Skip to content

Codex Desktop 26.611.8604.0 causes intermittent system input lag on Windows despite clean logs/plugins disabled #28855

Description

@Yassycodes

What issue are you seeing?

Codex Desktop on Windows causes intermittent whole-system input lag, especially immediately after launch/reopen and during later bursts. Mouse movement and typing become visibly laggy while Codex is open. The same machine otherwise handles heavier workloads normally.

This reproduced after Codex Desktop updated from 26.611.8273.0 to 26.611.8604.0.

This appears different from the earlier local logs_2.sqlite bloat problem: after clearing/rotating logs and reproducing on 26.611.8604.0, the log DB stayed small while Codex Desktop/Electron main/GPU/renderer processes continued to spike CPU.

Environment

  • Platform: Windows desktop app from Microsoft Store
  • Codex before update: OpenAI.Codex 26.611.8273.0
  • Codex after update: OpenAI.Codex 26.611.8604.0
  • Install path after update: C:\Program Files\WindowsApps\OpenAI.Codex_26.611.8604.0_x64__2p2nqsd0c76g0
  • Machine has 32 logical processors
  • AC processor max state was verified/restored to 100%

Steps to reproduce

  1. Start from Codex Desktop on Windows with clean/small local logs.
  2. Update Codex Desktop from 26.611.8273.0 to 26.611.8604.0.
  3. Reopen Codex Desktop.
  4. Open/continue an existing local Codex thread.
  5. Observe mouse/typing lag immediately after launch/reopen and intermittently afterward.

Expected behavior

Codex Desktop should remain responsive enough that mouse movement and typing do not lag system-wide, especially on a high-spec Windows machine, even while opening a thread or doing background startup/reconnect work.

Actual behavior

Codex Desktop causes intermittent system input lag while open. The lag is worst on launch/reopen, can settle, then later recurs in bursts.

Evidence collected

A controlled update watch captured the package changing from 26.611.8273.0 to 26.611.8604.0 around 2026-06-18 12:47 AWST.

Before the update watch, a separate issue was found and mitigated:

  • %USERPROFILE%\.codex\logs_2.sqlite had grown to about 1.56 GB
  • %USERPROFILE%\.codex\logs_2.sqlite-wal had grown to about 174 MB
  • Read-only SQLite query showed 232,280 retained log rows
  • Moving logs_2.sqlite* aside while Codex was closed allowed Codex to recreate clean logs and temporarily improved responsiveness

After updating to 26.611.8604.0 and reproducing the lag:

  • logs_2.sqlite stayed around 15.98 MB
  • logs_2.sqlite-wal stayed around 7.77 MB
  • Log growth did not explain the renewed lag

During active lag after the update, a 10-second CPU delta showed:

  • Codex.exe process using about 81.4% of one core
  • another Codex.exe process using about 38.8% of one core
  • another Codex.exe process using about 20.2% of one core
  • codex.exe app-server using about 4.2% of one core

A faster 2-second burst sampler observed:

  • combined Codex/codex/node CPU bursts up to about 200% of one core
  • single Codex renderer bursts up to about 100% of one core
  • log files stayed small during these bursts

A later recurrence showed:

  • main/high-memory Codex.exe process around 81.2% of one core
  • GPU-process Codex.exe around 37.5% of one core
  • renderer Codex.exe around 12.1% of one core
  • codex.exe app-server --analytics-default-enabled around 6.2% of one core

Mitigations / A-B checks tried

  • Restored AC processor max state to 100%
  • Disabled all always-on plugin bundles
  • Prevented full-home-folder scans in local Codex instructions
  • Rotated the huge logs_2.sqlite* files so Codex recreated clean logs
  • Temporarily set ambient-suggestions-enabled = false
  • Removed active node_repl MCP block before restart
  • Set Codex/codex processes to BelowNormal, then Idle
  • Restricted Codex/codex processes to 4 of 32 logical processors
  • Killed only the Electron GPU child process so Electron could respawn it, then re-applied the guard

Result: lag still reproduced. Runtime priority/affinity containment did not fully protect system input responsiveness.

Additional observation

After removing the active [mcp_servers.node_repl] block before restart, Codex appeared to reintroduce an active node_repl MCP block on launch/update while a disabled copy also remained. However, during lag, the main measured CPU pressure came from Codex.exe main/GPU/renderer processes rather than node_repl.exe.

Related issues found before filing

This appears related to prior performance reports, but the version/timeline and the clean-log reproduction look specific enough to file separately:

Logs / session sharing

The troubleshooting page says to use in-app / feedback and share the resulting session ID. I have not included raw local logs/config here because they may contain sensitive local paths or tokens. Sanitized evidence summaries are available, and a feedback/session ID can be added if needed.

Metadata

Metadata

Assignees

No one assigned

    Labels

    appIssues related to the Codex desktop appbugSomething isn't workingperformancewindows-osIssues related to Codex on Windows systems

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions