Opencode Free models

[woss]

Hello,

I've been trying to set up headroom for opencode with free models like big-pickle and deepseek-v4-flash-free but it doesn't work, i get error in the logs 'unsupported provider litellm'.

I've also seen that there are lot of gh issues regarding the opencode plugin. Did anyone got this thing working?

My reasoning is: If i'm going to use free models for learning, why not save on tokens as well, fair is fair.

Thanks

[173787247]

unsupported provider litellm means OpenCode is not sending an Anthropic/OpenAI-shaped request

headroom wrap opencode injects config and starts the proxy. Free catalog models (big-pickle, deepseek-v4-flash-free, etc.) often go through OpenCode’s LiteLLM provider, whose request body is not what Headroom’s proxy parses — hence skip/error instead of compression.

Working patterns people actually get

1. Don’t use the LiteLLM provider for the Headroom hop. Point OpenCode at Headroom’s OpenAI-compatible /v1 (or Anthropic /v1/messages if the model is Claude-shaped), then let Headroom talk to the real upstream.
2. If the free model only exists behind LiteLLM, Headroom currently cannot wrap that provider. Use a model OpenCode can call as openai / anthropic compatible, or run Headroom in library/compress-only on tool output rather than as a full proxy.
3. Confirm wrap actually rewrote OpenCode config: after headroom wrap opencode, check that the provider base URL is http://127.0.0.1:8787 (or whatever port you used), not still litellm.

Minimal isolate

headroom proxy --port 8787 --log-file /tmp/headroom.jsonl
# point one OpenAI-compatible client at http://127.0.0.1:8787/v1

If that client works and OpenCode-via-LiteLLM does not, it’s the provider adapter, not “free models are banned”.

Free models + token saving is a reasonable goal; the blocker is LiteLLM request shape, not the model being free. Track the OpenCode plugin issues, but this error is expected until Headroom adds a LiteLLM translator.