Ability to Save a Session State and Branch New Sessions From It

What feature would you like to see?

Summary

I am requesting a feature that enables saving a session’s internal state at a specific point and then creating new branches (new sessions) from that saved state. This would allow users to avoid repeating lengthy initialization or context-building prompt sequences across multiple sessions, significantly reducing token usage and improving workflow efficiency.

Problem Statement

For certain tasks, especially multi-step or complex workflows, I often go through a series of prompts to “prepare” the model:

• Loading context
• Establishing the working conventions
• Setting architectural or project background
• Reaching a stable, useful conversational state

These steps may require hundreds or thousands of tokens and must be repeated whenever I want to start a new, separate conversation following the same assumptions or context. This leads to:

• Excessive token consumption
• Repetitive prompting
• Increased time to reach a usable working state
• Difficulty maintaining consistency across parallel sessions

Proposed Solution

Introduce the capability to save a session snapshot (the model’s internal conversation state at a given point in time) and then create new sessions that start from that snapshot.

This could take the form of features such as:

• “Save State”: Serialize an internal representation of the current session context.
• “Branch Session”: Start a new session that inherits the prior state while allowing its own independent evolution.
• Optional: “List saved states” / “Delete saved state” management interface.
• Optional: Programmatic API support so developers can automate branching and state management.

Benefits

1. Reduced token usage: Eliminates the need to resend large initialization prompts.
2. Improved consistency: Ensures multiple sessions can start from an identical baseline.
3. Faster workflows: Users can immediately begin working at the intended context depth.
4. Better multi-branch experimentation: Enables parallel explorations from a shared root state without redoing setup each time.

Example Workflow

1. I start a session and guide the model through 10–20 prompts establishing architectural rules, coding conventions, domain details, etc.
2. I save the session state as “project-setup-v1.”
3. Later, I open a new conversation and choose “Start from saved state → project-setup-v1.”
4. I now have a clean, fresh session with all the prior context loaded, allowing me to explore different directions without repeating setup prompts.

Additional Notes

• This feature aligns with common version-control mental models (e.g., branching from a baseline).
• It would be especially valuable for developers using Codex as part of iterative or modular project workflows.
• The capability does not need to reveal any internal model representation; it simply needs to restore the conversational context.

Request

Please consider adding this feature to improve usability, efficiency, and productivity for advanced users who rely on reproducible conversational baselines and multi-branch workflows.

Additional information

_No response_