Support .codex/skills as a Codex-specific skills discovery directory
What variant of Codex are you using?
App, IDE Extesion, CLI
What feature would you like to see?
Codex currently discovers shared Agent Skills from .agents/skills, but there is no Codex-private skill directory.
Please support .codex/skills as a Codex-specific discovery root, alongside .agents/skills.
Motivation:
.agents/skillsis a shared, cross-agent convention..codex/skillswould allow Codex-specific skill variants.- Some skills should prefer Codex-native repo tools such as
rg,git grep,find, package scripts, tests, and compiler checks. - Generic skills may include fallback Python or PowerShell scripts that are less suitable for Codex.
- The LLM should not be expected to infer whether it is running inside Codex from visible tools.
Suggested behavior:
- Discover repo-local
.codex/skills - Discover user-level
~/.codex/skills - Keep existing
.agents/skillsbehavior unchanged - If
.codex/skills/fooand.agents/skills/fooboth exist, prefer the Codex-specific one - Show the skill source in
/skills
Why this matters
Without a Codex-private directory, users have three bad options:
Option 1: Put Codex-specific instructions into .agents/skills
This pollutes the shared skill with Codex-specific assumptions.
Bad example:
If you are Codex, use rg. If you are another agent, use scripts/analyze.py.
This assumes the LLM knows which agent invoked it. That is not reliable.
Option 2: Put multiple variants into .agents/skills
Example:
.agents/skills/dependency-analyzer-codex
.agents/skills/dependency-analyzer-generic
.agents/skills/dependency-analyzer-windows
This forces the model to choose between near-duplicate skills. That makes discovery less stable.
Option 3: Avoid Codex-specific optimization entirely
This makes Codex less effective, because generic skills may overuse bundled scripts instead of Codex-native repository inspection.
Additional information
_No response_
2 Comments
We're unlikely to add this. We previously used this path but shifted away from it when the industry rallied around standardization.
Anyway I'll tell you what I'm thinking.
I don't think Agents will be standardized. You may have some common sense
but the choices will be completely different from each other. Less harness
or heavy harness? Who controls the memory: the Agent, the LLM or the
user's responsibility? What is safe, opening the shells or limiting the
shells? Safeguard the tool call, or safeguard each file/folder/url?
The interface will be the same, but the content will differ for each
agent. Users will always want to add different skills for different agents.
Each agent is like an individual guide; you may only standardize the
textbook, never the guide.
Unless you think there will never be any agents. Every internal tool and
every internal skill can be acquired on the open market. Cursor or Codex
only sells its own skill on the market.
On Thu, May 14, 2026 at 11:52 AM Eric Traut @.***>
wrote: