omics-skills

Agents and skills for bioinformatics, literature discovery, scientific writing, and data visualization. Runs under Claude Code and the Codex CLI.

Install

git clone https://github.com/fmschulz/omics-skills.git
cd omics-skills
make install

make install builds catalog/ and symlinks skills and agents into ~/.claude/ and ~/.codex/. Use make install-claude or make install-codex for a single runtime, and make status to report what's installed. make install INSTALL_METHOD=copy installs copies instead of symlinks.

Detailed install and troubleshooting notes: INSTALL.md.

Enable the routing hook

The router picks an agent and an ordered set of skills for a task. Without the hook it is a manual command; with the hook it runs on every user prompt and prepends a short routing hint to the model context.

make install-hook        # Claude Code + Codex CLI
make hook-status         # check install state
make uninstall-hook      # remove

Set OMICS_SKILLS_AUTOROUTE=0 in the environment to suppress the hint for a session without uninstalling.

Using it

Invoke an agent directly:

claude --agent omics-scientist
codex --system-prompt ~/.codex/agents/omics-scientist.md

Query the router manually:

python3 scripts/skill_index.py route \
  "assemble a metagenome and recover MAGs"

Skills are invocable individually as /<skill-name>. Agent files list which skills each agent exposes and how they compose.

Agents

Agent	Focus	Skills
`omics-scientist`	Sequencing reads, assembly, binning, annotation, phylogenomics, MAG recovery, JGI access	16
`literature-expert`	PMC full text, arXiv and bioRxiv preprints, DOI metadata, citation impact	7
`science-writer`	Manuscript drafting, multi-reviewer critique, proposal review, AI-output evaluation	7
`dataviz-artist`	marimo notebooks, matplotlib/seaborn figures, Plotly Dash dashboards, widget and badge helpers	9
`codexloop`	Plan-driven implementation harness with resumable Codex runs	2

Run python3 scripts/skill_index.py route --agent <agent> "<task>" to see how a specific agent routes a given task.

CodexLoop

codexloop is a separate harness for long-running coding work that needs durable progress tracking. The agent file is agents/codexloop.md; the runtime lives in skills/codexloop/. After make install, use the launcher:

codexloop init /path/to/project      # scaffolds docs/plans/, MEMORY.md, .codexloop/
codexloop plan --repo /path/to/project
codexloop run  --repo /path/to/project
codexloop resume --repo /path/to/project

init creates the project-local scaffold (implementation plan, doctor check, runtime state, per-project agent file). Planning uses Codex structured output to produce a task backlog; each task runs in its own Git worktree. Verification failures retry automatically until the task passes or the retry budget runs out.

Repository layout

agents/                     5 agent definitions
skills/                     36 skill directories; each has SKILL.md
catalog/                    generated router artifacts (catalog, relationships, routing)
scripts/
  skill_index.py            router and catalog builder
  routing_benchmark.py      regression harness (37 tasks)
  emit_routing_hint.py      hook payload generator
  install_hook.py           idempotent hook installer
  install.sh                shell-script install (Makefile-free)
  uninstall.sh, test-install.sh, validate-skills.py
tests/
  test_skill_index.py       unit tests for catalog and router
  test_routing_benchmark.py harness sanity tests
  test_emit_routing_hint.py hook-script tests
  routing_benchmark.yaml    routing regression suite
docs/
  ROUTING_IMPROVEMENTS.md   per-PR router deltas
  SKILL_GRAPH.md            routing model and graph
  routing_baseline.json     benchmark baseline
Makefile                    install, catalog, hook, benchmark, uninstall targets

Development

python3 -m unittest discover tests              # unit tests
make benchmark                                  # routing regression vs baseline
python3 scripts/skill_index.py build            # rebuild catalog artifacts

Adding or modifying a skill:

Create or edit skills/<name>/SKILL.md. The YAML name field must match the directory name.
Add the skill to the relevant agent's Mandatory Skill Usage and Task Recognition Patterns.
Rebuild the catalog and run the test suite.
Add a benchmark row in tests/routing_benchmark.yaml if the skill is non-trivially discoverable by the router.

See AGENTS.md for structural conventions and docs/SKILL_GRAPH.md for how the router scores and composes skills.

Compatibility

Platform	Notes
Claude Code	Primary runtime; agents installed to `~/.claude/agents/`, skills to `~/.claude/skills/`.
Codex CLI	Agents installed to `~/.codex/agents/`; skills to `~/.codex/skills/`; hook uses `~/.codex/hooks.json` with `[features] codex_hooks = true` in `~/.codex/config.toml`.
Claude API	Agent markdown files can be loaded as system prompts. Skills remain readable as reference.

License

MIT. See LICENSE.