Feature idea: User-defined context axes for long-running Codex tasks

[yousaki11]

I would like to suggest a feature idea for Codex and long-context AI workflows.
When working on long-running tasks, users often include different types of information in the same conversation, such as goals, constraints, assumptions, decisions, examples, temporary notes, and output preferences.
In these cases, the model can sometimes mix these information types unintentionally. A temporary assumption may be treated as a final decision, an example may be treated as a requirement, or an output preference may influence the substance of the answer.
My suggestion is to let users define context axes for a task.
The idea is:
Split user input into smaller semantic units.
Assign each unit to a user-defined context axis, such as goal, constraint, assumption, decision, example, note, or output format.
Let Codex keep these axes separate when needed.
Let the user choose when to merge selected axes into a unified answer.
Before final output, check whether any axis was mixed, omitted, or incorrectly merged.
This could work as a reasoning and output-control layer on top of Codex, especially for complex long-running tasks with many changing requirements.
Possible benefits:
Fewer context-mixing errors
Better separation between assumptions, decisions, examples, and requirements
More reliable output formatting
Easier review of which part of the answer came from which type of information
Better support for complex long-running projects
I am not presenting this as a finished technical solution, but as a product/research direction based on repeated user experience with long-context AI workflows.