Skip to content

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

fix(social-publish): bind nested media intent

Barack Sokulluacb9c6fdcb73
diff --git a/src/chrome/src/agent/agent.js b/src/chrome/src/agent/agent.js
index b4222c9f1..dfca5369e 100644
--- a/src/chrome/src/agent/agent.js
+++ b/src/chrome/src/agent/agent.js
@@ -3194,6 +3194,15 @@ export class Agent extends LoopDetector {
if (!segmentImage && !segmentVideo && !segmentGif) minScopedGeneric = true;
}
}
+ for (const postfixMinimum of text.matchAll(
+ /(?<![\p{L}\p{N}_])or\s+more\s+(images?|photos?|pictures?|pics?|videos?|clips?|recordings?|gifs?|attachments?|files?|media|uploads?|items?|assets?)(?![\p{L}\p{N}_])/giu,
+ )) {
+ const noun = postfixMinimum[1] || '';
+ if (RAW_GIF_NOUN_REGEX.test(noun)) minScopedGif = true;
+ else if (RAW_IMAGE_NOUN_REGEX.test(noun)) minScopedImage = true;
+ else if (RAW_VIDEO_NOUN_REGEX.test(noun)) minScopedVideo = true;
+ else minScopedGeneric = true;
+ }
if (hasBoundedCountRange) {
if (boundedRangeKind === 'image') minScopedImage = true;
else if (boundedRangeKind === 'video') minScopedVideo = true;
@@ -3280,7 +3289,9 @@ export class Agent extends LoopDetector {
}
const DISJUNCTION_REGEX = /(?:\b(?:or|oder|ou|o|oppure|или|либо|veya|ya\s+da)\b|[或或者]|(?:또는|혹은)|(?:または|それとも))/i;
- const hasDisjunction = isGeneric && DISJUNCTION_REGEX.test(text);
+ // "one or more" is one lower-bound phrase, not a choice between counts.
+ const disjunctionText = text.replace(MIN_ATTACHMENT_COUNT_STRIP_REGEX, ' ');
+ const hasDisjunction = isGeneric && DISJUNCTION_REGEX.test(disjunctionText);
const requestedTypeCount = Number(wantsImage) + Number(wantsOrdinaryVideo) + Number(wantsGif);
const alternativeCounts = hasDisjunction && requestedTypeCount <= 1
? [...countText.matchAll(new RegExp(`(?<![${SOCIAL_WORD_EDGE}])(${countPrefix})(?![${SOCIAL_WORD_EDGE}])`, 'giu'))]
@@ -3578,10 +3589,10 @@ export class Agent extends LoopDetector {
const type = attachmentType(att);
if (type === 'animated_gif' || type === 'gif') return true;
const src = String(att?.src || att?.url || '');
- const alt = String(att?.alt || att?.description || '');
+ const name = String(att?.name || '');
return /\.gif(?:\.mp4)?(?:[?#]|$)/i.test(src)
|| /\/tweet_video(?:_thumb)?\//i.test(src)
- || /\.gif(?:[?#\s]|$)/i.test(alt);
+ || /\.gif(?:[?#]|$)/i.test(name);
};
const canMatchSpecificTargets = (targets, candidates) => {
@@ -17211,7 +17222,11 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d
if (SOCIAL_READ_VERBS.test(beforeVerbClause)) continue;
if (socialNegationGovernsPublish(beforeVerbClause)) continue;
const verbWord = verb[0] || '';
- const afterVerbText = targetText.slice(verbIndex + verbWord.length);
+ const unscopedAfterVerbText = targetText.slice(verbIndex + verbWord.length);
+ const nextPublish = unscopedAfterVerbText.match(SOCIAL_PUBLISH_VERBS);
+ const afterVerbText = nextPublish
+ ? unscopedAfterVerbText.slice(0, nextPublish.index)
+ : unscopedAfterVerbText;
if (socialPostNegationGovernsPublish(afterVerbText)) continue;
if (SOCIAL_NOUN_LIKE_PUBLISH.test(verbWord) && (
SOCIAL_READ_VERBS.test(afterVerbText)
@@ -17219,7 +17234,7 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d
|| NON_SOCIAL_DESTINATION_IN_TEXT.test(beforeVerbClause)
)) continue;
if (/^shares?$/i.test(verbWord) && /(?<![\p{L}\p{N}_])(?:market|mind|revenue|profit|traffic|audience|wallet|fair|lion's|stock|equity|file|screen|time)\s+$/iu.test(beforeVerbClause)) continue;
- const afterVerb = targetText.slice(verbIndex + verbWord.length);
+ const afterVerb = afterVerbText;
const beforeVerb = targetText.slice(0, verbIndex);
for (const matchAfter of afterVerb.matchAll(platformPattern)) {
diff --git a/src/firefox/src/agent/agent.js b/src/firefox/src/agent/agent.js
index 7355f320b..0deef186a 100644
--- a/src/firefox/src/agent/agent.js
+++ b/src/firefox/src/agent/agent.js
@@ -2909,6 +2909,15 @@ export class Agent extends LoopDetector {
if (!segmentImage && !segmentVideo && !segmentGif) minScopedGeneric = true;
}
}
+ for (const postfixMinimum of text.matchAll(
+ /(?<![\p{L}\p{N}_])or\s+more\s+(images?|photos?|pictures?|pics?|videos?|clips?|recordings?|gifs?|attachments?|files?|media|uploads?|items?|assets?)(?![\p{L}\p{N}_])/giu,
+ )) {
+ const noun = postfixMinimum[1] || '';
+ if (RAW_GIF_NOUN_REGEX.test(noun)) minScopedGif = true;
+ else if (RAW_IMAGE_NOUN_REGEX.test(noun)) minScopedImage = true;
+ else if (RAW_VIDEO_NOUN_REGEX.test(noun)) minScopedVideo = true;
+ else minScopedGeneric = true;
+ }
if (hasBoundedCountRange) {
if (boundedRangeKind === 'image') minScopedImage = true;
else if (boundedRangeKind === 'video') minScopedVideo = true;
@@ -2995,7 +3004,9 @@ export class Agent extends LoopDetector {
}
const DISJUNCTION_REGEX = /(?:\b(?:or|oder|ou|o|oppure|или|либо|veya|ya\s+da)\b|[或或者]|(?:또는|혹은)|(?:または|それとも))/i;
- const hasDisjunction = isGeneric && DISJUNCTION_REGEX.test(text);
+ // "one or more" is one lower-bound phrase, not a choice between counts.
+ const disjunctionText = text.replace(MIN_ATTACHMENT_COUNT_STRIP_REGEX, ' ');
+ const hasDisjunction = isGeneric && DISJUNCTION_REGEX.test(disjunctionText);
const requestedTypeCount = Number(wantsImage) + Number(wantsOrdinaryVideo) + Number(wantsGif);
const alternativeCounts = hasDisjunction && requestedTypeCount <= 1
? [...countText.matchAll(new RegExp(`(?<![${SOCIAL_WORD_EDGE}])(${countPrefix})(?![${SOCIAL_WORD_EDGE}])`, 'giu'))]
@@ -3293,10 +3304,10 @@ export class Agent extends LoopDetector {
const type = attachmentType(att);
if (type === 'animated_gif' || type === 'gif') return true;
const src = String(att?.src || att?.url || '');
- const alt = String(att?.alt || att?.description || '');
+ const name = String(att?.name || '');
return /\.gif(?:\.mp4)?(?:[?#]|$)/i.test(src)
|| /\/tweet_video(?:_thumb)?\//i.test(src)
- || /\.gif(?:[?#\s]|$)/i.test(alt);
+ || /\.gif(?:[?#]|$)/i.test(name);
};
const canMatchSpecificTargets = (targets, candidates) => {
@@ -15010,7 +15021,11 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d
if (SOCIAL_READ_VERBS.test(beforeVerbClause)) continue;
if (socialNegationGovernsPublish(beforeVerbClause)) continue;
const verbWord = verb[0] || '';
- const afterVerbText = targetText.slice(verbIndex + verbWord.length);
+ const unscopedAfterVerbText = targetText.slice(verbIndex + verbWord.length);
+ const nextPublish = unscopedAfterVerbText.match(SOCIAL_PUBLISH_VERBS);
+ const afterVerbText = nextPublish
+ ? unscopedAfterVerbText.slice(0, nextPublish.index)
+ : unscopedAfterVerbText;
if (socialPostNegationGovernsPublish(afterVerbText)) continue;
if (SOCIAL_NOUN_LIKE_PUBLISH.test(verbWord) && (
SOCIAL_READ_VERBS.test(afterVerbText)
@@ -15018,7 +15033,7 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d
|| NON_SOCIAL_DESTINATION_IN_TEXT.test(beforeVerbClause)
)) continue;
if (/^shares?$/i.test(verbWord) && /(?<![\p{L}\p{N}_])(?:market|mind|revenue|profit|traffic|audience|wallet|fair|lion's|stock|equity|file|screen|time)\s+$/iu.test(beforeVerbClause)) continue;
- const afterVerb = targetText.slice(verbIndex + verbWord.length);
+ const afterVerb = afterVerbText;
const beforeVerb = targetText.slice(0, verbIndex);
for (const matchAfter of afterVerb.matchAll(platformPattern)) {
diff --git a/test/run.js b/test/run.js
index 18a6bfd5b..6f4ebb255 100644
--- a/test/run.js
+++ b/test/run.js
@@ -90791,6 +90791,8 @@ test('social publication workflow follows the live X or Bluesky destination and
'explicit affirmative publish clause inherited negation from X'],
['Post this not on X but on Bluesky', ['bluesky'],
'post-verb destination negation selected X instead of contrastive Bluesky'],
+ ['Post on X without posting on Bluesky', ['twitter'],
+ 'a nested negated publish clause added its forbidden Bluesky destination'],
['Post or publish this on X', ['twitter'],
'affirmative coordinated publish verbs were wrongly rejected for X'],
['Post "Do not panic" on X', ['twitter'],
@@ -93519,6 +93521,7 @@ test('attachment verification matches specific attachment names without substrin
// A GIF is a video, but an ordinary video is not a GIF.
const gifMedia = {
mp4: { type: 'video', src: 'https://video.twimg.com/media/clip.mp4' },
+ altNamedMp4: { type: 'video', src: 'https://video.twimg.com/media/clip.mp4', alt: 'demo.gif' },
typed: { type: 'animated_gif', src: 'https://video.twimg.com/tweet_video/loop.mp4' },
named: { type: 'video', src: 'https://video.twimg.com/tweet_video/animation.gif.mp4', alt: 'animation.gif' },
plain: { type: 'video', src: 'https://cdn.example.com/loop.gif' },
@@ -93551,6 +93554,19 @@ test('attachment verification matches specific attachment names without substrin
true,
AgentClass.name + ': an ordinary mp4 should pass a negated GIF requirement',
);
+ assert.equal(
+ agent._workflowSocialPublishedAttachmentObserved({ value: 'a GIF' }, { attachments: [gifMedia.altNamedMp4] }),
+ false,
+ AgentClass.name + ': alt text ending in .gif reclassified an ordinary mp4',
+ );
+ assert.equal(
+ agent._workflowSocialPublishedAttachmentObserved(
+ { value: 'one video and no GIFs' },
+ { attachments: [gifMedia.altNamedMp4] },
+ ),
+ true,
+ AgentClass.name + ': GIF-like alt text violated a no-GIF requirement for an ordinary mp4',
+ );
const twoGifRequirement = agent._parseWorkflowAttachmentRequirement({ value: 'two GIFs' });
assert.equal(twoGifRequirement.expectedGifCount, 2,
AgentClass.name + ': the GIF count was collapsed into the general video count');
@@ -93631,6 +93647,7 @@ test('attachment verification matches specific attachment names without substrin
// Minimum-count qualifiers bound the count from below instead of naming a file.
for (const [minReq, minCount] of [
+ ['one or more images', 1],
['at least two images', 2],
['at least 2 images', 2],
['minimum of 3 photos', 3],
@@ -93701,6 +93718,23 @@ test('attachment verification matches specific attachment names without substrin
{ attachments: [image(1), image(2), video(1), video(2)] },
), false, AgentClass.name + ': the video minimum incorrectly loosened the exact image count');
+ const postfixVideoMinimum = { value: 'one image and one or more videos' };
+ const parsedPostfixVideoMinimum = agent._parseWorkflowAttachmentRequirement(postfixVideoMinimum);
+ assert.equal(parsedPostfixVideoMinimum.isAlternative, false,
+ AgentClass.name + ': "or more" was parsed as a media alternative');
+ assert.equal(parsedPostfixVideoMinimum.isVideoMinimum, true,
+ AgentClass.name + ': postfix video minimum scope was lost');
+ assert.equal(parsedPostfixVideoMinimum.isImageMinimum, false,
+ AgentClass.name + ': postfix video minimum leaked onto the image count');
+ assert.equal(agent._workflowSocialPublishedAttachmentObserved(
+ postfixVideoMinimum,
+ { attachments: [image(1), video(1), video(2)] },
+ ), true, AgentClass.name + ': extra video did not satisfy the postfix minimum');
+ assert.equal(agent._workflowSocialPublishedAttachmentObserved(
+ postfixVideoMinimum,
+ { attachments: [image(1)] },
+ ), false, AgentClass.name + ': missing videos satisfied the postfix minimum');
+
const alternativeScopedMinimum = { value: 'at least two images or one video' };
const parsedAlternativeScopedMinimum = agent._parseWorkflowAttachmentRequirement(alternativeScopedMinimum);
assert.equal(parsedAlternativeScopedMinimum.isImageMinimum, true,