Unexpectedly fast context/token depletion and sudden usage drops across Codex project chats

What version of the Codex App are you using (From “About Codex” dialog)?

26.623.42026

What subscription do you have?

Codex Pro Plan

What platform is your computer?

Windows nt x64

What issue are you seeing?

Across multiple long-running Codex project chats, context/token usage appears to burn down unusually fast. In some cases, the available context seems to drop suddenly by a large amount, even when the visible user/assistant messages do not appear large enough to explain the drop.
At first this looked like delay or recovery overhead, but the behavior appears broader than recovery workflows. It happens across project chats generally, including during normal discussion, planning, QA, and tool-assisted work.

What steps can reproduce the bug?

Use several long-running Codex project chats.
Continue normal project work involving planning, repository instructions, tool calls, and occasional recovery/checkpoint prompts.
Watch the context/usage window over time.
Observe that usage sometimes decreases much faster than expected or drops suddenly.
Compare the visible amount of new text/tool work against the size of the apparent context reduction.

What is the expected behavior?

Context usage should decline in a way that roughly corresponds to visible conversation, tool output, attached context, and known recovery reads. If large hidden overhead is being counted, the UI should make that easier to understand.
Actual Behavior
Available context appears to deplete rapidly and sometimes suddenly across multiple chats. The cause is not clear from the visible transcript. This makes it hard to predict when compaction will occur or manage project workflows efficiently.
Impact
This causes frequent compactions, interrupts long-running project work, increases recovery overhead, and makes it difficult to understand whether token usage is coming from visible conversation, tool output, hidden context, app behavior, or a possible accounting/display issue.

Additional information

This is not limited to one specific project thread or one recovery workflow. Recovery/reanchor flows can make the issue more noticeable, but the broader problem appears to be unexplained fast context depletion and sudden usage drops during normal Codex project-chat use.
Useful debugging information to expose might include:
visible transcript tokens
tool-output tokens
hidden/system/context overhead
attachments or injected file-context size
compaction threshold behavior
before/after token meter changes per assistant turn