Expose active model and reasoning effort to Codex voice/mobile sessions

Open 💬 3 comments Opened Aug 2, 2026 by Hamzablm

💡 Likely answer: A maintainer (github-actions[bot], contributor) responded on this thread — see the highlighted reply below.

Summary

When using Codex through voice or mobile, the active model and reasoning-effort setting are not visible to the agent. The user may be away from their desktop and unable to inspect the model picker, while the agent also cannot read the active session configuration or operate the Codex app through Computer Use.

This makes it impossible to answer a basic question accurately: which model and reasoning effort are active in the current conversation?

Actual behavior

The session’s model and reasoning effort are not exposed to the tool-backed agent context.
The agent can only describe documented defaults, not confirm the active setting for the current session.
Desktop automation cannot operate the Codex app to inspect the picker.
On mobile/voice, the user may not have access to the desktop control at all.

Expected behavior

Expose a read-only, structured session-metadata capability that returns at least:

active model
reasoning effort
speed or service tier, when relevant
whether the setting is inherited, defaulted, or explicitly selected

This should be available to the active Codex agent and ideally visible in mobile/voice UI as well.

User impact

Users cannot verify whether they are using the intended model or reasoning level while away from the desktop. It also prevents the agent from giving a trustworthy answer about its current configuration.

This appears adjacent to #32291, which requests resolved model and reasoning metadata for tool-backed child sessions. This report concerns visibility of the currently active parent session, particularly for mobile and voice use.

View original on GitHub ↗

3 Comments

github-actions[bot] contributor · 25 days ago

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

#36636

Powered by Codex Action

ropepop · 17 days ago

Another data point for exposing session metadata to mobile: with a custom provider on the macOS Codex app, the desktop session shows the active model and renders the full reasoning summary + tool-call middle section. The same session opened from iOS remote neither exposes the active model/reasoning config nor renders the middle steps (only a flattened thinking cell) — see #37598 and #32185. Exposing structured session metadata (model, reasoning effort, tier) AND the reasoning/tool items to the remote surface would close both visibility gaps for custom-provider users.

ropepop · 15 days ago

A data point that narrows this further to a client-side render gate (2026-08-12): the token data is not missing from what reaches the phone — it's the client never drawing it.

On ChatGPT remote (iOS), thread/tokenUsage/updated fires and is delivered during connected turns (verified live here); the session rollout carries growing token_count events. The context-window meter still never appears because the mobile client gates it on resumeState and composerMode conditions that remote threads simply don't satisfy. That is independent of which model is in use — native gpt-5.6 threads show the same absence as custom-provider threads.

So for custom-provider users, the needed change isn't "include the data" (it already is) but "let the client render on remote" — same ask, sharper framing: the remote surface receives model identity, reasoning effort, token usage/tier, and reasoning/tool items, but renders only the message and summary-view items. Exposing the rest on the mobile path (as already happens on desktop) would cover this issue across all models.