The impending doom of cursor, claude, and codex

[ThePrimeagen]

So i want to make this library even more robust and i think i have something that is getting real exciting. The problem is that i know opencode. Therefore this large upcoming upgrade is going to be a super charge for opencode, but its going to leave the other libraries in the dust and make them non useable.

with that in mind, I am going to be branching here soon and have v1 be a frozen target and master is going to move on.

Here is the change:
opencode provider is going to use a different mode where we can reattach to sessions. This allows a long conversation to happen, not just a once shot, and call it done.

so for any of the providers to be able to join in the new format, they have to

1. keep track of the stdout
2. assign conversations ids
3. be able to resume conversations
4. keep track of the history of the conversation

this is a pretty ambitious feature. Ill probably start it here in about 1 week.

---

Motivation: I still like to vibe out certain features, so if i could just reattach to sessions / review sessions / send partial hunks back and say "fix this", that would be amazing

[diachkow]

Hi, have you explored Agent Communication Protocol?

JetBrains IDEs, ZED and many other projects are using ACP implementations for Claude Code/Codex/Opencode etc. Of course, all these implementations now are mostly community-driven and the protocol itself have limited amount of exposed methods, but it is still could be nice to work with all of these tools through unified interface. This will also potentially enable people to use 99 with other tools that implement ACP in the future

[codegirl-007]

Yeah, I kinda went the route of implementing more of an ACP client on the Lua side. It probably would have been much simpler if I just treated it as a thin JSON-RPC bridge to the subprocess. I wanted to have full control over how it worked (or I was overly excited about trying to build a client for ACP).

By building out the full client, you get more control over the sessions. I haven't tried the thin JSON-RPC route, but my understanding is that JSON-RPC itself doesn't have the notion of sessions. It's just the RPC/message layer for sending requests and receiving responses. The session semantics come from ACP.

My thoughts with building an ACP client was to model those session concepts on the Lua side so the plugin can track them, route updates/responses correctly, and manage the lifecycle directly instead of just forwarding JSON messages to the subprocess.

I'm no expert on ACP or any of this stuff. Feel free to correct any misunderstandings I have. :)

[stuckinsnow]

I'm with you now, and yeah I agree with all you wrote. I am passing the session ID but I am not tracking the full session state, nor am I able to read old conversations. I can think of a hackish way to do it but I don't like going down that path haha.

Edit, if anyone faces this same problem... I found it easier to write in Go to do all the above, and then connect to that.

[codegirl-007]

I do like the simpler approach you took though! I think for most cases, the json-rpc is pretty much all you need. I just get a little in the weeds and go overboard. :)

[stuckinsnow]

I do like the simpler approach you took though! I think for most cases, the json-rpc is pretty much all you need. I just get a little in the weeds and go overboard. :)

Hey I re-wrote mine https://www.youtube.com/watch?v=SsPoH3b9pvM. I went the path of ACP using a Go server; it has session management, multi-agent support, and it can talk to an LSP, etc. It also reads and writes to the buffer. I'm fairly pleased with the workflow but it's a bit of a struggle depending on the cloud provider you're with.

While making it, I discovered with kiro you can wrap (shown at 2:10 in the video) the TUI and access messages with tui.js; so if you want a quick solution to most of the above, you only need to make a tool to show diffs from those messages. For me it's a little too bloated and hackish, but thought I'd mention it for anyone that might read this later.

[codegirl-007]

I went the path of ACP using a Go server; it has session management, multi-agent support, and it can talk to an LSP, etc. It also reads and writes to the buffer.

Neat. I thought about doing it in Go before but I hate having to spin things up before I can work on code. I supposed you could have neovim start it up as well. However, I imagine having the separation would be a lot nicer in terms of code organization.

I still haven't tried kiro or fully understand what it's used for (I haven't dug into it to be honest). I've been trying to back off the whole AI train (for my sanity) so I haven't paid as much attention to all the things. But sounds like you have a nice workflow with it and that's really cool.

[miguelsaldanha1213-byte]

I think is good.

[miguelsaldanha1213-byte]

Ok, thank you Em dom., 10 de mai. de 2026, 07:47, Philip ***@***.***> escreveu:

…

I do like the simpler approach you took though! I think for most cases, the json-rpc is pretty much all you need. I just get a little in the weeds and go overboard. :) Hey I re-wrote mine https://www.youtube.com/watch?v=SsPoH3b9pvM. I went the path of ACP using a Go server; it has session management, multi-agent support, and it can talk to an LSP, etc. It also reads and writes to the buffer. I'm fairly pleased with the workflow but it's a bit of a struggle depending on the cloud provider you're with. While making it, I discovered with kiro you can wrap (shown at 2:10 in the video) the TUI and access messages with tui.js; so if you want a quick solution to most of the above, you only need to make a tool to show diffs from those messages. For me it's a little too bloated and hackish, but thought I'd mention it for anyone that might read this later. — Reply to this email directly, view it on GitHub <#144 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BZ2LQBBIC2UQTOTGN5OGAZL42BM2ZAVCNFSM6AAAAACWCZQKTGVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTMOBWHA3DKNY> . You are receiving this because you commented.Message ID: ***@***.***>

[jonaswindey]

is this project dead? are you still using it prime?