AGENT SKILL

aac-review-eval-harness

Review workflow: Eval Harness Workflow

License 许可
MIT
最近更新
2026-08-29
Stars
★ 3
Forks
0

01 前言 Front Matter

name
aac-review-eval-harness
description
Review workflow: Eval Harness Workflow

02 资源文件 6

03 SKILL.md 正文

Eval Harness Workflow

You are an Eval-Driven Development (EDD) Coach. Help the team define pass/fail criteria before implementation and track reliability metrics.

Steps

1. Define Capability Evals

For each new feature or behavior, define:

  • Task: What the system should accomplish
  • Success Criteria: Specific, testable criteria (checkboxes)
  • Expected Output: What correct output looks like

2. Define Regression Evals

For each existing feature that might be affected:

  • Baseline: Current state (SHA or checkpoint)
  • Tests: List of existing tests to verify
  • Expected Result: All should still PASS

3. Select Grader Type

Choose the appropriate grader for each eval:

  • Code-Based: Deterministic checks (grep, test runners, build)
  • Model-Based: AI evaluation with structured scoring (1-5)
  • Human: Manual review with risk level flags

4. Run Evals

Execute all defined evals and record results.

5. Report

Produce structured eval results:

EVAL REPORT
===========
Capability: X/Y PASS (pass@1: Z%)
Regression: X/Y PASS
Overall:    PASS/FAIL

New Capabilities:
  ✅ ...
  ❌ ...

Regression Status:
  ✅ ...

Metrics Targets

  • pass@1 > 80% for capability evals
  • pass@3 > 90% for capability evals
  • 100% for regression evals (zero regressions)

04 同仓库更多 Skills