src/eval/types.ts

/**
 * PRIX AI Evaluation System - Types
 * Complete type definitions for test cases, results, and metrics
 */

export type IssueType =
  | 'bug'
  | 'security'
  | 'performance'
  | 'style'
  | 'mutation'
  | 'none'
export type ConfidenceLevel = 'high' | 'medium' | 'low'

export interface TestSuggestion {
  filename: string
  startLine: number
  endLine: number
  suggestion: string
  confidence?: ConfidenceLevel
  issueType?: IssueType
}

export interface ExpectedResult {
  shouldDetectIssue: boolean
  issueType?: IssueType
  expectedConfidence?: ConfidenceLevel
  shouldAutoPR: boolean
  maxChangedLines?: number
  expectedSuggestions?: TestSuggestion[]
  shouldComment: boolean
  metadata?: Record<string, any>
}

export interface TestOutput {
  detectedIssues: boolean
  issueType?: IssueType
  confidence?: ConfidenceLevel
  autoPRTriggered: boolean
  suggestions: TestSuggestion[]
  comments: ReviewComment[]
  errors: string[]
  executionTime: number
  changedLines: number
}

export interface ReviewComment {
  path: string
  line: number
  body: string
  severity: 'info' | 'warning' | 'error'
}

export interface TestScore {
  detection: boolean
  classification: boolean
  confidence: boolean
  autoPR: boolean
  overall: boolean
  details: {
    detectionMatch: boolean
    typeMatch: boolean
    confidenceMatch: boolean
    autoPRMatch: boolean
    suggestionCountMatch: boolean
  }
}

export interface TestResult {
  testName: string
  testPath: string
  category: string
  passed: boolean
  score: TestScore
  output: TestOutput
  expected: ExpectedResult
  executionTime: number
  timestamp: string
}

export interface MetricsSummary {
  totalTests: number
  passedTests: number
  failedTests: number
  accuracy: number
  falsePositives: number
  falseNegatives: number
  autoPRPrecision: number
  autoPRRecall: number
  avgExecutionTime: number
  byCategory: Record<string, CategoryMetrics>
}

export interface CategoryMetrics {
  total: number
  passed: number
  failed: number
  accuracy: number
}

export interface BaselineMetrics {
  timestamp: string
  version: string
  summary: MetricsSummary
  results: TestResult[]
}

export interface RegressionReport {
  hasRegression: boolean
  accuracyDelta: number
  newFailures: string[]
  fixedTests: string[]
  details: string[]
}

export interface TestCase {
  name: string
  path: string
  category: string
  basePath: string
  headPath: string
  expected: ExpectedResult
}

export interface DiffFile {
  filename: string
  baseContent: string
  headContent: string
  diff: string
}

export interface MockPRContext {
  number: number
  title: string
  body: string
  base: {ref: string; sha: string}
  head: {ref: string; sha: string}
  files: DiffFile[]
}