nothumanallowed 14.1.51 → 14.1.52

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "nothumanallowed",
-  "version": "14.1.51",
+  "version": "14.1.52",
   "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 = '14.1.51';
+export const VERSION = '14.1.52';
 export const BASE_URL = 'https://nothumanallowed.com/cli';
 export const API_BASE = 'https://nothumanallowed.com/api/v1';
 

package/src/services/llm.mjs

@@ -59,7 +59,7 @@ const BOUNDARY_WORDS = new Set([
 ]);
 
 // Single characters commonly produced by BPE fragmentation in Italian/English
-const BPE_SINGLE_CHARS = new Set(['a','c','d','e','i','l','m','n','o','p','r','s','t','u','v']);
+const BPE_SINGLE_CHARS = new Set('abcdefghijklmnopqrstuvwxyzàèéìòù'.split(''));
 
 /**
  * Enterprise-grade linguistic scorer for Italian/English word candidates.
@@ -693,8 +693,8 @@ export async function callNHA(apiKey, model, systemPrompt, userMessage, stream =
   let content = data.choices?.[0]?.message?.content || '';
   // Strip thinking tags if present
   content = content.replace(/<think>[\s\S]*?<\/think>/g, '').trim();
-  // Fix Qwen3 BPE tokenizer artifacts (words fused without spaces)
-  content = fixQwen3BPE(content);
+  // NOTE: Do NOT apply fixQwen3BPE here. With stream:false, vLLM returns
+  // correctly-spaced text. The BPE repair regex corrupts normal words.
   return content;
 }
 
@@ -962,13 +962,9 @@ 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);
+    // NOTE: Do NOT apply fixQwen3BPE here. With stream:false, vLLM returns
+    // correctly-spaced text. The BPE repair regex is too aggressive and
+    // corrupts normal Italian words (e.g. "assistente" → "ass ist ente").
     if (onToken) onToken(fullNhaText);
     return fullNhaText;
   }
@@ -1233,8 +1229,6 @@ async function streamSSEWithCallback(res, format, onToken) {
     }
   }
 
-  // Apply BPE repair to the final accumulated text before returning
-  fullText = fixQwen3BPE(fullText);
   return fullText;
 }