@tiens.nguyen/gu-cli 1.0.749 → 1.0.750

package/gonext_agent_chat.py

@@ -3813,6 +3813,12 @@ def _normalize_code_tags(text: str) -> str:
          <code> tag is present, so we don't touch backticks inside a real <code> block).
       5. opened-but-never-closed (exactly one "<code>", no "</code>"): append "</code>"
          so the parser can still extract the code instead of failing the whole step.
+      6. NO DELIMITERS AT ALL, but the reply is Python that calls something — an optional
+         "Thought:" line, then e.g. `print(read_file_lines(...))`. Measured on Claude: 13
+         such replies in one turn, each thrown away as "no runnable code block" and re-asked,
+         about a quarter of the turn's steps. Wrapped only when the code (after the Thought
+         line) COMPILES and contains a call — prose does not parse as Python, so a sentence
+         is never mistaken for code.
     Idempotent — already-correct "<code>…</code>" is returned unchanged."""
     if not text:
         return text
@@ -3823,9 +3829,36 @@ def _normalize_code_tags(text: str) -> str:
         fixed = _MD_FENCE_BLOCK.sub(lambda m: f"<code>\n{m.group(1)}</code>", fixed, count=1)
     if fixed.count("<code>") == 1 and "</code>" not in fixed:
         fixed = fixed + "\n</code>"
+    if "<code>" not in fixed and "</code>" not in fixed:
+        bare = _bare_code_block(fixed)
+        if bare is not None:
+            thought, code = bare
+            fixed = (thought + "\n" if thought else "") + "<code>\n" + code + "\n</code>"
     return fixed
 
 
+def _bare_code_block(text: str):
+    """(thought, code) when `text` is Python with no delimiters — an optional leading
+    "Thought: …" line, then code that compiles and calls something. Else None."""
+    import ast
+    t = (text or "").strip()
+    if not t:
+        return None
+    thought = ""
+    lines = t.splitlines()
+    if lines and lines[0].lstrip().lower().startswith("thought:"):
+        thought, t = lines[0].strip(), "\n".join(lines[1:]).strip()
+    if not t:
+        return None
+    try:
+        tree = ast.parse(t)
+    except SyntaxError:
+        return None
+    if not any(isinstance(n, ast.Call) for n in ast.walk(tree)):
+        return None
+    return thought, t
+
+
 def _escape_raw_newlines_in_strings(code: str, defuse_tags: bool = False) -> str:
     """Turn RAW newlines/tabs that sit INSIDE a single/double-quoted string literal into
     \\n/\\t escapes. Triple-quoted strings, comments, and everything outside a string are
@@ -10310,7 +10343,11 @@ def run_agent_chat(cfg):
                         r"send_email)\s*\("
                     )
                     fence_is_toolcall = any(re.search(_tool_call_re, b) for b in fence_bodies)
-                    if has_code_tag or fence_is_toolcall:
+                    # Bare Python with no delimiters at all (case 6 of _normalize_code_tags) —
+                    # it used to fall through to the prose branch and cost a whole step.
+                    bare_code = (not has_code_tag and not fence_bodies
+                                 and _bare_code_block(content) is not None)
+                    if has_code_tag or fence_is_toolcall or bare_code:
                         normalized = _normalize_code_tags(content)
                         if normalized != content:
                             _log("normalized code-block tags (<code >/fence → <code>)")
@@ -10462,13 +10499,16 @@ def run_agent_chat(cfg):
         # from the prompt, not writing code. Deliberately carries NO literal delimiter (the
         # system prompt already shows the format with a worked example), so there is nothing
         # to echo. Applied only when kind=openai; Ollama keeps the stock message.
+        # STRONG MODELS (2026-09-25): the delimiters are named again. Leaving them out was a
+        # workaround for Kimi echoing them; on Claude it did the opposite — told to use "the
+        # delimiters shown in the system prompt", it replied with bare code, which failed the
+        # same way, 13 times in one turn. And "call exactly one tool" contradicted batching.
         _PARSE_ERROR_REPLACEMENT = (
             "Error: your last reply contained no runnable code block, so nothing executed "
             "and no observation was produced.\n"
-            "Reply with ONE 'Thought:' line, then the code block using the exact opening and "
-            "closing delimiters shown in the system prompt, with ONLY Python between them — "
-            "no prose inside the block, and do not repeat these instructions back to me. "
-            "Call exactly one tool, or call final_answer(...) if you are done."
+            "Reply with ONE 'Thought:' line, then your Python between <code> and </code> — "
+            "nothing else inside the block. Batch independent calls as usual, or call "
+            "final_answer(...) if you are done."
         )
 
         def _sanitize_parse_errors(self, msgs):

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tiens.nguyen/gu-cli",
-  "version": "1.0.749",
+  "version": "1.0.750",
   "description": "Gu CLI — the gu terminal plus the local worker that runs agent / OCR / PDF / embedding jobs on your Mac (Ollama / MLX / OpenAI-compatible).",
   "type": "module",
   "license": "MIT",