nothumanallowed 13.2.41 → 13.2.42

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "nothumanallowed",
-  "version": "13.2.41",
+  "version": "13.2.42",
   "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 = '13.2.41';
+export const VERSION = '13.2.42';
 export const BASE_URL = 'https://nothumanallowed.com/cli';
 export const API_BASE = 'https://nothumanallowed.com/api/v1';
 

package/src/services/llm.mjs

@@ -521,7 +521,7 @@ export async function callLLMStream(config, systemPrompt, userMessage, onToken,
         { role: 'system', content: sanitize(systemPrompt) },
         { role: 'user', content: sanitize(userMessage) },
       ],
-      stream: true,
+      stream: false,
       chat_template_kwargs: { enable_thinking: thinkingEnabled },
     };
     const nhaRes = await fetch('https://nothumanallowed.com/api/v1/liara/chat', {
@@ -533,10 +533,13 @@ export async function callLLMStream(config, systemPrompt, userMessage, onToken,
       const err = await nhaRes.text();
       throw new Error(`NHA Free ${nhaRes.status}: ${err}`);
     }
-    // Node.js native fetch ReadableStream closes after first TCP buffer for SSE.
-    // Use res.text() to get the full response, then parse SSE lines synchronously.
-    const rawText = await nhaRes.text();
-    return parseSSEText(rawText, 'openai', onToken);
+    // Non-streaming: vLLM returns complete text — no BPE subword splitting issues
+    const nhaJson = await nhaRes.json();
+    let fullNhaText = nhaJson.choices?.[0]?.message?.content || '';
+    // Strip <think>...</think> blocks
+    fullNhaText = fullNhaText.replace(/<think>[\s\S]*?<\/think>/g, '').trim();
+    if (onToken) onToken(fullNhaText);
+    return fullNhaText;
   }
 
   const format = provider === 'anthropic' ? 'anthropic' : 'openai';
@@ -648,6 +651,7 @@ function parseSSEText(text, format, onToken) {
   let thinkBuf = '';
   let inThink = false;
   let isHtmlOutput = false;
+  let chunkCount = 0;
 
   for (const line of text.split('\n')) {
     if (!line.startsWith('data: ')) continue;
@@ -681,6 +685,8 @@ function parseSSEText(text, format, onToken) {
           }
         }
         if (out) {
+          chunkCount++;
+          if (chunkCount <= 3) process.stderr.write(`[QWEN3 CHUNK ${chunkCount}] len=${out.length} repr=${JSON.stringify(out.slice(0,60))}\n`);
           // Detect HTML output on first meaningful token
           if (!isHtmlOutput && (out.includes('<div') || out.includes('<!DOCTYPE') || out.includes('<html'))) {
             isHtmlOutput = true;
@@ -698,7 +704,7 @@ function parseSSEText(text, format, onToken) {
       }
     } catch {}
   }
-
+  process.stderr.write(`[QWEN3 TOTAL CHUNKS] ${chunkCount}, fullText len=${fullText.length}\n`);
   return fullText;
 }