Open-source AI browser agent for Chrome and Firefox (monorepo) 🧠
JavaScript ai-agentai-agentsai-toolsbrowser-aibrowser-automationbrowser-extensionchrome-aichrome-devtoolschrome-extensionchrome-extensionsfirefox-addonfirefox-addonsfirefox-extensionfirefox-extensionsgemmagemma4minimaxoffline-aioffline-llmqwen
fix(agent): match social destination across full masked clause and skip incidental command colons
Emre Sokullu9db7d8fbbb02
diff --git a/src/chrome/src/agent/agent.js b/src/chrome/src/agent/agent.jsindex 7c1551304..9a6082465 100644--- a/src/chrome/src/agent/agent.js+++ b/src/chrome/src/agent/agent.js@@ -15369,14 +15369,15 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout dif (SOCIAL_POST_NEGATION.test(afterVerbText.trim())) continue;if (SOCIAL_NOUN_LIKE_PUBLISH.test(verbWord) && SOCIAL_READ_VERBS.test(afterVerbText)) 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 = clause.text.slice(verbIndex, verbIndex + verbWord.length + 60);- const beforeVerb = clause.text.slice(Math.max(0, verbIndex - 60), verbIndex);+ const afterVerb = targetText.slice(verbIndex + verbWord.length);+ const beforeVerb = targetText.slice(0, verbIndex);platformPattern.lastIndex = 0;if (platformPattern.test(afterVerb) || platformPattern.test(beforeVerb)) return true;if (clauseIdx > 0) {const prevClause = clauses[clauseIdx - 1];+ const prevTargetText = prevClause.maskedText || prevClause.text;platformPattern.lastIndex = 0;- if (platformPattern.test(prevClause.text) && !SOCIAL_READ_VERBS.test(prevClause.text) && !prevClause.isNegated) {+ if (platformPattern.test(prevTargetText) && !SOCIAL_READ_VERBS.test(prevTargetText) && !prevClause.isNegated) {return true;}}@@ -22424,20 +22425,98 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout dconst colonPattern = new RegExp(`${publishVerbPattern}[\\s\\S]*?(?<!https?|ftp|sftp)[::](?!//)\\s*([\\s\\S]+)$`, 'iu');for (const text of candidates) {let body = '';- const colonMatch = text.match(colonPattern);+ // Scan for the colon delimiter that introduces the post body.+ // We look for colons occurring after a publish verb, skipping:+ // 1. URL schemes (e.g. "https://", "ftp://")+ // 2. Colons inside parentheses, brackets, or quotes (e.g. "(account: @acme):")+ // 3. Incidental metadata key labels (e.g. "account: @acme")+ const verbMatches = [...text.matchAll(new RegExp(publishVerbPattern, 'giu'))];let colonBody = '';let colonDelimIndex = -1;- if (colonMatch && colonMatch[1]?.trim()) {- let candidate = colonMatch[1].trim();- const innerQuote = candidate.match(/^(?:“([\s\S]+)”|「([\s\S]+)」|『([\s\S]+)』|«([\s\S]+)»|"([\s\S]+)"|'([\s\S]+)')$/);- const innerContent = innerQuote ? (innerQuote[1] || innerQuote[2] || innerQuote[3] || innerQuote[4] || innerQuote[5] || innerQuote[6]) : null;- if (innerContent && innerContent.trim()) {- candidate = innerContent.trim();- }- colonBody = candidate;- const prefix = colonMatch[0].slice(0, colonMatch[0].length - colonMatch[1].length);- const matchDelim = prefix.search(/[::][^\S\r\n]*$/);- colonDelimIndex = colonMatch.index + (matchDelim >= 0 ? matchDelim : prefix.length - 1);+ for (const verbMatch of verbMatches) {+ const verbStart = verbMatch.index ?? 0;+ const verbEnd = verbStart + (verbMatch[0] || '').length;+ let parenDepth = 0;+ let bracketDepth = 0;+ let braceDepth = 0;+ let inDoubleQuote = false;+ let inSingleQuote = false;+ let inSmartQuote = 0;+ let inCjkQuote = 0;+ let inGuillemet = 0;++ for (let i = verbEnd; i < text.length; i++) {+ const ch = text[i];+ const prevChar = i > 0 ? text[i - 1] : '';+ const nextChar = i + 1 < text.length ? text[i + 1] : '';++ if (ch === '(' || ch === '\uff08') parenDepth++;+ else if (ch === ')' || ch === '\uff09') parenDepth = Math.max(0, parenDepth - 1);+ else if (ch === '[' || ch === '\u3010' || ch === '\uff3b') bracketDepth++;+ else if (ch === ']' || ch === '\u3011' || ch === '\uff3d') bracketDepth = Math.max(0, bracketDepth - 1);+ else if (ch === '{' || ch === '\uff5b') braceDepth++;+ else if (ch === '}' || ch === '\uff5d') braceDepth = Math.max(0, braceDepth - 1);+ else if (ch === '"') inDoubleQuote = !inDoubleQuote;+ else if (ch === '“') inSmartQuote++;+ else if (ch === '”') inSmartQuote = Math.max(0, inSmartQuote - 1);+ else if (ch === '「' || ch === '『') inCjkQuote++;+ else if (ch === '」' || ch === '』') inCjkQuote = Math.max(0, inCjkQuote - 1);+ else if (ch === '«') inGuillemet++;+ else if (ch === '»') inGuillemet = Math.max(0, inGuillemet - 1);+ else if (ch === "'") {+ const isLetterBefore = /\p{L}/u.test(prevChar);+ const isLetterAfter = /\p{L}/u.test(nextChar);+ if (!inSingleQuote && !isLetterBefore && isLetterAfter) {+ inSingleQuote = true;+ } else if (inSingleQuote && isLetterBefore && !isLetterAfter) {+ inSingleQuote = false;+ }+ }++ if (ch === ':' || ch === '\uff1a') {+ if (text.slice(i, i + 3) === '://' || text.slice(i, i + 2) === ':\u2044\u2044') continue;+ if (/(?:https?|ftp|sftp|file|mailto)$/i.test(text.slice(Math.max(0, i - 10), i))) continue;++ if (parenDepth > 0 || bracketDepth > 0 || braceDepth > 0+ || inDoubleQuote || inSingleQuote || inSmartQuote > 0 || inCjkQuote > 0 || inGuillemet > 0) {+ continue;+ }++ const beforeColon = text.slice(verbEnd, i);+ if (/(?<![\p{L}\p{N}_])(?:account|user|username|handle|profile|channel|destination|target|via|date|time|tag|tags|label|media|image|video|cuenta|compte|benutzer|kullanıcı|hesap|аккаунт|пользователь|账号|帳號|ユーザー|アカウント|계정)\s*$/iu.test(beforeColon)) {+ continue;+ }++ const candidate = text.slice(i + 1).trim();+ if (candidate) {+ colonDelimIndex = i;+ colonBody = candidate;+ const innerQuote = candidate.match(/^(?:“([\s\S]+)”|「([\s\S]+)」|『([\s\S]+)』|«([\s\S]+)»|"([\s\S]+)"|'([\s\S]+)')$/);+ const innerContent = innerQuote ? (innerQuote[1] || innerQuote[2] || innerQuote[3] || innerQuote[4] || innerQuote[5] || innerQuote[6]) : null;+ if (innerContent && innerContent.trim()) {+ colonBody = innerContent.trim();+ }+ break;+ }+ }+ }+ if (colonBody) break;+ }++ if (!colonBody) {+ const colonMatch = text.match(colonPattern);+ if (colonMatch && colonMatch[1]?.trim()) {+ let candidate = colonMatch[1].trim();+ const innerQuote = candidate.match(/^(?:“([\s\S]+)”|「([\s\S]+)」|『([\s\S]+)』|«([\s\S]+)»|"([\s\S]+)"|'([\s\S]+)')$/);+ const innerContent = innerQuote ? (innerQuote[1] || innerQuote[2] || innerQuote[3] || innerQuote[4] || innerQuote[5] || innerQuote[6]) : null;+ if (innerContent && innerContent.trim()) {+ candidate = innerContent.trim();+ }+ colonBody = candidate;+ const prefix = colonMatch[0].slice(0, colonMatch[0].length - colonMatch[1].length);+ const matchDelim = prefix.search(/[::][^\S\r\n]*$/);+ colonDelimIndex = colonMatch.index + (matchDelim >= 0 ? matchDelim : prefix.length - 1);+ }}const quotedMatch = text.match(quotedPattern);diff --git a/src/firefox/src/agent/agent.js b/src/firefox/src/agent/agent.jsindex a23fd584d..ce30559ff 100644--- a/src/firefox/src/agent/agent.js+++ b/src/firefox/src/agent/agent.js@@ -13169,14 +13169,15 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout dif (SOCIAL_POST_NEGATION.test(afterVerbText.trim())) continue;if (SOCIAL_NOUN_LIKE_PUBLISH.test(verbWord) && SOCIAL_READ_VERBS.test(afterVerbText)) continue;if (/^shares?$/i.test(verbWord) && /(?<![\p{L}\p{N}_])(?:market|mind|revenue|profit|traffic|audience|wallet|fair|lion's|stock|equity|file|screen|time)\s+$/iu.test(beforeVerbClause)) continue;- const afterVerb = clause.text.slice(verbIndex, verbIndex + verbWord.length + 60);- const beforeVerb = clause.text.slice(Math.max(0, verbIndex - 60), verbIndex);+ const afterVerb = targetText.slice(verbIndex + verbWord.length);+ const beforeVerb = targetText.slice(0, verbIndex);platformPattern.lastIndex = 0;if (platformPattern.test(afterVerb) || platformPattern.test(beforeVerb)) return true;if (clauseIdx > 0) {const prevClause = clauses[clauseIdx - 1];+ const prevTargetText = prevClause.maskedText || prevClause.text;platformPattern.lastIndex = 0;- if (platformPattern.test(prevClause.text) && !SOCIAL_READ_VERBS.test(prevClause.text) && !prevClause.isNegated) {+ if (platformPattern.test(prevTargetText) && !SOCIAL_READ_VERBS.test(prevTargetText) && !prevClause.isNegated) {return true;}}@@ -20027,20 +20028,98 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout dconst colonPattern = new RegExp(`${publishVerbPattern}[\\s\\S]*?(?<!https?|ftp|sftp)[::](?!//)\\s*([\\s\\S]+)$`, 'iu');for (const text of candidates) {let body = '';- const colonMatch = text.match(colonPattern);+ // Scan for the colon delimiter that introduces the post body.+ // We look for colons occurring after a publish verb, skipping:+ // 1. URL schemes (e.g. "https://", "ftp://")+ // 2. Colons inside parentheses, brackets, or quotes (e.g. "(account: @acme):")+ // 3. Incidental metadata key labels (e.g. "account: @acme")+ const verbMatches = [...text.matchAll(new RegExp(publishVerbPattern, 'giu'))];let colonBody = '';let colonDelimIndex = -1;- if (colonMatch && colonMatch[1]?.trim()) {- let candidate = colonMatch[1].trim();- const innerQuote = candidate.match(/^(?:“([\s\S]+)”|「([\s\S]+)」|『([\s\S]+)』|«([\s\S]+)»|"([\s\S]+)"|'([\s\S]+)')$/);- const innerContent = innerQuote ? (innerQuote[1] || innerQuote[2] || innerQuote[3] || innerQuote[4] || innerQuote[5] || innerQuote[6]) : null;- if (innerContent && innerContent.trim()) {- candidate = innerContent.trim();- }- colonBody = candidate;- const prefix = colonMatch[0].slice(0, colonMatch[0].length - colonMatch[1].length);- const matchDelim = prefix.search(/[::][^\S\r\n]*$/);- colonDelimIndex = colonMatch.index + (matchDelim >= 0 ? matchDelim : prefix.length - 1);+ for (const verbMatch of verbMatches) {+ const verbStart = verbMatch.index ?? 0;+ const verbEnd = verbStart + (verbMatch[0] || '').length;+ let parenDepth = 0;+ let bracketDepth = 0;+ let braceDepth = 0;+ let inDoubleQuote = false;+ let inSingleQuote = false;+ let inSmartQuote = 0;+ let inCjkQuote = 0;+ let inGuillemet = 0;++ for (let i = verbEnd; i < text.length; i++) {+ const ch = text[i];+ const prevChar = i > 0 ? text[i - 1] : '';+ const nextChar = i + 1 < text.length ? text[i + 1] : '';++ if (ch === '(' || ch === '\uff08') parenDepth++;+ else if (ch === ')' || ch === '\uff09') parenDepth = Math.max(0, parenDepth - 1);+ else if (ch === '[' || ch === '\u3010' || ch === '\uff3b') bracketDepth++;+ else if (ch === ']' || ch === '\u3011' || ch === '\uff3d') bracketDepth = Math.max(0, bracketDepth - 1);+ else if (ch === '{' || ch === '\uff5b') braceDepth++;+ else if (ch === '}' || ch === '\uff5d') braceDepth = Math.max(0, braceDepth - 1);+ else if (ch === '"') inDoubleQuote = !inDoubleQuote;+ else if (ch === '“') inSmartQuote++;+ else if (ch === '”') inSmartQuote = Math.max(0, inSmartQuote - 1);+ else if (ch === '「' || ch === '『') inCjkQuote++;+ else if (ch === '」' || ch === '』') inCjkQuote = Math.max(0, inCjkQuote - 1);+ else if (ch === '«') inGuillemet++;+ else if (ch === '»') inGuillemet = Math.max(0, inGuillemet - 1);+ else if (ch === "'") {+ const isLetterBefore = /\p{L}/u.test(prevChar);+ const isLetterAfter = /\p{L}/u.test(nextChar);+ if (!inSingleQuote && !isLetterBefore && isLetterAfter) {+ inSingleQuote = true;+ } else if (inSingleQuote && isLetterBefore && !isLetterAfter) {+ inSingleQuote = false;+ }+ }++ if (ch === ':' || ch === '\uff1a') {+ if (text.slice(i, i + 3) === '://' || text.slice(i, i + 2) === ':\u2044\u2044') continue;+ if (/(?:https?|ftp|sftp|file|mailto)$/i.test(text.slice(Math.max(0, i - 10), i))) continue;++ if (parenDepth > 0 || bracketDepth > 0 || braceDepth > 0+ || inDoubleQuote || inSingleQuote || inSmartQuote > 0 || inCjkQuote > 0 || inGuillemet > 0) {+ continue;+ }++ const beforeColon = text.slice(verbEnd, i);+ if (/(?<![\p{L}\p{N}_])(?:account|user|username|handle|profile|channel|destination|target|via|date|time|tag|tags|label|media|image|video|cuenta|compte|benutzer|kullanıcı|hesap|аккаунт|пользователь|账号|帳號|ユーザー|アカウント|계정)\s*$/iu.test(beforeColon)) {+ continue;+ }++ const candidate = text.slice(i + 1).trim();+ if (candidate) {+ colonDelimIndex = i;+ colonBody = candidate;+ const innerQuote = candidate.match(/^(?:“([\s\S]+)”|「([\s\S]+)」|『([\s\S]+)』|«([\s\S]+)»|"([\s\S]+)"|'([\s\S]+)')$/);+ const innerContent = innerQuote ? (innerQuote[1] || innerQuote[2] || innerQuote[3] || innerQuote[4] || innerQuote[5] || innerQuote[6]) : null;+ if (innerContent && innerContent.trim()) {+ colonBody = innerContent.trim();+ }+ break;+ }+ }+ }+ if (colonBody) break;+ }++ if (!colonBody) {+ const colonMatch = text.match(colonPattern);+ if (colonMatch && colonMatch[1]?.trim()) {+ let candidate = colonMatch[1].trim();+ const innerQuote = candidate.match(/^(?:“([\s\S]+)”|「([\s\S]+)」|『([\s\S]+)』|«([\s\S]+)»|"([\s\S]+)"|'([\s\S]+)')$/);+ const innerContent = innerQuote ? (innerQuote[1] || innerQuote[2] || innerQuote[3] || innerQuote[4] || innerQuote[5] || innerQuote[6]) : null;+ if (innerContent && innerContent.trim()) {+ candidate = innerContent.trim();+ }+ colonBody = candidate;+ const prefix = colonMatch[0].slice(0, colonMatch[0].length - colonMatch[1].length);+ const matchDelim = prefix.search(/[::][^\S\r\n]*$/);+ colonDelimIndex = colonMatch.index + (matchDelim >= 0 ? matchDelim : prefix.length - 1);+ }}const quotedMatch = text.match(quotedPattern);diff --git a/test/run.js b/test/run.jsindex 6ef5412b0..86bf74b76 100644--- a/test/run.js+++ b/test/run.js@@ -86332,6 +86332,12 @@ test('social publication workflow follows the live X or Bluesky destination and'without attachments in post metadata was wrongly treated as publish negation'],['Without delay, post on X', ['twitter'],'without delay was wrongly treated as publish negation'],+ ['Post "This is a very long announcement body that contains far more than fifty or sixty characters and explains everything clearly" on X', ['twitter'],+ 'long quoted body before on X was cut off by character window'],+ ['Post "This is another long announcement body with more than sixty characters" to Bluesky', ['bluesky'],+ 'long quoted body before to Bluesky was cut off by character window'],+ ['Post "Check our thoughts on X" to Bluesky', ['bluesky'],+ 'platform mentioned inside quoted body leaked into target detection'],]) {assert.deepEqual([...agent._trustedSocialPublishTargetAdapters({ taskText })],@@ -87888,6 +87894,42 @@ test('post body extraction prefers colon-introduced body when colon precedes inc}});+test('post body extraction skips incidental command colons and parenthesized metadata', () => {+ for (const AgentClass of [AgentCh, AgentFx]) {+ const agent = new AgentClass({});+ assert.equal(+ agent._extractWorkflowTaskBody('Post on X (account: @acme): Hello world'),+ 'Hello world',+ AgentClass.name + ': parenthesized account metadata with colon should not be captured as body'+ );+ assert.equal(+ agent._extractWorkflowTaskBody('Post on X [account: @acme]: Hello world'),+ 'Hello world',+ AgentClass.name + ': bracketed account metadata with colon should not be captured as body'+ );+ assert.equal(+ agent._extractWorkflowTaskBody('Post on X (account: @acme): Here is a long announcement containing "important news" and more updates today'),+ 'Here is a long announcement containing "important news" and more updates today',+ AgentClass.name + ': parenthesized metadata with colon preceding inner quotes should extract full body'+ );+ assert.equal(+ agent._extractWorkflowTaskBody('Post on X, account: @acme: Hello world'),+ 'Hello world',+ AgentClass.name + ': metadata key prefix with colon should be skipped in favor of payload colon'+ );+ assert.equal(+ agent._extractWorkflowTaskBody('Xに投稿(アカウント:@acme):こんにちは世界'),+ 'こんにちは世界',+ AgentClass.name + ': Japanese full-width parenthesized account with colon should not be captured as body'+ );+ assert.equal(+ agent._extractWorkflowTaskBody('Post on X (account: @acme): Breaking: Version 2 is released'),+ 'Breaking: Version 2 is released',+ AgentClass.name + ': body containing colon should preserve colon after payload delimiter'+ );+ }+});+test('post body extraction preserves nested quotation marks and does not overwrite full body with prefix', async () => {for (const [index, AgentClass] of [AgentCh, AgentFx].entries()) {const agent = new AgentClass({});