Codex Desktop profile permanently undercounts skill/plugin usage and labels skills as plugins
What version of the Codex App are you using (From “About Codex” dialog)?
Codex Desktop 26.715.3651.0
Bundled runtime/plugin release observed locally: 26.715.31251
What subscription do you have?
ChatGPT subscription (exact tier omitted for privacy)
What platform is your computer?
Windows x64
What issue are you seeing?
Settings → Profile → Activity Insights has remained stuck for more than one week:
- Skills explored: 1
- Total skills used: 1
- Most used plugins:
$openai-docs— 1 run - Total tasks: 260
This account has used multiple Codex skills and bundled plugins across many tasks, so reporting only one unique skill and one total invocation does not reflect actual usage.
There is also a labeling problem: the $ prefix identifies openai-docs as a skill, but it is displayed under the heading “Most used plugins”. Skills and plugins therefore appear to be mixed under a misleading label.
Other profile statistics load normally, including total tasks, reasoning-effort distribution, and fast-mode percentage. Desktop logs contain no profile_usage_query_failed or stats_error entries. This suggests the problem is specific to skill/plugin usage aggregation rather than a general profile request, authentication, or rendering failure.
The values remain unchanged after full app restarts and for much longer than the client-side profile cache period.
What steps can reproduce the bug?
- Use several different Codex skills/plugins across multiple tasks.
- Open Settings → Profile → Activity Insights.
- Check Skills explored, Total skills used, and Most used plugins.
- Wait more than six hours and/or fully restart Codex.
- Observe that the values remain stuck at one skill and one invocation.
What is the expected behavior?
- Distinct and total skill/plugin invocations should update consistently with actual Codex usage.
- Skills should have their own section, or the heading should accurately describe both skills and plugins.
What is the actual behavior?
The profile reports only $openai-docs with one invocation for more than a week despite 260 tasks and repeated use of multiple skills/plugins.
Additional information
- The rest of the profile statistics render normally.
- Restarting Codex does not fix the problem.
- This does not appear related to local marketplace source paths or plugin installation state.
- The symptom is related to, but more specific than, #25479: that report concerns broadly stale usage statistics, while this report concerns the skill/plugin invocation counters and their UI classification.
- No auth tokens, account IDs, local usernames, or personal identifiers are included.
3 Comments
New finding: only skill/plugin metrics fail to aggregate Windows usage after signing in on macOS
I confirmed a cross-platform condition that narrows this issue significantly.
After signing into the same account on Codex Desktop for macOS:
unique_skills_used/ Skills exploredtotal_skills_used/ Total skills usedtop_invocations/ Most used pluginsThis makes a general Profile cache or account-authentication problem unlikely. The issue appears specific to cross-platform aggregation of skill/plugin invocation telemetry.
Possible failure points include:
Because the other Profile metrics still aggregate normally, this seems isolated to the pipeline producing
unique_skills_used,total_skills_used, andtop_invocations, rather than the entire/wham/profiles/meresponse.Confirming the same issue on another Windows installation.
Environment
26.715.10079.00.145.0-alpha.30Observed profile values
117$impeccable — 17 runsI inspected the cached response from
/backend-api/wham/profiles/me. The incorrect values already originate from the backend:This rules out a renderer or stale local-cache problem: the response was freshly generated, had no stats error, and already omitted the other skills.
Local session records contain uses of additional skills before the aggregation cutoff, including
grill-me,playwright-interactive, and plugin-provided skills. These do not appear in the backend aggregate.The likely failure point is therefore the skill-usage ingestion or identity-normalization pipeline. It may be recording only a subset of invocation types or skill origins while dropping implicit, bundled, personal, repository, or plugin-owned skill usage.
There is also a labeling issue: the API identifies
$impeccableastype: "skill", but the UI displays it under Most used plugins.Suggested areas to investigate:
Confirming that this is still reproducible on a newer Windows build.
Environment
26.721.4979.0codex-cli 0.146.0-alpha.3.1Current Profile → Activity Insights
사용한 스킬):10총 사용한 스킬 수):13204No plugins used yetThis account has repeatedly used multiple skills and plugin-provided capabilities across tasks, but the reported skill invocation total remains far below the observed usage and the plugin section remains empty.
Local skill/plugin discovery is healthy: skills are loaded and plugins are installed and enabled without local configuration warnings. Other profile metrics render normally. This is consistent with the backend skill/plugin aggregation problem described in this issue rather than a local installation or discovery failure.
The app version above is newer than the
26.715.*builds already reported, so the issue appears to persist in26.721.4979.0.