@tiens.nguyen/gonext-local-worker 1.0.73 → 1.0.75

package/gonext-local-worker.mjs

@@ -1061,8 +1061,10 @@ async function correctOcrText(extractedText, modelOverride = "") {
   if (override.startsWith("http://") || override.startsWith("https://")) {
     const baseUrl = /\/v1\/?$/i.test(override) ? override : `${override}/v1`;
     console.log(`[gonext-worker] OCR correction via server ${baseUrl}`);
+    console.log(`[gonext-worker] OCR correction input: ${extractedText.slice(0, 300)}`);
     try {
       const corrected = await correctOcrTextViaServer(extractedText, baseUrl);
+      console.log(`[gonext-worker] OCR correction output: ${corrected.slice(0, 300)}`);
       console.log(`[gonext-worker] OCR correction done: ${extractedText.length} → ${corrected.length} chars`);
       return corrected;
     } catch (e) {
@@ -1081,6 +1083,7 @@ async function correctOcrText(extractedText, modelOverride = "") {
     "Return only the corrected text without any explanation.\n\n" +
     `Text:\n${extractedText}`;
   console.log(`[gonext-worker] OCR correction via CLI model=${modelPath}`);
+  console.log(`[gonext-worker] OCR correction input: ${extractedText.slice(0, 300)}`);
   // Write prompt to a temp Python script that calls mlx_lm programmatically,
   // avoiding CLI arg length/escaping limits and the missing --prompt-file flag.
   const scriptFile = join(tmpdir(), `gonext-ocr-correct-${Date.now()}.py`);
@@ -1104,6 +1107,7 @@ print(result)
       console.log(`[gonext-worker] OCR correction CLI stderr: ${stderr.trim().slice(0, 300)}`);
     }
     const corrected = normalizeCorrection(stdout, extractedText);
+    console.log(`[gonext-worker] OCR correction output: ${corrected.slice(0, 300)}`);
     console.log(`[gonext-worker] OCR correction done: ${extractedText.length} → ${corrected.length} chars`);
     return corrected;
   } catch (e) {

package/gonext_agent_chat.py

@@ -128,10 +128,12 @@ def _route(task_text: str, base_url: str, api_key: str, model_id: str) -> bool:
             model=model_id,
             messages=[
                 {"role": "system", "content": (
-                    "You are a task classifier. Reply with YES or NO only, no punctuation."
+                    "You are a task classifier. Reply YES or NO only, no punctuation. "
+                    "Answer NO if the task can be solved with Python stdlib (datetime, math, etc.) "
+                    "or is just conversation. Answer YES only if it requires calling an EXTERNAL URL or web API."
                 )},
                 {"role": "user", "content": (
-                    f"Does this task require making an HTTP request to a URL or API?\n\n"
+                    f"Does this task specifically require making an HTTP request to an external URL or web API?\n\n"
                     f"Task: {task_text}\n\nYES or NO:"
                 )},
             ],
@@ -246,10 +248,17 @@ def run_agent_chat(cfg):
         "You have ONE built-in function: `http_request(method, url, headers='', body='')`. "
         "It returns a STRING with the HTTP status and body. "
         "Call it, then immediately call `final_answer(response)` inside the same code block.\n"
-        "Do NOT put final_answer outside the code block.\n\n"
+        "IMPORTANT RULES:\n"
+        "- Python's `datetime` module is available — use it for date/time tasks, no HTTP needed.\n"
+        "- If http_request returns an error starting with 'Error:', the URL failed. "
+        "Do NOT retry the same URL. Try a DIFFERENT URL or API, or use Python stdlib.\n"
+        "- Do NOT put final_answer outside the code block.\n\n"
     )
     task_with_hint = tool_hint + "Task: " + task_text
 
+    # Track URLs that have already failed so we don't retry dead endpoints across steps.
+    _failed_urls: set = set()
+
     @tool
     def http_request(method: str, url: str, headers: str = "", body: str = "") -> str:
         """Perform an HTTP request and return the status code and body preview.
@@ -266,11 +275,25 @@ def run_agent_chat(cfg):
                 parsed_headers = json.loads(headers)
             except Exception:  # noqa: BLE001
                 pass
-        # Retry once on timeout — gorok tunnels can be flaky on the first attempt.
+        url_key = f"{method.upper()}:{url}"
+        if url_key in _failed_urls:
+            msg = f"Error: {url} already failed — try a different URL or use Python stdlib."
+            _emit({"type": "step", "text": f"HTTP {method.upper()} {url} → (skipped, already failed)"})
+            _log(f"http_request skipped (already failed): {url_key}")
+            return msg
         result = _http_request_impl(method, url, parsed_headers, body or None)
         if result.startswith("Error:"):
+            # Retry once for flaky connections (e.g. gorok tunnels).
             _log(f"http_request retry {method.upper()} {url}")
             result = _http_request_impl(method, url, parsed_headers, body or None)
+        if result.startswith("Error:"):
+            # Both attempts failed — mark URL as dead so model tries something else.
+            _failed_urls.add(url_key)
+            result = (
+                f"{result}\n"
+                "Note: This URL failed twice. Do NOT retry it. "
+                "Try a DIFFERENT URL or use Python's datetime/math/etc. module instead."
+            )
         status_line = result.split("\n")[0][:150] if result else "no response"
         _emit({"type": "step", "text": f"HTTP {method.upper()} {url} → {status_line}"})
         _log(f"http_request {method.upper()} {url} → {result[:80]}")

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tiens.nguyen/gonext-local-worker",
-  "version": "1.0.73",
+  "version": "1.0.75",
   "description": "Polls GoNext cloud API for async local LLM jobs and runs them against Ollama/OpenAI-compatible servers on this Mac",
   "type": "module",
   "license": "MIT",