Add interactive shell mode

[aki-colt]

Created a new mode called interactive shell, which inherits all the features of interactive mode and can execute shell commands.
Implementation:

• Use system prompt to force LLM generate commands wrapped in XML <tgpt_command> tags.
• Stream the natural language output except for the commands. Finally, ask whether to execute the command (of course, automatic execution is also supported).
• If the execution is successful, interaction can continue; if it fails, it will exit (inheriting the characteristics of shell mode).

[aandrew-me]

Let it just be <cmd>

[aandrew-me]

Looks good

Phind likely doesn't follow instructions properly

[aandrew-me]

Great job!
Just change to and I will merge

[aki-colt]

Thanks for review. Sometimes, it does fail to follow instructions. Perhaps adding some self-correction would be a good solution, where the program prompts the large model to regenerate if it detects incomplete tags. Additionally, I noticed that commands like ping or similar ones that don't self-terminate cannot be stopped in interactive-shell mode. I think this is a bug, and I’ll try to fix it.

[aki-colt]

I’ll first push a commit to fix the issue where commands cannot be stopped. Currently, using Ctrl+C during command execution stops both the command and tgpt. I tried stopping only the command while continuing the conversation, but it caused strange errors. I’ll keep trying, but I don’t think this is a high priority. Additionally, adding self-correction is quite complex under the current code structure. The LLM's error rate is not high, and users can continue inputting to correct the model. So, I think self-correction may not be necessary.

[aandrew-me]

I am not able to merge because of conflicts

[aki-colt]

I didn't see conflicts between main branch and mine 🤔.

[aandrew-me]

All good, thanks for your work!

[aandrew-me]

What do you think about implementing this using mcp?