@tiens.nguyen/gonext-local-worker 1.0.123 → 1.0.124

package/gonext_agent_chat.py

@@ -904,20 +904,33 @@ def run_agent_chat(cfg):
     raw_coding_base = (cfg.get("codingBaseURL") or "").strip()
     raw_coding_model = (cfg.get("codingModelId") or "").strip()
     if raw_coding_base:
-        # A dedicated coding server is configured. If no model name was given,
-        # ask the server which model it serves (mlx_lm.server otherwise tries to
-        # download a mismatched name from HF and 404s).
-        detected = raw_coding_model or _detect_model_id(raw_coding_base, agent_api_key)
-        if detected:
-            coding_base_url = raw_coding_base
-            coding_model_id = detected
-        else:
+        same_server = raw_coding_base.rstrip("/") == (agent_base_url or "").rstrip("/")
+        if same_server and not raw_coding_model:
+            # The coding URL points at the SAME server as the chat model. Do NOT
+            # auto-detect: /v1/models can list several cached models and return a
+            # SMALLER one first (e.g. a 3B), silently downgrading the agent's reasoning.
+            # Reuse the chat model id directly — it's the model actually loaded here.
             _log(
-                f"coding model id unresolved for {raw_coding_base!r}; "
-                "falling back to chat model"
+                f"coding base == chat base ({raw_coding_base}); "
+                f"reusing chat model {agent_model_id!r} (skipping auto-detect)"
             )
             coding_base_url = agent_base_url
             coding_model_id = agent_model_id
+        else:
+            # A DIFFERENT dedicated coding server. If no model name was given, ask the
+            # server which model it serves (mlx_lm.server otherwise tries to download a
+            # mismatched name from HF and 404s).
+            detected = raw_coding_model or _detect_model_id(raw_coding_base, agent_api_key)
+            if detected:
+                coding_base_url = raw_coding_base
+                coding_model_id = detected
+            else:
+                _log(
+                    f"coding model id unresolved for {raw_coding_base!r}; "
+                    "falling back to chat model"
+                )
+                coding_base_url = agent_base_url
+                coding_model_id = agent_model_id
     else:
         coding_base_url = agent_base_url
         coding_model_id = agent_model_id
@@ -1187,15 +1200,13 @@ def run_agent_chat(cfg):
         "\n"
         "http_request 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"
+        "AUTH — ONLY when the TASK itself provides credentials or a token. If it does "
+        "NOT, call http_request with NO username/password/headers. NEVER invent or copy "
+        "the placeholder values below.\n"
+        "  BASIC AUTH — pass username= and password= (auto base64; never build a 'Basic ' header):\n"
+        "      http_request('GET', url, username=<username from the task>, password=<password from the task>)\n"
+        "  BEARER TOKEN — pass it in headers:\n"
+        "      http_request('GET', url, headers='{\"Authorization\": \"Bearer <token from the task>\"}')\n"
         "\n"
         "CHOOSING A TOOL (match the TASK, not these examples):\n"
         "- ONLY a date/time question (e.g. 'what is the date today') -> get_current_datetime().\n"
@@ -1203,6 +1214,8 @@ def run_agent_chat(cfg):
         "knowledge -> web_search(query).\n"
         "- 'read this page' / 'open this link' / 'what does <URL> say' -> fetch_url(url).\n"
         "- any arithmetic, percentage, or 'how much is' math -> calculate(expression).\n"
+        "- a live/current NUMBER with no URL given (price, exchange rate, weather, score, "
+        "stock/crypto) -> web_search(query). Do NOT guess an API URL for these.\n"
         "- 'create/make/generate/export a PDF' of some text/data -> create_pdf(text, title).\n"
         + _pdf_read_choose_line + _email_choose_line +
         "- A specific known API/URL was given -> http_request().\n"
@@ -1212,8 +1225,11 @@ def run_agent_chat(cfg):
         "If nothing works, call final_answer explaining what you need — do NOT make up an answer.\n"
         "- Only report what a tool ACTUALLY returned. Never fabricate a response, body, or status code.\n"
         "- Pass an http_request response DIRECTLY to final_answer — do NOT split, parse, or index it.\n"
-        "- If a response starts with 'HTTP 2' it SUCCEEDED — call final_answer immediately.\n"
-        "- If a tool returns 'Error:' or HTTP 4xx/5xx, try a DIFFERENT approach, not the same URL.\n"
+        "- When a call's result is UNCERTAIN (any network call), do NOT call final_answer "
+        "in the SAME block. Call the tool alone, READ the Observation, THEN answer next step.\n"
+        "- If a response starts with 'HTTP 2' it SUCCEEDED — answer with it next step.\n"
+        "- NEVER pass an 'Error:' / HTTP 4xx / 5xx result to final_answer. On an error, try a "
+        "DIFFERENT approach next step (e.g. web_search); only give up after a few tries.\n"
         "- Do NOT put final_answer outside the code block.\n\n"
     )
     # Lead with the TASK so the weak model anchors on what's actually being asked —

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tiens.nguyen/gonext-local-worker",
-  "version": "1.0.123",
+  "version": "1.0.124",
   "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",