evals-implement

어떤 스킬인가요?

Use when a published goldset is ready to become executable graders and configs — Python graders and metrics with auto-run unit tests verifying grader correctness.

설치 방법

프로젝트 폴더에서 아래 명령어를 실행하고, 설치할 에이전트를 선택하세요.

설치 명령을 직접 실행해야 적용됩니다. 지원 에이전트와 필요한 권한·라이선스는 제작자의 안내를 확인하세요.

스킬 원문 파일

#!/usr/bin/env bash
# setup-evals-implement.sh — Setup for evals-init (self-contained)
set -euo pipefail

resolve_project_root() {
  local dir
  dir="$(pwd)"
  while [[ "$dir" != "/" ]]; do
    if [[ -d "${dir}/.adlc" ]]; then
      echo "$dir"
      return
    fi
    dir="$(dirname "$dir")"
  done
  git rev-parse --show-toplevel 2>/dev/null || pwd
}

resolve_team_ai_directives() {
  local project_root="$1"
  local td="${TEAM_AI_DIRECTIVES:-}"
  [[ -n "$td" ]] && { echo "$td"; return; }
  if [[ -f "${project_root}/.adlc/init-options.json" ]]; then
    td=$(jq -r '.team_ai_directives // empty' "${project_root}/.adlc/init-options.json" 2>/dev/null || true)
    [[ -n "$td" ]] && { echo "$td"; return; }
  fi
  echo "${project_root}/team-ai-directives"
}

resolve_branch() {
  git branch --show-current 2>/dev/null || echo "unknown"
}

PROJECT_ROOT=$(resolve_project_root)
TEAM_AI_DIRECTIVES=$(resolve_team_ai_directives "$PROJECT_ROOT")
BRANCH=$(resolve_branch)

jq -n --arg rr "$PROJECT_ROOT" --arg td "$TEAM_AI_DIRECTIVES" --arg br "$BRANCH" \
  '{REPO_ROOT:$rr,TEAM_AI_DIRECTIVES:$td,BRANCH:$br}'

#!/usr/bin/env pwsh
# setup-evals-implement.ps1 — Setup for evals-init (self-contained)
$ErrorActionPreference = "Stop"

function Resolve-ProjectRoot {
    $dir = $PSScriptRoot
    while ($dir -ne "") {
        if (Test-Path (Join-Path $dir ".adlc")) { return $dir }
        $parent = Split-Path $dir -Parent
        if ($parent -eq $dir) { break }
        $dir = $parent
    }
    $gitRoot = git rev-parse --show-toplevel 2>$null
    if ($gitRoot) { return $gitRoot }
    return (Get-Location).Path
}

function Resolve-TeamAiDirectives {
    param([string]$ProjectRoot)
    $td = $env:TEAM_AI_DIRECTIVES
    if ($td) { return $td }
    $initOpts = Join-Path $ProjectRoot ".adlc/init-options.json"
    if (Test-Path $initOpts) {
        try {
            $config = Get-Content $initOpts -Raw | ConvertFrom-Json
            if ($config.team_ai_directives) { return $config.team_ai_directives }
        } catch {}
    }
    return (Join-Path $ProjectRoot "team-ai-directives")
}

function Resolve-Branch {
    $b = git branch --show-current 2>$null
    if ($b) { return $b }
    return "unknown"
}

$ProjectRoot = Resolve-ProjectRoot
$TeamAiDirectives = Resolve-TeamAiDirectives $ProjectRoot
$Branch = Resolve-Branch

$result = @{
    REPO_ROOT = $ProjectRoot
    TEAM_AI_DIRECTIVES = $TeamAiDirectives
    BRANCH = $Branch
}

$result | ConvertTo-Json -Compress

---
name: evals-implement
description: Use when a published goldset is ready to become executable graders and configs — Python graders and metrics with auto-run unit tests verifying grader correctness.
disable-model-invocation: true
---

# evals-implement

## What this skill does

Generates the **complete executable evaluation implementation** following **EDD Principle VIII** (Close Production Loop) from the published goldset, with automated unit testing to verify evaluator correctness.

**Output**:
1. **Grader/Metric Implementation** - Python evaluators for each goldset criterion with binary pass/fail
   - PromptFoo: Python grader functions with JSON output in `evals/{system}/graders/`
   - DeepEval: Custom metric classes inheriting from `BaseMetric`
2. **Evaluator Unit Tests** - Automated tests (`evals/{system}/tests/test_check_*.py`) that run the goldset pass/fail examples against the generated graders to ensure the evaluator itself is accurate
3. **Evaluation Configuration** - Complete config file (`config.js` or `config.py`) with Tier 1 + Tier 2 evaluation structure
4. **Auto-handoff** to `/evals-validate` to run validation

**Key EDD Principles Applied**:
- **Principle VIII**: Close Production Loop - Failure type gates route to appropriate actions
- **Principle II**: Binary Pass/Fail - Ensure graders return strictly 1.0 (pass) or 0.0 (fail)
- **Principle IX**: Test Data as Code - Unit test generated code against dataset examples

## When to use

- **After `/evals-clarify`**: Convert accepted goldset criteria into executable code
- **Regenerating configs**: Re-build evaluator suite after adding new goldset criteria
- **Adding unit tests**: Hardening the evaluator itself against regression or bugs

## When NOT to use

- **Goldset not published**: Run `/evals-clarify` to generate `goldset.json` first
- **Running evaluations**: Use `/evals-validate` to run the suite against application outputs

## Process

### User Input
```text
$ARGUMENTS
```
- `--system SYSTEM` — Override active evaluation framework (`promptfoo` or `deepeval`)
- `--no-tests` — Skip automated unit test generation for graders (not recommended)

### Execution Steps

#### Phase 1: Trace-to-Grader Synthesis (Automated Eval Engineering)
- Reads `evals/{system}/goldset.json`.
- Maps rich evidence fields from the goldset criteria into grader logic (Trace-to-Grader Synthesis):
  - Uses `pass_condition` and `fail_condition` as the grader's core rubric.
  - Extracts pass/fail examples to act as raw data anchors and few-shot classification anchors inside the grader logic.
  - Injects `Root Cause Analysis` and `axial_coding` notes as contextual prompt guidelines or regex patterns to catch exact failure manifestations.
- For PromptFoo: Generates Python grader functions (`evals/{system}/graders/check_*.py`) containing specialized, dynamic LLM-judge templates or regex checks compiled from these goldset inputs.
- For DeepEval: Generates Custom Metric classes inheriting from `BaseMetric` compiled from these goldset inputs.
- All graders conform strictly to the binary pass/fail standard (returning only `1.0` or `0.0`, with zero Likert scale leakage).

#### Phase 2: Unit Test Generation
- Generates matching unit tests (`evals/{system}/tests/test_check_*.py`) for each grader.
- Unit tests verify the grader correctly identifies the goldset's training pass and fail examples.

#### Phase 2b: Closed-Loop Grader Self-Tuning
- Executes generated unit tests (`pytest evals/{system}/tests/`) to verify evaluator accuracy.
- **Grader Calibration Loop**:
  1. Inspects test results to detect any misclassifications (false positives/negatives) on the training cases.
  2. If any test fails, triggers a feedback edit step that parses the failure reasons and automatically adjusts the grader's internal prompt rubric, regex stubs, or score thresholds.
  3. Re-runs pytest to check accuracy.
  4. Repeats for up to **3 iterations** (the hard circuit-breaker limit).
- **Holdout Locking**: Ensure the holdout validation set (`holdout.json`) remains completely isolated and is never loaded or exposed to the self-tuning loop (to prevent overfitting).
- **Failure Escalation**: If the grader does not converge to 100% training accuracy within 3 iterations, the loop halts, surfaces the failing test case details, and raises an error rather than passing silently.

#### Phase 3: Config Generation
- Generates the unified framework configuration file (`config.js` or `config.py`).
- Configures separate Tier 1 (fast checks, <30s, deterministic) and Tier 2 (semantic checks, <5min, LLM-judge) pipelines.

#### Phase 4: Auto-Handoff
Trigger `/evals-validate` to run validation.

## Verification
- `evals/{system}/graders/` contains Python grader scripts for each criterion compiled dynamically from goldset pass/fail examples and root-cause analyses
- `evals/{system}/tests/` contains matching unit test files
- Framework config (`config.js` or `config.py`) successfully generated
- Grader calibration self-tuning loop ran and converged to 100% training accuracy within the 3-iteration cap (or raised explicit non-convergence errors)
- Holdout dataset protection confirmed (validation `holdout.json` remained completely isolated and untouched during tuning)
- All grader unit tests pass locally (`pytest evals/{system}/tests/`)
- Handover summary lists generated graders, self-tuning iterations, and test results