Title: Codex Desktop background ambient/personalized suggestions appear to consume quota without explicit user action

[CSP-Han]

Hi everyone,

I’d like to report a quota-consumption issue I observed in Codex Desktop.

Summary

My Codex quota dropped by around 20% within a recent 5-hour window, even though I did not actively start any Codex chat or task during that time.

After inspecting local Codex logs and SQLite data, I found that the usage appears to come from Codex Desktop’s built-in background suggestion feature, something like ambient suggestions, hyperpersonalized suggestions, or proactive suggestions.

These requests do not appear in the normal thread list, so from the user’s perspective it looks like “nothing happened.” However, the local logs show that real model requests were made.

Evidence

Environment:

• Windows
• Codex Desktop
• Local project path: D:\01-codex_projection\07-多目标点
• Time: morning of June 29, 2026

From C:\Users\<user>\.codex\logs_2.sqlite, I found multiple response.create requests:

• Time range: 2026-06-29 08:30:20 to 2026-06-29 09:48:22
• 69 parsed requests using gpt-5.4 with medium reasoning effort
• client_metadata showed:
  • thread_source = system
  • request_kind = turn

The request prompt included:

Generate 0 to 3 hyperpersonalized suggestions for what this user can do with Codex in this local project...
Get an understanding of the user's intent and goals by deeply viewing their connected apps.

So these do not look like idle heartbeats or local indexing. They appear to be actual background model requests.

Additional Observations

1. These system threads do not appear in the normal threads table or the sidebar thread list.
2. agent_jobs was empty, so this was not a user-created agent job.
3. The automations directory was empty, so this was not a user-created automation.
4. There were no new Claude task files from that day.
5. These requests started shortly after Codex Desktop launched:
   • Codex Desktop started around 08:29
   • The first background suggestion requests started around 08:30
6. The current project triggered similar background suggestion requests again around 10:06, suggesting that this feature can continue to run repeatedly.

User-facing Issues

This behavior is concerning for a few reasons:

1. Quota is consumed without an explicit user-submitted task.
2. These background system threads are not visible in the normal Activity or thread list.
3. It is difficult to know which background features are calling models.
4. It is unclear where to disable ambient suggestions, personalized suggestions, or proactive suggestions.
5. If these requests inspect connected apps or project context, users should be clearly informed when this happens, how much quota it consumes, and how to turn it off.

Suggested Improvements

I think Codex Desktop should provide:

1. A clear setting such as:
   Enable proactive / personalized / ambient suggestions

2. Visibility in Activity/Usage for background requests, including:

   • timestamp
   • model
   • trigger source, e.g. system / ambient suggestions
   • project path
   • token or quota usage
   • whether connected apps were used

3. A first-run notice explaining that this feature may call models and consume quota.

4. A low-cost/default-off mode for this feature, especially for quota-sensitive users.

5. Clear labeling of system threads in local logs and ideally in the UI, rather than having them only discoverable through logs_2.sqlite.

My Interpretation

This may not be a crash-style bug, but rather a visibility and quota-control issue around background intelligent suggestions. The feature may be useful, but users should know when it is running, whether it consumes quota, and how to disable it.

Could the Codex team clarify:

• Is ambient / hyperpersonalized suggestions an official built-in Codex Desktop feature?
• What triggers it?
• Does it count against user quota?
• Is there currently an official way to disable it?
• Will these system requests be shown in Activity/Usage in the future?

Thanks.