Merge pull request #139 from cagostino/chris/shell_upgrades

upgrading check llm command dand code executor

Author: cagostino

npcsh/llm_funcs.py

@@ -802,15 +802,20 @@ def check_llm_command(
 
     prompt = f"""
     A user submitted this query: {command}
+
     Determine the nature of the user's request:
-    1. Is it a specific request for a task that could be accomplished via a bash command or a simple python script that could be executed in a single bash call?
-    2. Should a tool be invoked to fulfill the request?
-    3. Is it a general question that requires an informative answer or a highly specific question that
+
+    1. Should a tool be invoked to fulfill the request?
+
+    2. Is it a general question that requires an informative answer or a highly specific question that
         requires inforrmation on the web?
-    4. Would this question be best answered by an alternative NPC?
-    5. Is it a complex request that actually requires more than one
-    tool to be called, perhaps in a sequence?
-    6. is there a need for the user to provide additional input to fulfill the request?
+
+    3. Would this question be best answered by an alternative NPC?
+
+    4. Is it a complex request that actually requires more than one
+        tool to be called, perhaps in a sequence?
+
+    5. is there a need for the user to provide additional input to fulfill the request?
 
 
 
@@ -877,8 +882,12 @@ def check_llm_command(
 
     prompt += f"""
     In considering how to answer this, consider:
-    - Whether it can be answered via a bash command on the user's computer. e.g. if a user is curious about file sizes within a directory or about processes running on their computer, these are likely best handled by a bash command.
-    - Whether more context from the user is required to adequately answer the question. e.g. if a user asks for a joke about their favorite city but they don't include the city , it would be helpful to ask for that information. Similarly, if a user asks to open a browser and to check the weather in a city, it would be helpful to ask for the city and which website or source to use.
+
+    - Whether more context from the user is required to adequately answer the question.
+        e.g. if a user asks for a joke about their favorite city but they don't include the city ,
+        it would be helpful to ask for that information. Similarly, if a user asks to open a browser
+        and to check the weather in a city, it would be helpful to ask for the city and which website
+        or source to use.
     - Whether a tool should be used.
 
 
@@ -887,14 +896,17 @@ def check_llm_command(
     extra tools or agent passes.
     Only use tools or pass to other NPCs
     when it is obvious that the answer needs to be as up-to-date as possible. For example,
-    a question about where mount everest is does not necessarily need to be answered by a tool call or an agent pass.
+        a question about where mount everest is does not necessarily need to be answered by a tool call or an agent pass.
     Similarly, if a user asks to explain the plot of the aeneid, this can be answered without a tool call or agent pass.
-    If a user were to ask for the current weather in tokyo or the current price of bitcoin or who the mayor of a city is, then a tool call or agent pass may be appropriate. If a user asks about the process using the most ram or the biggest file in a directory, a bash command will be most appropriate.
+
+    If a user were to ask for the current weather in tokyo or the current price of bitcoin or who the mayor of a city is,
+        then a tool call or agent pass may be appropriate.
+
     Tools are valuable but their use should be limited and purposeful to
     ensure the best user experience.
 
     Respond with a JSON object containing:
-    - "action": one of ["execute_command", "invoke_tool", "answer_question", "pass_to_npc", "execute_sequence", "request_input"]
+    - "action": one of ["invoke_tool", "answer_question", "pass_to_npc", "execute_sequence", "request_input"]
     - "tool_name": : if action is "invoke_tool": the name of the tool to use.
                      else if action is "execute_sequence", a list of tool names to use.
     - "explanation": a brief explanation of why you chose this action.
@@ -907,15 +919,17 @@ def check_llm_command(
 
     The format of the JSON object is:
     {{
-        "action": "execute_command" | "invoke_tool" | "answer_question" | "pass_to_npc" | "execute_sequence" | "request_input",
+        "action": "invoke_tool" | "answer_question" | "pass_to_npc" | "execute_sequence" | "request_input",
         "tool_name": "<tool_name(s)_if_applicable>",
         "explanation": "<your_explanation>",
         "npc_name": "<npc_name(s)_if_applicable>"
     }}
 
     If you execute a sequence, ensure that you have a specified NPC for each tool use.
 
-    Remember, do not include ANY ADDITIONAL MARKDOWN FORMATTING. There should be no prefix 'json'. Start straight with the opening curly brace.
+    Remember, do not include ANY ADDITIONAL MARKDOWN FORMATTING.
+    There should be no leading ```json.
+
     """
 
     if docs_context:
@@ -932,11 +946,6 @@ def check_llm_command(
         {context}
 
         """
-
-    # print(prompt)
-
-    # For action determination, we don't need to pass the conversation messages to avoid confusion
-    # print(npc, model, provider)
     action_response = get_llm_response(
         prompt,
         model=model,
@@ -965,12 +974,11 @@ def check_llm_command(
     else:
         response_content_parsed = response_content
 
-    # Proceed according to the action specified
     action = response_content_parsed.get("action")
     explanation = response_content["explanation"]
-    # Include the user's command in the conversation messages
     print(f"action chosen: {action}")
     print(f"explanation given: {explanation}")
+
     if response_content_parsed.get("tool_name"):
         print(f"tool name: {response_content_parsed.get('tool_name')}")
 
@@ -1316,8 +1324,9 @@ def handle_tool_call(
             stream=stream,
             messages=messages,
         )
-        if "Error" in tool_output:
-            raise Exception(tool_output)
+        if not stream:
+            if "Error" in tool_output:
+                raise Exception(tool_output)
     except Exception as e:
         # diagnose_problem = get_llm_response(
         ##    f"""a problem has occurred.

npcsh/npc_compiler.py

@@ -551,6 +551,7 @@ def execute(
 
         # Process Steps
         for i, step in enumerate(self.steps):
+
             context = self.execute_step(
                 step,
                 context,
@@ -564,6 +565,7 @@ def execute(
             # if i is the last step and the user has reuqested a streaming output
             # then we should return the stream
             if i == len(self.steps) - 1 and stream:  # this was causing the big issue X:
+                print("tool successful, passing output to stream")
                 return context
         # Return the final output
         if context.get("output") is not None:
@@ -592,8 +594,14 @@ def execute_step(
         except Exception as e:
             print(f"Error rendering template: {e}")
             rendered_code = code
-
-        if engine == "natural":
+        # render engine if necessary
+        try:
+            template = jinja_env.from_string(engine)
+            rendered_engine = template.render(**context)
+        except:
+            print("error rendering engine")
+            rendered_engine = engine
+        if rendered_engine == "natural":
             if len(rendered_code.strip()) > 0:
                 # print(f"Executing natural language step: {rendered_code}")
                 if stream:
@@ -610,7 +618,7 @@ def execute_step(
                     context["llm_response"] = response_text
                     context["results"] = response_text
 
-        elif engine == "python":
+        elif rendered_engine == "python":
             exec_globals = {
                 "__builtins__": __builtins__,
                 "npc": npc,
@@ -639,12 +647,23 @@ def execute_step(
             exec_env = context.copy()
             try:
                 exec(rendered_code, exec_globals, new_locals)
+                exec_env.update(new_locals)
+
+                context.update(exec_env)
+
                 exec_env.update(new_locals)
                 context.update(exec_env)
-                # If output is set, also set it as results
+
+                # Add this line to explicitly copy the output
+                if "output" in new_locals:
+                    context["output"] = new_locals["output"]
+
+                # Then your existing code
                 if "output" in exec_env:
                     if exec_env["output"] is not None:
                         context["results"] = exec_env["output"]
+                        print("result from code execution: ", exec_env["output"])
+
             except NameError as e:
                 tb_lines = traceback.format_exc().splitlines()
                 limited_tb = (

npcsh/npc_team/tools/bash_executer.tool

@@ -0,0 +1,32 @@
+tool_name: bash_executor
+description: Execute bash queries.
+inputs:
+  - bash_command
+  - user_request
+steps:
+  - engine: python
+    code: |
+      import subprocess
+      import os
+      cmd = '{{bash_command}}'  # Properly quote the command input
+      def run_command(cmd):
+        process = subprocess.Popen(cmd, shell=True, stdout=subprocess.PIPE, stderr=subprocess.PIPE)
+        stdout, stderr = process.communicate()
+        if stderr:
+          print(f"Error: {stderr.decode('utf-8')}")
+          return stderr
+        return stdout
+      result = run_command(cmd)
+      output = result.decode('utf-8')
+
+  - engine: natural
+    code: |
+
+      Here is the result of the bash command:
+      ```
+      {{ output }}
+      ```
+      This was the original user request: {{ user_request }}
+
+      Please provide a response accordingly.
+

npcsh/npc_team/tools/code_executor.tool

@@ -0,0 +1,16 @@
+tool_name: code_executor
+description: Execute scripts with a specified language. choose from python, bash, R, or javascript. Set the ultimate result as the "output" variable. It must be a string. Do not add unnecessary print statements.
+inputs:
+  - code
+  - language
+steps:
+  - engine: '{{ language }}'
+    code: |
+      {{code}}
+  - engine: natural
+    code: |
+      Here is the result of the code execution that an agent ran.
+      ```
+      {{ output }}
+      ```
+      please provide a response accordingly.

npcsh/npc_team/tools/npcsh_executor.tool

@@ -0,0 +1,9 @@
+tool_name: npcsh_executor
+description: Execute npcsh commands. Use the macro commands.
+inputs:
+  - code
+  - language
+steps:
+  - engine: "{{language}}"
+    code: |
+      {{code}}

npcsh/npc_team/tools/sql_executor.tool

@@ -1,5 +1,5 @@
-tool_name: sql_executor
-description: Execute SQL queries on the ~/npcsh_history.db and display the result. The database contains only information about conversations and other user-provided data. It does not store any information about individual files.
+tool_name: data_pull
+description: Execute queries on the ~/npcsh_history.db to pull data. The database contains only information about conversations and other user-provided data. It does not store any information about individual files.
 inputs:
   - sql_query
   - interpret: false  # Note that this is not a boolean, but a string

npcsh/shell.py

@@ -378,6 +378,7 @@ def main() -> None:
 
                 current_npc = result["current_npc"]
             output = result.get("output")
+
             conversation_id = result.get("conversation_id")
             model = result.get("model")
             provider = result.get("provider")
@@ -404,17 +405,16 @@ def main() -> None:
                 npc=npc_name,
                 attachments=attachments,
             )
-            if NPCSH_STREAM_OUTPUT and (
-                isgenerator(output)
-                or (hasattr(output, "__iter__") and hasattr(output, "__next__"))
-            ):
-                str_output = ""
+
+            str_output = ""
+            if NPCSH_STREAM_OUTPUT and hasattr(output, "__iter__"):
+
                 buffer = ""
                 in_code = False
                 code_buffer = ""
 
                 for chunk in output:
-                    # Get chunk content based on provider
+
                     if provider == "anthropic":
                         chunk_content = (
                             chunk.delta.text
@@ -434,7 +434,7 @@ def main() -> None:
                         continue
 
                     str_output += chunk_content
-
+                    # print(str_output, "str_output")
                     # Process the content character by character
                     for char in chunk_content:
                         buffer += char
@@ -490,16 +490,18 @@ def main() -> None:
                 if str_output:
                     output = str_output
             print("\n")
-            save_conversation_message(
-                command_history,
-                conversation_id,
-                "assistant",
-                output,
-                wd=current_path,
-                model=model,
-                provider=provider,
-                npc=npc_name,
-            )
+
+            if isinstance(output, str):
+                save_conversation_message(
+                    command_history,
+                    conversation_id,
+                    "assistant",
+                    output,
+                    wd=current_path,
+                    model=model,
+                    provider=provider,
+                    npc=npc_name,
+                )
 
             # if there are attachments in most recent user sent message, save them
             # save_attachment_to_message(command_history, message_id, # file_path, attachment_name, attachment_type)