Cut a long page's text between characters too, not only a control's name

CHANGELOG.md

@@ -8,6 +8,13 @@ Newest first. `Unreleased` is what is on `main` and not yet tagged.
 
 ## Unreleased
 
+### A long page's text is cut between characters, not through one
+
+A navigation hands the Bot the first 6000 UTF-16 code units of the page's readable text. When that
+limit fell between the two halves of an emoji, the Bot was handed text ending on half a character,
+which reads as U+FFFD: a broken character that is not on the page. It now stops one code unit short
+in that case, which is what a control's name and value in a page snapshot already did.
+
 ### Tool selection reads a skill choice the model wrapped in a code fence
 
 Before a run, the deployment's model picks which of a Bot's skills the message needs, so a Bot holding

agent-computer/src/aria-snapshot.ts

@@ -163,8 +163,12 @@ export function parseDescriptor(text: string): Descriptor | null {
  * high surrogate last: JSON carries it as a bare `\ud83d` and UTF-8 as U+FFFD, and the Bot reads a
  * broken character that is not on the page. The server's `cutAtCodeUnits` is the same rule; this
  * process shares no code with the server, so it is repeated here rather than imported.
+ *
+ * Exported for `index.ts`, which cuts the readable page text the same way, and so that the rule has
+ * its own tests. It lives here rather than beside that caller because this module imports no
+ * Playwright: `index.ts` does, at load, so a helper declared there could not be tested at all.
  */
-function cutAtCodeUnits(text: string, limit: number): string {
+export function cutAtCodeUnits(text: string, limit: number): string {
   const sliced = text.slice(0, limit);
   const last = sliced.charCodeAt(sliced.length - 1);
   return last >= 0xd800 && last <= 0xdbff ? sliced.slice(0, -1) : sliced;

agent-computer/src/index.ts

@@ -1,14 +1,18 @@
 import { serve } from "bun";
 import type { Page } from "playwright";
-import { parseAriaSnapshot, type SnapshotElement } from "./aria-snapshot";
-import { browserModeFromEnv } from "./browser-mode";
+import {
+  cutAtCodeUnits,
+  parseAriaSnapshot,
+  type SnapshotElement,
+} from "./aria-snapshot";
 import {
   actsOnTheComputer,
   isOpenPath,
   matchesToken,
   offeredToken,
 } from "./authorisation";
 import { isPlainBotId } from "./bot-id";
+import { browserModeFromEnv } from "./browser-mode";
 import {
   type Control,
   ControlError,
@@ -289,7 +293,18 @@ async function readablePageText(
 
   const collapsed = raw.replace(/\n{3,}/g, "\n\n").trim();
   return {
-    text: collapsed.slice(0, TEXT_EXTRACT_LIMIT),
+    /*
+     * Cut between characters, not through one. #539 fixed this for a control's name and value in the
+     * snapshot and left the page text, which is the same bug at thirty times the length: `slice`
+     * counts UTF-16 code units, an emoji is two, and a limit landing between the halves hands the Bot
+     * a lone high surrogate that reads as U+FFFD — a character that is not on the page.
+     *
+     * More likely to bite here than there, for the reason the limit is bigger. A 200-unit control
+     * name rarely reaches an emoji; 6000 units of somebody's page usually passes through several, and
+     * whether the cut lands mid-character is decided by whatever was above it.
+     */
+    text: cutAtCodeUnits(collapsed, TEXT_EXTRACT_LIMIT),
+    // Unaffected by the line above: dropping one more unit cannot make an over-limit string fit.
     truncated: collapsed.length > TEXT_EXTRACT_LIMIT,
   };
 }

agent-computer/tests/aria-snapshot.test.ts

@@ -1,5 +1,9 @@
 import { describe, expect, test } from "bun:test";
-import { parseAriaSnapshot, parseDescriptor } from "../src/aria-snapshot";
+import {
+  cutAtCodeUnits,
+  parseAriaSnapshot,
+  parseDescriptor,
+} from "../src/aria-snapshot";
 
 /**
  * The parser, tested against captured Playwright output.
@@ -200,6 +204,46 @@ describe("a name or value too long to keep whole", () => {
   });
 });
 
+/**
+ * The cut itself, at any limit.
+ *
+ * Tested directly as well as through the parser because `index.ts` cuts the readable page text with
+ * it at 6000 rather than 200, and that caller imports Playwright at load, so there is no test that
+ * can reach it. The rule is the thing worth pinning, so it is pinned where it is declared.
+ */
+describe("cutting at code units", () => {
+  const EMOJI = "\u{1F600}";
+
+  test("a limit landing between the halves of a character drops the character", () => {
+    const text = `${"a".repeat(5999)}${EMOJI}tail`;
+    const cut = cutAtCodeUnits(text, 6000);
+    expect(cut).toBe("a".repeat(5999));
+    // Not a lone high surrogate, which is what a bare `slice` leaves and what reads as U+FFFD.
+    expect(cut.charCodeAt(cut.length - 1)).toBeLessThan(0xd800);
+  });
+
+  test("a character that ends exactly at the limit is kept whole", () => {
+    const text = `${"a".repeat(5998)}${EMOJI}tail`;
+    expect(cutAtCodeUnits(text, 6000)).toBe(`${"a".repeat(5998)}${EMOJI}`);
+  });
+
+  test("text that fits is returned unchanged, emoji and all", () => {
+    const text = `hello ${EMOJI} world`;
+    expect(cutAtCodeUnits(text, 6000)).toBe(text);
+  });
+
+  test("empty text is empty rather than a thrown index", () => {
+    // `charCodeAt(-1)` is NaN and every comparison against it is false, so this returns "".
+    expect(cutAtCodeUnits("", 6000)).toBe("");
+  });
+
+  test("a low surrogate last is a whole character and is kept", () => {
+    // The guard must look only for an UNPAIRED high surrogate. A complete pair ends on its low half,
+    // and dropping that would cost a character the limit had room for.
+    expect(cutAtCodeUnits(EMOJI, 2)).toBe(EMOJI);
+  });
+});
+
 describe("values a real parser handles and a pattern got wrong", () => {
   test("a quoted numeric value is not left with its quotes", () => {
     // Numeric-looking text remains a string, so one-time codes are not coerced.