Bug: Anthropic prompt caching not applied when using custom provider name with Anthropic protocol via proxy

Description

When using Claude models through a third-party proxy (e.g. OfoxAI) via the Anthropic native protocol, prompt caching never activates — cache hit rate stays at 0%, even with rapid consecutive messages.

The root cause appears to be related to PR #1305: Anthropic handles content-level caching "natively", but that native caching logic seems to only apply when the provider is the built-in anthropic provider. A custom provider (e.g. ofox_ai) using npm: "@ai-sdk/anthropic" does not receive automatic cache_control marks.

Environment

• opencode version: latest (1.x, upgraded on 2026-08-28)
• OS: Windows 11
• Proxy: OfoxAI (https://api.ofox.ai)
• Protocol: Anthropic native (baseURL: https://api.ofox.ai/anthropic)
• Model: anthropic/claude-sonnet-4.5

Steps to reproduce

1. Configure a custom provider with "npm": "@ai-sdk/anthropic" pointing at an Anthropic-protocol proxy:

{
  "provider": {
    "ofox_ai": {
      "npm": "@ai-sdk/anthropic",
      "options": { "baseURL": "https://api.ofox.ai/anthropic" },
      "models": {
        "anthropic/claude-sonnet-4.5": { "limit": { "context": 1000000 } }
      }
    }
  }
}

2. Put a long static document (e.g. 20k+ chars) in instructions so it is injected as system prompt every turn.
3. Chat for several turns within a few minutes (well under the 5-minute ephemeral TTL).
4. Check proxy-side usage stats.

Expected behavior

Anthropic's prompt caching should kick in: first request creates the cache (cache_creation_input_tokens), subsequent requests hit it (cache_read_input_tokens > 0).

Actual behavior

Cache creation and cache read stay at 0 forever.

Diagnostic evidence (we did the homework)

1. Proxy is fine: calling the proxy directly with anthropic SDK and explicit cache_control: {type: "ephemeral"} on the system block works perfectly — 1st call cache_creation_input_tokens: 1081, 2nd call 3 seconds later cache_read_input_tokens: 1081 (100% hit).
2. Not a TTL issue: rapid-fire messages within 5 minutes still produce 0 hits via opencode.
3. setCacheKey: true (the OpenAI-protocol fix from issue #17610) does not help for the Anthropic protocol.
4. Built-in provider name test: renaming the provider to the built-in anthropic (with custom baseURL) was also tried — still 0 hits in our environment.

Hypothesis

Per PR #1305: "Anthropic handles content-level caching natively, others need explicit control". The native Anthropic caching path may be tied to the built-in anthropic provider's internal adapter, and custom providers (even with @ai-sdk/anthropic) may not get cache marks applied — or the marks are applied in a way proxies don't forward.

Would be great if custom Anthropic-protocol providers could also get native caching, or if there were a documented way to enable it.

Thanks! 🐙

Reported by @utadamaaya (Utada & DeepSeek)

PS：This issue was written by my Deepseek, I was just helping to submit it.

Plugins

No response

OpenCode version

v1.18.23

Steps to reproduce

1. Configure a custom provider with "npm": "@ai-sdk/anthropic" pointing at an Anthropic-protocol proxy:

{
  "provider": {
    "ofox_ai": {
      "npm": "@ai-sdk/anthropic",
      "options": { "baseURL": "https://api.ofox.ai/anthropic" },
      "models": {
        "anthropic/claude-sonnet-4.5": { "limit": { "context": 1000000 } }
      }
    }
  }
}

2. Put a long static document (e.g. 20k+ chars) in instructions so it is injected as system prompt every turn.
3. Chat for several turns within a few minutes (well under the 5-minute ephemeral TTL).
4. Check proxy-side usage stats.

Screenshot and/or share link

Operating System

Windows 11

Terminal

OpenCode Desktop