Where does the bug appear (feature/product)?
Cursor IDE
Describe the Bug
Environment: Cursor 3.21.13 (build 18 Sep 2026), macOS recently updated to “Golden Gate”, Apple Silicon Mac.
Symptom: Agent chat works normally for conversation, but freezes indefinitely the moment it needs to execute a tool call involving file search (grep) or file reading. Status stays stuck on “Exploring / Grepping …” forever, with no timeout, no error, no progress. New agent chats can be opened and typed into fine (the UI itself stays fully responsive) - only the specific in-progress tool call never resolves. Eventually surfaces: “Agent stopped retrying - The connection was interrupted repeatedly without saving progress…” (Request ID: d7b68362-177e-402f-960e-a2e2789a4891).
Relevant logs (Output panel, “Cursor Agent Exec” / “Cursor Grep Service” channels):
[error] Failed to initialize GrepClient cannot index repository without Git
[info] using external crepectl binary at ‘…/Cursor.app/Contents/Resources/app/resources/helpers/crepectl’
[warning] [watchdog, LocalReadExecutor] permissionsService.shouldBlockFileRead still not completed after 3000ms
[info] [agent-store-sync] sync failed: [unavailable] getaddrinfo ENOTFOUND api2.cursor.sh
The shouldBlockFileRead warning recurs every few minutes and never resolves to “completed”.
What we’ve ruled out (extensive troubleshooting over several hours):
- Not network/DNS: HTTP/1.0 and HTTP/1.1 pass Cursor’s own Network Diagnostics fully green.
- Not VPN (all profiles disconnected) or firewall (inactive).
- Not macOS file permissions: Full Disk Access already granted to Cursor.
- Not workspace size or missing git: reproduced identically after switching from a large multi-project workspace to a small, isolated single-project folder.
- Not orphaned processes: fully quit via Cmd+Q (verified zero Cursor/Cursor Helper processes in Activity Monitor), clean relaunch - issue recurred at the exact same step.
- Not CPU-bound: Cursor main process sits at under 1% CPU while “hung” (looks blocked/waiting, not spinning).
- Not third-party AV/monitoring: fully quit CleanMyMac X.
- Not Gatekeeper/quarantine on the app bundle: ran xattr -dr com.apple.quarantine on Cursor.app, no change.
- Not Spotlight reindexing: mds/mds_stores CPU usage negligible.
- Not shell environment: removed a problematic “source <(ng completion script)” line from .zshrc, no change.
- Already on latest build (checked via Check for Updates).
Interesting clue: after one full restart, the UI instantly rendered several previously-stuck steps as “completed” (with timestamps like “12m ago”) the moment the app reopened - suggesting the backend agent process may have actually finished those steps, but the UI/renderer never received or displayed the completion event until restart. This may point to an IPC/state-sync bug between the agent backend and the renderer, rather than (or in addition to) an actual hang in the grep/file-read execution itself.
Timeline: issue began right after updating macOS to “Golden Gate”; no other environment changes were made around that time.
Steps to Reproduce
- Open Cursor with any workspace (reproduced both on a large multi-project folder and a small single-project folder). 2. Start a new Agent chat and give it a task that requires searching/reading files (e.g. “find where cookies are saved and update the logic”). 3. Watch the agent progress to a step like “Exploring / Grepping …” - it freezes there indefinitely, never completes, no error surfaces until eventually “Agent stopped retrying” appears. 4. Fully quit Cursor (Cmd+Q) and relaunch - issue recurs at the same type of step (grep/file read) even in a brand new chat.
Expected Behavior
The agent tool call (file search/read) should complete normally within seconds, like it did before the macOS “Golden Gate” update, instead of hanging indefinitely and eventually forcing “Agent stopped retrying”.
Operating System
MacOS
Version Information
Cursor IDE 3.21.13 (build 18 Sep 2026), macOS (Golden Gate)
For AI issues: which model did you use?
Not model-specific — the hang happens on the tool-call/file-search layer regardless of which model is selected (reproduced with Claude Sonnet 4.5 and others).
Does this stop you from using Cursor
Sometimes - I can sometimes use Cursor