Codex app 100% CPU Usage even when only one agent is running

Potential duplicates detected. Please review them and close your issue if it is a duplicate.

• #11596
• #11161
• #11124
• #10530
• #11584

Powered by Codex Action

@beladiyaraj, are you still seeing this with the latest version of the app? We've fixed a number of performance issues since then. If you're still seeing this, please provide additional details about how we might be able to repro the problem. You mentioned "this is happening in the cloud". Are you saying that you're using a cloud thread (where the tool calls are happening on Codex cloud), or is it a local thread? Are you seeing this with all threads or just some? For example, does it happen only with really long threads?

Having the same issue, syspolicyd spikes at 100 and i can't do anything. Even when i don't do anything, it's still happening.

Observing the same thing that I mentioned in this other issue: https://github.com/openai/codex/issues/11179#issuecomment-3965877219 it sems that Codex is doing posix_spawn on a directory(!) at root of my project (not a binary) repeatedly many times per second and triggering massive syspolicyd cpu spikes.

<img width="733" height="684" alt="Image" src="https://github.com/user-attachments/assets/a7e36345-8b64-4079-bfdf-ca875c20648d" />
same situation as in others, latest version

chip: Apple M1 Pro

Adding a correlated data point from local diagnostics (Mar 5, 2026, macOS).

What looks wrong:

• CPU is mostly consumed by desktop client processes (Renderer/GPU), not a single backend call.
• Symptoms match a cluster of issues tied to:
• long-session renderer growth (#11984, #12709)
• large diff / generated-file tracking pressure (#11861, #11179, #12861)
• heavy local file I/O (#12644)

Local measurements:

• Current snapshot: total Codex app CPU ~58%, RSS ~1.3 GB (Renderer + GPU processes dominate).
• Earlier today (09:17 session): WARN=6, ERROR=360, ENOENT=339.
• Later today (09:56 session): WARN=8, ERROR=14, ENOENT=8.

What improved usage-wise (correlation, not proof):

• Keeping active change set smaller,
• avoiding large generated/build artifacts in tracked diffs,
• restarting into a fresh session when UI starts lagging.

This looks like multiple client-side perf paths rather than one single bug.

@etraut-openai I have not encountered this issue after the update has been done

@beladiyaraj, thanks for the confirmation. I'm going to leave this bug open a little longer to wait for confirmation from others who have posted here.

Yes, problem solved and the new model is awesome (GPT-5.4)!

Edit: not solved. Perhaps worktrees cause the issue...after automations have been running...

!Image
!Image
!Image

My mac is boiling running one prompt 🔥

I have that issue when running codex on RustRover on Ubuntu

Have same issue

same issue in apple m3

Similar issue with M3, syspolicyd running at very high CPU % and 'Open Files and Ports' points to the Codex app (26.309.31024 (962))

Same issue MacBook Air m2

Same issue MacBook M4 Pro

same issue, ubuntu 24.04
If deep searching is not required - there are no problem with cpu consumption, but whenever there is a task to analyze, etc after finishing it the cpu load is still near 95%+ and does not go down

Same issue, Ubuntu 24

Same Issue, Macbook cpu M5 Pro, macos 26.4, VS Code Version: 1.113.0, codex 26.324.21329, the problem wasn't present for me before last week, so now i have to stop using the app, it tries to melt the computer.

Edit: Comand line tools updated later on and reset everything in the ide, and it seems to work somewhat better now, not perfect, still heats very much, but not trying to melt it at least.

Same issue in Macbook Pro M4 pro. The problem wasn't present for me before last week, so now i have to stop using the app, it tries to melt the computer.

<img width="225" height="142" alt="Image" src="https://github.com/user-attachments/assets/3c924e10-b9c7-4211-afdf-088a455f8c4f" />

<img width="1293" height="220" alt="Image" src="https://github.com/user-attachments/assets/333f193f-80c1-4bec-99b8-f4001da59cba" />

<img width="231" height="244" alt="Image" src="https://github.com/user-attachments/assets/0828cfb6-4124-4492-bef4-31f3955a9a91" />

Same issue (High CPU usage)

Ubuntu: 25.10 Codex: 26.324.21329 Vscode: 1.113.0

For now I've reverted codex extension to version 0.4.69 and the problem doesn't occur.

Same issue here with v26.325.21211

Same in VS code on Ubuntu 24.04 LTS, plugin version 26.5325.31654

same issue on my macbook

<img width="892" height="206" alt="Image" src="https://github.com/user-attachments/assets/e8916da3-f4e4-4b7a-9af2-bc94288c58eb" />

<img width="478" height="65" alt="Image" src="https://github.com/user-attachments/assets/41ab0a97-accf-4cd1-bde5-b86c61a25326" /> same issue on my macbook m4pro

<img width="849" height="973" alt="Image" src="https://github.com/user-attachments/assets/59ef44ba-0f36-4164-bacb-6ad01e655ef5" />

<img width="845" height="983" alt="Image" src="https://github.com/user-attachments/assets/49824e3d-cb2e-4e24-b392-4d137a783913" />

Happening to me too.

First pic is when codex is open. second pic is about 30 seconds after it's closed. It's so bad that it stops Windows opening new processes, like notepad, paint, taskmanager, screensnip etc. Firefox eventually freezes. All clears up as soon as Codex is closed.

It’s been going on for a month, and OpenAI still hasn’t fixed it.

It’s been going on for a month, and OpenAI still hasn’t fixed it.

They just don't care cause it doesn't cause them problems, users still use it, even if it's pain in the ass to do so.

Probably though they don't understand why does this happen, as the entire tool is vibe-coded

It's becoming unusable. As soon as I open it up, all programs start to freeze. I can't even use the clipboard selector with win+v, or open notepad. Youtube starts stuttering. Started happening after I updated

Confirming this on Apple Silicon/MacOS and Ubuntu.

This is clearly bigger than a one-off machine-specific issue. There are so many open reports now across the app and the VS Code extension describing the same behavior: high CPU, heat, and regressions in newer versions.

At that point, telling users to keep threads short or work around it manually is not acceptable. That is not a workaround; that is asking people to avoid normal usage because the product cannot handle it reliably.

A coding tool that drives sustained ~100% CPU during ordinary use is not production-ready from a developer experience standpoint.

Please stop treating these as isolated bug reports and provide one consolidated update:

• root cause
• affected versions
• reliable mitigation
• ETA for a fix

Right now the only consistent message users are getting is “downgrade and hope” which is not good enough.

It seems this bug has been fixed in the latest preview version of the extension.

Couldn't confirm

<img width="1508" height="308" alt="Image" src="https://github.com/user-attachments/assets/54fc6670-5c8c-4f11-a560-e4c8b87273bf" />

(before starting the chat was ~3% load)

This is just ridiculous. I’m too afraid to do anything—the computer gets unbearably hot. I wonder if there’s any other plugin I can use.

<img width="758" height="402" alt="Image" src="https://github.com/user-attachments/assets/c013dd99-c162-4f0e-a85c-e329b03e3612" />

This is still happening in 26.406.31014. In fact it's now worse now as Codex doesn't close when exiting out, and it must be force shut from the Task Manager processes list (it spawns multiple processes when attempting to shutdown). It creates a massive memory leak, filling up my RAM in about 30 minutes of being open, requiring a restart to free.

There is no way to downgrade on Windows, as it's controlled by Microsoft store. My system is basically unusable when Codex is open. I have:

• 4090Ti
• 3080
• 64GB of DDR5 RAM
• 2x RAID 0 NVME drives (4TB each)
• Ryzen 9 7950X.

It grinds my system to a halt almost instantly.

<img width="1207" height="1195" alt="Image" src="https://github.com/user-attachments/assets/e9d3ceac-ae07-42db-9ee8-c34254f70bdc" />

Adding another macOS data point.

Environment:

• Mac: MacBook Pro (Mac14,9)
• Chip: Apple M2 Pro
• Memory: 32 GB
• macOS: 15.6.1 (24G90)
• Codex desktop app: 26.406.31014 (1395)
• VS Code: 1.114.0 (arm64)
• Codex extension: 26.406.31014

Behavior:

• Codex desktop app: CPU spikes hard and my MacBook gets unusually hot, even with a single prompt
• VS Code extension: same behavior
• Codex CLI in iTerm2: normal

I tested against the same project/workspace. While Codex CLI was reading/searching files, I monitored the machine
with Stats. During CLI usage, CPU stayed under ~15% and RAM usage barely changed.

Also, when I quit the Codex desktop app, the machine immediately returned to normal.

So in my case, this does not look like “Codex is inherently heavy on this repo”.
It looks much more specific to the macOS desktop app / VS Code extension local client layer.

Try running through openrouter the openai processing is almost nothing for large projects

Same issue with me on my MacBook Air M4. it is using over 100%, so more than one full core's worth of CPU power constantly. Mac is getting super hot and battery is draining.

Everything is updated. This issue still happens on M4 pro. Laptop getting hot, and CPU usage of Code Helper is crazy- around 200%

This Codex overheating issue makes the user experience terrible. And they still haven’t fixed it quickly—seriously speechless.

When I open Codex, there are two Codex Helper Renderer processes, and both hit 100% CPU, even when I’m not doing anything.

I've gone back to the CLI which doesn't have the issue. The x2 usage limit
has finished anyway on the app.

On Sun, 19 Apr 2026, 12:51 pm zhouqinsong, @.***> wrote:

gitryson left a comment (openai/codex#11981) <https://github.com/openai/codex/issues/11981#issuecomment-4275038523> When I open Codex, there are two Codex Helper Renderer processes, and both hit 100% CPU, even when I’m not doing anything. — Reply to this email directly, view it on GitHub <https://github.com/openai/codex/issues/11981?email_source=notifications&email_token=AAUAM7U3EFL63SCOCLIKS4T4WQ5KDA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTIMRXGUYDGOBVGIZ2M4TFMFZW63VHMNXW23LFNZ2KKZLWMVXHJLDGN5XXIZLSL5RWY2LDNM#issuecomment-4275038523>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAUAM7VSVZWAK232J3DGSSL4WQ5KDAVCNFSM6AAAAACVMLETFCVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHM2DENZVGAZTQNJSGM> . You are receiving this because you commented.Message ID: @.***>

I had this issue for weeks with the vs-code extension. This worked for me

https://github.com/jxhnx/codex/tree/fix/code-extension-cpu-regression/patches/openai-chatgpt-vscode-extension-26.5415.20818-broadcast-warning-hotfix

Pinned to a recent version. No guarantees. Platform macOS Tahoe 26.4.1.

Same on Windows... the perf of the desktop app is horrible. The only reason I don't recommend it to more people.

My feedback id if it helps:
d1467c55-28f3-4c72-954c-e3331663e76c

I had this issue for weeks with the vs-code extension. This worked for me https://github.com/jxhnx/codex/tree/fix/code-extension-cpu-regression/patches/openai-chatgpt-vscode-extension-26.5415.20818-broadcast-warning-hotfix Pinned to a recent version. No guarantees. Platform macOS Tahoe 26.4.1.

Just did this, didn't do anything for me.

Edit: Actually, the CPU usage is still high but my Mac isn't making as much noise. Progress I think?

Edit 2: Nah, still the same thing.

Figured out that using Codex (beta).app makes the cpu and gpu usage about 30-35% less than with the main app. Worth it for now.

Got the same problem for few weeks, now just open codex app makes my Mac overheated. No agents are running, simply open it. M4 chip with 24Gb ram, sequoia 15.4.1. Codex 26.422.30944 (2080)

Got the same problem for few weeks, now just open codex app makes my Mac overheated. No agents are running, simply open it. M4 chip with 24Gb ram, sequoia 15.4.1. Codex 26.422.30944 (2080)

I coded my own codex app. Its 9x more cpu efficient. No heating, no jet engine sound, super smooth. Highly recommend.

High CPU usage with a single agent is usually caused by one of three things in agent architectures:

1. Polling loops without backoff: The agent loop polls for state changes (file system events, model responses, background tasks) at maximum frequency instead of using event-driven notifications. This consumes 100% of one CPU core even when idle.

1. Context window management thrashing: When context approaches the window limit, the system enters a compact-expand cycle. It compacts context to free space, which triggers a state recalculation, which expands context again, which triggers compaction. We see this in our multi-agent deployment when the compaction threshold is too close to the context limit.

1. Unbounded memory growth leading to GC pressure: Agent state (conversation history, tool results, cached embeddings) grows over time. Without proactive cleanup, the garbage collector runs continuously trying to free memory, consuming CPU.

What we do to keep agents efficient:

1. Event-driven agent loop: The agent sleeps until there is work to do (user message, tool result, timer event). No polling. CPU usage is near zero when idle.

1. Progressive compaction with hysteresis: Compact when context reaches 75% of limit. Do not re-expand until context drops below 60%. The 15% gap prevents the compact-expand thrashing cycle.

1. Per-agent resource budgets: Each agent has a CPU time limit per turn. If a single reasoning step exceeds the limit, it is interrupted and the agent reports partial results. This prevents one complex prompt from monopolizing the CPU.

1. Memory management: Tool results are compacted after processing (full result compressed to structured summary). Working memory is bounded with an LRU eviction policy.

Context management architecture: https://blog.kinthai.ai/why-character-ai-forgets-you-persistent-memory-architecture

Multi-agent resource governance: https://blog.kinthai.ai/openclaw-multi-tenancy-why-vm-per-user-doesnt-scale

Found the problem for me and a workaround in #20213

<img width="1716" height="188" alt="Image" src="https://github.com/user-attachments/assets/6c007357-4e27-4b03-8658-0ed71fe342bd" />

Really want to use Codex app but keep going back to the CLI as the performance is just crazy still

Macbook(M4 pro with 48G ram, sequoia 15.6.1). Just open Codex(Version 26.506.31421 (2620)) with a chat, cpu usage raise to 200% and always keep the weird status unless I quit the app😭