Codex Desktop ambient_suggestions consumes ~70k tokens in the background
Codex version
Codex Desktop for Windows 26.623.5546.0
Environment
- Windows 11 Pro
- Codex Desktop signed in with ChatGPT
- Date observed: 2026-06-28
- No app-server reconnects were recorded during the affected period
Problem
Codex Desktop generated home-page ambient suggestions in background/ephemeral turns and consumed a very large amount of included usage despite no corresponding user-requested response.
The desktop log recorded two successful ephemeral_generation_token_usage events for feature=ambient_suggestions:
| Time (UTC) | Model | Input | Cached input | Output | Reasoning output | Total |
|---|---:|---:|---:|---:|---:|---:|
| 2026-06-28 00:28:15 | gpt-5.4 | 34,478 | 34,176 | 157 | 140 | 34,635 |
| 2026-06-28 00:28:24 | gpt-5.4 | 35,776 | 34,688 | 15 | 0 | 35,791 |
| Total | | 70,254 | 68,864 | 172 | 140 | 70,426 |
The app also logged the generated turns as unknown conversations:
Received turn/started for unknown conversation
ephemeral_generation_token_usage feature=ambient_suggestions ...
Received turn/completed for unknown conversation
The installed app bundle indicates that ambient-suggestions-enabled defaults to true.
Steps to reproduce
- Open Codex Desktop with existing long conversations/projects.
- Visit or remain on the home page.
- Wait for ambient suggestions to appear.
- Inspect the desktop log for
ephemeral_generation_token_usage feature=ambient_suggestions.
Expected behavior
Background home-page suggestions should not silently consume a large amount of the user's included Codex quota. At minimum, the feature should clearly disclose its usage impact and expose an obvious opt-out before generation.
Actual behavior
Two unsolicited background generations consumed 70,426 total tokens in roughly nine seconds. The visible generated output was only 172 tokens.
Additional notes
- This was not caused by network reconnect retries: the desktop log showed zero non-zero reconnect attempts during the affected app session.
- Please confirm whether
ambient_suggestionsusage is intended to count against user limits. - Please consider disabling the feature by default, limiting context size, or making the opt-out prominent.
- Please advise whether quota consumed by these background generations can be restored.
Sensitive account information and conversation text have intentionally been omitted.
This issue has 3 comments on GitHub. Read the full discussion on GitHub ↗