feat(byok): per-user provider config — proxy URL + pinned model

[harshithmullapudi]

Summary

Extends the BYOK settings surface so users can configure their own proxy endpoint and pinned model per provider — not just an API key. Chats that run against any of these keys still skip the aura debit; the gate is `usedBYOK`, which stays true across all four slots.

Four provider rows on the settings card:

Slot	Key	Base URL	Model
Anthropic	required	—	optional (defaults to `claude-haiku-4-5-20251001`)
OpenAI	required	—	optional (defaults to `gpt-5.4-mini`)
OpenAI Proxy	required	required	optional (proxy routing key)
Ollama Cloud	required	optional (defaults to `https://ollama.com/v1\`)	optional (defaults to `gpt-oss:120b-cloud`)

Changes

• DB (`packages/db/prisma/schema.prisma`) — two nullable columns on `ModelKey`: `baseUrl`, `model`. Migration `20260728084341_modelkey_proxy_fields` (safe — only adds columns).
• Store (`apps/web/src/lib/byok/store.ts`) — adds `openai_proxy` to `BYOK_PROVIDERS`. Enforces baseUrl on `openai_proxy`, drops baseUrl silently for providers that don't accept it, always accepts a pinned model. `resolveByokForUser` walks all four slots with `LLM_PROVIDER` as an explicit override and a natural priority chain (anthropic → openai_proxy → openai → ollama).
• Chat model (`apps/web/src/lib/chat-model.ts`) — four BYOK branches. `anthropic`/`openai` stay direct against their vendor endpoints; `openai_proxy` uses `createOpenAI({ baseURL, apiKey })`; `ollama` uses `createOpenAICompatible` so a user can point at a self-hosted daemon or Ollama Cloud with their own key.
• UI (`apps/web/src/ui/BYOKSection.tsx`) — four rows. `baseUrl` input renders for `openai_proxy` (required) and `ollama` (optional). `model` input renders for every provider. Saved rows show `url · ` and/or `model · ` beneath the status pill. Client-side validation: `http(s)://` on any provided baseUrl, and the `openai_proxy` baseUrl requirement.
• API (`apps/web/src/app/api/byok/route.ts`) — already accepted optional `baseUrl` / `model`; store enforces per-provider rules.
• Ollama helper (`apps/web/src/lib/ollama.ts`) — exports `DEFAULT_OLLAMA_BASE_URL` so the BYOK Ollama branch shares the same fallback.

Fallback behaviour

Unchanged: if a user has no BYOK row, `resolveByokForUser` returns `null` and `getChatModel` falls through to the platform env path (`OPENAI_BASE_URL`, `OPENAI_CHAT_MODEL`, `OPENAI_STRUCTURED_MODEL`, `ANTHROPIC_API_KEY`, etc.). Aura debits normally on that path.

Migration required

```bash
pnpm db:migrate
```

Applies `20260728084341_modelkey_proxy_fields` — two `ALTER TABLE` statements adding nullable columns. No data loss risk; existing rows keep working with `baseUrl = null, model = null` (falls back to defaults).

Test plan

• `pnpm typecheck` — passes locally
• Store a BYOK row per each of the four provider slots via the UI; confirm chat routes through it and no aura debit occurs
• Verify OpenAI Proxy row refuses to save without a baseUrl (client + server validation)
• Verify pinned model id is used verbatim by the proxy call
• With no BYOK stored, confirm the platform env path still works and aura debits
• Ollama BYOK: pin `gpt-oss:120b-cloud` and a self-hosted baseUrl, confirm chat streams

🤖 Generated with Claude Code