model-registry-refresh

# Model Registry Refresh ## Doctrine `catalog/model-registry.json` is the single source of truth `scripts/model-policy.mjs` fails closed against. Every model name and reasoning-effort value it accepts must trace to official documentation with a citation — never to memory, never to a plausible-sounding guess at a slug. This skill is the repeatable workflow for keeping that registry accurate without letting research cost dominate the orchestrator's context. ## When to run - `npm run model-policy:check` fails with an error naming a model "not in the verified model registry" — the registry is missing a model the policy (or an operator) wants to use. - A provider (OpenAI, Anthropic, Cursor) ships new models or retires old ones and the catalog needs to reflect current reality. - Quarterly staleness check — `last_refreshed` in `catalog/model-registry.json` is more than ~3 months old. ## Step 1 — delegate research to Haiku Explore agents Fan out one Haiku `Explore` agent per harness (or per namespace, for codex) using the Context7 MCP tools (`mcp__Context7__resolve-library-id` then `mcp__Context7__query-docs`) plus official docs URLs already cited in the registry. Each research task must: - Ask for **exact slugs/IDs**, not families — `gpt-5.5` not "the gpt-5 line". - Ask for **reasoning-effort support per model**, not per harness — some models in a family predate newer effort levels (see `o1`/`o3`/`o4-mini` lacking `none`/`minimal`/`xhigh` in the current registry). -...