[FEATURE]: Add Aperture gateway provider

[wedow]

Feature hasn't been suggested before.

• I have verified this feature I'm about to request hasn't been suggested before.

Describe the enhancement you want to request

Aperture is Tailscale's AI gateway. It sits in front of Anthropic, OpenAI, Bedrock, Vertex, and other upstream providers, handling auth and routing for organizations.

You can configure Aperture as a custom provider today, but it doesn't work well. Each model has to be declared manually, and the correct AI SDK depends on which upstream provider is behind a given model. A Claude model routed through Bedrock needs @ai-sdk/anthropic, a GPT model needs @ai-sdk/openai, a Gemini model needs @ai-sdk/openai-compatible, etc. Users can't express that through config, and the model list changes as admins enable or disable upstream providers.

Aperture's /v1/models endpoint returns the full model catalog with upstream provider metadata, so a built-in provider can discover models at startup and route each one to the right SDK automatically. No hardcoded model list to maintain.

I work on the Aperture team at Tailscale. I've built this on my fork and have been running it for the past week. Opening a PR with that implementation as well.

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• [FEATURE]: Add Omniroute provider with dynamic model discovery #25947: Very similar request — adding a gateway/proxy provider (Omniroute) with dynamic model discovery via /v1/models, same core pattern as Aperture
• Auto-discover models from OpenAI-compatible provider endpoints #6231: Covers the general auto-discovery of models from OpenAI-compatible /v1/models endpoints, which is part of what this feature requires

If your request is distinct from those (e.g., Aperture-specific auth, routing logic, or SDK selection based on upstream provider metadata), it would help to note that difference explicitly.

[wedow]

Those issues are related but distinct. #6231 covers fetching model IDs from /v1/models but Aperture's response includes upstream provider metadata, which is what drives SDK routing as described above.

The PR does generalize the discovery loop (currently hardcoded to GitLab) so any provider can register a discoverModels function, which would benefit #6231 and #25947 as future work.