Often exceeding context window (but still 84% context left)

What version of Codex is running?

0.42

Which model were you using?

gpt-5-codex

What platform is your computer?

Darwin 24.6.0 x86_64 i386

What steps can reproduce the bug?

In the latest version, after a big task with gpt-5-codex medium I'm often running into:

⚠️ stream error: stream disconnected before completion: Your input exceeds the context window of this
model. Please adjust your input and try again.; retrying 5/5 in 3.505s…

■ stream disconnected before completion: Your input exceeds the context window of this model. Please
adjust your input and try again.

▌ Summarize recent commits

⏎ send   ⌃J newline   ⌃T transcript   ⌃C quit   119K tokens used   84% context left

I try /compact and the error still there.

Accessing through API_KEY.

What is the expected behavior?

I used to be able to go very long before running out context and actually not error out, instead just started not working as well and then /compact ed and improved behavior, or started a new session.

What do you see instead?

Erroring out quickly.

Additional information

■ stream disconnected before completion: Your input exceeds the context window of this model. Please
adjust your input and try again.
Token usage: total=119,076 input=107,201 (+ 1,871,616 cached) output=11,875 (reasoning 7,744)
To continue this session, run codex resume 01999579-12ad-7601-9d1d-0d6d24b20214.