Bio Foundation Housekeeping

Add validated metadata models and a queryable catalog to an existing bioinformatics project. This is an independent entry point when the layout already exists; if it does not, complete bioinformatics-project as a separate setup task.

Instructions

1. Confirm bioinformatics-project has established input/output boundaries, project records, and a pinned environment. Do not create a competing project layout.
2. Adapt schemas/project-metadata.yaml for the sample, run, file, result, and provenance records the project needs. Keep identifiers stable and declare types, required fields, enumerations, and patterns in the schema.
3. Generate Pydantic models with scripts/generate_models.py. The command pins LinkML and Pydantic, rejects changed outputs, imports the generated module, and can assert that expected classes exist.
4. Validate the complete metadata bundle with the generated MetadataBundle model. Reject unexpected fields and malformed types before checking relationships.
5. Check unique identifiers and foreign keys across record collections before writing outputs. At minimum, verify run-to-sample, file-to-run, result-to-input/output-file, and provenance-to-result links.
6. Normalize validated records into one Parquet table per record class plus bridge tables for multivalued relationships. Register the tables and their relative paths, row counts, and SHA-256 values in DuckDB.
7. Exercise the full boundary with scripts/build_metadata_catalog.py and the bundled valid, model-invalid, and foreign-key-invalid fixtures. Use scripts/build_sample_catalog.py only for the smaller sample-only smoke path.
8. Before adopting a project-specific extension or migrating stored records, run scripts/check_schema_compatibility.py. Optional slots and new classes are compatible; required additions and constraint changes are reported as breaking. Keep a versioned input/expected migration fixture like fixtures/schema-migration-v1-to-v1.1.json for every supported transition.

Quick Reference

Input Requirements

Prerequisites:

• Tools declared in the project's pinned Pixi environment. See docs/README.md for expected tools.
• Target project root is writable.

Inputs:

• A project root already organized by bioinformatics-project.
• Representative valid, model-invalid, duplicate-ID, and broken-foreign-key metadata records.
• Required identifiers, fields, types, enumerations, and cross-record constraints.

Output

• schemas/project-metadata.yaml
• schemas/generated/project_metadata.py
• data/catalog.duckdb
• data/normalized/{samples,runs,files,results,result_inputs,provenance}.parquet
• results/bio-foundation-housekeeping/report.md (status, schema and Parquet SHA-256 values, row counts, and LinkML, Pydantic, and DuckDB versions)
• results/bio-foundation-housekeeping/rejections.json (failed runs only: field locations and error classes)
• results/bio-foundation-housekeeping/logs/

Quality Gates

• [ ] Model generation succeeds, the generated module imports, --expect-class finds every expected class, and the source contains no build-machine absolute paths.
• [ ] Raw metadata validates against the LinkML schema and generated Pydantic models before DuckDB ingestion.
• [ ] Duplicate identifiers and broken foreign keys fail before any artifact is published.
• [ ] report.md records the schema hash, Parquet hashes, row counts, and exact LinkML, Pydantic, and DuckDB versions.
• [ ] The valid fixture produces six verified Parquet tables and a readable DuckDB catalog that points at them; invalid fixtures exit non-zero without publishing schema, model, Parquet, or DuckDB artifacts.
• [ ] On failure, the report records the field location and error class without copying rejected values.
• [ ] Project-specific schema extensions have a compatibility report, and each supported schema-version transition has an input/expected migration fixture.

Examples

Example 1: Generate and check Pydantic models

SKILL_ROOT=~/.agents/skills/bio-foundation-housekeeping

uv run --script "$SKILL_ROOT/scripts/generate_models.py" \
  --schema "$SKILL_ROOT/schemas/project-metadata.yaml" \
  --output ./coastal-metagenomes/schemas/generated/project_metadata.py \
  --expect-class MetadataBundle

Run the same command with --check in CI. The generator exits non-zero if the output is missing or differs from the schema-derived model.

Example 2: Build the linked metadata catalog

SKILL_ROOT=~/.agents/skills/bio-foundation-housekeeping

uv run --script "$SKILL_ROOT/scripts/build_metadata_catalog.py" \
  --schema "$SKILL_ROOT/schemas/project-metadata.yaml" \
  --input "$SKILL_ROOT/fixtures/valid-project-metadata.json" \
  --project-root ./coastal-metagenomes

The bundled driver generates and imports the model in a temporary directory, validates every field and cross-record relationship, then publishes schema, model, Parquet, and DuckDB artifacts. Adapt the schema and fixtures together for project-specific records.

Example 3: Exercise the sample-only boundary

SKILL_ROOT=~/.agents/skills/bio-foundation-housekeeping

uv run --script "$SKILL_ROOT/scripts/build_sample_catalog.py" \
  --input "$SKILL_ROOT/fixtures/valid-samples.jsonl" \
  --project-root ./coastal-metagenomes

The bundled invalid fixtures demonstrate that malformed types, unexpected keys, duplicate identifiers, and missing foreign keys fail before ingestion. Rejection reports record locations and error classes, not rejected values.

Troubleshooting

Issue: A Pydantic field validator cannot see another field Solution: Pydantic validates fields in declaration order. Declare the dependency first or move the cross-field rule to @model_validator(mode="after").

Issue: A constructed model is accepted without validation Solution: Do not use model_construct() for external input. Parse raw records with model_validate(), or configure revalidate_instances="always" when existing model instances must be checked again.

Issue: LinkML generation passes but the bundle has missing references Solution: Schema validation checks record shape, while foreign keys span collections. Run the cross-record checks before creating output directories or opening DuckDB.

Issue: A generated model already exists and differs from the schema Solution: Review the schema and model diff. Remove or relocate the generated file only after confirming it is disposable; the generator does not overwrite changed files.

Issue: A killed process leaves some generated targets but no catalog Solution: Treat data/catalog.duckdb as the completion marker. Compare the remaining target paths with the run log and remove only artifacts from the interrupted run after confirming they are not project-owned; the next run refuses partial targets instead of guessing.