performance issues

[solotim]

Hi,
I frequently see nvim slowdown when the size of the conversation grow big. In one session I asked ai diagnose and it found below. Any idea if it's related to agentic.nvim?

performance analysis of Neovim, specifically identifying high CPU usage related to LuaJIT Garbage Collection. Below is the transcribed text:
14.96% gc_sweep \leftarrow LuaJIT GC sweeping
9.19% lj_alloc_free \leftarrow LuaJIT memory freeing
8.70% propagatemark \leftarrow LuaJIT GC marking
5.62% marktree_lookup \leftarrow nvim extmark tree lookups
~33% of CPU is LuaJIT Garbage Collection. Some Lua plugin(s) are creating massive numbers of short-lived objects, causing the GC to run constantly and burn a full CPU core.
marktree_lookup at 5.6% on top of that points to too many extmarks — the data structure nvim uses for treesitter highlights, LSP diagnostics, signs, virtual text.
The two likely culprits together:
A plugin with a hot autocmd (on CursorMoved, TextChanged, etc.) allocating lots of Lua tables — check plugins like lualine, noice, nvim-cmp, or any status/hint plugins
Extmark flood from two clangd instances both pushing diagnostics/inlay hints into the same buffer simultaneously
To pinpoint the Lua hotspot, run this inside nvim:
:lua vim.uv.new_timer():start(0, 100, function() print(collectgarbage('count')) end)

[carlos-algms]

I'm sad to read you're facing slowdowns on your Neovim.

On our side, we don't use CursorMoved nor TextChanged, we use a TextChangedI, only on the Prompt buffer, which doesn't have many changes.

Most of the activity in this plugin happens on the Chat buffer, which doesn't have auto commands.

I would suggest following your AI suggestions and disabling plugins like lualine, noice, nvim-cmp, or any status/hint plugins, and check for performance again on a long session.

And maybe run :lua vim.uv.new_timer():start(0, 100, function() print(collectgarbage('count')) end) also.

On my day-to-day, I'm running very tool-call intensive sessions, with many blocks, and I don't feel this slowdown.
Some days, I run 2 tabs with 2 Agentic sessions and still no slowdown.

[solotim]

Thanks Carlos.
It turned out the culprit is render-markdown. Once i disable it for AgenticChat buffer, things are smoothier.
Do you personally use any md render plugin for the chat buffer? Any recommendations?

[carlos-algms]

Nice that you found it.

I don't use any markdown render, I've tried 2 of them, and at the end of the day, I gave up.
They were too distracting for me.

[solotim]

Thanks.
I also noticed that the show/hide agentic panes are slow even I disabled all markdown render pligins. I guess it's just because the session buffer is too large and drags down nvim. What's your suggestion when you have huge session buffer? I can save context to external me and use new session to resume from it. But I'd like to know the recommendation from the author.

[carlos-algms]

Yes, that was a known issue, sorry for that.

It should have been fixed by:

• fix(ui): auto-scroll on tool call updates to prevent flicker #210
• full exploration and reasoning: Tool call streaming flicker: why we use manual folds with anchor pads #212

I migrated to manual folds instead of function that evaluates the whole buffer for folding sections.

Please, update and check if it's better now 😊