Description

applyCaching() in transform.ts sets cache hints for Anthropic, Bedrock, OpenRouter, Copilot, and OpenAI-compatible providers, but skips Google Vertex AI and Google AI entirely. This means Gemini models via @ai-sdk/google-vertex or @ai-sdk/google never get the AI SDK's cachePoint annotation, so implicit context caching can't kick in for the system prompt prefix.

The SDK already supports cachePoint for these providers — it just needs to be set.

Impact: ~5-15K tokens of system prompt + tool declarations re-billed as new input on every request. Gemini charges 75% less for cached tokens, so this is meaningful for heavy users.

Reproduction

1. Configure opencode with a Google Vertex AI or Google AI provider
2. Send a message that includes the system prompt
3. Observe that no cachePoint annotation is set on the system message
4. Compare with Anthropic/Bedrock providers which DO get cache annotations

Expected Behavior

Google Vertex AI and Google AI providers should receive google: { cachePoint: { type: "default" } } annotations on the system prompt, matching the pattern used by Anthropic (cacheControl) and Bedrock (cachePoint).

Environment

• opencode version: latest main
• Provider: @ai-sdk/google-vertex and @ai-sdk/google
• File: packages/opencode/src/provider/transform.ts