Skip to content

omics-skills v2.0.0

Choose a tag to compare

@github-actions github-actions released this 11 Nov 16:02
· 4 commits to main since this release

omics-skills 2.0.0

Version 2.0.0 cuts the pack from 27 skills to 12 and from three agents to one, and removes the
keyword router. Current models pick a skill from its description and know the standard tools, so
the pack now holds what a model cannot derive: house rules, the owner's own tools, a hosted
literature corpus, and small tested helpers. This release removes skills, scripts and Makefile
targets; read "Upgrade from 1.x" before you update.

Skills

Twelve skills remain: omics-pipeline, bioinformatics-project, polars-dovmed,
crossref-lookup, public-db-lookup, beautiful-data-viz, notebooks, bio-logic, and the
stage skills tracking-taxonomy-updates, bio-annotation, bio-viromics and
bio-interdomain-hgt.

Folded into omics-pipeline. Six stage skills are now one decision page each under
skills/omics-pipeline/references/: reads QC and mapping, assembly, binning, gene calling,
phylogenomics and structure. A page gives the default tool per case, the rules a model would not
guess, command lines, and the QC gates to record. Tool notes stay only for tools that models
know poorly: QuickBin, Autocycler, pyrodigal-gv, Infernal with its Rfam model table, Boltz and
the pinned TM-Vec fork. The drivers, fixtures and general tool manuals of these six stages are
gone; a project writes its own restartable driver as bioinformatics-project describes.

Removed. exploratory-data-analysis, bio-workflow-methods-docwriter,
bio-foundation-housekeeping, bio-stats-ml-reporting, bio-protein-clustering-pangenome,
bio-fasta-database-curator and preprint-search.

Moved out of the pack. proposal-review and ai-scientist-evaluator are review tools, not
omics skills, and are now maintained outside this repository.

Slimmed. bio-logic keeps the hypothesis register, the reflection format and the
claim-against-evidence check, without its textbook reference pages. notebooks drops four
copied marimo documents. bio-annotation and bio-viromics drop the general manuals for
DIAMOND, eggNOG-mapper, InterProScan, TaxonKit, geNomad and CheckV and keep one decision page
each.

Agents

literature-expert remains. omics-scientist and science-writer are removed; the discovery
and reporting rules of omics-scientist are now rules 8 and 9 of omics-pipeline.

Router, catalog and hint hook

scripts/skill_index.py, the generated catalog/catalog.json, the routing benchmark and its
baseline, and the per-prompt routing hint hook are removed, with the Makefile targets
build-catalog, install-catalog, install-hook, uninstall-hook, hook-status and
benchmark. In 90 days of session logs the router was consulted in fewer than one in ten of the
sessions that used a pack skill, and it recommended a skill for prompts such as "how do I rename
a git branch". make test now runs the three validators and the unit tests.

The installer keeps install, install-claude, install-codex, install-selected,
uninstall, status, validate and clean. A skill selection (SELECTED_SKILL_DIRS) installs
exactly the selection with every install target; a selection that holds omics-pipeline or a
stage skill gets the whole stage group, and the installer prints what it added. An install
removes a link to a retired skill or agent only when the link points into this checkout. It no
longer archives copied skill directories or rendered Codex agents. A directory under skills/
without a SKILL.md, such as one that only ignored __pycache__ files keep alive after the
update, is skipped.

Other changes since 1.7.1

Version 1.8.0 was never tagged; its changes ship here:

  • beautiful-data-viz makes greyscale the default for every figure and uses color only to
    encode information.
  • A version in a skill records what its commands were checked against. It is not an install
    pin: a project takes the newest release at setup and locks it. No skill, script or example
    holds a real email address, and a test fails on one.
  • bioinformatics-project uses pixi instead of conda, works on main, commits
    tasks/METHODS.md, and submits Slurm drivers from the project root.
  • polars-dovmed helpers run from the installed skill directory.
  • The installer renders and installs agents idempotently, and LINK_CLIENT_SKILLS=no leaves
    ~/.claude/skills and ~/.codex/skills to another tool.
  • The docs site has seven pages instead of fifteen.

Fixes since 1.7.1

A full audit preceded the restructure. The fixes that remain relevant after it:

  • bio-annotation, bio-viromics and bio-interdomain-hgt: a failed driver run left partial
    output, so the rerun stopped on "output directory is not empty". Validation now runs before
    the first write, and a table row with the wrong field count is rejected on read.
  • bio-interdomain-hgt: tree support on a percent scale passed the 0.9 gate, and domain labels
    outside the vocabulary were accepted.
  • tracking-taxonomy-updates: Bacteriophage routed to GTDB-Tk, and the EukCC job skipped
    bins that do not end in .fa (new variable EUKCC_SUFFIX).
  • polars-dovmed: the helper refused to run without an API key although the hosted API is
    public.
  • public-db-lookup: MGnify moved to API v2; esearch reports the hit count; text without a
    charset is decoded as UTF-8; URL-encoded keys and addresses are redacted from error messages.
  • crossref-lookup: the bibliography audit counted booktitle fields as missing DOIs.
  • beautiful-data-viz: slides fonts were smaller than notebook fonts when seaborn was
    installed, and finalize_axes raised a warning under constrained_layout.
  • notebooks: execute_notebook.py failed without --kernel and kept the kernel placeholder.
  • bioinformatics-project: --check failed once a scaffolded file was edited.
  • scripts/validate-citations.py: a citation that names an author or year now fails when its
    cache record holds neither.

Upgrade from 1.x

  1. If you installed the routing hint hook, run make uninstall-hook in the 1.x checkout before
    you update. Version 2.0.0 has no such target, and a hook that stays installed calls a script
    that no longer exists.
  2. Update the checkout and run make install. Links to removed skills and agents that point
    into the checkout are removed, and so are the router files under ~/.agents/omics-skills/.
  3. Delete by hand what the installer cannot attribute to this pack: the rendered Codex agents
    ~/.codex/agents/omics-scientist.toml and ~/.codex/agents/science-writer.toml, and, after
    a copy install, the directories of removed skills under ~/.agents/skills/.
  4. Scripts that projects called by path are gone with their skills, among them
    run_reads_qc_mapping.py, run_assembly_qc.py, run_gene_calling.py,
    run_phylogenomics.py, run_structure_annotation.py, build_bin_qc_tables.py,
    build_pangenome_artifacts.py, curate_fasta.py, eda_analyzer.py and the workflow-run
    extractors. Check out tag v1.7.1 to copy a script into a project that still needs it.
    The 1.7.1 copies do not contain the fixes made after that tag.

Validation

The truth-set registry under validation/ now lists the three remaining drivers (annotation,
viromics, interdomain HGT). All three truth sets are candidates; no validation job is defined.