0.2.0 context window issue and lack of features
Resolved 💬 7 comments Opened Jul 4, 2025 by itsmehatef Closed Jul 29, 2025
💡 Likely answer: A maintainer (bolinfest, collaborator)
responded on this thread — see the highlighted reply below.
awhile i appreciate the effort, i'm switching back to npm version 0.1.2.x. there are many issues introduced with 0.2.0, here is the list of the issues that i've observed:
- context window drops on a smallest tasks compared to older version. i simply asked the agent to give me summary of a component, and it dropped the context window from 100 to 41%. while in older version it would barely consume 2 or 3% tops.
- most commands are missing from the menu which makes it less intuitive and harder to use. /compact is being one.
- can't clear prompt after typing in it. must delete everything with a backspace. hitting escape should clear the prompt.
- How many time does a user have to auto approve a read command? Right now, if the client is trying to read a large file, it is asking for approval to read every X lines instead of seeking approval to read the entire file. I know they're technically not the same command because line numbers are different, but this is not efficient.
I hope we get those features back and context issue is resolved.
7 Comments
@hatefkasraei Thanks for the detailed issue! I'll reply to your points individually:
I have also noticed the context window issue, though I am not certain I am reporting the percentage correctly in Rust because even after it hits 100%, I can often continue the conversation for awhile (even though there is no
/compact). For reference, https://github.com/openai/codex/pull/1388 is the PR where I introduced this new behavior. Note that I link to the TypeScript implementation of the equivalent feature, _which approximates the token count using a heuristic instead of using the metrics returned by the model_. As such, I would have expected the Rust version to be more accurate.One piece of functionality that could change things considerably is how much output from shell commands is sent to the model. While the implementations are different, the default values for truncation appear to be the same in both versions:
https://github.com/openai/codex/blob/e0c08cea4fcf8d71ca189a495c0094fbbda79f06/codex-cli/src/utils/config.ts#L52-L53
https://github.com/openai/codex/blob/e0c08cea4fcf8d71ca189a495c0094fbbda79f06/codex-rs/core/src/exec.rs#L29-L30
Out of curiosity, are you using the same model and roughly the same configurations for both versions of the CLI?
Admittedly, the Rust CLI has not caught up in terms of commands. This work is being tracked in https://github.com/openai/codex/issues/1262.
In terms of clearing the prompt, I think that
ctrl+uis the "correct" thing, though in the widget we are using, that defaults to Undo as discussed here:https://github.com/openai/codex/issues/1399
As a shortcut for now, you can use
ctrl+ato move the cursor to the beginning of the line and thenctrl+kto kill the line. Also,alt+backspacewill kill backwards one word at a time.As of the 0.3.0 release https://github.com/openai/codex/releases/tag/rust-v0.3.0, maybe you want to run with
-s read-only -a on-failure?It's either misreporting the context window % or blowing through it way faster. I am doing some simple things with same settings as before (0.1.x) and seeing it (0.5.0) just consume the whole context window in a short bit of back and forth.
@bolinfest yeah the parameters are actually what i needed. thanks for the info. although trying on 0.7.0 now.
regarding the context window, i feel like its behaving better. either way even if it shows that it dropped, it doesn't feel/behave like it has compacted or restarted the context. im pretty sure this is a display issue more than being an actual problem.
closing this as version is no longer relevant.