New in v8.6.0 - ✨Improved Agentic Workflows

[olimorris]

I've been passionate about agentic workflows for some time. Giving the LLM the ability to self-reflect leads to some interesting results.

In this release I've overhauled how the plugin handles agentic workflows...infact you can create them yourself by adding them to the prompt library (check out the RECIPES and the config.lua file)

Workflows.mp4

In this video, you'll notice that my next prompts are auto-populated (they could also be auto-submitted but I like to press that <return> key myself). The wording of these prompts also give the LLM the opportunity to self-reflect and revise its answer.

I'm keen to implement some of the features that Aider has in the coming weeks and I think workflows will be a key part of this.

Please play around with workflows and share yours with the wider community.

[e2r2fx]

@olimorris see you're playing with the @img tool thing. I can submit that as a PR np, the downside is that its hardcoded to work with OpenAI's vision API with a curl command, but other than that I find it really useful on my end

[olimorris]

Ahhh I was just testing custom tooling. I would love to see your img implementation though

[e2r2fx]

sure, here's a gist

The way I use it is quite rudimentary, I just tell the llm to use the @img tool and do X with it, and it builds a query for the vision API.

For providing an image from a screenshot I use a quick n dirty plugin I wrote some time ago
https://github.com/e2r2fx/cloudy.nvim

[e2r2fx]

I like better the approach of locally base64 encoding the image and sending that to the vision API. But since I had that plugin laying around... it gets the job done :)