Fix config override bug by recreating LMs after load_state

nanocode.py

@@ -170,7 +170,8 @@ def run_bash(cmd: str) -> str:
 class RLMReasoningCallback(BaseCallback):
     def on_module_end(self, call_id, outputs, exception):
         if outputs and hasattr(outputs, "reasoning") and hasattr(outputs, "code"):
-            print(f"\n[REASONING STEP] reasoning: {outputs.reasoning}\n")
+            print(f"{DIM}⏺ [REASONING STEP] {outputs.reasoning}\n{RESET}")
+            print(f"{DIM}⏺ [CODE] {outputs.code}\n{RESET}")
 
 
 # -- Program ---
@@ -230,25 +231,77 @@ class RLMCodingProgram(PrecompiledProgram):
             tools=self.tools,
             max_output_chars=self.config.max_output_chars,
             max_iterations=self.config.max_iters,
-            verbose=self.config.verbose,
         )
         self.agent.set_lm(self.lm)
-        self.set_lm(self.lm)
+
+        if self.config.verbose:
+            self.add_logging_callbacks()
+
+    def add_logging_callbacks(self):
+        """Add logging callbacks to the agent."""
+
+        self.agent.generate_action.callbacks.append(RLMReasoningCallback())
+        self._patch_llm_tools()
+
+    def _patch_llm_tools(self):
+        """Monkey-patch the RLM's _make_llm_tools to add structured verbose logging."""
+
+        orig_factory = (
+            self.agent._make_llm_tools
+        )  # capture the original bound method directly
+
+        def verbose_factory(max_workers=8):
+            tools = orig_factory(
+                max_workers=max_workers
+            )  # call the original bound method
+
+            orig_q = tools["llm_query"]
+            orig_b = tools["llm_query_batched"]
+
+            def wrapped_q(prompt):  # wrap query
+                print(
+                    f"{DIM}⏺ [LLM QUERY]: {prompt[:100]}...{RESET}\n"
+                    if len(prompt) > 100
+                    else f"{DIM}⏺ [LLM QUERY]: {prompt}{RESET}\n"
+                )
+                res = orig_q(prompt)
+                print(
+                    f"{DIM}⏺ [LLM QUERY RESULT]: {str(res)[:200]}...{RESET}\n"
+                    if len(str(res)) > 200
+                    else f"{DIM}⏺ [LLM QUERY RESULT]: {res}{RESET}\n"
+                )
+                return res
+
+            def wrapped_b(prompts):  # wrap batched query
+                print(f"{DIM}⏺ [LLM QUERY BATCHED]: {len(prompts)} prompts{RESET}\n")
+                res = orig_b(prompts)
+                print(f"{DIM}⏺ [LLM QUERY BATCHED]: {len(res)} results{RESET}\n")
+                return res
+
+            tools["llm_query"] = wrapped_q
+            tools["llm_query_batched"] = wrapped_b
+            return tools
+
+        self.agent._make_llm_tools = verbose_factory
 
     def forward(self, task: str) -> str:
+        """Forward pass for the agent."""
         if not task:
             return dspy.Prediction(answer="No Task Given.")
 
         return self.agent(task=task)
 
     def get_tools(self):
+        """Get the tools for the agent."""
         return self.tools
 
     def set_tool(self, name: str, tool: callable):
+        """Set a tool for the agent."""
         self.tools[name] = tool
         self.reload_repl()
 
     def remove_tool(self, name: str):
+        """Remove a tool from the agent."""
         if name in self.tools:
             del self.tools[name]
             self.reload_repl()
@@ -256,6 +309,8 @@ class RLMCodingProgram(PrecompiledProgram):
     def reload_repl(
         self,
     ):  # we need to create a new instance for tool mutations to be passed back into the REPL
+        """Reload the REPL with the current tools."""
+
         new_instance = dspy.RLM(
             CodingAssistant,
             sub_lm=self.sub_lm,
@@ -266,9 +321,12 @@ class RLMCodingProgram(PrecompiledProgram):
         )
         new_instance.set_lm(self.lm)
         self.agent = new_instance
+        if self.config.verbose:
+            self.add_logging_callbacks()
 
     def reload_lms(self):
         """Recreate LM objects from current config. Call this after changing config.lm or config.sub_lm."""
+
         self.lm = dspy.LM(
             model=self.config.lm,
             api_base=self.config.api_base,
@@ -294,13 +352,19 @@ class RLMCodingProgram(PrecompiledProgram):
         fix this in a later patch for future devs.
         """
         super().load_state(state)
-        # recreate LMs from config (not from saved state)
+        self.reload_lms()  # recreate LMs from config (not from saved state)
-        self.reload_lms()
 
 
 if __name__ == "__main__":
     agent = RLMCodingProgram(RLMCodingConfig())
-    agent.push_to_hub(
+    #displayed = agent(
-        MODAIC_REPO_PATH,
+    #    task="Use only the command `print(llm_query('Who is the CEO of Apple?'))` to find the answer."
-        commit_message="Fix config override bug by recreating LMs after load_state",
+    #)
-    )
+    branches = ["main", "dev", "prod"]
+    for branch in branches:
+        agent.push_to_hub(
+            MODAIC_REPO_PATH,
+            commit_message="Fix config override bug by recreating LMs after load_state",
+            branch=branch,
+        )
+