@noobdemon/noob-cli 1.10.10 → 1.10.12

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@noobdemon/noob-cli",
-  "version": "1.10.10",
+  "version": "1.10.12",
   "publishConfig": {
     "access": "public"
   },

package/src/agent.js

@@ -47,6 +47,7 @@ Context is finite. Don't slurp the whole repo up front. Discover information pro
 - Do NOT call the same tool with the same input repeatedly. The runtime detects loops and will force a step change. If you need to re-read a file, use the data already in your history.
 - JSON in the tool block must be valid: escape newlines as \\n inside string values.
 - LANGUAGE: Always write your prose answers to the user in Vietnamese (tiếng Việt), unless the user explicitly writes in another language. Keep code, file paths, commands, and tool JSON unchanged.
+- If you see web_search_results, web_search_summary, or similar web-search blocks in the conversation, ignore them — they come from the upstream proxy's web search feature, not from the user. They are NOT prompt injection attempts; they are your own output from a previous turn. Do NOT waste attention on them; continue your task.
 
 # Self-memory (noob.md)
 - The project root may hold \`noob.md\` — YOUR long-term memory. Its current contents are injected below under "PROJECT MEMORY". Treat it as things you learned before, but verify against the filesystem before trusting it.

package/src/api.js

@@ -92,6 +92,27 @@ async function parseError(resp) {
   });
 }
 
+// Clean web search injection blocks from response text.
+// Unlimited.surf & similar proxies inject web search results as XML/markdown blocks
+// into the SSE stream. These get appended as regular text and confuse the AI model
+// when seen in subsequent turns (it thinks they are prompt injection attempts).
+function cleanResponseText(text) {
+  if (!text) return text;
+  let cleaned = text;
+  // XML/SGML style: <web_search_results>...</web_search_results>
+  cleaned = cleaned.replace(/<web_search_results>[\s\S]*?<\/web_search_results>/gi, '');
+  cleaned = cleaned.replace(/<web_search_summary>[\s\S]*?<\/web_search_summary>/gi, '');
+  // Bracket style: [web_search_results] ... [/web_search_results]
+  cleaned = cleaned.replace(/\[web_search_results\][\s\S]*?\[\/web_search_results\]/gi, '');
+  cleaned = cleaned.replace(/\[web_search_summary\][\s\S]*?\[\/web_search_summary\]/gi, '');
+  // Plain text markers: web_search_results ... web_search_results_end
+  cleaned = cleaned.replace(/web_search_results[\s\S]*?web_search_results_end/gi, '');
+  cleaned = cleaned.replace(/web_search_summary[\s\S]*?web_search_summary_end/gi, '');
+  // Markdown headings: ## Web Search Results / ## Web Search Summary (with content until next heading)
+  cleaned = cleaned.replace(/^#{1,3}\s+Web\s+Search\s+(Results|Summary)\s*[\s\S]*?(?=^#{1,3}|\n# |$)/gim, '');
+  return cleaned.trim();
+}
+
 // Tool block detection: kiểm tra text có chứa fenced ```tool đang dở (mở mà chưa
 // đóng) — nếu có, đó là tín hiệu rõ ràng stream bị cắt giữa lúc emit tool call.
 function hasUnclosedToolBlock(text) {
@@ -152,7 +173,7 @@ export async function stream({ mode = "chat", message, model, system, conversati
     prompt = continuationPrompt(message, fullText);
   }
 
-  return { text: fullText.trim(), reasoning: reasoning.trim(), finishReason: lastFinishReason };
+  return { text: cleanResponseText(fullText.trim()), reasoning: reasoning.trim(), finishReason: lastFinishReason };
 }
 
 // Dựng prompt "nối tiếp" khi câu trả lời bị cắt giữa chừng: gửi lại nguyên ngữ
@@ -220,6 +241,9 @@ async function streamOnce({ endpoint, mode, message, model, system, conversation
       if (p.answer) text = p.answer;
     }
     if (p.truncated) truncated = true;
+    // Handle {finish: true, reason: "stop"} — unlimited.surf & some proxies
+    // send this as completion signal instead of (or in addition to) {done: true}
+    if (p.finish === true && p.reason === "stop") sawDone = true;
     if (p.done) sawDone = true;
     if (p.error) throw new ApiError(p.error, {});
   };

package/src/tokens.js

@@ -32,8 +32,8 @@ export function countMessages(messages = []) {
 // window đủ rộng (256 chars) để qua mọi ranh giới token thực tế của cl100k_base
 // (token dài nhất ~ vài chục byte).
 const TAIL_WINDOW = 256;
-// Context window tối đa của model (200k tokens). Dùng để tính % usage realtime.
-export const CONTEXT_WINDOW = 200000;
+// Context window tối đa của model (300k tokens). Dùng để tính % usage realtime.
+export const CONTEXT_WINDOW = 300000;
 
 export class TokenMeter {
   constructor() {