Nice work on the local-first approach. One thing worth considering as you scale this: session persistence becomes critical once users run long agentic tasks. The painful failure mode is losing context mid-run during a network hiccup or browser refresh. Teams doing serious Claude Code work often need instant output replay and the ability to resume conversations exactly where they left off, not restart from scratch. That architectural decision early saves enormous frustration later.
Thank you for the insight! I think the project is in a good place with persistence. Chats are threaded, history is stored in JSON on disk + VectorDB for RAG. I have seen issues with long tool runs timing out on the client side, but complete through the LLM server side, just needs a page refresh to show the latest output, which could be a reverse proxy issue. But enhancing the workflow of the agent is high on my todo list!
The timeout issue sounds like it's probably a proxy buffer or keepalive setting. If you're running nginx, bumping proxy_read_timeout and enabling chunked transfer helps a lot for long-running streams. SSE or websockets can also help surface partial progress rather than making users wait for a full refresh. Good call prioritizing the agent workflow - that's usually where the real friction lives once persistence is solid.
Nice work on the local-first approach. One thing worth considering as you scale this: session persistence becomes critical once users run long agentic tasks. The painful failure mode is losing context mid-run during a network hiccup or browser refresh. Teams doing serious Claude Code work often need instant output replay and the ability to resume conversations exactly where they left off, not restart from scratch. That architectural decision early saves enormous frustration later.
Thank you for the insight! I think the project is in a good place with persistence. Chats are threaded, history is stored in JSON on disk + VectorDB for RAG. I have seen issues with long tool runs timing out on the client side, but complete through the LLM server side, just needs a page refresh to show the latest output, which could be a reverse proxy issue. But enhancing the workflow of the agent is high on my todo list!
The timeout issue sounds like it's probably a proxy buffer or keepalive setting. If you're running nginx, bumping proxy_read_timeout and enabling chunked transfer helps a lot for long-running streams. SSE or websockets can also help surface partial progress rather than making users wait for a full refresh. Good call prioritizing the agent workflow - that's usually where the real friction lives once persistence is solid.