Make Codex and Claude Code subscriptions go further: a local model router that gives routine agent work to economical models and keeps your strongest model for the hard parts.
Current model metadata
Run python3 assets/model-data/current/generate.py; --check rejects stale output. This directory
is the single maintained MVP model-data source. The generator pins the complete source metadata and
exact client-discovery runs under inputs/ by byte digest and emits model-data.json, runtime-projection.json, and
coverage.json; no network or wall-clock input participates.
Schema and endpoint identifiers may retain a numeric wire-format revision. Those suffixes do not
identify model-metadata editions: the only maintained and runtime-loaded catalog is current.
The catalog contains source-scoped Provider and model records alongside access products, normalized publisher models, endpoint bindings, and dynamic-route attribution. Each maintained provider-scoped field closes to an observed fact, a traceable conservative inference, or explicit unsupported/not-applicable. Cost hints remain display-only and cannot become runtime price or free facts. Source, product, region, protocol, account, and upstream model ID scopes remain distinct, so same-name models from different Providers never share endpoints, credentials, prices, or entitlements.
The generated runtime projection promotes only records with a complete executable endpoint and adapter contract. Models newly observed from an already-qualified account/Endpoint may use the separately marked conservative runtime fallback; that does not create canonical identity, rating, price, free eligibility, or cross-Provider facts. Known image-only and internal-agent records stay outside the native text execution slice.
Claude subscription capabilities use explicit product/model/upstream-ID bindings in the same
source catalog. subscription_projection.py translates their reasoning_projection into existing
runtime contracts; it does not infer family members or account access. Check must still observe
the exact ID in the selected account's inventory before these static facts can qualify a saved
binding. Haiku 4.5 retains its manual-budget source fact and exposes only the explicit toggle
subset: disabled, or enabled with 1024 thinking tokens and an output budget greater than 1024.
Ratings are an independent slice. Existing measured/estimated records retain their exact evidence; missing ratings stay unknown rather than being filled from metadata inference.