Skip to content

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

fix(social-publish): preserve scoped publication clauses

Barack Sokullu998ab61a1c11
diff --git a/src/chrome/src/agent/agent.js b/src/chrome/src/agent/agent.js
index 68dfa3dbf..904f99a7e 100644
--- a/src/chrome/src/agent/agent.js
+++ b/src/chrome/src/agent/agent.js
@@ -261,7 +261,7 @@ const SOCIAL_POST_DESTINATION_NEGATION = new RegExp(
'iu',
);
const SOCIAL_CONTRASTIVE_EXCLUSION = /(?<![\p{L}\p{N}_])(?:rather\s+than|instead\s+of)(?:\s+(?:post(?:ing)?|publish(?:ing)?|shar(?:e|ing)|tweet(?:ing)?|send(?:ing)?))?(?:\s+(?:on|onto|to|via|in|at))?\s*$/iu;
-const SOCIAL_DESTINATION_EXCLUSION = /^\s*(?:exclude|excluding|avoid|avoiding|no)\s+(?:(?:post(?:ing)?|publish(?:ing)?|shar(?:e|ing)|tweet(?:ing)?|send(?:ing)?)\s+)?(?:(?:on|onto|to|via|in|at)\s+)?(?:the\s+)?(?:x|twitter|bluesky|bsky(?:\.app)?)(?![\p{L}\p{N}_.-])\s*$/iu;
+const SOCIAL_DESTINATION_EXCLUSION = /^\s*(?:exclude|excluding|avoid|avoiding|skip|skipping|no)\s+(?:(?:post(?:ing)?|publish(?:ing)?|shar(?:e|ing)|tweet(?:ing)?|send(?:ing)?)\s+)?(?:(?:on|onto|to|via|in|at)\s+)?(?:the\s+)?(?:x|twitter|bluesky|bsky(?:\.app)?)(?![\p{L}\p{N}_.-])\s*$/iu;
const socialPostNegationGovernsPublish = value => (
SOCIAL_POST_NEGATION.test(String(value || '').trim())
|| SOCIAL_POST_DESTINATION_NEGATION.test(String(value || ''))
@@ -3474,6 +3474,7 @@ export class Agent extends LoopDetector {
'giu',
);
const attachmentKindForAlternativeNoun = noun => {
+ if (new RegExp(`^(?:${IMAGE_ATTACHMENT_FORMAT}|${VIDEO_ATTACHMENT_FORMAT})$`, 'iu').test(noun)) return '';
if (RAW_GIF_NOUN_REGEX.test(noun)) return 'gif';
if (RAW_IMAGE_NOUN_REGEX.test(noun)) return 'image';
if (RAW_VIDEO_NOUN_REGEX.test(noun)) return 'video';
@@ -17428,11 +17429,22 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d
_socialPublicationClauses(text) {
const raw = String(text || '');
const masked = this._maskQuotedPayload(raw);
- const parts = masked.split(SOCIAL_CLAUSE_DELIMITER);
+ // URL punctuation is content, not clause structure. Hide only the
+ // delimiter characters from the splitter and slice the original masked
+ // text below, so platform discovery still receives intact URLs.
+ const delimiterMasked = masked.replace(
+ /https?:\/\/[^\s<>"'`\u3002\u3001\uff0c\uff1b\uff1a\uff01\uff1f\u2026\u2025]+/gi,
+ (rawUrl) => {
+ const url = this._workflowTrimUrlPunctuation(rawUrl);
+ return url.replace(/[.!?;:,]/g, '\u0001') + rawUrl.slice(url.length);
+ },
+ );
+ const parts = delimiterMasked.split(SOCIAL_CLAUSE_DELIMITER);
const clauses = [];
let offset = 0;
for (let i = 0; i < parts.length; i += 2) {
- const maskedClause = parts[i] || '';
+ const clauseLength = (parts[i] || '').length;
+ const maskedClause = masked.slice(offset, offset + clauseLength);
const delim = i > 0 ? (parts[i - 1] || '').trim() : '';
const rawClause = raw.slice(offset, offset + maskedClause.length);
clauses.push({ text: rawClause, maskedText: maskedClause, delim });
diff --git a/src/firefox/src/agent/agent.js b/src/firefox/src/agent/agent.js
index 550010eb6..d0a6b3598 100644
--- a/src/firefox/src/agent/agent.js
+++ b/src/firefox/src/agent/agent.js
@@ -195,7 +195,7 @@ const SOCIAL_POST_DESTINATION_NEGATION = new RegExp(
'iu',
);
const SOCIAL_CONTRASTIVE_EXCLUSION = /(?<![\p{L}\p{N}_])(?:rather\s+than|instead\s+of)(?:\s+(?:post(?:ing)?|publish(?:ing)?|shar(?:e|ing)|tweet(?:ing)?|send(?:ing)?))?(?:\s+(?:on|onto|to|via|in|at))?\s*$/iu;
-const SOCIAL_DESTINATION_EXCLUSION = /^\s*(?:exclude|excluding|avoid|avoiding|no)\s+(?:(?:post(?:ing)?|publish(?:ing)?|shar(?:e|ing)|tweet(?:ing)?|send(?:ing)?)\s+)?(?:(?:on|onto|to|via|in|at)\s+)?(?:the\s+)?(?:x|twitter|bluesky|bsky(?:\.app)?)(?![\p{L}\p{N}_.-])\s*$/iu;
+const SOCIAL_DESTINATION_EXCLUSION = /^\s*(?:exclude|excluding|avoid|avoiding|skip|skipping|no)\s+(?:(?:post(?:ing)?|publish(?:ing)?|shar(?:e|ing)|tweet(?:ing)?|send(?:ing)?)\s+)?(?:(?:on|onto|to|via|in|at)\s+)?(?:the\s+)?(?:x|twitter|bluesky|bsky(?:\.app)?)(?![\p{L}\p{N}_.-])\s*$/iu;
const socialPostNegationGovernsPublish = value => (
SOCIAL_POST_NEGATION.test(String(value || '').trim())
|| SOCIAL_POST_DESTINATION_NEGATION.test(String(value || ''))
@@ -3189,6 +3189,7 @@ export class Agent extends LoopDetector {
'giu',
);
const attachmentKindForAlternativeNoun = noun => {
+ if (new RegExp(`^(?:${IMAGE_ATTACHMENT_FORMAT}|${VIDEO_ATTACHMENT_FORMAT})$`, 'iu').test(noun)) return '';
if (RAW_GIF_NOUN_REGEX.test(noun)) return 'gif';
if (RAW_IMAGE_NOUN_REGEX.test(noun)) return 'image';
if (RAW_VIDEO_NOUN_REGEX.test(noun)) return 'video';
@@ -15227,11 +15228,22 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d
_socialPublicationClauses(text) {
const raw = String(text || '');
const masked = this._maskQuotedPayload(raw);
- const parts = masked.split(SOCIAL_CLAUSE_DELIMITER);
+ // URL punctuation is content, not clause structure. Hide only the
+ // delimiter characters from the splitter and slice the original masked
+ // text below, so platform discovery still receives intact URLs.
+ const delimiterMasked = masked.replace(
+ /https?:\/\/[^\s<>"'`\u3002\u3001\uff0c\uff1b\uff1a\uff01\uff1f\u2026\u2025]+/gi,
+ (rawUrl) => {
+ const url = this._workflowTrimUrlPunctuation(rawUrl);
+ return url.replace(/[.!?;:,]/g, '\u0001') + rawUrl.slice(url.length);
+ },
+ );
+ const parts = delimiterMasked.split(SOCIAL_CLAUSE_DELIMITER);
const clauses = [];
let offset = 0;
for (let i = 0; i < parts.length; i += 2) {
- const maskedClause = parts[i] || '';
+ const clauseLength = (parts[i] || '').length;
+ const maskedClause = masked.slice(offset, offset + clauseLength);
const delim = i > 0 ? (parts[i - 1] || '').trim() : '';
const rawClause = raw.slice(offset, offset + maskedClause.length);
clauses.push({ text: rawClause, maskedText: maskedClause, delim });
diff --git a/test/run.js b/test/run.js
index 20b2d459d..dd8bc8e2a 100644
--- a/test/run.js
+++ b/test/run.js
@@ -90825,6 +90825,10 @@ test('social publication workflow follows the live X or Bluesky destination and
'an avoiding clause adopted its forbidden X destination'],
['Post on X, avoid Bluesky', ['twitter'],
'an avoid clause adopted its forbidden Bluesky destination'],
+ ['Post on Bluesky, skip X', ['bluesky'],
+ 'a skip clause adopted its forbidden X destination'],
+ ['Post on X, skipping Bluesky', ['twitter'],
+ 'a skipping clause adopted its forbidden Bluesky destination'],
['Post on Bluesky rather than posting on X', ['bluesky'],
'rather-than clause adopted its expressly excluded X destination'],
['Post on Bluesky rather than on X', ['bluesky'],
@@ -92459,6 +92463,16 @@ test('extracting post body skips URL scheme colons', () => {
'Announcing v2!',
AgentClass.name + ': Bluesky URL scheme colon halted body extraction',
);
+ assert.equal(
+ agent._extractWorkflowTaskBody('Post on X: https://example.com/path'),
+ 'https://example.com/path',
+ AgentClass.name + ': punctuation inside an unquoted URL truncated the post body',
+ );
+ assert.equal(
+ agent._extractWorkflowTaskBody('Post on X: Visit https://example.com/path today'),
+ 'Visit https://example.com/path today',
+ AgentClass.name + ': an embedded unquoted URL truncated the surrounding post body',
+ );
}
});
@@ -92996,6 +93010,12 @@ test('upper-bound attachment qualifiers verify as maximum counts', () => {
assert.equal(agent._workflowSocialPublishedAttachmentObserved(
{ value: 'one PNG or JPEG image' }, { attachments: [{ type: 'image', src: 'https://cdn.example/a.webp' }] }), false,
AgentClass.name + ': WebP satisfied a PNG-or-JPEG format choice');
+ const unrestrictedOrPngImage = agent._parseWorkflowAttachmentRequirement('one image or one PNG image');
+ assert.equal(unrestrictedOrPngImage.mediaAlternativeBranches.length, 2,
+ AgentClass.name + ': a format token consumed the noun slot of a whole-media alternative');
+ assert.equal(agent._workflowSocialPublishedAttachmentObserved(
+ { value: 'one image or one PNG image' }, { attachments: [{ type: 'image', src: 'https://cdn.example/a.jpg' }] }), true,
+ AgentClass.name + ': the unrestricted image branch incorrectly required PNG');
const pngAndJpegImages = agent._parseWorkflowAttachmentRequirement('one PNG image and one JPEG image');
assert.equal(pngAndJpegImages.expectedImageCount, 2,
AgentClass.name + ': conjunctive image-format counts were not summed');