Windows Desktop 26.707.9981: system cursor visually jitters when Codex becomes idle/background
What happened?
Codex Desktop for Windows causes the system cursor to visually jitter or flicker after the app enters an idle/background state. The issue is visible over Codex and other applications, including browsers.
This is visual composition jitter rather than real pointer input. During a 75-second affected sample, 2,367 cursor samples had zero cursor-handle changes and zero coordinate changes while the cursor still visibly jittered.
The behavior is timing-sensitive:
- Opening Codex Desktop is the trigger.
- Active message processing/streaming can temporarily stop the jitter.
- The jitter usually returns when the task finishes, the app becomes idle, or focus moves to another application.
- At other times it spreads to every visible page/window.
- Closing every packaged Codex Desktop process is the only repeatable immediate recovery.
- Rebooting clears the condition until Codex Desktop is opened again.
Reproduction
- Reboot Windows and confirm that the cursor is stable.
- Open Codex Desktop.
- Use a task normally, then leave the app idle or switch focus to a browser.
- Move or hold the cursor over Codex and other windows.
- Observe visual cursor jitter/flicker.
- Send a message to an active Codex task; the jitter may temporarily stop while output is streaming, then return after the task becomes idle.
Environment
- Codex package:
OpenAI.Codex_26.707.9981.0_x64__2p2nqsd0c76g0 - Bundled Chromium observed:
150.0.7871.115 - Windows 11 x64, build
26300 - GPU: Intel Arc A380
- Intel graphics driver:
32.0.101.8861 - Two physical 2560x1440 displays, approximately 120 Hz
- Codex CLI:
0.144.2
Controlled tests and exclusions
The symptom remained after these reversible, one-variable tests:
- TRACE database logging stopped and WAL truncated
- Empty
.gitprobe storm repaired; repeated short-livedgit.exeprocesses reduced to zero - UU Remote/GameViewer app, services, virtual display/input devices, and three driver packages fully uninstalled, followed by reboot verification
- Wacom Professional Service stopped/disabled with no Wacom device connected
- Old and current Intel graphics drivers
- Both 60 Hz and 120 Hz display modes
- DWM restart
- MPO and HAGS mitigations
- Mouse trails/software-cursor compatibility path
- Windows Magnifier at 100% to force an accessibility composition path
- Chromium background/occlusion throttling switches
--disable-direct-composition- Windows per-process power throttling disabled for every Codex main/GPU/renderer/app-server process
No-GPU/WARP launch modes are not viable because they produce a blank white window and very slow startup on this machine.
Several actions appeared to stop the issue briefly, but the jitter returned while the tested component remained disabled. Those correlations were therefore rejected as root causes.
Expected behavior
Codex Desktop should become idle without affecting system cursor composition in Codex or other applications.
Diagnostic interpretation
The stable cursor coordinates and handle during visible jitter point to a Windows DWM/GPU cursor-composition or presentation interaction triggered by the packaged Electron/Chromium desktop app, rather than mouse hardware or actual HID input.
Please investigate the Desktop app's idle/background renderer and GPU presentation behavior on Windows. A supported diagnostic switch or software-cursor fallback would also help affected users while the underlying issue is fixed.
Related reports: #32516, #30021, #30920, #16374.
Raw traces contain local paths and session metadata, so they are not attached publicly. Sanitized counter summaries and additional bounded traces can be provided if maintainers specify what is needed.
4 Comments
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
Powered by Codex Action
Additional independent public reports found after filing:
26.707.9981.0package. That reporter uses an NVIDIA RTX 4060 Ti and reproduced after a clean Windows 11 reinstall, lowering mouse polling rate, disabling a virtual display adapter, changing GPU preferences, disabling GPU rendering, and clearing Chromium shader/profile caches. Their WPR trace retained roughly 75-78% idle CPU, so the symptom was not explained by CPU/disk saturation.Compositionhandle growth after Codex tool-call tasks on the same 26.707 package / Electron 42 / Chromium 150 stack, including unreleased DirectComposition resource batches.Together these reports show that this is not isolated to Intel Arc, UU Remote/GameViewer, one mouse, or one Windows installation. The common component is the current Windows Codex/ChatGPT Desktop rendering/runtime path.
Adding a Windows 11 reproduction from the same cursor-composition failure family, with before-and-after counters from Codex Desktop
26.721.4979.0.Environment
26200OpenAI.Codex_26.721.4979.0_x64__2p2nqsd0c76g032.0.101.5542, dated2024-06-080.45.0.0, dated2024-01-25Symptoms
Application Hang 1002forChatGPT.exeon2026-07-28 15:10:25and again on2026-07-30 02:16:06.Controls before changing the display path
I first removed the large local-history variable:
4.07 GiB.262 MiB.466.15 MiBCodex log database. Codex recreated it at1.69 MiB.The typing and cursor lag remained.
A live sample during the degraded state did not show an SSD, paging, or machine-capacity bottleneck:
0.26 MiB/s0.93 MiB/s0.019.61 GiB0/s0.11%average0.21%averageThe direct process counters showed the composition imbalance:
dwm.exe:136%of one CPU core on average,199.2%peakChatGPT.exeprocesses combined:3%of one core on average,24.7%peak524 MiB1.86 GiBDisplay-driver intervention
Windows showed two active display adapters:
Parsec was not needed on this machine. I removed its three registered components:
All three uninstallers returned exit code
0. Windows required a reboot to unload the display and USB drivers.Post-reboot verification found:
Result after reboot
Codex typing and navigation immediately felt much more responsive. The broad cursor jitter disappeared. The remaining brief pauses occurred only while normal post-boot processes were starting.
Measured results:
| Counter | Before removal/reboot | After startup settled |
| --- | ---: | ---: |
| DWM CPU, percentage of one core |
136%average | about4.6%average || DWM CPU, peak |
199.2%| about12.4%|| DWM private memory | about
524 MiB| about148 MiBin the first post-reboot sample || Physical-disk writes | low during affected control |
0.35 MiB/saverage after startup || ChatGPT total-machine CPU | low between stalls |
0.41%average after startup |The post-reboot sample also showed
dwm.exeat only0.23%of total machine CPU across 20 logical processors.Interpretation and limit
This does not prove Parsec is the sole root cause. The recovery step included a Windows reboot, which also resets DWM and DirectComposition state. It does show:
4.07 GiBof unrelated rollout history and recreation of the large Codex log database.Parsec may be an amplifier for a Codex/Electron/Chromium presentation or DirectComposition lifetime problem rather than the original source. This also appears related to #33192 and #33776.
Useful diagnostics from the app would include:
No project names, paths, prompts, thread IDs, client information, credentials, or raw rollout contents are included.
Additional Windows reproduction report:
This may be related to the Windows cursor/composition behavior described in this issue. The report is being added to help establish that the problem can also occur specifically during the Codex startup period.