Provider-level npm field silently ignored unless models is also configured #33888

Description

Summary

When configuring a custom provider with a different npm package (e.g. switching ZAI/ZhipuAI from the default @ai-sdk/openai-compatible to @ai-sdk/anthropic), setting only provider.<id>.npm and provider.<id>.options.baseURL has no effect — opencode silently keeps using the catalog/discovery-provided npm package (e.g. @ai-sdk/openai-compatible) and sends requests in the wrong protocol. There is no warning, error, or log line indicating that provider.npm was ignored.

The user must additionally enumerate every model under provider.<id>.models: { ... } for provider.npm to take effect on those entries. Provider-level npm is only read inside the loop over provider.models, so models pre-populated by the discovery plugin keep their original npm package.

Reproduction

Configure ZAI to use Anthropic's /v1/messages endpoint:

// opencode.jsonc
{
  "provider": {
    "zai": {
      "npm": "@ai-sdk/anthropic",
      "options": {
        "baseURL": "https://open.bigmodel.cn/api/anthropic/v1"
      }
    }
  }
}

Run opencode run -m zai/glm-4.5 "hi".

Expected

provider.npm applies provider-wide; all zai models (including those from the discovery plugin) use @ai-sdk/anthropic and hit /v1/messages.

Actual

• The npm field is silently ignored.
• Requests go to https://open.bigmodel.cn/api/anthropic/v1/chat/completions (OpenAI path, because @ai-sdk/openai-compatible is still in use).
• For this URL ZAI's gateway returns either:
  • HTTP 200 with body {"code":500,"msg":"404 NOT_FOUND","success":false} (business-layer 500 that the SSE parser silently swallows), or
  • HTTP 404 {"detail":"Not Found"} (when /v1 is in the path) — which surfaces as a stream error.
• No warning is emitted that provider.npm had no effect.

Workaround

Enumerate every model:

"models": {
  "glm-4.5": {}, "glm-4.6": {}, "glm-5": {}, "glm-5.2": {}
}

Cumbersome when the provider exposes many models.

Root cause

packages/opencode/src/provider/provider.ts:1394-1414:

for (const [providerID, provider] of configProviders) {
  const existing = database[providerID]
  const parsed: Info = {
    // ...
    models: existing?.models ?? {},          // catalog/discovery models, with their own npm
  }

  for (const [modelID, model] of Object.entries(provider.models ?? {})) {  // ← only iterates user-listed models
    const apiNpm =
      model.provider?.npm ??
      provider.npm ??                          // ← only consulted inside this loop
      existingModel?.api.npm ??
      modelsDev[providerID]?.npm ??
      "@ai-sdk/openai-compatible"
    // ...
  }
}

Proposal

Either (or both):

1. Make provider.npm provider-wide. When set, override api.npm (and optionally api.url from provider.api) for all existing.models entries that don't have an explicit per-model override.
2. Warn when provider.npm is set but provider.models is empty (the user almost certainly intended provider-wide application and will get silently wrong behavior).

Environment

• opencode 1.17.10
• platform: linux
• relevant source: packages/opencode/src/provider/provider.ts

Plugins

No response

OpenCode version

No response

Steps to reproduce

No response

Screenshot and/or share link

No response

Operating System

No response

Terminal

No response