@ai-sdk/openai-compatible provider fails to connect to LAN addresses — FailedToOpenSocket errno:0 #28418

OpenCode 1.15.5 (Homebrew, macOS arm64) cannot connect to an @ai-sdk/openai-compatible provider on a LAN address. The embedded Bun runtime's fetch() fails with FailedToOpenSocket / errno: 0 for any non-localhost address. This affects HTTP and HTTPS equally.

Configuration:

"llama.cpp": {
  "npm": "@ai-sdk/openai-compatible",
  "name": "llama-server (remote)",
  "options": {
    "baseURL": "http://192.168.4.9:8080/v1",
    "apiKey": "no-key-needed"
  },
  "models": {
    "qwen3.6-27b:q8_0": {
      "name": "Qwen3.6-27B Q8_0 (local)",
      "limit": { "context": 262144, "output": 131072 }
    }
  }
}

Error (from log):

ERROR service=llm providerID=llama.cpp modelID=qwen3.6-27b:q8_0 error={"error":{"name":"AI_APICallError","cause":{"code":"FailedToOpenSocket","path":"http://192.168.4.9:8080/v1/chat/completions","errno":0}}}

Verified working from the same machine:

• curl http://192.168.4.9:8080/v1/models — 200 OK
• Node.js fetch() — 200 OK
• System Bun 1.3.8 fetch() — 200 OK, including POST to /v1/chat/completions with streaming

The server responds correctly to every client except OpenCode's embedded runtime.

Also tested:

• Hostname instead of IP — same failure
• HTTPS via reverse proxy — same failure
• NODE_TLS_REJECT_UNAUTHORIZED=0 — OpenCode logged the TLS warning (confirming it read the env var) but still failed with FailedToOpenSocket
• apiKey in provider options vs. auth.json — no difference

Prior reports of the same bug:

• Remote Ollama stopped working #2667 (Sep 2025) — remote Ollama, same error, confirmed localhost works
• Unable to connect local Ollama #10961 (Jan 2026) — remote Ollama on LAN IP, same error

The defect appears to be in how the Bun runtime is embedded in the OpenCode binary. Standalone Bun connects to LAN addresses without issue. Providers using their own HTTP stacks (e.g., @ai-sdk/amazon-bedrock via AWS SDK v3) are unaffected — only @ai-sdk/openai-compatible, which relies on the runtime's native fetch(), is broken.

This blocks use of OpenCode with any remote inference server on a LAN, which is a common deployment for GPU-equipped build machines separate from the development workstation.