Windows daemon becomes unresponsive with many active workspaces

[paseo-bot]

Imported from Discord.

---

Edolce (@edolce), 2026-07-27 11:10 UTC

Summary

After ~50 minutes of uptime the Paseo daemon becomes effectively unusable. It stays
alive and /api/health answers instantly, but the Node event loop is blocked ~100% of
the time, so every WebSocket request queues behind it. Opening an agent timeline takes
up to 14 minutes. Sending a message to an agent took 794 seconds.

The machine itself is idle (22% CPU, 21 GB RAM free of 63 GB) — this is internal
serialization in the daemon, not resource starvation.

Restarting the daemon does not fix it: it was restarted at 10:49 today and the same
state came back within ~50 minutes.

---

Environment

Paseo daemon      0.2.2
Paseo app         0.1.110
Electron          41.2.0  (Chrome 146.0.7680.179)
OS                Windows 11 Pro, build 26200
CPU               Intel Core Ultra 9 285K — 24 physical / 24 logical cores
RAM               63.5 GB total, 21.5 GB free at time of capture
Disk              SSD, 1422 GB free of 1862 GB
git               2.53.0.windows.3
node              v24.15.0
Defender          Real-time protection ENABLED (exclusions not readable as non-admin)

Edolce (@edolce), 2026-07-27 11:12 UTC

I will make some manual clearing of the worktrees and other things that are causing this problems, just wanted to let you know this bottleneck

Edolce (@edolce), 2026-07-27 11:37 UTC

I think this issue #2451 is linked to this

Evidence

The attached fullreport.txt was inspected as UTF-8 plain text. It records:

• 202 ws_slow_request entries, including fetch_agent_timeline_request up to 846 seconds, send_agent_message_request up to 794 seconds, and provider/subagent requests up to 763 seconds.
• Sustained event-loop delay of 1000–1220 ms at p50 and 5000–15552 ms at p99, with 10 sustained in-flight requests and heap usage around 136 MB of 146 MB.
• 302 workspace records, including 30 active workspaces, with 60 Git watched directories and 60 Git subscriptions. Twenty-four active workspaces had not been touched for several days.
• Repeated 60-second provider availability timeouts across 15 working directories: 50 Codex, 44 Claude, and 14 Codex Scheduler timeouts. Direct provider version checks completed in 142–1702 ms.
• A 43 GB repository with about 906,000 files and an empty .git directory; git status --porcelain took 1639 ms before failing there. Four active workspaces referenced that directory.
• The same degraded state returning about 50 minutes after a daemon restart.

This occurrence on v0.2.2 postdates the fixes recorded in #304 (v0.1.55) and #2365 (v0.2.0). #2451 requests a Git-integration configuration escape hatch rather than tracking this runtime failure. PR #2747 contains related, unmerged Git-read performance work.