There's also ELK that I found gave the best results. Checkout some diagrams generated here - https://arcaide.foo/app/projects

Which ones of these are generated by ELK? All of them?

Yes, they all are.

I keep going back to xfce-terminal after trying anything else. Ridiculously small footprint and never drew attention to itself.

Not at this time, but down the line I probably will.

They're generated from source code using a combination of static analysis + llm analysis. They're not editable per se, but you can update them when you edit your code.

I think you found a very good prompt/technique to extract mostly the high level stuff without making the diagrams too complicated. And then elk.js seems to do the layout.

Deepwiki.com also has some usable autogenerated architecture diagrams worth a look.

Something I do have in mind. Do you have an idea of the kind of UX you'd like to see?

I would be interested in this too if you ever build it.

Only naïve ones I'm afraid, such as a heap view.

It identifies telemetry, accessors, and other trivial code and strips them out of the graph. The idea is to retain architecturally relevant parts of the code.

What are the memory requirements for pgGraph like? Does the entire graph for my dataset need to fit in memory?

You can try it out, whats in memory is a compressed mathematical map of your data, typically it uses about 30-34x less memory than the total dataset size. Currently the only ram hungry part is during the graph build (which we're working on lowering too.)

You might be interested to look up the Williamson-Van der Mark model. It says that an electron is just light that loops in on itself and shows how properties like mass, charge and spin can emerge from such a geometry. It's fascinating because it opens up the question whether every particle could be described in terms of different ways light can loop in on itself.

Talking to insurance agents I realised, they don't bother to read the policy documents and have a very superficial knowledge of the policies they are selling. You can glean lot more information feeding the docs to an LLM and asking questions.

How often do they sell to someone interested in the details?

LSP is not great for non-editor use cases. Everything is cursor position oriented.

Yes, something like TreeSitter would seem to be of more value - able to lookup symbols by name, and find the spans of source code where they are defined and used.

I don't see ast-grep as being very useful to an agent.

What a coding agent needs is to be able to locate portions of source code relevant to what it has been tasked with, and preferably in more context-efficient fashion than just grepping and loading entire source files into context. One way to do this is something like Cursor's vector index of code chunks, and another would be something like TreeSitter (or other identifier-based tools) that knows where identifiers (variables, functions) are defined and used.

Language servers (LSP) are not useful for this task since they can't tell the agent "where is function foo() defined" (but TreeSitter can), since as someone else noted language servers are based on location (line number) not content (symbols). Language servers are designed to help editors.

It's possible that ast-grep might be some some use to a coding agent, but looking for syntax/AST patterns rather than just identifier definitions and usages seems a much more niche facility.

There are actions that don't require cursor position, like document/workspace symbols, that could be useful.