I’d like to propose an improvement specifically for the official OpenAI Codex extension for Visual Studio Code (Codex – OpenAI’s coding agent).
Right now, when this extension performs code edits, it applies changes directly to the file. That works fine for small adjustments, but it quickly becomes hard to see exactly what was modified in medium or larger edits.
This lowers visibility and control over the changes the AI makes, especially compared to other tools like GitHub Copilot, which shows a diff and lets you accept or reject edits before they’re applied.
Feature request
Please consider adding a workflow where:
The extension shows a diff preview before applying edits
Users can accept or reject changes per hunk or block
It integrates with VS Code’s native diff/patch UI
Even a basic “Preview changes before applying” option would already be a big usability improvement.
This would enhance transparency, trust, and control — especially in professional workflows where reviewing code changes is critical.
Hi @isaacT, this is a really thoughtful suggestion, and honestly, a fair one.
Right now the Codex extension applies edits directly, which is fast, but as you pointed out, it can get hard to track what changed once the edits get larger. Especially in professional workflows, being able to review a diff before changes land is pretty important.
There isn’t currently a built-in “preview before apply” workflow in the extension, but the idea of diff preview + selective acceptance makes a lot of sense. I’ll make sure this feedback is captured, particularly the integration with VS Code’s native diff UI, which would feel very natural.
Appreciate you laying it out clearly. This is the kind of UX improvement that can make a big difference in day-to-day usage.
Helo, I just switched over from another IDE/extension where this was a native feature. Without it, I can’t clearly see what Codex has changed, and I need a way to track those edits. My codebase is 50,000 lines long; when it modifies 20 different locations, I can’t verify if the changes are correct or even see the logic behind the edits (e.g., if it started at the right place). Currently, I can only see the output in the chat window. I’m using Codex in Windsurf with a Plus subscription and considering upgrading to Pro, but I don’t feel safe using it without this feature – even though I’ve noticed its work is very precise.
Is there any movement here? This is the only thing that prevents me from entirely switching from Cursor to Codex. Are there workarounds? Hunk-level Accept/Reject is a must-have and VSCode seem to already support it.
+1, all agentic coding providers have this (AntiGravity, Copilot, Claude Code, Cursor…), besides Codex…
I am also missing that when I enter the plan mode, it auto-creates the bullets, and I can clearly see which bullet it is currently implementing in the agent mode.
+1. Trying to switch to codex as well, and this is an absolute must. Copilot showed the differences intuitively, and I was able to reject anything I didn’t like. Has there been any progress on this?
same issue with me I use windsurf along with codex but I need more control with codex due to this I only use codex in limited capacity to review the code changes only.
The entire Codex has this mentality of exactly what OpenAI says.
Right now the Codex extension applies edits directly, which is fast,
Guys, that’s NOT ok. The goal of an AI is to work WITH the human, the human reviews and guides the code in real time. An AI that makes rapid changes with out human interaction is a danger, just look at any time AIs have torpedo database, or just delete changes Hell I’ve already seen it delete whole sections of code and render functionality broken. Code Review tools MUST be important and they MUST be done BEFORE changes are made. It’s been 3 months, but what I’m observing from using it makes me think the goal is “Speed over interactivity”.
That’s great for vibe coders, but no serious programmer is willing to trust any AI that much , at least not at first.
Three things that need to change if Codex is to be taken seriously.
A. Request based READs, not writes, but even reading files. Opening a directory doesn’t mean “Read anything you want in it” unless the user choose that.
B. Default to permission based systems rather than giving AIs carte blanche to do what they want. Also make those options able to be set up.
C. Allow users to properly code review changes BEFORE they are implemented, not after they are implimented. This is so obvious, I can’t believe anyone thought this tiny diff is acceptable.
Maybe OpenAI doesn’t want a real programmer using this, maybe you only want people who blindly trust the AI and allow it to do anything, but I’ve already seen enough destroyed projects because an AI had that much control. All you’re doing is creating a bad product that will get a bad reputation not because of user action but because of user inaction, and hidden features that are standardized by the industry already.
