model-radarlisted
Install: claude install-skill Shalice-in-AIland/claude-language-assessment-research-kit
# model-radar — the maintenance watch on pinned models
The audit layers pin exact model IDs and prompt versions for reproducibility (the `## Audit` section of `review-conventions.md`). But vendors retire model aliases, reprice tiers, and revise their prompting guidance — and a dead pin silently breaks a project's audits months later. This skill is the standing check that catches that early. It changes nothing itself: findings are proposals; the user edits their own conventions.
## The three homes (check all, every run)
1. **Pinned-ID liveness — the priority check.** For each model pinned in the conventions (`model:`, `manuscript-model:`): verify the ID still exists — query the provider's `/v1/models` endpoint using the user's own key (the environment variable named in `key-env`; test presence only, never print or echo any part of a key; if unset in this session, say "verify manually" and fall back to the vendor's models/deprecations page via web search). Also search for announced retirements of the pinned IDs. **A dead or retiring pin is the highest-priority finding.**
2. **Prompting-guide deltas.** The vendor guides that govern the project's prompts — the guide for the audit scripts' model family, and the Claude prompting guidance that governs skill authoring. Diff against the log's previous entry; report material changes only. If guidance changed enough to warrant a prompt revision, recommend a review and remind: **any prompt change bumps the pinned `prompt-version`**.
3