SuperAPI AI Directory
Partner中文AI newsFavorites
Skillsdym-scientific-agent-skills-chay-script-bundled-tu-shell

dym-scientific-agent-skills-chay-script-bundled-tu-shell

Agents & models

View repository

Updated 10-18 01:22

About this skill

Skill: dym-scientific-agent-skills-chay-script-bundled-tu-shell — Run bundled scripts straight from the shell (no agent needed)

Everything in this file is typed into the terminal. The prompt for the agent is in 03-dung-skill-trong-agent.md.

Why use it: 105/163 skills carry a scripts/directory, and 90 of those skills have scripts written with argparse. This is a real CLI, runs deterministically, costs no tokens — usable in a Makefile, cron, CI, or just to check quickly without opening the agent. What it produces: JSON/TSV on stdout or in a file specified by -o , plus an exit code that you must branch on.

Find a runnable script

# skill nào có scripts/
ls -d skills/*/scripts | cut -d/ -f2

# script nào là CLI argparse
grep -l argparse skills/*/scripts/*.py | head -20
every argparse script answers --help (this is the repo's CI contract)

python skills/paper-lookup/scripts/paginate.py --help


Hợp đồng "`--help` luôn chạy" không phải lời hứa suông: `tests/_contract/cli.py` sinh test case cho **mọi** script argparse trong repo, và `tests/_meta` chạy nó lên toàn bộ 163 skill trong mỗi PR.

## Ví dụ 1 — chuyển toạ độ genomic (thuần stdlib, offline, đã chạy thật)

```bash
python skills/genomic-coordinates/scripts/convert_coords.py --list
python skills/genomic-coordinates/scripts/convert_coords.py chr1 1000 2000 --from bed --to gff

Real output:

contig	input	output	length	status	detail
chr1	1000-2000	1001-2000	1000	ok

And the convention table from --list:

format         convention            notes
bed            0-based half-open     BED3/6/12, narrowPeak, broadPeak
gff            1-based inclusive     GFF3
gtf            1-based inclusive     GTF/GFF2, GENCODE, Ensembl
vcf            1-based inclusive     POS..POS+len(REF)-1
sam            1-based inclusive     SAM text POS (BAM stores it 0-based)
interval-list  1-based inclusive     Picard/GATK

Scripts in the same skill: audit_intervals.py, check_contigs.py, normalize_variant.py. All four use only the standard library — nothing extra to install.

Example 2 — paginating a literature API, with count reconciliation

# xem kế hoạch gọi mà không gọi mạng
python skills/paper-lookup/scripts/paginate.py \
  --api openalex --query "search=crispr" --max-records 200 --dry-run

Real output of --dry-run (exit 0, no request left the machine):

{
  "api": "openalex",
  "first_url": "https://api.openalex.org/works?...&per-page=100&cursor=%2A",
  "delay_seconds": 0.2,
  "query_format": "query is a raw parameter string, e.g. 'search=crispr' or 'filter=publication_year:2024'"
}

Supported APIs: biorxiv, crossref, europepmc, medrxiv, openalex (see --list-apis).

The sibling scripts all read from a file or - (stdin), so they can be piped together:

python skills/paper-lookup/scripts/arxiv_atom.py feed.xml --ids-only
python skills/paper-lookup/scripts/openalex_abstract.py works.json --text-only
python skills/paper-lookup/scripts/jats_to_text.py article.xml

Example 3 — sub-command client (ARAX knowledge graph)

python skills/ncats-arax/scripts/arax_client.py --help
python skills/ncats-arax/scripts/arax_client.py preflight

Sub-commands: preflight (checks OpenAPI + production version), normalize (normalizes free-text terms into CURIEs, for review only), one-hop, two-hop, summarize. No API key needed. Note the constraint the skill itself states: queries may be public, use only for non-sensitive research questions.

Exit codes — READ them, don't assume 0/1

The repo deliberately uses non-standard exit codes in places where the API returns HTTP 200 for a silent failure. Quoted from the docstring of fail() in skills/paper-lookup/scripts/_common.py:219:

Scripts here exit non-zero on silent-failure conditions — a JATS document with no <body>, an arXiv Error entry, a pagination shortfall — precisely because the APIs return HTTP 200 for them.

Script Exit Meaning Source
paper-lookup/arxiv_atom.py 3 feed is arXiv error response (arrives as HTTP 200, totalResults 1, an entry named 'Error') arxiv_atom.py:166
paper-lookup/paginate.py 4 shortfall when reconciling counts — returns fewer records than the API itself claims paginate.py:478
paper-lookup/paginate.py 2 missing --api/--query paginate.py:426
analytical-method-validation, iso-standards-readiness, pkpd-modeling (scripts/_common.py) 0 / 1 / 2 EXIT_OK / EXIT_FINDINGS / EXIT_INPUT_ERROR — 1 means "there is a finding", not "crash" _common.py:32-34
exa-search, lab-hardware-cad, pyopenms 2 input error sys.exit(2) literal

This means the automation wrapper script must branch explicitly:

python skills/paper-lookup/scripts/paginate.py --api crossref --query "query=crispr" -o out.json
rc=$?
case $rc in
  0) echo "ok" ;;
  4) echo "count reconciliation shortfall — kết quả KHÔNG đầy đủ, đừng dùng"; exit 1 ;;
  *) echo "lỗi (exit $rc)"; exit $rc ;;
esac

With the three skills using EXIT_FINDINGS = 1, do the opposite: 1 is a valid result ("found an issue"), merging it with real errors is wrong.

How to check the exit code of any script before trusting it:

grep -rn 'sys.exit(\|SystemExit(\|EXIT_[A-Z_]* *=\|code=' skills/<name>/scripts/*.py
python skills/<name>/scripts/<script>.py --help   # nhiều script ghi thẳng mã lạ vào description

Dependency

Scripts install nothing themselves. 34 skills are purely standard library — run immediately with python3 and you're done:

adaptyv, analytical-method-validation, bgpt-paper-search, clinical-decision-support,
clinical-reports, consciousness-council, dhdna-profiler, generate-image, genomic-coordinates,
ginkgo-cloud-lab, hypothesis-generation, iso-standards-readiness, markdown-mermaid-writing,
market-research-reports, ncats-arax, ontology-term-resolution, optimize-for-gpu, pacsomatic,
paper-lookup, paperzilla, parallel-web, pathml, pathogen-variant-surveillance, peer-review,
pi-agent, protocolsio-integration, research-grants, scholar-evaluation, scientific-brainstorming,
scientific-critical-thinking, scientific-writing, treatment-plans, venue-templates, what-if-oracle

(Source: the sections packages = [] in tests/skill-requirements.toml.)

For skills that need packages, tests/skill-requirements.toml is exactly the list to install — use it instead of guessing from import:

python3 - <<'PY'
import tomllib
cfg = tomllib.load(open("tests/skill-requirements.toml", "rb"))
entry = cfg["skills"]["scanpy"]
print("python:", entry.get("python", cfg["defaults"].get("python")))
print("packages:", " ".join(entry.get("packages", [])))
PY

Then run in a temporary environment, without touching the system environment:

uv run --with scanpy --with anndata python skills/scanpy/scripts/<script>.py --help

Pitfalls

  • Don't merge the scripts/ of multiple skills into the same sys.path. 32 skills all ship a file named scripts/_common.py; importing them into one process means _common points to whichever skill loaded first. Run using the full path to each script, don't sys.path.append multiple skill directories.
  • 18 packages are recorded as not installable anywhere in the section [unavailable] of tests/skill-requirements.toml, with reasons — matlabengine requires MATLAB preinstalled, omero-py requires a C++ build, flash-attn/deepspeed/bitsandbytes requires a CUDA toolchain, rsgislib/pdal is only on conda-forge. Read that section before wasting a whole session.
  • A few skills need an interpreter older than 3.13 (pytdc, molfeat, deepchem, histolab, vaex, ete3). The key python in the entry tests/skill-requirements.toml of that skill states which version.