Skills

DSAgt enables an agent to discover and install skills from an external corpus during a session.

Installed skills, the three base skills included, are located in <project>/skills/. Each is a directory containing a SKILL.md file and optional reference documents.

Corpus and installed skills

Skills fall into two sets, the searchable corpus and the project's installed skills, and one class, SkillRouter, routes every skill operation between them:

• Corpus: skills in external repositories, available to install. DSAgt federates many sources (the known names below, or any git URL); each is fetched and indexed into its own collection. The agent browses the corpus with search_skills and manages sources with add_skill_source / list_skill_sources.
• Installed skills: the skills in <project>/skills/: the base skills (skill-creator, datacard-generator, aidrin), skills installed from the corpus with install_skill, and skills the agent authors with skill-creator. dsagt links each into the agent's native skills directory for the project (a relative symlink), so the agent finds them the way it finds any skill.

Rationale

• Corpus search. Every supported agent (Claude, Codex, Goose, Cline, opencode) discovers SKILL.md folders natively, so installed skills are found by the platform. search_skills covers the uninstalled ones: a corpus of thousands of skills, searchable while none of them is in context. The corpus is indexed on name, description, and tags, which keeps each entry compact and the embedding focused on what the skill does.
• Keyword fallback. When no embedding model is configured, search_skills uses a keyword match over the local source cache.
• Federated and provenance-preserving. Each source is an independent per-source collection, so re-syncing one never disturbs another; installing a skill from the corpus preserves its upstream LICENSE/NOTICE and stamps a PROVENANCE.txt into the installed directory.

Sources

dsagt init enables the genesis source by default. The others are enabled at init (the interactive checkbox, or --include <name>) or during a session with the add_skill_source tool, which also accepts any git URL. A source is a git repository holding SKILL.md directories, fetched as a GitHub tarball over HTTPS, so the end-user install needs no git; git is the fallback for a repository the GitHub API refuses (a private one the user's ssh key can reach) and for a URL on another host. Discovery is recursive under the configured subdirectory, so a skill added upstream appears after the source is re-synced: add_skill_source with force: true re-fetches a cached source.

Name	Repository	Contents
genesis (default)	github.com/AI-ModCon/genesis-skills, skills/	HPC job and site skills (slurm, pbs, perlmutter, aurora, frontier); BaseData skills (datacard-generator, croissant-validator, hdmf-schema-builder, well-convert, skill-creator); BaseEval and BaseSAFE skills; plasma simulation (gkeyll, gs2); AmSC skills (Globus Compute, IRI API, data movement); academy, literature-search.
k-dense-ai	github.com/K-Dense-AI/scientific-agent-skills	140+ chemistry, biology, medicine, and materials skills.
anthropic	github.com/anthropics/skills, skills/	Anthropic document-editing and design skills.
antigravity	github.com/sickn33/antigravity-awesome-skills	1,500+ cross-platform agent skills.
composio	github.com/ComposioHQ/awesome-claude-skills	Workflow skills for SaaS applications.

The genesis source is the ModCon aggregation point: skills contributed by ModCon and AmSC teams are located there and become searchable on the next sync.

Base and authored skills

Every dsagt init installs three base skills into <project>/skills/ from the repositories that maintain them. A source is fetched once into the shared cache at ~/dsagt-projects/.skill_sources/ and reused by every later init; a cache held at another branch or tag than the one asked for is re-fetched, and add_skill_source with force re-fetches on request.

Skill	Source
skill-creator	genesis, skills/basedata-skills/skill-creator/
datacard-generator	genesis, skills/basedata-skills/datacard-generator/
aidrin	github.com/idtlab/AIDRIN, .claude/skills/aidrin/, at the release tag of the installed aidrin package (v2026.08.2 for 2026.8.2), so the skill describes the CLI dsagt installs. The AI-readiness check runs its quality baseline at every tabular stage, and aidrin is registered as a code so every call is recorded.

The scripts a base skill's workflow runs, and the CLI the aidrin skill documents, are registered as codes at the same init, so the agent runs them through dsagt-run and every run is recorded. The installed aidrin skill shows that CLI in its registered form (dsagt-run --code aidrin -- aidrin …), and its PROVENANCE.txt records the rewrite; the rest of the text is upstream's. datacard-generator registers datacard-introspect, datacard-validate, and datacard-convert-v1; each executes its script in place under <project>/skills/datacard-generator/scripts/, and the two that need PyYAML and Pydantic run under uv run --with. The vectors search_registry matches these codes by are embedded once, with the built-in tools, into the shared ~/dsagt-projects/kb_index/codes collection that every init copies, so dsagt init loads no embedding model; the collection rebuilds when the dsagt version or a base-skill code spec changes.

A skill declares the codes its scripts are registered as in a codes list in its SKILL.md frontmatter, one entry per script with the code's name, the script's path inside the skill, its description, its parameters with cli and role, and its dependencies; the entry is the code's spec, so the script runs under uv run --with its packages and every run records the files its parameters name. A script with no entry is registered from its argparse calls, with no dependencies and no roles. save_skill takes the same list in its spec, and the base-skill specs are dsagt's own table, since those skills are maintained upstream.

The agent discovers installed SKILL.md folders natively, so search_skills indexes the corpus only. Other domain skills, such as the BaseData croissant-validator, are installed from the corpus and stay current upstream.

To add a skill by hand, place a new directory under <project>/skills/ with a SKILL.md describing the workflow; the next dsagt start links it into the agent's native skills directory, after which the agent discovers and invokes it.

Try it

dsagt init            # follow the prompts: name it `demo`, then pick your agent
dsagt start demo      # launch the agent in the project

Then, in the agent:

1. List the skill sources and their sync status.

2. Sync the genesis source and search it for a data-card skill.

3. Install the one that fits, then use it on this project.