Skip to content

Open-source AI browser agent for Chrome and Firefox (monorepo) 🧠

fix(agent): split target lists around finished quoted attachment names

Emre Sokullub7ea1ee9bf62
diff --git a/src/chrome/src/agent/agent.js b/src/chrome/src/agent/agent.js
index 83bc830fc..e3fc13595 100644
--- a/src/chrome/src/agent/agent.js
+++ b/src/chrome/src/agent/agent.js
@@ -313,6 +313,12 @@ const GENERIC_ATTACHMENT_WORDS = new Set([
'와', '과', '및', '그리고', '하고', '도',
]);
+// Quoted attachment names are parked behind these placeholders while a target
+// list is split, so a conjunction inside a quoted name is never a separator.
+const QUOTED_NAME_PLACEHOLDER = /\u0000(\d+)\u0000/;
+const QUOTED_NAME_PLACEHOLDER_TAIL = /\u0000\d+\u0000\s*$/;
+const QUOTED_NAME_PLACEHOLDER_HEAD = /^\s*\u0000\d+\u0000/;
+
// "at least two images" bounds the count from below instead of naming a file,
// so the qualifier is lifted out before the generic-media test and its
// lower-bound meaning is carried on the parsed requirement.
@@ -2544,8 +2550,9 @@ export class Agent extends LoopDetector {
}
// "cat.png and dog.jpg" lists two files, but "research and development.png"
- // names one. A bare conjunction only separates targets when what precedes it
- // already ends in an extension, or when what follows it does not.
+ // names one. Punctuation always separates; a bare conjunction only separates
+ // when the name on its left is already finished -- it ends in an extension or
+ // in a closing quote -- or when what follows is not the tail of a filename.
_splitAttachmentTargetList(text, splitRegex) {
const flags = splitRegex.flags.includes('g') ? splitRegex.flags : `${splitRegex.flags}g`;
const scanner = new RegExp(splitRegex.source, flags);
@@ -2557,6 +2564,8 @@ export class Agent extends LoopDetector {
}
if (boundaries.length === 0) return [text];
const endsWithExtension = str => /\.[a-z0-9]{2,5}$/i.test(String(str).trim());
+ const endsWithQuotedName = str => QUOTED_NAME_PLACEHOLDER_TAIL.test(String(str));
+ const startsWithQuotedName = str => QUOTED_NAME_PLACEHOLDER_HEAD.test(String(str));
const segments = [];
let cursor = 0;
for (let i = 0; i < boundaries.length; i++) {
@@ -2564,7 +2573,9 @@ export class Agent extends LoopDetector {
const left = text.slice(cursor, boundary.start);
const next = text.slice(boundary.end, boundaries[i + 1] ? boundaries[i + 1].start : text.length);
const punctuated = /[,;\n,、]/.test(boundary.delimiter);
- if (!punctuated && !endsWithExtension(left) && endsWithExtension(next)) continue;
+ const leftIsFinished = endsWithExtension(left) || endsWithQuotedName(left);
+ const nextIsFilenameTail = endsWithExtension(next) && !startsWithQuotedName(next);
+ if (!punctuated && !leftIsFinished && nextIsFilenameTail) continue;
segments.push(left);
cursor = boundary.end;
}
@@ -2590,7 +2601,7 @@ export class Agent extends LoopDetector {
},
);
const restoreQuoted = str => String(str).replace(
- /\u0000(\d+)\u0000/g,
+ new RegExp(QUOTED_NAME_PLACEHOLDER.source, 'g'),
(whole, index) => (quotedNames[Number(index)] !== undefined ? quotedNames[Number(index)] : whole),
);
text = text.replace(/^['"“”«»`]+|['"“”«»`]+$/g, '').trim();
diff --git a/src/firefox/src/agent/agent.js b/src/firefox/src/agent/agent.js
index d1bf22810..307c52408 100644
--- a/src/firefox/src/agent/agent.js
+++ b/src/firefox/src/agent/agent.js
@@ -246,6 +246,12 @@ const GENERIC_ATTACHMENT_WORDS = new Set([
'와', '과', '및', '그리고', '하고', '도',
]);
+// Quoted attachment names are parked behind these placeholders while a target
+// list is split, so a conjunction inside a quoted name is never a separator.
+const QUOTED_NAME_PLACEHOLDER = /\u0000(\d+)\u0000/;
+const QUOTED_NAME_PLACEHOLDER_TAIL = /\u0000\d+\u0000\s*$/;
+const QUOTED_NAME_PLACEHOLDER_HEAD = /^\s*\u0000\d+\u0000/;
+
// "at least two images" bounds the count from below instead of naming a file,
// so the qualifier is lifted out before the generic-media test and its
// lower-bound meaning is carried on the parsed requirement.
@@ -2258,8 +2264,9 @@ export class Agent extends LoopDetector {
}
// "cat.png and dog.jpg" lists two files, but "research and development.png"
- // names one. A bare conjunction only separates targets when what precedes it
- // already ends in an extension, or when what follows it does not.
+ // names one. Punctuation always separates; a bare conjunction only separates
+ // when the name on its left is already finished -- it ends in an extension or
+ // in a closing quote -- or when what follows is not the tail of a filename.
_splitAttachmentTargetList(text, splitRegex) {
const flags = splitRegex.flags.includes('g') ? splitRegex.flags : `${splitRegex.flags}g`;
const scanner = new RegExp(splitRegex.source, flags);
@@ -2271,6 +2278,8 @@ export class Agent extends LoopDetector {
}
if (boundaries.length === 0) return [text];
const endsWithExtension = str => /\.[a-z0-9]{2,5}$/i.test(String(str).trim());
+ const endsWithQuotedName = str => QUOTED_NAME_PLACEHOLDER_TAIL.test(String(str));
+ const startsWithQuotedName = str => QUOTED_NAME_PLACEHOLDER_HEAD.test(String(str));
const segments = [];
let cursor = 0;
for (let i = 0; i < boundaries.length; i++) {
@@ -2278,7 +2287,9 @@ export class Agent extends LoopDetector {
const left = text.slice(cursor, boundary.start);
const next = text.slice(boundary.end, boundaries[i + 1] ? boundaries[i + 1].start : text.length);
const punctuated = /[,;\n,、]/.test(boundary.delimiter);
- if (!punctuated && !endsWithExtension(left) && endsWithExtension(next)) continue;
+ const leftIsFinished = endsWithExtension(left) || endsWithQuotedName(left);
+ const nextIsFilenameTail = endsWithExtension(next) && !startsWithQuotedName(next);
+ if (!punctuated && !leftIsFinished && nextIsFilenameTail) continue;
segments.push(left);
cursor = boundary.end;
}
@@ -2304,7 +2315,7 @@ export class Agent extends LoopDetector {
},
);
const restoreQuoted = str => String(str).replace(
- /\u0000(\d+)\u0000/g,
+ new RegExp(QUOTED_NAME_PLACEHOLDER.source, 'g'),
(whole, index) => (quotedNames[Number(index)] !== undefined ? quotedNames[Number(index)] : whole),
);
text = text.replace(/^['"“”«»`]+|['"“”«»`]+$/g, '').trim();
diff --git a/test/run.js b/test/run.js
index b927f6209..199364f4b 100644
--- a/test/run.js
+++ b/test/run.js
@@ -89137,6 +89137,17 @@ test('attachment verification matches specific attachment names without substrin
assert.deepEqual(quotedListReq.specificTargets, ['research and development.png', 'q1 report.png'],
AgentClass.name + ': quoted names split on the delimiter but not on their own conjunctions');
+ // A finished quoted name still ends a target, so the list around it splits.
+ for (const [mixedValue, mixedTargets] of [
+ ['"research and development.png" and cover.jpg', ['research and development.png', 'cover.jpg']],
+ ['cover.jpg and "research and development.png"', ['cover.jpg', 'research and development.png']],
+ ['research and "development.png"', ['research', 'development.png']],
+ ['"my notes" and cover.jpg', ['my notes', 'cover.jpg']],
+ ]) {
+ assert.deepEqual(agent._parseSpecificAttachmentTargets(mixedValue), mixedTargets,
+ AgentClass.name + `: "${mixedValue}" should split around the quoted name`);
+ }
+
// An unquoted conjunction still separates two named files.
const mixedListReq = agent._parseWorkflowAttachmentRequirement({
value: 'research and development.png and cover.jpg',