product-audit-swarm

name

product-audit-swarm

description

Use when the user asks to audit, sweep, re-review, re-score, compile a report on, or check the consistency of DLinRT.eu products. Triggers on "audit the products", "run the product swarm", "compile a product review", "re-score evidence for all products", "check products for inconsistencies", "product QA pass". Runs a multi-role agent swarm over `src/data/products/<category>/*.ts`, aligned to `docs/FIELD_REFERENCE.md`, `docs/review/GUIDE.md`, the inclusion policy, and the dual-axis evidence rubric, and emits Markdown + CSV audit artefacts under /mnt/documents/.

Product Audit Swarm

Coordinated audit of every active ProductDetails entry. Specialised roles each check one slice of the schema; a Compiler merges their findings into one report per product. No product files are auto-edited — output is a worklist for a separate, confirmed edit pass.

When to use

• "Audit the products" / "run the product audit swarm" / "compile a product report".
• "Re-score evidence for all products" after the rubric changes.
• "Check products for inconsistencies" or pre-release QA sweeps.
• Optionally scoped to one --category folder or one product id.

Do not use for company-only audits, in-app feature work, or auto-editing product .ts files.

Workflow

1. Bootstrap — read once per session and cache:
   • docs/FIELD_REFERENCE.md (field semantics)
   • docs/review/GUIDE.md (review rubric)
   • src/data/evidence-impact-levels.ts (E/I/R definitions)
   • references/scoring-rubric.md, references/inclusion-policy.md, references/roles.md, references/output-format.md from this skill
2. Enumerate the worklist:

   code--copy knowledge://skill/product-audit-swarm/scripts/enumerate_products.sh /tmp/enumerate_products.sh
   code--exec bash /tmp/enumerate_products.sh             # all categories
   code--exec bash /tmp/enumerate_products.sh auto-contouring   # one category
   code--exec bash /tmp/enumerate_products.sh --id raysearch-raystation
   

   Excludes archived/ and examples/.
3. Per product, run roles (definitions in references/roles.md). Roles independent of each other run in parallel within a single code--exec batch:
   • Independent: Identity, Inclusion, Regulatory, Technical, Structures, Transparency.
   • Sequential: Evidence (after websearch--web_search PubMed/DOI lookup), then Cross-check (uses cached registry).
4. Compile one Markdown block per product following references/output-format.md. Each finding carries a severity (error|warn|info) per the rubric below.
5. Aggregate three artefacts to /mnt/documents/:
   • product-audit-<YYYY-MM-DD>.md — full per-product report
   • product-audit-<YYYY-MM-DD>.csv — one row per finding
   • product-audit-<YYYY-MM-DD>-rescoring.csv — id, current E/I/R, proposed E/I/R, rationale
6. Hand off the artefacts with <presentation-artifact> tags. Stop. Do not edit product .ts files in the same turn — the Minimal Intervention policy requires a separate, user-confirmed edit pass, ideally one PR per category wave.

Severity rubric

• error — invalid enum, broken URL, duplicate id, archived-company reference from a live entry, AI/DL inclusion violation, certification summary disagrees with regulatory.*, missing required field.
• warn — recommended field missing for the assigned (E,I,R) (e.g. keyPapers empty at E ≥ E1), lastRevised > 180 days, DICOM nomenclature drift, structure name not in Region: Structure Name form.
• info — neutral observation or a re-scoring suggestion where the current score is defensible.

Composability

When re-scoring requires summarising several abstracts, defer to the bundled AI Gateway skill: knowledge://skill/ai-gateway/SKILL.md. Fall back to websearch--web_search and code--fetch_website when the gateway is unavailable. Keep all queries timeboxed; record the search date in the per-product Sources Consulted section.

Success criteria

• Worklist enumerated and product count printed.
• For every product in scope, the per-product Markdown block contains all eight role sections (or an explicit "n/a — reason" stub).
• Both CSVs parse cleanly and the row counts match the Markdown report.
• Artefacts surfaced via <presentation-artifact>; no product .ts files modified in this turn.