Skip to content

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

fix(social-publish): proper-name scope, polite requests, plus ellipsis

Barack Sokulluda2fac1134fe
diff --git a/src/chrome/src/agent/agent.js b/src/chrome/src/agent/agent.js
index d3c4629fb..915348442 100644
--- a/src/chrome/src/agent/agent.js
+++ b/src/chrome/src/agent/agent.js
@@ -288,6 +288,10 @@ const SOCIAL_COORDINATING_DELIMITER = new RegExp(
// destination, but it must not carry polarity: "do not post on X, then post on
// Bluesky" starts a new command while "post on X, then on Bluesky" names two.
const SOCIAL_SEQUENTIAL_DELIMITER = /^(?:then|luego|puis|ensuite|dann|poi|sonra|затем|然后|然後|接着|そして|それから)$/iu;
+// A proper name can masquerade as a leading publication verb: "Post Malone
+// is on Bluesky." is prose about a person, not a command. Post + Name +
+// copula/reporting verb stays body prose instead of opening a new command.
+const SOCIAL_PROPER_NAME_PROSE_LEAD = /^post\s+[A-Z][a-z]+(?:'[a-z]+)?\s+(?:is|are|was|were|be|been|being|has|have|had|say|says|said|announce|announces|announced)\b/i;
const SOCIAL_CLAUSE_BREAK = new RegExp(
`[.!?;:,\\n]|(?<![${SOCIAL_WORD_EDGE}])(?:and|then|but|or|nor|after|before|y|e|ed|luego|puis|et|ensuite|und|dann|poi|sonra|ve|затем|и)(?![${SOCIAL_WORD_EDGE}])|然后|然後|接着|そして|それから|または|、|。`,
'iu',
@@ -3096,17 +3100,19 @@ export class Agent extends LoopDetector {
// Elliptical conjunction ("one PNG and one JPEG image") omits the first
// media noun. Restore it so each counted format parses as its own
// conjunctive clause; "or" choices keep sharing one noun and are handled
- // by the format-qualifier grammar instead.
+ // by the format-qualifier grammar instead. The other unambiguously
+ // conjunctive separators ("plus", "also", "as well as") restore alike.
{
const ellipticalCount = '(?:\\d+|zero|one|two|three|four|five|six|seven|eight|nine|ten|a|an|single|both)';
+ const ellipticalConjunction = '(and|plus|also|as\\s+well\\s+as)';
text = text
.replace(
- new RegExp(`(?<![${SOCIAL_WORD_EDGE}])(${ellipticalCount})\\s+(${IMAGE_ATTACHMENT_FORMAT})\\s+and\\s+(${ellipticalCount})\\s+(${IMAGE_ATTACHMENT_FORMAT})\\s+(images?|photos?|pictures?|pics?)(?![${SOCIAL_WORD_EDGE}])`, 'giu'),
- '$1 $2 $5 and $3 $4 $5',
+ new RegExp(`(?<![${SOCIAL_WORD_EDGE}])(${ellipticalCount})\\s+(${IMAGE_ATTACHMENT_FORMAT})\\s+${ellipticalConjunction}\\s+(${ellipticalCount})\\s+(${IMAGE_ATTACHMENT_FORMAT})\\s+(images?|photos?|pictures?|pics?)(?![${SOCIAL_WORD_EDGE}])`, 'giu'),
+ '$1 $2 $6 $3 $4 $5 $6',
)
.replace(
- new RegExp(`(?<![${SOCIAL_WORD_EDGE}])(${ellipticalCount})\\s+(${VIDEO_ATTACHMENT_FORMAT})\\s+and\\s+(${ellipticalCount})\\s+(${VIDEO_ATTACHMENT_FORMAT})\\s+(videos?|clips?|recordings?)(?![${SOCIAL_WORD_EDGE}])`, 'giu'),
- '$1 $2 $5 and $3 $4 $5',
+ new RegExp(`(?<![${SOCIAL_WORD_EDGE}])(${ellipticalCount})\\s+(${VIDEO_ATTACHMENT_FORMAT})\\s+${ellipticalConjunction}\\s+(${ellipticalCount})\\s+(${VIDEO_ATTACHMENT_FORMAT})\\s+(videos?|clips?|recordings?)(?![${SOCIAL_WORD_EDGE}])`, 'giu'),
+ '$1 $2 $6 $3 $4 $5 $6',
);
}
@@ -18062,7 +18068,9 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d
}
if (inColonBodyScope) {
const clauseDelim = clause.delim || '';
- const leadsWithPublish = startsWithPublishVerb.test(String(clause.maskedText || clause.text || '').trimStart());
+ const leadText = String(clause.maskedText || clause.text || '').trimStart();
+ const leadsWithPublish = startsWithPublishVerb.test(leadText)
+ && !SOCIAL_PROPER_NAME_PROSE_LEAD.test(leadText);
if (!SOCIAL_COORDINATING_DELIMITER.test(clauseDelim)
&& !SOCIAL_SEQUENTIAL_DELIMITER.test(clauseDelim)
&& clauseDelim.trim() !== ','
@@ -26798,6 +26806,7 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d
const startsNewPublication = (clauseMasked) => {
const text = String(clauseMasked || '');
if (!SOCIAL_PUBLISH_VERBS.test(text)) return false;
+ if (SOCIAL_PROPER_NAME_PROSE_LEAD.test(text.trimStart())) return false;
for (const platformMatch of text.matchAll(new RegExp(anyPlatformPattern.source, 'giu'))) {
const before = text.slice(0, platformMatch.index || 0);
const prep = before.match(new RegExp(`(?:${destinationPrepPattern})\\s+(?:the\\s+)?$`, 'iu'));
@@ -26821,12 +26830,14 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d
);
const metadataAttachmentVerb = /(?<![\p{L}\p{N}_])(?:attach(?:ed|ing|ment)?s?|upload(?:ed|ing|s)?|detach(?:ed|ing)?)(?![\p{L}\p{N}_])/iu;
const metadataNarrativeSubject = /(?<![\p{L}\p{N}_])(?:i|we|you|he|she|they|it)(?![\p{L}\p{N}_])[\s,]*$/iu;
+ const metadataPoliteRequest = /(?<![\p{L}\p{N}_])(?:could|can|would|will|shall|should|may|might|please)(?:\s+you)?(?![\p{L}\p{N}_])[\s,]*$/iu;
const startsMetadataInstruction = (clauseMasked) => {
const text = String(clauseMasked || '');
if (!metadataInstructionPattern.test(text)) return false;
const verb = text.match(metadataAttachmentVerb);
if (!verb) return true;
- return !metadataNarrativeSubject.test(text.slice(0, verb.index));
+ const before = text.slice(0, verb.index);
+ return !metadataNarrativeSubject.test(before) || metadataPoliteRequest.test(before);
};
for (const rawCandidate of rawCandidates) {
const clauses = this._socialPublicationClauses(rawCandidate);
diff --git a/src/firefox/src/agent/agent.js b/src/firefox/src/agent/agent.js
index 843b1f8c7..cd1368df1 100644
--- a/src/firefox/src/agent/agent.js
+++ b/src/firefox/src/agent/agent.js
@@ -222,6 +222,10 @@ const SOCIAL_COORDINATING_DELIMITER = new RegExp(
// destination, but it must not carry polarity: "do not post on X, then post on
// Bluesky" starts a new command while "post on X, then on Bluesky" names two.
const SOCIAL_SEQUENTIAL_DELIMITER = /^(?:then|luego|puis|ensuite|dann|poi|sonra|затем|然后|然後|接着|そして|それから)$/iu;
+// A proper name can masquerade as a leading publication verb: "Post Malone
+// is on Bluesky." is prose about a person, not a command. Post + Name +
+// copula/reporting verb stays body prose instead of opening a new command.
+const SOCIAL_PROPER_NAME_PROSE_LEAD = /^post\s+[A-Z][a-z]+(?:'[a-z]+)?\s+(?:is|are|was|were|be|been|being|has|have|had|say|says|said|announce|announces|announced)\b/i;
const SOCIAL_CLAUSE_BREAK = new RegExp(
`[.!?;:,\\n]|(?<![${SOCIAL_WORD_EDGE}])(?:and|then|but|or|nor|after|before|y|e|ed|luego|puis|et|ensuite|und|dann|poi|sonra|ve|затем|и)(?![${SOCIAL_WORD_EDGE}])|然后|然後|接着|そして|それから|または|、|。`,
'iu',
@@ -2811,17 +2815,19 @@ export class Agent extends LoopDetector {
// Elliptical conjunction ("one PNG and one JPEG image") omits the first
// media noun. Restore it so each counted format parses as its own
// conjunctive clause; "or" choices keep sharing one noun and are handled
- // by the format-qualifier grammar instead.
+ // by the format-qualifier grammar instead. The other unambiguously
+ // conjunctive separators ("plus", "also", "as well as") restore alike.
{
const ellipticalCount = '(?:\\d+|zero|one|two|three|four|five|six|seven|eight|nine|ten|a|an|single|both)';
+ const ellipticalConjunction = '(and|plus|also|as\\s+well\\s+as)';
text = text
.replace(
- new RegExp(`(?<![${SOCIAL_WORD_EDGE}])(${ellipticalCount})\\s+(${IMAGE_ATTACHMENT_FORMAT})\\s+and\\s+(${ellipticalCount})\\s+(${IMAGE_ATTACHMENT_FORMAT})\\s+(images?|photos?|pictures?|pics?)(?![${SOCIAL_WORD_EDGE}])`, 'giu'),
- '$1 $2 $5 and $3 $4 $5',
+ new RegExp(`(?<![${SOCIAL_WORD_EDGE}])(${ellipticalCount})\\s+(${IMAGE_ATTACHMENT_FORMAT})\\s+${ellipticalConjunction}\\s+(${ellipticalCount})\\s+(${IMAGE_ATTACHMENT_FORMAT})\\s+(images?|photos?|pictures?|pics?)(?![${SOCIAL_WORD_EDGE}])`, 'giu'),
+ '$1 $2 $6 $3 $4 $5 $6',
)
.replace(
- new RegExp(`(?<![${SOCIAL_WORD_EDGE}])(${ellipticalCount})\\s+(${VIDEO_ATTACHMENT_FORMAT})\\s+and\\s+(${ellipticalCount})\\s+(${VIDEO_ATTACHMENT_FORMAT})\\s+(videos?|clips?|recordings?)(?![${SOCIAL_WORD_EDGE}])`, 'giu'),
- '$1 $2 $5 and $3 $4 $5',
+ new RegExp(`(?<![${SOCIAL_WORD_EDGE}])(${ellipticalCount})\\s+(${VIDEO_ATTACHMENT_FORMAT})\\s+${ellipticalConjunction}\\s+(${ellipticalCount})\\s+(${VIDEO_ATTACHMENT_FORMAT})\\s+(videos?|clips?|recordings?)(?![${SOCIAL_WORD_EDGE}])`, 'giu'),
+ '$1 $2 $6 $3 $4 $5 $6',
);
}
@@ -15861,7 +15867,9 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d
}
if (inColonBodyScope) {
const clauseDelim = clause.delim || '';
- const leadsWithPublish = startsWithPublishVerb.test(String(clause.maskedText || clause.text || '').trimStart());
+ const leadText = String(clause.maskedText || clause.text || '').trimStart();
+ const leadsWithPublish = startsWithPublishVerb.test(leadText)
+ && !SOCIAL_PROPER_NAME_PROSE_LEAD.test(leadText);
if (!SOCIAL_COORDINATING_DELIMITER.test(clauseDelim)
&& !SOCIAL_SEQUENTIAL_DELIMITER.test(clauseDelim)
&& clauseDelim.trim() !== ','
@@ -24405,6 +24413,7 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d
const startsNewPublication = (clauseMasked) => {
const text = String(clauseMasked || '');
if (!SOCIAL_PUBLISH_VERBS.test(text)) return false;
+ if (SOCIAL_PROPER_NAME_PROSE_LEAD.test(text.trimStart())) return false;
for (const platformMatch of text.matchAll(new RegExp(anyPlatformPattern.source, 'giu'))) {
const before = text.slice(0, platformMatch.index || 0);
const prep = before.match(new RegExp(`(?:${destinationPrepPattern})\\s+(?:the\\s+)?$`, 'iu'));
@@ -24428,12 +24437,14 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d
);
const metadataAttachmentVerb = /(?<![\p{L}\p{N}_])(?:attach(?:ed|ing|ment)?s?|upload(?:ed|ing|s)?|detach(?:ed|ing)?)(?![\p{L}\p{N}_])/iu;
const metadataNarrativeSubject = /(?<![\p{L}\p{N}_])(?:i|we|you|he|she|they|it)(?![\p{L}\p{N}_])[\s,]*$/iu;
+ const metadataPoliteRequest = /(?<![\p{L}\p{N}_])(?:could|can|would|will|shall|should|may|might|please)(?:\s+you)?(?![\p{L}\p{N}_])[\s,]*$/iu;
const startsMetadataInstruction = (clauseMasked) => {
const text = String(clauseMasked || '');
if (!metadataInstructionPattern.test(text)) return false;
const verb = text.match(metadataAttachmentVerb);
if (!verb) return true;
- return !metadataNarrativeSubject.test(text.slice(0, verb.index));
+ const before = text.slice(0, verb.index);
+ return !metadataNarrativeSubject.test(before) || metadataPoliteRequest.test(before);
};
for (const rawCandidate of rawCandidates) {
const clauses = this._socialPublicationClauses(rawCandidate);
diff --git a/test/run.js b/test/run.js
index e79659c18..7bd870e9e 100644
--- a/test/run.js
+++ b/test/run.js
@@ -91027,6 +91027,8 @@ test('social publication workflow follows the live X or Bluesky destination and
'per-destination bodies should still discover both platforms'],
['Post on X: I posted on Bluesky yesterday.', ['twitter'],
'colon-scoped body prose adopted Bluesky as a destination'],
+ ['Post on X: Hello. Post Malone is on Bluesky.', ['twitter'],
+ 'proper-name prose left colon scope and adopted Bluesky'],
['Do not publish anything; open https://x.com/compose/post to inspect it, then submit findings in the form', [],
'a prohibited composer URL was adopted as a publish destination'],
['Post on X: hello. Post on Bluesky: world', ['twitter', 'bluesky'],
@@ -91199,6 +91201,10 @@ test('social destination rebinding refreshes platform-specific payload and uploa
AgentClass.name + ': prose mentioning an upload without media ended the post body');
assert.equal(agent._extractWorkflowTaskBody('Post on X: Hello. I upload photos every weekend.', '', 'twitter'), 'Hello. I upload photos every weekend.',
AgentClass.name + ': narrative prose naming media ended the post body');
+ assert.equal(agent._extractWorkflowTaskBody('Post on X: Hello. Could you attach image.png?', '', 'twitter'), 'Hello',
+ AgentClass.name + ': a polite attachment request contaminated the post body');
+ assert.equal(agent._extractWorkflowTaskBody('Post on X: Hello. Post Malone is on Bluesky.', '', 'twitter'), 'Hello. Post Malone is on Bluesky.',
+ AgentClass.name + ': proper-name prose truncated the post body');
assert.equal(agent._extractWorkflowTaskBody('Post on X: Hello world. Then let me know when it is done', '', 'twitter'), 'Hello world',
AgentClass.name + ': sequential follow-up instruction contaminated the post body');
assert.equal(agent._workflowExtractedBodySupersedesClassified('Alpha beta', 'Alpha beta gamma delta'), true,
@@ -93276,6 +93282,13 @@ test('upper-bound attachment qualifiers verify as maximum counts', () => {
assert.equal(agent._workflowSocialPublishedAttachmentObserved(
{ value: 'one PNG and one JPEG image' }, { attachments: [image(1)] },
), false, AgentClass.name + ': one PNG satisfied a two-format elliptical conjunction');
+ const ellipticalPlusFormats = agent._parseWorkflowAttachmentRequirement('one PNG plus one JPEG image');
+ assert.equal(ellipticalPlusFormats.isGeneric, true,
+ AgentClass.name + ': an elliptical plus-format conjunction was parsed as filenames');
+ assert.deepEqual(ellipticalPlusFormats.imageFormatCounts, [
+ { format: 'png', count: 1 },
+ { format: 'jpeg', count: 1 },
+ ], AgentClass.name + ': elliptical plus-format counts were not distributed');
for (const [attachments, expected, message] of [
[[{ type: 'image', src: 'https://cdn.example/a.jpg' }], true, 'the required JPEG was rejected by a separate PNG prohibition'],
[[{ type: 'image', src: 'https://cdn.example/a.webp' }], false, 'an unrequested WebP satisfied a required JPEG'],