amplicon-phylo-tree

amplicon-phylo-tree

An agent skill for building defensible, publishable phylogenetic trees from second-generation amplicon data (16S / 18S / ITS / nifH / amoA / COI and other marker genes).

It is a workflow document plus parameterised helper scripts, not a black box: every threshold is chosen for the study at hand and written down, every gate has an exit code, and every override leaves a trace.

The skill itself lives in skills/amplicon-phylo-tree/. Point your agent at SKILL.md inside it.

Install

Any of 42+ agents through the universal Skills CLI (this repository is indexed by skills.sh):

# install into every agent it detects, keeping the skill available everywhere
npx skills add mumu-sci/amplicon-phylo-tree@amplicon-phylo-tree -g

# or pick the target yourself
npx skills add mumu-sci/amplicon-phylo-tree@amplicon-phylo-tree --agent claude-code

Claude Code plugin marketplace:

/plugin marketplace add mumu-sci/amplicon-phylo-tree

By hand — copy the folder into your agent's skills directory:

cp -r skills/amplicon-phylo-tree <your-skills-dir>/amplicon-phylo-tree

The folder must stay named amplicon-phylo-tree: that name has to match the name field in SKILL.md.

Then verify the copy you installed (three commands, no external tools needed):

cd <your-skills-dir>/amplicon-phylo-tree
python scripts/selftest.py            # coordinate, stop and genetic-code contract
python scripts/check_consistency.py   # documentation vs scripts agree
python scripts/regression_tests.py    # black-box regression tests (37)

Where each client looks

Basic SKILL.md skills are portable, but every client scans its own directories — this table is the reason the same folder can serve all of them:

.agents/skills/ is the de-facto shared location: OpenCode, Codex, Cursor, Cline and DSH all read it. Installing there once covers all of them.

Installing per client — install.sh / install.ps1 (repository root) put the folder in the right root for you:

bash install.sh                 # auto-detect a skills root
bash install.sh --project       # ./.agents/skills

pwsh -File install.ps1

They refuse to overwrite a copy you have edited, strip __pycache__ (a .pyc embeds the absolute path of the machine that created it), and verify what they installed.

Per-client directory details (including DSH's scan order and the frontmatter keys it accepts): INSTALL.md.

What it is, and what it is not

It is a workflow with gates, not a one-shot script: it refuses to hand you a tree whose model does not match the data, whose tip set drifted from the taxon list, or whose alignment broke — and when it cannot check something it says so and fails instead of passing quietly.

It is not a pipeline that decides science for you: it never deletes a taxon on its own, never judges which topology is "true", and never computes your ecological metrics. It tells you what you must state in the Methods, and gives you the artefacts to back it up.

Repository layout

amplicon-phylo-tree/
├── README.md / README.zh.md      this file, in English and Chinese
├── INSTALL.md                    per-client install locations and DSH specifics
├── LICENSE                       MIT
├── CHANGELOG.md                  what changed, including the defects that were fixed
├── PRE_UPLOAD_CHECKLIST.md       the checklist this repository is released against
├── install.sh / install.ps1      install into the right skills root and verify it
├── .github/workflows/verify.yml  CI: self-test, consistency, regression, smoke test
├── .claude-plugin/plugin.json    Claude Code plugin metadata
└── skills/amplicon-phylo-tree/   the skill: SKILL.md, scripts/, references/, assets/

Contributing

Run the three verify commands above before opening a pull request; CI runs them again on 3.9 and 3.12. If you change a documented default, change it in both the reference document and the script — check_consistency.py fails the build when the two disagree, and that property is deliberate.

What it does

probe the environment -> fix the taxon rule -> audit the inputs -> normalise frame/strand
   -> align -> QC the alignment -> infer the tree -> root it -> verify -> deliver

At each step it produces the artefacts a reviewer will ask for:

What it deliberately does NOT do

• No automatic taxon deletion. The skill never silently drops a taxon. When an alignment fragments, the remedies are a reference backbone or block alignment -- both keep every tip.
• No truth adjudication. It does not decide which topology is "correct"; it decides whether the analysis you ran is defensible as stated.
• No downstream ecology. PD / UniFrac / NRI-NTI / betaNTI are not computed. It produces the tree and the statements you must make about it.
• No upstream denoising or OTU clustering. DADA2 and friends are upstream of it.
• No bundled binaries or databases. Reference databases are yours to supply.
• No fragment insertion implementation. SEPP / EPA-ng routing and documentation only (an explicit design decision).

Requirements

• Python 3.9+ (standard library only).
• MAFFT, and FastTree or IQ-TREE.
• HMMER and/or BLAST+ for the input audit of coding markers.
• R 4.x with ape and phangorn for rooting (scripts/root_tree.R).
• vsearch (recommended) for chimera screening.

python skills/amplicon-phylo-tree/scripts/probe_tools.py --require mafft,tree reports what is missing and records versions.

Known limitations

• FastTree is approximate ML. Its support values are SH-like local tests and are not comparable with UFBoot2 / SH-aLRT. The skill forces you to say so.
• Midpoint rooting is a convention, not evidence. With no outgroup the skill runs a rooting-sensitivity test by default and refuses to let a NOT RUN report pass the acceptance gate.
• ITS is not a community-phylogenetics marker. Cross-genus ITS trees are refused for metric use; within-genus trees need narrowing plus structure-aware alignment and belong in supplementary material.
• The memory heuristics are planning aids. One of them (alpha, B/cell) has been contradicted by measurement; the documents say so rather than hiding it.
• Measured anchors are from a developer machine and are labelled as such: they calibrate an order of magnitude, nothing more. Run the benchmark.

Licence

MIT. See LICENSE.

Version

1.0.1 — see CHANGELOG.md.