Add support to run commands that are written by ai in the chat.

[Autonvilkku]

feature:

• Specially in multiline mode tgpt would ask if the user does he want to run the command if the ai has written an command.

usecases

• In multiline mode this would help a lot, becouse the user doesn't have to exit the conversation.
• Also in the normal mode this would help, so the user doesn't have to copy and paste it, tgpt would just ask if the user wants to execute that command.

example how it would work

user: write a command that make a folder and its name is "folder1"
bot: mkdir folder1
tget: do you want to execute that command(Y,n):

[ekkinox]

Tgpt is a great project!

But indeed it doesn't offer the possibility to execute generated commands.

You can check Yai for this, but it's only working with Open AI for the moment (I have no time right now to make it evolve).

If it is a direction @aandrew-me wants to take for tgpt, maybe we should make it able to run commands reusing what was done in Yai ?

Yai: https://github.com/ekkinox/yai

[aandrew-me]

I will take a look

[aki-colt]

I would like to work on this enhancement. I think implementing it with Tool Call is a good approach, but not all models support this feature. Perhaps having the model generate structured data like JSON or XML would be better. What do you think? @aandrew-me

[aandrew-me]

I would like to work on this enhancement. I think implementing it with Tool Call is a good approach, but not all models support this feature. Perhaps having the model generate structured data like JSON or XML would be better. What do you think? @aandrew-me

Earlier providers in tgpt didn't follow instructions properly, but I have tried recently and now they do obey the commands. I have tested the json approach and you can check the prompts here https://github.com/ekkinox/yai/blob/main/ai/engine.go

I like the approach. If you want to work on it, thats great. Create a new interactive mode for that, and no need to stream text, since the api will be sending in json format. Also if possible, its better to pass the instructions as system prompt. So it should be implemented separately for each provider

[aki-colt]

Thank you for your suggestion, but I think we can still use streaming text. We can have the model return both the response and the action to execute a command, distinguished by a special marker, such as XML. The following is an example response for listing files.

Yes, I would use 'ls' command to help you list files.
<tgpt_run_command>
ls -l
</tgpt_run_command>

This way, by parsing the XML portion of the model's response (using regular expressions), we can simultaneously stream the model's thought process and confirm with the user whether to execute the command. We can use prompts to limit the model to outputting only one command at a time, and after user confirmation, feed the execution result back into the model's context, achieving an effect similar to a tool call.
This allows users to see the model's response immediately without having to wait for the model to generate a complete reply. And I think this approach requires less from the model's ability to follow instructions compared to having it generate complete and valid JSON.

[aandrew-me]

Thank you for your suggestion, but I think we can still use streaming text. We can have the model return both the response and the action to execute a command, distinguished by a special marker, such as XML. The following is an example response for listing files.

Yes, I would use 'ls' command to help you list files.
<tgpt_run_command>
ls -l
</tgpt_run_command>

This way, by parsing the XML portion of the model's response (using regular expressions), we can simultaneously stream the model's thought process and confirm with the user whether to execute the command. We can use prompts to limit the model to outputting only one command at a time, and after user confirmation, feed the execution result back into the model's context, achieving an effect similar to a tool call. This allows users to see the model's response immediately without having to wait for the model to generate a complete reply. And I think this approach requires less from the model's ability to follow instructions compared to having it generate complete and valid JSON.

Sure! As long as it works, all good

[aki-colt]

I've made some progress. I defined a new mode called interactive shell, and as shown in the demo video below, it can execute commands during a conversation. I tested most of the providers. However, I still encountered some issues. I couldn't successfully register with groq so I haven't tested it. Isou and Koboldai seem unable to support passing more context, so I didn't include support for them. Does this meet our expectations now? If so, a PR is on the way.

tgpt-demo.mp4

[aandrew-me]

Looks good! Create a PR and I will test

[aki-colt]

Great, pushed a PR #363 to the main branch.

[Autonvilkku]

Thanks for this feature. This is why I love open source. la 10. toukok. 2025 klo 21.09 Andrew ***@***.***> kirjoitti:

…

Closed #281 <#281> as completed. — Reply to this email directly, view it on GitHub <#281 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BCTJV44L7HLS6TUKQVCQHDT25Y6FBAVCNFSM6AAAAAB3V6JEC6VHI2DSMVQWIX3LMV45UABCJFZXG5LFIV3GK3TUJZXXI2LGNFRWC5DJN5XDWMJXGYYDGMBZHE3TQMI> . You are receiving this because you authored the thread.Message ID: ***@***.***>