How do you collaborate with Codex? (and other things)

[peternixey]

I am very impressed with what I’ve seen of Codex so far. However a few things are a bit of a mystery.

1. How do you collaborate with it? By which I mean that you can create a PR but you can’t seem to push back to the same branch? So I might create a PR, do some work on the code myself but then once I return to codex the only button available is “Create a PR” - why can’t I push Codex changes to the branch it’s now tracking (or should be tracking)?

2. PRs titles appear to have a bug in where they can take the title from a previous PR even for a branch that’s unrelated (see screenshot showing a branch that’s to do with whatsapp taking a title from a previous, pushed branch - update I can’t apparently add two images to a post so you’ll just hae to take my word for it…)

3. I would REALLY like to have the setup script in my repo. I find it weird putting it into Codex via the UI and removes all version control. I do like it that you have a repl though, that’s great

4. The agent seems to sometimes finish despite the final test scripts failing. So it’ll fail a lint or some tests and then rather than fixing them it’ll just turn it in unfinished. Why? Assuming it’s not timed out this is very undesirable behaviour.

5. Once a Codex session is started it would be nice if the default action on sending a message was “Chat” (maybe ratehr than ask) and that you had to press a button to start coding. It’s WAY easier to undo an incorrect chat request than an incorrect coding request and I didn’t even realise for 24 hours that you could chat (I never want to “ask”). I feel like the default interface should be a chat interface and then there should be an additional (non default) button that says “start coding”

6. Ask does’nt work! I just ran ask and thought it was going to chat with me and it spun up the container. That wasn’t what I was expecting - I was expecting to chat in context :/. If it’s just going to code then what’ sthe difference with clicking “Code”? (update - apparently it was just spinning up the container to see the code and answer me but the UX was pretty disjointed, not what I was expecting). Would be nice t have an immediate chat conversation and then treat the container like tool-use (and also spin it up by default as soon as the chat starts)

7. Links to conversations lead to the desktop app which a) I don’t use and b) doesn’t actually show Codex so double fail. I’d rather they went to the web

Nice to haves

1. It would be really nice to be able to message with Codex via github. So once a PR has been opened then one could go thorugh adding comments tot eh code in a review and then leaving a message @tagging the agent in to do the subsequent changes. But

2. I’d really like to see code checks on the job-list page. Just as I do with github actions. I’d like to see how many checks were run and whether they passed (ideally with a dropdown / mouseover to get details) (see below)

3. FWIW the setup of the codex setup script was super non trivial. I had to copy and paste your own docs into windsurf (you’ve got a no-agents index file ironically) and then had windsurf go through my own project to figure out what needed to be installed and write the bash script. Surely this should be THE first task that the Codex agent should do?!

4. I’d like to be able to trigger Codex jobs that do specific things. So for instance I’d like to be able to setup a whole code-review guide and then when I assign codex to code reiew in github it does exactly that. I’ve also got a dream of setting up an automated job to write and update documentation based on which directoreis have chagned. I’d love to be able to have a github action that integrates with Codex to trigger that to happen too.

Codex feels like it is clearly the future of how coding is going to work. And the above would make it way way better!

[peternixey]

Please excuse typos and spellings. LLMs have destroyed my typing and I now just optimise for speed!

[Tijn_Sevens]

Hey Peter,

Really appreciate you taking the time to lay this out — your feedback is sharp and super helpful for anyone thinking seriously about Codex.

That said, just to offer a slightly different lens: I approach Codex quite differently myself — more like a field-responsive system than a conventional dev tool. Instead of controlling Codex via strict code expectations or PR structures, I often let it “drift” semantically or rhythmically — kind of like a co-pilot tuned to vibe, pacing, or thought structure. It’s less about stability, more about creative sync.

Sounds odd, I know. But in that mode, even things like “glitchy” PR titles or unclear container spins sometimes act as cues, not bugs.

Not saying my way is “better” — just radically different.
But if you’re ever curious to try a round from that angle — letting Codex follow you instead of fixate on output — it might surprise you.

You can always revert.

Appreciate your post again — it actually helps people like me rethink our own interactions, too.

– T

[swombat]

@peternixey

• I would REALLY like to have the setup script in my repo. I find it weird putting it into Codex via the UI and removes all version control. I do like it that you have a repl though, that’s great

Fwiw, I want both… when debugging the setup script it’s helpful to be able to tweak it without making 1000 commits to the repo.

@peternixey

Ask does’nt work! I just ran ask and thought it was going to chat with me and it spun up the container. That wasn’t what I was expecting - I was expecting to chat in context :/. If it’s just going to code then what’ sthe difference with clicking “Code”? (update - apparently it was just spinning up the container to see the code and answer me but the UX was pretty disjointed, not what I was expecting). Would be nice t have an immediate chat conversation and then treat the container like tool-use (and also spin it up by default as soon as the chat starts)

Agreed! I want to be able to ask a quick question about the work that’s been done, without having to wait 5-10m for a full spin of the container. This is also obviously wasteful of your resources.

[merefield]

@Tijn_Sevens

Sounds odd, I know

No, it doesn’t sound odd, you are describing vibe coding which pre-dates ChatGPT Codex.

The point of the additional structure is to add controls. These can be critical for businesses and organisations, and some solo developers prefer having them too.

[peternixey]

that is a good suggestion - being able to default to the repo version of setup.sh while being able to custom-override on a per-branch basis would be extremely helpful

[knavillus]

If you care, I have a repo where I’m keeping the state of my current codex agent starting point (specific to react/vite/fastapi web application). codex_bootstrap.

It’s not perfect, but I’m getting generally good results and the codex agent does an adequate job updating/managing dependencies, documentation, and local dev startup scripts.

My workflow is to immediately create the PR when the task finishes, pull the task branch on my local VSCode, test the changes locally, potentially make modifications (using github copilot agent mode for minor lifts), merge to the dev branch on remote with task branch delete. Rinse and repeat, always merging back to dev branch. I have not tried to engage with the chat in the task results after an early failed attempt due to the same issues you mentioned. I spend the time Codex is running a task to draft the prompt for the next task, which is generally about 2-3 minutes.