feat(ai): honor per-provider base URL from env

[kundeng]

Problem

ai() and assistant() resolve models via PROVIDERS[targetProvider](modelId) (src/lib/ai.ts), which uses the bare provider singletons. Those pin the vendor endpoint — @ai-sdk/openai hardcodes https://api.openai.com/v1 and reads no base-URL env var:

const baseURL = withoutTrailingSlash(options.baseURL) ?? "https://api.openai.com/v1"

options.baseURL only exists if you construct the provider yourself. So KIT_AI_DEFAULT_PROVIDER and KIT_AI_DEFAULT_MODEL let you choose who and what, but there is no way to choose where.

The effect: Kit's built-in AI helpers cannot be used with any OpenAI-compatible endpoint that isn't the vendor's own — LiteLLM, Ollama, vLLM, a self-hosted router, or a corporate proxy. The only workaround is to skip ai()/assistant() and hand-construct a provider in every script, which throws away the multi-turn state assistant() exists to manage.

Change

Read an optional per-provider base URL from the environment; when set, build the provider with the matching create* factory.

Provider	Existing key var	New (optional)
openai	OPENAI_API_KEY	OPENAI_BASE_URL
anthropic	ANTHROPIC_API_KEY	ANTHROPIC_BASE_URL
google	GOOGLE_API_KEY	GOOGLE_BASE_URL
xai	XAI_API_KEY	XAI_BASE_URL

Names follow each SDK's own convention and sit alongside the existing getProviderEnvVar map.

KIT_AI_DEFAULT_PROVIDER=openai
KIT_AI_DEFAULT_MODEL=my-model
OPENAI_BASE_URL=http://localhost:4000/v1
OPENAI_API_KEY=sk-local

const chat = assistant("You are helpful.")   // no per-script wiring

Compatibility

• No behaviour change when the var is unset — PROVIDERS[provider] is used exactly as today.
• Read at call time, not module load, so a script that sets it via env() before its first ai() call is honored. (KIT_AI_DEFAULT_PROVIDER/KIT_AI_DEFAULT_MODEL still cache at module load; unchanged here.)
• openrouter left commented out to match the existing v5 TODOs.

Verification

tsc --noEmit -p . reports 0 errors both before and after the patch.

Two cases added to src/lib/ai-env-integration.test.ts (base URL set → constructs; unset → unchanged).

Heads-up: those tests don't currently run. test/ava.config.mjs globs src/**/*.test.js, so every *.test.ts under src/lib/ — including the existing ai-env-integration.test.ts and ai.test.ts — is excluded. Happy to wire that up here or in a separate PR, whichever you prefer.

🤖 Generated with Claude Code

---

Update: added a custom provider

Per-provider *_BASE_URL covers "I reach a vendor through a gateway." It doesn't cover the more common self-hosted case: the endpoint is its own provider, serving many vendors' models under one namespace. Pointing OPENAI_BASE_URL at a router that answers for Claude, Gemini and GPT works, but misnames what it is.

So this branch also adds a custom provider backed by @ai-sdk/openai-compatible — already in the tree transitively, now a direct dependency:

KIT_AI_DEFAULT_PROVIDER=custom
KIT_AI_CUSTOM_BASE_URL=http://localhost:4000/v1
KIT_AI_CUSTOM_API_KEY=...
KIT_AI_DEFAULT_MODEL=whatever/the-endpoint-serves

Also works in prefix form: resolveModel('custom:some-model').

Unlike the vendor providers, custom has no default endpoint, so selecting it without KIT_AI_CUSTOM_BASE_URL throws an error naming the missing variable rather than failing later against an undefined URL.

Verified live: with KIT_AI_DEFAULT_PROVIDER=custom, resolveModel(undefined) returns custom.chat / the configured model id and completes a real generateText call against a local OpenAI-compatible router. tsc --noEmit -p . stays at 0 errors.

The two mechanisms compose — use *_BASE_URL to redirect a vendor, or custom to declare an endpoint that stands on its own. Happy to split them into separate PRs if you'd rather review them independently.