Codex can't read files

Open 💬 4 comments Opened Feb 12, 2026 by davidgilbertson

💡 Likely answer: A maintainer (github-actions[bot], contributor) responded on this thread — see the highlighted reply below.

I think this is new since 0.99, Codex in Windows runs Get-Content README.md and gets nothing back.
Sometimes, if pushed it will say that it can do it with escalated permissions. When asked, it said "I then retried with sandbox_permissions: "require_escalated" (you approved it)." Although I didn't see any approval request.

This had been working fine for quite a while, I don't think I've changed any permissions that should break it. My config.toml:

sandbox_mode = "workspace-write"
personality = "pragmatic"
model = "gpt-5.3-codex"
model_reasoning_effort = "medium"

[sandbox_workspace_write]
network_access = true

[features]
powershell_utf8 = true
elevated_windows_sandbox = true

[tui]
status_line = ["model-with-reasoning", "current-dir", "context-remaining", "five-hour-limit", "weekly-limit"]

Another quote from Codex: "I’m currently unable to read files from this session because every shell command is failing with no output (including Get-Content, Get-ChildItem, and pwd)"

View original on GitHub ↗

4 Comments

github-actions[bot] contributor · 6 months ago

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

#11482
#9341

Powered by Codex Action

etraut-openai contributor · 6 months ago

Thanks for the bug report. Can you confirm that if you downgrade from 0.99.0 to 0.98.0 that the problem goes away? If this is a regression, that increases the priority significantly, so confirmation would be helpful.

It's possible that this is a model behavior issue, but I can't verify without additional details.

If you are able to repro this, it would be helpful if you could use the /feedback slash command to upload your logs and session details and post the thread ID here.

davidgilbertson contributor · 6 months ago

With:

experimental_windows_sandbox = false
elevated_windows_sandbox = true

It's failing back as far as I tested 0.99, 0.98, 0.97, 0.96, so not a regression.

With experimental_windows_sandbox = true it works.

I think what happened is that I mostly use it through an IDE but today used the CLI for the first time in a while.

I did log this via /feedback and clicked the URL to create this issue which looks like it should embed the thread ID. Anyhoo, ran it again and another thread ID is 019c4f6d-99b7-7411-9df2-891ebb3dd748.

BTW I assume it's known that the VS Code extension becomes super laggy to the point of being unusable after a dozen or so prompts (hence me using the CLI).

kky42 · 6 months ago

Models' behaviors in 0.99 are very strange. They barely work, just behave like a chatbot.