User Action Node

User Action Node#

Overview#

UserActionNode provides user interaction capabilities for supervising and controlling workflow execution. It currently handles proceed, feedback, edit, resample, and record operations.

Its key features include:

  • User Actions Selection

    • Proceed (p): action == "p" - Continue workflow execution

    • Feedback (f): action == "f" - Provide feedback to LLM

    • Edit (e): action == "e" - Edit specific step content

    • Resample (n): action == "n" - Regenerate step content

    • Record (r): action == "r" - Record turning points

  • Command Processing

    • Subtask commands: cmd_type == "subtask" - Opens substack with data.open_substack()

    • Context commands: cmd_type == "context" - Executes Python code for context modification

    • Terminal commands: Executes terminal commands via TerminalNode

Context Commands#

  • Summarize: summarize(start_idx, end_idx, summary) - Summarizes conversation segments

  • Dynamic execution: Uses exec(cmd, {}, available_objects) for context modification

Edit Functionality#

def _edit(self, message: Message, data: ContextData, resample=False):
    if resample:
        new_content = ""
    else:
        # Open editor with temp file
        with tempfile.NamedTemporaryFile(mode="w+", delete=False) as f:
            f.write(message.content)
        os.system(f"{self.editor} {f.name}")
        # Read edited content

    if new_content != message.content:
        data.branch_at(message.step - 1)  # Create branch
        extra_info = EditExtraInfo(edit_prefix=new_content, content_before_edit=message.content)
        return self.llm_node, extra_info

Implementation#

def run(self, data: ContextData, extra_info: Optional[ExtraInfo] = None):
    if isinstance(extra_info, DirectActionExtraInfo):
        action = extra_info.action
        target_step = extra_info.target_step
    else:
        action, target_step = self._user_input(data)

    if action == "f":
        return self._feedback(data)
    elif action == "e":
        target_message = data.current_branch[target_step]
        return self._edit(target_message, data)
    elif action == "n":
        target_message = data.current_branch[target_step]
        return self._edit(target_message, data, resample=True)
    elif action == "r":
        return self._record(data)
    elif action == "p":
        return self._proceed(data)
Contents