Never split an emoji at the edge of a find snippet or a cut read

src/act.js

@@ -157,7 +157,7 @@ export function read(n, { session = DEFAULT_SESSION, budget = 2000 } = {}) {
     // budget it still gets cut: 'up to N tokens' is a promise the page must
     // not be able to break.
     if (!lines.length && cost > budget) {
-      lines.push(`${line.slice(0, budget * 4)} ... cut at ~${budget} tokens, raise --budget for the rest`);
+      lines.push(`${line.slice(0, whole(line, budget * 4))} ... cut at ~${budget} tokens, raise --budget for the rest`);
       spent += budget;
       continue;
     }
@@ -211,6 +211,12 @@ export function submit(n) {
 const BEFORE = 60;
 const SNIPPET = 200;
 
+// A character past U+FFFF (an emoji, a rare CJK ideograph) is two UTF-16
+// units, and an edge between them prints half of one, which a terminal shows
+// as the replacement character. The compact view's cap steps back for the
+// same reason (render.js); a snippet's edges and read's cut do the same.
+const whole = (s, i) => (/[\uD800-\uDBFF]/.test(s[i - 1] ?? '') ? i - 1 : i);
+
 /**
  * Where a string appears on the current page. This is the answer to "the page
  * is long and I only care about one thing in it": one command, no fetch, and
@@ -307,8 +313,9 @@ function search(blocks, terms) {
     // One number, one line: a run of short blocks under the same handle would
     // otherwise report the same place several times.
     if (out.length && out.at(-1).n === n) continue;
-    const start = Math.max(0, Math.min(...found) - BEFORE);
-    const end = Math.min(block.text.length, start + SNIPPET);
+    const from = Math.max(0, Math.min(...found) - BEFORE);
+    const start = whole(block.text, from);
+    const end = whole(block.text, Math.min(block.text.length, from + SNIPPET));
     // One line per match is the promise the list makes, and a code block is the
     // one kind of block that carries lines of its own. They survive where they
     // are read rather than indexed: the whole-match mode above, and `read`.

tests/act.test.js

@@ -212,6 +212,32 @@ test('find opens the snippet on the match, not on the start of a long block', ()
   assert.ok(!out.includes('x'.repeat(250)), `snippet was not trimmed:\n${out.slice(0, 200)}`);
 });
 
+test('neither edge of a find snippet nor a cut read splits an emoji', () => {
+  // An emoji is two UTF-16 units. The snippet window opens 60 units before
+  // the match and runs 200, so here it opened inside the first 😀 and closed
+  // inside the second, and printed a lone half of each.
+  const text = `ab😀${'x'.repeat(59)}needle${'y'.repeat(133)}😀${'z'.repeat(50)}`;
+  saveSession('emoji', {
+    url: 'https://example.test/emoji',
+    blocks: [1, 2, 3].map((n) => ({ n, type: 'text', text: n === 1 ? text : `needle ${'w'.repeat(1000)}` })),
+    cursor: null,
+  });
+  const out = find('needle', { session: 'emoji', budget: 40 });
+  const line = out.split('\n')[1];
+  assert.ok(line.isWellFormed(), `half an emoji in ${JSON.stringify(line)}`);
+  // The opening edge takes the whole 😀, the closing one leaves it out.
+  assert.ok(line.startsWith('[1] ... 😀x') && line.endsWith('y ...'), `the window moved: ${line}`);
+  // read cuts a block over its whole budget at budget * 4 characters; the
+  // 😀 here straddles character 40, after the "[1] " tag.
+  saveSession('emoji-read', {
+    url: 'https://example.test/emoji',
+    blocks: [{ n: 1, type: 'text', text: `${'x'.repeat(35)}😀${'y'.repeat(100)}` }],
+    cursor: null,
+  });
+  const cut = read(1, { session: 'emoji-read', budget: 10 });
+  assert.ok(cut.isWellFormed(), `half an emoji in ${JSON.stringify(cut)}`);
+});
+
 test('find answers with the whole match when the matches fit', () => {
   open();
   // The point of the whole path: the text an agent would have spent a `read