Recipes

Short patterns that use tools already on this list. Each was run on a Linux box with local models only; commands and output are in evidence/recipes.md.

Trace any CLI harness through a gateway. Point Claude Code (through ANTHROPIC_BASE_URL), OpenCode, Pi, or another CLI that accepts a custom Anthropic- or OpenAI-compatible base URL at a LiteLLM proxy with the otel callback writing into an OpenTelemetry Collector. You get spans with per-request input and output token counts, model, latency, and user agent, without harness-specific instrumentation. Claude Code’s own monitoring can send to the same collector, but on local models it may report a phantom cost and only session-level metrics.

Run MCP Conformance through a proxy. Point MCP Conformance at the proxy or gateway in front of the server, not only at the server itself, so buffering and protocol bugs in the middle layer show up as failures that the direct run does not have.

Run a harness eval without Docker. Harbor and Inspect SWE expect container sandboxes. Inspect can also run a task with no sandbox, or with its built-in local sandbox (the host file system, no isolation), so a custom solver can shell out to each CLI on the host and score with hidden tests and model_graded_qa; the agent then runs with your user’s permissions.

Promptfoo Node version. Promptfoo 0.122 and later require Node 22.22 or newer (engines.node). Node 22.20 fails; use a newer Node if npx promptfoo errors on startup.