feat: add browser_read_page reader-mode tool

URL (or owned targetId) → clean main-article text with nav/ads/footer
boilerplate stripped, via a dependency-free readability heuristic. DOM walk
lives in an in-page capture expression; scoring/selection is a pure, fixture-
tested function (link-density + boilerplate-ancestor filtering, body-text
fallback for structure-less pages). Reuses the isolated-attach lifecycle;
registered unserialized (safe concurrent read). 10 extractor unit scenarios.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

4 files changed, 247 insertions, 0 deletions

diff --git a/src/domains/readpage/capture.ts b/src/domains/readpage/capture.ts
new file mode 100644
index 0000000..1ee4500
--- /dev/null
+++ b/src/domains/readpage/capture.ts
@@ -0,0 +1,46 @@
+/**
+ * The JavaScript source evaluated in a loaded page to capture candidate content
+ * blocks + metadata for the pure readability extractor (readability.ts). Crosses
+ * the CDP boundary via Runtime.evaluate (returnByValue) and returns a JSON
+ * string shaped like PageCapture.
+ *
+ * The DOM walk lives here (unavoidable); all scoring/selection is pure and lives
+ * in readability.ts, so the decision logic stays unit-testable.
+ */
+
+/** Max bytes of body-text fallback captured for thin/structure-less pages. */
+const BODY_TEXT_LIMIT = 20_000;
+
+/** Build the capture expression. No interpolated values ⇒ no injection surface. */
+export const buildPageCaptureExpr = (): string => `
+  (() => {
+    const BOILERPLATE = new Set(['NAV', 'HEADER', 'FOOTER', 'ASIDE']);
+    const inBoilerplate = (el) => {
+      for (let n = el; n; n = n.parentElement) {
+        if (BOILERPLATE.has(n.tagName)) return true;
+        const role = n.getAttribute && n.getAttribute('role');
+        if (role === 'navigation' || role === 'banner' || role === 'contentinfo') return true;
+      }
+      return false;
+    };
+    const kindOf = (tag) => {
+      if (/^H[1-6]$/.test(tag)) return 'heading';
+      if (tag === 'LI') return 'listitem';
+      if (tag === 'BLOCKQUOTE') return 'blockquote';
+      if (tag === 'P') return 'paragraph';
+      return 'other';
+    };
+    const selector = 'p, h1, h2, h3, h4, h5, h6, li, blockquote';
+    const root = document.querySelector('article, main, [role=main]') || document.body;
+    const blocks = [];
+    for (const el of root.querySelectorAll(selector)) {
+      const text = (el.innerText || '').replace(/\\s+/g, ' ').trim();
+      if (!text) continue;
+      let linkTextLength = 0;
+      for (const a of el.querySelectorAll('a')) linkTextLength += (a.innerText || '').length;
+      blocks.push({ kind: kindOf(el.tagName), text, linkTextLength, inBoilerplate: inBoilerplate(el) });
+    }
+    const bodyText = (document.body.innerText || '').replace(/\\n{3,}/g, '\\n\\n').slice(0, ${BODY_TEXT_LIMIT});
+    return JSON.stringify({ url: location.href, title: document.title, blocks, bodyText });
+  })()
+`;
diff --git a/src/domains/readpage/read-page.ts b/src/domains/readpage/read-page.ts
new file mode 100644
index 0000000..7bf2fdb
--- /dev/null
+++ b/src/domains/readpage/read-page.ts
@@ -0,0 +1,109 @@
+import { Type } from "typebox";
+import { type Result, err, ok } from "../../util/result";
+import { applyTruncation } from "../../util/truncate";
+import { defineBrowserTool, type ToolErr, type ToolOk } from "../../util/tool";
+import {
+  closeIsolatedTab,
+  evalInIsolatedTab,
+  type IsolatedTab,
+  navigateIsolatedTab,
+  openIsolatedTab,
+  waitForIsolatedLoad,
+} from "../isolated-tab";
+import { buildPageCaptureExpr } from "./capture";
+import { extractReadable, type PageCapture, type ReadablePage } from "./readability";
+
+const LOAD_TIMEOUT_MS = 15_000;
+
+const ReadPageArgs = Type.Object({
+  url: Type.Optional(Type.String({ description: "URL to open in an isolated tab, read, then close." })),
+  targetId: Type.Optional(
+    Type.String({ description: "Read an already-open tab this session owns instead of opening a URL." }),
+  ),
+});
+
+const isPageCapture = (value: unknown): value is PageCapture => {
+  if (typeof value !== "object" || value === null) return false;
+  const v = value as Record<string, unknown>;
+  return typeof v["url"] === "string" && typeof v["title"] === "string" && Array.isArray(v["blocks"]) && typeof v["bodyText"] === "string";
+};
+
+const safeParse = (text: string): unknown => {
+  try {
+    return JSON.parse(text) as unknown;
+  } catch {
+    return undefined;
+  }
+};
+
+const captureToResult = async (raw: unknown): Promise<Result<ToolOk, ToolErr>> => {
+  const parsed = typeof raw === "string" ? safeParse(raw) : raw;
+  if (!isPageCapture(parsed)) return err({ kind: "internal", message: "page capture returned an unexpected shape" });
+  const page: ReadablePage = extractReadable(parsed);
+  const header = `# ${page.title}\n${page.url}\n(${page.wordCount} words)\n\n`;
+  const truncated = await applyTruncation(header + page.text, "readpage");
+  return ok({
+    text: truncated.text,
+    details: {
+      title: page.title,
+      url: page.url,
+      wordCount: page.wordCount,
+      ...(truncated.fullOutputPath !== undefined ? { fullOutputPath: truncated.fullOutputPath } : {}),
+    },
+  });
+};
+
+const readOpenedUrl = async (
+  client: Parameters<typeof openIsolatedTab>[0],
+  url: string,
+  signal: AbortSignal | undefined,
+): Promise<Result<ToolOk, ToolErr>> => {
+  const opened = await openIsolatedTab(client);
+  if (!opened.success) return err(opened.error);
+  const tab: IsolatedTab = opened.data;
+  try {
+    const navigated = await navigateIsolatedTab(client, tab, url);
+    if (!navigated.success) return err(navigated.error);
+    const loaded = await waitForIsolatedLoad(client, tab, LOAD_TIMEOUT_MS, signal);
+    if (!loaded.success) return err(loaded.error);
+    const captured = await evalInIsolatedTab(client, tab, buildPageCaptureExpr());
+    if (!captured.success) return err(captured.error);
+    return captureToResult(captured.data);
+  } finally {
+    await closeIsolatedTab(client, tab);
+  }
+};
+
+const readOwnedTab = async (
+  client: Parameters<typeof openIsolatedTab>[0],
+  targetId: string,
+): Promise<Result<ToolOk, ToolErr>> => {
+  if (!client.owns(targetId)) {
+    return err({ kind: "invalid_state", message: `Tab ${targetId} is not owned by this session.` });
+  }
+  const attached = await client.session().callBrowser("Target.attachToTarget", { targetId, flatten: true });
+  if (!attached.success) return err({ kind: "cdp_error", message: attached.error.message });
+  const sessionId = (attached.data as { sessionId: string }).sessionId;
+  const captured = await client.evaluateJs(buildPageCaptureExpr(), sessionId);
+  if (!captured.success) return err({ kind: "cdp_error", message: captured.error.message });
+  return captureToResult(captured.data);
+};
+
+export const readPageTool = defineBrowserTool({
+  name: "browser_read_page",
+  label: "Browser Read Page",
+  description:
+    "Read a page as clean article text — main content with nav/ads/boilerplate stripped. Pass a url (opened in an isolated tab, read, then closed) or a targetId of an owned tab. Reader-mode counterpart to browser_web_search.",
+  promptSnippet: "Read a page's main content as clean text",
+  promptGuidelines: [
+    "Pass url to read an arbitrary page (opened + closed in its own tab, never disturbing your current tab), or targetId to read an already-open owned tab.",
+    "Returns readable main-article text — use this over browser_snapshot/browser_execute_js when you want an article's content for reading or research.",
+    "Boilerplate-heavy or structure-less pages fall back to bounded body text rather than erroring.",
+  ],
+  parameters: ReadPageArgs,
+  async handler(args, { client, signal }): Promise<Result<ToolOk, ToolErr>> {
+    if (args.url !== undefined && args.url.length > 0) return readOpenedUrl(client, args.url, signal);
+    if (args.targetId !== undefined && args.targetId.length > 0) return readOwnedTab(client, args.targetId);
+    return err({ kind: "invalid_state", message: "provide either url or targetId" });
+  },
+});
diff --git a/src/domains/readpage/readability.ts b/src/domains/readpage/readability.ts
new file mode 100644
index 0000000..37f2a38
--- /dev/null
+++ b/src/domains/readpage/readability.ts
@@ -0,0 +1,89 @@
+/**
+ * Pure reader-mode extraction — no CDP, no DOM. The in-page script
+ * (see capture.ts) serializes candidate content blocks; these pure functions
+ * score them and assemble clean article text, so the selection logic is
+ * unit-testable against saved fixtures (test/deep-research/readability-test.ts).
+ *
+ * Dependency-free by design (no jsdom / Readability): the harness ships no new
+ * runtime deps. The heuristic favors blocks with high text density and low link
+ * density, which is what separates article body from nav/ads/boilerplate.
+ */
+
+/** One candidate content block captured in-page. */
+export type ContentBlock = {
+  /** Semantic tag: paragraph | heading | listitem | blockquote | other. */
+  readonly kind: "paragraph" | "heading" | "listitem" | "blockquote" | "other";
+  readonly text: string;
+  /** Chars of text that sit inside <a> tags — high ratio ⇒ nav/menu, not prose. */
+  readonly linkTextLength: number;
+  /** True when an ancestor is nav/header/footer/aside — structural boilerplate. */
+  readonly inBoilerplate: boolean;
+};
+
+/** The JSON payload capture.ts returns from the page. */
+export type PageCapture = {
+  readonly url: string;
+  readonly title: string;
+  readonly blocks: ReadonlyArray<ContentBlock>;
+  /** Bounded innerText fallback when block extraction finds nothing usable. */
+  readonly bodyText: string;
+};
+
+/** The tool's public output shape. */
+export type ReadablePage = {
+  readonly title: string;
+  readonly url: string;
+  readonly text: string;
+  readonly wordCount: number;
+};
+
+const MIN_PARAGRAPH_CHARS = 25;
+const MAX_LINK_DENSITY = 0.5;
+const MIN_ARTICLE_WORDS = 40;
+
+const linkDensity = (block: ContentBlock): number => {
+  if (block.text.length === 0) return 1;
+  return block.linkTextLength / block.text.length;
+};
+
+/** A block is article prose if it is substantive, text-dense, and not chrome. */
+const isArticleBlock = (block: ContentBlock): boolean => {
+  if (block.inBoilerplate) return false;
+  if (linkDensity(block) > MAX_LINK_DENSITY) return false;
+  if (block.kind === "heading" || block.kind === "blockquote" || block.kind === "listitem") {
+    return block.text.trim().length > 0;
+  }
+  return block.text.trim().length >= MIN_PARAGRAPH_CHARS;
+};
+
+const render = (block: ContentBlock): string => {
+  const text = block.text.trim();
+  if (block.kind === "heading") return `\n## ${text}\n`;
+  if (block.kind === "listitem") return `- ${text}`;
+  if (block.kind === "blockquote") return `> ${text}`;
+  return text;
+};
+
+const wordCountOf = (text: string): number => {
+  const trimmed = text.trim();
+  return trimmed.length === 0 ? 0 : trimmed.split(/\s+/).length;
+};
+
+/**
+ * Assemble readable article text from a page capture. Keeps text-dense,
+ * non-boilerplate blocks; falls back to the bounded body text when the article
+ * is too thin (e.g. a page with no clear content structure).
+ */
+export const extractReadable = (capture: PageCapture): ReadablePage => {
+  const articleText = capture.blocks.filter(isArticleBlock).map(render).join("\n\n").replace(/\n{3,}/g, "\n\n").trim();
+
+  const useFallback = wordCountOf(articleText) < MIN_ARTICLE_WORDS;
+  const text = useFallback ? capture.bodyText.trim() : articleText;
+
+  return {
+    title: capture.title.trim(),
+    url: capture.url,
+    text,
+    wordCount: wordCountOf(text),
+  };
+};
diff --git a/src/registry.ts b/src/registry.ts
index c7ff5c8..a063c9f 100644
--- a/src/registry.ts
+++ b/src/registry.ts
@@ -20,6 +20,7 @@ import { snapshotTool } from "./domains/snapshot";
 import { executeJsTool, runScriptTool } from "./domains/js";
 import { setupTool } from "./domains/setup";
 import { webSearchTool } from "./domains/search/web-search";
+import { readPageTool } from "./domains/readpage/read-page";
 
 // Mutation tools that modify page / browser state are marked `serialized: true`
 // so they run through a shared async mutex.  Observation / read-only tools are
@@ -64,6 +65,8 @@ const TOOLS: ReadonlyArray<AnyBrowserToolDefinition> = [
   runScriptTool,
   // Opens and owns an isolated tab to scrape a SERP — same class as openUrlsTool.
   { ...webSearchTool, serialized: true },
+  // Pure isolated read (own sessionId, own tab) — safe concurrent, like executeJsTool.
+  readPageTool,
 ];
 
 export const registerAllTools = (pi: ExtensionAPI, client: BrowserClient): void => {