nothumanallowed 13.2.71 → 13.2.73

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "nothumanallowed",
-  "version": "13.2.71",
+  "version": "13.2.73",
   "description": "NotHumanAllowed — 38 AI agents, 80 tools, Studio (visual agentic workflows). Email, calendar, browser automation, screen capture, canvas, cron/heartbeat, Alexandria E2E messaging, GitHub, Notion, Slack, voice chat, free AI (Liara), 28 languages. Zero-dependency CLI.",
   "type": "module",
   "bin": {

package/src/commands/ui.mjs

@@ -2692,9 +2692,10 @@ export async function cmdUI(args) {
 
         // ── Fast keyword-based planning (no LLM call needed for common patterns) ──────
         const taskLow = task.toLowerCase();
+        const hasPdf        = !!(body.hasPdf) || /pdf|allegat|catalogo|scheda\s*tecnic|document/i.test(taskLow);
         const hasEmail      = /email|mail|inbox|posta/i.test(taskLow);
         const hasCalendar   = /calendar|agenda|calendari|eventi|schedule/i.test(taskLow);
-        const hasSearch     = /cerca|search|notizie|news|ultime|latest|web|internet|tendenz|trend/i.test(taskLow);
+        const hasSearch     = /cerca|search|notizie|news|ultime|latest|web|internet|tendenz|trend|acquista|compra|dove\s+trovare|where\s+to\s+buy|similar|simile/i.test(taskLow);
         const hasCanvas     = /html|dashboard|visua|report|grafico|chart/i.test(taskLow);
         const hasGitHub     = /github|git|issue|pr|pull request/i.test(taskLow);
         const hasSlack      = /slack|channel|messag/i.test(taskLow);
@@ -2726,13 +2727,23 @@ export async function cmdUI(args) {
         // Build plan directly from keywords — reliable, fast, no SENTINEL risk
         const buildKeywordPlan = () => {
           const steps = [];
+          // PDF attachment: always read document first to extract specs/data before any web search
+          if (hasPdf) {
+            const pdfName = body.pdfName || 'documento allegato';
+            steps.push({icon:'\u{1F4C4}',agent:'DocumentReaderAgent',label:it?'Leggi documento':'Read document',prompt:`Extract all technical specifications, model numbers, part codes, product names, manufacturer, dimensions, ratings, and any other key data from the attached document "${pdfName}". List every technical detail precisely.`});
+          }
           if (hasEmail)    steps.push({icon:'\u{1F4E7}',agent:'EmailAgent',   label:it?'Controlla email':'Check emails',       prompt:'Read the latest unread emails and identify urgent items, deadlines, and required actions'});
           if (hasCalendar) steps.push({icon:'\u{1F4C5}',agent:'CalendarAgent', label:it?'Rivedi calendario':'Review calendar',   prompt:'Check today\'s events and identify any scheduling conflicts or important meetings'});
           if (hasGitHub)   steps.push({icon:'\u{1F4BB}',agent:'GitHubAgent',   label:'GitHub',                                  prompt:'Read open issues and pull requests, identify what needs attention'});
           if (hasSlack)    steps.push({icon:'\u{1F4AC}',agent:'SlackAgent',    label:'Slack',                                   prompt:'Check recent Slack messages and identify important conversations'});
           if (hasNotion)   steps.push({icon:'\u{1F4DD}',agent:'NotionAgent',   label:'Notion',                                  prompt:'Search Notion for relevant pages and notes'});
-          if (hasSearch || hasReputation || (!hasEmail && !hasCalendar && !hasGitHub && !hasSlack)) {
-            steps.push({icon:'\u{1F50D}',agent:'WebSearchAgent',label:it?'Ricerca web':'Web search',prompt:searchQuery});
+          // When PDF is present: always search web (to find where to buy, similar products etc.)
+          // The search query will be refined at runtime using the extracted PDF specs as context
+          if (hasPdf || hasSearch || hasReputation || (!hasEmail && !hasCalendar && !hasGitHub && !hasSlack)) {
+            const searchPrompt = hasPdf
+              ? (it ? 'Usando le specifiche tecniche estratte dal documento (codice prodotto, modello, costruttore, caratteristiche), cerca online dove acquistare il prodotto o articoli equivalenti. Usa i codici esatti dal documento come query di ricerca.' : 'Using the technical specifications extracted from the document (product code, model, manufacturer, specs), search online for where to buy this product or equivalent alternatives. Use exact codes from the document as search queries.')
+              : searchQuery;
+            steps.push({icon:'\u{1F50D}',agent:'WebSearchAgent',label:it?'Ricerca web':'Web search',prompt:searchPrompt});
           }
           // Specialist agents — can stack multiple
           if (hasSecurity)   steps.push({icon:'\u{1F6E1}',agent:'cassandra',   label:it?'CASSANDRA \u2014 Rischi sicurezza':'CASSANDRA \u2014 Security risks',    prompt:'Analyze the collected data and identify security risks, vulnerabilities and concrete recommendations'});
@@ -2758,7 +2769,7 @@ export async function cmdUI(args) {
 
         // Use keyword plan directly — only fall back to LLM for genuinely ambiguous tasks
         const keywordSteps = buildKeywordPlan();
-        const taskIsComplex = !hasEmail && !hasCalendar && !hasSearch && !hasGitHub && !hasSlack && !hasBriefing && !hasStrategy && !hasReputation && !hasCode && !hasWriting && !hasData && keywordSteps.length <= 1;
+        const taskIsComplex = !hasPdf && !hasEmail && !hasCalendar && !hasSearch && !hasGitHub && !hasSlack && !hasBriefing && !hasStrategy && !hasReputation && !hasCode && !hasWriting && !hasData && keywordSteps.length <= 1;
 
         try {
           let steps;
@@ -2840,7 +2851,79 @@ export async function cmdUI(args) {
           let toolData = '';
 
           // ── Fetch REAL data for each agent type ──────────────────────
-          if (agent === 'EmailAgent') {
+          if (agent === 'DocumentReaderAgent') {
+            // Extract text from attached PDF, then ask the LLM to structure it cleanly.
+            // The structured output becomes context for all subsequent steps.
+            sendToken('[Reading attached document...] ');
+            let rawText = '';
+            if (stepPdfBase64) {
+              try {
+                const b64 = stepPdfBase64.includes(',') ? stepPdfBase64.split(',')[1] : stepPdfBase64;
+                const pdfBuffer = Buffer.from(b64, 'base64');
+                rawText = extractTextFromPdf(pdfBuffer) || '';
+                if (!rawText || rawText.length < 20) {
+                  // Scanned PDF — use vision OCR
+                  sendToken('[No text layer — using vision OCR...] ');
+                  try {
+                    rawText = await callLLMVision(config, stepPdfBase64, 'application/pdf',
+                      'Extract ALL text from this document exactly as printed, preserving all numbers, codes, and values.');
+                  } catch (ve) { rawText = ''; }
+                }
+              } catch (e) { rawText = ''; }
+            }
+            if (!rawText) {
+              sendToken('Could not extract text from the attached document.');
+              clearInterval(keepalive);
+              sendEvent({ done: true, usage: { input: 0, output: 0 } });
+              res.end();
+              logRequest(method, pathname, 200, Date.now() - start);
+              return;
+            }
+            // Ask LLM to structure the raw extracted text into readable markdown
+            sendToken('[Structuring document content...] ');
+            const LANG_MAP_DOC = {en:'English',it:'Italian',es:'Spanish',fr:'French',de:'German',pt:'Portuguese',zh:'Chinese',ja:'Japanese',ar:'Arabic',hi:'Hindi',ru:'Russian',nl:'Dutch',pl:'Polish',tr:'Turkish',ko:'Korean',sv:'Swedish',da:'Danish',fi:'Finnish',no:'Norwegian',cs:'Czech'};
+            const docLang = LANG_MAP_DOC[(config?.language||'it').toLowerCase().slice(0,2)] || 'Italian';
+            const docSys = `You are a technical document analyst. Extract and structure the content of this document into clear, readable markdown. Respond in ${docLang}.
+Rules:
+- List ALL technical specifications with their exact values (codes, voltages, pressures, temperatures, dimensions, flow rates, etc.)
+- Use markdown headers (##), bullet points (-), and tables where appropriate
+- Do NOT invent, interpret, or add anything not present in the raw text
+- Include all product/part codes exactly as written
+- Keep all numeric values with their units`;
+            const docUser = `Here is the raw text extracted from "${stepPdfName || 'document.pdf'}". Structure it into clean, readable markdown:\n\n${rawText.slice(0, 18000)}`;
+            let structuredOutput = '';
+            let inThink = false;
+            try {
+              await withTimeout(
+                callLLMStream(config, docSys, docUser,
+                  (token) => {
+                    // Strip <think> blocks
+                    let buf = token;
+                    if (inThink) {
+                      const ci = buf.indexOf('</think>');
+                      if (ci >= 0) { buf = buf.slice(ci + 8); inThink = false; }
+                      else return;
+                    }
+                    const oi = buf.indexOf('<think>');
+                    if (oi >= 0) { buf = buf.slice(0, oi); inThink = true; }
+                    if (buf) { structuredOutput += buf; sendToken(buf); }
+                  },
+                  { max_tokens: 3000 }
+                ),
+                90000
+              );
+            } catch (e) {
+              // LLM failed — fall back to raw text
+              structuredOutput = `## ${stepPdfName || 'Document'}\n\n${rawText.slice(0, 8000)}`;
+              sendToken(structuredOutput);
+            }
+            clearInterval(keepalive);
+            sendEvent({ done: true, usage: { input: Math.ceil(rawText.length / 4), output: Math.ceil(structuredOutput.length / 4) } });
+            res.end();
+            logRequest(method, pathname, 200, Date.now() - start);
+            return;
+
+          } else if (agent === 'EmailAgent') {
             sendToken('[Reading emails...] ');
             try {
               const emails = await withTimeout(getUnreadImportant(config, 10), 'EmailAgent');
@@ -2863,7 +2946,23 @@ export async function cmdUI(args) {
             try {
               // Extract a concise search query from the step prompt
               let searchQuery = stepPrompt;
-              if (searchQuery.length > 120) {
+
+              // If context contains extracted PDF data, extract the best search query from it:
+              // product codes, model numbers, part numbers etc. are better search terms than the task text.
+              if (context && context.length > 50) {
+                // Look for product codes: alphanumeric codes, e.g. 321k63, VFD-001, 4WE6D6X
+                const codeMatch = context.match(/\b([A-Z0-9]{2,}[-\/]?[A-Z0-9]{2,}(?:[-\/][A-Z0-9]+)*)\b/g);
+                const productCodes = codeMatch ? [...new Set(codeMatch)].slice(0, 3) : [];
+                // Also grab manufacturer name if present
+                const mfrMatch = context.match(/(?:Marca|Marchio|Manufacturer|Brand|Produttore|Costruttore)[:\s]+([A-Za-z0-9 &]{2,40})/i);
+                const mfr = mfrMatch ? mfrMatch[1].trim() : '';
+                if (productCodes.length > 0) {
+                  searchQuery = (mfr ? mfr + ' ' : '') + productCodes.join(' ') + ' buy acquista distributore';
+                  sendToken(`[Search query from document: "${searchQuery}"] `);
+                }
+              }
+
+              if (searchQuery.length > 120 && !context) {
                 const keywordMatch = searchQuery.match(/(?:cerca|search|find|ricerca|notizie su|news about|latest on|aggiornamenti su)\s+(.{5,80}?)(?:\s+(?:e|and|per|for|poi|then)|$)/i);
                 if (keywordMatch) {
                   searchQuery = keywordMatch[1].trim();

package/src/constants.mjs

@@ -5,7 +5,7 @@ import { fileURLToPath } from 'url';
 const __filename = fileURLToPath(import.meta.url);
 const __dirname = path.dirname(__filename);
 
-export const VERSION = '13.2.71';
+export const VERSION = '13.2.73';
 export const BASE_URL = 'https://nothumanallowed.com/cli';
 export const API_BASE = 'https://nothumanallowed.com/api/v1';
 

package/src/services/web-ui.mjs

@@ -3532,7 +3532,13 @@ async function runStudio() {
     var taskForPlan = studioState.attachmentContext
       ? task + nl + nl + '[User has attached a file: ' + studioState.attachmentName + '. Agents will receive the full content.]'
       : task;
-    var planRes = await apiPost('/api/studio/plan', {task: taskForPlan});
+    // Include PDF info in plan request so server can add DocumentReaderAgent step
+    var planBody = {task: taskForPlan};
+    if (studioState.attachmentContext && studioState.attachmentContext.indexOf('[ATTACHED PDF:') === 0) {
+      planBody.hasPdf = true;
+      planBody.pdfName = studioState.attachmentName || 'document.pdf';
+    }
+    var planRes = await apiPost('/api/studio/plan', planBody);
     if (!planRes || !planRes.steps || !planRes.steps.length) {
       studioLog('Studio', '⚠', 'Could not plan workflow. Check your LLM provider config.', 'error');
       studioState.running = false;