You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Codex Desktop on macOS keeps renderer/GPU active while visible, causing high WindowServer CPU and large memory footprint #18567
On macOS, Codex Desktop appears to keep its renderer/compositor pipeline active while the app window is visible but not frontmost. With only two chat threads open, this results in sustained WindowServer CPU usage, sustained Codex Helper (GPU) / Codex Helper (Renderer) activity, and a total Codex-related memory footprint around 1.6–1.7 GB.
This does not seem proportional to the current workload.
Environment
App: Codex Desktop
Version: 26.415.40636
Build: 1799
Electron: 41.2.0
macOS: 26.3.2 (25D2150)
Hardware: Apple Silicon Mac
GPU: Apple M5 Pro, 16-core GPU
Observed behavior
Codex was:
visible
not frontmost
not minimized
Even while not actively using the app, its rendering-related processes remained active.
This suggests the app continues participating in display refresh/compositing while visible.
Codex support files
I also checked ~/Library/Application Support/Codex/.
GPUCache, DawnWebGPUCache, and DawnGraphiteCache were open, but during a short observation window they did not appear to be the primary source of active churn. The only obvious file changes were to:
sentry/scope_v3.json
Cookies
So this currently looks more like a sustained rendering/compositing issue than heavy cache-write pressure.
Expected behavior
If Codex is visible but not actively used, renderer/GPU activity should drop significantly, or at least should not remain this high for such a light workload.
Actual behavior
Codex remains visually/compositionally active while merely visible, causing:
sustained WindowServer load
sustained renderer/GPU load
disproportionately high memory footprint
Impact
noticeable GPU / WindowServer spikes
heavier-than-expected system load
memory usage that feels too large for only two open threads
I am seeing a likely related rendering/compositor issue on an Intel Mac.
Environment:
Codex App: 26.422.30944 (CFBundleVersion2080)
Electron: 41.2.0 (from the running crashpad process annotation)
macOS: 13.7.8 (22H730)
Darwin: 22.6.0x86_64
CPU: Intel(R) Core(TM) i7-7700HQ CPU @ 2.80GHz
Observed behavior:
The Codex window intermittently turns fully gray, as if the current rendered window is hidden behind an opaque/translucent gray layer.
It can later render normally again and remain interactive, so it does not look like a hard app crash.
At the time I checked processes, Codex rendering processes were active: Codex Helper (Renderer) was using high CPU and Codex Helper GPU process was also active.
This feels consistent with a visible-window compositor/GPU path issue rather than model/session-specific behavior.
I can reproduce something very similar, but with Stage Manager already OFF.
In my case, the key finding was that ~/.codex/state_5.sqlite contained many threads.title values that were absurdly large (some over 200k chars). These were effectively full prompt bodies, not normal short chat titles.
After trimming those long threads.title values locally, Codex became responsive again.
So this may not only be a compositor / visible-window issue. There may also be a bug where very large first prompts get persisted into thread titles or related UI metadata, and the desktop app keeps reprocessing them.
Activity
github-actions commentedon Apr 19, 2026
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
Powered by Codex Action
ybratec commentedon Apr 25, 2026
I am seeing a likely related rendering/compositor issue on an Intel Mac.
Environment:
26.422.30944(CFBundleVersion2080)41.2.0(from the running crashpad process annotation)13.7.8(22H730)22.6.0x86_64Intel(R) Core(TM) i7-7700HQ CPU @ 2.80GHzObserved behavior:
Codex Helper (Renderer)was using high CPU andCodex HelperGPU process was also active.This feels consistent with a visible-window compositor/GPU path issue rather than model/session-specific behavior.
sakataka commentedon May 6, 2026
I can reproduce something very similar, but with Stage Manager already OFF.
In my case, the key finding was that
~/.codex/state_5.sqlitecontained manythreads.titlevalues that were absurdly large (some over 200k chars). These were effectively full prompt bodies, not normal short chat titles.After trimming those long
threads.titlevalues locally, Codex became responsive again.So this may not only be a compositor / visible-window issue. There may also be a bug where very large first prompts get persisted into thread titles or related UI metadata, and the desktop app keeps reprocessing them.
Environment: