Skip to content

Feature request: Include model and reasoning-level recommendations in ChatGPT-to-Codex handoffs #36251

Description

@scruffyOrc

What variant of Codex are you using?

App

What feature would you like to see?

When ChatGPT hands a task to Codex, it should be able to include a recommended model and reasoning level based on the context of the conversation.

ChatGPT often already understands whether a task is routine implementation, repository exploration, cross-system work, difficult debugging, architectural design, or consequential review. That context should be passed into the Codex handoff instead of requiring the user to make the model-selection decision again manually.

The handoff could show:

Recommended model: Luna
Recommended reasoning: High
Reason: The behavior is already defined and the remaining work is primarily implementation and testing.

The user should be able to accept the recommendation, override it, or disable recommendations entirely.

This would improve cost efficiency, reliability, and transparency while helping users take advantage of lower-cost models when they are appropriate.

Additional information

I regularly use ChatGPT to clarify product behavior and technical intent before handing repository-level work to Codex.

The appropriate model often depends on where the uncertainty lies:

  • Luna when the task is implementing already-defined behavior
  • Terra when the task requires understanding the existing system and making moderate implementation decisions
  • Sol when the task involves architecture, unresolved product decisions, difficult diagnosis, or high-consequence review

Today, ChatGPT may have enough context to make that distinction, but the recommendation is lost during the handoff. The user must remember to reconsider the model and reasoning level before starting the task.

A later extension could allow Codex to recommend escalation when a lower-cost model discovers unexpected architectural uncertainty, repeated failures, or risks beyond the original task classification. The initial feature, however, could simply pass an optional model and reasoning recommendation with the task.

Related issue

This is related to #35062, but concerns a different workflow stage. #35062 proposes secure model dispatch originating within Codex. This request proposes carrying ChatGPT’s existing task context into the initial Codex handoff as an optional model and reasoning-level recommendation. The capabilities could complement each other.

Metadata

Metadata

Assignees

No one assigned

    Labels

    appIssues related to the Codex desktop appenhancementNew feature or request

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions