harness-eval-contract

Defines capability, regression, command, and Judge evaluation criteria for Agent OS specs.

Updated Jul 29, 2026
One-click install
npx skills add https://github.com/MaiconGambini/opencode-harness-guide --skill harness-eval-contract-maicongambini
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: harness-eval-contract
Source: https://github.com/MaiconGambini/opencode-harness-guide/tree/main/skills/harness-eval-contract
Command: npx skills add https://github.com/MaiconGambini/opencode-harness-guide --skill harness-eval-contract-maicongambini

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve? Medium and high risk implementation work often starts without explicit acceptance criteria, causing specs to drift from quality gates and leaving no shared definition of what passing means. This Skill enforces an eval contract before any implementation begins. ## Core Features & Use Cases - Capability and Regression Evals: Define what must become possible and what must remain true, with each bug fix requiring a regression test. - Command and Judge Evals: Specify exact checks with expected outputs, plus a rubric and acceptance threshold for Judge evaluation. - Single Source of Thresholds: All numeric expectations cite agent-os/quality-thresholds.json instead of restating values, preventing drift between documents. - Use Case: Before implementing a medium-risk feature in an Agent OS project, add an evals.md section to the spec that records capability goals, regression coverage, command checks, Judge rubric, and where evidence will be stored. ## Quick Start Ask the agent to add an eval contract with capability, regression, command, and Judge evals to the current spec before starting implementation.

Frequently Asked Questions about harness-eval-contract

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I define acceptance criteria before implementing a feature spec?

Add an evals.md section to the spec covering capability evals, regression evals, command evals with expected outputs, and Judge evals with a rubric and threshold. Implementation should not start until this eval contract is explicit.

What should an eval contract for a software spec include?

It should include capability evals describing what must become possible, regression evals describing what must remain true with a covering test, command evals with exact checks and expected outputs, Judge evals with rubric and threshold, and an evidence location.

Why should eval thresholds reference a single quality-thresholds file?

Citing agent-os/quality-thresholds.json as the single source prevents specs from hardcoding values that drift from the quality gate. When two documents restate thresholds independently, they eventually disagree about what passing means.

When is an eval contract required for a spec?

An eval contract is required for medium or high risk implementation work. Low risk changes may not need the full contract, but any bug fix must add a regression eval with a corresponding regression test.

What happens if a bug fix has no regression test?

The contract requires every bug fix to add a regression eval covered by a regression test. Without it, the fixed behavior has no guard against reintroduction and the spec's regression section remains incomplete.