Feature: forking conversations

[jgarvin]

The web UIs for most of the popular models now let you edit earlier answers, which then deletes everything that came after and lets you pretend the conversation went a different way. This is very useful for conserving tokens, you often go down an unhelpful path and realize you need to back up several prompts to edit something. comint lets you go to earlier prompts, and you can type on them, but when you hit enter it will repeat the original version of the prompt from that line, not the edited one. Example:

First I just say "testing":

Claude(4-opus/General)> testing
<shell-maker-end-of-prompt>
Hello! I'm here to help you with programming questions, debugging, code explanations, or any other technical topics you'd like to explore. 

What would you like to test or work on today? Here are some things I can help with:

- **Code Review**: Share code for feedback and improvements
- **Debugging**: Help track down and fix bugs
- **Algorithm Design**: Work through problem-solving approaches
- **Language-Specific Questions**: Python, JavaScript, Java, C++, and more
- **System Design**: Discuss architecture and design patterns
- **Testing Strategies**: Unit tests, integration tests, TDD approaches

Feel free to share your code or describe what you're working on!

Then I go back to edit the answer and hit enter:

Claude(4-opus/General)> testing can I change this?
<shell-maker-end-of-prompt>
Hello! I'm here to help you with programming questions, debugging, code explanations, or any other technical topics you'd like to explore. 

What would you like to test or work on today? Here are some things I can help with:

- **Code Review**: Share code for feedback and improvements
- **Debugging**: Help track down and fix bugs
- **Algorithm Design**: Work through problem-solving approaches
- **Language-Specific Questions**: Python, JavaScript, Java, C++, and more
- **System Design**: Discuss architecture and design patterns
- **Testing Strategies**: Unit tests, integration tests, TDD approaches

Feel free to share your code or describe what you're working on!

Claude(4-opus/General)> testing
<shell-maker-end-of-prompt>
I see you're testing! Yes, you can interact with me and I'll respond to help with your programming needs. 

Here's a quick example to show I'm working - let's say you wanted to test a simple function:

```python
def greet(name):
    """A simple greeting function"""
    return f"Hello, {name}! Welcome to the coding session."

# Test it out
print(greet("Developer"))
# Output: Hello, Developer! Welcome to the coding session.

Feel free to:
- **Ask specific questions** about code
- **Share code snippets** for review
- **Request examples** in any programming language
- **Explore concepts** you're learning

What would you like to test or work on specifically?

As you can see it just repeated the "testing" prompt

[xenodium]

Thanks for filing! If I'm understanding correctly, there are a couple of features to consider here...

1. Forking chats
2. Rewriting last request

Forking chats

I've been considering a command M-x chatgpt-shell-fork-chat-at-point, which starts a new chat including interactions up to point location.

Rewriting last request

For this one, I've been considering enabling you to rewrite the last submission and imediately delete it from history after new one is submitted. The rationale here is to make it easier to understand what is actually included in the context/history.

In the mean time

While I've been considering the above, I kinda haven't done them as I've gotten away it with a couple of things that probably aren't that discoverable.

M-x chatgpt-shell-delete-interaction-at-point

Hope it conveys that it does on the tin. This enables you to wipe any pair (request and response) from the context/history.

Emacs undo!

This one is kinda a hidden gem. The shell buffer is, well just another buffer, so I often undo (and sometimes even redo) the last interaction (sometimes multiple interactions), and so on.

ps. Thanks for sponsoring! Much appreciated.