[Windows][Desktop 26.810.4967.0] Opening or switching conversations can trigger a persistent 350-800 MiB/s read loop and system-wide stutter

What version of the Codex App are you using (From “About Codex” dialog)?

26.810.4967.0 (x64 MSIX)

What subscription do you have?

Unknown / not relevant to this local desktop performance issue

What platform is your computer?

Microsoft Windows NT 10.0.26200.0 x64 (Windows 11 Home, 25H2)

What issue are you seeing?

Opening or switching to a local conversation in the Codex Windows desktop app can immediately cause system-wide UI and mouse stutter and start a persistent high-rate read loop in the app's main ChatGPT.exe process.

The first reproduction followed opening a very long local conversation (approximately 25 MiB local session transcript). A later independent reproduction showed the same read-loop signature without a clearly identifiable long-conversation trigger. Long history may increase the likelihood, but it does not appear to be required; conversation navigation or an associated state refresh/hydration path is the broader trigger surface.

Once triggered, the read loop continues after the active task finishes and after switching to a different conversation. Fully quitting and restarting Codex clears it.

Read-only performance samples:

• Normal state immediately after restart (8.12-second sample): main ChatGPT.exe read rate 0.03 MiB/s; CPU 0.22%.
• First triggered state (10.37-second sample): main ChatGPT.exe logical read rate 695.82 MiB/s; CPU 7.51% of total host CPU.
• First triggered physical-disk sample: 356.8 MiB/s average read, 438.3 MiB/s maximum; queue 0.50 average, 0.85 maximum.
• Later reproduction (12.44-second sample): main ChatGPT.exe logical read rate 685.07 MiB/s; CPU 7.98%.
• Later physical-disk sample: 417.47 MiB/s average read, 693.93 MiB/s maximum; queue 0.40 average, 2.00 maximum.
• Immediate 6.01-second whole-system follow-up: the main ChatGPT.exe process read 792.61 MiB/s at 7.93% host CPU. The next-highest process read only 2.71 MiB/s, and the codex.exe backend read 1.26 MiB/s.
• About 16.9-17.3 GB of RAM remained available during the later reproduction.
• DWM, GPU/renderer processes, Defender, indexing, and the codex.exe backend were not saturated.

No transcript content, local path, account data, or logs are attached.

What steps can reproduce the bug?

1. Fully quit Codex and start it again.
2. Confirm that Codex and the rest of Windows are smooth.
3. Open or switch between existing local conversations. A very long conversation triggered the first reproduction, but a later reproduction was not tied to one clearly specific long conversation.
4. Observe immediate stutter across the desktop, not only inside Codex.
5. Switch to another conversation or wait for the active task to finish.
6. Observe that the high-rate read activity and stutter continue.
7. Fully quit and restart Codex; the stutter disappears.
8. Continue opening or switching conversations; the issue can retrigger, but the exact triggering conversation is not deterministic.

Reproducibility: observed repeatedly. Restarting clears the issue, but avoiding one particular conversation is not a reliable prevention method.

What is the expected behavior?

Opening or switching conversations should use bounded or lazy loading and should not continuously reread local session data or degrade system-wide responsiveness. Navigating away from a conversation or completing its task should cancel or stop associated hydration, replay, or refresh work.

Additional information

Only reliable workaround observed: fully quit and restart Codex. Continuing in a fresh, focused conversation may delay recurrence, but it does not guarantee prevention.

Related issues:

• openai/codex#38506 reports a broader post-update regression with high memory, high disk activity, and system-wide mouse lag. This report may share a root cause, but adds a repeatable main-process logical read loop, ample available memory, and restart/retrigger A/B evidence.
• openai/codex#33230 reports transient freezes when opening history conversations after local history stores become large.
• openai/codex#33074 reports system-wide mouse stutter during startup and task switching without a measured persistent read loop.
• openai/codex#32154 reports persisted-history replay and retained MCP stacks during navigation. In this reproduction, the dominant I/O was in the Windows main ChatGPT.exe process, and the loop persisted after navigation and task completion.

No session ID, thread ID or title, transcript content, local file path, account name or email, project name, screenshot, or log is included. Additional sanitized performance counters can be provided if needed.

Start by reproducing the issue through the Windows desktop conversation-opening and switching flow while monitoring the main ChatGPT.exe read loop. Trace the associated hydration, replay, or refresh work and verify that navigation or task completion stops it; done means conversation changes no longer cause persistent rereading or system-wide stutter.