New in v19.4.0 - ✨ CLI interactions

[olimorris]

CodeCompanion v19.4.0 adds a new interaction type - CLI - enabling you to interact with agents like Claude Code and Opencode from within Neovim.

A quick 2 min intro can be seen in the video below:

Note

The video has audio and no captions. I talk about how "performant" this is when switching between terminal and CLI buffers - it would have been more useful, visually, if I recorded at 120fps and not 60!

CLI.Interactions.mp4

Over the last month I've found that features in Claude Code are outpacing what goes into the Claude SDK and thus what Zed can implement in their ACP adapter, by a serious rate. Whilst I don't think there has been any features of note that we've missed out on, I'm keen to ensure that Neovim devs can always interact with the newest agents and LLMs from within the editor.

So, having the ability to share files and context with an agent via a swift keymap feels like a sensible addition. As does giving you devs the ability to share editor context and prompt from anywhere within a buffer.

The usage part of the docs covers everything in full and I'd recommend you take a look at that for more information alongside how you configure it.

---

I'm also super keen to hear how you would want to use this new interaction type in your workflow. So please post them below. I'm now so close to this PR that I can't see the wood for the trees 😆.

[daviareias]

I've using CodeCompanion only through Chat, but I heard so many things about ClaudeCode that I'm curious on how people are using it nowadays.

[olimorris]

It's pretty much become the only AI tool I use. The other one being Augment Code. Whilst it gets a ton of praise as an agent CLI, the real star is Opus 4.6 which, as a model, is the most impressive that I've used. Claude Code is wonderful in that its use of tools to manage context and memory make it feel like it has near infinite context (N.B. it doesn't and context rot is real).

Most of the greenfield features I shipped since January in CodeCompanion have been aided by Claude Code.

[daviareias]

not sure what I did, but using the API version with chat was a better experience for me, claude code takes a lof ot time to answer.

I guess you have to tweak the settings to make it work I guess.

Edit: it seems the chat version of claude_code is really superior experience compared to running the CLI so far for me.

[pixlmint]

I've tried it, and seems to work pretty well, thank you as always for your great work Oli

One thing I was wondering and couldn't deduce from the documentation, how can we customize the buffers filetype? Right now the filetype is empty, buftype is terminal. I'd appreciate it if the created buffer had a unique filetype for easy integration with edgy.nvim.

[olimorris]

Great suggestion. I've added that in #2894. codecompanion_terminal is the filetype

[pixlmint]

Thanks for the quick fix, works great

[vuongdanghuy]

Hi, I’ve been trying out the new CLI interaction with Opencode and ran into an issue.

The response shown in the terminal buffer appears to be truncated—I can only see the last few lines of the output. I’m also unable to scroll up to view the earlier parts of the response.

Here is an example image. As you can see the cursor is already place at the first line of the terminal buffer.

Is this expected behavior, or is there a way to access the full output?

Thanks!

[vuongdanghuy]

If i enable mouse in nvim then i can scroll up to see the earlier response in the terminal mode using the mouse. But if i disable the mouse and use ESC to change from Terminal to Normal mode to navigate around then i can't.

[olimorris]

I'm not familiar enough with Opencode to know how their TUI works and if there's keyboard shortcuts.

But if i disable the mouse and use ESC to change from Terminal to Normal mode to navigate around then i can't.

Yep this is expected behaviour. You're scrolling the buffer. When you're in terminal-mode and use your mouse then you're scrolling the TUI.

[vuongdanghuy]

Thank you

[vuongdanghuy]

in terminal mode, i can use the PageUp/PageDown to scroll. Just thinking it'd be nice if we can navigate using Normal mode key binding like in the chat buffer.
Anyway, nice work :)))

[olimorris]

in terminal mode, i can use the PageUp/PageDown to scroll. Just thinking it'd be nice if we can navigate using Normal mode key binding like in the chat buffer.

Anyway, nice work :)))

Sure you can. But this should be a custom keymap in your config. Not something CodeCompanion should set by default - far too opinionated.

[pixlmint]

One feature I really appreciate about claudecode.nvim is that when I'm working on a project with multiple CLAUDE.md files (like is often the case with monorepos), that it automatically loads the CLAUDE.md that's relevant to the buffer I was in.

For example, if I have a CLAUDE.md for the backend and also one for the frontend, and I'm working on the backend, it automatically injects the backend CLAUDE.md

I've asked claude how the plugin does it, and turns out it's a claude-code standard, I believe their vscode plugin does the exact same thing.

Here's the summary claude gave me on how it works:

Details

1. Plugin starts a WebSocket server and writes a lock file to
   ~/.claude/ide/.lock containing workspaceFolders — an array of
   directory paths (lockfile.lua:115-121).
2. Workspace folders are determined by (lockfile.lua:258-292):
   • Neovim's current working directory (vim.fn.getcwd())
   • Any additional folders reported by active LSP clients via
     client.config.workspace_folders
3. Claude CLI discovers the plugin by reading the lock file, connects
   over WebSocket, and can also call the getWorkspaceFolders MCP tool
   (get_workspace_folders.lua:16) which currently just returns
   vim.fn.getcwd().
4. Claude CLI then handles CLAUDE.md loading — it walks up from the
   workspace folder paths to find .claude/CLAUDE.md or CLAUDE.md files.
   This logic lives in the Claude CLI binary, not in this plugin.

So the key link is vim.fn.getcwd() — whatever directory Neovim was
started in (or changed to via :cd) becomes the workspace root that
Claude CLI uses to locate the correct CLAUDE.md. The plugin itself
never reads or parses CLAUDE.md; it just tells Claude CLI "here are
my workspace folders" and Claude CLI does the rest.

[olimorris]

If it's a claude-code standard then this should automatically apply when using it via the CLI?

[pixlmint]

Well, a claude-code standard in that, it provides a way for editors to tell it what files are currently opened