v0.1.0

• Distribution: MIT license for Frost code, separate source-data notices, contributor guidance, and both frost and catalog-build in local installs, release archives, and the Homebrew formula. Archives include configuration and producer assets and are smoke-tested before release.
• frost config check reports advisory, effort-aware warnings when a profile's latency class is more than one tier faster than the producer suggestion. Missing suggestion files remain optional.
• Capacity-aware selection. frost route and replay accept neutral capacity snapshots, --availability exclude|demote|ignore controls fresh exhausted pools, frost capacity check explains pool and profile state, and the optional CodexBar wrapper publishes snapshots independently of Frost. Capacity never raises the quality floor or turns missing telemetry into unavailability.
• Public catalog producer. catalog-build refreshes source-attributed model data from models.dev and SWE-bench Verified, supports restricted Artificial Analysis enrichment, applies reviewed identity mappings and durable local overrides, prints diffs, preserves last-known-good source data, and publishes atomically with reproducible fixtures. It also emits effort-aware latency suggestions as configuration priors.
• Model recommendation CLI. frost route (positional, --file, --stdin, --request), profiles list, config check, explain, and --record/--replay; deterministic router core with offline policy tests; TypeSafe adapter; TOML operator config and JSON catalog with portable examples and question spec v3.
• Runtime review hardening: invalid record destinations fail before analysis, and post-analysis append or sync failures return a recoverable structured record. The capacity producer honors source bindings, preserves the prior snapshot when every selected request fails, and publishes explicit failed-pool states on partial success. Catalog refresh retains last-good Artificial Analysis measurements only in the restricted file, rejects aliased output and backup paths before writing, and verifies recorded fixture digests.
• Evidence remains incomplete: the September 2026 SWE-bench Verified snapshot contains no applicable rows for the current configured models, the bundled catalog example has no measurements, latency suggestions are not end-to-end measurements, and no paired completion-outcome evaluation has been run. Recommendations therefore remain provisional.