feat: add forms domain and durable window binding (v0.9.0) (#14)

- Add forms domain with a universal field setter that auto-detects element
  type and fires framework-compatible input events (React native-setter trick)
- Add window-binding helper to capture and bind the real Chrome windowId as the
  session's durable window identity on every new-window code path
- Wire supporting changes across CDP session/ownership, client, navigate,
  isolated-tab, prompt, and state
- Add manual tests for forms and window binding
- Bump version 0.8.3 -> 0.9.0

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

10 files changed, 633 insertions, 140 deletions

diff --git a/src/cdp/ownership.ts b/src/cdp/ownership.ts
index f221d6a..727b3d8 100644
--- a/src/cdp/ownership.ts
+++ b/src/cdp/ownership.ts
@@ -20,14 +20,24 @@ export type OwnershipRegistry = {
   replaceAll(ids: ReadonlyArray<string>): void;
   setHarnessWindow(targetId: string | undefined): void;
   harnessWindow(): string | undefined;
+  /** The real Chrome windowId of the dedicated harness window — the durable
+   *  identity of "the window this session initialized", surviving seed-tab
+   *  closure. Undefined until the window is created (or after teardown). */
+  setHarnessWindowId(windowId: number | undefined): void;
+  harnessWindowId(): number | undefined;
   onChange(cb: () => void): void;
 };
 
 export const createOwnershipRegistry = (
-  initial?: { readonly ownedTargetIds?: ReadonlyArray<string>; readonly harnessWindowTargetId?: string },
+  initial?: {
+    readonly ownedTargetIds?: ReadonlyArray<string>;
+    readonly harnessWindowTargetId?: string;
+    readonly harnessWindowId?: number;
+  },
 ): OwnershipRegistry => {
   const owned = new Set<string>(initial?.ownedTargetIds ?? []);
   let harnessWindow: string | undefined = initial?.harnessWindowTargetId;
+  let harnessWindowId: number | undefined = initial?.harnessWindowId;
   let listener: (() => void) | null = null;
 
   const notify = (): void => {
@@ -65,6 +75,14 @@ export const createOwnershipRegistry = (
     harnessWindow() {
       return harnessWindow;
     },
+    setHarnessWindowId(windowId) {
+      if (harnessWindowId === windowId) return;
+      harnessWindowId = windowId;
+      notify();
+    },
+    harnessWindowId() {
+      return harnessWindowId;
+    },
     onChange(cb) {
       listener = cb;
     },
diff --git a/src/cdp/session.ts b/src/cdp/session.ts
index e090858..03410ba 100644
--- a/src/cdp/session.ts
+++ b/src/cdp/session.ts
@@ -92,6 +92,17 @@ export const createCdpSession = (
         const tab = targetId ? tabs.get(targetId) : undefined;
         if (tab) tab.pageInfoDirty = true;
       }
+      if (ev.method === "Target.targetCreated" && ownership) {
+        // Adopt tabs opened BY a tab we already own (window.open, target=_blank,
+        // popups). Chrome sets openerId to the opener target; if that opener is
+        // owned, the child belongs to this session's window and must be
+        // controllable + cleaned up. Targets the user opens themselves have no
+        // owned opener, so they are never adopted.
+        const info = (ev.params as { targetInfo?: { targetId?: string; type?: string; openerId?: string } } | undefined)?.targetInfo;
+        if (info?.type === "page" && info.targetId && info.openerId && ownership.has(info.openerId)) {
+          ownership.add(info.targetId);
+        }
+      }
       if (ev.method === "Target.targetDestroyed" && ownership) {
         const params = ev.params as { targetId?: string } | undefined;
         if (params?.targetId) {
@@ -179,6 +190,43 @@ export const createCdpSession = (
         if (survivors.length !== ownership.list().length) ownership.replaceAll(survivors);
         const hw = ownership.harnessWindow();
         if (hw && !live.has(hw)) ownership.setHarnessWindow(undefined);
+
+        if (survivors.length === 0) {
+          // No owned tab survived — the harness window is gone (user closed it,
+          // or Chrome restarted and reassigned all ids). The persisted windowId
+          // is now meaningless and could even collide with one of the user's
+          // windows, so drop it. attachFirstPage will mint a fresh window below.
+          ownership.setHarnessWindowId(undefined);
+        } else {
+          // Self-heal ownership from the window itself. We anchor on a tab we
+          // KNOW we own (a survivor) and re-adopt only its live window-mates —
+          // popups/children from last session, tabs reopened in our window.
+          // Ownership is thus DERIVED from the real window, never from the bare
+          // persisted integer (which Chrome may reassign to a user window on
+          // restart — anchoring on a survivor makes that misfire impossible).
+          const anchorId = survivors[0]!;
+          const anchorWin = await transport.request("Browser.getWindowForTarget", { targetId: anchorId }, { sessionId: null });
+          if (anchorWin.success) {
+            const anchorWindowId = (anchorWin.data as { windowId?: number }).windowId;
+            if (typeof anchorWindowId === "number") {
+              ownership.setHarnessWindowId(anchorWindowId); // refresh to the live id
+              const windows = await Promise.all(
+                allPages.map((p) =>
+                  p.targetId === anchorId
+                    ? Promise.resolve(anchorWin)
+                    : transport.request("Browser.getWindowForTarget", { targetId: p.targetId }, { sessionId: null }),
+                ),
+              );
+              allPages.forEach((p, i) => {
+                const r = windows[i];
+                if (r?.success && (r.data as { windowId?: number }).windowId === anchorWindowId) {
+                  ownership.add(p.targetId);
+                }
+              });
+            }
+          }
+          if (!ownership.harnessWindow()) ownership.setHarnessWindow(anchorId);
+        }
       }
       // Prune per-tab state for pages that no longer exist
       const liveTargetIds = new Set(allPages.map((p) => p.targetId));
@@ -208,6 +256,13 @@ export const createCdpSession = (
         if (ownership) {
           ownership.setHarnessWindow(c.targetId);
           ownership.add(c.targetId);
+          // Capture the real Chrome windowId — the durable identity of the
+          // window this session initialized (survives seed-tab closure).
+          const win = await transport.request("Browser.getWindowForTarget", { targetId: c.targetId }, { sessionId: null });
+          if (win.success) {
+            const wid = (win.data as { windowId?: number }).windowId;
+            if (typeof wid === "number") ownership.setHarnessWindowId(wid);
+          }
         }
       }
 
diff --git a/src/cdp/window-binding.ts b/src/cdp/window-binding.ts
new file mode 100644
index 0000000..4ae1ac3
--- /dev/null
+++ b/src/cdp/window-binding.ts
@@ -0,0 +1,15 @@
+import type { BrowserClient } from "../client";
+
+/**
+ * Capture the real Chrome windowId of a freshly-created harness window and bind
+ * it as the session's durable window identity. Best-effort: a failed CDP call
+ * leaves the previous binding untouched. Call this on every code path that
+ * spawns a NEW dedicated window (newWindow:true), so the windowId never points
+ * at a window that has since been replaced.
+ */
+export const bindHarnessWindowId = async (client: BrowserClient, targetId: string): Promise<void> => {
+  const r = await client.session().callBrowser("Browser.getWindowForTarget", { targetId });
+  if (!r.success) return;
+  const wid = (r.data as { windowId?: number }).windowId;
+  if (typeof wid === "number") client.ownership().setHarnessWindowId(wid);
+};
diff --git a/src/client.ts b/src/client.ts
index 724a711..9a97369 100644
--- a/src/client.ts
+++ b/src/client.ts
@@ -16,10 +16,12 @@ export type BrowserClientOptions = {
   readonly initialOwnership?: {
     readonly ownedTargetIds?: ReadonlyArray<string>;
     readonly harnessWindowTargetId?: string;
+    readonly harnessWindowId?: number;
   };
   readonly onOwnershipChange?: (snapshot: {
     readonly ownedTargetIds: ReadonlyArray<string>;
     readonly harnessWindowTargetId: string | undefined;
+    readonly harnessWindowId: number | undefined;
   }) => void;
 };
 
@@ -32,6 +34,9 @@ export type BrowserClient = {
   /** Detach from the current page target (removes the "Chrome is being controlled" banner)
    *  while keeping the transport connection alive. Call on session shutdown. */
   detach(): Promise<void>;
+  /** Close every tab this session owns and clear the window binding. Best-effort —
+   *  used on session shutdown so no stale harness tabs survive. */
+  closeOwnedTabs(): Promise<void>;
   evaluateJs(expression: string, sessionId?: string): Promise<Result<unknown, CdpError>>;
   pageInfo(): Promise<Result<PageInfo | { readonly dialog: DialogInfo }, CdpError>>;
   takeDialog(): DialogInfo | null;
@@ -82,13 +87,16 @@ const parsePageInfoPayload = (v: unknown): Result<PageInfo, CdpError> => {
 
 export const createBrowserClient = (opts: BrowserClientOptions): BrowserClient => {
   const transport = opts.transport ?? createCdpTransport();
-  const ownershipInit: { ownedTargetIds?: ReadonlyArray<string>; harnessWindowTargetId?: string } = {};
+  const ownershipInit: { ownedTargetIds?: ReadonlyArray<string>; harnessWindowTargetId?: string; harnessWindowId?: number } = {};
   if (opts.initialOwnership?.ownedTargetIds !== undefined) {
     ownershipInit.ownedTargetIds = opts.initialOwnership.ownedTargetIds;
   }
   if (opts.initialOwnership?.harnessWindowTargetId !== undefined) {
     ownershipInit.harnessWindowTargetId = opts.initialOwnership.harnessWindowTargetId;
   }
+  if (opts.initialOwnership?.harnessWindowId !== undefined) {
+    ownershipInit.harnessWindowId = opts.initialOwnership.harnessWindowId;
+  }
   const ownership = createOwnershipRegistry(ownershipInit);
   if (opts.onOwnershipChange) {
     const cb = opts.onOwnershipChange;
@@ -96,6 +104,7 @@ export const createBrowserClient = (opts: BrowserClientOptions): BrowserClient =
       cb({
         ownedTargetIds: ownership.list(),
         harnessWindowTargetId: ownership.harnessWindow(),
+        harnessWindowId: ownership.harnessWindowId(),
       });
     });
   }
@@ -256,6 +265,15 @@ export const createBrowserClient = (opts: BrowserClientOptions): BrowserClient =
     return ok(undefined);
   };
 
+  // Ask Chrome for the real windowId hosting a target. Undefined on failure —
+  // callers treat the window binding as best-effort.
+  const getWindowId = async (targetId: string): Promise<number | undefined> => {
+    const r = await session.callBrowser("Browser.getWindowForTarget", { targetId });
+    if (!r.success) return undefined;
+    const wid = (r.data as { windowId?: number }).windowId;
+    return typeof wid === "number" ? wid : undefined;
+  };
+
   const newTab = async (url?: string): Promise<Result<string, CdpError>> => {
     // Verify the recorded harness window still exists. Querying the live
     // target list also gives listTabs's reconciliation a chance to run.
@@ -263,22 +281,35 @@ export const createBrowserClient = (opts: BrowserClientOptions): BrowserClient =
     if (!tabsResult.success) return tabsResult;
     const live = new Set(tabsResult.data.map((t) => t.targetId));
     const hw = ownership.harnessWindow();
+    // Pick an opener that keeps us inside the harness window. Prefer the seed
+    // tab; if the user closed it but other owned tabs survive, reuse one of
+    // those instead of spawning a second window (avoids window scatter).
+    const opener = hw && live.has(hw)
+      ? hw
+      : ownership.list().find((id) => live.has(id));
 
     const params: Record<string, unknown> = { url: "about:blank" };
-    if (hw && live.has(hw)) {
-      // Open as a child of the harness window's seed tab — Chrome places
-      // it in the same window, giving us visual grouping for free.
-      params["openerId"] = hw;
+    if (opener) {
+      // Open as a child of an owned tab — Chrome places it in the same window,
+      // giving us visual grouping for free and keeping the session single-window.
+      params["openerId"] = opener;
     } else {
-      // Either fresh session or the harness window was closed by the user.
-      // Spawn a new dedicated window.
+      // Fresh session, or the whole harness window was closed by the user.
+      // Spawn a new dedicated window and rebind to it.
       params["newWindow"] = true;
     }
     const created = await session.callBrowser("Target.createTarget", params);
     if (!created.success) return created;
     const c = created.data as { targetId: string };
 
-    if (params["newWindow"]) ownership.setHarnessWindow(c.targetId);
+    if (params["newWindow"]) {
+      ownership.setHarnessWindow(c.targetId);
+      const wid = await getWindowId(c.targetId);
+      if (wid !== undefined) ownership.setHarnessWindowId(wid);
+    } else if (opener && opener !== hw) {
+      // Re-seeded onto a surviving owned tab; keep the seed pointer live.
+      ownership.setHarnessWindow(opener);
+    }
     ownership.add(c.targetId);
 
     const switched = await switchTab(c.targetId);
@@ -297,6 +328,18 @@ export const createBrowserClient = (opts: BrowserClientOptions): BrowserClient =
     return ok(undefined);
   };
 
+  const closeOwnedTabs = async (): Promise<void> => {
+    // Snapshot first — remove() mutates the list, and targetDestroyed events
+    // may also prune concurrently. Failures are swallowed: a tab the user
+    // already closed, or a dead transport, must not block shutdown.
+    for (const id of ownership.list()) {
+      await session.callBrowser("Target.closeTarget", { targetId: id }).catch(() => {});
+      ownership.remove(id);
+    }
+    ownership.setHarnessWindow(undefined);
+    ownership.setHarnessWindowId(undefined);
+  };
+
   const status = (): DaemonStatus => ({
     alive: transport.state() === "open" && session.current() !== null,
     sessionId: session.current()?.sessionId ?? null,
@@ -316,7 +359,7 @@ export const createBrowserClient = (opts: BrowserClientOptions): BrowserClient =
 
   return {
     namespace: opts.namespace,
-    ensureAlive, status, start, stop, detach,
+    ensureAlive, status, start, stop, detach, closeOwnedTabs,
     evaluateJs, pageInfo,
     takeDialog: () => session.takeDialog(),
     listTabs, switchTab, newTab, closeTab,
diff --git a/src/domains/forms.ts b/src/domains/forms.ts
new file mode 100644
index 0000000..1d83ba6
--- /dev/null
+++ b/src/domains/forms.ts
@@ -0,0 +1,355 @@
+import { Type } from "typebox";
+import type { BrowserClient } from "../client";
+import { type Result, err, ok } from "../util/result";
+import { defineBrowserTool, type ToolErr, type ToolOk } from "../util/tool";
+
+// ── Page-side functions ──────────────────────────────────────────────────────
+// These run inside the page via Runtime.callFunctionOn with `this` bound to the
+// resolved element. They are NOT string-interpolated with user data — the value
+// is passed as a structured `arguments` entry, so safeJs is unnecessary here.
+
+/**
+ * Universal field setter. Auto-detects the element type and applies the right
+ * strategy, then reads the resulting value back so the caller can confirm.
+ *
+ * The <input>/<textarea> branch uses the "native setter" trick: React (and
+ * other frameworks) overload the element's `value` setter to track state, so a
+ * plain `el.value = x` updates the DOM but not the framework. Calling the
+ * prototype's original setter and then dispatching a bubbling `input` event is
+ * what makes React's onChange fire.
+ */
+const FILL_FN = `
+function(value) {
+  const el = this;
+  if (!el || el.nodeType !== 1) return { ok: false, reason: "ref does not point to an element" };
+  if (el.disabled) return { ok: false, reason: "element is disabled" };
+  const tag = (el.tagName || "").toLowerCase();
+  const type = (el.type || "").toLowerCase();
+  try { el.focus(); } catch (e) {}
+  const fire = function(t) { el.dispatchEvent(new Event(t, { bubbles: true })); };
+
+  if (tag === "select") {
+    const want = String(value);
+    let matched = null;
+    for (let i = 0; i < el.options.length; i++) {
+      const o = el.options[i];
+      if (o.value === want || o.label === want || o.text === want) { matched = o; break; }
+    }
+    if (!matched) {
+      const opts = [];
+      for (let i = 0; i < el.options.length; i++) opts.push({ value: el.options[i].value, text: el.options[i].text });
+      return { ok: false, reason: "no matching option", kind: "select", options: opts };
+    }
+    el.value = matched.value;
+    fire("input"); fire("change");
+    return { ok: true, kind: "select", value: el.value, text: matched.text };
+  }
+
+  if (tag === "input" && (type === "checkbox" || type === "radio")) {
+    const want = value === true || value === "true" || value === "on" || value === 1;
+    if (el.checked !== want) {
+      try { el.click(); } catch (e) {}
+      if (el.checked !== want) { el.checked = want; fire("input"); fire("change"); }
+    }
+    return { ok: true, kind: type, checked: el.checked };
+  }
+
+  if (el.isContentEditable) {
+    el.textContent = String(value);
+    fire("input"); fire("change");
+    return { ok: true, kind: "contenteditable", value: el.textContent };
+  }
+
+  if (tag === "input" || tag === "textarea") {
+    const proto = tag === "textarea" ? window.HTMLTextAreaElement.prototype : window.HTMLInputElement.prototype;
+    const desc = Object.getOwnPropertyDescriptor(proto, "value");
+    if (desc && desc.set) desc.set.call(el, String(value));
+    else el.value = String(value);
+    fire("input"); fire("change");
+    return { ok: true, kind: tag, value: el.value };
+  }
+
+  return { ok: false, reason: "element is not a fillable field (tag=" + tag + ")" };
+}`;
+
+/** Specialized <select> setter — matches by value first, then by label/text. */
+const SELECT_FN = `
+function(value, label) {
+  const el = this;
+  if (!el || (el.tagName || "").toLowerCase() !== "select") return { ok: false, reason: "ref is not a <select> element" };
+  if (el.disabled) return { ok: false, reason: "select is disabled" };
+  let matched = null;
+  for (let i = 0; i < el.options.length; i++) {
+    const o = el.options[i];
+    if (value != null && o.value === String(value)) { matched = o; break; }
+    if (label != null && (o.text === String(label) || o.label === String(label))) { matched = o; break; }
+  }
+  if (!matched) {
+    const opts = [];
+    for (let i = 0; i < el.options.length; i++) opts.push({ value: el.options[i].value, text: el.options[i].text });
+    return { ok: false, reason: "no matching option", options: opts };
+  }
+  try { el.focus(); } catch (e) {}
+  el.value = matched.value;
+  el.dispatchEvent(new Event("input", { bubbles: true }));
+  el.dispatchEvent(new Event("change", { bubbles: true }));
+  return { ok: true, value: el.value, text: matched.text };
+}`;
+
+/** Specialized checkbox/radio setter — only acts when the state differs. */
+const CHECK_FN = `
+function(checked) {
+  const el = this;
+  const tag = (el.tagName || "").toLowerCase();
+  const type = (el.type || "").toLowerCase();
+  if (tag !== "input" || (type !== "checkbox" && type !== "radio")) return { ok: false, reason: "ref is not a checkbox or radio" };
+  if (el.disabled) return { ok: false, reason: "element is disabled" };
+  const want = checked === true;
+  let changed = false;
+  if (el.checked !== want) {
+    try { el.click(); } catch (e) {}
+    if (el.checked !== want) { el.checked = want; el.dispatchEvent(new Event("change", { bubbles: true })); }
+    changed = true;
+  }
+  return { ok: true, checked: el.checked, changed: changed };
+}`;
+
+// ── Resolver ─────────────────────────────────────────────────────────────────
+
+/** Shape returned by the page-side functions above. */
+type PageResult = {
+  readonly ok: boolean;
+  readonly reason?: string;
+  readonly kind?: string;
+  readonly value?: unknown;
+  readonly text?: string;
+  readonly checked?: boolean;
+  readonly changed?: boolean;
+  readonly options?: ReadonlyArray<{ value: string; text: string }>;
+};
+
+/**
+ * Resolve a ref (CDP backendDOMNodeId) to a live JS handle and invoke a
+ * page-side function on it. DOM + Runtime are enabled on every session
+ * (see cdp/session.ts enableDomains), so no extra setup is required.
+ */
+const resolveAndCall = async (
+  client: BrowserClient,
+  ref: number,
+  functionDeclaration: string,
+  args: ReadonlyArray<unknown>,
+): Promise<Result<PageResult, ToolErr>> => {
+  const session = client.session();
+
+  const resolved = await session.call("DOM.resolveNode", { backendNodeId: ref });
+  if (!resolved.success) {
+    return err({
+      kind: "invalid_state",
+      message: `ref ${ref} could not be resolved — the page may have changed. Re-run browser_snapshot to get fresh refs.`,
+      details: { ref },
+    });
+  }
+  const objectId = (resolved.data as { object?: { objectId?: string } }).object?.objectId;
+  if (objectId === undefined) {
+    return err({
+      kind: "invalid_state",
+      message: `ref ${ref} resolved to no JS object — re-run browser_snapshot to get fresh refs.`,
+      details: { ref },
+    });
+  }
+
+  try {
+    const called = await session.call("Runtime.callFunctionOn", {
+      objectId,
+      functionDeclaration,
+      arguments: args.map((v) => ({ value: v })),
+      returnByValue: true,
+      awaitPromise: true,
+    });
+    if (!called.success) return err({ kind: "cdp_error", message: called.error.message, details: { ref } });
+    const data = called.data as { result?: { value?: unknown }; exceptionDetails?: unknown };
+    if (data.exceptionDetails !== undefined) {
+      return err({ kind: "cdp_error", message: `page function threw: ${JSON.stringify(data.exceptionDetails)}`, details: { ref } });
+    }
+    const value = data.result?.value;
+    if (typeof value !== "object" || value === null) {
+      return err({ kind: "internal", message: "page function returned no result object", details: { ref } });
+    }
+    return ok(value as PageResult);
+  } finally {
+    // Best-effort: release the remote handle so it can be GC'd. Failure is
+    // harmless (the page may have navigated), so we ignore the result.
+    await session.call("Runtime.releaseObject", { objectId });
+  }
+};
+
+/** Drop the `ok` flag before stashing a page result in tool `details` — the
+ *  tool-result envelope reserves `ok` for its own success discriminant. */
+const detailsOf = (ref: number, res: PageResult, extra?: Readonly<Record<string, unknown>>): Readonly<Record<string, unknown>> => {
+  const { ok: _ok, ...rest } = res;
+  return { ref, ...rest, ...(extra ?? {}) };
+};
+
+// ── Tools ────────────────────────────────────────────────────────────────────
+
+const FieldValue = Type.Union([Type.String(), Type.Boolean()], {
+  description: "Value to set. String for text inputs / textareas / selects / contenteditable; boolean for checkboxes & radios.",
+});
+
+const FillArgs = Type.Object({
+  ref: Type.Integer({ description: "Element ref from browser_snapshot ([ref=N])." }),
+  value: FieldValue,
+});
+
+export const fillTool = defineBrowserTool({
+  name: "browser_fill",
+  label: "Browser Fill",
+  description:
+    "Fill a single form field by ref. Auto-detects input/textarea/select/checkbox/radio/contenteditable, clears then sets the value, fires input+change events (drives React/Vue controlled inputs correctly), and reads the value back to confirm. Get the ref from browser_snapshot's [ref=N].",
+  promptSnippet: "Fill a form field by ref (text, select, checkbox) — React-safe, self-confirming",
+  promptGuidelines: [
+    "Get the ref from browser_snapshot's [ref=N] for the target field. No click-to-focus needed.",
+    "value is a string for text fields/selects/contenteditable, or a boolean for checkboxes/radios.",
+    "The result reports the value the field actually holds after filling — check it matches what you intended.",
+    "For autocomplete/typeahead that needs real keystrokes, click the field's @(x,y) then use browser_type instead.",
+    "If a ref no longer resolves, re-run browser_snapshot to get fresh refs.",
+  ],
+  parameters: FillArgs,
+  async handler(args, { client }): Promise<Result<ToolOk, ToolErr>> {
+    const r = await resolveAndCall(client, args.ref, FILL_FN, [args.value]);
+    if (!r.success) return r;
+    const res = r.data;
+    if (!res.ok) {
+      return err({ kind: "invalid_state", message: `Could not fill ref ${args.ref}: ${res.reason ?? "unknown reason"}`, details: detailsOf(args.ref, res) });
+    }
+    const wanted = typeof args.value === "boolean" ? undefined : String(args.value);
+    const got = res.value !== undefined ? String(res.value) : undefined;
+    const mismatch = wanted !== undefined && got !== undefined && got !== wanted;
+    let text: string;
+    if (mismatch) {
+      text = `Filled ref ${args.ref}, but it now reads ${JSON.stringify(got)} (wanted ${JSON.stringify(wanted)})`;
+    } else if (got !== undefined) {
+      text = `Filled ref ${args.ref} = ${JSON.stringify(got)}`;
+    } else if (res.checked !== undefined) {
+      text = `Filled ref ${args.ref} (checked=${res.checked})`;
+    } else {
+      text = `Filled ref ${args.ref}`;
+    }
+    return ok({ text, details: detailsOf(args.ref, res, { mismatch }) });
+  },
+});
+
+const FillFormArgs = Type.Object({
+  fields: Type.Array(
+    Type.Object({
+      ref: Type.Integer({ description: "Element ref from browser_snapshot ([ref=N])." }),
+      value: FieldValue,
+    }),
+    { minItems: 1, description: "Fields to fill in one batch, each by ref." },
+  ),
+});
+
+export const fillFormTool = defineBrowserTool({
+  name: "browser_fill_form",
+  label: "Browser Fill Form",
+  description:
+    "Fill multiple form fields in one call — the efficient way to complete a form. Each field is { ref, value }; refs come from browser_snapshot's [ref=N]. Handles text inputs, textareas, selects, checkboxes, radios, and contenteditable, drives React/Vue controlled inputs correctly, and reports the resulting value of every field.",
+  promptSnippet: "Fill many form fields at once by ref — React-safe, self-confirming",
+  promptGuidelines: [
+    "Preferred for forms: snapshot once, then fill all fields in a single browser_fill_form call.",
+    "Each field: { ref, value }. value is a string for text/select/contenteditable, boolean for checkbox/radio.",
+    "The result summarizes how many fields filled cleanly and flags any value mismatches or per-field errors.",
+    "Re-run browser_snapshot afterwards to verify the form state if needed.",
+  ],
+  parameters: FillFormArgs,
+  async handler(args, { client }): Promise<Result<ToolOk, ToolErr>> {
+    const results: Array<Readonly<Record<string, unknown>>> = [];
+    let okCount = 0;
+    const issues: string[] = [];
+    for (const f of args.fields) {
+      const r = await resolveAndCall(client, f.ref, FILL_FN, [f.value]);
+      if (!r.success) {
+        results.push({ ref: f.ref, ok: false, error: r.error.message });
+        issues.push(`ref ${f.ref} (${r.error.message})`);
+        continue;
+      }
+      const res = r.data;
+      const wanted = typeof f.value === "boolean" ? undefined : String(f.value);
+      const got = res.value !== undefined ? String(res.value) : undefined;
+      const mismatch = wanted !== undefined && got !== undefined && got !== wanted;
+      const fieldOk = res.ok === true && !mismatch;
+      if (fieldOk) okCount++;
+      else issues.push(`ref ${f.ref} (${res.ok === false ? (res.reason ?? "failed") : "value mismatch"})`);
+      results.push({ ref: f.ref, ok: fieldOk, mismatch, ...detailsOf(f.ref, res) });
+    }
+    const text =
+      `Filled ${okCount}/${args.fields.length} fields` +
+      (issues.length > 0 ? `; issues: ${issues.join(", ")}` : "");
+    return ok({ text, details: { results } });
+  },
+});
+
+const SelectOptionArgs = Type.Object({
+  ref: Type.Integer({ description: "Element ref of a <select> from browser_snapshot ([ref=N])." }),
+  value: Type.Optional(Type.String({ description: "Option value attribute to select." })),
+  label: Type.Optional(Type.String({ description: "Visible option text/label to select (used if value is omitted)." })),
+});
+
+export const selectOptionTool = defineBrowserTool({
+  name: "browser_select_option",
+  label: "Browser Select Option",
+  description:
+    "Select an option in a native <select> dropdown by ref. Matches by option value, then by visible label/text, and fires change. Native selects cannot be driven by clicking coordinates — use this instead. Get the ref from browser_snapshot's [ref=N].",
+  promptSnippet: "Choose an option in a <select> dropdown by ref",
+  promptGuidelines: [
+    "Provide value (the option's value attribute) or label (the visible text). value wins if both given.",
+    "On no match, the result lists the available options so you can retry with a valid one.",
+    "Use this for native <select> elements — clicking coordinates does not work for OS dropdowns.",
+  ],
+  parameters: SelectOptionArgs,
+  async handler(args, { client }): Promise<Result<ToolOk, ToolErr>> {
+    if (args.value === undefined && args.label === undefined) {
+      return err({ kind: "invalid_state", message: "Provide value or label to select an option." });
+    }
+    const r = await resolveAndCall(client, args.ref, SELECT_FN, [args.value ?? null, args.label ?? null]);
+    if (!r.success) return r;
+    const res = r.data;
+    if (!res.ok) {
+      return err({ kind: "invalid_state", message: `Could not select option on ref ${args.ref}: ${res.reason ?? "unknown reason"}`, details: detailsOf(args.ref, res) });
+    }
+    return ok({
+      text: `Selected ${JSON.stringify(res.text)} (value=${JSON.stringify(res.value)}) on ref ${args.ref}`,
+      details: detailsOf(args.ref, res),
+    });
+  },
+});
+
+const SetCheckedArgs = Type.Object({
+  ref: Type.Integer({ description: "Element ref of a checkbox or radio from browser_snapshot ([ref=N])." }),
+  checked: Type.Boolean({ description: "Desired checked state." }),
+});
+
+export const setCheckedTool = defineBrowserTool({
+  name: "browser_set_checked",
+  label: "Browser Set Checked",
+  description:
+    "Set a checkbox or radio to a desired checked state by ref. Only acts if the current state differs, fires change, and reports the final state. Get the ref from browser_snapshot's [ref=N].",
+  promptSnippet: "Set a checkbox/radio checked state by ref",
+  promptGuidelines: [
+    "checked:true ticks the box, checked:false unticks it. Idempotent — no-op if already in the desired state.",
+    "For radios, set checked:true on the option you want; the group deselects the others.",
+  ],
+  parameters: SetCheckedArgs,
+  async handler(args, { client }): Promise<Result<ToolOk, ToolErr>> {
+    const r = await resolveAndCall(client, args.ref, CHECK_FN, [args.checked]);
+    if (!r.success) return r;
+    const res = r.data;
+    if (!res.ok) {
+      return err({ kind: "invalid_state", message: `Could not set checked on ref ${args.ref}: ${res.reason ?? "unknown reason"}`, details: detailsOf(args.ref, res) });
+    }
+    return ok({
+      text: `Set ref ${args.ref} checked=${res.checked}` + (res.changed === false ? " (already)" : ""),
+      details: detailsOf(args.ref, res),
+    });
+  },
+});
diff --git a/src/domains/isolated-tab.ts b/src/domains/isolated-tab.ts
index c759cec..73ed158 100644
--- a/src/domains/isolated-tab.ts
+++ b/src/domains/isolated-tab.ts
@@ -13,6 +13,7 @@
 import type { BrowserClient } from "../client";
 import { type Result, err, ok } from "../util/result";
 import { safeJs } from "../util/js-template";
+import { bindHarnessWindowId } from "../cdp/window-binding";
 import type { ToolErr } from "../util/tool";
 
 /** A tab attached by its own sessionId, ready to drive via callOnTarget. */
@@ -40,7 +41,10 @@ export const openIsolatedTab = async (client: BrowserClient): Promise<Result<Iso
   if (!created.success) return err(toToolErr(created.error.message));
   const { targetId } = created.data as { targetId: string };
 
-  if (createParams["newWindow"]) client.ownership().setHarnessWindow(targetId);
+  if (createParams["newWindow"]) {
+    client.ownership().setHarnessWindow(targetId);
+    await bindHarnessWindowId(client, targetId);
+  }
   client.ownership().add(targetId);
 
   const attached = await client.session().callBrowser("Target.attachToTarget", { targetId, flatten: true });
diff --git a/src/domains/navigate.ts b/src/domains/navigate.ts
index 4477485..e402efc 100644
--- a/src/domains/navigate.ts
+++ b/src/domains/navigate.ts
@@ -4,6 +4,7 @@ import { type Result, err, ok } from "../util/result";
 import { defineBrowserTool, type ToolErr, type ToolOk } from "../util/tool";
 import { applyTruncation } from "../util/truncate";
 import { safeJs } from "../util/js-template";
+import { bindHarnessWindowId } from "../cdp/window-binding";
 
 const NavigateArgs = Type.Object({
   url: Type.String({ description: "Full URL to navigate to (e.g. https://github.com)" }),
@@ -75,7 +76,7 @@ export const openUrlsTool = defineBrowserTool({
   description: "Open multiple URLs in parallel new tabs. Returns per-URL outcomes.",
   promptSnippet: "Open multiple URLs in new tabs (parallel)",
   promptGuidelines: [
-    "Use after web_search to open citations in parallel.",
+    "Use after browser_web_search to open citations in parallel.",
     "After opening, use browser_list_tabs / browser_switch_tab to navigate, then browser_snapshot to inspect each tab.",
     "Use browser_wait_for_load on a tab before extracting data from SPAs.",
   ],
@@ -83,17 +84,31 @@ export const openUrlsTool = defineBrowserTool({
   renderCall: (a) => new Text(`🌐 Opening ${a.urls.length} URL${a.urls.length !== 1 ? "s" : ""}…`, 0, 0),
   async handler(args, { client, onUpdate }): Promise<Result<ToolOk, ToolErr>> {
     const total = args.urls.length;
-    // Phase 1: create tabs in parallel.  Own them so they appear in
-    // browser_list_tabs scope:'owned' and group in the harness window.
-    const hw = client.ownership().harnessWindow();
-    const created = await Promise.all(args.urls.map(async (url): Promise<TabResult> => {
-      const params: Record<string, unknown> = { url: "about:blank" };
-      if (hw) params["openerId"] = hw;
-      else params["newWindow"] = true;
-      const r = await client.session().callBrowser("Target.createTarget", params);
+    // Phase 1: create tabs. All must land in ONE window. If the harness window
+    // already exists, open every url as its child in parallel. If it doesn't
+    // (fresh/closed), create the FIRST url as a new dedicated window (binding
+    // its windowId) and open the rest as children of it — a parallel
+    // newWindow:true per url would otherwise scatter tabs across N windows.
+    let seed = client.ownership().harnessWindow();
+    let seedResult: TabResult | undefined;
+    if (!seed) {
+      const url0 = args.urls[0];
+      if (url0 === undefined) return err({ kind: "invalid_state", message: "no URLs provided" });
+      const r0 = await client.session().callBrowser("Target.createTarget", { url: "about:blank", newWindow: true });
+      if (!r0.success) return err({ kind: "cdp_error", message: r0.error.message });
+      seed = (r0.data as { targetId: string }).targetId;
+      client.ownership().setHarnessWindow(seed);
+      await bindHarnessWindowId(client, seed);
+      client.ownership().add(seed);
+      seedResult = { url: url0, targetId: seed, ok: true };
+    }
+    const seededWindow = seed;
+    const created = await Promise.all(args.urls.map(async (url, i): Promise<TabResult> => {
+      // Reuse the seed tab (created above) for the first url instead of duplicating it.
+      if (i === 0 && seedResult) return seedResult;
+      const r = await client.session().callBrowser("Target.createTarget", { url: "about:blank", openerId: seededWindow });
       if (!r.success) return { url, targetId: "", ok: false, error: r.error.message };
       const c = r.data as { targetId: string };
-      if (params["newWindow"]) client.ownership().setHarnessWindow(c.targetId);
       client.ownership().add(c.targetId);
       return { url, targetId: c.targetId, ok: true };
     }));
diff --git a/src/index.ts b/src/index.ts
index d5f3435..7109ea0 100644
--- a/src/index.ts
+++ b/src/index.ts
@@ -109,9 +109,10 @@ export default function browserHarnessExtension(pi: ExtensionAPI): void {
     if (!client) {
       const transport = createDaemonTransport(state.namespace);
 
-      const initialOwnership: { ownedTargetIds?: ReadonlyArray<string>; harnessWindowTargetId?: string } = {};
+      const initialOwnership: { ownedTargetIds?: ReadonlyArray<string>; harnessWindowTargetId?: string; harnessWindowId?: number } = {};
       if (state.ownedTargetIds !== undefined) initialOwnership.ownedTargetIds = state.ownedTargetIds;
       if (state.harnessWindowTargetId !== undefined) initialOwnership.harnessWindowTargetId = state.harnessWindowTargetId;
+      if (state.harnessWindowId !== undefined) initialOwnership.harnessWindowId = state.harnessWindowId;
 
       client = createBrowserClient({
         namespace: state.namespace,
@@ -124,6 +125,9 @@ export default function browserHarnessExtension(pi: ExtensionAPI): void {
             ...(snap.harnessWindowTargetId !== undefined
               ? { harnessWindowTargetId: snap.harnessWindowTargetId }
               : {}),
+            ...(snap.harnessWindowId !== undefined
+              ? { harnessWindowId: snap.harnessWindowId }
+              : {}),
           };
           // Ownership changes can fire from CDP events at any time,
           // including after session replacement when pi's ctx is stale.
@@ -147,20 +151,27 @@ export default function browserHarnessExtension(pi: ExtensionAPI): void {
   });
 
   pi.on("session_shutdown", async () => {
-    persistState(pi, state);
     if (client) {
       try {
         // Detach from the page target (removes the "Chrome is being controlled"
         // banner) but keep the transport alive for the next session.
         await client.detach();
+        // Close every tab this session opened so no stale harness tabs linger
+        // in the user's browser between sessions. This also clears the persisted
+        // ownership + window binding via onOwnershipChange, so the next session
+        // starts with a fresh dedicated window rather than reattaching to leftovers.
+        await client.closeOwnedTabs();
       } catch (e) {
-        console.warn("[pi-browser-harness] client.detach() failed during shutdown:", e);
+        console.warn("[pi-browser-harness] browser teardown failed during shutdown:", e);
       }
       // ponytail: keep the transport alive across sessions.
       // Do NOT call client.stop() or null out client — the daemon connection
       // persists and eliminates the per-session "Allow Remote Debugging" prompt.
       // Only Chrome restart or daemon death triggers a new prompt.
     }
+    // Persist last — closeOwnedTabs mutates `state` through onOwnershipChange,
+    // so this captures the cleared ownership set.
+    persistState(pi, state);
     await cleanupTempDirs();
   });
 
diff --git a/src/prompt.ts b/src/prompt.ts
index 7483730..c336c0c 100644
--- a/src/prompt.ts
+++ b/src/prompt.ts
@@ -2,150 +2,126 @@
  * System prompt guidance for browser usage.
  *
  * Injects a compact browser control reference into the pi system prompt.
- * Per-tool guidelines are injected via promptGuidelines on each tool registration.
- * This prompt provides only the cross-cutting patterns and tool reference table.
+ * This prompt carries only cross-cutting knowledge — the mental model, tool
+ * routing, workflows, and gotchas. Per-tool specifics (schemas, detailed usage)
+ * are injected separately via promptGuidelines on each tool registration, so this
+ * file intentionally does NOT restate full tool descriptions.
  */
 
 export function getBrowserSystemPrompt(): string {
   return `
 ## Browser Control
 
-You have full control of a real Chrome browser via \`browser_*\` tools.
-The browser connects to the user's running Chrome (not a headless instance).
-
-**These tools coexist with all other standard pi tools** — read, bash, edit,
-write, and any other built-in tools are still available. Use browser_* tools
-to interact with pages visually, and other tools alongside them.
-
-### Browser-Specific Tools
-
-| Tool | Purpose |
-|------|---------|
-| browser_snapshot | Accessibility-tree snapshot. Assigns every interactive element a stable **ref** (\`[eN]\`) plus \`@(x,y)\`. Refs are the primary handle for interaction tools — they survive re-renders, coordinates don't. |
-| browser_screenshot | Capture a PNG or JPEG screenshot of the current page (rendered inline in the TUI; the LLM reads the image via the saved file path) |
-| browser_click | Click an element by **ref** (preferred — re-resolves position, survives re-renders) or viewport x/y. Appends a compact diff of page changes. |
-| browser_type | Type text into the focused element (real keystrokes; requires a focused field) |
-| browser_fill | Fill an input/textarea by **ref** (preferred) or selector — updates React/Vue controlled components, verifies the value, appends a page-changes diff |
-| browser_select_option | Select an option in a native <select> by **ref** (preferred) or selector, choosing by value/label/index (fires change) |
-| browser_focus | Focus an element by **ref** (preferred) or selector (deterministic; use before browser_type) |
-| browser_press_key | Press a keyboard key (Enter, Tab, Escape, arrows, etc.) |
-| browser_scroll | Scroll the page at coordinates |
-| browser_navigate | Navigate to a URL (result details.outcome.kind is "in_place" or "new_tab_created") |
-| browser_new_tab | Open a new tab, optionally navigate |
-| browser_open_urls | Open multiple URLs in new tabs (parallel) |
-| browser_go_back / browser_go_forward / browser_reload | History navigation |
-| browser_page_info | Get page URL, title, viewport, scroll position, or dialog info |
-| browser_list_tabs | List all open browser tabs |
-| browser_current_tab | Get current tab info |
-| browser_switch_tab | Switch to a different tab by targetId |
-| browser_execute_js | Execute JavaScript and return the result |
-| browser_http_get | Direct HTTP GET (outside browser, for APIs) |
-| browser_console | Read JS errors / console output on the current tab (diagnostic — use when an action looks broken; pass sinceSeq from the previous nextCursor to see only new messages) |
-| browser_wait | Wait N seconds |
-| browser_wait_for | Wait until a selector/text appears (or a selector disappears) — use for SPA content before interacting |
-| browser_wait_for_load | Wait for document.readyState === 'complete' (returns a typed timeout error if the page doesn't reach readyState=complete in N seconds, default 15) |
-| browser_handle_dialog | Accept or dismiss a JS dialog |
-| browser_run_script | Execute a temporary script file with daemon access (write script to disk, then run) |
-
-### Temporary Scripts
-
-When the built-in tools aren't enough for a multi-step workflow, write a temporary
-script to disk and execute it with browser_run_script. The script runs in the
-harness process with direct access to the browser daemon and Node.js APIs.
+You drive a **real Chrome** (the user's running instance, not headless) through
+\`browser_*\` tools. These coexist with every standard pi tool — read, bash, edit,
+write — so mix them freely: \`browser_*\` for the page, everything else alongside.
 
-\`\`\`
-write("/tmp/scrape-pages.js", \`
-  const results = [];
-  for (const url of params.urls) {
-    await daemon.cdp("Page.navigate", { url });
-    await new Promise(r => setTimeout(r, 2000));
-    const data = await daemon.evaluateJS("document.title");
-    results.push({ url, title: data });
-  }
-  return { content: [{ type: "text", text: JSON.stringify(results, null, 2) }] };
-\`)
-browser_run_script("/tmp/scrape-pages.js", { urls: [...] })
-\`\`\`
+### The core handle: refs
 
-Script bindings: params, daemon, require, signal, onUpdate, ctx, console, fetch, JSON, Buffer, setTimeout, clearTimeout.
+Interaction is ref-first. Call \`browser_snapshot\` to get an accessibility tree
+where every interactive element has a stable ref (\`[eN]\`) and \`@(x,y)\`. **Target
+elements by ref, not by coordinates or guessed selectors** — refs are keyed to
+element identity, so they survive the re-renders that React/Vue forms trigger on
+every edit and save. Coordinates go stale; guessed selectors miss.
 
-### Parallel Execution
+- Prefer \`browser_fill\` for text — it fires input/change so controlled components
+  update, and returns the value for verification. Use \`browser_click\` + \`browser_type\`
+  only for keystroke-sensitive widgets (autocomplete, masked inputs).
+- Mutating calls append a compact **"Page changes"** diff — read it to confirm the
+  change landed (form closed, new \`*[eN]\` appeared).
+- \`"ref is stale"\` means the page changed: re-run \`browser_snapshot\` for fresh refs.
 
-Observation tools (browser_screenshot, browser_page_info, browser_execute_js, browser_list_tabs, browser_http_get, etc.) can run in parallel with each other and with mutation tools. The harness automatically serializes mutation tools (click, type, scroll, navigate, switch_tab, etc.) so they never race on shared state. When operations are independent, emit them in the same turn for better performance.
+### Choosing a tool
 
-**Examples of safe parallel calls:**
-\`\`\`
-browser_screenshot() + browser_page_info() + browser_execute_js("document.title")
-browser_http_get("https://api.example.com/data") + browser_click(x, y)
-\`\`\`
+| To… | Reach for |
+|-----|-----------|
+| See the page structure (for interaction) | \`browser_snapshot\` (refs) |
+| See the page visually | \`browser_screenshot\` |
+| Search the web | \`browser_web_search\` → ranked {title, url, snippet} |
+| Read an article's clean text | \`browser_read_page\` (reader mode; url or owned tab) |
+| Extract specific DOM values | \`browser_execute_js\` |
+| Hit a JSON/API endpoint | \`browser_http_get\` (raw GET, outside the browser) |
+| Click / fill / type / select / focus / press a key | \`browser_click\` · \`browser_fill\` · \`browser_type\` · \`browser_select_option\` · \`browser_focus\` · \`browser_press_key\` |
+| Upload a file · drag · resize viewport | \`browser_upload_file\` · \`browser_drag_and_drop\` · \`browser_viewport_resize\` |
+| Go to a URL / open many / manage tabs | \`browser_navigate\` · \`browser_new_tab\` · \`browser_open_urls\` · \`browser_list_tabs\` · \`browser_switch_tab\` · \`browser_go_back\`/\`go_forward\`/\`reload\` |
+| Wait for content / load / a fixed delay | \`browser_wait_for\` (selector/text) · \`browser_wait_for_load\` (readyState) · \`browser_wait\` |
+| Inspect page / handle a dialog | \`browser_page_info\` · \`browser_current_tab\` · \`browser_handle_dialog\` |
+| Diagnose a broken action | \`browser_console\` (JS errors) · \`browser_get_network_log\` |
+| Save output / config downloads | \`browser_print_to_pdf\` · \`browser_download\` |
+| Anything the tools can't express | \`browser_run_script\` (see Extending) |
 
-**Multi-agent note**: When multiple subagents use the browser, tab switching by one agent changes the active tab for all agents. Per-tab data (console buffers, network traces, dialogs, page info cache) is isolated — switching tabs does not destroy another agent's collected data. Call \`browser_current_tab\` before mutation tools to confirm you're on the expected tab. If \`browser_page_info\` returns a dialog, handle it with \`browser_handle_dialog\` promptly — dialogs block page interaction and are not queued across agents.
+Each tool carries its own detailed guidelines — this table is only for routing.
 
-### Common Patterns
+### Workflows
 
-**Navigation:**
+**Research (search → read):**
 \`\`\`
-browser_new_tab("https://example.com") → browser_wait_for_load() → browser_screenshot()
+browser_web_search("query")        // ranked links, no page content
+→ browser_read_page(url)           // clean article text for the top hits
+// fall back to browser_open_urls + browser_execute_js only when reader mode misses
 \`\`\`
 
-**Form filling (ref-first — this is the reliable path on React/SPA forms):**
+**Form filling (ref-first — the reliable path on SPA forms):**
 \`\`\`
-browser_wait_for({ selector: "#email" })                    // ensure SPA field is rendered
-browser_snapshot()                                          // each interactive element gets a [eN] ref
-→ browser_fill({ ref: "e7", value: "a@b.com" })            // target by ref, not a guessed selector
-→ browser_select_option({ ref: "e9", label: "India" })     // native <select> by ref
-→ browser_click({ ref: "e12" })                            // e.g. Save — re-resolves position even if the form reflowed
-// each mutating call appends a compact "Page changes" diff — read it to confirm
-// the change landed (e.g. a save closed the form, a new field appeared as *[eN]).
+browser_wait_for({ selector: "#email" })   // ensure the field is rendered
+browser_snapshot()                          // interactive elements get [eN] refs
+→ browser_fill({ ref: "e7", value: "a@b.com" })
+→ browser_select_option({ ref: "e9", label: "India" })
+→ browser_click({ ref: "e12" })             // re-resolves position even after reflow
+// read each "Page changes" diff to confirm the step landed
 \`\`\`
-**Why ref over selector/coordinates:** refs are keyed to the element's identity, so
-they survive the re-renders that React/Vue forms trigger on every edit and save —
-coordinates go stale and selectors you guess often don't match. If a call returns
-"ref is stale", the page changed: re-run browser_snapshot for fresh refs.
-
-Prefer browser_fill for text inputs/textareas — it fires input/change so controlled
-components update, and returns the field's value for verification. Use browser_click +
-browser_type only for keystroke-sensitive widgets (autocomplete, masked inputs);
-browser_type requires an already-focused field and errors if none is focused.
 
-**Data extraction:**
+**Navigate + verify:**
 \`\`\`
-browser_execute_js("document.querySelector('.price').innerText")
-// or for APIs:
-browser_http_get("https://api.example.com/data")
+browser_new_tab("https://example.com") → browser_wait_for_load() → browser_screenshot()
 \`\`\`
 
-**Scrolling:**
-\`\`\`
-browser_screenshot() → browser_scroll({ deltaY: 500 }) → browser_screenshot()
-\`\`\`
-Note: deltaY follows W3C wheel-event convention: positive=down, negative=up. Default deltaY=300 scrolls down.
+**Scroll:** \`browser_scroll({ deltaY: 500 })\` — W3C wheel convention: positive = down,
+negative = up (default 300, down).
 
-**Research Workflow (search + browser):**
+### Parallelism
+
+Observation tools (\`browser_screenshot\`, \`browser_page_info\`, \`browser_execute_js\`,
+\`browser_list_tabs\`, \`browser_http_get\`, \`browser_web_search\`, \`browser_read_page\`, …)
+run in parallel with each other and with mutations. The harness serializes mutations
+(click, type, scroll, navigate, switch_tab) so they never race. Emit independent calls
+in one turn:
 \`\`\`
-browser_navigate("https://google.com/search?q=...") → search engine in a tab
-browser_open_urls(urls: ["url1", "url2", ...]) → open result pages in parallel tabs
-browser_list_tabs() → see all open tabs with targetIds
-browser_switch_tab(targetId: "...") → switch to a tab
-browser_screenshot() → visually inspect the page
-browser_execute_js("document.querySelector('.main').innerText") → extract content
+browser_screenshot() + browser_page_info() + browser_execute_js("document.title")
 \`\`\`
 
-**Temporary Scripts (extending the harness):**
+**Multi-agent:** tab switching by one agent changes the active tab for all agents,
+but per-tab data (console, network, dialogs) stays isolated. Call \`browser_current_tab\`
+before a mutation to confirm you're on the expected tab. Handle any dialog reported by
+\`browser_page_info\` promptly — dialogs block interaction and aren't queued across agents.
+
+**Session boundary & tab hygiene:** you operate only inside this session's own Chrome
+window. \`browser_switch_tab\`/\`browser_close_tab\` refuse tabs this session didn't open,
+and \`browser_list_tabs\` defaults to owned tabs — never reach into the user's other tabs
+or windows. Close tabs when you're done with them: call \`browser_close_tab\` on any tab
+you opened (via \`browser_new_tab\`/\`browser_open_urls\`) once you've extracted what you
+need. Don't leave a pile of stale tabs behind — keep only what a later step still needs.
+(Leftover tabs are closed automatically when the session ends, but close them yourself
+as you go.)
+
+### Extending: temporary scripts
+
+When built-in tools can't express a multi-step workflow, write a script to a temp file
+and run it with \`browser_run_script\`. It executes in the harness process with direct
+daemon + Node access.
 \`\`\`
-write("/tmp/extract.js", "...script with daemon access...") → browser_run_script("/tmp/extract.js")
+write(<tempfile>.js, \`
+  const results = [];
+  for (const url of params.urls) {
+    await daemon.cdp("Page.navigate", { url });
+    await new Promise(r => setTimeout(r, 2000));
+    results.push({ url, title: await daemon.evaluateJS("document.title") });
+  }
+  return { content: [{ type: "text", text: JSON.stringify(results, null, 2) }] };
+\`)
+browser_run_script(<tempfile>.js, { urls: [...] })
 \`\`\`
-
-### Additional Tools
-
-| Tool | Purpose |
-|------|---------|
-| browser_upload_file | Upload a file to a file input by **ref** (preferred) or selector (bypasses file picker) |
-| browser_dispatch_key | Dispatch a DOM KeyboardEvent on an element by **ref** (preferred) or selector (for React/Vue inputs); returns details.matched |
-| browser_download | Configure download directory and disable save-as prompts |
-| browser_viewport_resize | Resize the viewport for responsive testing |
-| browser_drag_and_drop | Perform drag-and-drop from one coordinate to another |
-| browser_print_to_pdf | Print the current page to a PDF file |
-| browser_get_network_log | Get buffered network request/response events |
+Script bindings: params, daemon, require, signal, onUpdate, ctx, console, fetch, JSON,
+Buffer, setTimeout, clearTimeout. Use a real temp path from the OS — don't hardcode \`/tmp\`.
 `;
 }
diff --git a/src/state.ts b/src/state.ts
index 0629f40..d4b1534 100644
--- a/src/state.ts
+++ b/src/state.ts
@@ -12,6 +12,7 @@ export type BrowserState = {
   readonly remoteBrowserId?: string;
   readonly ownedTargetIds?: ReadonlyArray<string>;
   readonly harnessWindowTargetId?: string;
+  readonly harnessWindowId?: number;
 };
 
 export const defaultState = (namespace = "default"): BrowserState => ({