DigitalOcean Anthropic models (Claude Fable, etc.) never hit prompt cache in v2 #51557

Description

In OpenCode v2, DigitalOcean models (cataloged with "npm": "@ai-sdk/openai-compatible") are routed through @opencode/ai/providers/openai-compatible.

Because openai-compatible-chat is not in RESPECTS_INLINE_HINTS in packages/ai/src/cache-policy.ts, and OpenAIChat.protocol does not pass cacheControl by default, requests are sent to DigitalOcean's /v1/chat/completions with plain text messages and no cache_control breakpoints.

While open-source models on DigitalOcean (like DeepSeek V4.1 Flash) perform automatic KV-cache prefix matching on raw prompts, Anthropic models served through DigitalOcean (Claude Fable 5.1, Claude Fable 5, Sonnet, Opus, Haiku) strictly require explicit cache_control: { type: "ephemeral" } markers to activate prompt caching.

Without cache_control, every single turn and tool step in an agent session resends the entire conversation transcript without cache markers, meaning 100% of input tokens are invoiced at the full rate ($10.00 / 1M on Fable instead of $0.25 / 1M cache read rate).

In OpenCode v1, applyCaching() in packages/opencode/src/provider/transform.ts checked if the model ID contained "anthropic" or "claude" and attached openaiCompatible.cache_control, which worked.

OpenCode version

v2

Steps to reproduce

1. Configure DigitalOcean provider with an API access key.
2. Select anthropic-claude-fable-5.1 or another Anthropic model on DigitalOcean.
3. Run a session with several tool turns.
4. Observe token usage on DigitalOcean or in response metadata: cache_read_input_tokens is 0 on all steps and all input tokens are billed at full price.