Model catalog (models-api.json)

src/escapement/llm/models-api.json is a minified models.dev dump, keyed by provider. Loaded at runtime by escapement.llm.catalog-source (CLJ/bb) and baked in at CLJS compile time by escapement.llm.catalog-macros.

Update

Refresh from the live models.dev API (fresher than opencode's vendored packages/opencode/test/tool/fixtures/models-api.json, which is a stale pin):

curl -fsSL https://models.dev/api.json | jq -c '.' > src/escapement/llm/models-api.json
bb test   # catalog_test asserts exact values (e.g. gpt-5 ctx/pricing) — confirms the schema still parses

Keep it minified (jq -c). Query with jq, never read the whole file.

Provider coverage

The dump is only half the story: catalog-source's allowlist decides which providers are surfaced, so refreshing the JSON does not by itself make a new provider visible. The other half is the hand-curated local-providers overlay in escapement.llm.catalog, which is the correct source for providers models.dev does not carry — or carries in a way that would mislead.

Every provider in providers/provider-templates is covered by one of the two, and catalog_source_test asserts that no template is invisible to the catalog.

Two mappings are deliberate and should not be "fixed" into the allowlist:

• :codex / :openai-codex stay on the overlay. The ChatGPT-account endpoint serves OpenAI models, so mapping it to the openai dump entry looks right and is not: the dump lists 48 models and that path accepts six. Advertising the dump's list would let a :needs gate pick a model guaranteed to fail with "The '' model is not supported when using Codex with a ChatGPT account".
• :claude-cli has no models.dev provider at all, and its --model takes the CLI's own aliases (sonnet), not Anthropic ids — so borrowing Anthropic's table would attach real prices and context windows to identifiers that provider does not accept. The overlay carries the aliases.