OpenCode Go Qwen Cache Fix

Pi extension that fixes prompt caching for Qwen models served through OpenCode Go.

What it does

OpenCode Go's Qwen backend expects Anthropic-style cache_control markers, even though the endpoint is OpenAI Chat Completions compatible. Pi's built-in OpenCode Go Qwen model definitions did not emit those markers, causing cache misses.

This extension re-registers the built-in opencode-go provider and patches only Qwen models with:

compat: {
  cacheControlFormat: "anthropic"
}

It also adds debug/status commands:

• /opencode-cache-status — show how many OpenCode Go Qwen models are patched
• /opencode-cache-debug — inspect the next OpenCode Go Qwen provider payload for cache_control markers

Install

pi install git:github.com/TreptowerPark/opencode-go-qwen-cache-fix

Or test without installing:

pi -e git:github.com/TreptowerPark/opencode-go-qwen-cache-fix

Then reload pi or run:

/reload

Requirements

You need OpenCode Go credentials configured in pi:

export OPENCODE_API_KEY="..."

Or via /login / ~/.pi/agent/auth.json with provider key opencode-go.

Verify

In pi, select an OpenCode Go Qwen model such as opencode-go/qwen3.6-plus, then run:

/opencode-cache-status
/opencode-cache-debug

The next provider request should report one or more cache_control markers.

Notes

• Only patches provider opencode-go
• Only patches models whose id contains qwen
• Does not modify non-Qwen OpenCode Go models
• Does not store or read API keys

License

MIT