@tiens.nguyen/gonext-local-worker 1.0.79 → 1.0.81

package/gonext-local-worker.mjs

@@ -1376,6 +1376,8 @@ async function runAgentChatJob(job) {
       agentBaseURL: payload?.agentBaseURL ?? "",
       agentApiKey: payload?.agentApiKey ?? "",
       agentModelId: payload?.agentModelId ?? "",
+      codingBaseURL: payload?.codingBaseURL ?? "",
+      codingModelId: payload?.codingModelId ?? "",
       tools: payload?.tools ?? ["http_request"],
       maxSteps: payload?.maxSteps ?? 5,
     });

package/gonext_agent_chat.py

@@ -8,6 +8,8 @@ Reads on stdin:
     "agentBaseURL": str,
     "agentApiKey": str,
     "agentModelId": str,
+    "codingBaseURL": str,        # optional: dedicated coding/reasoning model for the
+    "codingModelId": str,        #   CodeAgent's tool-use loop; empty = reuse agentModelId
     "tools": ["http_request"],   # v1: only http_request
     "maxSteps": int              # default 10
   }
@@ -204,18 +206,27 @@ def _summarize_result(task_text: str, agent_output: str,
         return agent_output
 
 
-def _plain_reply(task_text: str, base_url: str, api_key: str, model_id: str) -> str:
-    """Plain chat completion without any tools."""
+def _plain_reply(messages: list, base_url: str, api_key: str, model_id: str) -> str:
+    """Plain chat completion using the full conversation history."""
+    _THINK_RE_LOCAL = re.compile(r"<think>.*?</think>", re.DOTALL | re.IGNORECASE)
+    chat_messages = [{"role": "system", "content": "You are a helpful assistant."}]
+    for m in messages:
+        role = m.get("role", "")
+        content = m.get("content", "")
+        if role not in ("user", "assistant"):
+            continue
+        if role == "assistant":
+            content = _THINK_RE_LOCAL.sub("", content).strip()
+            if not content:
+                continue
+        chat_messages.append({"role": role, "content": content})
     try:
         from openai import OpenAI
         client = OpenAI(base_url=base_url, api_key=api_key or "local",
                         max_retries=0, timeout=60)
         resp = client.chat.completions.create(
             model=model_id,
-            messages=[
-                {"role": "system", "content": "You are a helpful assistant."},
-                {"role": "user", "content": task_text},
-            ],
+            messages=chat_messages,
             temperature=0.7,
             max_tokens=512,
         )
@@ -235,9 +246,17 @@ def run_agent_chat(cfg):
     agent_base_url = cfg.get("agentBaseURL") or ""
     agent_api_key = cfg.get("agentApiKey") or "local"
     agent_model_id = cfg.get("agentModelId") or ""
+    # Optional dedicated coding/reasoning model for the CodeAgent's tool-use loop.
+    # Routing, plain replies and summarization stay on the chat model (better at
+    # natural language); the code model only drives http_request reasoning.
+    coding_base_url = (cfg.get("codingBaseURL") or "").strip() or agent_base_url
+    coding_model_id = (cfg.get("codingModelId") or "").strip() or agent_model_id
     max_steps = int(cfg.get("maxSteps") or 5)
 
-    _log(f"start model={agent_model_id!r} base={agent_base_url!r} maxSteps={max_steps}")
+    _log(
+        f"start model={agent_model_id!r} base={agent_base_url!r} "
+        f"codeModel={coding_model_id!r} codeBase={coding_base_url!r} maxSteps={max_steps}"
+    )
 
     # Build task from the conversation history.
     # Include prior USER messages as context so the agent has conversational memory,
@@ -268,7 +287,7 @@ def run_agent_chat(cfg):
 
     if not needs_agent:
         _log("router: plain chat (no HTTP needed)")
-        answer = _plain_reply(task_text, agent_base_url, agent_api_key, agent_model_id)
+        answer = _plain_reply(messages, agent_base_url, agent_api_key, agent_model_id)
         _log(f"plain reply: {len(answer)} chars")
         _emit({"type": "final", "text": answer})
         return
@@ -282,20 +301,24 @@ def run_agent_chat(cfg):
     tool_hint = (
         "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"
+        "\n"
+        "RETURN FORMAT: 'HTTP 200\\n{body}' — first line is 'HTTP <code>', body follows.\n"
+        "\n"
+        "BASIC AUTH — ALWAYS use username= and password=, NEVER construct headers manually:\n"
+        "  response = http_request('GET', 'https://api.example.com/data',\n"
+        "                          username='alice@example.com', password='secret123')\n"
+        "  final_answer(response)\n"
+        "The function handles base64 encoding automatically. NEVER write 'Basic ' + anything.\n"
+        "\n"
+        "BEARER TOKEN — use headers:\n"
+        "  response = http_request('GET', url, headers='{\"Authorization\": \"Bearer TOKEN\"}')\n"
         "  final_answer(response)\n"
-        "Do NOT try to parse or split the response string.\n"
-        "IMPORTANT RULES:\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"
+        "\n"
+        "RULES:\n"
+        "- Pass response DIRECTLY to final_answer — do NOT split, parse, or index the string.\n"
+        "- If the response starts with 'HTTP 2' it SUCCEEDED — call final_answer immediately.\n"
+        "- If http_request returns 'Error:' or HTTP 4xx/5xx, try a different approach.\n"
+        "- Python's datetime module is available for date/time tasks (no HTTP needed).\n"
         "- Do NOT put final_answer outside the code block.\n\n"
     )
     task_with_hint = tool_hint + "Task: " + task_text
@@ -385,8 +408,8 @@ def run_agent_chat(cfg):
 
     try:
         model = OpenAIServerModel(
-            model_id=agent_model_id,
-            api_base=agent_base_url,
+            model_id=coding_model_id,
+            api_base=coding_base_url,
             api_key=agent_api_key,
         )
         agent = CodeAgent(

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tiens.nguyen/gonext-local-worker",
-  "version": "1.0.79",
+  "version": "1.0.81",
   "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",