Expand description
eval-magic — a CLI for running skill evals: it measures whether an agent
skill actually shifts behavior.
The crate is organized into seven submodules, ordered roughly by dependency:
core (domain types) underpins everything; validation, adapters,
sandbox, pipeline, and workspace each own one concern; cli dispatches
the subcommands and orchestrates run.
Modules§
- adapters
- The harness adapter layer.
- cli
- CLI surface: command-tree definition and dispatch.
- core
- Shared kernel used by nearly every other module.
- pipeline
- The post-dispatch processing chain: stateless JSON-in/JSON-out stages.
- sandbox
- Execution sandbox: shared write-guard machinery and write-boundary policy.
- validation
- JSON-Schema validation of
evals.jsonand pipeline artifacts. - workspace
- Baseline management and workspace cleanup.