Problem
I frequently use ChatGPT to discuss software design, architecture, debugging ideas, and implementation plans before asking Codex to actually modify my code.
Currently, this creates an awkward workflow:
- Discuss the problem in normal ChatGPT.
- Reach a concrete implementation plan.
- Open Codex in VS Code or the Codex app.
- Re-explain or copy the relevant context from the ChatGPT conversation.
- Ask Codex to inspect the repository and implement the plan.
Alternatively, I can have the entire discussion inside Codex, but then I am using Codex/agentic usage for conversations that do not actually require an agent to edit files or run commands.
Proposed Feature
Allow normal ChatGPT to have optional read-only awareness of an active VS Code workspace or selected local repository.
Then add a “Send to Codex” / “Implement with Codex” action.
The workflow could look like:
VS Code repository
→ ChatGPT reads repository context
→ User discusses architecture/design/debugging with normal ChatGPT
→ ChatGPT develops an implementation plan
→ User clicks “Send to Codex”
→ Codex receives the relevant conversation, plan, and repository context
→ Codex implements and tests the changes
Example
I could ask normal ChatGPT:
“Look at how error handling currently works in this project. I’m thinking about moving this responsibility from the interpreter to the controller. Does that architecture make sense?”
ChatGPT could inspect the relevant files in read-only mode and help me reason about the change.
Once we reach a decision, I could say:
“Implement this with Codex.”
Codex would then receive the relevant context and begin modifying the repository without me needing to reconstruct the prompt.
Why This Would Be Useful
This would create a clearer separation between:
- ChatGPT: reasoning, brainstorming, architecture, explanation, planning
- Codex: implementation, file modification, terminal commands, testing, debugging
It would also make Codex usage more efficient because users would not need to consume agentic usage simply to have extended conversations about their code.
Most importantly, it would eliminate the repeated context transfer between ChatGPT and Codex.
Optional Extensions
This could eventually support:
- “Send this plan to Codex”
- Selecting which parts of a ChatGPT conversation Codex receives
- Read-only VS Code workspace context for normal ChatGPT
- Returning Codex results back into the original ChatGPT discussion
- Multiple Codex tasks created from one higher-level ChatGPT conversation
- A visible summary of exactly what context will be handed to Codex before execution
Essentially, ChatGPT could serve as the reasoning/planning layer, while Codex becomes the execution layer, with a seamless handoff between the two.