Module config 

Configuration: which provider runs skills, the default platforms and models a run fans out across, and the model used for natural-language evals.

Config is loaded from a YAML file (default skilltest.yaml) and then refined by CLI overrides (see Config::apply_overrides).

Structs

ApiJudgeConfig

Settings for judging evals and the simulated user with a direct model API call instead of running them through a harness. This trades the harness’s auth-portability for a single fast HTTP round trip per judge call (no agent-loop cold start), with normalized token usage surfaced into the report.

CommandConfig

Settings for a custom provider command speaking the JSON-lines protocol (see docs/protocol.md). Used by the bundled skilltest-fake-provider and any provider you write yourself.

Config

The full configuration for a run.

OneharnessConfig

Settings for the default oneharness provider, which runs each prompt on a harness via oneharness run.

Overrides

CLI-supplied overrides. None/empty fields leave the config value in place.

Enums

ApiVendor

Which model vendor’s API the direct-API judge talks to.

JudgeConfig

How evals and the simulated user are judged, independent of the provider that runs the skill. Absent (the default) means the run’s provider judges too (e.g. the oneharness judge_harness).

ProviderConfig

Which provider backs a run.

Constants

ONEHARNESS_BIN_ENV

Environment variable the SDKs set to the oneharness binary they bundle (the oneharness-cli PyPI/npm package installed alongside the SDK), so a run resolves oneharness with no PATH setup. It is only a default: an explicit config provider.bin (or the --oneharness-bin flag) still wins.