T-MJWQ-oss-skill-fixtures-corpus
State: planning/proposed · Impact: medium · Complexity: medium
Every fixture the import, publish and spec-conformance work tests against today is hand-authored,
so the messy edge cases a real published skill has — odd frontmatter ordering, extra keys, nested
resources — never show up in a test run. This task vendors a small corpus of permissively
licensed, real-world skills and wires it into the Rust, CLI and e2e suites plus a skills-ref validate CI gate, closing the “no validator oracle” gap the conformance research names.
| Location | Role today |
|---|---|
solutions/determined/tests/fixtures/agent-pants-recommended/ | Synthetic marketplace fixture (one plugin, one marketplace note); hand-authored |
solutions/determined/tests/fixtures/agent-pants-freeform/ | Synthetic freeform vault fixture (skills, references, two marketplaces); hand-authored |
solutions/agent-pants/examples/vault/ | Hand-written example vault used in docs, not a test corpus |
solutions/agent-pants/crates/agent-adapt-core/tests/fixtures/agent-skills-vault/ | Hand-authored Rust integration-test fixture vault |
solutions/agent-pants/crates/agent-adapt-core/src/discover.rs | Discover tests run against the hand-authored fixture vault; no real-world skill in the set |
solutions/agent-pants/crates/agent-adapt-core/src/adopt.rs | Adopt tests run against synthetic sources; none carry real published frontmatter |
solutions/determined/products/agent-pants/tests/e2e/authoring.e2e.ts | Playwright spec against the synthetic fixtures; no real-world import path |
solutions/agent-pants/crates/agent-pants/tests/cli.rs | CLI suite against the native binary; no fixture corpus of real skills |
.github/workflows/agent-pants-ci.yml | Runs core tests, lint, the codegen drift gate and the CLI suite; no skills-ref validate step |
Proposed
Section titled “Proposed”solutions/agent-pants/fixtures/oss/<org>-<repo>/ holds a small set of vendored, permissively
licensed real-world skills (candidates: anthropics/skills, agentskills.io’s own published
examples), each directory carrying the upstream LICENSE and a PROVENANCE.md recording the
source URL, the vendored commit sha, and the date. agent-adapt-core’s discover and adopt tests
run against the corpus alongside the existing hand-authored fixtures; a new Playwright spec
imports one corpus entry into a scratch vault end to end; the CLI suite adopts and publishes one
entry against the committed wasm. skills-ref validate runs over every rendered target for the
corpus in a new CI step, skipped rather than failed when the tool cannot install.
Approach
Section titled “Approach”- Pick three to five candidate skills from
anthropics/skillsand/or agentskills.io’s published examples, checking each one’s license permits vendoring; skip anything without a clear permissive license. - Vendor each into
solutions/agent-pants/fixtures/oss/<org>-<repo>/verbatim, with the upstreamLICENSEand aPROVENANCE.md(source URL, vendored sha, vendor date). - Extend
discover.rs’s andadopt.rs’s existing#[cfg(test)]modules to walk the corpus and assert discovery succeeds and adoption round-trips for every entry. - Add a Playwright spec that imports one corpus entry through the real import page against a
scratch vault, separate from the GitHub-source spec
import-from-github-and-source-pickerowns. - Extend
cli.test.ts(or add a sibling test file) to adopt and publish one corpus entry against the committed wasm. - Add a
skills-ref validatestep to.github/workflows/agent-pants-ci.yml: install the validator and run it over the corpus rendered through every target; the step is skipped, not red, when the tool cannot install — decide the install method (pipvspipx) as an open item. - Note in the fixture directory’s own README which fixtures are real-world versus synthetic, so a future contributor does not conflate the two.
| Area | Note |
|---|---|
solutions/agent-pants/fixtures/oss/ | Vendored real-world skill corpus, one directory per source repo with LICENSE and PROVENANCE.md |
solutions/agent-pants/crates/agent-adapt-core/src/discover.rs | Corpus-driven discover test cases alongside the existing fixture tests |
solutions/agent-pants/crates/agent-adapt-core/src/adopt.rs | Corpus-driven adopt round-trip test cases |
solutions/determined/products/agent-pants/tests/e2e/import-oss-corpus.e2e.ts | Playwright import spec over one corpus entry |
solutions/agent-pants/crates/agent-pants/tests/cli.rs | CLI adopt-and-publish assertion over one corpus entry |
.github/workflows/agent-pants-ci.yml | skills-ref validate step over the corpus |
Acceptance criteria
Section titled “Acceptance criteria”- AC-1:
solutions/agent-pants/fixtures/oss/holds at least three vendored skill directories, each with an upstreamLICENSEand aPROVENANCE.mdnaming its source URL, vendored sha, and date. - AC-2:
cargo test -p agent-adapt-corediscovers and adopts every corpus entry with no change to the vendored content. - AC-3: The new Playwright spec imports one corpus entry into a scratch vault and passes.
- AC-4: The CLI suite adopts and publishes one corpus entry against the committed wasm and passes.
- AC-5:
.github/workflows/agent-pants-ci.ymlrunsskills-ref validateover every rendered target for the corpus, and the step is skipped rather than red when the tool is absent.
Out of scope
Section titled “Out of scope”- The CI path-filter and Playwright-in-CI gap for the determined product and experience — T-I7N0-agent-manager-screenshots-docs-ledger owns that.
- Rewriting the vendored skills’ own content; the corpus is copied verbatim.
- The GitHub-source import mechanism itself — T-7V6K-import-from-github-and-source-picker.
Dependencies
Section titled “Dependencies”- T-7V6K-import-from-github-and-source-picker: the import path this corpus’s Playwright spec exercises end to end.
- T-GLK0-spec-fields-and-skills-ref-oracle: the
skills-ref validateoracle this task’s CI step reuses.
Discovery context
Section titled “Discovery context”From the user’s planning request of 2026-09-07. Closes gap G10 (“Validator oracle: none —
skills-ref validate not run anywhere”) in
solutions/agent-pants/docs/vault/research/agent-skills-conformance.md.