Cursor-agent pegs CPU (~100%) / silent freezes for months — still on 2026.11.11-b912975

Where does the bug appear (feature/product)?

Cursor CLI

Describe the Bug

cursor-agent repeatedly locks up in two ways that have been hitting me for months
(since at least June 2026), still on latest CLI:

  1. CPU peg (“pegged”) — process sticks at ~100% on one core, no children, no output.
    Often after tools finish. SIGTERM ignored; needs kill -9. Stack samples show hot
    Array.filter + RegExp.test with merkle/ignore matching over large trees.
  2. Silent mid-turn freeze — agent just stops. Diagnosed once as HTTP/2 stream drop; I
    set network.useHttp1ForAgent: true on 2026-11-05. Pegs still happen.
    I have had to build external kill tooling (wedged-cursor-agents.ts: CPU≥80%, etime≥5m,
    no children → SIGKILL) because the CLI does not recover. A CPU watcher is not an
    acceptable product fix.
    Steps to Reproduce
    Intermittent, but highly repeatable over weeks:
  3. Run interactive cursor-agent in Ghostty (or headless via a long-lived worker /
    twist run q) on a non-trivial repo.
  4. Work normally for a while (tools, edits, commits). Large generated trees (e.g.
    cdk.out with 100k+ files) make the peg class much worse; it also happens on ordinary
    repos.
  5. Eventually the session stops progressing. ps shows cursor-agent at ~100% CPU with
    no children, or the process is idle with no further output.
  6. Attempt soft kill: SIGTERM often ignored. Need SIGKILL.
  7. Orphaned/stale agents can remain and leave the machine already hot for the next
    launch.

Steps to Reproduce

Intermittent, but highly repeatable over weeks:

  1. Run interactive cursor-agent in Ghostty (or headless via a long-lived worker /
    twist run q) on a non-trivial repo.
  2. Work normally for a while (tools, edits, commits). Large generated trees (e.g.
    cdk.out with 100k+ files) make the peg class much worse; it also happens on ordinary
    repos.
  3. Eventually the session stops progressing. ps shows cursor-agent at ~100% CPU with
    no children, or the process is idle with no further output.
  4. Attempt soft kill: SIGTERM often ignored. Need SIGKILL.
  5. Orphaned/stale agents can remain and leave the machine already hot for the next
    launch.

Expected Behavior

Agent completes, errors clearly, or times out and exits. No sustained single-core 100%
spin. SIGTERM terminates the process. Indexing/ignore matching cannot busy-loop forever.
Transport failures surface and fail/retry the turn.
Actual Behavior
Pegged cores and silent freezes. Mid-task work lost. External kill scripts required to
keep CLI usable. Today died mid JobRunr YAML commit on a Java monorepo.
Screenshots / Screen Recordings
Can attach Activity Monitor / ps snapshot and a process sample on next peg. Prior Oct 9
peg: ~100% for ~27 minutes; SIGTERM ignored; sample showed Array.filter +
RegExp.test + merkle thrash.

Operating System

MacOS

Version Information

Version Information

CLI Version 2026.11.11-b912975
Latest 2026.11.11-b912975 (up to date)
Model Auto
Subscription Tier Enterprise
OS darwin (arm64)

For AI issues: add Request ID with privacy disabled

Cant its corporate account.

Additional Information

Config already applied without solving it:

  1. Fix pegged-core ignore/merkle thrash with an internal timeout.
  2. Fix silent mid-turn freezes even with HTTP/1 forced.
  3. Make SIGTERM reliably kill stuck agents.
  4. Tell me what dump you want on the next live peg (sample / spindump / agent log).

Does this stop you from using Cursor

Yes - Cursor is unusable

Hey @jeengland, From our side, the turns of yours that died recently (today’s included) all look the same: the CLI stopped answering while a tool call was still in flight, then the connection closed.

They don’t look like dropped connections, so the HTTP/1 setting wouldn’t be expected to change them. Some of your sessions are also still connecting over HTTP/2, so it’s worth checking that network.useHttp1ForAgent is in the ~/.cursor/cli-config.json those sessions read.

SIGTERM is handled on the CLI’s main JavaScript thread, so while that thread is spinning the handler never gets to run. That’s why only kill -9 works during a peg, and it’s part of what I passed along.

For the next live peg, before your kill script fires:

sample <pid> 10 -file /tmp/cursor-agent-peg.txt
ps -o pid,ppid,etime,%cpu,command -p <pid>

The full sample file and that ps line, attached here, are what we need.