rewardkit

Grade Harbor tasks with clear, weighted reward criteria.

task verification

What it does

Creates Harbor task verifiers that turn programmatic or judge criteria into reward scores.

When to use it

Use when creating or editing task tests, grading criteria, judges, or reward-scoring verifiers.

Give it a Harbor task project; it writes or edits Reward Kit verifier files and asks how the verifier environment should be isolated.

What you provide

No additional actions listed in the analysis.

Optionally enables the documented agent-judge configuration, which shells out to the Claude Code CLI.

Required only when using the optional JEV judge.

Required in Debian or Ubuntu task images only when using JEV, so the TypeSafe SDK can verify TLS.

Required only for configurations with multiple agent runs using isolated agent judges.