =
@@ -64,7 +63,7 @@ function installChrome() {
   vi.stubGlobal("chrome", {
     runtime: { onMessage: runtimeOnMessage },
     tabs: {
-      onActivated: chromeEvent(),
+      onActivated: tabsOnActivated,
       onCreated: chromeEvent(),
       onUpdated: chromeEvent(),
     },
@@ -73,7 +72,7 @@ function installChrome() {
       onCommitted: webNavigationOnCommitted,
     },
   });
-  return { runtimeOnMessage, webNavigationOnCompleted, webNavigationOnCommitted };
+  return { runtimeOnMessage, tabsOnActivated, webNavigationOnCompleted, webNavigationOnCommitted };
 }
 
 function fakeManager() {
@@ -161,9 +160,9 @@ function makeFakeCdp(
       return { result: { value: { idleMs, readyState: "complete" } } };
     },
     "Accessibility.enable": () => ({}),
-    "Accessibility.getFullAXTree": (_params, tabId) => {
+    "Accessibility.getFullAXTree": (_params, _tabId) => {
       if (failing) throw new Error("Execution context was destroyed");
-      const tabTree = options?.treesByTab?.[tabId];
+      const tabTree = options?.treesByTab?.[_tabId];
       if (tabTree) return tabTree;
       if (script.length === 0) return axTree("Example Domain", ["Submit"]);
       return script.length === 1 ? script[0] : script.shift();
@@ -228,8 +227,13 @@ function makeMultiTabsApi(tabs: chrome.tabs.Tab[]) {
           (query.windowId === undefined || tab.windowId === query.windowId) &&
           (query.active === undefined || tab.active === query.active),
       ),
-    close(tabId: number) {
-      byId.delete(tabId);
+    activate(tabId: number) {
+      for (const tab of byId.values()) tab.active = tab.id === tabId;
+    },
+    goTo(tabId: number, url: string, title: string) {
+      const tab = byId.get(tabId);
+      if (!tab) throw new Error(`tab ${tabId} closed`);
+      Object.assign(tab, { url, title });
     },
   };
 }
@@ -243,18 +247,26 @@ function runtimeOnMessageEmit(
   chromeApi: ReturnType<typeof installChrome>,
   requestId: string,
   step: unknown,
-): void {
+  tabId = TAB_ID,
+  active = true,
+): ReturnType<typeof vi.fn> {
+  const producerId = `test-producer-${tabId}`;
+  const producerKey = `${requestId}:${producerId}`;
+  const sequence = (stepSequenceByProducer.get(producerKey) ?? 0) + 1;
+  stepSequenceByProducer.set(producerKey, sequence);
+  const sendResponse = vi.fn();
   chromeApi.runtimeOnMessage.emit(
-    { type: RECORD_STEP, requestId, step },
-    { tab: { id: TAB_ID } } as chrome.runtime.MessageSender,
-    () => {},
+    { type: RECORD_STEP, requestId, producerId, sequence, step },
+    { tab: { id: tabId, active } } as chrome.runtime.MessageSender,
+    sendResponse,
   );
+  return sendResponse;
 }
 
 describe("recorded user steps reach the exported trace", () => {
   afterEach(() => {
+    stepSequenceByProducer.clear();
     resetBrowserObservationForTests();
-    resetStateIdCounterForTests();
     vi.unstubAllGlobals();
   });
 
@@ -284,15 +296,14 @@ describe("recorded user steps reach the exported trace", () => {
       {
         type: RECORD_STEP,
         requestId,
+        producerId: "test-producer-4",
+        sequence: 1,
         step: {
           op: "click",
           page_url: START_URL,
           target: { role: "button", name: "Submit", tag: "button" },
           geometry: {
             rect: { x: 0, y: 0, w: 10, h: 10 },
-            scrollX: 0,
-            scrollY: 0,
-            position: "static",
             tag: "button",
           },
         },
@@ -328,21 +339,37 @@ describe("recorded user steps reach the exported trace", () => {
     await handleRecordStart(manager, RECORD_START_V3, { tabsApi, sendToTab, cdp: makeFakeCdp() });
     attachRecordStepListener({ tabsApi, sendToTab });
 
+    let sequence = 0;
     const emit = (step: unknown) => {
+      sequence += 1;
       runtimeOnMessage.emit(
-        { type: RECORD_STEP, requestId, step },
+        { type: RECORD_STEP, requestId, producerId: "test-producer-4", sequence, step },
         { tab: { id: TAB_ID } } as chrome.runtime.MessageSender,
         () => {},
       );
     };
 
-    emit({
+    const fill = {
       op: "fill",
       page_url: START_URL,
       target: { role: "textbox", name: "Search", tag: "input" },
       value: "hello",
       commit: "enter",
-    });
+    };
+    emit(fill);
+    const duplicateAck = vi.fn();
+    runtimeOnMessage.emit(
+      {
+        type: RECORD_STEP,
+        requestId,
+        producerId: "test-producer-4",
+        sequence: 1,
+        step: fill,
+      },
+      { tab: { id: TAB_ID } } as chrome.runtime.MessageSender,
+      duplicateAck,
+    );
+    expect(duplicateAck).toHaveBeenCalledWith({ ok: true, sequence: 1 });
     emit({
       op: "click",
       page_url: START_URL,
@@ -411,13 +438,13 @@ describe("recorded user steps reach the exported trace", () => {
     });
     attachRecordStepListener({ tabsApi, sendToTab });
 
-    const commit = (url: string, transitionType: string) => {
+    const commit = (url: string, transitionType: string, transitionQualifiers: string[] = []) => {
       chromeApi.webNavigationOnCommitted.emit({
         tabId: TAB_ID,
         frameId: 0,
         url,
         transitionType,
-        transitionQualifiers: [],
+        transitionQualifiers,
       } as unknown as chrome.webNavigation.WebNavigationTransitionCallbackDetails);
     };
 
@@ -428,7 +455,7 @@ describe("recorded user steps reach the exported trace", () => {
     await settleWait();
 
     tabsApi.goTo("https://iwiki.woa.com/dashboard", "工作台 - 腾讯iWiki");
-    commit("https://iwiki.woa.com/dashboard", "link");
+    commit("https://iwiki.woa.com/dashboard", "link", ["server_redirect"]);
     await settleWait();
 
     runtimeOnMessageEmit(chromeApi, requestId, {
@@ -1079,61 +1106,52 @@ describe("recorded user steps reach the exported trace", () => {
     expect(asTraceV3((stopped as RecordStopResult).trace).stopped_by).toBe("user_finish");
   });
 
-  it("retries STOP only on armed tabs whose final delivery did not ack", async () => {
+  it("returns a shared finish failure to a concurrent CLI stop", async () => {
     const chromeApi = installChrome();
     const manager = fakeManager();
-    const firstTab = {
-      id: TAB_ID,
-      windowId: AGENT_WINDOW_ID,
-      active: true,
-      status: "complete",
-      url: START_URL,
-      title: "First",
-    } as chrome.tabs.Tab;
-    const secondTab = {
-      id: 5,
-      windowId: AGENT_WINDOW_ID,
-      active: false,
-      status: "complete",
-      url: "https://example.com/second",
-      title: "Second",
-    } as chrome.tabs.Tab;
-    const tabsApi = makeMultiTabsApi([firstTab, secondTab]);
-    const stopTabs: number[] = [];
-    let secondStopAttempts = 0;
-    const sendToTab = vi.fn(async (tabId: number, msg: unknown) => {
-      const typed = msg as { type?: string };
+    const tabsApi = makeTabsApi();
+    let requestId = "";
+    let releaseStop!: () => void;
+    const stopReleased = new Promise<void>((resolve) => {
+      releaseStop = resolve;
+    });
+    let announceStop!: () => void;
+    const stopStarted = new Promise<void>((resolve) => {
+      announceStop = resolve;
+    });
+    const sendToTab = vi.fn(async (_tabId: number, msg: unknown) => {
+      const typed = msg as { type?: string; requestId?: string };
+      if (typed.type === RECORD_START && typed.requestId) requestId = typed.requestId;
       if (typed.type === RECORD_STOP) {
-        stopTabs.push(tabId);
-        if (tabId === 5 && secondStopAttempts++ === 0) {
-          return { ok: false, error: "failed to deliver final fill" };
-        }
+        announceStop();
+        await stopReleased;
+        return { ok: false, error: "final capture failed" };
       }
       return { ok: true };
     });
     const deps = { tabsApi, sendToTab, cdp: makeFakeCdp() };
 
     await handleRecordStart(manager, RECORD_START_V3, deps);
-    attachRecordQueryListener(deps);
-    await new Promise<void>((resolve) => {
-      chromeApi.runtimeOnMessage.emit(
-        { type: RECORD_QUERY },
-        { tab: secondTab } as chrome.runtime.MessageSender,
-        () => resolve(),
-      );
-    });
+    attachRecordFinishListener(deps);
+    chromeApi.runtimeOnMessage.emit(
+      { type: RECORD_FINISH, requestId },
+      { tab: { id: TAB_ID } } as chrome.runtime.MessageSender,
+      () => {},
+    );
+    await stopStarted;
 
-    const firstStop = await handleRecordStop(manager, { session_id: "abcd" }, deps);
-    expect(firstStop).toMatchObject({ code: "protocol_error" });
-    expect(stopTabs).toEqual([TAB_ID, 5]);
+    const cliStop = handleRecordStop(manager, { session_id: "abcd" }, deps);
+    releaseStop();
+    const result = await Promise.race([
+      cliStop,
+      new Promise<"timed_out">((resolve) => setTimeout(() => resolve("timed_out"), 500)),
+    ]);
 
-    const retry = await handleRecordStop(manager, { session_id: "abcd" }, deps);
-    const retryTrace = (retry as RecordStopResult).trace;
-    expect("version" in retryTrace && retryTrace.version).toBe(3);
-    expect(stopTabs).toEqual([TAB_ID, 5, 5]);
+    expect(result).not.toBe("timed_out");
+    expect(result).toMatchObject({ code: "protocol_error" });
   });
 
-  it("defaults to trace v2 when trace_version is omitted", async () => {
+  it("defaults to legacy trace v2 without starting observation capture", async () => {
     const { runtimeOnMessage } = installChrome();
     const manager = fakeManager();
     const tabsApi = makeTabsApi();
@@ -1151,6 +1169,23 @@ describe("recorded user steps reach the exported trace", () => {
       {
         type: RECORD_STEP,
         requestId,
+        producerId: "test-producer-4",
+        sequence: 1,
+        step: {
+          op: "hover",
+          page_url: START_URL,
+          target: { role: "button", name: "Submit", tag: "button" },
+        },
+      },
+      { tab: { id: TAB_ID } } as chrome.runtime.MessageSender,
+      () => {},
+    );
+    runtimeOnMessage.emit(
+      {
+        type: RECORD_STEP,
+        requestId,
+        producerId: "test-producer-4",
+        sequence: 2,
         step: {
           op: "click",
           page_url: START_URL,
@@ -1165,6 +1200,8 @@ describe("recorded user steps reach the exported trace", () => {
     const trace = (stopped as RecordStopResult).trace;
     expect("version" in trace).toBe(false);
     expect("pages" in trace && trace.pages.length).toBeGreaterThan(0);
+    // `hover` has no v2 counterpart on peers that negotiate v2.
+    expect("steps" in trace && trace.steps.map((step) => step.op)).toEqual(["click"]);
   });
 
   it("rejects unsupported trace_version values", async () => {
@@ -1177,142 +1214,37 @@ describe("recorded user steps reach the exported trace", () => {
     expect(result).toMatchObject({ code: "invalid_params" });
   });
 
-  it("does not let a closed armed tab block finish", async () => {
+  it("records tab transitions and binds actions to each tab's observation", async () => {
     const chromeApi = installChrome();
     const manager = fakeManager();
-    const firstTab = {
-      id: TAB_ID,
-      windowId: AGENT_WINDOW_ID,
-      active: true,
-      status: "complete",
-      url: START_URL,
-    } as chrome.tabs.Tab;
-    const secondTab = {
-      id: 5,
-      windowId: AGENT_WINDOW_ID,
-      active: false,
-      status: "complete",
-      url: "https://example.com/second",
-    } as chrome.tabs.Tab;
-    const tabsApi = makeMultiTabsApi([firstTab, secondTab]);
-    const stopTabs: number[] = [];
-    const sendToTab = vi.fn(async (tabId: number, msg: unknown) => {
-      if ((msg as { type?: string }).type === RECORD_STOP) stopTabs.push(tabId);
-      return { ok: true };
-    });
-    const deps = { tabsApi, sendToTab, cdp: makeFakeCdp() };
-
-    await handleRecordStart(manager, RECORD_START_V3, deps);
-    attachRecordQueryListener(deps);
-    await new Promise<void>((resolve) => {
-      chromeApi.runtimeOnMessage.emit(
-        { type: RECORD_QUERY },
-        { tab: secondTab } as chrome.runtime.MessageSender,
-        () => resolve(),
-      );
-    });
-    tabsApi.close(5);
-
-    const stopped = await handleRecordStop(manager, { session_id: "abcd" }, deps);
-
-    const stoppedTrace = (stopped as RecordStopResult).trace;
-    expect("version" in stoppedTrace && stoppedTrace.version).toBe(3);
-    expect(stopTabs).toEqual([TAB_ID]);
-  });
-
-  it("waits for an acknowledged in-flight rearm before stopping armed tabs", async () => {
-    const chromeApi = installChrome();
-    const manager = fakeManager();
-    const firstTab = {
-      id: TAB_ID,
-      windowId: AGENT_WINDOW_ID,
-      active: true,
-      status: "complete",
-      url: START_URL,
-    } as chrome.tabs.Tab;
-    const secondTab = {
-      id: 5,
-      windowId: AGENT_WINDOW_ID,
-      active: false,
-      status: "complete",
-      url: "https://example.com/second",
-    } as chrome.tabs.Tab;
-    const tabsApi = makeMultiTabsApi([firstTab, secondTab]);
-    let releaseSecondStart!: () => void;
-    const secondStartGate = new Promise<void>((resolve) => {
-      releaseSecondStart = resolve;
-    });
-    let announceSecondStart!: () => void;
-    const secondStartBegan = new Promise<void>((resolve) => {
-      announceSecondStart = resolve;
-    });
-    const stopTabs: number[] = [];
-    const sendToTab = vi.fn(async (tabId: number, msg: unknown) => {
-      const type = (msg as { type?: string }).type;
-      if (type === RECORD_START && tabId === 5) {
-        announceSecondStart();
-        await secondStartGate;
-      }
-      if (type === RECORD_STOP) stopTabs.push(tabId);
-      return { ok: true };
-    });
-    const deps = { tabsApi, sendToTab, cdp: makeFakeCdp() };
-
-    await handleRecordStart(manager, RECORD_START_V3, deps);
-    attachRecordQueryListener(deps);
-    chromeApi.runtimeOnMessage.emit(
-      { type: RECORD_QUERY },
-      { tab: secondTab } as chrome.runtime.MessageSender,
-      () => {},
-    );
-    await secondStartBegan;
-
-    const stopping = handleRecordStop(manager, { session_id: "abcd" }, deps);
-    await new Promise((resolve) => setTimeout(resolve, 50));
-    releaseSecondStart();
-    const stopped = await stopping;
-
-    const stoppedTrace = (stopped as RecordStopResult).trace;
-    expect("version" in stoppedTrace && stoppedTrace.version).toBe(3);
-    expect(stopTabs).toEqual([TAB_ID, 5]);
-  });
-
-  it("drains a navigation callback received while finish is settling", async () => {
-    const chromeApi = installChrome();
-    const manager = fakeManager();
-    const firstTab = {
-      id: TAB_ID,
-      windowId: AGENT_WINDOW_ID,
-      active: true,
-      status: "complete",
-      url: START_URL,
-      title: "Editor",
-    } as chrome.tabs.Tab;
-    const secondUrl = "https://example.com/late-navigation";
-    const secondTab = {
-      id: 5,
-      windowId: AGENT_WINDOW_ID,
-      active: false,
-      status: "complete",
-      url: secondUrl,
-      title: "Late page",
-    } as chrome.tabs.Tab;
-    let releaseLookup!: () => void;
-    const lookupGate = new Promise<void>((resolve) => {
-      releaseLookup = resolve;
-    });
-    const baseTabsApi = makeMultiTabsApi([firstTab, secondTab]);
-    const tabsApi = {
-      ...baseTabsApi,
-      get: async (tabId: number) => {
-        if (tabId === 5) await lookupGate;
-        return baseTabsApi.get(tabId);
+    const secondUrl = "https://example.com/second";
+    const tabsApi = makeMultiTabsApi([
+      {
+        id: TAB_ID,
+        windowId: AGENT_WINDOW_ID,
+        active: true,
+        status: "complete",
+        url: START_URL,
+        title: "First tab",
+      } as chrome.tabs.Tab,
+      {
+        id: 5,
+        windowId: AGENT_WINDOW_ID,
+        active: false,
+        status: "complete",
+        url: secondUrl,
+        title: "Second tab",
+      } as chrome.tabs.Tab,
+    ]);
+    const cdp = makeFakeCdp(undefined, {
+      treesByTab: {
+        [TAB_ID]: axTree("First tab", ["First action"]),
+        5: axTree("Second tab", ["Second action"]),
       },
-    };
-    const cdp = makeFakeCdp([axTree("Editor", ["Save"]), axTree("Saved")]);
+    });
     let requestId = "";
-    const sendToTab = vi.fn(async (_tabId: number, msg: unknown) => {
-      const typed = msg as { type?: string; requestId?: string };
+    const sendToTab = vi.fn(async (_tabId: number, message: unknown) => {
+      const typed = message as { type?: string; requestId?: string };
       if (typed.type === RECORD_START && typed.requestId) requestId = typed.requestId;
       return { ok: true };
     });
@@ -1320,80 +1252,352 @@ describe("recorded user steps reach the exported trace", () => {
 
     await handleRecordStart(manager, RECORD_START_V3, deps);
     attachRecordStepListener(deps);
+    tabsApi.activate(5);
+    chromeApi.tabsOnActivated.emit({ tabId: 5, windowId: AGENT_WINDOW_ID });
+    await settleWait();
+    runtimeOnMessageEmit(
+      chromeApi,
+      requestId,
+      {
+        op: "click",
+        page_url: secondUrl,
+        target: { role: "button", name: "Second action", tag: "button" },
+      },
+      5,
+    );
+    await settleWait();
+
+    tabsApi.activate(TAB_ID);
+    chromeApi.tabsOnActivated.emit({ tabId: TAB_ID, windowId: AGENT_WINDOW_ID });
+    await settleWait();
     runtimeOnMessageEmit(chromeApi, requestId, {
       op: "click",
       page_url: START_URL,
-      target: { role: "button", name: "Save", tag: "button" },
+      target: { role: "button", name: "First action", tag: "button" },
     });
+    await settleWait();
 
-    const stopping = handleRecordStop(manager, { session_id: "abcd" }, deps);
-    chromeApi.webNavigationOnCommitted.emit({
-      tabId: 5,
-      frameId: 0,
-      url: secondUrl,
-      transitionType: "link",
-      transitionQualifiers: [],
-    } as unknown as chrome.webNavigation.WebNavigationTransitionCallbackDetails);
-    const earlyStop = await Promise.race([
-      stopping.then((result) => ({ result })),
-      new Promise<null>((resolve) => setTimeout(() => resolve(null), 500)),
-    ]);
-    releaseLookup();
-    const stopped = earlyStop?.result ?? (await stopping);
+    const stopped = await handleRecordStop(manager, { session_id: "abcd" }, deps);
     const trace = asTraceV3((stopped as RecordStopResult).trace);
+    const stateUrl = (stateId: string) => trace.states.find((state) => state.id === stateId)?.url;
+    const switches = trace.steps.filter((step) => step.op === "switch_tab");
+    const clicks = trace.steps.filter((step) => step.op === "click");
 
-    expect(trace.steps.some((step) => step.op === "navigate" && step.to === secondUrl)).toBe(true);
-    const navigation = trace.steps.find((step) => step.op === "navigate" && step.to === secondUrl);
-    expect(navigation?.result.state).toBeTruthy();
-  }, 15_000);
+    expect(switches).toHaveLength(2);
+    expect(clicks).toHaveLength(2);
+    expect(stateUrl(switches[0]!.state)).toBe(START_URL);
+    expect(stateUrl(switches[0]!.result.state)).toBe(secondUrl);
+    expect(stateUrl(clicks[0]!.state)).toBe(secondUrl);
+    expect(stateUrl(switches[1]!.result.state)).toBe(START_URL);
+    expect(stateUrl(clicks[1]!.state)).toBe(START_URL);
+  });
 
-  it("captures a new-tab navigation from the emitting tab", async () => {
+  it("keeps switch_tab internal when the v3 caller did not advertise support", async () => {
     const chromeApi = installChrome();
     const manager = fakeManager();
-    const firstTab = {
-      id: TAB_ID,
-      windowId: AGENT_WINDOW_ID,
-      active: true,
-      status: "complete",
-      url: START_URL,
-      title: "Old tab",
-    } as chrome.tabs.Tab;
-    const secondUrl = "https://example.com/new-tab";
-    const secondTab = {
-      id: 5,
-      windowId: AGENT_WINDOW_ID,
-      active: false,
-      status: "complete",
-      url: secondUrl,
-      title: "New tab",
-    } as chrome.tabs.Tab;
-    const tabsApi = makeMultiTabsApi([firstTab, secondTab]);
-    const cdp = makeFakeCdp(undefined, {
-      treesByTab: {
-        [TAB_ID]: axTree("Old tab", ["Old action"]),
-        5: axTree("New tab", ["New action"]),
-      },
+    const tabsApi = makeMultiTabsApi([
+      {
+        id: TAB_ID,
+        windowId: AGENT_WINDOW_ID,
+        active: true,
+        status: "complete",
+        url: START_URL,
+      } as chrome.tabs.Tab,
+      {
+        id: 5,
+        windowId: AGENT_WINDOW_ID,
+        active: false,
+        status: "complete",
+        url: "https://example.com/second",
+      } as chrome.tabs.Tab,
+    ]);
+    const deps = { tabsApi, sendToTab: vi.fn(async () => ({ ok: true })), cdp: makeFakeCdp() };
+
+    await handleRecordStart(
+      manager,
+      { session_id: "abcd", url: START_URL, trace_version: 3 },
+      deps,
+    );
+    tabsApi.activate(5);
+    chromeApi.tabsOnActivated.emit({ tabId: 5, windowId: AGENT_WINDOW_ID });
+    await settleWait();
+
+    const stopped = await handleRecordStop(manager, { session_id: "abcd" }, deps);
+    const trace = asTraceV3((stopped as RecordStopResult).trace);
+    expect(trace.steps.some((step) => step.op === "switch_tab")).toBe(false);
+  });
+
+  it("keeps the latest rapid tab activation when an earlier rearm acknowledges late", async () => {
+    const chromeApi = installChrome();
+    const manager = fakeManager();
+    const tabsApi = makeMultiTabsApi(
+      [TAB_ID, 5, 6].map(
+        (id) =>
+          ({
+            id,
+            windowId: AGENT_WINDOW_ID,
+            active: id === TAB_ID,
+            status: "complete",
+            url: id === TAB_ID ? START_URL : `https://example.com/tab-${id}`,
+          }) as chrome.tabs.Tab,
+      ),
+    );
+    let releaseSecond!: () => void;
+    const secondGate = new Promise<void>((resolve) => {
+      releaseSecond = resolve;
     });
-    const sendToTab = vi.fn(async () => ({ ok: true }));
-    const deps = { tabsApi, sendToTab, cdp };
+    let announceSecond!: () => void;
+    const secondStarted = new Promise<void>((resolve) => {
+      announceSecond = resolve;
+    });
+    const sendToTab = vi.fn(async (tabId: number, message: unknown) => {
+      if ((message as { type?: string }).type === RECORD_START && tabId === 5) {
+        announceSecond();
+        await secondGate;
+      }
+      return { ok: true };
+    });
+    const deps = { tabsApi, sendToTab, cdp: makeFakeCdp() };
 
     await handleRecordStart(manager, RECORD_START_V3, deps);
+    tabsApi.activate(5);
+    chromeApi.tabsOnActivated.emit({ tabId: 5, windowId: AGENT_WINDOW_ID });
+    await secondStarted;
+    tabsApi.activate(6);
+    chromeApi.tabsOnActivated.emit({ tabId: 6, windowId: AGENT_WINDOW_ID });
+    await settleWait();
+    releaseSecond();
+    await new Promise((resolve) => setTimeout(resolve, 50));
+
+    const stopped = await handleRecordStop(manager, { session_id: "abcd" }, deps);
+    const trace = asTraceV3((stopped as RecordStopResult).trace);
+    const switches = trace.steps.filter((step) => step.op === "switch_tab");
+    const stateUrl = (stateId: string) => trace.states.find((state) => state.id === stateId)?.url;
+    expect(switches).toHaveLength(1);
+    expect(stateUrl(switches[0]!.result.state)).toBe("https://example.com/tab-6");
+  });
+
+  it("does not retry an in-flight tab rearm after recording stops", async () => {
+    const chromeApi = installChrome();
+    const manager = fakeManager();
+    const tabsApi = makeMultiTabsApi([
+      {
+        id: TAB_ID,
+        windowId: AGENT_WINDOW_ID,
+        active: true,
+        status: "complete",
+        url: START_URL,
+      } as chrome.tabs.Tab,
+      {
+        id: 5,
+        windowId: AGENT_WINDOW_ID,
+        active: false,
+        status: "complete",
+        url: "https://example.com/second",
+      } as chrome.tabs.Tab,
+    ]);
+    let failRearm!: () => void;
+    const rearmGate = new Promise<never>((_resolve, reject) => {
+      failRearm = () => reject(new Error("rearm failed after stop"));
+    });
+    let announceRearm!: () => void;
+    const rearmStarted = new Promise<void>((resolve) => {
+      announceRearm = resolve;
+    });
+    const sendToTab = vi.fn(async (tabId: number, message: unknown) => {
+      if ((message as { type?: string }).type === RECORD_START && tabId === 5) {
+        announceRearm();
+        await rearmGate;
+      }
+      return { ok: true };
+    });
+    const deps = { tabsApi, sendToTab, cdp: makeFakeCdp() };
+
+    await handleRecordStart(manager, RECORD_START_V3, deps);
+    tabsApi.activate(5);
+    chromeApi.tabsOnActivated.emit({ tabId: 5, windowId: AGENT_WINDOW_ID });
+    await rearmStarted;
+
+    const stoppedPromise = handleRecordStop(manager, { session_id: "abcd" }, deps);
+    await vi.waitFor(() => {
+      expect(sendToTab).toHaveBeenCalledWith(5, expect.objectContaining({ type: RECORD_STOP }));
+    });
+    failRearm();
+    const stopped = await stoppedPromise;
+    await new Promise((resolve) => setTimeout(resolve, 600));
+
+    const startsForSecondTab = sendToTab.mock.calls.filter(
+      ([tabId, message]) => tabId === 5 && (message as { type?: string }).type === RECORD_START,
+    );
+    expect(startsForSecondTab).toHaveLength(1);
+    expect(asTraceV3((stopped as RecordStopResult).trace).steps).toHaveLength(0);
+  });
+
+  it("keeps pending action settles scoped to their tab during an immediate switch", async () => {
+    const chromeApi = installChrome();
+    const manager = fakeManager();
+    const secondUrl = "https://example.com/second";
+    const tabsApi = makeMultiTabsApi([
+      {
+        id: TAB_ID,
+        windowId: AGENT_WINDOW_ID,
+        active: true,
+        status: "complete",
+        url: START_URL,
+      } as chrome.tabs.Tab,
+      {
+        id: 5,
+        windowId: AGENT_WINDOW_ID,
+        active: false,
+        status: "complete",
+        url: secondUrl,
+      } as chrome.tabs.Tab,
+    ]);
+    let requestId = "";
+    const sendToTab = vi.fn(async (_tabId: number, message: unknown) => {
+      const typed = message as { type?: string; requestId?: string };
+      if (typed.type === RECORD_START && typed.requestId) requestId = typed.requestId;
+      return { ok: true };
+    });
+    const deps = { tabsApi, sendToTab, cdp: makeFakeCdp() };
+
+    await handleRecordStart(manager, RECORD_START_V3, deps);
+    attachRecordStepListener(deps);
+    runtimeOnMessageEmit(chromeApi, requestId, {
+      op: "click",
+      page_url: START_URL,
+      target: { role: "button", name: "First action", tag: "button" },
+    });
+    tabsApi.activate(5);
+    chromeApi.tabsOnActivated.emit({ tabId: 5, windowId: AGENT_WINDOW_ID });
+    runtimeOnMessageEmit(
+      chromeApi,
+      requestId,
+      {
+        op: "click",
+        page_url: secondUrl,
+        target: { role: "button", name: "Second action", tag: "button" },
+      },
+      5,
+    );
+    await settleWait();
+
+    const stopped = await handleRecordStop(manager, { session_id: "abcd" }, deps);
+    const trace = asTraceV3((stopped as RecordStopResult).trace);
+    const stateUrl = (stateId: string) => trace.states.find((state) => state.id === stateId)?.url;
+
+    expect(trace.steps.map((step) => step.op)).toEqual(["click", "switch_tab", "click"]);
+    expect(stateUrl(trace.steps[0]!.result.state)).toBe(START_URL);
+    expect(stateUrl(trace.steps[1]!.result.state)).toBe(secondUrl);
+    expect(stateUrl(trace.steps[2]!.state)).toBe(secondUrl);
+  });
+
+  it("tracks the same navigation URL independently for each tab", async () => {
+    const chromeApi = installChrome();
+    const manager = fakeManager();
+    const tabsApi = makeMultiTabsApi([
+      {
+        id: TAB_ID,
+        windowId: AGENT_WINDOW_ID,
+        active: true,
+        status: "complete",
+        url: START_URL,
+      } as chrome.tabs.Tab,
+      {
+        id: 5,
+        windowId: AGENT_WINDOW_ID,
+        active: false,
+        status: "complete",
+        url: "https://example.com/second",
+      } as chrome.tabs.Tab,
+    ]);
+    const deps = { tabsApi, sendToTab: vi.fn(async () => ({ ok: true })), cdp: makeFakeCdp() };
+
+    await handleRecordStart(manager, RECORD_START_V3, deps);
+    tabsApi.activate(5);
+    chromeApi.tabsOnActivated.emit({ tabId: 5, windowId: AGENT_WINDOW_ID });
+    await settleWait();
+    tabsApi.goTo(5, START_URL, "Same URL in second tab");
     chromeApi.webNavigationOnCommitted.emit({
       tabId: 5,
       frameId: 0,
-      url: secondUrl,
-      transitionType: "link",
-      transitionQualifiers: [],
+      url: START_URL,
+      transitionType: "typed",
+      transitionQualifiers: ["from_address_bar"],
     } as unknown as chrome.webNavigation.WebNavigationTransitionCallbackDetails);
     await settleWait();
 
     const stopped = await handleRecordStop(manager, { session_id: "abcd" }, deps);
     const trace = asTraceV3((stopped as RecordStopResult).trace);
-    const navigation = trace.steps.find((step) => step.op === "navigate" && step.to === secondUrl);
-    const landed = trace.states.find((state) => state.id === navigation?.result.state);
+    expect(
+      trace.steps.filter((step) => step.op === "navigate" && step.to === START_URL),
+    ).toHaveLength(1);
+  });
 
-    expect(landed?.url).toBe(secondUrl);
-    expect(landed?.body).toContain("New tab");
-    expect(landed?.body).not.toContain("Old action");
-  }, 15_000);
+  it("drops late input and frame navigation from an inactive tab", async () => {
+    const chromeApi = installChrome();
+    const manager = fakeManager();
+    const secondUrl = "https://example.com/second";
+    const tabsApi = makeMultiTabsApi([
+      {
+        id: TAB_ID,
+        windowId: AGENT_WINDOW_ID,
+        active: true,
+        status: "complete",
+        url: START_URL,
+      } as chrome.tabs.Tab,
+      {
+        id: 5,
+        windowId: AGENT_WINDOW_ID,
+        active: false,
+        status: "complete",
+        url: secondUrl,
+      } as chrome.tabs.Tab,
+    ]);
+    let requestId = "";
+    const sendToTab = vi.fn(async (_tabId: number, message: unknown) => {
+      const typed = message as { type?: string; requestId?: string };
+      if (typed.type === RECORD_START && typed.requestId) requestId = typed.requestId;
+      return { ok: true };
+    });
+    const deps = { tabsApi, sendToTab, cdp: makeFakeCdp() };
+
+    await handleRecordStart(manager, RECORD_START_V3, deps);
+    attachRecordStepListener(deps);
+    tabsApi.activate(5);
+    chromeApi.tabsOnActivated.emit({ tabId: 5, windowId: AGENT_WINDOW_ID });
+    await settleWait();
+    const lateAck = runtimeOnMessageEmit(
+      chromeApi,
+      requestId,
+      {
+        op: "fill",
+        page_url: START_URL,
+        target: { role: "textbox", name: "Draft", tag: "input" },
+        value: "late",
+      },
+      TAB_ID,
+      false,
+    );
+    expect(lateAck).toHaveBeenCalledWith({ ok: true, sequence: 1 });
+    chromeApi.webNavigationOnCommitted.emit({
+      tabId: TAB_ID,
+      frameId: 0,
+      url: "https://example.com/late",
+      transitionType: "link",
+      transitionQualifiers: [],
+    } as unknown as chrome.webNavigation.WebNavigationTransitionCallbackDetails);
+    chromeApi.webNavigationOnCommitted.emit({
+      tabId: 5,
+      frameId: 7,
+      url: "https://frame.example/late",
+      transitionType: "link",
+      transitionQualifiers: [],
+    } as unknown as chrome.webNavigation.WebNavigationTransitionCallbackDetails);
+    await new Promise((resolve) => setTimeout(resolve, 50));
+
+    const stopped = await handleRecordStop(manager, { session_id: "abcd" }, deps);
+    const trace = asTraceV3((stopped as RecordStopResult).trace);
+    expect(trace.steps.map((step) => step.op)).toEqual(["switch_tab"]);
+  });
 });

apps/extension/src/tools/__tests__/session.test.ts

@@ -66,6 +66,52 @@ function makeApis(
 }
 
 describe("handleSessionStop with auto-return", () => {
+  it("leaves the session untouched when cancellation arrived before teardown", async () => {
+    const aw = fakeAgentWindow([100]);
+    const sm = new SessionManager({ agentWindow: aw });
+    await sm.start("aa11");
+    const cdp = { detachSession: vi.fn(async () => {}) };
+    const controller = new AbortController();
+    controller.abort();
+
+    const res = await handleSessionStop(
+      sm,
+      { session_id: "aa11" },
+      { cdp, signal: controller.signal },
+    );
+
+    expect(res).toMatchObject({ code: "cancelled" });
+    expect(sm.has("aa11")).toBe(true);
+    expect(cdp.detachSession).not.toHaveBeenCalled();
+    expect(aw.remove).not.toHaveBeenCalled();
+  });
+
+  it("does not close the Agent Window when cancellation lands during CDP detach", async () => {
+    const aw = fakeAgentWindow([100]);
+    const sm = new SessionManager({ agentWindow: aw });
+    await sm.start("aa11");
+    const controller = new AbortController();
+    let finishDetach: () => void = () => {};
+    const detach = new Promise<void>((resolve) => {
+      finishDetach = resolve;
+    });
+    const cdp = { detachSession: vi.fn(() => detach) };
+
+    const stopping = handleSessionStop(
+      sm,
+      { session_id: "aa11" },
+      { cdp, signal: controller.signal },
+    );
+    await vi.waitFor(() => expect(cdp.detachSession).toHaveBeenCalledWith("aa11"));
+    controller.abort();
+    finishDetach();
+
+    const res = await stopping;
+    expect(res).toMatchObject({ code: "cancelled" });
+    expect(sm.has("aa11")).toBe(true);
+    expect(aw.remove).not.toHaveBeenCalled();
+  });
+
   it("returns every borrowed tab and closes the Agent Window in the right order", async () => {
     const aw = fakeAgentWindow([100]);
     const sm = new SessionManager({ agentWindow: aw });

apps/extension/src/tools/__tests__/snapshot-ref.test.ts

@@ -24,10 +24,14 @@ describe("lookupSnapshotRef", () => {
     expect(lookupSnapshotRef(ctx, "@e3", 4)).toEqual({
       backendNodeId: 1234,
       refKey: "e3",
+      kind: "dom",
+      capabilities: ["interact", "screenshot"],
     });
     expect(lookupSnapshotRef(ctx, "e3", 4)).toEqual({
       backendNodeId: 1234,
       refKey: "e3",
+      kind: "dom",
+      capabilities: ["interact", "screenshot"],
     });
   });
 
@@ -46,6 +50,29 @@ describe("lookupSnapshotRef", () => {
 });
 
 describe("resolveSnapshotRef", () => {
+  it("preserves the frame and child session needed to route iframe refs", async () => {
+    const sm = new SessionManager({ agentWindow: fakeAgentWindow([100]) });
+    const ctx = await sm.start("aa11");
+    ctx.refStore.set("e3", 1234, {
+      tabId: 4,
+      frameId: "child-frame",
+      cdpSessionId: "child-session",
+      kind: "dom",
+      capabilities: ["interact", "screenshot"],
+    });
+
+    const expected = {
+      backendNodeId: 1234,
+      refKey: "e3",
+      frameId: "child-frame",
+      cdpSessionId: "child-session",
+      kind: "dom" as const,
+      capabilities: ["interact", "screenshot"] as const,
+    };
+    expect(lookupSnapshotRef(ctx, "@e3", 4)).toEqual(expected);
+    expect(resolveSnapshotRef(ctx, "@e3", 4)).toEqual(expected);
+  });
+
   it("returns not_found for unknown ref", async () => {
     const sm = new SessionManager({ agentWindow: fakeAgentWindow([100]) });
     const ctx = await sm.start("aa11");
@@ -86,6 +113,28 @@ describe("resolveSnapshotRef", () => {
     expect(resolveSnapshotRef(ctx, "@e3", 4)).toEqual({
       backendNodeId: 1234,
       refKey: "e3",
+      kind: "dom",
+      capabilities: ["interact", "screenshot"],
+    });
+  });
+
+  it("rejects an operation outside the ref's declared capabilities", async () => {
+    const sm = new SessionManager({ agentWindow: fakeAgentWindow([100]) });
+    const ctx = await sm.start("aa11");
+    ctx.refStore.set("e3", 1234, {
+      tabId: 4,
+      kind: "surface",
+      capabilities: ["screenshot"],
+    });
+
+    expect(resolveSnapshotRef(ctx, "@e3", 4, "interact")).toMatchObject({
+      code: "permission_denied",
+      message: "ref @e3 does not support interact",
+      data: {
+        reason: "ref_capability_denied",
+        required_capability: "interact",
+        kind: "surface",
+      },
     });
   });
 });

apps/extension/src/tools/borrow-confirmation.ts

@@ -250,9 +250,10 @@ const DEFAULT_NOTIFICATION_COPY: BorrowNotificationCopy = {
 // URL injection-eligibility test
 // ---------------------------------------------------------------------------
 
-const CHROME_WEB_STORE_RES = [
+const EXTENSION_STORE_RES = [
   /^https:\/\/chrome\.google\.com\/webstore/i,
   /^https:\/\/chromewebstore\.google\.com/i,
+  /^https:\/\/microsoftedge\.microsoft\.com\/addons/i,
 ];
 
 /**
@@ -264,14 +265,15 @@ const CHROME_WEB_STORE_RES = [
  *   - `ftp://` no longer hosts content scripts reliably in modern Chrome.
  *   - `about:` / `chrome:` / `chrome-extension:` / `edge:` / `devtools:` /
  *     `view-source:` / `data:` / `blob:` are blocked by the platform.
- *   - The Chrome Web Store (both legacy and new domains) is also blocked.
+ *   - Extension storefronts are blocked by their own browser: the Chrome Web
+ *     Store (legacy and new domains) under Chrome, and Edge Add-ons under Edge.
  *
  * Mirrors the implicit scheme list documented in
  * https://developer.chrome.com/docs/extensions/develop/concepts/match-patterns.
  */
 export function isInjectableContentScriptUrl(url: string | undefined): boolean {
   if (!url) return false;
-  for (const re of CHROME_WEB_STORE_RES) {
+  for (const re of EXTENSION_STORE_RES) {
     if (re.test(url)) return false;
   }
   return /^https?:\/\//i.test(url);

apps/extension/src/tools/capture-vom-observation.ts

@@ -1,101 +1,9 @@
-import { type RenderedRef, renderVom, type VomOptions } from "@browser-skill/vom";
-import type { CdpAxNode } from "./observation";
-import { buildVomScene } from "./observation";
-import type { CdpRunner } from "./shared";
-import {
-  type CapturedNode,
-  type CapturedSurfaceProbe,
-  type CapturedViewModel,
-  captureViewModel,
-  collectOverlayExcludedBackendIds,
-} from "./vom/capture";
-
-export interface CaptureVomObservationOptions extends VomOptions {
-  conditionalSurfaceProbe?: boolean;
-  hoverProbeBypassOverlay?: (tabId: number, enabled: boolean) => Promise<void>;
-}
-
-export interface CaptureVomObservationResult {
-  text: string;
-  refs: RenderedRef[];
-  truncated: boolean;
-  captured: CapturedNode[];
-  surfaceProbes?: CapturedSurfaceProbe[];
-}
-
-function emptyCapturedViewModel(viewport = { width: 0, height: 0 }): CapturedViewModel {
-  return { viewport, nodes: [], iframeNodes: new Map(), excludedBackendNodeIds: new Set() };
-}
-
-interface LayoutMetricsViewportReply {
-  cssLayoutViewport?: { clientWidth?: number; clientHeight?: number };
-  layoutViewport?: { clientWidth?: number; clientHeight?: number };
-}
-
-async function fallbackCapturedViewModel(
-  cdp: CdpRunner,
-  tabId: number,
-): Promise<CapturedViewModel> {
-  let viewport = { width: 0, height: 0 };
-  try {
-    const metrics = await cdp.send<LayoutMetricsViewportReply>(tabId, "Page.getLayoutMetrics", {});
-    const vpSrc = metrics.cssLayoutViewport ?? metrics.layoutViewport ?? {};
-    viewport = {
-      width: vpSrc.clientWidth ?? 0,
-      height: vpSrc.clientHeight ?? 0,
-    };
-  } catch {
-    // viewport stays zero-sized
-  }
-  const excludedBackendNodeIds = await collectOverlayExcludedBackendIds(cdp, tabId);
-  return { ...emptyCapturedViewModel(viewport), excludedBackendNodeIds };
-}
-
-async function captureForVom(
-  cdp: CdpRunner,
-  tabId: number,
-  conditionalSurfaceProbe: boolean,
-  hoverProbeBypassOverlay?: (tabId: number, enabled: boolean) => Promise<void>,
-): Promise<CapturedViewModel> {
-  try {
-    return await captureViewModel(cdp, tabId, {
-      conditionalSurfaceProbe,
-      hoverProbeBypassOverlay,
-    });
-  } catch {
-    return fallbackCapturedViewModel(cdp, tabId);
-  }
-}
-
-/** Pure VOM capture pipeline shared by snapshot, observe, and recording. */
-export async function captureVomObservation(
-  cdp: CdpRunner,
-  tabId: number,
-  url: string | undefined,
-  opts: CaptureVomObservationOptions = {},
-): Promise<CaptureVomObservationResult> {
-  await cdp.ensureAttachedToUrl?.(tabId, url);
-  await cdp.send<unknown>(tabId, "Accessibility.enable", {});
-  const result = await cdp.send<{ nodes: CdpAxNode[] }>(tabId, "Accessibility.getFullAXTree", {});
-  const axNodes = result.nodes ?? [];
-  const captured = await captureForVom(
-    cdp,
-    tabId,
-    opts.conditionalSurfaceProbe ?? false,
-    opts.hoverProbeBypassOverlay,
-  );
-  const scene = buildVomScene(axNodes, captured, { pageUrl: url });
-  const rendered = renderVom(scene, {
-    maxDepth: opts.maxDepth,
-    maxTokens: opts.maxTokens,
-    redactValues: opts.redactValues,
-    activeRegionPolicy: opts.activeRegionPolicy ?? true,
-  });
-  return {
-    text: rendered.text,
-    refs: rendered.refs,
-    truncated: rendered.truncated,
-    captured: captured.nodes,
-    surfaceProbes: captured.surfaceProbes,
-  };
-}
+export {
+  type CaptureVomObservationOptions,
+  captureVomObservation,
+} from "./observation";
+export {
+  type CaptureVomFrame,
+  type CaptureVomMatchNode,
+  type CaptureVomObservationResult,
+} from "./vom/record-safe-observation";

apps/extension/src/tools/dispatcher.ts

@@ -51,7 +51,12 @@ import {
   handleScreenshot,
   handleSnapshot,
 } from "./observation";
-import { handleRecordAwait, handleRecordStart, handleRecordStop } from "./record";
+import {
+  handleRecordAwait,
+  handleRecordStart,
+  handleRecordStop,
+  type RecordRuntimeDeps,
+} from "./record";
 import {
   handleSessionStart,
   handleSessionStop,
@@ -99,6 +104,7 @@ export interface DispatcherDeps {
   transport: Transport;
   sessions: SessionManager;
   cdp?: DispatcherCdpRunner;
+  recording?: RecordRuntimeDeps;
   /**
    * Invoked whenever a dispatched RPC may have changed the live
    * session set (currently `tool.session_start` and
@@ -134,6 +140,7 @@ export class ToolDispatcher {
   private readonly transport: Transport;
   private readonly sessions: SessionManager;
   private readonly cdp?: DispatcherCdpRunner;
+  private readonly recording?: RecordRuntimeDeps;
   private readonly onSessionsChanged?: () => void;
   private readonly onBrowserControlResumed?: (sessionId: string) => void;
   private readonly approveBorrow?: BorrowConfirmationApprover;
@@ -153,6 +160,7 @@ export class ToolDispatcher {
     this.transport = deps.transport;
     this.sessions = deps.sessions;
     this.cdp = deps.cdp;
+    this.recording = deps.recording;
     this.onSessionsChanged = deps.onSessionsChanged;
     this.onBrowserControlResumed = deps.onBrowserControlResumed;
     this.approveBorrow = deps.approveBorrow;
@@ -290,6 +298,7 @@ export class ToolDispatcher {
         await this.releaseHoverLatch((req.params as SessionStopParams).session_id);
         return handleSessionStop(this.sessions, req.params as SessionStopParams, {
           cdp: this.cdp,
+          signal,
         });
       }
       case "tool.tab_list":
@@ -530,44 +539,26 @@ export class ToolDispatcher {
           signal,
         });
       case "tool.record_start":
-        return handleRecordStart(this.sessions, req.params as RecordStartParams, {
-          tabsApi: chromeTabsApi,
-          sendToTab: (tabId, msg) => chrome.tabs.sendMessage(tabId, msg),
-          bypassOverlay: async (tabId, enabled) => {
-            try {
-              await chrome.tabs.sendMessage(tabId, {
-                type: OVERLAY_AUTOMATION_BYPASS,
-                enabled,
-              });
-            } catch {
-              // Content script may be unavailable on restricted pages.
-            }
-          },
-          ...(this.cdp ? { cdp: this.cdp } : {}),
-          signal,
-        });
+        return this.recording
+          ? handleRecordStart(this.sessions, req.params as RecordStartParams, {
+              ...this.recording,
+              signal,
+            })
+          : recordingRuntimeUnavailable();
       case "tool.record_stop":
-        return handleRecordStop(this.sessions, req.params as RecordStopParams, {
-          tabsApi: chromeTabsApi,
-          sendToTab: (tabId, msg) => chrome.tabs.sendMessage(tabId, msg),
-          bypassOverlay: async (tabId, enabled) => {
-            try {
-              await chrome.tabs.sendMessage(tabId, {
-                type: OVERLAY_AUTOMATION_BYPASS,
-                enabled,
-              });
-            } catch {
-              // Content script may be unavailable on restricted pages.
-            }
-          },
-          signal,
-        });
+        return this.recording
+          ? handleRecordStop(this.sessions, req.params as RecordStopParams, {
+              ...this.recording,
+              signal,
+            })
+          : recordingRuntimeUnavailable();
       case "tool.record_await":
-        return handleRecordAwait(this.sessions, req.params as RecordAwaitParams, {
-          tabsApi: chromeTabsApi,
-          sendToTab: (tabId, msg) => chrome.tabs.sendMessage(tabId, msg),
-          signal,
-        });
+        return this.recording
+          ? handleRecordAwait(this.sessions, req.params as RecordAwaitParams, {
+              ...this.recording,
+              signal,
+            })
+          : recordingRuntimeUnavailable();
       default:
         return {
           code: "unknown_method",
@@ -703,6 +694,13 @@ function isRpcError(v: unknown): v is RpcError {
   );
 }
 
+function recordingRuntimeUnavailable(): RpcError {
+  return {
+    code: "protocol_error",
+    message: "recording runtime is unavailable",
+  };
+}
+
 function sessionIdForBrowserControlMethod(req: RequestFrame): string | null {
   switch (req.method) {
     case "tool.tab_create":

apps/extension/src/tools/element-geometry.ts

@@ -1,9 +1,15 @@
-// Shared CDP helpers for scrolling nodes into view and computing
-// viewport-space bounding rectangles. Used by interaction tools (click
-// centre) and observation tools (element screenshot clip).
+// Resolve and scroll nodes inside one CDP target. Cross-frame projection
+// belongs to frame-geometry.ts.
 
 import type { RpcError } from "@/transport/types";
 import { rpcError } from "./errors";
+import {
+  parseCdpQuad,
+  polygonBounds,
+  type Region,
+  rectPolygon,
+  type ViewportRect,
+} from "./geometry";
 import { type CdpRunner, isRpcError } from "./shared";
 
 const ELEMENT_NOT_VISIBLE_MESSAGE =
@@ -13,45 +19,6 @@ function elementNotVisibleError(): RpcError {
   return rpcError("permission_denied", "element_not_visible", ELEMENT_NOT_VISIBLE_MESSAGE);
 }
 
-/** Viewport-space axis-aligned bounding box for CDP `Page.captureScreenshot` clip. */
-export interface ViewportRect {
-  x: number;
-  y: number;
-  width: number;
-  height: number;
-}
-
-/**
- * Compute the axis-aligned bounding box of an 8-double CDP content
- * quad / box-model polygon. Exported for unit tests.
- */
-export function quadBoundingRect(quad: number[]): ViewportRect | null {
-  if (quad.length !== 8) return null;
-  const xs = [quad[0], quad[2], quad[4], quad[6]];
-  const ys = [quad[1], quad[3], quad[5], quad[7]];
-  const minX = Math.min(...xs);
-  const maxX = Math.max(...xs);
-  const minY = Math.min(...ys);
-  const maxY = Math.max(...ys);
-  const width = maxX - minX;
-  const height = maxY - minY;
-  if (width <= 0 || height <= 0) return null;
-  return { x: minX, y: minY, width, height };
-}
-
-function rectToQuad(rect: ViewportRect): number[] {
-  return [
-    rect.x,
-    rect.y,
-    rect.x + rect.width,
-    rect.y,
-    rect.x + rect.width,
-    rect.y + rect.height,
-    rect.x,
-    rect.y + rect.height,
-  ];
-}
-
 /**
  * Resolve `backendNodeId` → CDP `objectId` so we can invoke
  * `Runtime.callFunctionOn` against the live JS object.
@@ -112,48 +79,24 @@ export async function scrollNodeIntoView(
   }
 }
 
-/**
- * Viewport-space bounding box for `backendNodeId`. Prefers
- * `DOM.getContentQuads`, then `DOM.getBoxModel`, then the union of
- * visible descendant client rects for zero-size overflow containers.
- */
-export async function nodeBoundingRect(
+function validCdpRegion(quads: number[][] | undefined): Region {
+  return (quads ?? []).flatMap((raw) => {
+    const polygon = parseCdpQuad(raw);
+    return polygon && polygonBounds(polygon) ? [polygon] : [];
+  });
+}
+
+async function resolveVisibleContentRegion(
   cdp: CdpRunner,
   tabId: number,
   backendNodeId: number,
-): Promise<ViewportRect | RpcError> {
-  const polygon = await visibleContentPolygon(cdp, tabId, backendNodeId);
-  if (isRpcError(polygon)) return polygon;
-  // visibleContentPolygon only returns quads with positive area.
-  return quadBoundingRect(polygon)!;
-}
-
-/**
- * Compute the centroid of a CDP content quad. Quads are reported as
- * 8 doubles in clockwise order: (x1, y1, x2, y2, x3, y3, x4, y4).
- */
-export function quadCentre(quad: number[]): { x: number; y: number } {
-  const x = (quad[0] + quad[4]) / 2;
-  const y = (quad[1] + quad[5]) / 2;
-  return { x, y };
-}
-
-/** Centre of a `DOM.getBoxModel` `content` polygon (also 8 doubles). */
-export function boxCentre(box: number[]): { x: number; y: number } {
-  return quadCentre(box);
-}
-
-async function visibleContentPolygon(
-  cdp: CdpRunner,
-  tabId: number,
-  backendNodeId: number,
-): Promise<number[] | RpcError> {
+): Promise<Region | RpcError> {
   try {
     const quads = await cdp.send<{ quads?: number[][] }>(tabId, "DOM.getContentQuads", {
       backendNodeId,
     });
-    const quad = quads.quads?.find((q) => q.length === 8 && quadBoundingRect(q) !== null);
-    if (quad) return quad;
+    const visible = validCdpRegion(quads.quads);
+    if (visible.length > 0) return visible;
   } catch (err) {
     console.debug("[bsk element-geometry] getContentQuads failed", err);
   }
@@ -162,15 +105,34 @@ async function visibleContentPolygon(
       backendNodeId,
     });
     const content = box.model?.content;
-    if (content && content.length === 8 && quadBoundingRect(content)) return content;
+    const fallback = validCdpRegion(content ? [content] : undefined);
+    if (fallback.length > 0) return fallback;
   } catch (err) {
     console.debug("[bsk element-geometry] getBoxModel failed", err);
   }
   const descendantRect = await descendantBoundingRect(cdp, tabId, backendNodeId);
-  if (descendantRect) return rectToQuad(descendantRect);
+  if (descendantRect) {
+    return [
+      rectPolygon({
+        x: descendantRect.x,
+        y: descendantRect.y,
+        w: descendantRect.width,
+        h: descendantRect.height,
+      }),
+    ];
+  }
   return elementNotVisibleError();
 }
 
+/** Resolve every visible content region in the node's target-local viewport. */
+export async function nodeContentRegion(
+  cdp: CdpRunner,
+  tabId: number,
+  backendNodeId: number,
+): Promise<Region | RpcError> {
+  return resolveVisibleContentRegion(cdp, tabId, backendNodeId);
+}
+
 async function descendantBoundingRect(
   cdp: CdpRunner,
   tabId: number,
@@ -241,20 +203,3 @@ function parseViewportRect(value: unknown): ViewportRect | null {
   if (![x, y, width, height].every(Number.isFinite) || width <= 0 || height <= 0) return null;
   return { x, y, width, height };
 }
-
-/**
- * Get the click point for `backendNodeId`. Prefers
- * `DOM.getContentQuads` (richer for rotated / transformed elements);
- * falls back to `DOM.getBoxModel`, then visible descendant bounds.
- * Returns `permission_denied` with "element not visible" when every
- * geometry path comes back empty / fails.
- */
-export async function nodeCentre(
-  cdp: CdpRunner,
-  tabId: number,
-  backendNodeId: number,
-): Promise<{ x: number; y: number } | RpcError> {
-  const polygon = await visibleContentPolygon(cdp, tabId, backendNodeId);
-  if (isRpcError(polygon)) return polygon;
-  return quadCentre(polygon);
-}

apps/extension/src/tools/frame-geometry.ts

@@ -0,0 +1,290 @@
+import { type CdpFrame, type CdpFrameGraph, type CdpTarget } from "@/browser-driver/frame-graph";
+import type { RpcError } from "@/transport/types";
+import { nodeContentRegion, scrollNodeIntoView } from "./element-geometry";
+import {
+  clipPolygon,
+  type GeometryProjection,
+  type Point,
+  type Polygon,
+  type ProjectiveEdge,
+  parseCdpQuad,
+  polygonArea,
+  polygonCentroid,
+  projectRegionToViewport,
+  type Quad,
+  type Region,
+  rectPolygon,
+  regionBounds,
+  type Size,
+  type ViewportRect,
+} from "./geometry";
+import { type CdpRunner, cdpRunnerForTarget, isRpcError, sendToCdpTarget } from "./shared";
+
+export interface NodeAddress {
+  target: CdpTarget;
+  backendNodeId: number;
+  frameId?: string;
+}
+
+export interface ResolvedNodeGeometry {
+  topVisibleRegions: Region;
+  topBounds: ViewportRect;
+  actionPoint: Point;
+}
+
+function geometryError(message: string): RpcError {
+  return { code: "cdp_failed", message };
+}
+
+function frameMap(graph: CdpFrameGraph): Map<string, CdpFrame> {
+  return new Map(graph.frames.map((frame) => [frame.frameId, frame]));
+}
+
+function targetRootFrame(byId: Map<string, CdpFrame>, frame: CdpFrame): CdpFrame | null {
+  const seen = new Set<string>();
+  let current = frame;
+  while (current.parentFrameId) {
+    if (seen.has(current.frameId)) return null;
+    seen.add(current.frameId);
+    const parent = byId.get(current.parentFrameId);
+    if (!parent || parent.target.sessionId !== current.target.sessionId) break;
+    current = parent;
+  }
+  return current;
+}
+
+function frameAncestry(graph: CdpFrameGraph, frameId: string): CdpFrame[] | null {
+  const byId = frameMap(graph);
+  const path: CdpFrame[] = [];
+  const seen = new Set<string>();
+  let current = byId.get(frameId);
+  if (!current) return null;
+  while (current.parentFrameId) {
+    if (seen.has(current.frameId)) return null;
+    seen.add(current.frameId);
+    path.push(current);
+    const parent = byId.get(current.parentFrameId);
+    if (!parent) return null;
+    current = parent;
+  }
+  return path;
+}
+
+async function targetViewport(cdp: CdpRunner, target: CdpTarget): Promise<Size | null> {
+  const metrics = await sendToCdpTarget<{
+    cssLayoutViewport?: { clientWidth?: number; clientHeight?: number };
+    layoutViewport?: { clientWidth?: number; clientHeight?: number };
+  }>(cdp, target, "Page.getLayoutMetrics", {});
+  const viewport = metrics.cssLayoutViewport ?? metrics.layoutViewport ?? {};
+  const width = viewport.clientWidth ?? 0;
+  const height = viewport.clientHeight ?? 0;
+  return width > 0 && height > 0 ? { width, height } : null;
+}
+
+async function ownerContentQuad(
+  cdp: CdpRunner,
+  parent: CdpFrame,
+  ownerBackendNodeId: number,
+): Promise<Quad | null> {
+  const result = await sendToCdpTarget<{ model?: { content?: number[] } }>(
+    cdp,
+    parent.target,
+    "DOM.getBoxModel",
+    { backendNodeId: ownerBackendNodeId },
+  );
+  return parseCdpQuad(result.model?.content);
+}
+
+async function sameTargetFrameClips(
+  cdp: CdpRunner,
+  byId: Map<string, CdpFrame>,
+  frame: CdpFrame,
+  targetRoot: CdpFrame,
+): Promise<Polygon[] | null> {
+  const clips: Polygon[] = [];
+  let current = frame;
+  const seen = new Set<string>();
+  while (current.frameId !== targetRoot.frameId) {
+    if (seen.has(current.frameId) || !current.parentFrameId) return null;
+    seen.add(current.frameId);
+    const parent = byId.get(current.parentFrameId);
+    if (!parent || current.ownerBackendNodeId === undefined) return null;
+    const clip = await ownerContentQuad(cdp, parent, current.ownerBackendNodeId);
+    if (!clip) return null;
+    clips.push(clip);
+    current = parent;
+  }
+  return clips;
+}
+
+export async function resolveFrameProjection(
+  cdp: CdpRunner,
+  graph: CdpFrameGraph,
+  frameId: string,
+): Promise<GeometryProjection | null> {
+  const byId = frameMap(graph);
+  const frame = byId.get(frameId);
+  if (!frame) return null;
+  let targetRoot = targetRootFrame(byId, frame);
+  if (!targetRoot) return null;
+  const sourceViewport = await targetViewport(cdp, targetRoot.target);
+  if (!sourceViewport) return null;
+  const sourceClips = await sameTargetFrameClips(cdp, byId, frame, targetRoot);
+  if (!sourceClips) return null;
+
+  const edges: ProjectiveEdge[] = [];
+  while (targetRoot.parentFrameId) {
+    const parent = byId.get(targetRoot.parentFrameId);
+    if (!parent || targetRoot.ownerBackendNodeId === undefined) return null;
+    const destinationQuad = await ownerContentQuad(cdp, parent, targetRoot.ownerBackendNodeId);
+    if (!destinationQuad) return null;
+    const source =
+      edges.length === 0 ? sourceViewport : await targetViewport(cdp, targetRoot.target);
+    if (!source) return null;
+    const parentTargetRoot = targetRootFrame(byId, parent);
+    if (!parentTargetRoot) return null;
+    const destinationClips = await sameTargetFrameClips(cdp, byId, parent, parentTargetRoot);
+    if (!destinationClips) return null;
+    edges.push({ sourceViewport: source, destinationQuad, destinationClips });
+    targetRoot = parentTargetRoot;
+  }
+
+  const topViewport =
+    edges.length === 0 ? sourceViewport : await targetViewport(cdp, targetRoot.target);
+  return topViewport ? { sourceClips, edges, topViewport } : null;
+}
+
+async function loadFrameGraph(cdp: CdpRunner, tabId: number): Promise<CdpFrameGraph | null> {
+  if (!cdp.getFrameGraph) return null;
+  try {
+    return await cdp.getFrameGraph(tabId);
+  } catch (error) {
+    console.debug("[bsk frame-geometry] frame graph resolution failed", error);
+    return null;
+  }
+}
+
+async function scrollFrameOwners(
+  cdp: CdpRunner,
+  tabId: number,
+  graph: CdpFrameGraph,
+  frameId: string,
+): Promise<RpcError | null> {
+  const byId = frameMap(graph);
+  const ancestry = frameAncestry(graph, frameId);
+  if (!ancestry) return geometryError(`could not resolve frame ancestry for ${frameId}`);
+  for (const child of [...ancestry].reverse()) {
+    const parent = child.parentFrameId ? byId.get(child.parentFrameId) : undefined;
+    if (!parent || child.ownerBackendNodeId === undefined) {
+      return geometryError(`could not resolve frame owner for ${child.frameId}`);
+    }
+    const error = await scrollNodeIntoView(
+      cdpRunnerForTarget(cdp, parent.target),
+      tabId,
+      child.ownerBackendNodeId,
+    );
+    if (error) return error;
+  }
+  return null;
+}
+
+async function scrollElementWithFrameGraph(
+  cdp: CdpRunner,
+  tabId: number,
+  target: CdpTarget,
+  backendNodeId: number,
+  frameId: string | undefined,
+  graph: CdpFrameGraph | null,
+): Promise<RpcError | null> {
+  if (frameId) {
+    if (!graph) return geometryError(`could not resolve frame graph for ${frameId}`);
+    const error = await scrollFrameOwners(cdp, tabId, graph, frameId);
+    if (error) return error;
+  } else if (target.sessionId) {
+    return geometryError("an OOPIF node address requires frameId");
+  }
+  return scrollNodeIntoView(cdpRunnerForTarget(cdp, target), tabId, backendNodeId);
+}
+
+export async function scrollElementAndFramesIntoView(
+  cdp: CdpRunner,
+  tabId: number,
+  target: CdpTarget,
+  backendNodeId: number,
+  frameId?: string,
+): Promise<RpcError | null> {
+  const graph = frameId ? await loadFrameGraph(cdp, tabId) : null;
+  return scrollElementWithFrameGraph(cdp, tabId, target, backendNodeId, frameId, graph);
+}
+
+function largestRegion(regions: Region): Polygon | null {
+  let largest: { polygon: Polygon; area: number } | null = null;
+  for (const polygon of regions) {
+    const area = polygonArea(polygon);
+    if (area <= 0) continue;
+    if (!largest || area > largest.area) largest = { polygon, area };
+  }
+  return largest?.polygon ?? null;
+}
+
+export async function resolveNodeGeometry(
+  cdp: CdpRunner,
+  tabId: number,
+  address: NodeAddress,
+  options: { scrollIntoView?: boolean } = {},
+): Promise<ResolvedNodeGeometry | RpcError> {
+  try {
+    if (address.target.sessionId && !address.frameId) {
+      return geometryError("an OOPIF node address requires frameId");
+    }
+    const graph = address.frameId ? await loadFrameGraph(cdp, tabId) : null;
+    if (address.frameId && !graph) {
+      return geometryError(`could not resolve frame graph for ${address.frameId}`);
+    }
+
+    if (options.scrollIntoView) {
+      const scrollError = await scrollElementWithFrameGraph(
+        cdp,
+        tabId,
+        address.target,
+        address.backendNodeId,
+        address.frameId,
+        graph,
+      );
+      if (scrollError) return scrollError;
+    }
+
+    const localRegion = await nodeContentRegion(
+      cdpRunnerForTarget(cdp, address.target),
+      tabId,
+      address.backendNodeId,
+    );
+    if (isRpcError(localRegion)) return localRegion;
+
+    let topVisibleRegions: Region;
+    if (address.frameId && graph) {
+      const projection = await resolveFrameProjection(cdp, graph, address.frameId);
+      if (!projection)
+        return geometryError(`could not resolve frame geometry for ${address.frameId}`);
+      topVisibleRegions = projectRegionToViewport(localRegion, projection);
+    } else {
+      const viewport = await targetViewport(cdp, address.target);
+      if (!viewport) return geometryError("could not resolve top viewport geometry");
+      topVisibleRegions = localRegion
+        .map((polygon) =>
+          clipPolygon(polygon, rectPolygon({ x: 0, y: 0, w: viewport.width, h: viewport.height })),
+        )
+        .filter((polygon) => polygon.length >= 3);
+    }
+
+    const topBounds = regionBounds(topVisibleRegions);
+    const actionRegion = largestRegion(topVisibleRegions);
+    const actionPoint = actionRegion ? polygonCentroid(actionRegion) : null;
+    if (!topBounds || !actionPoint) {
+      return { code: "permission_denied", message: "element not visible" };
+    }
+    return { topVisibleRegions, topBounds, actionPoint };
+  } catch (error) {
+    return geometryError(error instanceof Error ? error.message : String(error));
+  }
+}

apps/extension/src/tools/geometry.ts

@@ -0,0 +1,266 @@
+export interface Point {
+  x: number;
+  y: number;
+}
+
+export type Polygon = Point[];
+export type Region = Polygon[];
+export type Quad = [Point, Point, Point, Point];
+
+export interface Size {
+  width: number;
+  height: number;
+}
+
+export interface ViewportRect {
+  x: number;
+  y: number;
+  width: number;
+  height: number;
+}
+
+export interface ProjectiveEdge {
+  sourceViewport: Size;
+  destinationQuad: Quad;
+  destinationClips?: Polygon[];
+}
+
+export interface GeometryProjection {
+  sourceClips: Polygon[];
+  edges: ProjectiveEdge[];
+  topViewport: Size;
+}
+
+export function parseCdpQuad(raw: number[] | undefined): Quad | null {
+  if (!raw || raw.length !== 8 || !raw.every(Number.isFinite)) return null;
+  return [
+    { x: raw[0], y: raw[1] },
+    { x: raw[2], y: raw[3] },
+    { x: raw[4], y: raw[5] },
+    { x: raw[6], y: raw[7] },
+  ];
+}
+
+export function rectPolygon(rect: { x: number; y: number; w: number; h: number }): Quad {
+  return [
+    { x: rect.x, y: rect.y },
+    { x: rect.x + rect.w, y: rect.y },
+    { x: rect.x + rect.w, y: rect.y + rect.h },
+    { x: rect.x, y: rect.y + rect.h },
+  ];
+}
+
+export function viewportPolygon(viewport: Size): Quad {
+  return rectPolygon({ x: 0, y: 0, w: viewport.width, h: viewport.height });
+}
+
+/** Project a normalized unit-square point into an arbitrary content quad. */
+export function projectUnitPoint(point: Point, quad: Quad): Point {
+  const [p0, p1, p2, p3] = quad;
+  const dx1 = p1.x - p2.x;
+  const dx2 = p3.x - p2.x;
+  const dx3 = p0.x - p1.x + p2.x - p3.x;
+  const dy1 = p1.y - p2.y;
+  const dy2 = p3.y - p2.y;
+  const dy3 = p0.y - p1.y + p2.y - p3.y;
+
+  let a: number;
+  let b: number;
+  let c: number;
+  let d: number;
+  let e: number;
+  let f: number;
+  let g = 0;
+  let h = 0;
+
+  if (Math.abs(dx3) < 1e-9 && Math.abs(dy3) < 1e-9) {
+    a = p1.x - p0.x;
+    b = p3.x - p0.x;
+    c = p0.x;
+    d = p1.y - p0.y;
+    e = p3.y - p0.y;
+    f = p0.y;
+  } else {
+    const denominator = dx1 * dy2 - dx2 * dy1;
+    if (Math.abs(denominator) < 1e-9) {
+      return {
+        x: p0.x + (p1.x - p0.x) * point.x + (p3.x - p0.x) * point.y,
+        y: p0.y + (p1.y - p0.y) * point.x + (p3.y - p0.y) * point.y,
+      };
+    }
+    g = (dx3 * dy2 - dx2 * dy3) / denominator;
+    h = (dx1 * dy3 - dx3 * dy1) / denominator;
+    a = p1.x - p0.x + g * p1.x;
+    b = p3.x - p0.x + h * p3.x;
+    c = p0.x;
+    d = p1.y - p0.y + g * p1.y;
+    e = p3.y - p0.y + h * p3.y;
+    f = p0.y;
+  }
+
+  const scale = g * point.x + h * point.y + 1;
+  return {
+    x: (a * point.x + b * point.y + c) / scale,
+    y: (d * point.x + e * point.y + f) / scale,
+  };
+}
+
+export function projectPolygon(polygon: Polygon, edge: ProjectiveEdge): Polygon {
+  const { width, height } = edge.sourceViewport;
+  if (width <= 0 || height <= 0) return [];
+  return polygon.map((point) =>
+    projectUnitPoint({ x: point.x / width, y: point.y / height }, edge.destinationQuad),
+  );
+}
+
+function polygonSignedArea(points: Polygon): number {
+  let area = 0;
+  for (let index = 0; index < points.length; index += 1) {
+    const current = points[index];
+    const next = points[(index + 1) % points.length];
+    area += current.x * next.y - next.x * current.y;
+  }
+  return area / 2;
+}
+
+function edgeIntersection(start: Point, end: Point, clipStart: Point, clipEnd: Point): Point {
+  const subjectX = end.x - start.x;
+  const subjectY = end.y - start.y;
+  const clipX = clipEnd.x - clipStart.x;
+  const clipY = clipEnd.y - clipStart.y;
+  const denominator = subjectX * clipY - subjectY * clipX;
+  if (Math.abs(denominator) < 1e-9) return end;
+  const offsetX = clipStart.x - start.x;
+  const offsetY = clipStart.y - start.y;
+  const scale = (offsetX * clipY - offsetY * clipX) / denominator;
+  return { x: start.x + scale * subjectX, y: start.y + scale * subjectY };
+}
+
+/** Clip a polygon to a convex polygon, preserving every resulting vertex. */
+export function clipPolygon(subject: Polygon, clip: Polygon): Polygon {
+  if (subject.length < 3 || clip.length < 3) return [];
+  const orientation = polygonSignedArea(clip) >= 0 ? 1 : -1;
+  let output = [...subject];
+  for (let edge = 0; edge < clip.length && output.length > 0; edge += 1) {
+    const clipStart = clip[edge];
+    const clipEnd = clip[(edge + 1) % clip.length];
+    const input = output;
+    output = [];
+    const inside = (point: Point) =>
+      orientation *
+        ((clipEnd.x - clipStart.x) * (point.y - clipStart.y) -
+          (clipEnd.y - clipStart.y) * (point.x - clipStart.x)) >=
+      -1e-9;
+    let start = input[input.length - 1];
+    for (const end of input) {
+      const startInside = inside(start);
+      const endInside = inside(end);
+      if (endInside) {
+        if (!startInside) output.push(edgeIntersection(start, end, clipStart, clipEnd));
+        output.push(end);
+      } else if (startInside) {
+        output.push(edgeIntersection(start, end, clipStart, clipEnd));
+      }
+      start = end;
+    }
+  }
+  return output;
+}
+
+export function polygonBounds(points: Polygon): ViewportRect | null {
+  if (points.length < 3) return null;
+  const xs = points.map((point) => point.x);
+  const ys = points.map((point) => point.y);
+  const x = Math.min(...xs);
+  const y = Math.min(...ys);
+  const right = Math.max(...xs);
+  const bottom = Math.max(...ys);
+  if (right <= x || bottom <= y) return null;
+  return { x, y, width: right - x, height: bottom - y };
+}
+
+export function regionBounds(region: Region): ViewportRect | null {
+  return polygonBounds(region.flat());
+}
+
+export function polygonArea(points: Polygon): number {
+  return Math.abs(polygonSignedArea(points));
+}
+
+export function polygonCentroid(points: Polygon): Point | null {
+  if (points.length < 3) return null;
+  let twiceArea = 0;
+  let x = 0;
+  let y = 0;
+  for (let index = 0; index < points.length; index += 1) {
+    const current = points[index];
+    const next = points[(index + 1) % points.length];
+    const cross = current.x * next.y - next.x * current.y;
+    twiceArea += cross;
+    x += (current.x + next.x) * cross;
+    y += (current.y + next.y) * cross;
+  }
+  if (Math.abs(twiceArea) < 1e-9) return null;
+  return { x: x / (3 * twiceArea), y: y / (3 * twiceArea) };
+}
+
+export function projectAndClipRegion(
+  region: Region,
+  edges: ProjectiveEdge[],
+  topViewport: Size,
+): Region {
+  let projected = region;
+  for (const edge of edges) {
+    const sourceClip = viewportPolygon(edge.sourceViewport);
+    projected = projected
+      .map((polygon) => clipPolygon(polygon, sourceClip))
+      .filter((polygon) => polygon.length >= 3)
+      .map((polygon) =>
+        [edge.destinationQuad, ...(edge.destinationClips ?? [])].reduce(
+          (current, clip) => clipPolygon(current, clip),
+          projectPolygon(polygon, edge),
+        ),
+      )
+      .filter((polygon) => polygon.length >= 3);
+  }
+  const topClip = viewportPolygon(topViewport);
+  return projected
+    .map((polygon) => clipPolygon(polygon, topClip))
+    .filter((polygon) => polygon.length >= 3);
+}
+
+export function projectRegionToViewport(region: Region, projection: GeometryProjection): Region {
+  const clipped = region
+    .map((polygon) =>
+      projection.sourceClips.reduce((current, clip) => clipPolygon(current, clip), polygon),
+    )
+    .filter((polygon) => polygon.length >= 3);
+  return projectAndClipRegion(clipped, projection.edges, projection.topViewport);
+}
+
+export function projectRectToViewport(
+  rect: { x: number; y: number; w: number; h: number } | null,
+  projection: GeometryProjection,
+): ViewportRect | null {
+  if (!rect) return null;
+  return regionBounds(projectRegionToViewport([rectPolygon(rect)], projection));
+}
+
+export function childFrameProjection(
+  parent: GeometryProjection,
+  ownerRectInParent: { x: number; y: number; w: number; h: number },
+): GeometryProjection {
+  const destinationQuad = rectPolygon(ownerRectInParent);
+  return {
+    sourceClips: [],
+    edges: [
+      {
+        sourceViewport: { width: ownerRectInParent.w, height: ownerRectInParent.h },
+        destinationQuad,
+        destinationClips: parent.sourceClips,
+      },
+      ...parent.edges,
+    ],
+    topViewport: parent.topViewport,
+  };
+}

apps/extension/src/tools/human-loop.ts

@@ -8,6 +8,7 @@ import {
   HELP_REQUEST,
   type HelpCancelMessage,
   type HelpFinishMessage,
+  type HelpHighlightRect,
   type HelpQueryResponse,
   type HelpRequestMessage,
   isHelpAckMessage,
@@ -25,6 +26,7 @@ import type {
   ResolvedTarget,
   RpcError,
 } from "@/transport/types";
+import { resolveNodeGeometry } from "./frame-geometry";
 import {
   type CdpRunner,
   type ChromeTabsApi,
@@ -37,7 +39,6 @@ import {
 import { lookupSnapshotRef } from "./snapshot-ref";
 
 const DEFAULT_HELP_TIMEOUT_MS = 300_000;
-const HELP_ATTR = "data-bsk-help";
 const HELP_SEND_RETRIES = 3;
 const HELP_SEND_RETRY_DELAY_MS = 350;
 const HELP_REARM_DEBOUNCE_MS = 150;
@@ -74,6 +75,7 @@ interface ActiveHelpRequest {
   title?: string;
   targets: HelpTarget[];
   selectors: string[];
+  rects: HelpHighlightRect[];
   timeoutMs: number;
   notificationId: string;
   resolvedTargets?: ResolvedTarget[];
@@ -134,6 +136,7 @@ function helpRequestMessage(help: ActiveHelpRequest, tabId: number): HelpRequest
     ...(help.title ? { title: help.title } : {}),
     displayMode: isPrimaryTab ? "full" : "compact",
     selectors: isPrimaryTab ? help.selectors : [],
+    rects: isPrimaryTab ? help.rects : [],
     timeoutMs: help.timeoutMs,
   };
 }
@@ -165,33 +168,6 @@ async function sendHelpRequestWithAck(
   throw lastError ?? new Error("failed to show help overlay");
 }
 
-async function tagRefTarget(
-  cdp: CdpRunner,
-  tabId: number,
-  backendNodeId: number,
-  index: number,
-): Promise<string | null> {
-  let objectId: string | undefined;
-  try {
-    const resolved = await cdp.send<{ object?: { objectId?: string } }>(tabId, "DOM.resolveNode", {
-      backendNodeId,
-    });
-    objectId = resolved.object?.objectId;
-    if (!objectId) return null;
-    await cdp.send(tabId, "Runtime.callFunctionOn", {
-      objectId,
-      functionDeclaration: `function(){ this.setAttribute("${HELP_ATTR}", "${index}"); }`,
-    });
-    return `[${HELP_ATTR}="${index}"]`;
-  } catch {
-    return null;
-  } finally {
-    if (objectId) {
-      await cdp.send(tabId, "Runtime.releaseObject", { objectId }).catch(() => {});
-    }
-  }
-}
-
 async function selectorExists(
   cdp: CdpRunner,
   tabId: number,
@@ -209,36 +185,22 @@ async function selectorExists(
   }
 }
 
-async function clearRefTags(cdp: CdpRunner | undefined, tabId: number): Promise<void> {
-  if (!cdp) return;
-  try {
-    const doc = await cdp.send<{ root?: { nodeId?: number } }>(tabId, "DOM.getDocument", {
-      depth: 0,
-    });
-    const rootId = doc.root?.nodeId;
-    if (rootId === undefined) return;
-    const { nodeIds } = await cdp.send<{ nodeIds: number[] }>(tabId, "DOM.querySelectorAll", {
-      nodeId: rootId,
-      selector: `[${HELP_ATTR}]`,
-    });
-    for (const nodeId of nodeIds ?? []) {
-      await cdp.send(tabId, "DOM.removeAttribute", { nodeId, name: HELP_ATTR }).catch(() => {});
-    }
-  } catch {
-    // Best-effort cleanup.
-  }
-}
-
 async function resolveHelpTargets(
   ctx: ReturnType<SessionManager["get"]>,
   tabId: number,
   targets: HelpTarget[],
   deps: RequestHelpDeps,
-): Promise<{ selectors: string[]; resolvedTargets?: ResolvedTarget[] }> {
+  options: { scrollIntoView: boolean },
+): Promise<{
+  selectors: string[];
+  rects: HelpHighlightRect[];
+  resolvedTargets?: ResolvedTarget[];
+}> {
   const selectors: string[] = [];
+  const rects: HelpHighlightRect[] = [];
   const resolved: ResolvedTarget[] = [];
   let rootNodeId: number | undefined;
-  if (deps.cdp && targets.length > 0) {
+  if (deps.cdp && targets.some((target) => Boolean(target.selector))) {
     try {
       const doc = await deps.cdp.send<{ root?: { nodeId?: number } }>(tabId, "DOM.getDocument", {
         depth: 0,
@@ -262,28 +224,55 @@ async function resolveHelpTargets(
       resolved.push({ matched, selector: tgt.selector });
     } else if (tgt.ref) {
       const looked = ctx ? lookupSnapshotRef(ctx, tgt.ref, tabId) : null;
-      const backendNodeId = looked?.backendNodeId ?? null;
-      let sel: string | null = null;
-      if (backendNodeId !== null && deps.cdp) {
-        sel = await tagRefTarget(deps.cdp, tabId, backendNodeId, i);
+      let rect: HelpHighlightRect | null = null;
+      if (looked && deps.cdp) {
+        const target = {
+          tabId,
+          ...(looked.cdpSessionId ? { sessionId: looked.cdpSessionId } : {}),
+        };
+        const geometry = await resolveNodeGeometry(
+          deps.cdp,
+          tabId,
+          {
+            target,
+            backendNodeId: looked.backendNodeId,
+            ...(looked.frameId ? { frameId: looked.frameId } : {}),
+          },
+          { scrollIntoView: options.scrollIntoView },
+        );
+        if (!isRpcError(geometry)) {
+          rect = {
+            top: geometry.topBounds.y,
+            left: geometry.topBounds.x,
+            width: geometry.topBounds.width,
+            height: geometry.topBounds.height,
+          };
+        }
       }
-      if (sel) selectors.push(sel);
-      resolved.push({ matched: sel !== null, ref: tgt.ref });
+      if (rect) rects.push(rect);
+      resolved.push({
+        matched: rect !== null,
+        ref: tgt.ref,
+      });
     }
   }
 
-  return { selectors, resolvedTargets: resolved.length > 0 ? resolved : undefined };
+  return { selectors, rects, resolvedTargets: resolved.length > 0 ? resolved : undefined };
 }
 
-async function refreshHelpTargets(help: ActiveHelpRequest): Promise<void> {
-  await clearRefTags(help.deps.cdp, help.primaryTabId);
-  const { selectors, resolvedTargets } = await resolveHelpTargets(
+async function refreshHelpTargets(
+  help: ActiveHelpRequest,
+  options: { scrollIntoView: boolean } = { scrollIntoView: true },
+): Promise<void> {
+  const { selectors, rects, resolvedTargets } = await resolveHelpTargets(
     help.ctx,
     help.primaryTabId,
     help.targets,
     help.deps,
+    options,
   );
   help.selectors = selectors;
+  help.rects = rects;
   help.resolvedTargets = resolvedTargets;
 }
 
@@ -294,12 +283,7 @@ async function cleanupHelp(help: ActiveHelpRequest): Promise<void> {
   const tabsToCancel = new Set([help.primaryTabId, ...help.overlayTabIds]);
   await Promise.all(
     [...tabsToCancel].map((tabId) =>
-      Promise.all([
-        clearRefTags(help.deps.cdp, tabId),
-        help.deps
-          .sendToTab(tabId, { type: HELP_CANCEL, requestId: help.requestId })
-          .catch(() => {}),
-      ]),
+      help.deps.sendToTab(tabId, { type: HELP_CANCEL, requestId: help.requestId }).catch(() => {}),
     ),
   );
 }
@@ -476,7 +460,9 @@ function attachHelpRuntimeListener(deps: RequestHelpDeps): () => void {
           sendResponse({ active: false });
           return;
         }
-        if (tabId === help.primaryTabId) await refreshHelpTargets(help);
+        if (tabId === help.primaryTabId) {
+          await refreshHelpTargets(help, { scrollIntoView: false });
+        }
         help.overlayTabIds.add(tabId);
         sendResponse({
           active: true,
@@ -486,6 +472,7 @@ function attachHelpRuntimeListener(deps: RequestHelpDeps): () => void {
             ...(help.title ? { title: help.title } : {}),
             displayMode: tabId === help.primaryTabId ? "full" : "compact",
             selectors: tabId === help.primaryTabId ? help.selectors : [],
+            rects: tabId === help.primaryTabId ? help.rects : [],
             timeoutMs: help.timeoutMs,
           },
         });
@@ -686,7 +673,13 @@ export async function handleRequestHelp(
   if (deps.autoAttachLifecycle !== false) ensureHelpLifecycleListeners(deps);
   const tabId = target.tabId;
   const initialTargets = params.targets ?? [];
-  const { selectors, resolvedTargets } = await resolveHelpTargets(ctx, tabId, initialTargets, deps);
+  const { selectors, rects, resolvedTargets } = await resolveHelpTargets(
+    ctx,
+    tabId,
+    initialTargets,
+    deps,
+    { scrollIntoView: true },
+  );
 
   await deps.windows.update(target.windowId, { focused: true }).catch(() => {});
   await deps.activateTab(tabId).catch(() => {});
@@ -729,6 +722,7 @@ export async function handleRequestHelp(
       ...(params.title ? { title: params.title } : {}),
       targets: initialTargets,
       selectors,
+      rects,
       timeoutMs,
       notificationId,
       resolvedTargets,

apps/extension/src/tools/interaction.ts

@@ -3,14 +3,15 @@
 //
 // All interaction tools:
 // 1. Resolve target tab (sandbox: must be inside Agent Window).
-// 2. Resolve target element by `ref` (RefStore.resolve with tabId
-//    binding) or `selector` (DOM.querySelector + describeNode).
-// 3. Scroll the node into view, then dispatch the appropriate
-//    `Input.*` CDP events.
+// 2. Resolve target element by `ref` (compound frame/session identity)
+//    or `selector` (DOM.querySelector + describeNode).
+// 3. Scroll the node and its frame owners into view, project its live
+//    content quad into the top viewport, then dispatch `Input.*` events.
 // 4. Honour `AbortSignal` so canceled calls don't issue follow-up CDP
 //    commands.
 
 import { ChromiumCdp } from "@/browser-driver/chromium-cdp";
+import type { CdpTarget } from "@/browser-driver/frame-graph";
 import type { SessionContext, SessionManager } from "@/session-manager/manager";
 import type {
   ClickParams,
@@ -28,17 +29,13 @@ import type {
   SelectResult,
 } from "@/transport/types";
 import { attachDialogs, markDialogCursor } from "./dialogs";
-import {
-  backendNodeToObject,
-  boxCentre,
-  nodeCentre,
-  quadCentre,
-  scrollNodeIntoView,
-} from "./element-geometry";
+import { backendNodeToObject } from "./element-geometry";
 import { rpcError } from "./errors";
+import { resolveNodeGeometry, scrollElementAndFramesIntoView } from "./frame-geometry";
 import {
   type CdpRunner,
   type ChromeTabsApi,
+  cdpRunnerForTarget,
   chromeTabsApi,
   enforceAgentWindow,
   isRpcError,
@@ -137,7 +134,16 @@ async function resolveBackendNode(
   target: { tabId: number },
   params: { ref?: string; selector?: string },
   toolName: string,
-): Promise<{ backendNodeId: number; usedRef?: string; usedSelector?: string } | RpcError> {
+): Promise<
+  | {
+      backendNodeId: number;
+      cdpTarget: CdpTarget;
+      frameId?: string;
+      usedRef?: string;
+      usedSelector?: string;
+    }
+  | RpcError
+> {
   const hasRef = typeof params.ref === "string" && params.ref.length > 0;
   const hasSelector = typeof params.selector === "string" && params.selector.length > 0;
   if (hasRef && hasSelector) {
@@ -153,9 +159,17 @@ async function resolveBackendNode(
     };
   }
   if (hasRef) {
-    const resolved = resolveSnapshotRef(ctx, params.ref as string, target.tabId);
+    const resolved = resolveSnapshotRef(ctx, params.ref as string, target.tabId, "interact");
     if (isRpcError(resolved)) return resolved;
-    return { backendNodeId: resolved.backendNodeId, usedRef: resolved.refKey };
+    return {
+      backendNodeId: resolved.backendNodeId,
+      cdpTarget: {
+        tabId: target.tabId,
+        ...(resolved.cdpSessionId ? { sessionId: resolved.cdpSessionId } : {}),
+      },
+      ...(resolved.frameId ? { frameId: resolved.frameId } : {}),
+      usedRef: resolved.refKey,
+    };
   }
   // selector path
   try {
@@ -192,7 +206,7 @@ async function resolveBackendNode(
         message: "DOM.describeNode returned no backendNodeId",
       };
     }
-    return { backendNodeId, usedSelector: params.selector };
+    return { backendNodeId, cdpTarget: { tabId: target.tabId }, usedSelector: params.selector };
   } catch (err) {
     return {
       code: "cdp_failed",
@@ -228,19 +242,19 @@ export async function handleClick(
     return { code: "cancelled", message: "click aborted" };
   }
 
-  try {
-    deps.cdp.trackSessionTab?.(ctx.sessionId, target.tabId);
-    const scrollErr = await scrollNodeIntoView(deps.cdp, target.tabId, node.backendNodeId);
-    if (scrollErr) return scrollErr;
-  } catch (err) {
-    return {
-      code: "cdp_failed",
-      message: err instanceof Error ? err.message : String(err),
-    };
-  }
-
-  const centre = await nodeCentre(deps.cdp, target.tabId, node.backendNodeId);
-  if (isRpcError(centre)) return centre;
+  deps.cdp.trackSessionTab?.(ctx.sessionId, target.tabId);
+  const geometry = await resolveNodeGeometry(
+    deps.cdp,
+    target.tabId,
+    {
+      target: node.cdpTarget,
+      backendNodeId: node.backendNodeId,
+      ...(node.frameId ? { frameId: node.frameId } : {}),
+    },
+    { scrollIntoView: true },
+  );
+  if (isRpcError(geometry)) return geometry;
+  const centre = geometry.actionPoint;
 
   if (throwIfAborted(deps.signal)) {
     return { code: "cancelled", message: "click aborted" };
@@ -353,19 +367,19 @@ export async function handleHover(
   const node = await resolveBackendNode(deps.cdp, ctx, target, params, "hover");
   if (isRpcError(node)) return node;
 
-  try {
-    deps.cdp.trackSessionTab?.(ctx.sessionId, target.tabId);
-    const scrollErr = await scrollNodeIntoView(deps.cdp, target.tabId, node.backendNodeId);
-    if (scrollErr) return scrollErr;
-  } catch (err) {
-    return {
-      code: "cdp_failed",
-      message: err instanceof Error ? err.message : String(err),
-    };
-  }
-
-  const centre = await nodeCentre(deps.cdp, target.tabId, node.backendNodeId);
-  if (isRpcError(centre)) return centre;
+  deps.cdp.trackSessionTab?.(ctx.sessionId, target.tabId);
+  const geometry = await resolveNodeGeometry(
+    deps.cdp,
+    target.tabId,
+    {
+      target: node.cdpTarget,
+      backendNodeId: node.backendNodeId,
+      ...(node.frameId ? { frameId: node.frameId } : {}),
+    },
+    { scrollIntoView: true },
+  );
+  if (isRpcError(geometry)) return geometry;
+  const centre = geometry.actionPoint;
 
   if (throwIfAborted(deps.signal)) {
     return { code: "cancelled", message: "hover aborted" };
@@ -585,10 +599,11 @@ export async function handleFill(
 
   const node = await resolveBackendNode(deps.cdp, ctx, target, params, "fill");
   if (isRpcError(node)) return node;
+  const nodeCdp = cdpRunnerForTarget(deps.cdp, node.cdpTarget);
 
   try {
     deps.cdp.trackSessionTab?.(ctx.sessionId, target.tabId);
-    const described = await deps.cdp.send<{ node?: DescribedNode }>(
+    const described = await nodeCdp.send<{ node?: DescribedNode }>(
       target.tabId,
       "DOM.describeNode",
       {
@@ -602,12 +617,18 @@ export async function handleFill(
         `element ${described.node?.nodeName ?? "?"} not fillable (need input/textarea/contenteditable)`,
       );
     }
-    const scrollErr = await scrollNodeIntoView(deps.cdp, target.tabId, node.backendNodeId);
+    const scrollErr = await scrollElementAndFramesIntoView(
+      deps.cdp,
+      target.tabId,
+      node.cdpTarget,
+      node.backendNodeId,
+      node.frameId,
+    );
     if (scrollErr) return scrollErr;
     if (throwIfAborted(deps.signal)) {
       return { code: "cancelled", message: "fill aborted" };
     }
-    await deps.cdp.send(target.tabId, "DOM.focus", { backendNodeId: node.backendNodeId });
+    await nodeCdp.send(target.tabId, "DOM.focus", { backendNodeId: node.backendNodeId });
   } catch (err) {
     return {
       code: "cdp_failed",
@@ -619,7 +640,7 @@ export async function handleFill(
     return { code: "cancelled", message: "fill aborted" };
   }
 
-  const objectIdOrErr = await backendNodeToObject(deps.cdp, target.tabId, node.backendNodeId);
+  const objectIdOrErr = await backendNodeToObject(nodeCdp, target.tabId, node.backendNodeId);
   if (isRpcError(objectIdOrErr)) return objectIdOrErr;
   const objectId = objectIdOrErr;
   const clearBefore = params.clear_before ?? true;
@@ -628,7 +649,7 @@ export async function handleFill(
     if (clearBefore) {
       // Clear input/textarea value or wipe contenteditable innerText,
       // then fire `input` so frameworks observe the empty state.
-      await deps.cdp.send(target.tabId, "Runtime.callFunctionOn", {
+      await nodeCdp.send(target.tabId, "Runtime.callFunctionOn", {
         objectId,
         functionDeclaration: `function() {
           if (this.isContentEditable) { this.textContent = ''; }
@@ -655,7 +676,7 @@ export async function handleFill(
       return { code: "cancelled", message: "fill aborted" };
     }
     // Fire `input` + `change` so React / Vue controlled inputs commit.
-    await deps.cdp.send(target.tabId, "Runtime.callFunctionOn", {
+    await nodeCdp.send(target.tabId, "Runtime.callFunctionOn", {
       objectId,
       functionDeclaration: `function() {
         this.dispatchEvent(new Event('input', { bubbles: true }));
@@ -853,14 +874,21 @@ export async function handlePress(
   if (params.ref || params.selector) {
     const node = await resolveBackendNode(deps.cdp, ctx, target, params, "press");
     if (isRpcError(node)) return node;
+    const nodeCdp = cdpRunnerForTarget(deps.cdp, node.cdpTarget);
     try {
       deps.cdp.trackSessionTab?.(ctx.sessionId, target.tabId);
-      const scrollErr = await scrollNodeIntoView(deps.cdp, target.tabId, node.backendNodeId);
+      const scrollErr = await scrollElementAndFramesIntoView(
+        deps.cdp,
+        target.tabId,
+        node.cdpTarget,
+        node.backendNodeId,
+        node.frameId,
+      );
       if (scrollErr) return scrollErr;
       if (throwIfAborted(deps.signal)) {
         return { code: "cancelled", message: "press aborted" };
       }
-      await deps.cdp.send(target.tabId, "DOM.focus", { backendNodeId: node.backendNodeId });
+      await nodeCdp.send(target.tabId, "DOM.focus", { backendNodeId: node.backendNodeId });
     } catch (err) {
       return {
         code: "cdp_failed",
@@ -965,10 +993,11 @@ export async function handleSelect(
 
   const node = await resolveBackendNode(deps.cdp, ctx, target, params, "select");
   if (isRpcError(node)) return node;
+  const nodeCdp = cdpRunnerForTarget(deps.cdp, node.cdpTarget);
 
   try {
     deps.cdp.trackSessionTab?.(ctx.sessionId, target.tabId);
-    const described = await deps.cdp.send<{ node?: DescribedNode }>(
+    const described = await nodeCdp.send<{ node?: DescribedNode }>(
       target.tabId,
       "DOM.describeNode",
       { backendNodeId: node.backendNodeId },
@@ -991,12 +1020,18 @@ export async function handleSelect(
         "single-select <select> requires exactly one value",
       );
     }
-    const scrollErr = await scrollNodeIntoView(deps.cdp, target.tabId, node.backendNodeId);
+    const scrollErr = await scrollElementAndFramesIntoView(
+      deps.cdp,
+      target.tabId,
+      node.cdpTarget,
+      node.backendNodeId,
+      node.frameId,
+    );
     if (scrollErr) return scrollErr;
     if (throwIfAborted(deps.signal)) {
       return { code: "cancelled", message: "select aborted" };
     }
-    await deps.cdp.send(target.tabId, "DOM.focus", { backendNodeId: node.backendNodeId });
+    await nodeCdp.send(target.tabId, "DOM.focus", { backendNodeId: node.backendNodeId });
   } catch (err) {
     return {
       code: "cdp_failed",
@@ -1008,12 +1043,12 @@ export async function handleSelect(
     return { code: "cancelled", message: "select aborted" };
   }
 
-  const objectIdOrErr = await backendNodeToObject(deps.cdp, target.tabId, node.backendNodeId);
+  const objectIdOrErr = await backendNodeToObject(nodeCdp, target.tabId, node.backendNodeId);
   if (isRpcError(objectIdOrErr)) return objectIdOrErr;
   const objectId = objectIdOrErr;
 
   try {
-    const evaluated = await deps.cdp.send<{
+    const evaluated = await nodeCdp.send<{
       result?: { value?: SelectMutationResult | null };
     }>(target.tabId, "Runtime.callFunctionOn", {
       objectId,
@@ -1072,9 +1107,6 @@ export async function handleSelect(
 
 export const __testing__ = {
   DEFAULT_TIMEOUT_MS,
-  quadCentre,
-  boxCentre,
   resolveBackendNode,
-  nodeCentre,
   isFillable,
 };

apps/extension/src/tools/record.ts

@@ -15,31 +15,26 @@ import {
   type RecordQueryResponse,
   type RecordStartAck,
   type RecordStartMessage,
+  type RecordStepAck,
   type RecordStopMessage,
 } from "@/lib/record-bridge";
-import { DEFAULT_MAX_PAGE_TOKENS } from "@/lib/record-constants";
 import {
-  applyTargetMatching,
-  buildTraceV3,
-  cancelPendingSettles,
-  captureAndRegisterObservation,
-  clearPendingRedirectLanding,
-  createObservationState,
-  flushPendingRedirectLanding,
-  flushPendingSettles,
-  inferMissingPostStates,
-  type RecordingObservationState,
-  rememberStepOnPage,
-  scheduleDraftSettle,
-  scheduleRedirectLandingFlush,
-  settleUnsettledDrafts,
-} from "@/lib/record-observation";
-import { appendRecordedPayload, observeRecordedNavigation } from "@/lib/recording-step-buffer";
-import { buildTraceV2 } from "@/lib/trace-reducer-v2";
+  type RecordFrameCoordinator,
+  type RecordingCaptureScope,
+  recordFrameCoordinator,
+} from "@/lib/recording/frame-coordinator";
+import { RecordingObservationRuntime } from "@/lib/recording/recording-runtime";
+import {
+  appendRecordedPayload,
+  observeRecordedNavigation,
+  type RecordingStepBuffer,
+} from "@/lib/recording/step-buffer";
+import { RecordingTabCoordinator, type TabActivation } from "@/lib/recording/tab-coordinator";
+import { buildTraceV2 } from "@/lib/recording/trace-reducer-v2";
+import type { RecordingDraftStep } from "@/lib/recording/types";
 import type { SessionManager } from "@/session-manager/manager";
 import { EXTENSION_VERSION } from "@/transport/handshake";
 import type {
-  DraftTraceStep,
   RecordAwaitParams,
   RecordAwaitResult,
   RecordedTrace,
@@ -50,7 +45,7 @@ import type {
   RpcError,
   StopReason,
 } from "@/transport/types";
-import { TRACE_VERSION } from "@/transport/types";
+import { TRACE_VERSION_V3 } from "@/transport/types";
 import { handleNavigate } from "./navigation";
 import {
   type CdpRunner,
@@ -63,30 +58,22 @@ import {
 
 interface ActiveRecording {
   requestId: string;
-  tabId: number;
+  tabs: RecordingTabCoordinator;
   agentWindowId: number;
   startUrl?: string;
   purpose?: string;
-  steps: DraftTraceStep[];
+  steps: RecordingDraftStep[];
   startedAt: string;
   startedAtMs: number;
   traceVersion: 2 | 3;
+  supportsTabSwitchSteps: boolean;
   finishPromise: Promise<RecordedTrace>;
   resolveFinish: (trace: RecordedTrace) => void;
   rejectFinish: (err: Error) => void;
   settled: boolean;
-  finishing: boolean;
-  currentUrl?: string;
-  pendingNavigation: boolean;
-  pendingNavigationDeadline?: number;
-  observation: RecordingObservationState;
+  finishAttempt: Promise<RecordedTrace | null> | null;
+  observation: RecordingObservationRuntime | null;
   stoppedBy: StopReason;
-  maxPageTokens: number;
-  redactValues: boolean;
-  /** Tabs whose content capture acknowledged START and still need STOP. */
-  armedTabIds: Set<number>;
-  /** Rearms already sending START when finish begins. */
-  rearmCallbacks: Set<Promise<void>>;
   /** Navigation callbacks tracked from event receipt through action enqueue. */
   navigationCallbacks: Set<Promise<void>>;
   /** Synchronous intake gate closed only after finish drains to stability. */
@@ -96,16 +83,21 @@ interface ActiveRecording {
    * in `steps` before a same-turn `webNavigation` tries to annotate it.
    */
   actionQueue: Promise<void>;
-  /**
-   * CDP runner captured at `record_start`. The message listeners are attached
-   * once at service-worker startup, where no runner exists yet, so observation
-   * capture must go through the recording instead of the listener deps.
-   */
-  cdp?: CdpRunner;
+  /** Last accepted sequence for each content-script document producer. */
+  lastStepSequenceByProducer: Map<string, number>;
 }
 
-function enqueueRecordingAction(recording: ActiveRecording, task: () => Promise<void>): void {
-  recording.actionQueue = recording.actionQueue.then(task, task).catch(() => {});
+function enqueueRecordingAction(
+  recording: ActiveRecording,
+  task: () => Promise<void>,
+): Promise<void> {
+  const queued = recording.actionQueue.then(task, task);
+  recording.actionQueue = queued.catch(() => {});
+  return queued;
+}
+
+function isRecordingFinishing(recording: ActiveRecording): boolean {
+  return recording.settled || recording.finishAttempt !== null;
 }
 
 const recordings = new Map<string, ActiveRecording>();
@@ -188,9 +180,11 @@ async function sendRecordStartWithAck(
   tabId: number,
   msg: RecordStartMessage,
   sendToTab: RecordDeps["sendToTab"],
+  cancelled?: () => boolean,
 ): Promise<void> {
   let lastError: unknown;
   for (let attempt = 0; attempt < RECORD_START_RETRIES; attempt += 1) {
+    if (cancelled?.()) throw new Error("record start cancelled");
     try {
       const response = await sendToTab(tabId, msg);
       if (isRecordStartAck(response)) return;
@@ -198,6 +192,7 @@ async function sendRecordStartWithAck(
     } catch (err) {
       lastError = err;
     }
+    if (cancelled?.()) throw new Error("record start cancelled");
     if (attempt + 1 < RECORD_START_RETRIES) {
       await sleep(RECORD_START_RETRY_DELAY_MS);
     }
@@ -207,25 +202,28 @@ async function sendRecordStartWithAck(
 
 function negotiatedTraceVersion(params: RecordStartParams): 2 | 3 | RpcError {
   if (params.trace_version === undefined) return 2;
-  if (params.trace_version === TRACE_VERSION) return 3;
+  if (params.trace_version === TRACE_VERSION_V3) return 3;
   return {
     code: "invalid_params",
-    message: `unsupported trace_version ${params.trace_version}; supported values are omitted (v2) or ${TRACE_VERSION} (v3)`,
+    message: `unsupported trace_version ${params.trace_version}; supported values are omitted (v2) or ${TRACE_VERSION_V3} (v3)`,
   };
 }
 
 function buildTrace(recording: ActiveRecording): RecordedTrace {
-  if (recording.traceVersion === 3) {
-    return buildTraceV3({
-      obs: recording.observation,
-      steps: recording.steps,
+  if (recording.traceVersion === 3 && recording.observation) {
+    return recording.observation.buildTrace({
+      drafts: recording.steps,
       startedAt: recording.startedAt,
-      ...(recording.purpose ? { purpose: recording.purpose } : {}),
+      purpose: recording.purpose,
       startUrl: recording.startUrl,
       stoppedBy: recording.stoppedBy,
       bskVersion: BSK_TRACE_VERSION,
+      includeTabSwitches: recording.supportsTabSwitchSteps,
     });
   }
+  if (recording.traceVersion === 3) {
+    throw new Error("trace v3 observation runtime is unavailable");
+  }
   return buildTraceV2({
     steps: recording.steps,
     startedAt: recording.startedAt,
@@ -234,33 +232,49 @@ function buildTrace(recording: ActiveRecording): RecordedTrace {
   });
 }
 
+function stepBufferFor(
+  recording: ActiveRecording,
+  tabId: number,
+  fallbackUrl?: string,
+): RecordingStepBuffer {
+  return {
+    steps: recording.steps,
+    navigation: recording.tabs.navigation(tabId, fallbackUrl),
+  };
+}
+
+async function activateRecordingTab(
+  recording: ActiveRecording,
+  targetTabId: number,
+): Promise<void> {
+  const previousTabId = recording.tabs.currentTabId;
+  if (targetTabId === previousTabId) return;
+
+  const transition = recording.observation
+    ? await recording.observation.captureTabTransition(previousTabId, targetTabId)
+    : {};
+  const { targetUrl, ...stateLinks } = transition;
+  const draft: Extract<RecordingDraftStep, { op: "switch_tab" }> = {
+    op: "switch_tab",
+    ...stateLinks,
+  };
+  recording.steps.push(draft);
+  recording.observation?.bindTabTransition(draft, recording.steps.length);
+  recording.tabs.commit(targetTabId, targetUrl);
+}
+
 async function processRecordedStep(
   recording: ActiveRecording,
-  deps: RecordDeps,
   draftIndex: number,
+  tabId: number,
+  producerId?: string,
 ): Promise<void> {
-  const cdp = recording.cdp ?? deps.cdp;
-  if (!cdp) return;
-  const draft = recording.steps[draftIndex];
-  if (!draft) return;
-  const stepId = draftIndex + 1;
-
-  if (draft.op === "navigate") {
-    draft.preStateId = recording.observation.lastSettled?.stateId;
-  } else {
-    applyTargetMatching(recording.observation, draft, stepId);
+  if (!recording.observation) return;
+  try {
+    await recording.observation.processDraft(tabId, recording.steps, draftIndex, producerId);
+  } catch (err) {
+    console.warn(`[bsk record] observation failed for step ${draftIndex + 1}`, err);
   }
-  rememberStepOnPage(recording.observation, draft, stepId);
-
-  scheduleDraftSettle(
-    recording.observation,
-    draftIndex,
-    stepId,
-    cdp,
-    recording.tabId,
-    deps.tabsApi,
-    recording.steps,
-  );
 }
 
 export interface RecordDeps {
@@ -270,16 +284,25 @@ export interface RecordDeps {
     msg: RecordStartMessage | RecordStopMessage | RecordCancelMessage,
   ): Promise<unknown>;
   bypassOverlay?: (tabId: number, enabled: boolean) => Promise<void>;
+  frameCoordinator?: Pick<
+    RecordFrameCoordinator,
+    "begin" | "armTab" | "sourceFor" | "stop" | "cancel"
+  >;
   cdp?: CdpRunner;
   signal?: AbortSignal;
 }
 
+export type RecordRuntimeDeps = Omit<RecordDeps, "frameCoordinator" | "signal"> & {
+  frameCoordinator: NonNullable<RecordDeps["frameCoordinator"]>;
+};
+
 let defaultDeps: RecordDeps | null = null;
 function getDefaultDeps(): RecordDeps {
   if (!defaultDeps) {
     defaultDeps = {
       tabsApi: chromeTabsApi,
       sendToTab: (tabId, msg) => chrome.tabs.sendMessage(tabId, msg),
+      frameCoordinator: recordFrameCoordinator,
     };
   }
   return defaultDeps;
@@ -340,28 +363,55 @@ export function releaseBrowserObservationListenersIfIdle(): void {
 export function attachRecordStepListener(deps: RecordDeps = getDefaultDeps()): () => void {
   const listener = (
     message: unknown,
-    _sender: chrome.runtime.MessageSender,
-    _sendResponse: () => void,
+    sender: chrome.runtime.MessageSender,
+    sendResponse: (response: RecordStepAck) => void,
   ) => {
     if (!isRecordStepMessage(message)) return false;
     for (const recording of recordings.values()) {
       if (recording.requestId !== message.requestId) continue;
+      const source: RecordingCaptureScope | null | undefined = deps.frameCoordinator
+        ? deps.frameCoordinator.sourceFor(message.requestId, message.producerId, sender)
+        : undefined;
+      if (deps.frameCoordinator && !source) return false;
+      const sourceTabId = source?.tabId ?? sender.tab?.id ?? recording.tabs.currentTabId;
+      const sourceWasActive = sender.tab?.active ?? sourceTabId === recording.tabs.activeTabId;
+      const producerKey = source
+        ? `${source.tabId}:${source.documentId}:${source.producerId}`
+        : `${sourceTabId}:${message.producerId}`;
+      const expectedSequence = (recording.lastStepSequenceByProducer.get(producerKey) ?? 0) + 1;
+      if (message.sequence < expectedSequence) {
+        sendResponse({ ok: true, sequence: message.sequence });
+        return false;
+      }
+      if (message.sequence > expectedSequence) {
+        sendResponse({
+          ok: false,
+          expectedSequence,
+          error: `out-of-order recorded step ${message.sequence}`,
+        });
+        return false;
+      }
+
+      recording.lastStepSequenceByProducer.set(producerKey, message.sequence);
       enqueueRecordingAction(recording, async () => {
-        const cdp = recording.cdp ?? deps.cdp;
-        // Flush before appending so a coalesced OAuth navigate is ordered
-        // ahead of the click that landed on the destination page. The flush
-        // itself waits for that navigate's observation before returning.
-        await flushPendingRedirectLanding(
-          recording.observation,
-          recording.steps,
-          cdp,
-          recording.tabId,
-          deps.tabsApi,
-          { cancelled: () => recording.settled },
+        if (sourceTabId !== recording.tabs.currentTabId) {
+          if (!sourceWasActive) return;
+          if (sourceTabId !== recording.tabs.activeTabId)
+            recording.tabs.noteActivation(sourceTabId);
+          await activateRecordingTab(recording, sourceTabId);
+        }
+        await recording.observation?.flushRedirects(sourceTabId);
+        const targetHint = message.step.geometry ? { geometry: message.step.geometry } : undefined;
+        const draftIndex = appendRecordedPayload(
+          stepBufferFor(recording, sourceTabId, message.step.page_url),
+          message.step,
+          targetHint,
         );
-        const draftIndex = appendRecordedPayload(recording, message.step);
-        if (draftIndex !== null) await processRecordedStep(recording, deps, draftIndex);
+        if (draftIndex !== null) {
+          await processRecordedStep(recording, draftIndex, sourceTabId, source?.producerId);
+        }
       });
+      sendResponse({ ok: true, sequence: message.sequence });
       return false;
     }
     return false;
@@ -388,9 +438,7 @@ export function attachRecordFinishListener(deps: RecordDeps = getDefaultDeps()):
 
 function findRecordingByTabId(tabId: number): ActiveRecording | null {
   for (const recording of recordings.values()) {
-    if (!recording.settled && (recording.tabId === tabId || recording.armedTabIds.has(tabId))) {
-      return recording;
-    }
+    if (recording.tabs.currentTabId === tabId && !recording.settled) return recording;
   }
   return null;
 }
@@ -427,7 +475,7 @@ async function clearRearmTimersForRecording(
   recording: ActiveRecording,
   deps: RecordDeps,
 ): Promise<void> {
-  clearRearmTimer(recording.tabId);
+  clearRearmTimer(recording.tabs.currentTabId);
   try {
     const tabs = await deps.tabsApi.query({ windowId: recording.agentWindowId });
     for (const tab of tabs) {
@@ -442,25 +490,33 @@ async function stopRecordingOnAllAgentTabs(
   recording: ActiveRecording,
   deps: RecordDeps,
 ): Promise<void> {
+  if (deps.frameCoordinator && !(await deps.frameCoordinator.stop(recording.requestId))) {
+    throw new Error("failed to flush one or more recording documents");
+  }
   const stopMsg: RecordStopMessage = { type: RECORD_STOP, requestId: recording.requestId };
-  let failed = false;
-  for (const tabId of [...recording.armedTabIds]) {
-    try {
-      await deps.tabsApi.get(tabId);
-    } catch {
-      // A closed tab no longer has content capture to flush.
-      recording.armedTabIds.delete(tabId);
-      continue;
-    }
+  let tabIds = [recording.tabs.currentTabId];
+  try {
+    const tabs = await deps.tabsApi.query({ windowId: recording.agentWindowId });
+    tabIds = [
+      ...new Set([
+        recording.tabs.currentTabId,
+        ...tabs.flatMap((tab) => (typeof tab.id === "number" ? [tab.id] : [])),
+      ]),
+    ];
+  } catch {
+    // Fall back to the current recording tab.
+  }
+
+  for (const tabId of tabIds) {
     try {
       const response = await deps.sendToTab(tabId, stopMsg);
-      if (!isRecordStartAck(response)) {
+      if (tabId === recording.tabs.currentTabId && !isRecordStartAck(response)) {
         throw new Error("content script did not confirm recorded steps");
       }
-      recording.armedTabIds.delete(tabId);
     } catch {
-      failed = true;
-      continue;
+      if (tabId === recording.tabs.currentTabId) {
+        throw new Error("failed to flush recorded steps");
+      }
     }
     if (deps.bypassOverlay) {
       try {
@@ -470,53 +526,54 @@ async function stopRecordingOnAllAgentTabs(
       }
     }
   }
-  if (failed || recording.armedTabIds.size > 0) {
-    throw new Error("failed to flush recorded steps");
-  }
 }
 
 async function rearmRecording(
   recording: ActiveRecording,
   targetTabId: number,
   deps: RecordDeps,
+  activation?: TabActivation,
 ): Promise<boolean> {
-  let resolveTracked!: () => void;
-  const tracked = new Promise<void>((resolve) => {
-    resolveTracked = resolve;
-  });
-  recording.rearmCallbacks.add(tracked);
-
   // Do NOT toggle automation-bypass here: each retry used to increment the
   // content-script counter, and a single stop decrement left the ControlOverlay
   // stuck with pointer-events:none (page usable, Interrupt dead). RecordOverlay
   // already hides the control chrome while activeRecord is set.
-  try {
-    for (let attempt = 0; attempt < RECORD_REARM_MAX_ATTEMPTS; attempt += 1) {
-      if (recording.settled || recording.finishing) return false;
-      const startMsg: RecordStartMessage = {
-        type: RECORD_START,
-        requestId: recording.requestId,
-        startedAtMs: recording.startedAtMs,
-      };
-      try {
-        await sendRecordStartWithAck(targetTabId, startMsg, deps.sendToTab);
-        recording.armedTabIds.add(targetTabId);
-        recording.tabId = targetTabId;
-        return true;
-      } catch {
-        if (attempt + 1 < RECORD_REARM_MAX_ATTEMPTS) {
-          await sleep(RECORD_REARM_RETRY_DELAY_MS);
-        }
+  const isFinishing = () => isRecordingFinishing(recording);
+  for (let attempt = 0; attempt < RECORD_REARM_MAX_ATTEMPTS; attempt += 1) {
+    if (isFinishing()) return false;
+    const startMsg: RecordStartMessage = {
+      type: RECORD_START,
+      requestId: recording.requestId,
+      startedAtMs: recording.startedAtMs,
+    };
+    try {
+      if (deps.frameCoordinator) {
+        const frameStarted = await deps.frameCoordinator.armTab(recording.requestId, targetTabId);
+        if (!frameStarted) throw new Error("recording document did not start");
+      }
+      await sendRecordStartWithAck(targetTabId, startMsg, deps.sendToTab, isFinishing);
+      if (isFinishing()) return false;
+      if (activation) {
+        if (!recording.tabs.isLatest(activation)) return true;
+        await enqueueRecordingAction(recording, async () => {
+          if (isFinishing() || !recording.tabs.isLatest(activation)) {
+            return;
+          }
+          await activateRecordingTab(recording, targetTabId);
+        });
+      }
+      return true;
+    } catch {
+      if (isFinishing()) return false;
+      if (attempt + 1 < RECORD_REARM_MAX_ATTEMPTS) {
+        await sleep(RECORD_REARM_RETRY_DELAY_MS);
       }
     }
-    return false;
-  } finally {
-    recording.rearmCallbacks.delete(tracked);
-    resolveTracked();
   }
+  return false;
 }
 
-function scheduleRearmForTab(tabId: number, deps: RecordDeps): void {
+function scheduleRearmForTab(tabId: number, deps: RecordDeps, activation?: TabActivation): void {
   const existing = rearmTimers.get(tabId);
   if (existing) clearTimeout(existing);
   rearmTimers.set(
@@ -525,7 +582,7 @@ function scheduleRearmForTab(tabId: number, deps: RecordDeps): void {
       rearmTimers.delete(tabId);
       void (async () => {
         const current = await findRecordingForTab(tabId, deps);
-        if (current) await rearmRecording(current, tabId, deps);
+        if (current) await rearmRecording(current, tabId, deps, activation);
       })();
     }, RECORD_REARM_DEBOUNCE_MS),
   );
@@ -537,14 +594,20 @@ export function attachRecordTabListener(deps: RecordDeps = getDefaultDeps()): ()
     const windowId = tab.windowId;
     if (tabId === undefined || windowId === undefined) return;
     for (const recording of recordings.values()) {
-      if (recording.settled || recording.agentWindowId !== windowId) continue;
+      if (isRecordingFinishing(recording) || recording.agentWindowId !== windowId) continue;
       scheduleRearmForTab(tabId, deps);
       return;
     }
   };
 
   const onActivated = (activeInfo: chrome.tabs.TabActiveInfo) => {
-    scheduleRearmForTab(activeInfo.tabId, deps);
+    for (const recording of recordings.values()) {
+      if (isRecordingFinishing(recording) || recording.agentWindowId !== activeInfo.windowId)
+        continue;
+      const activation = recording.tabs.noteActivation(activeInfo.tabId);
+      scheduleRearmForTab(activeInfo.tabId, deps, activation);
+      return;
+    }
   };
 
   chrome.tabs.onCreated.addListener(onCreated);
@@ -564,80 +627,39 @@ export function attachRecordNavigationListener(deps: RecordDeps = getDefaultDeps
     transitionQualifiers?: string[],
   ) => {
     if (!url) return;
-    const direct = findRecordingByTabId(tabId);
-    const candidates = direct
-      ? direct.acceptingNavigation
-        ? [direct]
-        : []
-      : [...recordings.values()].filter(
-          (recording) => !recording.settled && recording.acceptingNavigation,
+    const candidates = [...recordings.values()].filter(
+      (recording) =>
+        !recording.settled && recording.acceptingNavigation && recording.tabs.activeTabId === tabId,
+    );
+    for (const recording of candidates) {
+      const queued = enqueueRecordingAction(recording, async () => {
+        if (tabId !== recording.tabs.currentTabId) {
+          await activateRecordingTab(recording, tabId);
+        }
+        const result = observeRecordedNavigation(
+          stepBufferFor(recording, tabId, url),
+          url,
+          causedByAction,
+          transitionType,
+          transitionQualifiers,
         );
-    if (candidates.length === 0) return;
-
-    let resolveTracked!: () => void;
-    const tracked = new Promise<void>((resolve) => {
-      resolveTracked = resolve;
-    });
-    for (const candidate of candidates) candidate.navigationCallbacks.add(tracked);
-
-    void (async () => {
-      try {
-        const recording = await findRecordingForTab(tabId, deps);
-        if (!recording || !recording.acceptingNavigation || !candidates.includes(recording)) {
+        if (result.kind === "coalesce_redirect") {
+          recording.observation?.scheduleRedirect(tabId, recording.steps, result.url);
           return;
         }
-        enqueueRecordingAction(recording, async () => {
-          const result = observeRecordedNavigation(
-            recording,
-            url,
-            causedByAction,
-            transitionType,
-            transitionQualifiers,
-          );
-          const cdp = recording.cdp ?? deps.cdp;
-          if (!cdp) return;
 
-          if (result.kind === "coalesce_redirect") {
-            scheduleRedirectLandingFlush(
-              recording.observation,
-              recording.steps,
-              cdp,
-              tabId,
-              deps.tabsApi,
-              result.url,
-              { cancelled: () => recording.settled },
-            );
-            return;
-          }
-
-          if (result.kind === "noop") return;
-
-          // A concrete navigation supersedes any in-flight redirect coalesce.
-          clearPendingRedirectLanding(recording.observation);
-
-          const draftIndex = result.index;
-          const lastDraft = recording.steps[draftIndex];
-          if (!lastDraft) return;
-
-          if (result.kind === "appended") {
-            lastDraft.preStateId = recording.observation.lastSettled?.stateId;
-            rememberStepOnPage(recording.observation, lastDraft, draftIndex + 1);
-          }
-          scheduleDraftSettle(
-            recording.observation,
-            draftIndex,
-            draftIndex + 1,
-            cdp,
-            tabId,
-            deps.tabsApi,
-            recording.steps,
-          );
-        });
-      } finally {
-        for (const candidate of candidates) candidate.navigationCallbacks.delete(tracked);
-        resolveTracked();
-      }
-    })();
+        if (result.kind === "noop") return;
+        recording.observation?.clearRedirect(tabId);
+        if (result.kind === "appended") {
+          await processRecordedStep(recording, result.index, tabId);
+        } else {
+          recording.observation?.scheduleSettle(tabId, recording.steps, result.index);
+        }
+      });
+      const tracked = queued.catch(() => {});
+      recording.navigationCallbacks.add(tracked);
+      void tracked.finally(() => recording.navigationCallbacks.delete(tracked));
+    }
   };
   const onMainFrameComplete = (tabId: number, url?: string) => {
     void (async () => {
@@ -683,34 +705,14 @@ export function attachRecordNavigationListener(deps: RecordDeps = getDefaultDeps
 
 const MAX_FINISH_DRAIN_ROUNDS = 10;
 
-async function drainRecordingToStability(
-  recording: ActiveRecording,
-  deps: RecordDeps,
-): Promise<boolean> {
-  const cdp = recording.cdp ?? deps.cdp;
+async function drainRecordingToStability(recording: ActiveRecording): Promise<boolean> {
   for (let round = 0; round < MAX_FINISH_DRAIN_ROUNDS; round += 1) {
     await Promise.all([...recording.navigationCallbacks]);
     const actionTail = recording.actionQueue;
     await actionTail;
-    await flushPendingRedirectLanding(
-      recording.observation,
-      recording.steps,
-      cdp,
-      recording.tabId,
-      deps.tabsApi,
-      { cancelled: () => recording.settled },
-    );
-    await flushPendingSettles(recording.observation);
-    const redirectTail = recording.observation.redirectFlushQueue;
-    const settleTail = recording.observation.settleQueue;
-    await Promise.all([redirectTail, settleTail]);
+    await recording.observation?.flush();
 
-    if (
-      recording.navigationCallbacks.size === 0 &&
-      recording.actionQueue === actionTail &&
-      recording.observation.redirectFlushQueue === redirectTail &&
-      recording.observation.settleQueue === settleTail
-    ) {
+    if (recording.navigationCallbacks.size === 0 && recording.actionQueue === actionTail) {
       // No event or promise continuation can interleave with this synchronous
       // check-and-close, so work accepted before the cutoff is fully drained.
       recording.acceptingNavigation = false;
@@ -739,7 +741,8 @@ export function attachRecordQueryListener(deps: RecordDeps = getDefaultDeps()):
         sendResponse({ active: false });
         return;
       }
-      await rearmRecording(recording, tabId, deps);
+      const activation = sender.tab?.active ? recording.tabs.noteActivation(tabId) : undefined;
+      await rearmRecording(recording, tabId, deps, activation);
       sendResponse({
         active: true,
         requestId: recording.requestId,
@@ -773,37 +776,38 @@ async function finishRecording(
 ): Promise<RecordedTrace | null> {
   const recording = recordings.get(sessionId);
   if (!recording || recording.settled) return null;
-  if (recording.finishing) return recording.finishPromise;
-  recording.finishing = true;
+  if (recording.finishAttempt) return recording.finishAttempt;
   recording.stoppedBy = stoppedBy;
 
+  const attempt = finishRecordingAttempt(sessionId, recording, deps);
+  recording.finishAttempt = attempt;
+  try {
+    return await attempt;
+  } finally {
+    if (!recording.settled) {
+      recording.finishAttempt = null;
+    }
+  }
+}
+
+async function finishRecordingAttempt(
+  sessionId: string,
+  recording: ActiveRecording,
+  deps: RecordDeps,
+): Promise<RecordedTrace | null> {
   await clearRearmTimersForRecording(recording, deps);
-  await Promise.all([...recording.rearmCallbacks]);
   try {
     // Disposing content capture may commit a final dirty fill. Flush content
     // first so all resulting RECORD_STEP messages enter actionQueue before it
     // and the observation queues are drained.
     await stopRecordingOnAllAgentTabs(recording, deps);
   } catch {
-    recording.finishing = false;
     return null;
   }
-  if (!(await drainRecordingToStability(recording, deps))) {
-    recording.finishing = false;
+  if (!(await drainRecordingToStability(recording))) {
     return null;
   }
-  cancelPendingSettles(recording.observation);
-  inferMissingPostStates(recording.steps);
-  const cdp = recording.cdp ?? deps.cdp;
-  if (cdp) {
-    await settleUnsettledDrafts(
-      recording.observation,
-      cdp,
-      recording.tabId,
-      deps.tabsApi,
-      recording.steps,
-    );
-  }
+  await recording.observation?.settleTrailing(recording.tabs.currentTabId, recording.steps);
 
   recording.settled = true;
   recordings.delete(sessionId);
@@ -820,6 +824,12 @@ export async function handleRecordStart(
 ): Promise<RecordStartResult | RpcError> {
   const traceVersionOrErr = negotiatedTraceVersion(params);
   if (typeof traceVersionOrErr !== "number") return traceVersionOrErr;
+  if (traceVersionOrErr === 3 && !deps.cdp) {
+    return {
+      code: "protocol_error",
+      message: "trace v3 recording requires an active CDP connection",
+    };
+  }
 
   const ctxOrErr = lookupSession(manager, params, "record_start");
   if (isRpcError(ctxOrErr)) return ctxOrErr;
@@ -849,7 +859,7 @@ export async function handleRecordStart(
   const redactValues = params.redact_values ?? false;
   recordings.set(params.session_id, {
     requestId,
-    tabId: target.tabId,
+    tabs: new RecordingTabCoordinator(target.tabId, navigateUrl),
     agentWindowId: ctx.agentWindowId,
     startUrl: navigateUrl,
     ...(params.purpose ? { purpose: params.purpose } : {}),
@@ -857,30 +867,44 @@ export async function handleRecordStart(
     startedAt: new Date(startedAtMs).toISOString(),
     startedAtMs,
     traceVersion: traceVersionOrErr,
+    supportsTabSwitchSteps: params.supports_tab_switch_steps === true,
     finishPromise,
     resolveFinish,
     rejectFinish,
     settled: false,
-    finishing: false,
-    currentUrl: navigateUrl,
-    pendingNavigation: false,
-    pendingNavigationDeadline: undefined,
-    observation: createObservationState({ maxPageTokens, redactValues }),
+    finishAttempt: null,
+    observation:
+      traceVersionOrErr === 3 && deps.cdp
+        ? new RecordingObservationRuntime({
+            cdp: deps.cdp,
+            tabsApi: deps.tabsApi,
+            maxTokens: maxPageTokens,
+            redactValues,
+          })
+        : null,
     stoppedBy: "user_finish",
-    maxPageTokens: maxPageTokens ?? DEFAULT_MAX_PAGE_TOKENS,
-    redactValues,
-    armedTabIds: new Set(),
-    rearmCallbacks: new Set(),
     navigationCallbacks: new Set(),
     acceptingNavigation: true,
     actionQueue: Promise.resolve(),
-    ...(deps.cdp ? { cdp: deps.cdp } : {}),
+    lastStepSequenceByProducer: new Map(),
+  });
+  deps.frameCoordinator?.begin(requestId, startedAtMs, target.tabId, async (scope) => {
+    const recording = recordings.get(params.session_id);
+    if (!recording || recording.requestId !== requestId || recording.settled) return;
+    try {
+      await recording.observation?.refreshDocument(scope.tabId, scope.producerId);
+    } catch {
+      // The normal action-time capture remains available if a child Document
+      // is replaced while its readiness refresh is in flight.
+    }
   });
   // Observe navigations for the whole recording lifetime; attach before
   // optional navigate so the destination load can rearm capture.
   ensureBrowserObservationListeners(deps);
 
   const abortPending = async (notifyContent: boolean) => {
+    recordings.get(params.session_id)?.observation?.cancel();
+    deps.frameCoordinator?.cancel(requestId);
     recordings.delete(params.session_id);
     releaseBrowserObservationListenersIfIdle();
     if (notifyContent) {
@@ -962,7 +986,7 @@ export async function handleRecordStart(
     return cancelledError();
   }
   active.startUrl = startUrl;
-  active.currentUrl = startUrl;
+  active.tabs.navigation(target.tabId, startUrl).currentUrl = startUrl;
 
   if (isContentScriptRestrictedUrl(startUrl)) {
     await abortPending(false);
@@ -970,7 +994,7 @@ export async function handleRecordStart(
       code: "invalid_params",
       message: params.url
         ? `cannot record on restricted URL (${startUrl}); use an http(s) page`
-        : `cannot record on restricted URL (${startUrl}); default start page must be injectable http(s)`,
+        : `cannot record on restricted URL (${startUrl}); default start page https://example.com/ did not load — pass --url with a page you can open`,
     };
   }
 
@@ -997,8 +1021,11 @@ export async function handleRecordStart(
   const startMsg: RecordStartMessage = { type: RECORD_START, requestId, startedAtMs };
 
   try {
+    if (deps.frameCoordinator) {
+      const frameStarted = await deps.frameCoordinator.armTab(requestId, target.tabId);
+      if (!frameStarted) throw new Error("top recording document did not start");
+    }
     await sendRecordStartWithAck(target.tabId, startMsg, deps.sendToTab);
-    active.armedTabIds.add(target.tabId);
   } catch {
     await abortPending(true);
     return {
@@ -1008,17 +1035,11 @@ export async function handleRecordStart(
     };
   }
 
-  if (deps.cdp) {
+  if (active.observation) {
     const activeRecording = recordings.get(params.session_id);
     if (activeRecording) {
       try {
-        await captureAndRegisterObservation(
-          activeRecording.observation,
-          deps.cdp,
-          target.tabId,
-          deps.tabsApi,
-          startUrl,
-        );
+        await activeRecording.observation?.captureInitial(target.tabId);
       } catch {
         // Proceed without initial observation; steps may be dropped by reducer.
       }
@@ -1123,6 +1144,8 @@ export function clearRecordingForSession(sessionId: string): void {
   }
   void clearRearmTimersForRecording(recording, getDefaultDeps());
   if (!recording.settled) {
+    getDefaultDeps().frameCoordinator?.cancel(recording.requestId);
+    recording.observation?.cancel();
     recording.settled = true;
     recording.rejectFinish(new Error("recording cleared"));
   }

apps/extension/src/tools/session.ts

@@ -76,6 +76,7 @@ export interface SessionStopResult {
 }
 
 export interface SessionStopDeps {
+  signal?: AbortSignal;
   cdp?: {
     detachSession(sessionId: string): Promise<void>;
   };
@@ -177,6 +178,9 @@ export async function handleSessionStop(
       message: `session ${params.session_id} unknown`,
     };
   }
+  if (deps.signal?.aborted) {
+    return { code: "cancelled", message: "session_stop aborted before teardown" };
+  }
 
   // Step 1: auto-return borrowed tabs. Iterate over a snapshot of the
   // ids so deletions during iteration do not break the Map iterator.
@@ -187,6 +191,7 @@ export async function handleSessionStop(
     try {
       const tabManagement = {
         ...(deps.tabManagement ?? {}),
+        signal: deps.signal,
         isAgentWindowId:
           deps.tabManagement?.isAgentWindowId ??
           ((windowId: number) => manager.findByWindowId(windowId) !== null),
@@ -214,6 +219,17 @@ export async function handleSessionStop(
     }
   }
 
+  if (deps.signal?.aborted) {
+    const nonCancelFailures = returnFailures?.filter((failure) => failure.code !== "cancelled");
+    if (nonCancelFailures && nonCancelFailures.length > 0) {
+      return {
+        ...(returnedTabIds.length > 0 ? { returned_tab_ids: returnedTabIds } : {}),
+        return_failures: nonCancelFailures,
+      };
+    }
+    return { code: "cancelled", message: "session_stop aborted during tab return" };
+  }
+
   const result: SessionStopResult = {};
   if (returnedTabIds.length > 0) result.returned_tab_ids = returnedTabIds;
   if (returnFailures && returnFailures.length > 0) {
@@ -233,6 +249,10 @@ export async function handleSessionStop(
   // Step 3: detach CDP sessions this session opened (no-op if none).
   await deps.cdp?.detachSession(params.session_id);
 
+  if (deps.signal?.aborted) {
+    return { code: "cancelled", message: "session_stop aborted before window close" };
+  }
+
   // Step 4: close the Agent Window and drop the context.
   await manager.stop(params.session_id);
 

apps/extension/src/tools/shared.ts

@@ -6,7 +6,8 @@
 // exactly the same sandbox + visibility rules as the M6 observation
 // handlers (review parity).
 
-import type { DialogCursor } from "@/browser-driver/chromium-cdp";
+import type { CdpDebuggee, DialogCursor } from "@/browser-driver/chromium-cdp";
+import type { CdpFrameGraph, CdpTarget } from "@/browser-driver/frame-graph";
 import type { SessionContext, SessionManager } from "@/session-manager/manager";
 import { normaliseRef } from "@/session-manager/ref-store";
 import type { ConsoleResult, JavaScriptDialogInfo, RpcError } from "@/transport/types";
@@ -51,9 +52,11 @@ export type { DialogCursor };
  */
 export interface CdpRunner {
   send<T = unknown>(tabId: number, method: string, params?: object): Promise<T>;
+  sendToTarget?<T = unknown>(target: CdpTarget, method: string, params?: object): Promise<T>;
+  getFrameGraph?(tabId: number): Promise<CdpFrameGraph>;
   ensureAttachedToUrl?(tabId: number, expectedUrl: string | undefined): Promise<void>;
   trackSessionTab?(sessionId: string, tabId: number): void;
-  onEvent?(handler: (source: chrome.debugger.Debuggee, method: string, params: unknown) => void): {
+  onEvent?(handler: (source: CdpDebuggee, method: string, params: unknown) => void): {
     dispose(): void;
   };
   dialogCursor?(tabId: number): DialogCursor;
@@ -68,6 +71,25 @@ export interface CdpRunner {
   ): ConsoleResult;
 }
 
+export function sendToCdpTarget<T = unknown>(
+  cdp: CdpRunner,
+  target: CdpTarget,
+  method: string,
+  params?: object,
+): Promise<T> {
+  if (target.sessionId && cdp.sendToTarget) {
+    return cdp.sendToTarget<T>(target, method, params);
+  }
+  return cdp.send<T>(target.tabId, method, params);
+}
+
+export function cdpRunnerForTarget(cdp: CdpRunner, target: CdpTarget): CdpRunner {
+  return {
+    send: <T = unknown>(_tabId: number, method: string, params?: object) =>
+      sendToCdpTarget<T>(cdp, target, method, params),
+  };
+}
+
 export interface BufferedReadBounds {
   since: number | undefined;
   limit: number;

apps/extension/src/tools/snapshot-ref.ts

@@ -1,15 +1,26 @@
 // Snapshot ref resolution — normalise `@eN` / `eN`, look up
-// `backendNodeId` via the session RefStore, and build stable
+// compound CDP node identity via the session RefStore, and build stable
 // `ref_not_found` errors for hard-failure tool paths.
 
+import type { Rect } from "@browser-skill/vom";
 import type { SessionContext } from "@/session-manager/manager";
-import { normaliseRef } from "@/session-manager/ref-store";
+import { normaliseRef, type RefCapability, type RefTargetKind } from "@/session-manager/ref-store";
 import type { RpcError } from "@/transport/types";
 import { rpcError } from "./errors";
 
 export interface SnapshotRefLookup {
   backendNodeId: number;
   refKey: string;
+  frameId?: string;
+  cdpSessionId?: string;
+  visibleRect?: Rect;
+  kind: RefTargetKind;
+  capabilities: RefCapability[];
+}
+
+function refEntryForTab(ctx: SessionContext, refKey: string, tabId: number) {
+  const entry = ctx.refStore.resolveEntry(refKey);
+  return entry?.tabId === tabId ? entry : null;
 }
 
 /**
@@ -23,9 +34,17 @@ export function lookupSnapshotRef(
   tabId: number,
 ): SnapshotRefLookup | null {
   const refKey = normaliseRef(ref);
-  const backendNodeId = ctx.refStore.resolve(refKey, { tabId });
-  if (backendNodeId === null) return null;
-  return { backendNodeId, refKey };
+  const entry = refEntryForTab(ctx, refKey, tabId);
+  if (!entry) return null;
+  return {
+    backendNodeId: entry.backendNodeId,
+    refKey,
+    ...(entry.frameId ? { frameId: entry.frameId } : {}),
+    ...(entry.cdpSessionId ? { cdpSessionId: entry.cdpSessionId } : {}),
+    ...(entry.visibleRect ? { visibleRect: entry.visibleRect } : {}),
+    kind: entry.kind,
+    capabilities: entry.capabilities,
+  };
 }
 
 /**
@@ -37,7 +56,8 @@ export function resolveSnapshotRef(
   ctx: SessionContext,
   ref: string,
   tabId: number,
-): { backendNodeId: number; refKey: string } | RpcError {
+  requiredCapability?: RefCapability,
+): SnapshotRefLookup | RpcError {
   const looked = lookupSnapshotRef(ctx, ref, tabId);
   if (looked === null) {
     return rpcError(
@@ -46,5 +66,13 @@ export function resolveSnapshotRef(
       `ref ${ref} unknown for tab ${tabId} in session ${ctx.sessionId}`,
     );
   }
+  if (requiredCapability && !looked.capabilities.includes(requiredCapability)) {
+    return rpcError(
+      "permission_denied",
+      "ref_capability_denied",
+      `ref ${ref} does not support ${requiredCapability}`,
+      { ref: looked.refKey, required_capability: requiredCapability, kind: looked.kind },
+    );
+  }
   return looked;
 }

apps/extension/src/tools/vom/__tests__/capture.test.ts

@@ -18,6 +18,8 @@ function fakeSnapshotReply() {
     "auto",
     "type",
     "password",
+    "value",
+    "hunter2",
   ];
   const i = (s: string) => S.indexOf(s);
   return {
@@ -29,11 +31,12 @@ function fakeSnapshotReply() {
           nodeType: [1, 1, 1, 1],
           nodeName: [i("html"), i("body"), i("div"), i("input")],
           backendNodeId: [10, 11, 12, 13],
-          attributes: [[], [], [], [i("type"), i("password")]],
+          attributes: [[], [], [], [i("type"), i("password"), i("value"), i("hunter2")]],
+          inputValue: { index: [3], value: [i("hunter2")] },
         },
         layout: {
           nodeIndex: [1, 2, 3],
-          // styles columns follow requested computedStyles order: [position, pointer-events]
+          // Legacy fixture: omitted style columns fall back to visible defaults.
           styles: [
             [i("static"), i("auto")],
             [i("fixed"), i("auto")],
@@ -217,7 +220,8 @@ function makeCdp(snapshot: unknown) {
                 {
                   state: "filled",
                   sensitive: true,
-                  defaultValue: "",
+                  defaultValue: "hunter2",
+                  value: "hunter2",
                   placeholder: "secret",
                 },
               ],
@@ -262,10 +266,11 @@ describe("captureViewModel", () => {
 
     expect(input).toMatchObject({
       formState: "filled",
-      formDefaultValue: "",
       formPlaceholder: "secret",
+      attrs: { type: "password" },
     });
     expect(input?.formValue).toBeUndefined();
+    expect(input?.formDefaultValue).toBeUndefined();
     expect(cdp.send).toHaveBeenCalledWith(
       4,
       "Runtime.evaluate",
@@ -677,7 +682,85 @@ describe("captureViewModel", () => {
     expect(nodes.find((n) => n.backendNodeId === 11)?.cursor).toBe("auto");
   });
 
-  it("preserves document rects while deriving viewport rects after scroll", async () => {
+  it("captures painted visibility as DOM fallback evidence", async () => {
+    const S = ["html", "body", "button", "static", "auto", "visible", "hidden", "1", "0"];
+    const i = (value: string) => S.indexOf(value);
+    const snapshot = {
+      strings: S,
+      documents: [
+        {
+          nodes: {
+            parentIndex: [-1, 0, 1, 1, 1],
+            nodeName: [i("html"), i("body"), i("button"), i("button"), i("button")],
+            backendNodeId: [10, 11, 12, 13, 14],
+            attributes: [[], [], [], [], []],
+          },
+          layout: {
+            nodeIndex: [1, 2, 3, 4],
+            styles: [
+              [i("static"), i("auto"), i("auto"), i("visible"), i("1")],
+              [i("static"), i("auto"), i("auto"), i("visible"), i("1")],
+              [i("static"), i("auto"), i("auto"), i("hidden"), i("1")],
+              [i("static"), i("auto"), i("auto"), i("visible"), i("0")],
+            ],
+            bounds: [
+              [0, 0, 1000, 800],
+              [10, 10, 100, 30],
+              [10, 50, 100, 30],
+              [10, 90, 100, 30],
+            ],
+            paintOrders: [0, 1, 2, 3],
+          },
+        },
+      ],
+    };
+
+    const { nodes } = await captureViewModel(makeCdp(snapshot), 4);
+
+    expect(nodes.find((node) => node.backendNodeId === 12)?.rendered).toBe(true);
+    expect(nodes.find((node) => node.backendNodeId === 13)?.rendered).toBe(false);
+    expect(nodes.find((node) => node.backendNodeId === 14)?.rendered).toBe(false);
+  });
+
+  it("propagates fully transparent ancestor suppression to painted descendants", async () => {
+    const S = ["html", "body", "div", "canvas", "static", "auto", "visible", "1", "0"];
+    const i = (value: string) => S.indexOf(value);
+    const snapshot = {
+      strings: S,
+      documents: [
+        {
+          nodes: {
+            parentIndex: [-1, 0, 1, 2],
+            nodeName: [i("html"), i("body"), i("div"), i("canvas")],
+            backendNodeId: [10, 11, 12, 13],
+            attributes: [[], [], [], []],
+          },
+          layout: {
+            nodeIndex: [1, 2, 3],
+            styles: [
+              [i("static"), i("auto"), i("auto"), i("visible"), i("1")],
+              [i("static"), i("auto"), i("auto"), i("visible"), i("0")],
+              [i("static"), i("auto"), i("auto"), i("visible"), i("1")],
+            ],
+            bounds: [
+              [0, 0, 1000, 800],
+              [10, 10, 200, 100],
+              [10, 10, 200, 100],
+            ],
+            paintOrders: [0, 1, 2],
+          },
+        },
+      ],
+    };
+
+    const { nodes } = await captureViewModel(makeCdp(snapshot), 4);
+    const canvas = nodes.find((node) => node.backendNodeId === 13);
+
+    expect(canvas?.rendered).toBe(true);
+    expect(canvas?.visuallySuppressed).toBe(true);
+  });
+
+  it("subtracts scroll offset to make rects viewport-relative", async () => {
     const cdp = {
       send: vi.fn(async (_t: number, method: string) => {
         if (method === "DOMSnapshot.enable") return {};
@@ -691,10 +774,9 @@ describe("captureViewModel", () => {
       }) as unknown as <T>(tabId: number, method: string, params?: object) => Promise<T>,
     };
     const { nodes } = await captureViewModel(cdp, 4);
-    const input = nodes.find((n) => n.backendNodeId === 13);
-    expect(input?.documentRect).toEqual({ x: 400, y: 300, w: 200, h: 40 });
-    expect(input?.localRect).toEqual({ x: 400, y: 100, w: 200, h: 40 });
-    expect(input?.rect).toEqual({ x: 400, y: 100, w: 200, h: 40 });
+    const div = nodes.find((n) => n.backendNodeId === 12);
+    expect(div?.localRect?.y).toBe(-200);
+    expect(div?.rect).toMatchObject({ y: 0, h: 600 });
   });
 
   it("normalizes device-pixel bounds by devicePixelRatio", async () => {
@@ -838,6 +920,7 @@ describe("captureViewModel", () => {
       strings: S,
       documents: [
         {
+          frameId: "main-frame",
           scrollOffsetX: 0,
           scrollOffsetY: 0,
           nodes: {
@@ -865,6 +948,7 @@ describe("captureViewModel", () => {
         },
         // documents[1]: the iframe's sub-document with a login input
         {
+          frameId: "child-frame",
           scrollOffsetX: 0,
           scrollOffsetY: 0,
           nodes: {
@@ -894,7 +978,14 @@ describe("captureViewModel", () => {
         throw new Error(method);
       }) as unknown as <T>(tabId: number, method: string, params?: object) => Promise<T>,
     };
-    const { nodes, iframeNodes } = await captureViewModel(cdp, 4);
+    const {
+      nodes,
+      iframeNodes,
+      frameNodes,
+      frameOwnerBackendNodeIds,
+      frameParentIds,
+      rootFrameId,
+    } = await captureViewModel(cdp, 4);
     // Main frame has 4 nodes; iframe is included
     expect(nodes.find((n) => n.backendNodeId === 13)?.tag).toBe("iframe");
     // iframeNodes keyed by the <iframe> element's backendNodeId=13
@@ -907,9 +998,13 @@ describe("captureViewModel", () => {
     expect(input?.ownerFrameBackendNodeId).toBe(13);
     expect(input?.localRect).toEqual({ x: 0, y: 0, w: 200, h: 40 });
     expect(input?.rect).toEqual({ x: 100, y: 300, w: 200, h: 40 });
+    expect(rootFrameId).toBe("main-frame");
+    expect(frameNodes?.get("child-frame")).toBe(subNodes);
+    expect(frameOwnerBackendNodeIds?.get("child-frame")).toBe(13);
+    expect(frameParentIds?.get("child-frame")).toBe("main-frame");
   });
 
-  it("clips iframe sub-document rects to the iframe viewport in top coordinates", async () => {
+  it("captures iframe documents without requiring owner-frame geometry", async () => {
     const S = [
       "html",
       "body",
@@ -937,16 +1032,10 @@ describe("captureViewModel", () => {
             contentDocumentIndex: { index: [2], value: [1] },
           },
           layout: {
-            nodeIndex: [1, 2],
-            styles: [
-              [i("static"), i("auto")],
-              [i("static"), i("auto")],
-            ],
-            bounds: [
-              [0, 0, 1000, 800],
-              [950, 760, 200, 100],
-            ],
-            paintOrders: [0, 1],
+            nodeIndex: [1],
+            styles: [[i("static"), i("auto")]],
+            bounds: [[0, 0, 1000, 800]],
+            paintOrders: [0],
           },
         },
         {
@@ -984,7 +1073,7 @@ describe("captureViewModel", () => {
     const input = iframeNodes.get(13)?.find((n) => n.backendNodeId === 101);
 
     expect(input?.localRect).toEqual({ x: 20, y: 20, w: 120, h: 60 });
-    expect(input?.rect).toEqual({ x: 970, y: 780, w: 30, h: 20 });
+    expect(input?.rect).toBeNull();
   });
 
   it("recursively normalizes nested iframe sub-documents", async () => {
@@ -1079,8 +1168,10 @@ describe("captureViewModel", () => {
     const nestedIframe = iframeNodes.get(13)?.find((n) => n.backendNodeId === 23);
     const input = iframeNodes.get(23)?.find((n) => n.backendNodeId === 31);
 
+    expect(nestedIframe?.localRect).toEqual({ x: 5, y: 6, w: 100, h: 80 });
     expect(nestedIframe?.rect).toEqual({ x: 15, y: 26, w: 100, h: 80 });
     expect(input?.ownerFrameBackendNodeId).toBe(23);
+    expect(input?.localRect).toEqual({ x: 1, y: 2, w: 40, h: 20 });
     expect(input?.rect).toEqual({ x: 16, y: 28, w: 40, h: 20 });
   });
 

apps/extension/src/tools/vom/__tests__/frame-capture.test.ts

@@ -0,0 +1,198 @@
+import { describe, expect, it, vi } from "vitest";
+import type { CdpRunner } from "../../shared";
+import type { CapturedNode, CapturedViewModel } from "../capture";
+import { captureFrameData } from "../frame-capture";
+
+function ownerNode(backendNodeId: number, x: number): CapturedNode {
+  return {
+    backendNodeId,
+    parentBackendNodeId: 1,
+    frameId: "main",
+    ownerFrameBackendNodeId: null,
+    tag: "iframe",
+    attrs: {},
+    rect: { x, y: 100, w: 300, h: 200 },
+    localRect: { x, y: 100, w: 300, h: 200 },
+    paintOrder: 1,
+    position: "static",
+    pointerEvents: "auto",
+  };
+}
+
+function childSnapshot(frameId: string, backendNodeId: number) {
+  const strings = [frameId, "body", "button", "static", "auto", "pointer"];
+  return {
+    strings,
+    documents: [
+      {
+        frameId,
+        nodes: {
+          parentIndex: [-1, 0],
+          nodeName: [1, 2],
+          backendNodeId: [backendNodeId - 1, backendNodeId],
+          attributes: [[], []],
+        },
+        layout: {
+          nodeIndex: [0, 1],
+          styles: [
+            [3, 4, 4],
+            [3, 4, 5],
+          ],
+          bounds: [
+            [0, 0, 300, 200],
+            [10, 20, 100, 40],
+          ],
+          paintOrders: [0, 1],
+        },
+      },
+    ],
+  };
+}
+
+describe("captureFrameData", () => {
+  it("captures and positions multiple OOPIF documents missing from the root snapshot", async () => {
+    const leftOwner = ownerNode(10, 50);
+    const rightOwner = ownerNode(20, 500);
+    const captured: CapturedViewModel = {
+      nodes: [leftOwner, rightOwner],
+      viewport: { width: 1000, height: 800 },
+      iframeNodes: new Map(),
+      frameNodes: new Map([["main", [leftOwner, rightOwner]]]),
+      frameOwnerBackendNodeIds: new Map(),
+      rootFrameId: "main",
+      excludedBackendNodeIds: new Set(),
+    };
+    const sendToTarget = vi.fn(async (target, method) => {
+      if (method === "Page.getLayoutMetrics") {
+        return { cssLayoutViewport: { clientWidth: 300, clientHeight: 200, pageX: 0, pageY: 0 } };
+      }
+      if (method === "DOMSnapshot.enable" || method === "Accessibility.enable") return {};
+      if (method === "DOMSnapshot.captureSnapshot") {
+        return target.sessionId === "left-session"
+          ? childSnapshot("left", 101)
+          : childSnapshot("right", 201);
+      }
+      if (method === "Accessibility.getFullAXTree") return { nodes: [] };
+      throw new Error(`unexpected ${method}`);
+    });
+    const cdp: CdpRunner = {
+      send: vi.fn(async (_tabId, method, params) => {
+        if (method === "Accessibility.enable") return {};
+        if (method === "Accessibility.getFullAXTree") return { nodes: [] };
+        if (method === "DOM.getBoxModel") {
+          const backendNodeId = (params as { backendNodeId?: number })?.backendNodeId;
+          const x = backendNodeId === 10 ? 50 : 500;
+          return { model: { content: [x, 100, x + 300, 100, x + 300, 300, x, 300] } };
+        }
+        if (method === "Page.getLayoutMetrics") {
+          return { cssLayoutViewport: { clientWidth: 1000, clientHeight: 800 } };
+        }
+        throw new Error(`unexpected root ${method}`);
+      }) as CdpRunner["send"],
+      sendToTarget: sendToTarget as unknown as NonNullable<CdpRunner["sendToTarget"]>,
+      getFrameGraph: vi.fn(async () => ({
+        rootFrameId: "main",
+        frames: [
+          { frameId: "main", target: { tabId: 4 } },
+          {
+            frameId: "left",
+            parentFrameId: "main",
+            ownerBackendNodeId: 10,
+            target: { tabId: 4, sessionId: "left-session" },
+          },
+          {
+            frameId: "right",
+            parentFrameId: "main",
+            ownerBackendNodeId: 20,
+            target: { tabId: 4, sessionId: "right-session" },
+          },
+        ],
+      })),
+    };
+
+    const trees = await captureFrameData(cdp, 4, captured);
+
+    expect(trees.map((tree) => tree.frameId)).toEqual(["main", "left", "right"]);
+    expect(
+      captured.frameNodes?.get("left")?.find((node) => node.backendNodeId === 101)?.rect,
+    ).toEqual({ x: 60, y: 120, w: 100, h: 40 });
+    expect(
+      captured.frameNodes?.get("right")?.find((node) => node.backendNodeId === 201)?.rect,
+    ).toEqual({ x: 510, y: 120, w: 100, h: 40 });
+    expect(captured.frameOwnerBackendNodeIds).toEqual(
+      new Map([
+        ["left", 10],
+        ["right", 20],
+      ]),
+    );
+    expect(captured.frameParentIds).toEqual(
+      new Map([
+        ["left", "main"],
+        ["right", "main"],
+      ]),
+    );
+  });
+
+  it("keeps OOPIF semantics when viewport projection is unavailable", async () => {
+    const owner = { ...ownerNode(10, 50), rect: null, localRect: null };
+    const captured: CapturedViewModel = {
+      nodes: [owner],
+      viewport: { width: 1000, height: 800 },
+      iframeNodes: new Map(),
+      frameNodes: new Map([["main", [owner]]]),
+      rootFrameId: "main",
+      excludedBackendNodeIds: new Set(),
+    };
+    const cdp: CdpRunner = {
+      send: vi.fn(async (_tabId, method) => {
+        if (method === "Accessibility.enable") return {};
+        if (method === "Accessibility.getFullAXTree") return { nodes: [] };
+        if (method === "DOM.getBoxModel") throw new Error("owner geometry unavailable");
+        if (method === "Page.getLayoutMetrics") {
+          return { cssLayoutViewport: { clientWidth: 1000, clientHeight: 800 } };
+        }
+        throw new Error(`unexpected root ${method}`);
+      }) as CdpRunner["send"],
+      sendToTarget: vi.fn(async (_target, method) => {
+        if (method === "Page.getLayoutMetrics") throw new Error("viewport unavailable");
+        if (method === "DOMSnapshot.enable" || method === "Accessibility.enable") return {};
+        if (method === "DOMSnapshot.captureSnapshot") return childSnapshot("child", 101);
+        if (method === "Accessibility.getFullAXTree") {
+          return {
+            nodes: [
+              {
+                nodeId: "button",
+                backendDOMNodeId: 101,
+                role: { type: "role", value: "button" },
+                name: { type: "computedString", value: "Continue" },
+              },
+            ],
+          };
+        }
+        throw new Error(`unexpected ${method}`);
+      }) as unknown as NonNullable<CdpRunner["sendToTarget"]>,
+      getFrameGraph: vi.fn(async () => ({
+        rootFrameId: "main",
+        frames: [
+          { frameId: "main", target: { tabId: 4 } },
+          {
+            frameId: "child",
+            parentFrameId: "main",
+            ownerBackendNodeId: 10,
+            target: { tabId: 4, sessionId: "child-session" },
+          },
+        ],
+      })),
+    };
+
+    const documents = await captureFrameData(cdp, 4, captured);
+    const childDocument = documents.find((document) => document.frameId === "child");
+
+    expect(childDocument?.axNodes).toEqual([
+      expect.objectContaining({ backendDOMNodeId: 101, frameId: "child" }),
+    ]);
+    expect(childDocument?.domNodes.find((node) => node.backendNodeId === 101)).toEqual(
+      expect.objectContaining({ rect: null, localRect: { x: 10, y: 20, w: 100, h: 40 } }),
+    );
+  });
+});