Go side: eval-server HTTP endpoint, index caching keyed by (repo, commit),
and integration with the existing MCP server. Python side: full eval harness
with single/matrix/debug modes, Docker-based environments, agent wrapper
with augmentation support, composable YAML configs for models and modes
(baseline/native/native_augment), bridge scripts for CLI tool access,
and property-based tests. Tracks resolve rate, cost, tool call counts,
and augmentation hit rate.
Andrey Kumanyaev
fd89126884