Fresh-thread token usage is massively inflated by enabled plugins/skills/apps in Codex Desktop #18498

test


### What steps can reproduce the bug?

1. Open Codex Desktop on macOS.
2. Start a brand-new thread in a brand-new untitled Desktop folder project.
3. Send a one-word prompt: `test`.
4. Check token usage in the local Codex log database: `~/.codex/logs_2.sqlite`.
5. Repeat the same fresh-thread `test` while progressively disabling plugins/skills/apps.
6. Compare the estimated input tokens for the first real reply turn.

Observed results from my tests:

- Earlier enabled setup on a fresh thread:
  - main reply turn: estimated input around `24.2k-24.4k`, total around `32.1k`
  - title-generation pass: estimated input around `20.1k`, total around `24.9k`

- After removing `Superpowers` only:
  - main reply turn: estimated input around `21.3k`, total around `29.1k`
  - title-generation pass: estimated input around `18.8k`, total around `23.6k`

- After removing almost everything, leaving only GitHub + Google Drive plugins/apps plus remaining local/system skills:
  - main reply turn: estimated input around `10.9k`, total around `17.5k`
  - title-generation pass: estimated input around `8.5k`, total around `11.9k`

- Final stripped-down baseline with:
  - Plugins: `0`
  - Skills: `0`
  - Apps toggled off

  Run 1:
  - main reply turn: estimated input `6942`, total `14319`
  - title-generation pass: estimated input `4494`, total `8803`

  Run 2:
  - main reply turn: estimated input `6954`, total `14336`
  - title-generation pass: estimated


### What is the expected behavior?

A fresh-thread one-word prompt like `test` should have a small and fairly stable token baseline.

Enabled plugins/skills/apps should not add a large token cost unless they are actually invoked for that turn.

A title-generation side pass also should not consume several thousand input tokens for a trivial thread.


### Additional information

This new issue is about hidden prompt/context inflation and unusually high fresh-thread token usage.

I also saw malformed or weird auto-generated thread titles during some of these tests.

The final clean tests were run in a brand-new untitled Desktop folder project, so this was not just a huge old-thread-history problem.

This bug has burned usage much faster than expected, which is why I’m reporting it separately.