fix(providers): guide Coding Plan keys to the Anthropic endpoint their quota lives on #1069

Summary

One root problem for a GLM Coding Plan subscriber on first run, reproduced on a fresh raven web today (macOS arm64, main 43159933): the plan's quota lives on the Anthropic-compatible endpoint, and nothing tells the user that.

The default model resolves to bigmodel/glm-5.3-flash, which LiteLLM sends to the zai/bigmodel pay-as-you-go endpoints. The key is the same account in both cases; what differs is the billing track inside it: a Coding Plan's quota is consumed only through the Anthropic-compatible endpoint (https://open.bigmodel.cn/api/anthropic), while the chat-completions endpoints check the pay-as-you-go balance, which is zero. So the same key answers ZaiException - 余额不足或无可用资源包,请充值 / Insufficient balance or no resource package. Please recharge. on one track while its other track is full -- and nothing in onboarding or the docs distinguishes a plan key from pay-as-you-go credit, so the message reads as "your money is gone."

The dead end this misdirection leads to: following the balance message, the natural next move is to edit the bigmodel provider entry's baseURL to /api/anthropic. A provider's protocol does not follow its URL -- the entry still speaks OpenAI (chat/completions through LiteLLM), the Anthropic endpoint closes the stream immediately, and every retry logs the stream ended without the upstream's terminal chunk after 0s, surfacing as Error calling LLM (network): the model's reply was cut off by the connection. An anthropic-type provider entry with the same URL and key works immediately, which is the actual fix -- but a deterministic mismatch reported as a transport failure sends the user debugging sockets instead of finding it.

Steps to reproduce

1. raven web with a GLM Coding Plan key configured and the default model (bigmodel/glm-5.3-flash); send 你好 in a new conversation.
2. Observe the balance refusal from the pay-as-you-go endpoint (~/.raven/logs/gateway.log):

   09:59:20  Iteration 1/40 model=bigmodel/glm-5.3-flash
   09:59:25  Stream LLM error [rate_limit]: ZaiException - 余额不足或无可用资源包,请充值。
   

3. In Settings, edit the bigmodel provider's baseURL to https://open.bigmodel.cn/api/anthropic and send 你好 again:

   10:17:08  Iteration 1/40 model=bigmodel/glm-5.3-flash
   10:17:10  the stream ended without the upstream's terminal chunk after 2s (0 chars ...)
   10:17:56  Error calling LLM (network): the model's reply was cut off by the connection
   

4. Create an anthropic-type provider with that same URL and key: the turn completes.

Expected behavior

• The onboarding wizard and the providers docs say which endpoint each credential kind works with: a Coding Plan key belongs on an Anthropic-protocol provider pointed at /api/anthropic; pay-as-you-go keys belong on the bigmodel/zai entries.
• Along the way, so the dead end stops hiding the fix: a provider whose protocol cannot parse the endpoint's answer reports a protocol mismatch rather than a connection cut, or the settings probe catches it before a turn is spent on it.

Environment

macOS arm64, Raven main at 43159933, default model glm-5.3-flash, LiteLLM routing to provider = zai; log lines quoted from ~/.raven/logs/gateway.log on 2026-10-23.