Show HN: Whiz – A copilot for your terminal (opens in new tab)

I loved Warp for it's speed, but the keybindings and lack of configuration didn't work for me. For example, it doesn't support "ctrl+x ctrl+e" which I use daily for editing long prompts in vim.

I recently switched to Kitty (https://sw.kovidgoyal.net/kitty/) and I don't think I'm ever going back. It's not flashy and doesn't have any AI features. It's just a donation-funded, wickedly fast, highly configurable, dotfiles-friendly, modern terminal emulator.

Again, business model: Juicero for bash.

Collecting telemetry, and being able to tie it with people's GitHub profiles, is why Warp was able to secure millions of VC dollars to build a fucking mOdErN terminal emulator.

This seems to require joining a waitlist to gain access.

Highly recommend it, too. I've found it invaluable. Especially for some of the more arcane ffmpeg incantations.

Meh. I am fairly certain the number of people who would both use this and also fail to consider possible use cases half a second into the ls demo is zero.

Can it create regexes? That would be the thing that would sell it to me via a demo.

I think the point of the demo is to start with something dead simple that "everyone" knows and get more complex with each example.

I can reproduce this.

The issue is ChatGPT hallucinating a function that doesn't exist. I'm figuring out how to rail guard against this. Thanks for trying it out and reporting with an example.

If I switch to ChatGPT 4, it works fine.

Pretty easy to see in the code that it is not sending any of that to OpenAI. It's only sending the prompt, OS info, and shell type:

https://github.com/searchableguy/whiz/blob/main/src/commands...

https://github.com/searchableguy/whiz/blob/main/src/completi...

That said, OP, it's perhaps worth making this extremely clear in the README as it's a reasonable concern.

And I love just typing a key combination to get the suggestion, almost like I can change my mind during the command...

I was able to understand the Whiz source (~200 lines) in about sixty seconds, which is a big win in my opinion. The shell_gpt code is more convoluted. Would be interesting to compare how they perform.

That's a cool project.

I built this after copy pasting the same git command to list all the recently committed branches from chat.openai.com.

I have a couple more functions that I want to add to the CLI like [0] translating novels, web scraping or LLM powered grep.

0] https://gist.github.com/searchableguy/436933aa9f95a423d2f89d...

I was wondering if Whiz was using agent-style planning prompts, but it's much simpler and passes a single `function` parameter named `shell` to the OpenAI API.

I thought that was clever, because shell commands are highly composable, so it's likely that something can be done in "one shot".

It will also be way more efficient with token usage.

I love yours. It is entirely in shell script.

I recommend using chatgpt functions for more reliability. I'm making use of them[0]. I wanted to use javascript to extend it with a plugin system, haha.

0] https://github.com/searchableguy/whiz/blob/main/src/completi...

Looking at the code, the system prompt says "You MUST NOT use functions that are not available." but apart from the OS info, it's not giving GPT any info on what commands are/aren't available, so I'd imagine it's a bit of a crapshoot.

We need good locally installed LLMs (and cheap hardware to run it). I hope there can be some kind of breakthrough for this similar to what Stable Diffusion made for image-generators. I tried to generate some simple code using a few of the llama-models small enough to run on my computer and it did surprisingly well, but still far from good enough to be useful.

It appears whiz is foss, while fig is not.