codesoda

I've posted a full breakdown over here https://news.ycombinator.com/item?id=48421328 (including the source code and tracking changes they make in a changelog.md)

It uses their llm, which is at https://paxel-llm.ycombinator.com.

It's an Anthropic SDK based LLM proxy, though they use GPT5.5 to analyze the transcripts.

Basically it mounts your home directory, then reads your coding agent sessions from .claude, .codex etc. then uses their LLM-proxy to score the agent sessions against a rubric at https://github.com/codesoda/paxel/blob/main/rootfs/rails/db/..., to then score you out of 10.

It's a bit of a bias of my own. It's all local, agents seem very good at running and monitoring cli outputs. They know how to follow standards, e.g. `cli-name help` to work out how to use it. The cli is self documenting this way, and they only need to learn about the commands they want to use.

No reason MCP support couldn't be added.

I hadn't seen that one. Looks great also.

When I built this I specifically didn't want to have to provide feedback and then submit it back, then open again after the agent had considered my feedback.

The way discuss is setup, it's interacting forwards and backwards on the discussion, from the web page, but within the original agent session I was having.

Very convenient to ask a question "in the document" and get a response a few seconds later, in the same document.