Release v1.1.5 · ferro-labs/ai-gateway

[1.1.5] — 2026-06-23

Capability parity with upstream provider APIs (issue #148): implemented missing optional interfaces — embeddings on deepinfra, qwen, nvidia-nim, and ollama; live model discovery (/v1/models) on anthropic, mistral, groq, and together; azure-openai embeddings and image generation; and image generation on bedrock (Nova Canvas / Titan Image / SDXL), gemini and vertex-ai (Imagen), and xai (grok image). The shared discovery helper now supports custom auth headers (for Anthropic's x-api-key) and tolerates both wrapped and bare-array /models responses. (AI21 embeddings were excluded — AI21 exposes no embeddings endpoint.)

/v1/models and model routing now derive from the model catalog (issue #146): the model list reported per provider is sourced from the model catalog instead of hand-maintained SupportedModels() slices, eliminating drift. Precedence is live discovery > catalog > hardcoded fallback. Opt-in periodic live refresh from providers exposing a /models endpoint is enabled by setting FERRO_MODEL_DISCOVERY_INTERVAL to a positive Go duration (e.g. 6h); unset or 0 disables it (default).

Thanks to everyone who shipped this release:

@truffle-dev — #149: capability-miss now returns HTTP 404 (not 500), plus capability↔interface alignment tests
@nanookclaw — #147: SupportsModel prefix gaps (OpenAI codex/sora, Bedrock cross-region inference profiles)
@MitulShah1 — #146 catalog-derived /v1/models & routing, #148 provider capability parity, and release coordination (#225)

Full changelog: https://github.com/ferro-labs/ai-gateway/blob/v1.1.5/CHANGELOG.md