# Plugin marketplace sync runs \`git ls-remote\` with no timeout, leaking hundreds of stuck git/sh/ssh processes on Remote SSH (Windows)

**URL:** <https://forum.cursor.com/t/plugin-marketplace-sync-runs-git-ls-remote-with-no-timeout-leaking-hundreds-of-stuck-git-sh-ssh-processes-on-remote-ssh-windows/176416>\
**Category:** Bug Reports\
**Tags:** plugins, ssh, networking\
**Created:** [October 27, 2026, 7:13pm UTC](https://forum.cursor.com/t/plugin-marketplace-sync-runs-git-ls-remote-with-no-timeout-leaking-hundreds-of-stuck-git-sh-ssh-processes-on-remote-ssh-windows/176416 "2026-10-27T19:13:57Z")\
**Posts on this page:** 3\
**Page:** 1

<div class="post-metadata">

**Author:** ![AustinS](https://avatars.discourse-cdn.com/v4/letter/a/e8c25b/32.png) [@AustinS](https://forum.cursor.com/u/AustinS)\
**Post date:** [October 27, 2026, 7:13pm UTC](https://forum.cursor.com/t/plugin-marketplace-sync-runs-git-ls-remote-with-no-timeout-leaking-hundreds-of-stuck-git-sh-ssh-processes-on-remote-ssh-windows/176416/1 "2026-10-27T19:13:57Z")

</div>

### Where does the bug appear (feature/product)?

Cursor IDE

### Describe the Bug

**Summary**  
Each time the agent runs on the remote host, it resolves the git ref of a plugin marketplace repo using `git ls-remote` over SSH, with no timeout. On a network where TCP port 22 connects but the SSH greeting never arrives, the call hangs forever. Each hung attempt leaves a `git.exe`, `sh.exe` and `ssh.exe` process behind. I found 325 of each (975 processes) after about 28 hours, and the count was still rising while I watched.

**Root cause (from reading the shipped extension code)**

- The code is in `extensions/cursor-agent-exec/dist/main.js` and the same logic in `cursor-agent-host`. `MarketplaceCacheManager.ensureClonedImpl` rewrites an https marketplace URL to `git@github.com:...` and calls a git ref resolver that runs `git ls-remote`.
- The SSH attempt sets `sshBatchMode`, which only avoids prompts and does not limit how long the call can hang.
- `lsRemoteTimeoutMs()` returns `undefined` unless the `sparsePluginClones` option is on (30000 ms when it is).
- Because the call never fails, the `catch` block that falls back to HTTPS is never reached.
- Verified on the same machine: `git ls-remote https://github.com/octoplus-sa/claude-skills.git main` returned the commit in about 1 second.

**Suggested fix**

- Always set a timeout on `ls-remote`, for example by also passing `-o ConnectTimeout=10` or `ServerAliveInterval`, and kill the whole process tree when it expires.
- Apply the same timeout whether or not `sparsePluginClones` is on.
- Kill or reap child processes when the parent exits.

**Workaround found:** adding this to the remote host’s `~/.ssh/config` makes the SSH attempt fail after 10 seconds, so the fallback runs and nothing hangs:

```auto
Host github.com
    ConnectTimeout 10

```

**Impact:** The remote host accumulated about 1,000 stuck processes. Nothing broke, but this wastes process slots and memory and makes the process list hard to use.

### Steps to Reproduce

1. Install a plugin whose marketplace repo is hosted on GitHub. Mine is a JSL scripting plugin, and I never added a marketplace by hand. The marketplace is `octoplus-sa/claude-skills`. This is a public repo, and HTTPS access to it works.

2. Use a network where SSH to [github.com](http://github.com) hangs. The connection opens but no greeting comes back. `ssh -v git@github.com` reports “Connection timed out during banner exchange”. This happens on both port 22 and 443 (`ssh.github.com`). There was no SSH key on the remote host (`.ssh` held only `authorized_keys`).

3. Connect with Remote SSH and use the agent normally.

4. After a while, list processes on the remote host. The count of identical stuck commands keeps rising:

### Expected Behavior

**Expected:** The ref lookup times out after a reasonable period, kills the child process, and falls back to HTTPS. No processes are left behind.

**Actual:**

- The call never returns. It waits on `ssh.exe` indefinitely, so the HTTPS fallback never runs.
- When the launching node process exits, the `git` and `ssh` children are left running, with no parent.
- Batches appear every few minutes while the agent is in use, usually three at a time.

### Operating System

Windows 10/11

### Version Information

- Cursor 3.27.14, commit `69b724ebc67ec522c1b5d287a5025f7a7e81be90`, built 2026-10-23
- Remote host: Windows 11 (NT 10.0.26300), Git for Windows 2.56.0, OpenSSH\_for\_Windows\_9.5p2
- Connected via Remote SSH. The agent runs on the remote host.

### Does this stop you from using Cursor

Sometimes - I can sometimes use Cursor

---

<div class="post-metadata">

**Author:** ![mohitjain](https://sea3.discourse-cdn.com/cursor1/user_avatar/forum.cursor.com/mohitjain/32/100229_2.png) [@mohitjain](https://forum.cursor.com/u/mohitjain)\
**Post date:** [October 28, 2026, 5:26am UTC](https://forum.cursor.com/t/plugin-marketplace-sync-runs-git-ls-remote-with-no-timeout-leaking-hundreds-of-stuck-git-sh-ssh-processes-on-remote-ssh-windows/176416/9 "2026-10-28T05:26:52Z")

</div>

Hey @AustinS, thanks for the detailed write-up. I was able to reproduce this, and your `ConnectTimeout` workaround is the right thing to keep in place for now. Just keep the value under 30 seconds so the HTTPS fallback still runs.

It’s safe to stop the `git.exe`, `sh.exe`, and `ssh.exe` processes that are already stuck on `octoplus-sa/claude-skills`.

What you’re describing isn’t intended behavior, and we’ve let the team know. I’ll post here when there’s an update.

---

<div class="post-metadata">

**Author:** ![AustinS](https://avatars.discourse-cdn.com/v4/letter/a/e8c25b/32.png) [@AustinS](https://forum.cursor.com/u/AustinS)\
**Post date:** [October 28, 2026, 1:56pm UTC](https://forum.cursor.com/t/plugin-marketplace-sync-runs-git-ls-remote-with-no-timeout-leaking-hundreds-of-stuck-git-sh-ssh-processes-on-remote-ssh-windows/176416/10 "2026-10-28T13:56:40Z")

</div>

Thanks for the replies. Just a quick follow up on the workaround in case it’s helpful:

About an hour after I added `ConnectTimeout 10` to `~/.ssh/config`, stuck processes started building up again. In about 67 minutes, 20 new `ssh.exe` processes appeared, all running the same `git-upload-pack 'octoplus-sa/claude-skills.git'` call. They stayed alive for over an hour.

Each stuck `ssh.exe` had a socket in the `Bound` state with remote address `0.0.0.0:0`. It never reached `SynSent` or `Established`, and it used almost no CPU. The timeout only starts once ssh begins connecting, so it never fired. Run by hand, the same ssh call, and the same `git ls-remote` spawned from Node with pipes and no console, timed out cleanly after about 10 seconds. I could not reproduce the stuck state outside the agent’s own spawns. I don’t know what holds the connection (possibly a network or security layer on the host).

Some of the stuck `git` and `sh` processes did exit, but the `ssh.exe` child outlived them.

What worked: having ssh never open a socket for `github.com`:

```auto
Host github.com
    ConnectTimeout 10
    ProxyCommand cmd /c exit 1

```

With this, `ssh` fails in about 0.2 seconds and the agent’s `git ls-remote` fails in about 0.5 seconds with “Could not read from remote repository”. That error is what should trigger the fallback to HTTPS. After applying it and clearing the old processes, none had returned by the next morning. The obvious downside is that ssh cannot open a connection to [github.com](http://github.com) with this configuration.

The takeaway is that a ref lookup that depends on ssh-level options can’t be guaranteed to time out. Cursor could enforce the timeout itself, for example by killing the whole process tree (`taskkill /T`) after a fixed time. It could also set a timeout whether or not `sparsePluginClones` is enabled.
