Use with Codex

Why it works

Codex and Claude Code implement the same open agent-skills standard: a skill is a directory holding a SKILL.md whose frontmatter carries a name and a description. Nothing in that format is Claude-specific, so all 35 Researcher skills port straight across.

Codex looks for skills in three places, in priority order:

1. $CWD/.agents/skills
2. $REPO_ROOT/.agents/skills
3. $HOME/.agents/skills

Install

Clone the repository, then run the installer:

~/.agents/skills

git clone https://github.com/sokolmarek/researcher.git
cd researcher

python scripts/install-codex-skills.py

# Repo scope: ./.agents/skills, for a single project
python scripts/install-codex-skills.py --repo .

Two more flags:

python scripts/install-codex-skills.py --list       # preview what would be installed
python scripts/install-codex-skills.py --uninstall  # remove the installed skills

Invoking a skill

Skills install under a researcher- prefix, so literature-search becomes researcher-literature-search. In Codex you can call one explicitly:

$researcher-literature-search self-supervised learning for ECG classification

Or just describe what you want and let Codex match the skill from its description, the same way it works in Claude Code. Run /skills to list everything Codex can see.

What carries over

Carries over to Codex	Claude-only
All 35 skills	The 9 subagents. Codex has none, so the two skills that fork into the Sonnet code-agent under Claude simply run in the main session, and the installed copies say so.
Every Python script (bib-validator, latex-compile, install-git-hooks). They are agent-agnostic and run standalone, with no agent at all.	The in-session tool guards in hooks/hooks.json. Codex has no hook system.
The real git pre-commit citation guard	The namespaced slash commands. The installer rewrites each one to the skill it routed to, so /researcher:draft-section becomes $researcher-paper-drafting.
LaTeX compile checks, with any TeX engine
The figure style presets

Integrity in Codex

With no hook system, the backstop is the real git hook. Install it once per manuscript repository:

python scripts/install-git-hooks.py

That installs a git pre-commit that blocks commits containing dangling \cite keys, regardless of which agent made the commit, or whether an agent was involved at all.

The installer also writes an AGENTS.md template into the shared asset directory, carrying the integrity rules (never fabricate a citation, never invent data, no em dashes in generated text, compile-check LaTeX before delivery). Copy those rules into your project’s own AGENTS.md so Codex reads them every session.

Paths and refreshing

Skills ship with plugin-relative paths (references/..., templates/..., scripts/..., and the CLAUDE_PLUGIN_ROOT variable). The installer rewrites them to absolute paths under a shared asset directory (~/.agents/researcher), so an installed skill still resolves every reference, template, and script it points at. The installer’s test suite asserts that every referenced file exists after install.

Re-run the installer any time to refresh an existing install against a newer checkout.