Ollama feature-parity for ACP

[digitalextremist]

My target would be to do this in TypeScript versus Rust and rely on Bun rather than node from the start.

Required:

• Indexer ( Unsaved file synchronisation #50 touches ) which ought be git based, not file-system or open-tab
  • This keeps a vector-database up to date, and likely would hook-up to git smoothly to update on changes.
• Server seems to be a wrapper around Ollama more or less? Kind of a positive MITM which is a ducktype+RAG.
  • Retriever grabs and inserts augmentations while Server holds the phone of requests from Zed waiting for embeds.
  • Relay integrates with Ollama via its client package; delivers embeds along surface activity from Zed or other requests.

Parting thought:

If code modification is happening now, then there would also be a Commiter that makes the appropriate changes and essentially constructs the parts of a PR, then sends one. But that seems out of scope with the first step of ACP implementation at all.

---

Does that sound wrong? Am I missing something?

[benbrandt]

@digitalextremist this seems to be a lot of discussion rolled into one. In terms of ollama, if you wanted to create an Agent on top of ollama, this could use ACP, but doesn't feel like it needs to be an ACP concern directly. Am I missing something?

[digitalextremist]

Trimmed down the post ~90% to clarify!

[benbrandt]

Thanks, but I am still a bit confused. Apart from git vs fs, it seems like this is mostly around using ollama to keep a vector db up to date on the codebase, which to me at least, seems like a particular implementation detail of a given agent. i.e. how is the code chunked for the embedding? When are the embeddings used? which embedding model?

I feel like most agent authors would want to control these variables, which to me pushes this from being an ACP concern to that of the agent (though maybe there is something at the protocol level we would need to supply to make this happen). Am I missing something?

[digitalextremist]

The good news that I hear here is that ACP is likely easier to implement than I think.

I am going back through the reference implementation and refactoring my interpretation of ACP itself and will move forward on this until I hit a wall, to try and understand by doing... time allowing.

A jump-off point in the meantime is that the reference implementation immediately hits upon npx being in-path, which I keep running into in the editor itself. A key to my portion is Bun which might be out of scope for SCP but still falls within Zed because it seems node primacy was cherry-picked and hard-coded.

[ For example, my linter is always on the fritz, all my environments set for Bun in Typescript projects behave weirdly, but just enough to work... plus performance is subpar with node by a lot. ]

I feel like understanding ACP now and getting smoothly into that is essential so I am extremely grateful @benbrandt and do not mind haircutting my */OSS enthusiasm to focus on ACP directly.

Will be re-reading this 10x: https://github.com/zed-industries/agent-client-protocol/tree/main/typescript/examples