Using the skills in Codex

The skills follow the Agent Skills spec, so Codex reads them as well as Claude Code. Codex finds skills in .agents/skills/<name>/SKILL.md: ~/.agents/skills for your user, and from the working directory up to the repo root for a project.

rota skills install                                # user: ~/.agents/skills (and ~/.claude/skills)
rota skills install --scope project --agent codex  # this repo: .agents/skills

rota carries the skills inside the binary. install copies each rota-* skill, plus the references it cites, into the root and writes a .rota-manifest.json that lists what it wrote. These are copies, not symlinks, so they work without a checkout and can be committed. A file you edited, or one rota did not write, is kept and reported; --overwrite replaces it.

To keep installed skills from drifting behind the binary, rota skills install writes each SKILL.md as a short stub (the default) that tells the agent to run rota skills get <name>. That prints the text embedded in the installed binary, a first line Base directory for this skill: <dir>, and the absolute paths of the files the text links to (--references lists a skill’s reference files, --reference <file> prints one). The files live in a per-version tree in your user cache, so relative links resolve. Codex runs the command when it loads a skill. For a harness that cannot, rota skills install --full writes the full text and the references into the root. rota skills list names the embedded skills and the binary version.

A stub root follows the binary. Run rota skills update only when rota skills status (or rota doctor) reports it stale, which for a stub means a skill was added or removed or its stub changed (frontmatter or stub text); the report names the skill. A full root needs rota skills update after every upgrade. update keeps each root’s recorded mode unless you pass --stub or --full.

--references lists both the skill’s linked local files and its shared references, with paths relative to the skill directory. Use the listed path to fetch a file:

rota skills get rota-review --references
rota skills get rota-review --reference reviewer-briefs.md
rota skills get rota-review --reference references/test-quality.md

Bare shared names such as test-quality.md still work. When a linked local file has the same name, the bare name selects the local file; references/<name> selects the shared one. An unknown name exits with a list of valid paths.

To run Codex as a worker in a parallel round, see Codex workers.

In Codex, type $rota-pause where Claude Code uses /rota-pause. It is the same skill.

The skills use neutral instructions for file reads, edits, questions and subagent dispatch. If a question cannot use the current Codex picker, answer the numbered options in prose; approval gates still require an explicit answer. To invoke another skill, Codex reads its instructions and follows them with the supplied arguments. Subagents use round.tiers.codex.* when configured, otherwise the harness default. rota must be on PATH (see install).

Remaining Claude-only behaviour — 2026-10-08

rota-work: the shared wave-planning guidance names Claude Code’s “File has been modified since read” error and retains its retry with a byte-identical old_string. That tool-specific error and retry are labelled Claude-only.
rota-orchestrate: solo mode still launches Claude subagents through its Agent interface. Codex must use a supported terminal host for rounds; see Codex workers. This restriction applies to standing round workers, not ordinary subagents inside a skill.

Shared authoring guidance retains a Claude-only question-picker adapter and Claude task-title escaping advice. Neither is required for Codex skill execution.

The audit is enforced by a Go test running the tool-name grep, with exceptions for neutral prose and explicitly Claude-only instructions. Smoke section 144 installs the embedded skills and runs preview and capture through the scripted Codex worker in test/fakes: it checks output structure, read-only preview and capture’s created IDs. It does not call a model or prove live interpretation of every skill. The audit itself was implemented in a Codex worker with a default-model subagent; live coverage beyond that remains unverified.

Checking discovery

With CODEX_HOME unset, run codex debug prompt-input hi in a scratch repo. It prints the model-visible input, skills included, without starting a model session. Each rota-* skill should appear as rota-x: <description>. It reads your own ~/.codex and writes nothing to the project. rota doctor checks that your Codex version is in the supported range, and that installed skills match the binary.

Skill directories

rota skills install (above) is the route rota keeps up to date. npx skills add l4ci/rota also works.

skills.sh lists a repo once someone installs it with the skills CLI. It has no submission form. npx skills add l4ci/rota is the common install route and puts the skills in .agents/skills, where Codex reads them. It sends anonymous install telemetry, and writes no .rota-manifest.json, so rota skills status and update do not track the copies. rota is listed at skills.sh/l4ci/rota (checked 2026-10-08, after one clean-environment run of that command).
openai/skills is marked deprecated in its README, which points to openai/plugins and the Build plugins guide. Its skills/ tree has .curated and .system and no community tier. Checked on 2026-10-08.