@tiens.nguyen/gonext-local-worker 1.0.84 → 1.0.86

package/gonext_agent_chat.py

@@ -302,28 +302,64 @@ def run_agent_chat(cfg):
         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,
-    # but exclude prior ASSISTANT messages (they contain raw HTTP/thinking content
-    # that confuses small models).
-    task_text = ""
-    prior_user_msgs = []
-    for m in messages:
-        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:
+    # Build the task from the conversation history. We include the FULL conversation
+    # (both user AND assistant turns) so the agent remembers what it already did —
+    # e.g. data it fetched on a previous turn. Assistant turns are condensed (drop
+    # <think> reasoning; clip long raw HTTP dumps), and we keep the most recent turns
+    # within a character budget so we never overflow the model's context window.
+    # ~8000 chars ≈ 2k tokens, tiny against Qwen2.5-Coder-7B's 32k context, leaving
+    # ample room for smolagents' own system prompt + step memory (HTTP observations).
+    HISTORY_CHAR_BUDGET = 8000
+    think_re = re.compile(r"<think>.*?</think>", re.DOTALL | re.IGNORECASE)
+
+    def _condense(role, content):
+        text = (content or "").strip()
+        if role == "assistant":
+            text = think_re.sub("", text).strip()
+            # Raw HTTP dumps add little conversational value — keep only a snippet.
+            if text.startswith("HTTP "):
+                text = text[:500]
+        return text
+
+    # The latest user message is the current task; everything before it is history.
+    last_user_idx = -1
+    for i, m in enumerate(messages):
+        if m.get("role") == "user":
+            last_user_idx = i
+    if last_user_idx < 0:
         _emit({"type": "final", "text": "[No user message found in history]"})
         return
+    task_text = (messages[last_user_idx].get("content") or "").strip()
 
-    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}"
+    # Walk prior turns newest-first, keeping condensed lines until the budget is
+    # spent, then restore chronological (oldest→newest) order.
+    history_lines = []
+    used = 0
+    for m in reversed(messages[:last_user_idx]):
+        role = m.get("role", "")
+        if role not in ("user", "assistant"):
+            continue
+        text = _condense(role, m.get("content", ""))
+        if not text:
+            continue
+        line = f"{'User' if role == 'user' else 'Assistant'}: {text}"
+        if used + len(line) > HISTORY_CHAR_BUDGET:
+            break
+        history_lines.append(line)
+        used += len(line)
+    history_lines.reverse()
+
+    if history_lines:
+        convo = "\n".join(history_lines)
+        task_text = (
+            "Conversation so far (oldest to newest):\n"
+            f"{convo}\n\nCurrent task: {task_text}"
+        )
 
+    _log(
+        f"history: {len(history_lines)} prior turn(s), {used} chars "
+        f"(budget {HISTORY_CHAR_BUDGET})"
+    )
     _log(f"task={task_text[:120]!r}")
 
     # Route: ask the model if this task needs HTTP tool use.
@@ -450,8 +486,38 @@ def run_agent_chat(cfg):
         _log(f"step {step_num}: {text[:200]}")
         _emit({"type": "step", "text": text})
 
+    # Wrap the model so we can see EXACTLY what smolagents posts to the model
+    # server on every step — including its own system prompt, the task we passed,
+    # and any step memory it accumulates. completion_kwargs["messages"] here is the
+    # literal messages array sent to /v1/chat/completions.
+    class _LoggingModel(OpenAIServerModel):
+        def _prepare_completion_kwargs(self, *args, **kwargs):
+            ck = super()._prepare_completion_kwargs(*args, **kwargs)
+            try:
+                msgs = ck.get("messages", []) or []
+                _log(f"=== MODEL REQUEST: {len(msgs)} message(s) sent to the model ===")
+                for i, m in enumerate(msgs):
+                    role = m.get("role") if isinstance(m, dict) else getattr(m, "role", "?")
+                    content = (
+                        m.get("content") if isinstance(m, dict)
+                        else getattr(m, "content", "")
+                    )
+                    if isinstance(content, list):
+                        text = " ".join(
+                            (c.get("text", "") if isinstance(c, dict) else str(c))
+                            for c in content
+                        )
+                    else:
+                        text = str(content)
+                    text = text.replace("\n", " ")
+                    _log(f"  [{i}] {role} ({len(text)} chars): {text[:600]}")
+                _log("=== END MODEL REQUEST ===")
+            except Exception as e:  # noqa: BLE001
+                _log(f"MODEL REQUEST log error: {e}")
+            return ck
+
     try:
-        model = OpenAIServerModel(
+        model = _LoggingModel(
             model_id=coding_model_id,
             api_base=coding_base_url,
             api_key=agent_api_key,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tiens.nguyen/gonext-local-worker",
-  "version": "1.0.84",
+  "version": "1.0.86",
   "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",