providers: custom openai-compatible providers hang or send to undefined/chat/completions #36834

Summary

V2 next (opencode2 v0.0.0-next-15363) fails to start a chat completion for any
custom openai-compatible provider that the equivalent V1 config can use. The
session hangs with no error and only emits after timeout, and the documented
aisdk:@ai-sdk/openai-compatible package path is the only working shape — but
settings.baseURL is silently ignored when the V1-style options.baseURL is
used, producing a request to undefined/chat/completions. A V1 provider block
that worked for the same OpenAI-compatible gateway is rejected by V2.

Environment

• opencode version: opencode2 v0.0.0-next-15363
• OS: Linux 7.0.10-arch1-1 (linux x64)
• Terminal: tmux (TERM=tmux-256color, COLORTERM=truecolor)
• Shell: /usr/bin/bash
• Install/channel: next
• Active plugins: ~/.config/opencode/plugins/gk-hooks.js

Reproduction

1. Use this V1 config that works on V1 with an OpenAI-compatible gateway
   exposing a custom acme/acme-m3 model:

   {
     "$schema": "https://opencode.ai/config.json",
     "provider": {
       "acme": {
         "npm": "@ai-sdk/openai-compatible",
         "name": "Acme Gateway",
         "options": { "baseURL": "https://gateway.example.com/v1" },
         "models": {
           "acme-m3": { "name": "Acme M3" }
         }
       }
     }
   }

2. Translate to the V2 shape per
   https://v2.opencode.ai/providers.md#provider-packages:

   {
     "$schema": "https://opencode.ai/config.json",
     "providers": {
       "acme": {
         "name": "Acme Gateway",
         "env": ["ACME_API_KEY"],
         "package": "aisdk:@ai-sdk/openai-compatible",
         "settings": { "baseURL": "https://gateway.example.com/v1" },
         "models": {
           "acme-m3": {
             "name": "Acme M3",
             "capabilities": { "tools": true, "input": ["text"], "output": ["text"] },
             "limit": { "context": 200000, "output": 32000 }
           }
         }
       }
     }
   }

3. Run any request:

   opencode2 api post /api/generate \
     --data '{"model":{"providerID":"acme","id":"acme-m3"},"prompt":"ping"}'

Expected Behavior

• Either the V1 provider / options shape is accepted and translated, or
  V2 returns a clear migration error.
• The V2 config above sends the request to
  https://gateway.example.com/v1/chat/completions and returns a normal
  model response (the endpoint is reachable, GET /v1/models returns the
  model catalog, and curl with the same Bearer token works).

Actual Behavior

• First error reported by opencode2 api post /api/generate:

  Unsupported package for acme/acme-m3: @opencode-ai/llm/providers/openai-compatible
  

  The V2 docs say the native package is the recommended path, but the runtime
  refuses it. Switching to aisdk:@ai-sdk/openai-compatible is required but
  undocumented as the only working choice.

• After switching to aisdk:@ai-sdk/openai-compatible, the request hangs and
  the TUI never produces a response. A test against the same V2 build using
  the openai-compatible V2 native package fails with a parse error:

  "undefined/chat/completions" cannot be parsed as a URL.
  

  settings.baseURL is being dropped, so the package sends to undefined.

• Disabling reasoning via a none variant
  (settings: { reasoningEffort: "none" }) makes it worse — the upstream
  gateway rejects both reasoning_effort and thinking with
  UnsupportedParamsError: openai does not support parameters: ['reasoning_effort', 'thinking'], because OpenCode V2 still sends them
  even when the provider/model doesn't declare reasoning support.

Additional Context

• The same gateway endpoint responds fine to a direct curl against
  https://gateway.example.com/v1/chat/completions with the API key.
• V1 (opencode-ai v0.x) ran this provider block end-to-end without hangs.
• Workarounds attempted in V2:
  • package: "@opencode-ai/llm/providers/openai-compatible" (native) →
    rejected as "Unsupported package".
  • package: "aisdk:@ai-sdk/openai-compatible" + settings.baseURL →
    hangs / sends to undefined depending on which field is present.
  • options.baseURL (V1 spelling) → ignored entirely.
  • Adding a none variant with reasoningEffort: none → upstream rejects
    the unsupported parameters and returns 400.
• Likely root causes: (1) the V2 native openai-compatible package
  (@opencode-ai/llm/providers/openai-compatible) is documented but
  unregistered / unsupported at runtime; (2) settings.baseURL is not
  forwarded to the aisdk:@ai-sdk/openai-compatible provider, and the
  error surface for missing config is a silent hang or a confusing
  "undefined/chat/completions" parse error instead of a clear validation
  message; (3) reasoning-related fields are not stripped or detected
  per-provider.
• A real opencode.json (with the key redacted) is sufficient to reproduce
  using any OpenAI-compatible gateway and a custom model name.