Same silent-stall cluster on a purely local Windows workspace — a week of logs, including the shouldBlockFileRead watchdog signature
Adding a second data point to this thread, since the questions asked earlier were quite specific. My
setup hits the identical signature, but importantly on a local NTFS disk — no Remote-SSH, no UNC
path, no network share, so the corporate-network angle can be ruled out for this instance.
Answers to the questions asked earlier in the thread
-
Workspace location: local disk, C:\Users\...\Documents\PROJECTS\archive_v2. Not a network
share, not a mounted folder, not UNC.
-
Repo size: 52,926 files / 2.15 GB, containing 13 nested git repositories (several
sibling worktrees of the same projects). 27,286 of those files are under .git, 18,294 under
Rust/Java target/.
-
.cursorignore / .gitignore: neither existed at the time of the incidents. (I have since
added a .cursorignore that hides 87% of the tree — will report back whether it changes the
frequency.)
-
Small local test project: not yet tested, I will follow up.
Environment
-
Cursor 3.19.13, 3.19.19, 3.20.7, 3.20.10, 3.20.17 — the signature appears in all of them
-
Windows 10.0.26200 x64, Chromium 148.0.7778.280
-
Peak process memory 7.02 GB, peak activeComposersCount 73
Symptom
Agent runs stop mid-execution with no indication whatsoever in the UI — no error, no timeout,
no “taking longer than expected”, no abort. The run simply produces nothing further. Restarting the
run often lands in the same state.
What the logs show
1. Extension host goes UNRESPONSIVE — 731 events between 2026-09-08 and 2026-09-14.
[error] {"key":"composer","message":"Extension host became UNRESPONSIVE",
"client_version":"3.20.7","layout":"unifiedAgent",
"extensionHostKind":"LocalProcess","activeComposersCount":"73","extensionHosts":"[]"}
695 of them on a single day. 562 occurred between 19:00 and 23:00 local during continuous active
use — I checked the Windows Kernel-Power event log and there was no sleep/Modern-Standby
transition in that window, so these are not suspend artifacts. 548 carried
activeComposersCount: 73.
2. The same watchdog signature reported earlier in this thread — ~750 trips in one week:
Count
Message
275
[watchdog, LocalReadExecutor] permissionsService.shouldBlockFileRead still not completed after 3000ms
131
[watchdog, LocalReadExecutor] getFormatForFile still not completed after 3000ms
113
[watchdog, LocalWriteExecutor] writeFile still not completed after 3000ms
88
[watchdog, LocalReadExecutor] readTerminalFile still not completed after 3000ms
56
[watchdog, LocalReadExecutor] readTextStreaming still not completed after 7000ms
54
[watchdog, LocalWriteExecutor] getBufferForWrite still not completed after 3000ms
33
[watchdog, LocalWriteExecutor] permissionsService.shouldBlockWrite still not completed after 3000ms
The watchdog detects the hang and logs it — and then nothing recovers the operation. No timeout, no
fallback, no surfaced error. That gap is the actual bug.
3. The search is not slow — the layer wrapping it hangs. Span durations from
cursor.requestTraces.log:
Span
n
median
p90
max
>10 min
LocalGrepExecutor.executeRg (actual ripgrep)
1499
446 ms
2.0 s
25.8 s
0
LocalGrepExecutor.execute (its parent)
972
4.2 s
66 s
3.9 h
35
LocalGrepExecutor.runFilesMode
889
1.4 s
119 s
3.9 h
45
streamFromAgentBackend
343
186.7 s
18.1 min
2.6 h
78
networkPhase
635
52.7 s
13.1 min
2.6 h
83
agent.request
476
37.4 s
7.7 min
2.6 h
39
This is the part I would most like the team to look at: executeRg never exceeds 26 seconds,
while its own parent span execute runs for hours. Ripgrep completes fine every single time. The
stall lives entirely between the tool call and the search process, in the permission/orchestration
layer. A big workspace amplifies it but cannot be the cause.
(Caveat in the interest of accuracy: the multi-hour maxima are partly inflated by Modern Standby.
The medians, p90s, the 19:00–23:00 cluster and all watchdog counts are unaffected.)
4. Why nothing is visible — spans never reach a terminal state. Started and never completed,
across all sessions: 655 child.span, 36 exec.handle, 32 LocalGrepExecutor.execute,
19 AgentCompatService.runAgentLoop, 18 each of agent.request / streamFromAgentBackend /
getAgentStreamResponse, 17 networkPhase, 5 subagent.request.
An agent.request that neither completes nor errors is precisely the user-visible symptom.
5. First-token latency ladder: No first token received within Ns — 2s: 183×, 4s: 150×,
6s: 97×, 8s: 59×, 10s: 38×, 12s: 27×, 14s: 22×, 16s: 17×, 32s: 3×.
6. Transport errors: 2155 × Connect error in unary AI connect, 193 × Stream error after headers in extension host, 184 × Stream error reported from extension host, 103 × Agent exec stream start failed, 89 × Agent client stream abort. Also getaddrinfo ENOTFOUND api2.cursor.sh
after resuming from standby, with no recovery of in-flight runs.
A possible amplifier worth checking
The stack posted earlier in this thread shows _checkUnresponsive firing while the host is blocked
in getCurrentBranch → _refreshCachedBranch → $gitStatusWasRun. With 13 git repositories in
one workspace root, each git status refresh fans out into 13 branch lookups over the same
extension-host IPC channel that shouldBlockFileRead uses. That channel saturating would explain
why file reads, writes, terminal reads and search all stall together rather than independently.
Why this matters commercially
Because the stall is silent, the cost is not just waiting:
-
Every restart re-sends the full context, so the same work gets billed two or three times over.
-
Restarts frequently hit the same stall, so some work is never completed at all.
-
On one machine this consumed multiple hours of development time in a single week.
The same workflow against the same backend via the JetBrains/IntelliJ integration does not show
this behavior, which is what points me at the Cursor Desktop extension-host architecture rather
than the model backend.
What I would ask for
-
Make the watchdog act, not just log. The 3000/7000 ms watchdogs already detect the hang —
they should fail the operation so the agent can error out or retry.
-
Never leave an agent run without a terminal state. If the extension host goes UNRESPONSIVE,
surface it in the agent UI and fail the run.
-
Do not consume quota for runs that produced no output because the client froze.
-
Bound git status/branch refresh for workspaces with many nested repositories.
-
Warn on or virtualize large activeComposersCount — 73 live composers visibly degrades the
host.
Happy to supply raw logs, Process Explorer captures or specific request IDs if that helps.