Render markdown output about shell_gpt

[TheR1D]

It is possible to pipe output of sgpt to glow or any other markdown reader (which supports stdin).

sgpt "how to zip all files in current folder" | glow

[jeanlucthumm]

I found the DISABLE_STREAMING option and at the very least, it's possible to output markdown if that's enabled:

output.mov

Code: jeanlucthumm@eed0f7e

I'm not sure how I feel about the long waiting times though.

Other than that, I think you would need a full screen TUI to pull this off because you have to clear and repain the markdown file as words come in, but only from the current line down. The only thing I found was clearning the entire screen using console.clear(), and while that works, you lose all your history.

To deal with the streaming, one option is to detect beginning of special blocks, like a code block, and then automatically append a closing statement until we get the actual one from the llm. We'd have to identify all such blocks.

^ This isn't actually neccessary, it turns out that Markdown from rich still displays fine if there's no closing backicks fo a code block

[volkov]

True, at least for latest glow versions (1.4.0 does not support stdin).

Piping to glow makes outupt more readable, but I have to wait till end of generation to see anything.

[David-Else]

How about https://github.com/swsnr/mdcat for full output, or https://github.com/sharkdp/bat to get some colour highlighting? (bat is batcat in Debian)

This works, but for some reason there is no colour, do I need some --color type option?:

sgpt "show me some markdown that contains all markdown features" | batcat

[jeanlucthumm]

https://rich.readthedocs.io/en/stable/markdown.html

^ python lib for markdown rendering with code syntax highlighting, and it looks like it's already a dependency of this project:

shell_gpt/pyproject.toml

Line 34 in 1c58566

"rich >= 13.1.0, < 14.0.0",

A lot of my use case is with the --repl flag so piping is not an option

[jeanlucthumm]

The line that prints messages seems to be this one:

shell_gpt/sgpt/handlers/handler.py

Line 34 in 1c58566

typer.secho(word, fg=self.color, bold=True, nl=False)

Gonna try to change it to the Rich markdown handler and see what happens

I guess one issue is that the words stream in, and Rich requires a full text ready?

[jeanlucthumm]

To deal with the streaming, one option is to detect beginning of special blocks, like a code block, and then automatically append a closing statement until we get the actual one from the llm. We'd have to identify all such blocks.

Another consideration is how slow the markdown render is. It has to fully complete for every word streamed in which is multiple times per second.

[julian-hoch]

I use this alias:

sgpt_glow_replace() {
tmpfile=$(mktemp)
sgpt "$@" | tee $tmpfile
clear
cat $tmpfile | glow -
rm $tmpfile
}
alias s='sgpt_glow_replace'

[gwpl]

sounds like invitaiton for --glow and --no-glow options! (and default in config file "DEFAULT_GLOW=true|false"