Shape Guards

Compile-time assertions that keep ai-config's vocabulary compatible with ai-provider-bridge without an import edge between the two packages.

shape-guard.typecheck.ts is type-checked but never emitted. A build error there means the two packages' vocabularies have diverged. The active guards are:

• Provider IDs — BUILTIN_PROVIDER_IDS (ai-config) exactly matches PROVIDER_IDS (bridge).
• Model-override fields — MODEL_METADATA_FIELD_NAMES (ai-config) ⊆ keyof ModelInfo (bridge). Routing-only fields like baseUrl are intentionally excluded.
• Protocols — PROTOCOL_VALUES (ai-config) ⊆ Protocol (bridge).
• Client kinds — every CLIENT_KIND_VALUES entry (ai-config) resolves to a built-in provider id, either directly (identity) or via a non-identity mapping (aws → bedrock, snowflake → snowflake-cortex) maintained in the bridge's ProviderRegistry.

Running

Run from the repo root as part of npm run check-types, or directly:

tsc -p typechecks/tsconfig.json

tsconfig.json aliases ai-config and ai-provider-bridge to their src/index.ts, so the guard checks current source (no build step required). It is also intended to run on every PR to either package, so vocabulary divergence is caught at build time.

Adding a provider / protocol / client kind

A change to the shared vocabulary in one package fails the guard until the other is updated. Add the matching entry to both ai-config and ai-provider-bridge (see the bridge's memory-bank/providerGuide.md).