Replies to Codex's PR code review comments triggers new Codex tasks

The @codex at-mention is intended to be a signal that you want Codex to do something. If you don't want to invoke Codex, don't at-mention it.

Thanks for the context.

Based on the trigger text, there isn’t an actionable request for me to implement in code right now—it's a clarification about invocation behavior (“don’t @codex unless you want Codex to act”). So I’m not making repository changes, and I’m not opening a follow-up PR.

I also attempted the required web lookup step before responding, but outbound web access appears blocked in this environment (curl returns CONNECT tunnel failed, response 403).

View task →

@etraut-openai I think that this doesn't make sense given that people frequently tag others with an expectation that they see something / referencing them - not necessarily to get them to do something. Codex's response here to my @codex was great - my only suggestion essentially was that in such cases we don't need task bloat if it didn't change anything. Your response saying people just shouldn't tag it is kinda like "you're holding it wrong" from Apple IMHO...

I'm not sure what you're suggesting we do differently. When someone at-mentions codex in a github PR, it triggers a web hook that spins up a cloud-based codex environment, clones the repo, invokes the codex harness, and passes the agent the text after the at-mention. Are you saying that you want the at-mention to be silently ignored in some cases? That will require a classifier that inevitably results in false positives where codex simply doesn't respond to legitimate requests. That would be a much worse experience than what you're describing. We'd have more flexibility here if this was a UI surface that we controlled, but within github's UI, our options are rather limited.

I am suggesting not showing tasks without any diff (by default) in Codex Web. If you want to keep those tasks which did nothing around with a link to be able to debug them later more easily, then sure, but the idea is to avoid bloating the users context. I also wonder if the LLM should be able to choose if it thinks that it needs an environment to work, e.g. a thank you comment / a comment that is referencing them in passing might not need a code environment, but that isn't critical

This would break an important use case. It is not uncommon for users to respond to a codex-generated review comment with another question directed at codex — to challenge its analysis or ask a clarifying question to test its (or your) understanding of the reported problem.

Firstly: I don't want to waste any more of your time. I can "hold it differently". It isn't so hard to not do a @codex...

As a final thought on this, I will give more context:
I was using the GitHub PR comment resolve skill yesterday, and since I prefer to thank reviewers for their comments in the fixes, I instructed GPT-5.3-Codex to do so. It added the comments to GitHub (after asking for my approval to run those commands, but due to another bug in Codex, I literally could not see the whole command that the model was planning to run before approval) and in multiple of those comments it replied to the Codex code reviewer with @codex and thus triggered unnecessary Codex tasks... now of course I can instruct it not to respond to AI reviewers or Codex or something, but simply as a point of comparison, no other AI reviewer (we use 4 including Cursor's bugbot) kicked off a task run in response to the message. I am not telling you how to fix my issue or even to do so, but hopefully with the relevant context, the issue itself is clearer