opencode-svip-proxy

An OpenCode plugin that injects metadata.user_id into every model request body (/v1/messages and /v1/chat/completions).

A relay that routes by metadata.user_id can then pin all of a session's requests to one upstream account, keeping the prompt-cache hit rate high — so you get the caching benefit without running a separate request-rewriting proxy in front of OpenCode.

It wraps globalThis.fetch inside the plugin server() hook (the same mechanism @lucentia/opencode-trace uses), so it composes with other plugins.

Install

Add it to the plugin array in ~/.config/opencode/opencode.json:

{
  "plugin": ["opencode-svip-proxy"],
  "provider": {
    "svip": {
      "npm": "@ai-sdk/anthropic",
      "options": {
        "baseURL": "https://svip.xty.app/v1",
        "apiKey": "{env:LLM_API_KEY}"
      },
      "models": { "claude-opus-4-8": {} }
    }
  }
}

OpenCode fetches and caches the plugin from npm. Because the metadata is added in-process, the provider baseURL points straight at the remote relay — no local proxy needed.

Configure

export OPENCODE_METADATA_USER_ID=your-stable-user-or-session-id

Falls back to a built-in constant if unset.

Behaviour

Injection happens only when both hold (otherwise the request is passed through untouched — other providers/models see no change):

• the request host is the svip relay (OPENCODE_SVIP_HOST, default svip.xty.app) — i.e. it came through the svip provider; and
• the target model is a Claude or DeepSeek model (/claude|deepseek/i) — both are cache-routed by user_id on svip (claude via /v1/messages, deepseek via /v1/chat/completions).

Additionally:

• only POSTs to …/messages or …/chat/completions are touched;
• an existing metadata object is merged, not overwritten;
• all other request fields are preserved.

Run the tests: npm test.

License

MIT