@tiens.nguyen/gonext-local-worker 1.0.77 → 1.0.79

package/gonext_agent_chat.py

@@ -239,16 +239,28 @@ def run_agent_chat(cfg):
 
     _log(f"start model={agent_model_id!r} base={agent_base_url!r} maxSteps={max_steps}")
 
-    # Use only the latest user message as the agent task.
+    # Build task from the conversation history.
+    # Include prior USER messages as context so the agent has conversational memory,
+    # but exclude prior ASSISTANT messages (they contain raw HTTP/thinking content
+    # that confuses small models).
     task_text = ""
+    prior_user_msgs = []
     for m in messages:
-        if m.get("role") == "user":
-            task_text = m.get("content", "")
+        role = m.get("role", "")
+        content = m.get("content", "")
+        if role == "user":
+            if task_text:
+                prior_user_msgs.append(task_text)
+            task_text = content
 
     if not task_text:
         _emit({"type": "final", "text": "[No user message found in history]"})
         return
 
+    if prior_user_msgs:
+        context = "\n".join(f"- {t[:300]}" for t in prior_user_msgs[-4:])
+        task_text = f"Conversation context (previous user messages):\n{context}\n\nCurrent task: {task_text}"
+
     _log(f"task={task_text[:120]!r}")
 
     # Route: ask the model if this task needs HTTP tool use.
@@ -268,17 +280,22 @@ def run_agent_chat(cfg):
     # Prepend explicit tool instructions so small models use http_request correctly
     # and always terminate with final_answer() rather than looping forever.
     tool_hint = (
-        "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"
+        "You have ONE built-in function:\n"
+        "  `http_request(method, url, headers='', body='', username='', password='')`\n"
+        "RETURN FORMAT: the function returns a string like:\n"
+        "  'HTTP 200\\n{\"key\": \"value\"}'\n"
+        "  First line is 'HTTP <code>' (e.g. 'HTTP 200'). Body follows after the first newline.\n"
+        "CORRECT USAGE — always just pass response directly to final_answer:\n"
+        "  response = http_request('GET', url)\n"
+        "  final_answer(response)\n"
+        "Do NOT try to parse or split the response string.\n"
         "IMPORTANT RULES:\n"
-        "- `headers` must be a JSON string, e.g. headers='{\"Content-Type\": \"application/json\"}'\n"
-        "- For Basic Auth: import base64; "
-        "creds = base64.b64encode(b'user:password').decode(); "
-        "headers = '{\"Authorization\": \"Basic ' + creds + '\"}'\n"
-        "- For Bearer token: headers='{\"Authorization\": \"Bearer YOUR_TOKEN\"}'\n"
-        "- Python's `datetime` and `base64` modules are available.\n"
-        "- If http_request returns an error starting with 'Error:', do NOT retry the same URL. Try a different one.\n"
+        "- If you get HTTP 2xx (200, 201, etc.) in the first line, the request SUCCEEDED. "
+        "Call final_answer immediately — do NOT retry.\n"
+        "- For Basic Auth: username= and password= params handle encoding automatically.\n"
+        "- For Bearer token: headers='{\"Authorization\": \"Bearer TOKEN\"}'\n"
+        "- If http_request returns an error (starts with 'Error:'), try a different URL.\n"
+        "- Python's `datetime` module is available for date/time tasks.\n"
         "- Do NOT put final_answer outside the code block.\n\n"
     )
     task_with_hint = tool_hint + "Task: " + task_text
@@ -287,7 +304,8 @@ def run_agent_chat(cfg):
     _failed_urls: set = set()
 
     @tool
-    def http_request(method: str, url: str, headers: str = "", body: str = "") -> str:
+    def http_request(method: str, url: str, headers: str = "", body: str = "",
+                     username: str = "", password: str = "") -> str:
         """Perform an HTTP request and return the status code and body preview.
 
         Args:
@@ -295,11 +313,18 @@ def run_agent_chat(cfg):
             url: Full URL to request
             headers: Optional JSON object string of request headers
             body: Optional request body string
+            username: Username for Basic Authentication (avoids manual base64 encoding)
+            password: Password for Basic Authentication
         """
+        import base64 as _b64
         parsed_headers = {}
+        # Basic Auth: encode credentials automatically if provided.
+        if username or password:
+            creds = _b64.b64encode(f"{username}:{password}".encode()).decode()
+            parsed_headers["Authorization"] = f"Basic {creds}"
         if headers:
             try:
-                parsed_headers = json.loads(headers)
+                parsed_headers.update(json.loads(headers))
             except Exception:  # noqa: BLE001
                 pass
         url_key = f"{method.upper()}:{url}"
@@ -322,6 +347,9 @@ def run_agent_chat(cfg):
                 "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"
+        # Append a success tag to 2xx responses so the model knows to stop and call final_answer.
+        if result and result.startswith("HTTP 2"):
+            result = result + "\n[SUCCESS — call final_answer(response) now, do not parse or retry]"
         _emit({"type": "step", "text": f"HTTP {method.upper()} {url} → {status_line}"})
         _log(f"http_request {method.upper()} {url} → {result[:80]}")
         return result

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tiens.nguyen/gonext-local-worker",
-  "version": "1.0.77",
+  "version": "1.0.79",
   "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",