Custom provider models (via `@ai-sdk/openai-compatible`) fail in sub-agent context

Description

Custom provider models registered in config.json via the npm + options pattern do not work when assigned to a sub-agent (mode: subagent). The sub-agent returns a bare Error with no details — no crash log, no stack trace, no console output.

The same sub-agent works correctly when using the built-in opencode/deepseek-v4-flash-free model.

The custom provider itself works fine:

• The OpenAI-compatible API endpoint responds correctly
• The @ai-sdk/openai-compatible module resolves and creates a working provider from the config.json settings
• Direct API calls to the local model succeed

Steps to Reproduce

1. Create a custom provider in ~/.config/opencode/config.json:

   {
     "provider": {
       "Lemonade": {
         "name": "Lemonade (Local LLM)",
         "npm": "@ai-sdk/openai-compatible",
         "options": {
           "apiKey": "lemonade",
           "baseURL": "http://127.0.0.1:13305/v1"
         },
         "models": {
           "Qwen3.6-35B-A3B-MTP-GGUF": {
             "name": "Qwen3.6 35B (GPU)"
           }
         }
       }
     }
   }
   

2. Install the npm package:

   cd ~/.config/opencode
   bun install @ai-sdk/openai-compatible
   

3. Create a sub-agent with the custom model:

   ---
   description: Test agent with custom provider
   mode: subagent
   model: Lemonade:Qwen3.6-35B-A3B-MTP-GGUF
   temperature: 0.3
   ---
   You are a test agent.
   

4. Invoke the sub-agent via the task tool with subagent_type: "subagents/<name>".

5. Observe Error with no detail.

Expected Behavior

The sub-agent should use the custom provider model, or at minimum return a descriptive error (e.g., "Provider not found", "Model not available", "Failed to load provider package").

Actual Behavior

Bare Error with no details. No log entries in ~/.local/share/opencode/log/. The opencode providers list CLI command also does not list custom providers from config.json — it only shows credential-based providers (Nvidia API, DeepSeek API).

Diagnostic Checklist

Check	Result
Provider config.json syntax	✅ Valid JSON
@ai-sdk/openai-compatible installed	✅ Installed via bun in ~/.config/opencode/node_modules/
Module resolves via bun	✅ require.resolve('@ai-sdk/openai-compatible') works
Provider creates successfully	✅ createOpenAICompatible() returns a valid provider
Model API endpoint responds	✅ POST /v1/chat/completions returns 200 with valid response
Same sub-agent with default model	✅ Works with opencode/deepseek-v4-flash-free
Model name in Lemonade API	✅ Reported as Qwen3.6-35B-A3B-MTP-GGUF in /v1/models
Sub-agent tools section removed	✅ (Tools access confirmed working with default model)
opencode providers list	❌ Does NOT show Lemonade provider (only shows credential providers)
Error logs	❌ No relevant entries found

Additional Notes

• The issue is specifically with sub-agent model resolution for custom providers. The built-in providers work fine.
• Restarting OpenCode after config changes does not resolve the issue.
• The npm auto-install mechanism (referenced by the "npm" field in config) may not be triggering, or the module resolve path does not include the config directory's node_modules.
• This was tested with a local OpenAI-compatible server (Lemonade Server v10.9.0) and also may affect any custom provider using @ai-sdk/openai-compatible.

Related Files

• ~/.config/opencode/config.json — provider configuration
• ~/.config/opencode/package.json — dependencies (includes @ai-sdk/openai-compatible)
• ~/.config/opencode/bun.lock — lockfile after bun install
• ~/.config/opencode/agents/subagents/<name>.md — sub-agent definition

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