nothumanallowed 15.1.51 → 15.1.52

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "nothumanallowed",
-  "version": "15.1.51",
+  "version": "15.1.52",
   "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/constants.mjs

@@ -5,7 +5,7 @@ import { fileURLToPath } from 'url';
 const __filename = fileURLToPath(import.meta.url);
 const __dirname = path.dirname(__filename);
 
-export const VERSION = '15.1.51';
+export const VERSION = '15.1.52';
 export const BASE_URL = 'https://nothumanallowed.com/cli';
 export const API_BASE = 'https://nothumanallowed.com/api/v1';
 

package/src/server/routes/chat.mjs

@@ -17,6 +17,7 @@ import {
 import { callLLMStream, callLLM, callLLMVision, parseAgentFile } from '../../services/llm.mjs';
 import { buildMemoryContext } from '../../services/memory.mjs';
 import { parseActions, executeTool, buildSystemPrompt, stripOrphanFences } from '../../services/tool-executor.mjs';
+import { detectLanguage } from '../../services/message-responder.mjs';
 
 // Migrate on import (once)
 migrateOldHistory();
@@ -200,11 +201,17 @@ export function register(router) {
     try { const m = buildMemoryContext('chat', effectiveMsg); if (m) enrichedPrompt = enrichedPrompt + m; } catch {}
     try { const ic = await getImapAccountsContext(); if (ic) enrichedPrompt += ic; } catch {}
 
-    // Inject language instruction — always respects user's lang setting
+    // Inject language instruction — detect from the user's message first;
+    // fall back to config.language only when detection is inconclusive (very
+    // short messages, codes, etc.). Detected language is authoritative
+    // because users routinely interact in their natural language regardless
+    // of the dropdown setting.
     const LANG_MAP = { it:'Italian', en:'English', es:'Spanish', fr:'French', de:'German', pt:'Portuguese', nl:'Dutch', pl:'Polish', ru:'Russian', zh:'Chinese', ja:'Japanese', ko:'Korean', ar:'Arabic', hi:'Hindi', tr:'Turkish', sv:'Swedish', da:'Danish', fi:'Finnish', cs:'Czech' };
-    const userLang = LANG_MAP[(config?.language || config?.lang || 'en').slice(0,2)] || 'English';
+    const settingLang = LANG_MAP[(config?.language || config?.lang || 'en').slice(0,2)] || 'English';
+    const detectedFromMsg = detectLanguage(effectiveMsg);
+    const userLang = detectedFromMsg || settingLang;
     if (!enrichedPrompt.toLowerCase().includes('respond in') && !enrichedPrompt.toLowerCase().includes('rispondi in')) {
-      enrichedPrompt += `\n\nIMPORTANT: Always respond in ${userLang}.`;
+      enrichedPrompt += `\n\nIMPORTANT: The user just wrote their message in ${userLang}. Respond in ${userLang} — match the user's language exactly, do not switch to a different language mid-response.`;
     }
 
     // Rolling context window
@@ -521,7 +528,7 @@ export function register(router) {
           }
         };
         const toolContext = toolResults.map(t => `[${t.action} result]:\n${cleanResult(t.action, t.result)}`).join('\n\n---\n\n');
-        const synthesisPrompt = `${enrichedPrompt}\n\n## DATA FROM TOOLS (already executed — do NOT plan to call them again):\n${toolContext}\n\n## STRICT OUTPUT RULES:\n- The tools above HAVE ALREADY RUN. Their output is the ground truth.\n- NEVER say "Let me read…", "I'll search…", "I will fetch…", "leggerò", "cercherò", "ti dirò" — those tools are already done.\n- Answer the user's question DIRECTLY using the data above. If a specific detail (delivery date, item, total, etc.) is in the data, quote it verbatim.\n- If the data does not contain the answer, state plainly what is missing — do not announce intent to look further.\n- Write ONLY plain prose or markdown (headers, bullets, bold). NEVER use \`\`\`json or any fenced code block.\n- Format numbers/prices as plain text. Be concise and human-readable.`;
+        const synthesisPrompt = `${enrichedPrompt}\n\n## DATA FROM TOOLS (already executed — do NOT plan to call them again):\n${toolContext}\n\n## STRICT OUTPUT RULES:\n- LANGUAGE: respond ENTIRELY in ${userLang}. Do not switch languages mid-answer, do not mix English and ${userLang}.\n- The tools above HAVE ALREADY RUN. Their output is the ground truth.\n- NEVER say "Let me read…", "I'll search…", "I will fetch…", "leggerò", "cercherò", "ti dirò" — those tools are already done.\n- Answer the user's question DIRECTLY using the data above. If a specific detail (delivery date, item, total, etc.) is in the data, quote it verbatim.\n- If the data does not contain the answer, state plainly what is missing — do not announce intent to look further.\n- Write ONLY plain prose or markdown (headers, bullets, bold). NEVER use \`\`\`json or any fenced code block.\n- Format numbers/prices as plain text. Be concise and human-readable.`;
         const synthesisMsg = `${effectiveMsg}\n\nThe tools have already been executed and their results are in the system prompt. Answer the question DIRECTLY using that data. Do NOT announce further actions ("Let me…", "I'll…", "leggerò", "cercherò"). Plain prose/markdown — zero JSON, zero code blocks.`;
         sse('tool_synthesis', {});
         // Keep the pre-synthesis prose around. If the synthesis call returns
@@ -627,8 +634,10 @@ export function register(router) {
       let enrichedPrompt = chatSystemPrompt;
       try { const ic = await getImapAccountsContext(); if (ic) enrichedPrompt += ic; } catch {}
       const LANG_MAP = { it:'Italian', en:'English', es:'Spanish', fr:'French', de:'German', pt:'Portuguese', nl:'Dutch', pl:'Polish', ru:'Russian', zh:'Chinese', ja:'Japanese', ko:'Korean', ar:'Arabic', hi:'Hindi', tr:'Turkish', sv:'Swedish', da:'Danish', fi:'Finnish', cs:'Czech' };
-      const userLang = LANG_MAP[(config?.language || config?.lang || 'en').slice(0,2)] || 'English';
-      enrichedPrompt += `\n\nIMPORTANT: Always respond in ${userLang}.`;
+      const settingLang = LANG_MAP[(config?.language || config?.lang || 'en').slice(0,2)] || 'English';
+      const detectedFromMsg = detectLanguage(body.message);
+      const userLang = detectedFromMsg || settingLang;
+      enrichedPrompt += `\n\nIMPORTANT: The user wrote their message in ${userLang}. Respond in ${userLang}.`;
 
       let response;
 

package/src/services/message-responder.mjs

@@ -146,32 +146,58 @@ function routeMessage(text, useAutoRoute = true) {
 
 // ── Language detection from message text ─────────────────────────────────────
 
-const IT_WORDS = new Set(['il','lo','la','le','gli','un','una','che','di','da','in','con','su','per','tra','fra','non','ma','se','come','dove','quando','chi','cosa','ho','hai','ha','sono','sei','siamo','avere','essere','fare','dire','andare','mi','ti','ci','si','vi','li','le','gli','mio','tuo','suo','nostro','vostro','loro','questo','quello','questi','quelli','anche','già','ancora','sempre','mai','oggi','domani','ieri','adesso','ora','poi','dopo','prima','qui','qua','lì','là','più','meno','molto','poco','bene','male','sì','no','grazie','prego','ciao','buongiorno','buonasera','appuntamenti','appuntamento','calendario','riunione','meteo','temperatura','email','posta','notizie','del','dello','della','degli','delle','nel','nello','nella','negli','nelle','dal','dallo','dalla','dagli','dalle','sul','sullo','sulla','sugli','sulle','col','coi','quello','quella','quelli','quelle','cancella','cancellare','elimina','eliminare','crea','creare','sposta','spostare','aggiungi','aggiungere','modifica','modificare','ricerca','trovami','trovare','mostra','mostrami','dimmi','rispondimi','aiutami','puoi','voglio','vorrei','devo','posso','giorno','giorni','mese','mesi','anno','anni','settimana','settimane','ore','minuto','minuti','mattina','pomeriggio','sera','notte','veterinario','medico','dentista','dottore','riunioni','scadenza','scadenze']);
+const IT_WORDS = new Set(['il','lo','la','le','gli','un','una','che','di','da','in','con','su','per','tra','fra','non','ma','se','come','dove','quando','chi','cosa','ho','hai','ha','sono','sei','siamo','avere','essere','fare','dire','andare','mi','ti','ci','si','vi','li','le','gli','mio','tuo','suo','nostro','vostro','loro','questo','quello','questi','quelli','anche','già','ancora','sempre','mai','oggi','domani','ieri','adesso','ora','poi','dopo','prima','qui','qua','lì','là','più','meno','molto','poco','bene','male','sì','no','grazie','prego','ciao','buongiorno','buonasera','appuntamenti','appuntamento','calendario','riunione','meteo','temperatura','email','posta','notizie','del','dello','della','degli','delle','nel','nello','nella','negli','nelle','dal','dallo','dalla','dagli','dalle','sul','sullo','sulla','sugli','sulle','col','coi','quello','quella','quelli','quelle','cancella','cancellare','elimina','eliminare','crea','creare','sposta','spostare','aggiungi','aggiungere','modifica','modificare','ricerca','trovami','trovare','mostra','mostrami','dimmi','rispondimi','aiutami','puoi','voglio','vorrei','devo','posso','giorno','giorni','mese','mesi','anno','anni','settimana','settimane','ore','minuto','minuti','mattina','pomeriggio','sera','notte','veterinario','medico','dentista','dottore','riunioni','scadenza','scadenze',
+  // Apostrophe-truncated forms produced when we strip apostrophes during tokenization (`dell'email` → tokens `dell`, `email`).
+  'dell','all','nell','sull','dall','un','nell','quell',
+  // High-frequency content words missing from the original list — common in IMAP / calendar / order use cases.
+  'leggi','leggere','dammi','dammelo','dammela','ricevuto','ricevuta','ricevute','allegato','allegati','allegata','alle','articolo','articoli','ordine','ordini','consegna','consegne','fattura','fatture','preventivo','preventivi','offerta','offerte','richiesta','richieste','documento','documenti','prezzo','prezzi','totale','spedizione','cliente','clienti','fornitore','fornitori','data','date','codice','codici']);
 const ES_WORDS = new Set(['el','la','los','las','un','una','que','de','en','con','por','para','pero','como','donde','cuando','quien','qué','tengo','tienes','tiene','somos','soy','eres','hacer','decir','ir','me','te','se','nos','este','ese','estos','esos','también','ya','todavía','siempre','nunca','hoy','mañana','ayer','aquí','allí','más','menos','muy','bien','mal','sí','no','gracias','hola','buenos']);
 const FR_WORDS = new Set(['le','la','les','un','une','des','que','de','en','avec','pour','par','mais','comme','où','quand','qui','je','tu','il','elle','nous','vous','ils','elles','avoir','être','faire','dire','aller','me','te','se','ce','cet','cette','ces','aussi','déjà','toujours','jamais','aujourd','demain','hier','ici','là','plus','moins','très','bien','mal','oui','non','merci','bonjour','bonsoir']);
 const DE_WORDS = new Set(['der','die','das','ein','eine','und','oder','aber','nicht','mit','für','von','zu','an','auf','ist','sind','hat','haben','sein','werden','ich','du','er','sie','es','wir','ihr','mich','dich','sich','uns','euch','diesem','diesen','dieser','dieses','auch','schon','noch','immer','nie','heute','morgen','gestern','hier','dort','mehr','weniger','sehr','gut','schlecht','ja','nein','danke','hallo']);
 const PT_WORDS = new Set(['o','a','os','as','um','uma','que','de','em','com','por','para','mas','como','onde','quando','quem','eu','tu','ele','ela','nós','vós','eles','elas','ter','ser','fazer','dizer','ir','me','te','se','nos','este','esse','isso','aquele','também','já','ainda','sempre','nunca','hoje','amanhã','ontem','aqui','lá','mais','menos','muito','bem','mal','sim','não','obrigado','olá']);
 
-function detectLanguage(text) {
+export function detectLanguage(text) {
   if (!text || text.length < 6) return null;
-  const words = text.toLowerCase().replace(/[^a-zàáâãäèéêëìíîïòóôõöùúûüýñçàèìòù\s]/g, ' ').split(/\s+/).filter(w => w.length > 1);
-  if (words.length < 2) return null;
-
-  let it = 0, es = 0, fr = 0, de = 0, pt = 0, en = 0;
+  const raw = text.toLowerCase();
+  // Bonus signals — language-distinctive patterns that survive even when the
+  // message is dominated by codes, identifiers, or proper nouns (which is the
+  // common case for "leggi l'allegato dell'email NCSARMEMAIL.08/05"-style
+  // queries). Each pattern adds to its language's score.
+  const bonus = { it: 0, es: 0, fr: 0, de: 0, pt: 0, en: 0 };
+  // IT-distinctive: articulated prepositions with apostrophe — these forms
+  // exist in Italian but NOT in French (French uses `du / au / aux`, not `dell' / all'`).
+  if (/\b(dell'|nell'|sull'|dall'|all'|un'[aeiou]|com'è|dov'è|qual'è)\w/.test(raw)) bonus.it += 3;
+  if (/\b(c'è|c'era|d'accordo|po'\s|più\s|però\s|perché\s|cioè\s)/.test(raw)) bonus.it += 2;
+  if (/[àèéìòù]/.test(raw) && !/\bj'|qu'/.test(raw)) bonus.it += 1;
+  // FR-distinctive: pronoun apostrophes that don't exist in Italian.
+  if (/\b(j'ai|n'est|n'a\s|c'est|qu'il|qu'elle|qu'on|n'ont)/.test(raw)) bonus.fr += 4;
+  if (/\b(qu'|j'|n'|s')\w/.test(raw)) bonus.fr += 2;
+  if (/\b(ñ|ll\w+)\b/.test(raw)) bonus.es += 1;
+  if (/[äöüß]/.test(raw)) bonus.de += 2;
+  if (/\b(ção|ões|nh)\w*/.test(raw)) bonus.pt += 1;
+
+  const words = raw.replace(/[^a-zàáâãäèéêëìíîïòóôõöùúûüýñçàèìòù\s]/g, ' ').split(/\s+/).filter(w => w.length > 1);
+  if (words.length < 2 && Object.values(bonus).every(b => b === 0)) return null;
+
+  let it = bonus.it, es = bonus.es, fr = bonus.fr, de = bonus.de, pt = bonus.pt, en = bonus.en;
   for (const w of words) {
     if (IT_WORDS.has(w)) it++;
     if (ES_WORDS.has(w)) es++;
     if (FR_WORDS.has(w)) fr++;
     if (DE_WORDS.has(w)) de++;
     if (PT_WORDS.has(w)) pt++;
-    // Basic English common words
     if (['the','a','an','is','are','was','were','have','has','do','does','i','you','he','she','we','they','and','or','but','not','with','for','from','to','in','on','at','this','that','these','those','can','will','would','could','should','what','where','when','who','how'].includes(w)) en++;
   }
 
   const max = Math.max(it, es, fr, de, pt, en);
   if (max === 0) return null;
-  const threshold = Math.max(2, words.length * 0.15); // at least 15% of words or 2
+  // Threshold lowered for short or code-heavy messages — bonus signals are
+  // strong enough on their own. Still require a clear winner: if two
+  // languages tie, return null (caller falls back to setting).
+  const threshold = Math.max(2, Math.min(words.length * 0.12, 4));
   if (max < threshold) return null;
+  const tieCount = [it, es, fr, de, pt, en].filter(s => s === max).length;
+  if (tieCount > 1) return null;
 
   if (it === max) return 'Italian';
   if (es === max) return 'Spanish';