Context window size at launch should approximate effect of `AGENTS.md`
Resolved 💬 4 comments Opened Nov 22, 2025 by guidedways Closed Jan 14, 2026
💡 Likely answer: A maintainer (github-actions[bot], contributor)
responded on this thread — see the highlighted reply below.
What version of Codex is running?
0.63.0
What subscription do you have?
Pro
Which model were you using?
gpt-5.1-codex-max ex high
What platform is your computer?
macOS
What issue are you seeing?
Context window size displayed 100% at launch, then suddenly 95% after the first message - it seems the AGENTS.md consumed 5% but this wasn't immediately obvious at initial lunch without first sending a message
What steps can reproduce the bug?
Uploaded thread: 019aaa37-ec3e-73e2-a25c-35b7b598f82c
- Create a large AGENTS.md
- Start
codex - It says 100%
- Type 'Hello'
- Context window drops to 95% (my AGENTS.md is 40kb)
At launch:
<img width="1088" height="652" alt="Image" src="https://github.com/user-attachments/assets/a4a4b5cc-eb27-4caf-acc8-9d5957de23a0" />
After first message:
<img width="956" height="884" alt="Image" src="https://github.com/user-attachments/assets/48f39c18-ac01-47c4-af1d-c8dd08e3f03a" />
What is the expected behavior?
At launch the remaining context window size should be accurate, should reflect how much space has been consumed already by AGENTS.md
Additional information
_No response_
4 Comments
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
Powered by Codex Action
The AGENTS.md contents are included as part of the first message sent to the model. It wouldn't be accurate for the context window size to reflect a value less than 100% until after the first message is sent. Prior to sending the first message, the CLI has no way to determine the context window consumed that will be consumed by AGENTS.md. Tokenization and context window accounting are done server-side.
@etraut-openai agreed, I understand it's rather wasteful to send out a message at launch just to have a more accurate context count showing up front, which is why even if it's an
approximation- that would help.Right now it's jarring and confusing to see a 5% drop on
Hello. Context window percentage just needs to be closer to being accurate at different times (from launch onwards) because it truly helps with how one interacts with codex. For instance, if I see a ~5% drop at launch (without my first message) I'll know something's off (if that wasn't expected) and can fine-tune AGENTS.md or anything else such as an MCP server that's connected and consuming a lot of space.This feature request has received only one upvote, so I'm going to close it.