Code review skill produces wildly different results between Claude and Codex

[alessandrotesoro]

Hi there,

First of all, thank you for creating such a great plugin. I've been using it for some time now and it's been great.

I've primarily been a Claude user, but lately I've been testing Codex and I've noticed a big difference in how the /ce-code-review skill behaves between the two.

When used against the same task/commit diff, Codex tends to return single-digit findings, while Claude usually produces double-digit findings. The difference between the two is pretty significant.

When I asked Codex why there was such a difference between the results, it said:

I did not spawn the full applicable reviewer set for the final diff.
I did not create/read a fresh run artifact and then synthesize from every reviewer output.
I over-applied the “validate and drop weak findings” mindset before showing you the raw-to-final mapping.
Claude Code has a more direct implementation path for this CE skill; in Codex I had to manually emulate it with spawn_agent, and I under-emulated it.

This isn't a one-off, it consistently happens every time I run the skill. Is this a bug or a limitation of Codex? I'm also not sure if it's related to this #799 ?

---

Edit: here's some more comments from Codex:

Why results diverged:

Claude artifacts include raw reviewer output, not just final merged findings. Many items are testing_gaps or residual_risks, not primary findings. My final report only surfaced the main actionable findings and summarized coverage too tightly.

A lot of Claude’s raw findings violate the skill’s strict merge schema: severity values like high/medium/low, confidence values like 90/95/80. Under the skill’s Stage 5 validation, those are supposed to be dropped or normalized before final synthesis. I dropped some instead of salvaging/normalizing useful content.

I demoted or omitted several “coverage-only” items that Claude preserved, ....

Reviewer nondeterminism is real here. Different harness, model tiering, and agent implementations produced different issue discovery. But that does not fully excuse it: several Claude findings are valid and I should have caught or at least surfaced them as testing gaps/residual risks.

I also compressed the final report too much after applying the one safe fix. The skill asks for richer Coverage, Deployment Notes, and residual/actionable tables. My final answer was materially thinner than the Claude artifact set.

So the honest answer: I did spawn the required agents, but my synthesis pass was stricter and less complete than it should have been. The biggest actual miss is not agent coverage; it is that I failed to reconcile valid raw signals into the final report.

[tmchow]

I dno't think it's related to #799 as that is due to the OpenAI bug: openai/codex#16996

I haven't had the issue you're having with codex unfortunately so hard to say. You should ask it something more specific like look at the SKILL definition And to specifically point out where in the skill caused its behavior to go wrong, what it would recommend changing it to, and then posting something here.

[ace139]

I also have seen this issue over a couple of weeks now, for e.g. when running the /ce-code-review with Opus 4.7[1M], let's say it comes with 40 issues that needs to be fixed and when same is run in Codex gpt-5.5-high the issue count is 10-15. One more thing which I have noticed is that the number of sub-agents that Codex fire vs Claude is also different, 3-8 vs 14-16.

But even if you instruct and run all the sub-agents in Codex, it's still not able to find all the issues.

[alessandrotesoro]

Yes I've had the same, sometimes it'll fail to run some sub-agents and it'll say it queued them up and some times they actually execute (many times they just don't) but even when it says it ran them all, it still doesn't produce the same amount of findings.

I'm going to have to stop using Codex for now with this plugin 🥲