Add Wafer as a model provider

[ianye23301]

Summary

• Adds Wafer (https://wafer.ai) as a new endpoint type alongside cerebras, groq, together, etc.
• Wafer's serverless catalog is exposed at https://pass.wafer.ai/v1 and is OpenAI-compatible (chat completions, streaming, tools, /v1/models), so it slots into the existing OpenAI-format passthrough — no new request/response translator is needed.
• Registers the 7 publicly-listed Wafer models with available_providers: ["wafer"]. Pricing and context lengths come straight from https://pass.wafer.ai/v1/models on 2026-05-27.

Changes

• packages/proxy/schema/models.ts — add "wafer" to ModelEndpointType.
• packages/proxy/schema/secrets.ts — add "wafer" to the simple-types enum in APISecretSchema.
• packages/proxy/schema/index.ts — register WAFER_API_KEY → wafer in AISecretTypes and EndpointProviderToBaseURL.wafer = "https://pass.wafer.ai/v1".
• packages/proxy/schema/model_list.json — add the 7 Wafer-served public models:
  • GLM-5.1 ($1.20 / $3.60 per Mtok, 203K ctx)
  • Kimi-K2.6 ($0.88 / $3.84, 262K ctx, multimodal)
  • Qwen3.5-397B-A17B ($0.48 / $2.88, 262K ctx, multimodal)
  • Qwen3.6-35B-A3B ($0.15 / $1.00, 262K ctx, multimodal)
  • qwen3.7-max ($5.00 / $15.00, 256K ctx)
  • deepseek-v4-flash ($0.14 / $0.28, 1M ctx)
  • deepseek-v4-pro ($1.74 / $3.48, 1M ctx)
• packages/proxy/scripts/verify_proxy_models.ts — keep the local ModelEndpointType mirror in sync.

How it works at request time

A secret of type: "wafer" with secret: <WAFER_API_KEY> is routed through fetchOpenAI exactly like cerebras/groq/together: EndpointProviderToBaseURL.wafer supplies the base URL, the Authorization: Bearer <key> header is set by the shared code path, and the chat-completions body is forwarded unchanged. End users can call any of the model IDs above through the proxy and get OpenAI-formatted responses back.

Test plan

• pnpm install && pnpm --filter @braintrust/proxy build — clean (TS build succeeds, including the schema d.ts regeneration).
• pnpm test — same 89 pre-existing failures as main (all "No API keys found" against live providers; no new failures introduced).
• python3 -c 'import json; json.load(open("packages/proxy/schema/model_list.json"))' — model list still parses.
• Smoke-test end-to-end through a deployed proxy with a real WAFER_API_KEY against e.g. GLM-5.1. Happy to do this once a maintainer points me at the right staging entry point — or it can be done at review time.

Notes

• I'm posting this as a draft so you can shape the wiring (naming, where to register, whether you want an "available_regions" value, etc.) before I clean it up.
• Happy to add a WaferMetadataSchema later (e.g. for an api_base override so on-prem / EU edges can be pointed at directly), but didn't include one in this PR to keep the diff minimal — Wafer customers on the default region don't need any metadata.

[vercel]

ianye23301 is attempting to deploy a commit to the Braintrust Team on Vercel.

A member of the Team first needs to authorize it.