/add-interactive-chat

Interactive Chat Console Generator

Generate a chat.py file adapted to the user's agent for interactive testing with tool call logging, thread management, and optional context injection.

Workflow

Step 1: Detect Agent

Locate the agent creation function in the project.

If module path provided in arguments:

• Use the provided path directly
• Read the file to extract the function name and detect features

If no path provided:

• Search for create_deep_agent, create_react_agent, or def create_.*agent patterns across the project using Grep
• If multiple matches found: present them to the user with AskUserQuestion and ask which one to use
• If no matches found: ask the user for the module path and function name using AskUserQuestion
• If one match found: use it directly

Extract from the detected agent file:

• agent_module: the Python import path (e.g., src.myproject.agent)
• agent_function: the function name (e.g., create_agent)
• has_context_schema: whether the agent uses a context_schema parameter
• has_checkpointer: whether the agent already creates a MemorySaver or other checkpointer
• has_interrupt_on: whether the agent uses interrupt_on for HITL
• has_subagents: whether the agent defines subagents= dicts

Step 2: Determine Features

Ask the user using AskUserQuestion about desired features:

1. Context injection (only if context_schema detected): "Your agent uses a context_schema. Do you want the chat console to support user context switching? (yes/no)"
2. Verbose logging: "Show full tool parameters or summary only? (full/summary)"
3. Multi-user: "Support switching between test user IDs during the session? (yes/no)"

Step 3: Generate chat.py

Write the chat.py file at the project root, adapted to the detected agent. Use the following template, applying the adaptations described below.

"""Interactive chat console for testing your agent."""
import uuid
import sys

from langgraph.checkpoint.memory import MemorySaver  # noqa: used if agent needs checkpointer

# --- Agent import (adapted from detection) ---
from {agent_module} import {agent_function}


def format_tool_call(tool_call: dict, verbose: bool = {verbose}) -> str:
    """Format a tool call for display."""
    name = tool_call["name"]
    args = tool_call.get("args", {})
    if verbose:
        args_str = ", ".join(f"{k}={v!r}" for k, v in args.items())
        return f"  \033[36mTool:\033[0m {name}({args_str})"
    else:
        args_summary = ", ".join(args.keys()) if args else "no args"
        return f"  \033[36mTool:\033[0m {name}({args_summary})"


def format_tool_response(msg) -> str:
    """Format tool response summary."""
    content = str(msg.content) if hasattr(msg, "content") else str(msg)
    if len(content) > 200:
        content = content[:200] + "..."
    return f"  \033[33mResult:\033[0m {content}"


def main():
    # --- Agent setup ---
    agent = {agent_function}()

    # --- Thread management ---
    thread_id = str(uuid.uuid4())
    config = {"configurable": {"thread_id": thread_id}}

    # {context_setup_block}
    # If context_schema detected:
    #   user_id = "test-user-1"
    #   context = {context_factory}
    #   (included in config or invoke)

    # {user_id_block}
    # If multi-user enabled:
    #   user_id = "test-user-1"

    print(f"\n\033[1mAgent ready.\033[0m Thread: {thread_id[:8]}...")
    print("Commands: 'exit' | 'new' (new thread){user_commands_hint}")
    print("-" * 50)

    while True:
        try:
            user_input = input("\n\033[1mYou:\033[0m ").strip()
        except (EOFError, KeyboardInterrupt):
            print("\nBye!")
            break

        if not user_input:
            continue
        if user_input.lower() in ("exit", "quit", "salir"):
            print("Bye!")
            break
        if user_input.lower() in ("new", "nuevo"):
            thread_id = str(uuid.uuid4())
            config["configurable"]["thread_id"] = thread_id
            print(f"  \033[32mNew thread:\033[0m {thread_id[:8]}...")
            continue

        # {user_switch_block}
        # If multi-user enabled:
        #   if user_input.lower().startswith("user "):
        #       user_id = user_input.split(" ", 1)[1].strip()
        #       print(f"  Switched to user: {user_id}")
        #       continue

        # {context_switch_block}
        # If context injection enabled:
        #   if user_input.lower().startswith("context "):
        #       ... update context ...
        #       continue

        try:
            result = agent.invoke(
                {"messages": [{"role": "user", "content": user_input}]},
                config=config,
                # {context_invoke} — if context_schema: context=context
            )

            # Log tool calls and responses
            for msg in result["messages"]:
                if hasattr(msg, "tool_calls") and msg.tool_calls:
                    for tc in msg.tool_calls:
                        print(format_tool_call(tc))
                if hasattr(msg, "name") and msg.name:
                    print(format_tool_response(msg))

            # Final agent response
            final = result["messages"][-1].content
            print(f"\n\033[1mAgent:\033[0m {final}")

        except KeyboardInterrupt:
            print("\n  (interrupted)")
            continue
        except Exception as e:
            print(f"\n  \033[31mError:\033[0m {e}")


if __name__ == "__main__":
    main()

Adaptations

Apply the following adaptations when generating the final file:

With context_schema detected and context injection enabled:

• Add context creation block after agent setup:

  user_id = "test-user-1"
  context = {"user_id": user_id}  # Adapt to actual context_schema fields
  

• Add context=context to agent.invoke()
• Add a context <json> command for live context switching:

  if user_input.lower().startswith("context "):
      import json
      try:
          context = json.loads(user_input.split(" ", 1)[1])
          print(f"  Context updated: {context}")
      except json.JSONDecodeError:
          print("  Invalid JSON. Use: context {\"key\": \"value\"}")
      continue
  

Without checkpointer in agent:

• Wrap the agent creation to inject a MemorySaver:

  from langgraph.checkpoint.memory import MemorySaver  # noqa: used if agent needs checkpointer
  
  # Agent does not include a checkpointer; adding MemorySaver for thread persistence
  agent = {agent_function}(checkpointer=MemorySaver())
  

  Or if the function doesn't accept a checkpointer parameter, note this to the user and add MemorySaver inline if possible.

Multi-user enabled:

• Add user_id variable initialization
• Add the user <id> command handler:

  if user_input.lower().startswith("user "):
      user_id = user_input.split(" ", 1)[1].strip()
      config["configurable"]["user_id"] = user_id
      thread_id = str(uuid.uuid4())
      config["configurable"]["thread_id"] = thread_id
      print(f"  Switched to user: {user_id} (new thread: {thread_id[:8]}...)")
      continue
  

• Update the commands hint to include | 'user <id>'

Verbose vs summary logging:

• Set the verbose default in format_tool_call based on user preference (True for full, False for summary)

Step 4: Verify

After writing the file:

1. Read back chat.py to confirm it was written correctly
2. Print the file path to the user
3. Provide run instructions:

# Run the chat console
python chat.py

# Or with a specific module if needed
python -m chat

Step 5: Usage Instructions

Show the user what to expect:

$ python chat.py

Agent ready. Thread: a1b2c3d4...
Commands: 'exit' | 'new' (new thread) | 'user <id>'
--------------------------------------------------

You: What's my balance?
  Tool: get_account_balances(include_details=False)
  Result: 2 accounts found

Agent: You have 2 accounts. Your checking account has $1,234.56 and your savings has $5,678.90.

You: new
  New thread: e5f6g7h8...

You: exit
Bye!

Step 6: Next Steps

Suggest to the user:

• Use /validate-agent to check the agent for anti-patterns before testing
• Enable LangSmith tracing (LANGSMITH_TRACING=true) for full observability during chat sessions
• Add custom commands to the chat loop for agent-specific testing scenarios