Model Runtime API v0.1.0
Pre-release
Pre-release
Model Runtime API v0.1.0
First pre-alpha release of the provider-neutral Model Runtime API.
Included
- Versioned execution protocol, generated request schema, and OpenAPI surface
- Idempotent asynchronous submit, status, result, cancellation, and resumable SSE lifecycle
- Capability routing, bounded fallback, retry budget, provider-local concurrency, and EventStore SPI
- OpenAI Responses, Anthropic Messages, fal Queue, and deterministic Mock adapters
- Usage facts and metadata-only OpenTelemetry GenAI mapping without customer billing
- TypeScript, Go, and Python clients
- Versioned capability catalog and model diff
- Non-root read-only container and Kubernetes sidecar example
- Threat model, provider evidence registry, Gateway boundary, and public-data controls
Verification
- 9 TypeScript test files and 26 tests passed
- 2 Go SDK and 2 Python SDK tests passed
- Generated schema, OpenAPI/document links, Biome, typecheck, and build passed
- Custom public-data scan and Secretlint passed over 108 candidate files
- Dependency audit reported no known vulnerabilities
- Final container was healthy as non-root with a read-only root filesystem
- A fresh clone of GitHub
mainpassed frozen installation and the complete gate
See full verification evidence.
Important limits
- The reference server is unauthenticated and is not an internet-ready multi-tenant gateway.
- The bundled EventStore is in-memory; production durability is an SPI extension.
- Provider adapters are tested with local sanitized fixtures, not live-account credentials.
- Provider credentials, tenant control, commercial quotas, and customer billing remain deployment or
Gateway responsibilities. - The GitHub Actions workflow template is included under
docs/ci/. Activation is pending a GitHub
credential withworkflowscope; two official device-flow requests expired without account
confirmation.