nothumanallowed 13.2.70 → 13.2.72

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "nothumanallowed",
-  "version": "13.2.70",
+  "version": "13.2.72",
   "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;
@@ -2805,6 +2816,10 @@ export async function cmdUI(args) {
       if (pathname === '/api/studio/run' && method === 'POST') {
         const body = await parseBody(req);
         const { agent, task, context, stepDef } = body;
+        const stepPdfBase64 = body.pdfBase64 || null;
+        const stepPdfName = body.pdfName || null;
+        const stepImageBase64 = body.imageBase64 || null;
+        const stepImageMime = body.imageMimeType || 'image/jpeg';
         if (!agent || !task) { sendJSON(res, 400, { error: 'agent and task required' }); logRequest(method, pathname, 400, Date.now() - start); return; }
 
         res.writeHead(200, {
@@ -2836,7 +2851,41 @@ 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 and return it as the step output.
+            // This becomes context for all subsequent steps (WebSearchAgent etc.)
+            sendToken('[Reading attached document...] ');
+            if (stepPdfBase64) {
+              try {
+                const b64 = stepPdfBase64.includes(',') ? stepPdfBase64.split(',')[1] : stepPdfBase64;
+                const pdfBuffer = Buffer.from(b64, 'base64');
+                const extracted = extractTextFromPdf(pdfBuffer);
+                if (extracted && extracted.length > 20) {
+                  toolData = `## Document: ${stepPdfName || 'attached'}\n\n${extracted.slice(0, 20000)}`;
+                } else {
+                  // Fallback: ask vision model to describe/OCR the document
+                  sendToken('[No text found — using vision OCR...] ');
+                  try {
+                    const visionText = await callLLMVision(config, stepPdfBase64, 'application/pdf',
+                      `Extract ALL text, technical specifications, model numbers, part codes, product names, manufacturer details, dimensions, ratings, and any other data from this document. List every detail exactly as printed.`);
+                    toolData = `## Document (OCR): ${stepPdfName || 'attached'}\n\n${visionText}`;
+                  } catch (ve) {
+                    toolData = `Could not extract text from document: ${ve.message}`;
+                  }
+                }
+              } catch (e) { toolData = `Document read failed: ${e.message}`; }
+            } else {
+              toolData = 'No document attached.';
+            }
+            // Stream the extracted content as the step output directly — no LLM rewrite needed
+            sendToken(toolData);
+            clearInterval(keepalive);
+            sendEvent({ done: true, usage: { input: 0, output: Math.ceil(toolData.length / 4) } });
+            res.end();
+            logRequest(method, pathname, 200, Date.now() - start);
+            return;
+
+          } else if (agent === 'EmailAgent') {
             sendToken('[Reading emails...] ');
             try {
               const emails = await withTimeout(getUnreadImportant(config, 10), 'EmailAgent');
@@ -2859,7 +2908,23 @@ export async function cmdUI(args) {
             try {
               // Extract a concise search query from the step prompt
               let searchQuery = stepPrompt;
-              if (searchQuery.length > 120) {
+
+              // If context contains extracted PDF data, extract the best search query from it:
+              // product codes, model numbers, part numbers etc. are better search terms than the task text.
+              if (context && context.length > 50) {
+                // Look for product codes: alphanumeric codes, e.g. 321k63, VFD-001, 4WE6D6X
+                const codeMatch = context.match(/\b([A-Z0-9]{2,}[-\/]?[A-Z0-9]{2,}(?:[-\/][A-Z0-9]+)*)\b/g);
+                const productCodes = codeMatch ? [...new Set(codeMatch)].slice(0, 3) : [];
+                // Also grab manufacturer name if present
+                const mfrMatch = context.match(/(?:Marca|Marchio|Manufacturer|Brand|Produttore|Costruttore)[:\s]+([A-Za-z0-9 &]{2,40})/i);
+                const mfr = mfrMatch ? mfrMatch[1].trim() : '';
+                if (productCodes.length > 0) {
+                  searchQuery = (mfr ? mfr + ' ' : '') + productCodes.join(' ') + ' buy acquista distributore';
+                  sendToken(`[Search query from document: "${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();
@@ -3009,11 +3074,48 @@ RULES:
 - Use .priority-list for action items, .source-item for each email/news source, .bar-row for any percentage data
 - Output must start with <div class="header"> and end with <div class="footer">`;
 
+          // ── Handle PDF/image attachment on first step ─────────────────
+          let attachmentText = '';
+          if (stepPdfBase64 && !isLiveDataAgent) {
+            sendToken('[Reading PDF...] ');
+            try {
+              // Extract base64 payload (strip data URL prefix if present)
+              const b64 = stepPdfBase64.includes(',') ? stepPdfBase64.split(',')[1] : stepPdfBase64;
+              const pdfBuffer = Buffer.from(b64, 'base64');
+              const extracted = extractTextFromPdf(pdfBuffer);
+              if (extracted && extracted.length > 20) {
+                attachmentText = `## ATTACHED PDF: ${stepPdfName || 'document.pdf'}\n${extracted.slice(0, 15000)}`;
+              } else {
+                // PDF has no extractable text (scanned) — use vision
+                try {
+                  const { callLLMVision } = await import('../services/llm.mjs');
+                  const visionResult = await withTimeout(
+                    callLLMVision(config, 'Extract all text and information from this PDF document.', task, { base64: b64, mimeType: 'application/pdf' }),
+                    60000
+                  );
+                  if (visionResult) attachmentText = `## ATTACHED PDF: ${stepPdfName || 'document.pdf'}\n${visionResult.slice(0, 15000)}`;
+                } catch {}
+              }
+            } catch (e) { attachmentText = `[PDF read failed: ${e.message}]`; }
+          } else if (stepImageBase64 && !isLiveDataAgent) {
+            sendToken('[Reading image...] ');
+            try {
+              const { callLLMVision } = await import('../services/llm.mjs');
+              const b64 = stepImageBase64.includes(',') ? stepImageBase64.split(',')[1] : stepImageBase64;
+              const visionResult = await withTimeout(
+                callLLMVision(config, 'Describe and extract all information from this image in detail.', task, { base64: b64, mimeType: stepImageMime }),
+                45000
+              );
+              if (visionResult) attachmentText = `## ATTACHED IMAGE: ${body.imageName || 'image'}\n${visionResult.slice(0, 8000)}`;
+            } catch (e) { attachmentText = `[Image read failed: ${e.message}]`; }
+          }
+
           let sysPrompt, userMsg;
 
           if (isCanvasAgent) {
             sysPrompt = canvasSystemPrompt;
-            userMsg = `Create a professional dashboard report for this data. Output ONLY the inner HTML body content (starting with <div class="header">):\n\n${context}`;
+            const canvasData = [attachmentText, context].filter(Boolean).join('\n\n');
+            userMsg = `Create a professional dashboard report for this data. Output ONLY the inner HTML body content (starting with <div class="header">):\n\n${canvasData}`;
           } else if (isLiveDataAgent) {
             // These agents fetched real data — use a focused prompt (no tool definitions to avoid JSON output)
             const agentInstruction = `You are ${agent}, a specialist AI agent inside NHA Studio. Today is ${today}. Respond entirely in ${language}.
@@ -3025,7 +3127,7 @@ CRITICAL: Do NOT invent, hallucinate, or add any data not present in the DATA se
 Do NOT output JSON, tool calls, or code blocks. Write in plain text with markdown headers.
 Always apply your analysis specifically to the subject mentioned in the WORKFLOW GOAL.
 
-${toolData ? `## DATA FROM TOOLS:\n${toolData}\n` : '## DATA: No data was retrieved by this agent.\n'}
+${attachmentText ? `## ATTACHED FILE CONTENT:\n${attachmentText}\n` : ''}${toolData ? `## DATA FROM TOOLS:\n${toolData}\n` : '## DATA: No data was retrieved by this agent.\n'}
 ${context ? `## OUTPUT FROM PREVIOUS AGENTS:\n${context}\n` : ''}
 
 Your task: ${stepPrompt}`;
@@ -3052,7 +3154,7 @@ CRITICAL RULES:
 - Be thorough and specific — this is for an executive briefing based on REAL data only
 - Always keep the OVERALL WORKFLOW GOAL in mind — apply your analysis specifically to the subject mentioned
 
-${toolData ? `## LIVE DATA FROM TOOLS:\n${toolData}\n` : '## LIVE DATA: No tool data was fetched for this step.\n'}
+${attachmentText ? `## ATTACHED FILE CONTENT:\n${attachmentText}\n` : ''}${toolData ? `## LIVE DATA FROM TOOLS:\n${toolData}\n` : '## LIVE DATA: No tool data was fetched for this step.\n'}
 ${context ? `## OUTPUT FROM PREVIOUS AGENTS:\n${context}\n` : ''}`;
             userMsg = hasRealData
               ? `Based ONLY on the real data above, complete this task specifically for the subject in the WORKFLOW GOAL: ${stepPrompt}`

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 = '13.2.70';
+export const VERSION = '13.2.72';
 export const BASE_URL = 'https://nothumanallowed.com/cli';
 export const API_BASE = 'https://nothumanallowed.com/api/v1';
 

package/src/services/web-ui.mjs

@@ -3371,6 +3371,18 @@ function downloadStudioPDF() {
   var task = studioState.task || 'NHA Studio Report';
   var today = new Date().toLocaleDateString('it-IT', {day:'2-digit',month:'2-digit',year:'numeric'});
   var nodes = studioState.nodes || [];
+  var fileName = (task).slice(0, 60).replace(/[^a-z0-9\s]/gi,'').trim().replace(/\s+/g,'-') || 'NHA-Studio';
+
+  // If canvas exists, download the canvas HTML directly (preserves colors and layout)
+  if (studioState.canvas) {
+    var blob = new Blob([studioState.canvas], {type: 'text/html'});
+    var url = URL.createObjectURL(blob);
+    var a = document.createElement('a');
+    a.href = url; a.target = '_blank'; a.download = fileName + '.html';
+    document.body.appendChild(a); a.click(); document.body.removeChild(a);
+    setTimeout(function(){ URL.revokeObjectURL(url); }, 5000);
+    return;
+  }
 
   // Build sections for each agent
   function mdToPdfHtml(raw) {
@@ -3520,7 +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;
@@ -3780,11 +3798,31 @@ function runStudioStep(idx, node, task, context, stepDef, signal) {
   return new Promise(function(resolve) {
     var output = '';
     var canvasHtml = null;
-    // Inject attachment context into first step (idx===0) if present
-    var effectiveContext = (idx === 0 && studioState.attachmentContext)
-      ? (studioState.attachmentContext + (context ? String.fromCharCode(10) + String.fromCharCode(10) + context : ''))
-      : context;
-    var body = JSON.stringify({stepIdx: idx, agent: node.agent, task: task, context: effectiveContext, stepDef: stepDef});
+    // Inject attachment into first step only — pass PDF/image as dedicated fields,
+    // NOT as raw base64 in context (would cause 100k+ token overflow for any real PDF).
+    var bodyObj = {stepIdx: idx, agent: node.agent, task: task, context: context, stepDef: stepDef};
+    if (idx === 0 && studioState.attachmentContext) {
+      var ac = studioState.attachmentContext;
+      var isPdfAttach = ac.indexOf('[ATTACHED PDF:') === 0;
+      var isImgAttach = ac.indexOf('[ATTACHED IMAGE:') === 0;
+      // Extract base64 data URL from attachment context
+      var b64Match = ac.indexOf('Base64: ');
+      var dataUrl = b64Match >= 0 ? ac.slice(b64Match + 8).trim() : '';
+      if (isPdfAttach && dataUrl) {
+        // Pass PDF as dedicated field — agent/llm handles it natively
+        bodyObj.pdfBase64 = dataUrl;
+        bodyObj.pdfName = studioState.attachmentName;
+        // Add a short note in context instead of the full base64
+        bodyObj.context = '[User attached PDF: ' + studioState.attachmentName + ']' +
+          (context ? String.fromCharCode(10) + String.fromCharCode(10) + context : '');
+      } else if (isImgAttach && dataUrl) {
+        bodyObj.imageBase64 = dataUrl;
+        bodyObj.imageMimeType = dataUrl.indexOf('data:image/png') === 0 ? 'image/png' : 'image/jpeg';
+        bodyObj.context = '[User attached image: ' + studioState.attachmentName + ']' +
+          (context ? String.fromCharCode(10) + String.fromCharCode(10) + context : '');
+      }
+    }
+    var body = JSON.stringify(bodyObj);
     var fetchOpts = {method: 'POST', headers: {'Content-Type': 'application/json'}, body: body};
     if (signal) fetchOpts.signal = signal;