Codex treats skill injection via $ command as a request to read the skill.md file.

What version of Codex CLI is running?

0.112.0

What subscription do you have?

API

Which model were you using?

GPT-5.4 xhigh

What platform is your computer?

Windows 11

What terminal emulator and version are you using (if applicable)?

Windows Terminal

What issue are you seeing?

Here is a transcript:

› To clarify, I've seen you do this check a few times now, when we use a skill and you re-read the skill file to see if it matches what I pasted in. However, I am not pasting anything. I am calling the skill via the $ command and that injects the skill into context. Is there some confusion here?

• Yes. From my side, there is ambiguity.

In this thread I saw two things:

1. your $implementation-review-lite message
2. a separate <skill>...</skill> message containing the skill name, path, and full body

I cannot tell whether that second block was manually pasted by you or automatically injected by the harness. It just arrives as conversation content.

The extra re-read was not me thinking “the user pasted this, I should distrust it”. It was me following the skill workflow conservatively, because the active developer instructions say that when a skill applies I should open its SKILL.md. In practice that means:

• if I only have the skill name/path, I should read the file
• if the full skill body is already injected into context, re-reading the file is usually redundant

So the short answer is: not confusion about your intent, but uncertainty about provenance plus a conservative reading of the instructions.

What steps can reproduce the bug?

Inject a skill via $ and inserting it. The model may prompt itself to re-read the file rather than understand the injected skill tag <skill> IS the skill content.

What is the expected behavior?

The model understands that injected <skill> blocks are in fact the skill content, and does not need to re-read the SKILL.md

Additional information

_No response_