CommandCode terminal freezes entirely + flickering on large output

[johnnypernik]

CommandCode Freeze Report

Who am I

I'm a generalist all-rounder — I know enough to operate at a medium level across a lot of stuff (homelab, WSL, basic scripting, tooling) but I don't have the depth to human-diagnose a CLI tool's internals. This report is a self-test performed by DeepSeek v4 Pro running inside CommandCode, collecting its own environment and config to hand off to humans who know this codebase. Please take it from here.

---

What's happening

CommandCode becomes completely unresponsive during use. The tab freezes — no keypresses work, no output renders. I have to close the tab and open a new one. The session can be recovered (state is still there), but it's disruptive. Separate from this, there's a rendering issue with flickering. Specific things I notice:

• Tab freezes solid — no keypresses work, no output renders
• Must close tab and open new one — session recovers, but disruptive
• Screen flickers on large outputs — happens independently of freezing, fixes itself after the next shorter output renders (separate cosmetic bug, not related to the freeze)
• DeepSeek v4 Pro is extremely slow in general — even simple questions take very long
• Custom agents don't seem to work — I have 40 agent profiles set up but the tool doesn't route to them the way I'd expect, or they don't behave like their descriptions say they should

I'm running all defaults as installed. I haven't modified CommandCode internals. The project it freezes in has a larger-than-average .commandcode/ directory, but nothing unusual for someone using CommandCode seriously.

---

Self-collected environment info

DeepSeek v4 Pro collected this from inside my session using shell commands. It's the best I can give you without knowing what else to look for.

System:

• Windows 10, WSL2 Ubuntu 24.04.4 LTS, kernel 6.6.87.2
• AMD Ryzen 7 9800X3D (16 threads)
• 47GB RAM available to WSL
• 1TB NVMe disk
• No GPU visible inside WSL (Nvidia RTX4090)

Terminal:

• Windows Terminal (the default Microsoft one)
• Bash shell, TERM=xterm-256color

Node / CommandCode:

• Node.js v22.22.2 (installed via nvm)
• CommandCode v0.26.12 (global npm install)
• ~/.commandcode/config.json:

  {
    "provider": "command-code",
    "model": "deepseek/deepseek-v4-pro",
    "reasoningEffort": {
      "deepseek/deepseek-v4-pro": "max"
    }
  }

Project .commandcode/ directory (584KB total):

• COMMANDCODE.md (2.5KB) — routing file
• taste/taste.md — 8 learned preferences
• 18 skills in .commandcode/skills/ (all symlinked)
• 40 agent profiles in .commandcode/agents/ (483KB total, some profiles 20-27KB each)

MCP servers connected:

• AgentMemory (remote LXC, port 3111) — 50+ tools
• A2A Registry / MCP Bridge (port 3457) — 7 tools
• CodeGraph (remote LXC, port 3460) — https://github.com/colbymchenry/codegraph
• Cloakbrowser MCP (remote LXC, browser automation) — https://github.com/CloakHQ/CloakBrowser
• GitHub, CodeGraph Local

Token estimate by DeepSeek v4 Pro:
DeepSeek itself estimated that per-turn context overhead is roughly:

• 40 agent profiles: ~120,000 tokens
• 50+ AgentMemory MCP tool schemas: ~15,000-20,000 tokens
• AGENTS.md + COMMANDCODE.md: ~5,000 tokens
• Other MCP schemas: ~5,000 tokens
• Total: ~150,000 tokens before I type anything

---

My guesses (from a non-dev — may be wrong)

DeepSeek v4 Pro, while collecting this report, suggested these might be worth looking at. I'm including them since the model has more context than I do, but please treat these as starting points, not conclusions:

1. Way too much stuff loaded at once. 40 agent profiles + 50 MCP tools + project docs all in context before I even type. If there's a context window limit, I'm probably blowing past it.
2. DeepSeek v4 Pro is slow on its own and reasoningEffort: max makes it worse. Even outside this project, it's not snappy.
3. Large output rendering flickers — but this is a separate issue from freezing. Flicker happens on large responses and fixes itself after the next shorter output renders. This could be a TUI/ANSI rendering issue rather than a model problem.
4. Custom agents don't seem wired up. 40 profiles exist but the behavior doesn't match. Maybe the schema format I'm using isn't what CommandCode expects, or agent discovery isn't loading them correctly.
5. Remote MCP might be slow. The AgentMemory and A2A servers run on a separate machine on my local network. If network calls are blocking, that could explain hangs.

---

What I need from you

I don't know how to debug this further. I'm handing off:

• This report (everything I could self-collect)
• Happy to run any diagnostic commands you give me or test patches

Priority for me: stop the freezing so I can actually use the tool. If fixing agents or trimming MCP tools fixes the freezes too, great. If the freezes are a separate rendering bug, I need to know that.

Thanks for looking into this.

---

Related CommandCode GitHub Issues

This report touches on several existing open issues. Linking them here so devs can cross-reference:

Terminal / rendering:

• #322 — Terminal Flickers when choosing Models with /model — flickering symptom matches exactly
• #329 — Enter key and inputs not registering in interactive menus (WSL) — WSL-specific, may share root cause with terminal freeze

MCP hangs / crashes:

• #354 — /exit hangs when MCP servers are connected — processes not cleaned up — MCP-related hangs
• #345 — MCP Tool Crashes Command Code CLI on Connect — similar MCP startup crash (I use Cloakbrowser, but same pattern)
• #371 — MCP server processes not killed when terminal session closes — memory leak accumulates across sessions — could compound context load over time

Memory / heap:

• #342 — FATAL ERROR: heap out of memory — 40 agent profiles + 50 MCP tools may trigger this
• #361 — Crashing: JavaScript heap out of memory on long sessions — large config = possible trigger

DeepSeek / performance:

• #346 — deepseek-v4 cache hit rate much lower than OpenRouter — may explain why DeepSeek v4 Pro feels extremely slow

Agent routing:

• #380 — Feature Request: Per-Mode and Per-Agent LLM Routing — suggests agent routing isn't fully implemented yet, which would explain why 40 custom agents don't seem to work

If this self-diagnostic helps, I'm happy to provide more data or test patches.

---

Public tools I'm using (for context)

These are the third-party tools loaded into CommandCode that contribute to the token overhead:

• AgentMemory (50+ MCP tools) — session memory, actions, signals, sentinels for multi-agent coordination
• Google A2A protocol — https://github.com/a2aproject/A2A — agent-to-agent communication
• CodeGraph MCP — https://github.com/colbymchenry/codegraph
• Cloakbrowser MCP — https://github.com/CloakHQ/CloakBrowser — browser automation
• Matt Pocock Skills — https://github.com/mattpocock/skills — community skill pack (diagnose, tdd, triage, etc.)
• Agency Agents — https://github.com/msitarzewski/agency-agents — source of the 40 agent profiles (installed via install-agency-agents.sh)

[asharirfan]

@johnnypernik I’m curious, have you tried using Command Code through WSL?

You mentioned that you’re using Windows Terminal (the default Microsoft terminal). Right now, the recommended way to use Command Code on Windows is through WSL, since it provides a full Linux environment where everything works exactly as documented.

Native Windows terminal support is still in alpha, so you may run into issues depending on your setup. We’ve also noted this in our docs: https://commandcode.ai/docs/troubleshooting/windows

If you haven’t tried WSL yet, I’d recommend giving it a shot as it tends to provide the smoothest experience on Windows at the moment.

[johnnypernik]

@asharirfan i probably misinterpreted- I do use WSL2 as mentioned in system info, but I'm on Windows so latest (downloaded) Microsoft Terminal app is what I use to access WSL2 & commandcode. I even use native Linux filepaths and not mounting Windows drives to make stuff faster.

[asharirfan]

@johnnypernik Could you do me a favor and try it out with a more modern terminal like Alacritty OR hyper?

I have been using Alacrity myself w/ on a 4GB machine and I have not experienced any lag for now.

However, I will investigate the issues that you've mentioned in the issue. Thanks!