[Bug]: Mcode ends up using 100% of available RAM on every turn i use #508

Activity

1. paramcodes

   added

   bugSomething isn't workingSomething isn't working

   needs-triageAwaiting maintainer assessmentAwaiting maintainer assessment

   on Oct 19, 2026
2. github-actions

   added

   tuiInteractive terminal UI (TUI)Interactive terminal UI (TUI)

   on Oct 19, 2026
3. hetaoBackend commented on Oct 20, 2026

   

   hetaoBackend

   on Oct 20, 2026

   Collaborator

   More actions

   Hi @paramcodes, thanks for the report and the screenshot.

   We'd like to narrow this down. One area we're looking at is how 0.6.11 watches the workspace on Linux: for large directory trees, this can use a lot of memory. 0.6.12, already on npm, changes this code path on Linux. Could you help with the following?

   1. Try 0.6.12. Update with npm install -g @minimax-ai/code@latest --registry=https://registry.npmjs.org/ (or rerun the install script), then check whether the problem still happens.
   2. Working directory. Which directory do you start mcode from: a project, your home directory (~), or something else? Is it inside a Git repository (including a dotfiles repo in ~)?
   3. Size of that directory. Please share the output of:

      cd <the directory you start mcode in>
      find . -type d 2>/dev/null | wc -l
      git rev-parse --show-toplevel 2>/dev/null
      cat /proc/sys/fs/inotify/max_user_watches

   4. Empty-directory test. Does it still freeze and grow in memory if you start it in an empty directory?

      mkdir -p /tmp/mcode-test && cd /tmp/mcode-test && mcode

   5. When the memory grows. Does memory climb right after launch, before you send anything, or only after you send a prompt? Roughly how high does it get, and does the process exit or get OOM-killed?
   6. Environment. Your Node.js version (node -v) and how you installed MCode (install script or npm).

   If the TUI responds after a restart, you can also run /feedback <short description> in it. This submits a redacted diagnostic bundle. Please paste the Upload ID it shows here, so we can look at the logs on our side.

   Thanks!

   👀React with 👀1Reacted by Paramveer Singh
4. paramcodes commented on Oct 22, 2026

   

   paramcodes

   on Oct 22, 2026

   Author

   More actions

   Thanks @hetaoBackend. Diagnostics below.

   1. Version tested: updated to 0.6.12 (npm install -g @minimax-ai/code@latest --registry=https://registry.npmjs.org/). mcode --version → 0.6.12. The freeze/memory blowup still reproduces on the same workspace.

   2. Working directory: I start MCode from my home directory (~, /home/param). It is not inside a Git repository (git rev-parse --show-toplevel → "not a git repository"). I also have smaller project workspaces (~/Projects/noodle, ~/Work/maggi, ~/Work/poly-ai) which are git repos and behave fine.

   3. Size of that directory:

   cd /home/param
   find . -type d 2>/dev/null | wc -l   # 324193
   find . -type f 2>/dev/null | wc -l   # 2663015
   cat /proc/sys/fs/inotify/max_user_watches  # 524288
   

   4. Empty-directory test: mkdir -p /tmp/mcode-test && cd /tmp/mcode-test && mcode → TUI launches normally, no freeze, no memory growth. So the problem is tied to the size of the opened tree, not the empty case.

   5. When memory grows: It climbs right after launch in ~ (the TUI shows "Starting server…" for ~30s while indexing), and the process gets sluggish/OOM-prone on every turn; in an empty dir it stays flat. In ~ the TUI takes ~30s to settle before I can even send a prompt.

   6. Environment: Node.js v26.8.2 (mise). MCode installed via npm (0.6.12). I also use the MiniMax Code desktop app (3.1.4 unofficial Linux build), which has ~ configured as a workspace too, so the same watcher path likely applies there.

   Conclusion / ask: The watcher appears to recursively index the whole home tree (324k dirs). Could the Linux watcher (a) skip ignored/hidden dirs by default, (b) respect .gitignore/ignore files even outside a git repo, or (c) not watch ~ unless explicitly opted in? Happy to provide an Upload ID via /feedback if useful — just let me know if you'd like one.