[Bug]: Unable to use on a large repo

[aivanov93]

Description

I tried opening a large repo (3GB of code, 16GB of .git) and it basically took down my computer. I couldn't even take a look at processes, my Acitivity Monitor stalled. 🙃

Steps to Reproduce

Open a large repo as a project.

Expected Behavior

It works

Actual Behavior

It doesn't work

AI Engine

Not engine-specific

Operating System

macOS (Apple Silicon)

Harnss Version

0.17.4

Area

No response

Logs / Error Output

Additional Context

No response

[OpenSource03]

Did you make sure to check latest https://github.com/OpenSource03/harnss/releases/tag/v0.18.1?

I have a new release coming soon heavily focusing on performance too.

Meanwhile claude and codex will also be trying their best to see if there is anything else that may be causing this issue :)

[OpenSource03]

Also, do you have any public git repo that I may try to reproduce it on? Maybe the likes of Webkit?

[aivanov93]

It's my company's internal repo unfortunately. But here are some ideas from my Claude to yours 🙃

---

Title: Git operations cause system thrashing on large monorepos (no caching/debouncing)

Environment:

• macOS, Apple Silicon, 64GB RAM
• Repo: Large java monorepo (~large, thousands of files)

Problem:
Opening a large monorepo in Harnss and starting new chats causes extreme system load. My machine hit a load average of 359 (on 16 cores), used 63GB/64GB RAM, 20GB swap, and triggered macOS kernel_task thermal throttling. Fans run at max and the system becomes nearly unusable for ~30 seconds on every new chat.

Root cause:
Harnss's git IPC layer (electron/src/ipc/git.ts) runs raw git commands with no caching, debouncing, or incremental updates. Every call to git:status spawns a fresh git status --porcelain=v2 --branch on the full repo. On a large monorepo, each of these takes 10-30 seconds. Combined with the Claude Agent SDK's own git context gathering on session start (which also runs git status, git diff, git log), this creates a storm of concurrent git processes that exhaust CPU and memory.

Comparison with 1Code:
1Code (github.com/21st-dev/1Code) handles the same monorepo without issues because of three key differences:

1. LRU cache with TTL (src/main/lib/git/cache/git-cache.ts) - git status is cached for 5s, diffs for 1min, file contents for 5min. Repeated requests hit memory instead of spawning git processes.

2. Minimal file watcher (src/main/lib/git/watcher/git-watcher.ts) - watches only .git/index and .git/HEAD (2 file descriptors), then invalidates the cache on change. Harnss has no watcher and likely re-polls the full status on every UI refresh.

3. Debounced batch events - changes are batched with 100ms debounce so rapid git activity doesn't cascade.

Suggested fix:

• Add a short-lived cache (even 5s TTL) around git:status results in git.ts
• Debounce UI-triggered git IPC calls
• Use a lightweight .git/index + .git/HEAD file watcher to invalidate the cache instead of re-running full git status
• Consider adding a repo size heuristic to skip or defer expensive operations like git:discover-repos directory walking on large repos