First message in fresh session consumes ~100k input tokens immediatly, no matter how trivial the input

[afolkest]

What version of Codex CLI is running?

0.121.0

What subscription do you have?

Pro

Which model were you using?

gpt-5.4 xhigh and gpt-5.3-codex-spark high

What platform is your computer?

Darwin 24.6.0 arm64 arm

What terminal emulator and version are you using (if applicable)?

iTerm2

What issue are you seeing?

Whenever I open a fresh Codex session and send a simple message, it consumes about 102k-105k tokens immediately.

What steps can reproduce the bug?

Input:
codex exec "Say hello world"

Output:

OpenAI Codex v0.121.0 (research preview)
--------
workdir: <redacted>
model: gpt-5.4
provider: openai
approval: never
sandbox: workspace-write [workdir, /tmp, $TMPDIR, /Users/<redacted>/.codex/memories]
reasoning effort: xhigh
reasoning summaries: none
session id: 019da3a1-d1d5-7a22-950b-3839f72b8e20
--------
user
Say hello world
codex
hello world
tokens used
105,156

What is the expected behavior?

No response

Additional information

No response

[github-actions]

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

• Fresh-thread token usage is massively inflated by enabled plugins/skills/apps in Codex Desktop #18498
• Memory read path ignores cwd and injects entire global memory_summary.md into initial context for new sessions #17496
• Regression: Token usage significantly increased in v0.121.0 vs v0.116.0 for identical prompt and unchanged workspace #18345

Powered by Codex Action

[etraut-openai]

You must have a very large AGENTS.md or a large number of MCPs. You should look at the rollout file (you can ask codex to find it for you under ~/.codex/sessions).

When I try to repro the problem (with relatively few MCPs and a minimal AGENTS.md file), I see a value of about 22K tokens usage for the first prompt.

[afolkest]

I reproduced this issue in a freshly initialized empty repo with no AGENTS.md and with my MCPs removed from Codex.
I went ahead and downgraded Codex to 0.115 and the problem is gone. Looking at the rollout file I see nothing corresponding to 100k tokens.

The issue showed up immediately after upgrade to 0.121 with no other changes in my setup.

[quark-zju]

Could it be the "memories" feature? I noticed that after enabling "memories", codex tries to summarize past sessions into memories, which consumed a lot of tokens. The raw_memories.md is also 3.4k+ lines, which seems to be too long to be injected in context. I have to disable the memories feature entirely.

[etraut-openai]

@afolkest, are you still seeing this with the latest version of the CLI? We did discover a regression that made it in to 0.121 that could have explained what you were seeing. We fixed it shortly after, and the fix was in 0.122.0 and all subsequent versions.

[afolkest]

@etraut-openai It appears to be fixed now! Many thanks for getting back to me.