Skip to content

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

fix(social-publish): retain named attachment alternatives

Barack Sokullu8908d542a02b
diff --git a/src/chrome/src/agent/agent.js b/src/chrome/src/agent/agent.js
index 64b7ced1b..47d9ba507 100644
--- a/src/chrome/src/agent/agent.js
+++ b/src/chrome/src/agent/agent.js
@@ -2875,7 +2875,7 @@ export class Agent extends LoopDetector {
return segments;
}
- _parseSpecificAttachmentTargets(value) {
+ _parseSpecificAttachmentTargets(value, splitOverride = null) {
let text = String(value || '').trim();
if (!text) return [];
// A quoted span is one filename however many conjunctions it holds, so
@@ -2897,7 +2897,8 @@ export class Agent extends LoopDetector {
(whole, index) => (quotedNames[Number(index)] !== undefined ? quotedNames[Number(index)] : whole),
);
text = text.replace(/^['"“”«»`]+|['"“”«»`]+$/g, '').trim();
- const splitRegex = /(?:\s*[,;\n,、]\s*(?:and|und|et|e|y|ve|и|oder|or|as\s+well\s+as|&)\s*|\s*[,;\n,、]\s*|\s+(?:and|und|et|e|y|ve|и|oder|or|as\s+well\s+as|&)\s+|[와과및]\s*|[和与及以及]\s*|(?<=\.[a-z0-9]{2,5})\s*(?:と|や)\s*)/i;
+ const splitRegex = splitOverride
+ || /(?:\s*[,;\n,、]\s*(?:and|und|et|e|y|ve|и|oder|or|as\s+well\s+as|&)\s*|\s*[,;\n,、]\s*|\s+(?:and|und|et|e|y|ve|и|oder|or|as\s+well\s+as|&)\s+|[와과및]\s*|[和与及以及]\s*|(?<=\.[a-z0-9]{2,5})\s*(?:と|や)\s*)/i;
const parts = this._splitAttachmentTargetList(text, splitRegex)
.map(p => restoreQuoted(p).trim())
.filter(Boolean);
@@ -3249,6 +3250,7 @@ export class Agent extends LoopDetector {
let hasExplicitGenericCount = false;
let expectedCount = 1;
+ let namedTargetAlternatives = [];
if (isGeneric) {
const combinedVideoCount = expectedVideoCount + expectedGifCount;
if (isAlternative) {
@@ -3294,10 +3296,27 @@ export class Agent extends LoopDetector {
}
} else {
const specificTargets = this._parseSpecificAttachmentTargets(rawVal);
- if (specificTargets.length > 1) {
+ // Word joiners need surrounding whitespace so a filename such as
+ // report-or-draft.png remains one target. CJK joiners need no spaces.
+ const disjunctionSplit = /(?:\s+(?:or|oder|ou|o|oppure|или|либо|veya|ya\s+da)\s+|\s*(?:或者|或|または|それとも|또는|혹은)\s*)/iu;
+ const specificTargetAlternatives = DISJUNCTION_REGEX.test(text)
+ ? this._parseSpecificAttachmentTargets(rawVal, disjunctionSplit)
+ .map(group => this._parseSpecificAttachmentTargets(
+ group.replace(/^(?:either|one\s+of)\s+/iu, ''),
+ ))
+ .filter(group => group.length > 0)
+ : [];
+ if (specificTargetAlternatives.length > 1) {
+ expectedCount = Math.min(...specificTargetAlternatives.map(group => group.length));
+ hasExplicitGenericCount = true;
+ } else if (specificTargets.length > 1) {
expectedCount = Math.max(expectedCount, specificTargets.length);
hasExplicitGenericCount = true;
}
+ // Kept below and returned with the rest of the parsed contract.
+ namedTargetAlternatives = specificTargetAlternatives.length > 1
+ ? specificTargetAlternatives
+ : [];
}
const hasExplicitPositiveImageCount = hasExplicitImageCount && !isImageNegated;
@@ -3335,6 +3354,7 @@ export class Agent extends LoopDetector {
wantsOrdinaryVideo,
wantsGif,
specificTargets,
+ specificTargetAlternatives: namedTargetAlternatives || [],
normalized: text,
};
}
@@ -3462,6 +3482,40 @@ export class Agent extends LoopDetector {
|| /\.gif(?:[?#\s]|$)/i.test(alt);
};
+ const canMatchSpecificTargets = (targets, candidates) => {
+ const matchFrom = (targetIdx, usedIndices) => {
+ if (targetIdx >= targets.length) return true;
+ const targetNames = this._targetAttachmentNames(targets[targetIdx]);
+ for (let index = 0; index < candidates.length; index++) {
+ if (usedIndices.has(index)) continue;
+ const candidateNames = this._extractAttachmentCandidateNames(candidates[index]);
+ const matched = targetNames.some(targetName => {
+ const targetHasExt = /\.[a-z0-9]+$/i.test(targetName);
+ return candidateNames.some(candidateName => (
+ candidateName === targetName
+ || (!targetHasExt && candidateName.replace(/\.[a-z0-9]+$/i, '') === targetName)
+ ));
+ });
+ if (!matched) continue;
+ usedIndices.add(index);
+ if (matchFrom(targetIdx + 1, usedIndices)) return true;
+ usedIndices.delete(index);
+ }
+ return false;
+ };
+ return candidates.length >= targets.length && matchFrom(0, new Set());
+ };
+
+ // A named disjunction is authoritative across file types. Checking the
+ // aggregate image/video nouns first would make "chart.png or clip.mp4"
+ // require both types and defeat the alternative contract.
+ if (parsed.specificTargetAlternatives?.length) {
+ return parsed.specificTargetAlternatives.some(group => (
+ rawAttachments.length === group.length
+ && canMatchSpecificTargets(group, rawAttachments)
+ ));
+ }
+
const imageCount = rawAttachments.filter(isImageAttachment).length;
const videoCount = rawAttachments.filter(isVideoAttachment).length;
const gifCount = rawAttachments.filter(isGifAttachment).length;
@@ -3666,35 +3720,7 @@ export class Agent extends LoopDetector {
return false;
}
- const canMatchAll = (targetIdx, usedIndices) => {
- if (targetIdx >= targetsToCheck.length) return true;
- const target = targetsToCheck[targetIdx];
- const targetNames = this._targetAttachmentNames(target);
- for (let i = 0; i < matchingAttachments.length; i++) {
- if (usedIndices.has(i)) continue;
- const att = matchingAttachments[i];
- const candidates = this._extractAttachmentCandidateNames(att);
- const isMatch = targetNames.some(t => {
- const targetHasExt = /\.[a-z0-9]+$/i.test(t);
- return candidates.some(candidate => {
- if (candidate === t) return true;
- if (!targetHasExt) {
- const candidateBase = candidate.replace(/\.[a-z0-9]+$/i, '');
- if (candidateBase === t) return true;
- }
- return false;
- });
- });
- if (isMatch) {
- usedIndices.add(i);
- if (canMatchAll(targetIdx + 1, usedIndices)) return true;
- usedIndices.delete(i);
- }
- }
- return false;
- };
-
- return canMatchAll(0, new Set());
+ return canMatchSpecificTargets(targetsToCheck, matchingAttachments);
}
_workflowSocialPublishedAltTextObserved(requirement, record) {
diff --git a/src/firefox/src/agent/agent.js b/src/firefox/src/agent/agent.js
index f79956b77..7c6d936ea 100644
--- a/src/firefox/src/agent/agent.js
+++ b/src/firefox/src/agent/agent.js
@@ -2590,7 +2590,7 @@ export class Agent extends LoopDetector {
return segments;
}
- _parseSpecificAttachmentTargets(value) {
+ _parseSpecificAttachmentTargets(value, splitOverride = null) {
let text = String(value || '').trim();
if (!text) return [];
// A quoted span is one filename however many conjunctions it holds, so
@@ -2612,7 +2612,8 @@ export class Agent extends LoopDetector {
(whole, index) => (quotedNames[Number(index)] !== undefined ? quotedNames[Number(index)] : whole),
);
text = text.replace(/^['"“”«»`]+|['"“”«»`]+$/g, '').trim();
- const splitRegex = /(?:\s*[,;\n,、]\s*(?:and|und|et|e|y|ve|и|oder|or|as\s+well\s+as|&)\s*|\s*[,;\n,、]\s*|\s+(?:and|und|et|e|y|ve|и|oder|or|as\s+well\s+as|&)\s+|[와과및]\s*|[和与及以及]\s*|(?<=\.[a-z0-9]{2,5})\s*(?:と|や)\s*)/i;
+ const splitRegex = splitOverride
+ || /(?:\s*[,;\n,、]\s*(?:and|und|et|e|y|ve|и|oder|or|as\s+well\s+as|&)\s*|\s*[,;\n,、]\s*|\s+(?:and|und|et|e|y|ve|и|oder|or|as\s+well\s+as|&)\s+|[와과및]\s*|[和与及以及]\s*|(?<=\.[a-z0-9]{2,5})\s*(?:と|や)\s*)/i;
const parts = this._splitAttachmentTargetList(text, splitRegex)
.map(p => restoreQuoted(p).trim())
.filter(Boolean);
@@ -2964,6 +2965,7 @@ export class Agent extends LoopDetector {
let hasExplicitGenericCount = false;
let expectedCount = 1;
+ let namedTargetAlternatives = [];
if (isGeneric) {
const combinedVideoCount = expectedVideoCount + expectedGifCount;
if (isAlternative) {
@@ -3009,10 +3011,27 @@ export class Agent extends LoopDetector {
}
} else {
const specificTargets = this._parseSpecificAttachmentTargets(rawVal);
- if (specificTargets.length > 1) {
+ // Word joiners need surrounding whitespace so a filename such as
+ // report-or-draft.png remains one target. CJK joiners need no spaces.
+ const disjunctionSplit = /(?:\s+(?:or|oder|ou|o|oppure|или|либо|veya|ya\s+da)\s+|\s*(?:或者|或|または|それとも|또는|혹은)\s*)/iu;
+ const specificTargetAlternatives = DISJUNCTION_REGEX.test(text)
+ ? this._parseSpecificAttachmentTargets(rawVal, disjunctionSplit)
+ .map(group => this._parseSpecificAttachmentTargets(
+ group.replace(/^(?:either|one\s+of)\s+/iu, ''),
+ ))
+ .filter(group => group.length > 0)
+ : [];
+ if (specificTargetAlternatives.length > 1) {
+ expectedCount = Math.min(...specificTargetAlternatives.map(group => group.length));
+ hasExplicitGenericCount = true;
+ } else if (specificTargets.length > 1) {
expectedCount = Math.max(expectedCount, specificTargets.length);
hasExplicitGenericCount = true;
}
+ // Kept below and returned with the rest of the parsed contract.
+ namedTargetAlternatives = specificTargetAlternatives.length > 1
+ ? specificTargetAlternatives
+ : [];
}
const hasExplicitPositiveImageCount = hasExplicitImageCount && !isImageNegated;
@@ -3050,6 +3069,7 @@ export class Agent extends LoopDetector {
wantsOrdinaryVideo,
wantsGif,
specificTargets,
+ specificTargetAlternatives: namedTargetAlternatives || [],
normalized: text,
};
}
@@ -3177,6 +3197,40 @@ export class Agent extends LoopDetector {
|| /\.gif(?:[?#\s]|$)/i.test(alt);
};
+ const canMatchSpecificTargets = (targets, candidates) => {
+ const matchFrom = (targetIdx, usedIndices) => {
+ if (targetIdx >= targets.length) return true;
+ const targetNames = this._targetAttachmentNames(targets[targetIdx]);
+ for (let index = 0; index < candidates.length; index++) {
+ if (usedIndices.has(index)) continue;
+ const candidateNames = this._extractAttachmentCandidateNames(candidates[index]);
+ const matched = targetNames.some(targetName => {
+ const targetHasExt = /\.[a-z0-9]+$/i.test(targetName);
+ return candidateNames.some(candidateName => (
+ candidateName === targetName
+ || (!targetHasExt && candidateName.replace(/\.[a-z0-9]+$/i, '') === targetName)
+ ));
+ });
+ if (!matched) continue;
+ usedIndices.add(index);
+ if (matchFrom(targetIdx + 1, usedIndices)) return true;
+ usedIndices.delete(index);
+ }
+ return false;
+ };
+ return candidates.length >= targets.length && matchFrom(0, new Set());
+ };
+
+ // A named disjunction is authoritative across file types. Checking the
+ // aggregate image/video nouns first would make "chart.png or clip.mp4"
+ // require both types and defeat the alternative contract.
+ if (parsed.specificTargetAlternatives?.length) {
+ return parsed.specificTargetAlternatives.some(group => (
+ rawAttachments.length === group.length
+ && canMatchSpecificTargets(group, rawAttachments)
+ ));
+ }
+
const imageCount = rawAttachments.filter(isImageAttachment).length;
const videoCount = rawAttachments.filter(isVideoAttachment).length;
const gifCount = rawAttachments.filter(isGifAttachment).length;
@@ -3381,35 +3435,7 @@ export class Agent extends LoopDetector {
return false;
}
- const canMatchAll = (targetIdx, usedIndices) => {
- if (targetIdx >= targetsToCheck.length) return true;
- const target = targetsToCheck[targetIdx];
- const targetNames = this._targetAttachmentNames(target);
- for (let i = 0; i < matchingAttachments.length; i++) {
- if (usedIndices.has(i)) continue;
- const att = matchingAttachments[i];
- const candidates = this._extractAttachmentCandidateNames(att);
- const isMatch = targetNames.some(t => {
- const targetHasExt = /\.[a-z0-9]+$/i.test(t);
- return candidates.some(candidate => {
- if (candidate === t) return true;
- if (!targetHasExt) {
- const candidateBase = candidate.replace(/\.[a-z0-9]+$/i, '');
- if (candidateBase === t) return true;
- }
- return false;
- });
- });
- if (isMatch) {
- usedIndices.add(i);
- if (canMatchAll(targetIdx + 1, usedIndices)) return true;
- usedIndices.delete(i);
- }
- }
- return false;
- };
-
- return canMatchAll(0, new Set());
+ return canMatchSpecificTargets(targetsToCheck, matchingAttachments);
}
_workflowSocialPublishedAltTextObserved(requirement, record) {
diff --git a/test/run.js b/test/run.js
index 7a3c184d3..16433f7bb 100644
--- a/test/run.js
+++ b/test/run.js
@@ -93227,6 +93227,90 @@ test('attachment verification matches specific attachment names without substrin
);
assert.equal(wrongSecondAttachment, false, AgentClass.name + ': wrong second attachment should reject');
+ const namedAlternatives = agent._parseWorkflowAttachmentRequirement({
+ value: 'chart.png or graph.png',
+ });
+ assert.deepEqual(namedAlternatives.specificTargetAlternatives, [['chart.png'], ['graph.png']],
+ AgentClass.name + ': named attachment alternatives were flattened into one required set');
+ for (const allowedName of ['chart.png', 'graph.png']) {
+ assert.equal(
+ agent._workflowSocialPublishedAttachmentObserved(
+ { value: 'chart.png or graph.png' },
+ { attachments: [{ type: 'image', src: `https://pbs.twimg.com/media/${allowedName}` }] },
+ ),
+ true,
+ AgentClass.name + `: permitted attachment alternative ${allowedName} was rejected`,
+ );
+ }
+ assert.equal(
+ agent._workflowSocialPublishedAttachmentObserved(
+ { value: 'chart.png or graph.png' },
+ {
+ attachments: [
+ { type: 'image', src: 'https://pbs.twimg.com/media/chart.png' },
+ { type: 'image', src: 'https://pbs.twimg.com/media/graph.png' },
+ ],
+ },
+ ),
+ false,
+ AgentClass.name + ': both alternatives were accepted when exactly one was requested',
+ );
+ assert.equal(
+ agent._workflowSocialPublishedAttachmentObserved(
+ { value: 'chart.png or graph.png' },
+ { attachments: [{ type: 'image', src: 'https://pbs.twimg.com/media/other.png' }] },
+ ),
+ false,
+ AgentClass.name + ': unrelated media satisfied a named attachment alternative',
+ );
+ assert.deepEqual(
+ agent._parseWorkflowAttachmentRequirement('either chart.png or graph.png').specificTargetAlternatives,
+ [['chart.png'], ['graph.png']],
+ AgentClass.name + ': either modifier became part of a filename alternative',
+ );
+ assert.deepEqual(
+ agent._parseWorkflowAttachmentRequirement('report-or-draft.png').specificTargetAlternatives,
+ [],
+ AgentClass.name + ': disjunction text inside one hyphenated filename created alternatives',
+ );
+ const groupedAlternatives = agent._parseWorkflowAttachmentRequirement({
+ value: 'chart.png and logo.png or summary.jpg',
+ });
+ assert.deepEqual(groupedAlternatives.specificTargetAlternatives,
+ [['chart.png', 'logo.png'], ['summary.jpg']],
+ AgentClass.name + ': conjunctive groups inside attachment alternatives were not preserved');
+ assert.equal(
+ agent._workflowSocialPublishedAttachmentObserved(
+ { value: 'chart.png and logo.png or summary.jpg' },
+ {
+ attachments: [
+ { type: 'image', src: 'https://pbs.twimg.com/media/chart.png' },
+ { type: 'image', src: 'https://pbs.twimg.com/media/logo.png' },
+ ],
+ },
+ ),
+ true,
+ AgentClass.name + ': complete conjunctive alternative group was rejected',
+ );
+ assert.equal(
+ agent._workflowSocialPublishedAttachmentObserved(
+ { value: 'chart.png and logo.png or summary.jpg' },
+ { attachments: [{ type: 'image', src: 'https://pbs.twimg.com/media/chart.png' }] },
+ ),
+ false,
+ AgentClass.name + ': incomplete conjunctive alternative group was accepted',
+ );
+ for (const [allowedName, type] of [['chart.png', 'image'], ['clip.mp4', 'video']]) {
+ assert.equal(
+ agent._workflowSocialPublishedAttachmentObserved(
+ { value: 'chart.png or clip.mp4' },
+ { attachments: [{ type, src: `https://cdn.example/${allowedName}` }] },
+ ),
+ true,
+ AgentClass.name + `: cross-type named alternative ${allowedName} was rejected`,
+ );
+ }
+
// Multilingual conjunctions / delimiters in specific targets
for (const conjCase of [
'image1.png, image2.png',