Don't trust me bro: fixing GPT-OSS (3.49B tokens, 1k GPU hours, 1x3090)

burrito is a drop-in replacement for the OpenAI and Anthropic APIs, backed by a gpt-oss model. It accepts standard requests on /v1/chat/completions, /v1/responses, and /v1/messages (JSON and streamed) and delegates next-token prediction to an existing inference backend — llama.cpp or vLLM — over /v1/completions. It does not ship a model.

gpt-oss was trained with native python and browser tools and to work in distinct "channels" — interleaving reasoning, tool calls, and backtracking into reasoning before a final answer. Existing inference engines handle this pattern with compromises: llama.cpp forces tool calls through grammars and a hardcoded functions. prefix, while vLLM lacks /v1/messages for Anthropic-compatible clients and has no recovery from malformed output. burrito fills the other side of that specialization: it is a model-specific harness that renders conversations the way the model was trained to read them, executes the native python and browser tools inside the same process, recovers from hallucinated tool calls, and exposes production health checks and Prometheus metrics.

The inference pipeline tokenizes and renders requests with the harmony encoding, streams from the backend, and guards against repetition loops and malformed tool calls. Browser tools run on a bundled Playwright engine and SearXNG instance. Python tools run in-process Jupyter kernels or in isolated, pre-warmed burrito-kernel containers. Deployment uses a single docker-compose.yml that ships the app with SearXNG, Valkey, Prometheus, and Grafana. Evaluations were conducted with burrito-evals, running gpt-oss-20b across 320,192 runs — 3.49B tokens, 1,062 GPU hours on a single RTX 3090.