ARS-Codex: the Academic Research Skills suite ported to OpenAI Codex

What ARS-Codex actually packages

ARS-Codex is a Codex distribution of the Academic Research Skills (ARS) workflow. The README is explicit that it is "the Codex-native sibling" of the Claude Code ARS repository, with its own plugin identity, packaging, and runtime adapter. The release number follows the vendored ARS suite, which is why the current package version reads 3.22.0 while earlier releases used 0.1.x numbers.

The repository vendors the ARS workflow content as a single Codex skill under `skills/academic-research-suite/`. That directory holds `SKILL.md`, `manifest.json`, an `agents/openai.yaml` file, a `codex/` subtree with runtime manifests, agents, hooks and scripts, and an `ars/` subtree containing the actual workflow modules: `deep-research`, `academic-paper`, `academic-paper-reviewer`, `academic-pipeline`, `experiment-agent`, plus commands, hooks, docs, tests and shared code.

The intended audience is narrow. This is for someone running OpenAI Codex who wants the ARS research pipeline as one installed skill, and who is willing to treat the upstream Claude Code checkout as read-only reference material. The README states plainly that the original checkout is not modified and that upstream content is copied from fresh GitHub clones and adapted through the Codex router in `skills/academic-research-suite/SKILL.md`.

The vendoring model and the Codex router

The architecture is copy-and-adapt rather than fork-and-diverge. Upstream ARS content is cloned from GitHub, dropped into the `ars/` subtree, and reached through a router defined in the skill's `SKILL.md`. That router is the seam between the Claude-oriented workflow definitions and the Codex runtime.

A second layer sits in `codex/`: `full-runtime-manifest.json`, an `agents/` directory, `hooks/`, and `scripts/`. The README describes native adaptive execution as the default, where bounded subagents handle independent work and the lead continues with other work. The fixed full-runtime topology and its hooks are opt-in, not the default path. That distinction matters when you are deciding how much of the runtime machinery you want loaded.

The vendored source is pinned. The README records that this release tracks ARS v3.22.0 at commit `3c546bc08c56f79e0068f1ea4f0acedf5bf69b5e`, and says the exact upstream version, tag and commit live in `manifest.source_repositories[]`. If you need to know which upstream state you are running, that manifest field is the place to look, not the README prose.

Some upstream artifacts are preserved but inert. Nested `.github/` workflows and root `agents/` mirrors are kept for traceability and self-tests, and the README states they are not repo-level CI or Codex entrypoints. Claude and plugin loader files under `.claude/` and `.claude-plugin/` are intentionally excluded.

Installing ARS-Codex as a Codex plugin

The README gives two install paths. The plugin route adds the GitHub marketplace and installs the plugin through Codex CLI:

codex plugin marketplace add Imbad0202/academic-research-skills-codex --ref main
codex plugin add ars-codex@ars-codex

After that, open a new Codex conversation and invoke `$academic-research-suite`, or describe an academic research task that matches the bundled workflow. The README warns that existing Codex sessions may keep their old skill cache, so a fresh conversation is part of the procedure rather than an optional nicety.

Updating later uses the marketplace upgrade command followed by the same add command:

codex plugin marketplace upgrade ars-codex
codex plugin add ars-codex@ars-codex

Codex Desktop users can instead add the repository from the Plugins interface and install ARS-Codex, with marketplace source `https://github.com/Imbad0202/academic-research-skills-codex.git`, branch `main`, and plugin `ars-codex`. The README notes that the plugin root is `plugins/ars-codex/` and that its `skills/` directory holds a materialized copy rather than a symlink, specifically so Windows installs do not lose the bundled skill when caches turn symlinks into plain text files.

Direct skill install, and what /skills should show

If you would rather skip the plugin, the README documents installing the skill straight from the repo path with the skill installer script, using `--method git` so public and credentialed GitHub access behave the same:

python3 "$HOME/.codex/skills/.system/skill-installer/scripts/install-skill-from-github.py" \
  --repo Imbad0202/academic-research-skills-codex \
  --ref main \
  --path skills/academic-research-suite \
  --method git

The README notes that macOS and many Linux systems expose Python 3 as `python3` rather than `python`, and that if your system only has a `python` command and it is Python 3, you should use that instead. Updating means removing the installed directory and re-running the installer:

rm -rf "$HOME/.codex/skills/academic-research-suite"
python3 "$HOME/.codex/skills/.system/skill-installer/scripts/install-skill-from-github.py" \
  --repo Imbad0202/academic-research-skills-codex \
  --ref main \
  --path skills/academic-research-suite \
  --method git

The verification step is specific. Run `/skills` and you should see one ARS-Codex entry, `academic-research-suite` or `ARS-Codex`. You should not see separate `academic-paper`, `academic-pipeline`, `deep-research`, or `academic-paper-reviewer` skills from this package. If you do, the README says to reinstall with the update command and open a new Codex conversation. That is a useful failure signal: extra entries mean the single-suite packaging did not take effect.

Model settings, reasoning levels and the citation transport

The repository's `.codex/config.toml` selects `gpt-6-astra` with `xhigh` reasoning for new trusted Codex sessions. Installing the skill does not copy that project setting and does not change an already-running session, so the model policy in the repo is a reference configuration rather than something the install imposes on you.

The README describes the planner as suggesting `medium` for routine work and `xhigh` for complex work, and states directly that these are starting policies, not measured optimal settings. It also notes that `ultra` is optional in Codex for demanding work chosen explicitly, and that the contained citation transport rejects it. That last point is the most concrete constraint in the section: if your workflow depends on the citation transport, a reasoning level the transport will not accept is a hard stop, not a preference.

Two further documents are named for anyone who wants the details: the model runtime policy at `skills/academic-research-suite/codex/model-runtime-policy.md` and a system-card alignment audit under `skills/academic-research-suite/codex/audits/`. The README is careful about evidence here, stating that Claude plugin eval suites are reference material and not evidence of measured Codex performance. That is an honest boundary, and it means you should not read the vendored eval fixtures as a Codex benchmark.

Language output, Spanish triggers, and the Windows file-lock repairs

Release v3.22.0 adds an output-language-pair contract, Spanish intent routing, reviewer-calibration and plugin evaluation fixtures, and Windows file-lock repairs. The README then narrows the claim: the output-language-pair registry currently supports only `zh-tw-en`, and Spanish triggers do not install a Spanish output-locale pack.

That is a real limitation rather than a footnote. A Spanish trigger can route an intent without giving you a Spanish output locale, so a workflow that assumes Spanish output will not get it from the registry. If your research writing is in Spanish, the routing layer and the output layer are not the same thing here, and the README says so.

The Windows file-lock repairs connect to the packaging decision described earlier: the plugin's `skills/` directory is a materialized copy because Windows plugin caches may turn symlinks into plain text and skip bundled skill registration. Both the copy and the repair are aimed at the same platform problem.

When ARS-Codex is the wrong choice

The README draws the boundary itself. Use the Claude Code repository when you want the native Claude Code skill layout, Claude-specific agent-team behavior, or the original ARS development history. Use this repository when you want the Codex-native single-suite skill. If your team's workflow is built around Claude Code agent teams, ARS-Codex is a port of the content, not a substitute for that runtime behavior.

The single-suite packaging is also a constraint. Because the whole ARS workflow ships as one skill, you do not get separately installable `deep-research` or `academic-paper` entries. The README treats their appearance as a sign of a failed install. Anyone who wants to enable only the literature-review portion, or to version the modules independently, is working against the package's design.

Licensing is the other boundary. The README badge identifies the licence as CC BY-NC 4.0, while the repository metadata reports a NOASSERTION licence identifier. Those two signals do not agree, and the README does not reconcile them. If your use is commercial, that discrepancy is something to resolve with the project before you build on it, not something to assume away.

Finally, the repository is a vendored adapter. Upstream ARS changes arrive here through a sync process, and the version numbers move together. If you need to track upstream ARS commits directly, or to contribute fixes in the upstream development history, this is the wrong repository.

Upgrade cost and licence implications

Upgrades are not in place. The plugin path runs `codex plugin marketplace upgrade ars-codex` and then `codex plugin add ars-codex@ars-codex`. The direct skill path deletes `$HOME/.codex/skills/academic-research-suite` and re-runs the installer. Neither path is a patch operation, so plan for a reinstall whenever the vendored ARS version moves.

Because the release number follows the vendored suite, a version bump is a signal that upstream content changed, not just that the adapter was touched. The README says package-level changes are summarized in `CHANGELOG.md`, and that the exact upstream version, tag and commit are recorded in `manifest.source_repositories[]`. Those two files are the practical way to tell what an upgrade will bring before you run it.

On licensing, the README badge states CC BY-NC 4.0 and links to the Creative Commons deed, while the repository metadata reports NOASSERTION. The README does not explain the difference. CC BY-NC 4.0 is a non-commercial licence, so the practical question for anyone in a company or a funded project is whether their use counts as commercial, and that is a question for the project or for counsel rather than something this article can settle.

Editorial conclusion

Adopt ARS-Codex if your research work already happens in Codex and you want the ARS pipeline as a single skill rather than a set of separate ones. Do not adopt it if you need the native Claude Code agent-team layout, a stable API surface, or a permissive licence for commercial use, since the README points to CC BY-NC 4.0. Before committing, verify three things: that `/skills` lists exactly one ARS-Codex entry after install, that your Codex session is new rather than resumed, and that the vendored ARS tag recorded in `manifest.source_repositories[]` is the one you intend to work against.

Frequently asked questions

Official sources

1. Imbad0202/academic-research-skills-codex on GitHub
2. Issues
3. Project website
4. README
5. Releases