Critical: AI Unusable - NGHTTP2_REFUSED_STREAM, CursorCreds, Memory Leak on Remote SSH

[TheOneRecords]

Where does the bug appear (feature/product)?

Cursor IDE

Describe the Bug

AI features completely blocked by multiple errors:

1. HTTP/2 failure: NGHTTP2_REFUSED_STREAM in cursor-always-local extension
2. Auth error: CursorCreds is undefined - occurs with HTTP/2 both enabled AND disabled (tried toggling, error persists in both states)
3. Memory leak: 249+ event listeners accumulating during usage
   Previously worked perfectly. Started failing Oct 26, 2025 around 20:04 UTC.
   Troubleshooting attempts:

• Toggled HTTP/2 on/off - CursorCreds error in both states
• Cleared cache and workspace storage
• Signed out and back in multiple times
• Ran Network Diagnostics
• Restarted Cursor completely
  Environment: macOS + Remote SSH to Linux VPS, Cursor v1.7.54 (Universal)
  Is this a known server-side issue? The sudden failure across different settings suggests infrastructure changes.

Steps to Reproduce

1. Open Cursor with Remote SSH to Linux VPS
2. Send messages in Chat sidebar
3. Initial 2-4 prompts work correctly with normal AI responses
4. Subsequently, one of these failures occurs:
   • Infinite loading spinner (never completes)
   • Response cuts off mid-generation (e.g., stops halfway through a list)
   • Complete unresponsiveness to new prompts
5. Developer Console (Cmd+Shift+I) shows:
   • NGHTTP2_REFUSED_STREAM when HTTP/2 is enabled
   • CursorCreds is undefined regardless of HTTP/2 state
6. Only a full Cursor restart temporarily fixes the issue
7. Pattern repeats every session
   Note: HTTP/2 toggle has no effect - both configurations exhibit the same behavior.
   Expected: Stable connection throughout the session
   Actual:

• Progressive degradation after initial success
• Connection loss occurs unpredictably after 2-4 interactions
• Memory leak evident (249+ event listeners accumulating)
• Frequent restarts required for continued use

Expected Behavior

AI features should work consistently throughout the session without:

• Connection drops after initial prompts
• Mid-response truncation
• Authentication failures
• Memory leaks requiring restarts

Operating System

MacOS

Current Cursor Version (Menu → About Cursor → Copy)

Version: 1.7.54 (Universal)
VSCode Version: 1.99.3
Commit: 5c17eb2968a37f66bc6662f48d6356a100b67be0
Date: 2025-10-21T19:07:38.476Z
Electron: 34.5.8
Chromium: 132.0.6834.210
Node.js: 20.19.1
V8: 13.2.152.41-electron.0
OS: Darwin arm64 25.0.0

For AI issues: which model did you use?

Auto

For AI issues: add Request ID with privacy disabled

c062eecb-1e2d-4254-93f3-540d2bfb6eb9

Does this stop you from using Cursor

Yes - Cursor is unusable

[deanrie]

Hey, thanks for the report. The NGHTTP2_REFUSED_STREAM error is a known connection issue. Let’s try the following:

1. Properly disable HTTP/2:
   Cursor Settings Cmd + , → search “HTTP/2” → enable “Disable HTTP/2”
   Fully quit and relaunch Cursor (not just Reload Window)

2. Run Network Diagnostics:
   Settings → Network → Run Diagnostic
   Share the full output here (especially ping/chat/agent results)

3. Check MCP servers:
   If you have MCP servers configured, temporarily disable them (they can cause connection failures)

4. Test without Remote SSH:
   Open a local folder to see if the issue is specific to Remote SSH

[TheOneRecords]

It has already been deactivated. With my first prompt of the day, even short messages get stuck processing, although no errors appear in the network diagnostics. Locally, it does process, but still quite slowly, even for simple prompts. SSH had been working perfectly until now, but I left the initial “test” prompt running over SSH and it still hasn’t completed after 6–7 minutes. Could this be an issue with the SSH extension? I’m using macOS Tahoe 26.0.1.

The SSH session has plenty of resources available, and my macOS firewall is disabled and I’m not using a VPN.

Although when I had HTTP/2 set as HTTP compatibility, I did get an error once, now in network diagnostics I no longer get any error regardless of the HTTP compatibility setting, it just doesn’t process.

After several long prompts, it eventually timed out without showing an error.

Here is the request ID: f33fcae3-14e0-42bc-9cda-4d719f2f6b29.

[TheOneRecords]

I tried using the CLI version, and it worked better until now when I noticed that this one also got stuck.

5cb5bbe8-6fe9-4bf9-b9e7-eeb6bd2e6547

[Blake_McDonald]

Any resolution? Getting the same with SSH connection. In my case SSH is from AWS workspaces (Windows) to EC2 Amazon Linux, with Cursor + devcontainer.