nothumanallowed 14.1.26 → 14.1.28

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "nothumanallowed",
-  "version": "14.1.26",
+  "version": "14.1.28",
   "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 = '14.1.26';
+export const VERSION = '14.1.28';
 export const BASE_URL = 'https://nothumanallowed.com/cli';
 export const API_BASE = 'https://nothumanallowed.com/api/v1';
 

package/src/services/llm.mjs

@@ -5,6 +5,63 @@
  * Supports: Anthropic, OpenAI, Gemini, DeepSeek, Grok, Mistral, Cohere.
  */
 
+// ── Qwen3 BPE artifact repair ─────────────────────────────────────────────
+//
+// Qwen3-32B running on vLLM produces text where Italian/English function words
+// (articles, prepositions, conjunctions) are fused to adjacent content words
+// due to BPE tokenizer subword merging.  Examples:
+//   "lacorrelazione" → "la correlazione"
+//   "ilprezzodell'oro" → "il prezzo dell'oro"
+//   "deidati" → "dei dati"
+//   "nonesiste" → "non esiste"
+//
+// Strategy: insert a space before any known function word that immediately
+// follows a lowercase letter (i.e. is fused mid-word).  We run multiple
+// passes because chains like "dellacorrelazione" need two separations.
+// Table rows (lines starting with |) are intentionally skipped to avoid
+// corrupting cell separators.
+
+const IT_FUNCTION_WORDS = [
+  // Definite articles
+  'il','lo','la','i','gli','le',
+  // Indefinite articles
+  'un','uno','una',
+  // Prepositions + contractions
+  'di','del','dello','della','dei','degli','delle',
+  'a','al','allo','alla','ai','agli','alle',
+  'da','dal','dallo','dalla','dai','dagli','dalle',
+  'in','nel','nello','nella','nei','negli','nelle',
+  'su','sul','sullo','sulla','sui','sugli','sulle',
+  'con','col','per','tra','fra',
+  // Common conjunctions / adverbs that fuse
+  'che','non','ma','se','anche','come','dove','quando',
+  'perché','mentre','quindi','però','sia','né',
+  'questo','questa','questi','queste','quello','quella',
+  'e','o','è',
+];
+
+// Build a single regex that matches any of these words preceded by a lowercase letter
+const BPE_SPLIT_RE = new RegExp(
+  '([a-zàèéìòù])(' + IT_FUNCTION_WORDS.map((w) => w.replace(/[.*+?^${}()|[\]\\]/g, '\\$&')).join('|') + ')(?=[A-Za-zàèéìòù])',
+  'g',
+);
+
+function fixQwen3BPE(text) {
+  // Process line by line — skip table rows and code blocks
+  let inCode = false;
+  return text.split('\n').map((line) => {
+    if (line.trimStart().startsWith('```')) { inCode = !inCode; return line; }
+    if (inCode) return line;
+    if (line.trimStart().startsWith('|')) return line; // table row — untouched
+    // Three passes to handle chains
+    let out = line;
+    out = out.replace(BPE_SPLIT_RE, '$1 $2');
+    out = out.replace(BPE_SPLIT_RE, '$1 $2');
+    out = out.replace(BPE_SPLIT_RE, '$1 $2');
+    return out;
+  }).join('\n');
+}
+
 // ── Providers ──────────────────────────────────────────────────────────────
 
 export async function callAnthropic(apiKey, model, systemPrompt, userMessage, stream = false, opts = {}) {
@@ -574,6 +631,13 @@ export async function callLLMStream(config, systemPrompt, userMessage, onToken,
     let fullNhaText = nhaJson.choices?.[0]?.message?.content || '';
     // Strip <think>...</think> blocks
     fullNhaText = fullNhaText.replace(/<think>[\s\S]*?<\/think>/g, '').trim();
+    // Fix Qwen3 BPE tokenizer artifacts: words joined without spaces.
+    // Qwen3 streaming via vLLM occasionally produces subword tokens that arrive
+    // concatenated — e.g. "lacorrelazione" instead of "la correlazione",
+    // "ilprezzodell'oro" instead of "il prezzo dell'oro".
+    // We fix this by inserting spaces before Italian/English articles and
+    // prepositions that appear fused at word boundaries.
+    fullNhaText = fixQwen3BPE(fullNhaText);
     if (onToken) onToken(fullNhaText);
     return fullNhaText;
   }