@wix/eval-assertions 0.4.0 → 0.6.0

package/README.md

@@ -104,7 +104,7 @@ Runs a command in the working directory and checks the exit code. When the comma
 
 ### llm_judge
 
-Uses an LLM to evaluate the output with a customizable prompt.
+Uses an LLM to evaluate the output with a customizable prompt. The default system prompt instructs the judge to be strict on factual verification: when you ask to verify a specific fact, the judge must compare against the actual data and give 0 or near 0 if there is a mismatch. When the judge returns invalid JSON, the evaluator retries up to 3 times before failing.
 
 ```typescript
 {
@@ -118,6 +118,8 @@ Uses an LLM to evaluate the output with a customizable prompt.
 }
 ```
 
+**Tip:** When verifying file-related outcomes, include `{{changedFiles}}` in your prompt so the judge sees the actual files. Without it, the judge still receives this data in the system context, but making it explicit improves accuracy.
+
 **Available placeholders in prompts:**
 - `{{output}}` - The agent's final output text
 - `{{cwd}}` - Working directory path
@@ -134,10 +136,12 @@ The input data for assertion evaluation:
 interface EvaluationInput {
   outputText?: string;
   llmTrace?: LLMTrace;
-  fileDiffs?: Array<{ path: string }>;
+  fileDiffs?: Array<{ path: string; content?: string; status?: 'new' | 'modified' }>;
 }
 ```
 
+When `fileDiffs` items include `status`, the `{{modifiedFiles}}` and `{{newFiles}}` placeholders are populated for the LLM judge.
+
 ### AssertionContext
 
 Optional context for assertions:

package/build/index.js

@@ -61,7 +61,7 @@ var BuildPassedAssertionSchema = import_zod.z.object({
 });
 var LlmJudgeAssertionSchema = import_zod.z.object({
   type: import_zod.z.literal("llm_judge"),
-  /** Prompt template; placeholders: {{output}}, {{cwd}}, {{changedFiles}}, {{trace}} */
+  /** Prompt template; placeholders: {{output}}, {{cwd}}, {{changedFiles}}, {{modifiedFiles}}, {{newFiles}}, {{trace}} */
   prompt: import_zod.z.string(),
   /** Optional system prompt for the judge (default asks for JSON with score) */
   systemPrompt: import_zod.z.string().optional(),
@@ -363,7 +363,8 @@ function validateJudgeResult(parsed) {
 }
 var DEFAULT_MIN_SCORE = 70;
 var DEFAULT_MODEL = "claude-3-5-haiku-20241022";
-var DEFAULT_SYSTEM_PROMPT_CONTEXT = `You are judging a scenario run. Use these values:
+var DEFAULT_SYSTEM_PROMPT_CONTEXT = `You are judging a scenario run. The ACTUAL run data is provided below \u2014 use it to verify facts:
+
 - {{output}}: the agent's final output
 - {{cwd}}: working directory
 - {{changedFiles}}: list of all files changed (or "No files were changed")
@@ -371,7 +372,7 @@ var DEFAULT_SYSTEM_PROMPT_CONTEXT = `You are judging a scenario run. Use these v
 - {{newFiles}}: list of new files that were created (or "No new files were created")
 - {{trace}}: step-by-step trace (tool calls, completions) so you can check e.g. which tools were called and how many times
 
-Judge how well the output meets the acceptance criteria stated in the user prompt.`;
+CRITICAL: When the user asks you to verify a specific fact, compare it strictly against the actual data above. If the expected outcome does NOT match the actual outcome, you MUST give a score of 0 or near 0. Do not be lenient \u2014 factual mismatches are failures.`;
 var DEFAULT_SYSTEM_PROMPT = `You must respond only with a valid JSON object that conforms exactly to the following structure:
 
 {
@@ -431,52 +432,58 @@ var LlmJudgeEvaluator = class extends AssertionEvaluator {
         expected: String(minScore)
       };
     }
+    const maxParseAttempts = 3;
+    let lastParseError;
+    let lastRawText;
     try {
-      const result = generateTextStub ? await generateTextStub({
-        prompt: finalPrompt,
-        system: systemPrompt,
-        maxOutputTokens,
-        temperature
-      }) : await this.callGenerateText(
-        llmConfig,
-        assertion.model ?? DEFAULT_MODEL,
-        finalPrompt,
-        systemPrompt,
-        maxOutputTokens,
-        temperature
-      );
-      let judgeResult;
-      try {
-        const parsed = JSON.parse(result.text);
-        judgeResult = validateJudgeResult(parsed);
-      } catch (parseErr) {
-        return {
-          id: (0, import_crypto3.randomUUID)(),
-          assertionId,
-          assertionType: "llm_judge",
-          assertionName: "LLM judge",
-          status: "failed" /* FAILED */,
-          message: `Failed to parse judge response: ${parseErr instanceof Error ? parseErr.message : String(parseErr)}`,
-          expected: String(minScore),
-          actual: void 0,
-          details: { rawText: result.text.slice(0, 500) }
-        };
+      for (let attempt = 1; attempt <= maxParseAttempts; attempt++) {
+        const result = generateTextStub ? await generateTextStub({
+          prompt: finalPrompt,
+          system: systemPrompt,
+          maxOutputTokens,
+          temperature
+        }) : await this.callGenerateText(
+          llmConfig,
+          assertion.model ?? DEFAULT_MODEL,
+          finalPrompt,
+          systemPrompt,
+          maxOutputTokens,
+          temperature
+        );
+        lastRawText = result.text;
+        try {
+          const parsed = JSON.parse(result.text);
+          const judgeResult = validateJudgeResult(parsed);
+          const passed = judgeResult.score >= minScore;
+          return {
+            id: (0, import_crypto3.randomUUID)(),
+            assertionId,
+            assertionType: "llm_judge",
+            assertionName: "LLM judge",
+            status: passed ? "passed" /* PASSED */ : "failed" /* FAILED */,
+            message: passed ? `Judge score ${judgeResult.score} >= ${minScore}: ${judgeResult.text}` : `Judge score ${judgeResult.score} < ${minScore}: ${judgeResult.text}`,
+            expected: String(minScore),
+            actual: String(judgeResult.score),
+            details: {
+              score: judgeResult.score,
+              scoreReasoning: judgeResult.scoreReasoning,
+              text: judgeResult.text
+            }
+          };
+        } catch (parseErr) {
+          lastParseError = parseErr instanceof Error ? parseErr : new Error(String(parseErr));
+        }
       }
-      const passed = judgeResult.score >= minScore;
       return {
         id: (0, import_crypto3.randomUUID)(),
         assertionId,
         assertionType: "llm_judge",
         assertionName: "LLM judge",
-        status: passed ? "passed" /* PASSED */ : "failed" /* FAILED */,
-        message: passed ? `Judge score ${judgeResult.score} >= ${minScore}: ${judgeResult.text}` : `Judge score ${judgeResult.score} < ${minScore}: ${judgeResult.text}`,
+        status: "failed" /* FAILED */,
+        message: `Failed to parse judge response after ${maxParseAttempts} attempts: ${lastParseError?.message ?? "unknown"}`,
         expected: String(minScore),
-        actual: String(judgeResult.score),
-        details: {
-          score: judgeResult.score,
-          scoreReasoning: judgeResult.scoreReasoning,
-          text: judgeResult.text
-        }
+        actual: void 0,
+        details: { rawText: lastRawText?.slice(0, 500) }
       };
     } catch (err) {
       const message = err instanceof Error ? err.message : String(err);

package/build/index.js.map

@@ -1,7 +1,7 @@
 {
   "version": 3,
   "sources": ["../src/index.ts", "../src/types/assertions.ts", "../src/types/trace.ts", "../src/types/result.ts", "../src/evaluators/index.ts", "../src/evaluators/skill-was-called-evaluator.ts", "../src/evaluators/assertion-evaluator.ts", "../src/evaluators/build-passed-evaluator.ts", "../src/evaluators/llm-judge-evaluator.ts"],
-  "sourcesContent": ["/**\n * @wix/eval-assertions\n *\n * Assertion framework for AI agent evaluations.\n * Supports skill invocation checks, build validation, and LLM-based judging.\n */\n\n// Types\nexport {\n  // Assertion schemas and types\n  AssertionSchema,\n  SkillWasCalledAssertionSchema,\n  BuildPassedAssertionSchema,\n  LlmJudgeAssertionSchema,\n  type Assertion,\n  type SkillWasCalledAssertion,\n  type BuildPassedAssertion,\n  type LlmJudgeAssertion,\n  // Trace types\n  LLMTraceSchema,\n  LLMTraceStepSchema,\n  LLMTraceSummarySchema,\n  LLMBreakdownStatsSchema,\n  TokenUsageSchema,\n  LLMStepType,\n  type LLMTrace,\n  type LLMTraceStep,\n  type LLMTraceSummary,\n  type LLMBreakdownStats,\n  type TokenUsage,\n  // Result types\n  AssertionResultSchema,\n  AssertionResultStatus,\n  type AssertionResult,\n  // Input types\n  type EvaluationInput,\n  type FileDiff,\n} from \"./types/index.js\";\n\n// Evaluators\nexport {\n  evaluateAssertions,\n  registerEvaluator,\n  getEvaluator,\n  AssertionEvaluator,\n  SkillWasCalledEvaluator,\n  BuildPassedEvaluator,\n  LlmJudgeEvaluator,\n  formatTraceForJudge,\n  replacePlaceholders,\n  validateJudgeResult,\n  type AssertionContext,\n  type LlmConfig,\n  type LlmJudgeGenerateTextOptions,\n  type JudgeResult,\n} from \"./evaluators/index.js\";\n", "import { z } from \"zod\";\n\n/**\n * Assertion: the agent must have invoked a specific skill during the run.\n * Checked by inspecting the LLM trace for a \"Skill\" tool use with the given skill.\n * Data: skillName (the skill that must have been called).\n */\nexport const SkillWasCalledAssertionSchema = z.object({\n  type: z.literal(\"skill_was_called\"),\n  /** Name of the skill that must have been called (matched against trace Skill tool args) */\n  skillName: z.string(),\n});\n\nexport type SkillWasCalledAssertion = z.infer<\n  typeof SkillWasCalledAssertionSchema\n>;\n\n/**\n * Assertion: a build command must exit with the expected code (default 0).\n * Runs the command in the scenario working directory.\n */\nexport const BuildPassedAssertionSchema = z.object({\n  type: z.literal(\"build_passed\"),\n  /** Command to run (default: \"yarn build\") */\n  command: z.string().optional(),\n  /** Expected exit code (default: 0) */\n  expectedExitCode: z.number().int().optional(),\n});\n\nexport type BuildPassedAssertion = z.infer<typeof BuildPassedAssertionSchema>;\n\n/**\n * Assertion: an LLM judges the scenario output (score 0-100).\n * Prompt can use {{output}}, {{cwd}}, {{changedFiles}}, {{trace}}.\n * Passes if judge score >= minScore.\n */\nexport const LlmJudgeAssertionSchema = z.object({\n  type: z.literal(\"llm_judge\"),\n  /** Prompt template; placeholders: {{output}}, {{cwd}}, {{changedFiles}}, {{trace}} */\n  prompt: z.string(),\n  /** Optional system prompt for the judge (default asks for JSON with score) */\n  systemPrompt: z.string().optional(),\n  /** Minimum score to pass (0-100, default 70) */\n  minScore: z.number().int().min(0).max(100).optional(),\n  /** Model for the judge (e.g. claude-3-5-haiku) */\n  model: z.string().optional(),\n  maxTokens: z.number().int().optional(),\n  temperature: z.number().min(0).max(1).optional(),\n});\n\nexport type LlmJudgeAssertion = z.infer<typeof LlmJudgeAssertionSchema>;\n\n/**\n * Union of all assertion types.\n * Each assertion has a type and type-specific data.\n * Uses z.union (not z.discriminatedUnion) for Zod v4 compatibility when used as array element.\n */\nexport const AssertionSchema = z.union([\n  SkillWasCalledAssertionSchema,\n  BuildPassedAssertionSchema,\n  LlmJudgeAssertionSchema,\n]);\n\nexport type Assertion = z.infer<typeof AssertionSchema>;\n", "import { z } from \"zod\";\n\n/**\n * Token usage schema.\n */\nexport const TokenUsageSchema = z.object({\n  prompt: z.number(),\n  completion: z.number(),\n  total: z.number(),\n});\n\nexport type TokenUsage = z.infer<typeof TokenUsageSchema>;\n\n/**\n * LLM step type enum.\n */\nexport enum LLMStepType {\n  COMPLETION = \"completion\",\n  TOOL_USE = \"tool_use\",\n  TOOL_RESULT = \"tool_result\",\n  THINKING = \"thinking\",\n}\n\n/**\n * LLM trace step schema.\n */\nexport const LLMTraceStepSchema = z.object({\n  id: z.string(),\n  stepNumber: z.number(),\n  type: z.enum(LLMStepType),\n  model: z.string(),\n  provider: z.string(),\n  startedAt: z.string(),\n  durationMs: z.number(),\n  tokenUsage: TokenUsageSchema,\n  costUsd: z.number(),\n  toolName: z.string().optional(),\n  toolArguments: z.string().optional(),\n  inputPreview: z.string().optional(),\n  outputPreview: z.string().optional(),\n  success: z.boolean(),\n  error: z.string().optional(),\n});\n\nexport type LLMTraceStep = z.infer<typeof LLMTraceStepSchema>;\n\n/**\n * LLM breakdown stats schema.\n */\nexport const LLMBreakdownStatsSchema = z.object({\n  count: z.number(),\n  durationMs: z.number(),\n  tokens: z.number(),\n  costUsd: z.number(),\n});\n\nexport type LLMBreakdownStats = z.infer<typeof LLMBreakdownStatsSchema>;\n\n/**\n * LLM trace summary schema.\n */\nexport const LLMTraceSummarySchema = z.object({\n  totalSteps: z.number(),\n  totalDurationMs: z.number(),\n  totalTokens: TokenUsageSchema,\n  totalCostUsd: z.number(),\n  stepTypeBreakdown: z.record(z.string(), LLMBreakdownStatsSchema).optional(),\n  modelBreakdown: z.record(z.string(), LLMBreakdownStatsSchema),\n  modelsUsed: z.array(z.string()),\n});\n\nexport type LLMTraceSummary = z.infer<typeof LLMTraceSummarySchema>;\n\n/**\n * LLM trace schema.\n */\nexport const LLMTraceSchema = z.object({\n  id: z.string(),\n  steps: z.array(LLMTraceStepSchema),\n  summary: LLMTraceSummarySchema,\n});\n\nexport type LLMTrace = z.infer<typeof LLMTraceSchema>;\n", "import { z } from \"zod\";\nimport { LLMTraceStepSchema } from \"./trace.js\";\n\n/**\n * Assertion result status enum.\n */\nexport enum AssertionResultStatus {\n  PASSED = \"passed\",\n  FAILED = \"failed\",\n  SKIPPED = \"skipped\",\n  ERROR = \"error\",\n}\n\n/**\n * Assertion result schema.\n */\nexport const AssertionResultSchema = z.object({\n  id: z.string(),\n  assertionId: z.string(),\n  assertionType: z.string(),\n  assertionName: z.string(),\n  status: z.enum(AssertionResultStatus),\n  message: z.string().optional(),\n  expected: z.string().optional(),\n  actual: z.string().optional(),\n  duration: z.number().optional(),\n  details: z.record(z.string(), z.unknown()).optional(),\n  llmTraceSteps: z.array(LLMTraceStepSchema).optional(),\n});\n\nexport type AssertionResult = z.infer<typeof AssertionResultSchema>;\n", "import type { Assertion, AssertionResult } from \"../types/index.js\";\nimport { AssertionResultStatus } from \"../types/index.js\";\nimport { randomUUID } from \"crypto\";\nimport type { AssertionContext } from \"./assertion-evaluator.js\";\nimport { AssertionEvaluator } from \"./assertion-evaluator.js\";\nimport { SkillWasCalledEvaluator } from \"./skill-was-called-evaluator.js\";\nimport { BuildPassedEvaluator } from \"./build-passed-evaluator.js\";\nimport { LlmJudgeEvaluator } from \"./llm-judge-evaluator.js\";\nimport type { EvaluationInput } from \"../types/index.js\";\n\nconst llmJudgeEvaluator = new LlmJudgeEvaluator();\n\nconst evaluators: Record<string, AssertionEvaluator> = {\n  skill_was_called: new SkillWasCalledEvaluator(),\n  build_passed: new BuildPassedEvaluator(),\n  llm_judge: llmJudgeEvaluator,\n  // Custom assertions use the same LLM-based evaluation as llm_judge\n  custom: llmJudgeEvaluator,\n};\n\n/**\n * Register a custom assertion evaluator.\n *\n * @param type - The assertion type identifier\n * @param evaluator - The evaluator instance\n */\nexport function registerEvaluator(\n  type: string,\n  evaluator: AssertionEvaluator,\n): void {\n  evaluators[type] = evaluator;\n}\n\n/**\n * Get a registered evaluator by type.\n *\n * @param type - The assertion type identifier\n * @returns The evaluator or undefined if not found\n */\nexport function getEvaluator(type: string): AssertionEvaluator | undefined {\n  return evaluators[type];\n}\n\n/**\n * Evaluate all assertions against the input.\n *\n * @param input - Evaluation input (includes outputText, llmTrace, fileDiffs)\n * @param assertions - List of assertions to evaluate\n * @param context - Optional context (e.g. workDir for build_passed, llmConfig for llm_judge)\n * @returns Array of assertion results; empty if no assertions\n */\nexport async function evaluateAssertions(\n  input: EvaluationInput,\n  assertions: Assertion[],\n  context?: AssertionContext,\n): Promise<AssertionResult[]> {\n  if (assertions.length === 0) {\n    return [];\n  }\n  return Promise.all(\n    assertions.map(async (assertion) => {\n      const evaluator = evaluators[assertion.type];\n      if (!evaluator) {\n        return {\n          id: randomUUID(),\n          assertionId: randomUUID(),\n          assertionType: assertion.type,\n          assertionName: \"Unknown assertion\",\n          status: AssertionResultStatus.ERROR,\n          message: `Unsupported assertion type: ${assertion.type}`,\n          duration: 0,\n        };\n      }\n      const startMs = Date.now();\n      const result = await evaluator.evaluate(assertion, input, context);\n      const durationMs = Date.now() - startMs;\n      return { ...result, duration: durationMs };\n    }),\n  );\n}\n\n// Re-export evaluator classes and types\nexport { AssertionEvaluator } from \"./assertion-evaluator.js\";\nexport type {\n  AssertionContext,\n  LlmConfig,\n  LlmJudgeGenerateTextOptions,\n} from \"./assertion-evaluator.js\";\nexport { SkillWasCalledEvaluator } from \"./skill-was-called-evaluator.js\";\nexport { BuildPassedEvaluator } from \"./build-passed-evaluator.js\";\nexport {\n  LlmJudgeEvaluator,\n  formatTraceForJudge,\n  replacePlaceholders,\n  validateJudgeResult,\n  type JudgeResult,\n} from \"./llm-judge-evaluator.js\";\n", "import type {\n  SkillWasCalledAssertion,\n  AssertionResult,\n  LLMTrace,\n  EvaluationInput,\n} from \"../types/index.js\";\nimport { AssertionResultStatus } from \"../types/index.js\";\nimport { randomUUID } from \"crypto\";\nimport type { AssertionContext } from \"./assertion-evaluator.js\";\nimport { AssertionEvaluator } from \"./assertion-evaluator.js\";\n\n/**\n * Check if parsed Skill tool args reference the expected skill by name.\n */\nfunction toolArgsReferenceSkillByName(\n  args: unknown,\n  expectedSkillName: string,\n): boolean {\n  if (args === null || typeof args !== \"object\") {\n    return false;\n  }\n  const obj = args as Record<string, unknown>;\n  const name = obj.skill;\n  return typeof name === \"string\" && name === expectedSkillName;\n}\n\n/**\n * Evaluator for \"skill_was_called\" assertion: the LLM trace must contain a step\n * where the \"Skill\" tool was used with the expected skill (by name).\n */\nexport class SkillWasCalledEvaluator extends AssertionEvaluator<SkillWasCalledAssertion> {\n  readonly type = \"skill_was_called\" as const;\n\n  evaluate(\n    assertion: SkillWasCalledAssertion,\n    input: EvaluationInput,\n    // eslint-disable-next-line @typescript-eslint/no-unused-vars -- context not used for skill_was_called\n    _context?: AssertionContext,\n  ): AssertionResult {\n    const assertionId = randomUUID();\n\n    const llmTrace: LLMTrace | undefined = input.llmTrace;\n    if (!llmTrace?.steps?.length) {\n      return {\n        id: randomUUID(),\n        assertionId,\n        assertionType: \"skill_was_called\",\n        assertionName: \"Skill was called\",\n        status: AssertionResultStatus.FAILED,\n        message: \"No LLM trace steps to check for skill invocation\",\n        expected: assertion.skillName,\n      };\n    }\n\n    for (const step of llmTrace.steps) {\n      if (step.toolName !== \"Skill\") {\n        continue;\n      }\n      let args: unknown;\n      try {\n        args = step.toolArguments\n          ? (JSON.parse(step.toolArguments) as unknown)\n          : undefined;\n      } catch {\n        continue;\n      }\n      if (toolArgsReferenceSkillByName(args, assertion.skillName)) {\n        return {\n          id: randomUUID(),\n          assertionId,\n          assertionType: \"skill_was_called\",\n          assertionName: \"Skill was called\",\n          status: AssertionResultStatus.PASSED,\n          message: `Skill \"${assertion.skillName}\" was called`,\n          expected: assertion.skillName,\n        };\n      }\n    }\n\n    return {\n      id: randomUUID(),\n      assertionId,\n      assertionType: \"skill_was_called\",\n      assertionName: \"Skill was called\",\n      status: AssertionResultStatus.FAILED,\n      message: `Skill \"${assertion.skillName}\" was not called`,\n      expected: assertion.skillName,\n    };\n  }\n}\n", "import type {\n  Assertion,\n  AssertionResult,\n  EvaluationInput,\n} from \"../types/index.js\";\n\n/**\n * Options passed to the LLM for llm_judge. Used by the optional stub for testing.\n */\nexport interface LlmJudgeGenerateTextOptions {\n  prompt: string;\n  system: string;\n  maxOutputTokens: number;\n  temperature: number;\n}\n\n/**\n * Configuration for LLM calls (used by llm_judge assertion).\n */\nexport interface LlmConfig {\n  /** Base URL for the AI API (e.g., 'https://api.anthropic.com') */\n  baseUrl: string;\n  /** Headers to include in API requests (e.g., API key) */\n  headers: Record<string, string>;\n}\n\n/**\n * Optional context passed when evaluating assertions.\n */\nexport interface AssertionContext {\n  /** Working directory for the scenario (used by build_passed) */\n  workDir?: string;\n  /** LLM configuration (used by llm_judge) */\n  llmConfig?: LlmConfig;\n  /**\n   * Optional stub for llm_judge: when set, the evaluator uses this instead of the real AI call.\n   * Used only in tests to avoid hitting the API.\n   */\n  generateTextForLlmJudge?: (\n    options: LlmJudgeGenerateTextOptions,\n  ) => Promise<{ text: string }>;\n}\n\n/**\n * Abstract base for assertion evaluators.\n * Each assertion type has a concrete class that implements evaluate().\n * evaluate() may return a Promise for async assertions (e.g. llm_judge).\n */\nexport abstract class AssertionEvaluator<T extends Assertion = Assertion> {\n  abstract readonly type: T[\"type\"];\n\n  abstract evaluate(\n    assertion: T,\n    input: EvaluationInput,\n    context?: AssertionContext,\n  ): AssertionResult | Promise<AssertionResult>;\n}\n", "import type {\n  BuildPassedAssertion,\n  AssertionResult,\n  EvaluationInput,\n} from \"../types/index.js\";\nimport { AssertionResultStatus } from \"../types/index.js\";\nimport { randomUUID } from \"crypto\";\nimport { execSync } from \"child_process\";\nimport type { AssertionContext } from \"./assertion-evaluator.js\";\nimport { AssertionEvaluator } from \"./assertion-evaluator.js\";\n\nconst DEFAULT_COMMAND = \"yarn build\";\nconst DEFAULT_EXIT_CODE = 0;\n\n/**\n * Evaluator for \"build_passed\" assertion: runs a build command in the scenario\n * working directory and passes if the command exits with the expected code (default 0).\n */\nexport class BuildPassedEvaluator extends AssertionEvaluator<BuildPassedAssertion> {\n  readonly type = \"build_passed\" as const;\n\n  evaluate(\n    assertion: BuildPassedAssertion,\n    _input: EvaluationInput,\n    context?: AssertionContext,\n  ): AssertionResult {\n    const assertionId = randomUUID();\n    const workDir = context?.workDir;\n    const command = assertion.command ?? DEFAULT_COMMAND;\n    const expectedExitCode = assertion.expectedExitCode ?? DEFAULT_EXIT_CODE;\n\n    if (!workDir) {\n      return this.createResult(assertionId, {\n        status: AssertionResultStatus.FAILED,\n        message: \"No working directory provided for build_passed assertion\",\n        expected: String(expectedExitCode),\n      });\n    }\n\n    let exitCode: number | null = null;\n    let errorMessage: string | null = null;\n    let stdout: string | undefined;\n    let stderr: string | undefined;\n\n    console.log(`[build_passed] Running \"${command}\" in: ${workDir}`);\n\n    try {\n      execSync(command, {\n        cwd: workDir,\n        encoding: \"utf-8\",\n        stdio: [\"ignore\", \"pipe\", \"pipe\"],\n      });\n      exitCode = 0;\n    } catch (err) {\n      const error = err as Error & {\n        status?: number;\n        code?: number;\n        stdout?: string | Buffer;\n        stderr?: string | Buffer;\n      };\n      exitCode =\n        typeof error.status === \"number\"\n          ? error.status\n          : typeof error.code === \"number\"\n            ? error.code\n            : null;\n      errorMessage = error.message;\n      stdout = this.bufferToString(error.stdout);\n      stderr = this.bufferToString(error.stderr);\n    }\n\n    const passed = exitCode !== null && exitCode === expectedExitCode;\n\n    const details: Record<string, unknown> = { workDir, command };\n    if (stdout !== undefined && stdout !== \"\") {\n      details.stdout = stdout;\n    }\n    if (stderr !== undefined && stderr !== \"\") {\n      details.stderr = stderr;\n    }\n\n    return this.createResult(assertionId, {\n      status: passed\n        ? AssertionResultStatus.PASSED\n        : AssertionResultStatus.FAILED,\n      message: this.formatMessage(exitCode, expectedExitCode, errorMessage),\n      expected: String(expectedExitCode),\n      actual: exitCode !== null ? String(exitCode) : undefined,\n      details,\n    });\n  }\n\n  private createResult(\n    assertionId: string,\n    fields: Partial<AssertionResult>,\n  ): AssertionResult {\n    return {\n      id: randomUUID(),\n      assertionId,\n      assertionType: \"build_passed\",\n      assertionName: \"Build passed\",\n      status: AssertionResultStatus.FAILED,\n      ...fields,\n    };\n  }\n\n  private bufferToString(\n    value: string | Buffer | undefined,\n  ): string | undefined {\n    if (value === undefined || value === null) return undefined;\n    if (typeof value === \"string\") return value;\n    return value.toString(\"utf-8\");\n  }\n\n  private formatMessage(\n    exitCode: number | null,\n    expectedExitCode: number,\n    errorMessage: string | null,\n  ): string {\n    if (exitCode === null) {\n      return `Build failed: ${errorMessage}`;\n    }\n    if (exitCode === expectedExitCode) {\n      return `Build passed (exit code ${exitCode})`;\n    }\n    return `Build exited with ${exitCode}, expected ${expectedExitCode}`;\n  }\n}\n", "import type {\n  LlmJudgeAssertion,\n  AssertionResult,\n  LLMTrace,\n  EvaluationInput,\n} from \"../types/index.js\";\nimport { AssertionResultStatus } from \"../types/index.js\";\nimport { randomUUID } from \"crypto\";\nimport { createAnthropic } from \"@ai-sdk/anthropic\";\nimport { generateText } from \"ai\";\nimport type { AssertionContext, LlmConfig } from \"./assertion-evaluator.js\";\nimport { AssertionEvaluator } from \"./assertion-evaluator.js\";\n\nexport interface JudgeResult {\n  text: string;\n  score: number;\n  scoreReasoning: string;\n}\n\n/**\n * Format LLM trace as readable text for the judge (step number, type, tool name/args, output preview).\n */\nexport function formatTraceForJudge(llmTrace: LLMTrace | undefined): string {\n  if (!llmTrace?.steps?.length) {\n    return \"No trace available.\";\n  }\n  const lines: string[] = [];\n  for (const step of llmTrace.steps) {\n    const parts: string[] = [\n      `Step ${step.stepNumber}`,\n      `type: ${step.type}`,\n      `duration: ${step.durationMs}ms`,\n    ];\n    if (step.toolName) {\n      parts.push(`tool: ${step.toolName}`);\n      if (step.toolArguments) {\n        parts.push(`args: ${step.toolArguments}`);\n      }\n    }\n    if (step.outputPreview) {\n      parts.push(`output: ${step.outputPreview}`);\n    }\n    if (step.error) {\n      parts.push(`error: ${step.error}`);\n    }\n    lines.push(parts.join(\", \"));\n  }\n  return lines.join(\"\\n\");\n}\n\n/**\n * Context object for placeholder replacement.\n */\nexport interface PlaceholderContext {\n  output: string;\n  cwd: string;\n  changedFiles: string;\n  modifiedFiles: string;\n  newFiles: string;\n  trace: string;\n}\n\nexport function replacePlaceholders(\n  str: string,\n  ctx: PlaceholderContext,\n): string {\n  return str\n    .replace(/\\{\\{output\\}\\}/g, ctx.output)\n    .replace(/\\{\\{cwd\\}\\}/g, ctx.cwd)\n    .replace(/\\{\\{changedFiles\\}\\}/g, ctx.changedFiles)\n    .replace(/\\{\\{modifiedFiles\\}\\}/g, ctx.modifiedFiles)\n    .replace(/\\{\\{newFiles\\}\\}/g, ctx.newFiles)\n    .replace(/\\{\\{trace\\}\\}/g, ctx.trace);\n}\n\nexport function validateJudgeResult(parsed: unknown): JudgeResult {\n  if (parsed === null || typeof parsed !== \"object\") {\n    throw new Error(\"Judge result is not an object\");\n  }\n  const obj = parsed as Record<string, unknown>;\n  if (typeof obj.text !== \"string\") {\n    throw new Error(\"Judge result does not contain a valid text field\");\n  }\n  if (typeof obj.score !== \"number\") {\n    throw new Error(\"Judge result does not contain a valid score field\");\n  }\n  if (obj.score < 0 || obj.score > 100) {\n    throw new Error(\"Judge result score is not between 0 and 100\");\n  }\n  if (typeof obj.scoreReasoning !== \"string\") {\n    throw new Error(\n      \"Judge result does not contain a valid scoreReasoning field\",\n    );\n  }\n  return {\n    text: obj.text,\n    score: obj.score,\n    scoreReasoning: obj.scoreReasoning,\n  };\n}\n\nconst DEFAULT_MIN_SCORE = 70;\nconst DEFAULT_MODEL = \"claude-3-5-haiku-20241022\";\n\n/** Default system prompt context with placeholders; used when assertion.systemPrompt is empty. */\nconst DEFAULT_SYSTEM_PROMPT_CONTEXT = `You are judging a scenario run. Use these values:\n- {{output}}: the agent's final output\n- {{cwd}}: working directory\n- {{changedFiles}}: list of all files changed (or \"No files were changed\")\n- {{modifiedFiles}}: list of existing files that were modified (or \"No files were modified\")\n- {{newFiles}}: list of new files that were created (or \"No new files were created\")\n- {{trace}}: step-by-step trace (tool calls, completions) so you can check e.g. which tools were called and how many times\n\nJudge how well the output meets the acceptance criteria stated in the user prompt.`;\n\nconst DEFAULT_SYSTEM_PROMPT = `You must respond only with a valid JSON object that conforms exactly to the following structure:\n\n{\n  \"text\": string,\n  \"score\": number (0-100),\n  \"scoreReasoning\": string\n}\n\n- text: A brief textual verdict of the test result.\n- score: A number from 0 to 100 that reflects how well the answer meets the acceptance criteria.\n- scoreReasoning: A concise explanation justifying the assigned score.\n\nYour response must:\n- Contain only the JSON object above \u2014 no introductory text, no code formatting (e.g., no triple backticks), and no trailing comments.\n- Be valid and parseable by \\`JSON.parse\\`.\n- Use only double quotes for all keys and strings, as required by JSON.\n\nAny response that includes extra content or deviates from the specified format will cause parsing to fail. Follow these instructions exactly.`;\n\n/**\n * Evaluator for \"llm_judge\" assertion: an LLM judges the scenario output\n * (prompt with {{output}}, {{cwd}}, {{changedFiles}}, {{trace}}) and returns a score 0-100.\n * Passes if score >= minScore.\n */\nexport class LlmJudgeEvaluator extends AssertionEvaluator<LlmJudgeAssertion> {\n  readonly type = \"llm_judge\" as const;\n\n  async evaluate(\n    assertion: LlmJudgeAssertion,\n    input: EvaluationInput,\n    context?: AssertionContext,\n  ): Promise<AssertionResult> {\n    const assertionId = randomUUID();\n    const llmConfig = context?.llmConfig;\n    const workDir = context?.workDir ?? \"\";\n    const generateTextStub = context?.generateTextForLlmJudge;\n\n    const output = input.outputText ?? \"\";\n    const fileDiffs = input.fileDiffs ?? [];\n\n    // Compute file lists by status\n    const changedPaths = fileDiffs.map((d) => d.path);\n    const modifiedPaths = fileDiffs\n      .filter((d) => d.status === \"modified\")\n      .map((d) => d.path);\n    const newPaths = fileDiffs\n      .filter((d) => d.status === \"new\")\n      .map((d) => d.path);\n\n    const changedFiles =\n      changedPaths.length > 0\n        ? changedPaths.map((path: string) => `- ${path}`).join(\"\\n\")\n        : \"No files were changed\";\n    const modifiedFiles =\n      modifiedPaths.length > 0\n        ? modifiedPaths.map((path: string) => `- ${path}`).join(\"\\n\")\n        : \"No files were modified\";\n    const newFiles =\n      newPaths.length > 0\n        ? newPaths.map((path: string) => `- ${path}`).join(\"\\n\")\n        : \"No new files were created\";\n\n    const trace = formatTraceForJudge(input.llmTrace);\n    const ctx: PlaceholderContext = {\n      output,\n      cwd: workDir,\n      changedFiles,\n      modifiedFiles,\n      newFiles,\n      trace,\n    };\n    const replace = (s: string) => replacePlaceholders(s, ctx);\n\n    const finalPrompt = replace(assertion.prompt);\n    const systemPrompt =\n      assertion.systemPrompt != null && assertion.systemPrompt !== \"\"\n        ? replace(assertion.systemPrompt) + \"\\n\\n\" + DEFAULT_SYSTEM_PROMPT\n        : replace(DEFAULT_SYSTEM_PROMPT_CONTEXT) +\n          \"\\n\\n\" +\n          DEFAULT_SYSTEM_PROMPT;\n\n    const minScore = assertion.minScore ?? DEFAULT_MIN_SCORE;\n    const maxOutputTokens = assertion.maxTokens ?? 1024;\n    const temperature = assertion.temperature ?? 0;\n\n    if (!generateTextStub && !llmConfig) {\n      return {\n        id: randomUUID(),\n        assertionId,\n        assertionType: \"llm_judge\",\n        assertionName: \"LLM judge\",\n        status: AssertionResultStatus.FAILED,\n        message: \"No llmConfig for llm_judge assertion (AI gateway required)\",\n        expected: String(minScore),\n      };\n    }\n\n    try {\n      const result = generateTextStub\n        ? await generateTextStub({\n            prompt: finalPrompt,\n            system: systemPrompt,\n            maxOutputTokens,\n            temperature,\n          })\n        : await this.callGenerateText(\n            llmConfig!,\n            assertion.model ?? DEFAULT_MODEL,\n            finalPrompt,\n            systemPrompt,\n            maxOutputTokens,\n            temperature,\n          );\n\n      let judgeResult: JudgeResult;\n      try {\n        const parsed = JSON.parse(result.text);\n        judgeResult = validateJudgeResult(parsed);\n      } catch (parseErr) {\n        return {\n          id: randomUUID(),\n          assertionId,\n          assertionType: \"llm_judge\",\n          assertionName: \"LLM judge\",\n          status: AssertionResultStatus.FAILED,\n          message: `Failed to parse judge response: ${parseErr instanceof Error ? parseErr.message : String(parseErr)}`,\n          expected: String(minScore),\n          actual: undefined,\n          details: { rawText: result.text.slice(0, 500) },\n        };\n      }\n\n      const passed = judgeResult.score >= minScore;\n      return {\n        id: randomUUID(),\n        assertionId,\n        assertionType: \"llm_judge\",\n        assertionName: \"LLM judge\",\n        status: passed\n          ? AssertionResultStatus.PASSED\n          : AssertionResultStatus.FAILED,\n        message: passed\n          ? `Judge score ${judgeResult.score} >= ${minScore}: ${judgeResult.text}`\n          : `Judge score ${judgeResult.score} < ${minScore}: ${judgeResult.text}`,\n        expected: String(minScore),\n        actual: String(judgeResult.score),\n        details: {\n          score: judgeResult.score,\n          scoreReasoning: judgeResult.scoreReasoning,\n          text: judgeResult.text,\n        },\n      };\n    } catch (err) {\n      const message = err instanceof Error ? err.message : String(err);\n      return {\n        id: randomUUID(),\n        assertionId,\n        assertionType: \"llm_judge\",\n        assertionName: \"LLM judge\",\n        status: AssertionResultStatus.FAILED,\n        message: `LLM judge call failed: ${message}`,\n        expected: String(minScore),\n        details: { error: message },\n      };\n    }\n  }\n\n  private async callGenerateText(\n    llmConfig: LlmConfig,\n    modelId: string,\n    prompt: string,\n    system: string,\n    maxOutputTokens: number,\n    temperature: number,\n  ): Promise<{ text: string }> {\n    const anthropic = createAnthropic({\n      baseURL: llmConfig.baseUrl,\n      apiKey: \"dummy\",\n      headers: llmConfig.headers,\n    });\n    const result = await generateText({\n      model: anthropic(modelId),\n      prompt,\n      system,\n      maxOutputTokens,\n      temperature,\n    });\n    return { text: result.text };\n  }\n}\n"],
-  "mappings": ";;;;;;;;;;;;;;;;;;;;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;;;ACAA,iBAAkB;AAOX,IAAM,gCAAgC,aAAE,OAAO;AAAA,EACpD,MAAM,aAAE,QAAQ,kBAAkB;AAAA;AAAA,EAElC,WAAW,aAAE,OAAO;AACtB,CAAC;AAUM,IAAM,6BAA6B,aAAE,OAAO;AAAA,EACjD,MAAM,aAAE,QAAQ,cAAc;AAAA;AAAA,EAE9B,SAAS,aAAE,OAAO,EAAE,SAAS;AAAA;AAAA,EAE7B,kBAAkB,aAAE,OAAO,EAAE,IAAI,EAAE,SAAS;AAC9C,CAAC;AASM,IAAM,0BAA0B,aAAE,OAAO;AAAA,EAC9C,MAAM,aAAE,QAAQ,WAAW;AAAA;AAAA,EAE3B,QAAQ,aAAE,OAAO;AAAA;AAAA,EAEjB,cAAc,aAAE,OAAO,EAAE,SAAS;AAAA;AAAA,EAElC,UAAU,aAAE,OAAO,EAAE,IAAI,EAAE,IAAI,CAAC,EAAE,IAAI,GAAG,EAAE,SAAS;AAAA;AAAA,EAEpD,OAAO,aAAE,OAAO,EAAE,SAAS;AAAA,EAC3B,WAAW,aAAE,OAAO,EAAE,IAAI,EAAE,SAAS;AAAA,EACrC,aAAa,aAAE,OAAO,EAAE,IAAI,CAAC,EAAE,IAAI,CAAC,EAAE,SAAS;AACjD,CAAC;AASM,IAAM,kBAAkB,aAAE,MAAM;AAAA,EACrC;AAAA,EACA;AAAA,EACA;AACF,CAAC;;;AC7DD,IAAAA,cAAkB;AAKX,IAAM,mBAAmB,cAAE,OAAO;AAAA,EACvC,QAAQ,cAAE,OAAO;AAAA,EACjB,YAAY,cAAE,OAAO;AAAA,EACrB,OAAO,cAAE,OAAO;AAClB,CAAC;AAOM,IAAK,cAAL,kBAAKC,iBAAL;AACL,EAAAA,aAAA,gBAAa;AACb,EAAAA,aAAA,cAAW;AACX,EAAAA,aAAA,iBAAc;AACd,EAAAA,aAAA,cAAW;AAJD,SAAAA;AAAA,GAAA;AAUL,IAAM,qBAAqB,cAAE,OAAO;AAAA,EACzC,IAAI,cAAE,OAAO;AAAA,EACb,YAAY,cAAE,OAAO;AAAA,EACrB,MAAM,cAAE,KAAK,WAAW;AAAA,EACxB,OAAO,cAAE,OAAO;AAAA,EAChB,UAAU,cAAE,OAAO;AAAA,EACnB,WAAW,cAAE,OAAO;AAAA,EACpB,YAAY,cAAE,OAAO;AAAA,EACrB,YAAY;AAAA,EACZ,SAAS,cAAE,OAAO;AAAA,EAClB,UAAU,cAAE,OAAO,EAAE,SAAS;AAAA,EAC9B,eAAe,cAAE,OAAO,EAAE,SAAS;AAAA,EACnC,cAAc,cAAE,OAAO,EAAE,SAAS;AAAA,EAClC,eAAe,cAAE,OAAO,EAAE,SAAS;AAAA,EACnC,SAAS,cAAE,QAAQ;AAAA,EACnB,OAAO,cAAE,OAAO,EAAE,SAAS;AAC7B,CAAC;AAOM,IAAM,0BAA0B,cAAE,OAAO;AAAA,EAC9C,OAAO,cAAE,OAAO;AAAA,EAChB,YAAY,cAAE,OAAO;AAAA,EACrB,QAAQ,cAAE,OAAO;AAAA,EACjB,SAAS,cAAE,OAAO;AACpB,CAAC;AAOM,IAAM,wBAAwB,cAAE,OAAO;AAAA,EAC5C,YAAY,cAAE,OAAO;AAAA,EACrB,iBAAiB,cAAE,OAAO;AAAA,EAC1B,aAAa;AAAA,EACb,cAAc,cAAE,OAAO;AAAA,EACvB,mBAAmB,cAAE,OAAO,cAAE,OAAO,GAAG,uBAAuB,EAAE,SAAS;AAAA,EAC1E,gBAAgB,cAAE,OAAO,cAAE,OAAO,GAAG,uBAAuB;AAAA,EAC5D,YAAY,cAAE,MAAM,cAAE,OAAO,CAAC;AAChC,CAAC;AAOM,IAAM,iBAAiB,cAAE,OAAO;AAAA,EACrC,IAAI,cAAE,OAAO;AAAA,EACb,OAAO,cAAE,MAAM,kBAAkB;AAAA,EACjC,SAAS;AACX,CAAC;;;AChFD,IAAAC,cAAkB;AAMX,IAAK,wBAAL,kBAAKC,2BAAL;AACL,EAAAA,uBAAA,YAAS;AACT,EAAAA,uBAAA,YAAS;AACT,EAAAA,uBAAA,aAAU;AACV,EAAAA,uBAAA,WAAQ;AAJE,SAAAA;AAAA,GAAA;AAUL,IAAM,wBAAwB,cAAE,OAAO;AAAA,EAC5C,IAAI,cAAE,OAAO;AAAA,EACb,aAAa,cAAE,OAAO;AAAA,EACtB,eAAe,cAAE,OAAO;AAAA,EACxB,eAAe,cAAE,OAAO;AAAA,EACxB,QAAQ,cAAE,KAAK,qBAAqB;AAAA,EACpC,SAAS,cAAE,OAAO,EAAE,SAAS;AAAA,EAC7B,UAAU,cAAE,OAAO,EAAE,SAAS;AAAA,EAC9B,QAAQ,cAAE,OAAO,EAAE,SAAS;AAAA,EAC5B,UAAU,cAAE,OAAO,EAAE,SAAS;AAAA,EAC9B,SAAS,cAAE,OAAO,cAAE,OAAO,GAAG,cAAE,QAAQ,CAAC,EAAE,SAAS;AAAA,EACpD,eAAe,cAAE,MAAM,kBAAkB,EAAE,SAAS;AACtD,CAAC;;;AC1BD,IAAAC,iBAA2B;;;ACK3B,oBAA2B;;;ACyCpB,IAAe,qBAAf,MAAmE;AAQ1E;;;AD1CA,SAAS,6BACP,MACA,mBACS;AACT,MAAI,SAAS,QAAQ,OAAO,SAAS,UAAU;AAC7C,WAAO;AAAA,EACT;AACA,QAAM,MAAM;AACZ,QAAM,OAAO,IAAI;AACjB,SAAO,OAAO,SAAS,YAAY,SAAS;AAC9C;AAMO,IAAM,0BAAN,cAAsC,mBAA4C;AAAA,EAC9E,OAAO;AAAA,EAEhB,SACE,WACA,OAEA,UACiB;AACjB,UAAM,kBAAc,0BAAW;AAE/B,UAAM,WAAiC,MAAM;AAC7C,QAAI,CAAC,UAAU,OAAO,QAAQ;AAC5B,aAAO;AAAA,QACL,QAAI,0BAAW;AAAA,QACf;AAAA,QACA,eAAe;AAAA,QACf,eAAe;AAAA,QACf;AAAA,QACA,SAAS;AAAA,QACT,UAAU,UAAU;AAAA,MACtB;AAAA,IACF;AAEA,eAAW,QAAQ,SAAS,OAAO;AACjC,UAAI,KAAK,aAAa,SAAS;AAC7B;AAAA,MACF;AACA,UAAI;AACJ,UAAI;AACF,eAAO,KAAK,gBACP,KAAK,MAAM,KAAK,aAAa,IAC9B;AAAA,MACN,QAAQ;AACN;AAAA,MACF;AACA,UAAI,6BAA6B,MAAM,UAAU,SAAS,GAAG;AAC3D,eAAO;AAAA,UACL,QAAI,0BAAW;AAAA,UACf;AAAA,UACA,eAAe;AAAA,UACf,eAAe;AAAA,UACf;AAAA,UACA,SAAS,UAAU,UAAU,SAAS;AAAA,UACtC,UAAU,UAAU;AAAA,QACtB;AAAA,MACF;AAAA,IACF;AAEA,WAAO;AAAA,MACL,QAAI,0BAAW;AAAA,MACf;AAAA,MACA,eAAe;AAAA,MACf,eAAe;AAAA,MACf;AAAA,MACA,SAAS,UAAU,UAAU,SAAS;AAAA,MACtC,UAAU,UAAU;AAAA,IACtB;AAAA,EACF;AACF;;;AEnFA,IAAAC,iBAA2B;AAC3B,2BAAyB;AAIzB,IAAM,kBAAkB;AACxB,IAAM,oBAAoB;AAMnB,IAAM,uBAAN,cAAmC,mBAAyC;AAAA,EACxE,OAAO;AAAA,EAEhB,SACE,WACA,QACA,SACiB;AACjB,UAAM,kBAAc,2BAAW;AAC/B,UAAM,UAAU,SAAS;AACzB,UAAM,UAAU,UAAU,WAAW;AACrC,UAAM,mBAAmB,UAAU,oBAAoB;AAEvD,QAAI,CAAC,SAAS;AACZ,aAAO,KAAK,aAAa,aAAa;AAAA,QACpC;AAAA,QACA,SAAS;AAAA,QACT,UAAU,OAAO,gBAAgB;AAAA,MACnC,CAAC;AAAA,IACH;AAEA,QAAI,WAA0B;AAC9B,QAAI,eAA8B;AAClC,QAAI;AACJ,QAAI;AAEJ,YAAQ,IAAI,2BAA2B,OAAO,SAAS,OAAO,EAAE;AAEhE,QAAI;AACF,yCAAS,SAAS;AAAA,QAChB,KAAK;AAAA,QACL,UAAU;AAAA,QACV,OAAO,CAAC,UAAU,QAAQ,MAAM;AAAA,MAClC,CAAC;AACD,iBAAW;AAAA,IACb,SAAS,KAAK;AACZ,YAAM,QAAQ;AAMd,iBACE,OAAO,MAAM,WAAW,WACpB,MAAM,SACN,OAAO,MAAM,SAAS,WACpB,MAAM,OACN;AACR,qBAAe,MAAM;AACrB,eAAS,KAAK,eAAe,MAAM,MAAM;AACzC,eAAS,KAAK,eAAe,MAAM,MAAM;AAAA,IAC3C;AAEA,UAAM,SAAS,aAAa,QAAQ,aAAa;AAEjD,UAAM,UAAmC,EAAE,SAAS,QAAQ;AAC5D,QAAI,WAAW,UAAa,WAAW,IAAI;AACzC,cAAQ,SAAS;AAAA,IACnB;AACA,QAAI,WAAW,UAAa,WAAW,IAAI;AACzC,cAAQ,SAAS;AAAA,IACnB;AAEA,WAAO,KAAK,aAAa,aAAa;AAAA,MACpC,QAAQ;AAAA,MAGR,SAAS,KAAK,cAAc,UAAU,kBAAkB,YAAY;AAAA,MACpE,UAAU,OAAO,gBAAgB;AAAA,MACjC,QAAQ,aAAa,OAAO,OAAO,QAAQ,IAAI;AAAA,MAC/C;AAAA,IACF,CAAC;AAAA,EACH;AAAA,EAEQ,aACN,aACA,QACiB;AACjB,WAAO;AAAA,MACL,QAAI,2BAAW;AAAA,MACf;AAAA,MACA,eAAe;AAAA,MACf,eAAe;AAAA,MACf;AAAA,MACA,GAAG;AAAA,IACL;AAAA,EACF;AAAA,EAEQ,eACN,OACoB;AACpB,QAAI,UAAU,UAAa,UAAU,KAAM,QAAO;AAClD,QAAI,OAAO,UAAU,SAAU,QAAO;AACtC,WAAO,MAAM,SAAS,OAAO;AAAA,EAC/B;AAAA,EAEQ,cACN,UACA,kBACA,cACQ;AACR,QAAI,aAAa,MAAM;AACrB,aAAO,iBAAiB,YAAY;AAAA,IACtC;AACA,QAAI,aAAa,kBAAkB;AACjC,aAAO,2BAA2B,QAAQ;AAAA,IAC5C;AACA,WAAO,qBAAqB,QAAQ,cAAc,gBAAgB;AAAA,EACpE;AACF;;;ACxHA,IAAAC,iBAA2B;AAC3B,uBAAgC;AAChC,gBAA6B;AAatB,SAAS,oBAAoB,UAAwC;AAC1E,MAAI,CAAC,UAAU,OAAO,QAAQ;AAC5B,WAAO;AAAA,EACT;AACA,QAAM,QAAkB,CAAC;AACzB,aAAW,QAAQ,SAAS,OAAO;AACjC,UAAM,QAAkB;AAAA,MACtB,QAAQ,KAAK,UAAU;AAAA,MACvB,SAAS,KAAK,IAAI;AAAA,MAClB,aAAa,KAAK,UAAU;AAAA,IAC9B;AACA,QAAI,KAAK,UAAU;AACjB,YAAM,KAAK,SAAS,KAAK,QAAQ,EAAE;AACnC,UAAI,KAAK,eAAe;AACtB,cAAM,KAAK,SAAS,KAAK,aAAa,EAAE;AAAA,MAC1C;AAAA,IACF;AACA,QAAI,KAAK,eAAe;AACtB,YAAM,KAAK,WAAW,KAAK,aAAa,EAAE;AAAA,IAC5C;AACA,QAAI,KAAK,OAAO;AACd,YAAM,KAAK,UAAU,KAAK,KAAK,EAAE;AAAA,IACnC;AACA,UAAM,KAAK,MAAM,KAAK,IAAI,CAAC;AAAA,EAC7B;AACA,SAAO,MAAM,KAAK,IAAI;AACxB;AAcO,SAAS,oBACd,KACA,KACQ;AACR,SAAO,IACJ,QAAQ,mBAAmB,IAAI,MAAM,EACrC,QAAQ,gBAAgB,IAAI,GAAG,EAC/B,QAAQ,yBAAyB,IAAI,YAAY,EACjD,QAAQ,0BAA0B,IAAI,aAAa,EACnD,QAAQ,qBAAqB,IAAI,QAAQ,EACzC,QAAQ,kBAAkB,IAAI,KAAK;AACxC;AAEO,SAAS,oBAAoB,QAA8B;AAChE,MAAI,WAAW,QAAQ,OAAO,WAAW,UAAU;AACjD,UAAM,IAAI,MAAM,+BAA+B;AAAA,EACjD;AACA,QAAM,MAAM;AACZ,MAAI,OAAO,IAAI,SAAS,UAAU;AAChC,UAAM,IAAI,MAAM,kDAAkD;AAAA,EACpE;AACA,MAAI,OAAO,IAAI,UAAU,UAAU;AACjC,UAAM,IAAI,MAAM,mDAAmD;AAAA,EACrE;AACA,MAAI,IAAI,QAAQ,KAAK,IAAI,QAAQ,KAAK;AACpC,UAAM,IAAI,MAAM,6CAA6C;AAAA,EAC/D;AACA,MAAI,OAAO,IAAI,mBAAmB,UAAU;AAC1C,UAAM,IAAI;AAAA,MACR;AAAA,IACF;AAAA,EACF;AACA,SAAO;AAAA,IACL,MAAM,IAAI;AAAA,IACV,OAAO,IAAI;AAAA,IACX,gBAAgB,IAAI;AAAA,EACtB;AACF;AAEA,IAAM,oBAAoB;AAC1B,IAAM,gBAAgB;AAGtB,IAAM,gCAAgC;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAUtC,IAAM,wBAAwB;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAwBvB,IAAM,oBAAN,cAAgC,mBAAsC;AAAA,EAClE,OAAO;AAAA,EAEhB,MAAM,SACJ,WACA,OACA,SAC0B;AAC1B,UAAM,kBAAc,2BAAW;AAC/B,UAAM,YAAY,SAAS;AAC3B,UAAM,UAAU,SAAS,WAAW;AACpC,UAAM,mBAAmB,SAAS;AAElC,UAAM,SAAS,MAAM,cAAc;AACnC,UAAM,YAAY,MAAM,aAAa,CAAC;AAGtC,UAAM,eAAe,UAAU,IAAI,CAAC,MAAM,EAAE,IAAI;AAChD,UAAM,gBAAgB,UACnB,OAAO,CAAC,MAAM,EAAE,WAAW,UAAU,EACrC,IAAI,CAAC,MAAM,EAAE,IAAI;AACpB,UAAM,WAAW,UACd,OAAO,CAAC,MAAM,EAAE,WAAW,KAAK,EAChC,IAAI,CAAC,MAAM,EAAE,IAAI;AAEpB,UAAM,eACJ,aAAa,SAAS,IAClB,aAAa,IAAI,CAAC,SAAiB,KAAK,IAAI,EAAE,EAAE,KAAK,IAAI,IACzD;AACN,UAAM,gBACJ,cAAc,SAAS,IACnB,cAAc,IAAI,CAAC,SAAiB,KAAK,IAAI,EAAE,EAAE,KAAK,IAAI,IAC1D;AACN,UAAM,WACJ,SAAS,SAAS,IACd,SAAS,IAAI,CAAC,SAAiB,KAAK,IAAI,EAAE,EAAE,KAAK,IAAI,IACrD;AAEN,UAAM,QAAQ,oBAAoB,MAAM,QAAQ;AAChD,UAAM,MAA0B;AAAA,MAC9B;AAAA,MACA,KAAK;AAAA,MACL;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF;AACA,UAAM,UAAU,CAAC,MAAc,oBAAoB,GAAG,GAAG;AAEzD,UAAM,cAAc,QAAQ,UAAU,MAAM;AAC5C,UAAM,eACJ,UAAU,gBAAgB,QAAQ,UAAU,iBAAiB,KACzD,QAAQ,UAAU,YAAY,IAAI,SAAS,wBAC3C,QAAQ,6BAA6B,IACrC,SACA;AAEN,UAAM,WAAW,UAAU,YAAY;AACvC,UAAM,kBAAkB,UAAU,aAAa;AAC/C,UAAM,cAAc,UAAU,eAAe;AAE7C,QAAI,CAAC,oBAAoB,CAAC,WAAW;AACnC,aAAO;AAAA,QACL,QAAI,2BAAW;AAAA,QACf;AAAA,QACA,eAAe;AAAA,QACf,eAAe;AAAA,QACf;AAAA,QACA,SAAS;AAAA,QACT,UAAU,OAAO,QAAQ;AAAA,MAC3B;AAAA,IACF;AAEA,QAAI;AACF,YAAM,SAAS,mBACX,MAAM,iBAAiB;AAAA,QACrB,QAAQ;AAAA,QACR,QAAQ;AAAA,QACR;AAAA,QACA;AAAA,MACF,CAAC,IACD,MAAM,KAAK;AAAA,QACT;AAAA,QACA,UAAU,SAAS;AAAA,QACnB;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,MACF;AAEJ,UAAI;AACJ,UAAI;AACF,cAAM,SAAS,KAAK,MAAM,OAAO,IAAI;AACrC,sBAAc,oBAAoB,MAAM;AAAA,MAC1C,SAAS,UAAU;AACjB,eAAO;AAAA,UACL,QAAI,2BAAW;AAAA,UACf;AAAA,UACA,eAAe;AAAA,UACf,eAAe;AAAA,UACf;AAAA,UACA,SAAS,mCAAmC,oBAAoB,QAAQ,SAAS,UAAU,OAAO,QAAQ,CAAC;AAAA,UAC3G,UAAU,OAAO,QAAQ;AAAA,UACzB,QAAQ;AAAA,UACR,SAAS,EAAE,SAAS,OAAO,KAAK,MAAM,GAAG,GAAG,EAAE;AAAA,QAChD;AAAA,MACF;AAEA,YAAM,SAAS,YAAY,SAAS;AACpC,aAAO;AAAA,QACL,QAAI,2BAAW;AAAA,QACf;AAAA,QACA,eAAe;AAAA,QACf,eAAe;AAAA,QACf,QAAQ;AAAA,QAGR,SAAS,SACL,eAAe,YAAY,KAAK,OAAO,QAAQ,KAAK,YAAY,IAAI,KACpE,eAAe,YAAY,KAAK,MAAM,QAAQ,KAAK,YAAY,IAAI;AAAA,QACvE,UAAU,OAAO,QAAQ;AAAA,QACzB,QAAQ,OAAO,YAAY,KAAK;AAAA,QAChC,SAAS;AAAA,UACP,OAAO,YAAY;AAAA,UACnB,gBAAgB,YAAY;AAAA,UAC5B,MAAM,YAAY;AAAA,QACpB;AAAA,MACF;AAAA,IACF,SAAS,KAAK;AACZ,YAAM,UAAU,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAC/D,aAAO;AAAA,QACL,QAAI,2BAAW;AAAA,QACf;AAAA,QACA,eAAe;AAAA,QACf,eAAe;AAAA,QACf;AAAA,QACA,SAAS,0BAA0B,OAAO;AAAA,QAC1C,UAAU,OAAO,QAAQ;AAAA,QACzB,SAAS,EAAE,OAAO,QAAQ;AAAA,MAC5B;AAAA,IACF;AAAA,EACF;AAAA,EAEA,MAAc,iBACZ,WACA,SACA,QACA,QACA,iBACA,aAC2B;AAC3B,UAAM,gBAAY,kCAAgB;AAAA,MAChC,SAAS,UAAU;AAAA,MACnB,QAAQ;AAAA,MACR,SAAS,UAAU;AAAA,IACrB,CAAC;AACD,UAAM,SAAS,UAAM,wBAAa;AAAA,MAChC,OAAO,UAAU,OAAO;AAAA,MACxB;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF,CAAC;AACD,WAAO,EAAE,MAAM,OAAO,KAAK;AAAA,EAC7B;AACF;;;AJtSA,IAAM,oBAAoB,IAAI,kBAAkB;AAEhD,IAAM,aAAiD;AAAA,EACrD,kBAAkB,IAAI,wBAAwB;AAAA,EAC9C,cAAc,IAAI,qBAAqB;AAAA,EACvC,WAAW;AAAA;AAAA,EAEX,QAAQ;AACV;AAQO,SAAS,kBACd,MACA,WACM;AACN,aAAW,IAAI,IAAI;AACrB;AAQO,SAAS,aAAa,MAA8C;AACzE,SAAO,WAAW,IAAI;AACxB;AAUA,eAAsB,mBACpB,OACA,YACA,SAC4B;AAC5B,MAAI,WAAW,WAAW,GAAG;AAC3B,WAAO,CAAC;AAAA,EACV;AACA,SAAO,QAAQ;AAAA,IACb,WAAW,IAAI,OAAO,cAAc;AAClC,YAAM,YAAY,WAAW,UAAU,IAAI;AAC3C,UAAI,CAAC,WAAW;AACd,eAAO;AAAA,UACL,QAAI,2BAAW;AAAA,UACf,iBAAa,2BAAW;AAAA,UACxB,eAAe,UAAU;AAAA,UACzB,eAAe;AAAA,UACf;AAAA,UACA,SAAS,+BAA+B,UAAU,IAAI;AAAA,UACtD,UAAU;AAAA,QACZ;AAAA,MACF;AACA,YAAM,UAAU,KAAK,IAAI;AACzB,YAAM,SAAS,MAAM,UAAU,SAAS,WAAW,OAAO,OAAO;AACjE,YAAM,aAAa,KAAK,IAAI,IAAI;AAChC,aAAO,EAAE,GAAG,QAAQ,UAAU,WAAW;AAAA,IAC3C,CAAC;AAAA,EACH;AACF;",
+  "sourcesContent": ["/**\n * @wix/eval-assertions\n *\n * Assertion framework for AI agent evaluations.\n * Supports skill invocation checks, build validation, and LLM-based judging.\n */\n\n// Types\nexport {\n  // Assertion schemas and types\n  AssertionSchema,\n  SkillWasCalledAssertionSchema,\n  BuildPassedAssertionSchema,\n  LlmJudgeAssertionSchema,\n  type Assertion,\n  type SkillWasCalledAssertion,\n  type BuildPassedAssertion,\n  type LlmJudgeAssertion,\n  // Trace types\n  LLMTraceSchema,\n  LLMTraceStepSchema,\n  LLMTraceSummarySchema,\n  LLMBreakdownStatsSchema,\n  TokenUsageSchema,\n  LLMStepType,\n  type LLMTrace,\n  type LLMTraceStep,\n  type LLMTraceSummary,\n  type LLMBreakdownStats,\n  type TokenUsage,\n  // Result types\n  AssertionResultSchema,\n  AssertionResultStatus,\n  type AssertionResult,\n  // Input types\n  type EvaluationInput,\n  type FileDiff,\n} from \"./types/index.js\";\n\n// Evaluators\nexport {\n  evaluateAssertions,\n  registerEvaluator,\n  getEvaluator,\n  AssertionEvaluator,\n  SkillWasCalledEvaluator,\n  BuildPassedEvaluator,\n  LlmJudgeEvaluator,\n  formatTraceForJudge,\n  replacePlaceholders,\n  validateJudgeResult,\n  type AssertionContext,\n  type LlmConfig,\n  type LlmJudgeGenerateTextOptions,\n  type JudgeResult,\n} from \"./evaluators/index.js\";\n", "import { z } from \"zod\";\n\n/**\n * Assertion: the agent must have invoked a specific skill during the run.\n * Checked by inspecting the LLM trace for a \"Skill\" tool use with the given skill.\n * Data: skillName (the skill that must have been called).\n */\nexport const SkillWasCalledAssertionSchema = z.object({\n  type: z.literal(\"skill_was_called\"),\n  /** Name of the skill that must have been called (matched against trace Skill tool args) */\n  skillName: z.string(),\n});\n\nexport type SkillWasCalledAssertion = z.infer<\n  typeof SkillWasCalledAssertionSchema\n>;\n\n/**\n * Assertion: a build command must exit with the expected code (default 0).\n * Runs the command in the scenario working directory.\n */\nexport const BuildPassedAssertionSchema = z.object({\n  type: z.literal(\"build_passed\"),\n  /** Command to run (default: \"yarn build\") */\n  command: z.string().optional(),\n  /** Expected exit code (default: 0) */\n  expectedExitCode: z.number().int().optional(),\n});\n\nexport type BuildPassedAssertion = z.infer<typeof BuildPassedAssertionSchema>;\n\n/**\n * Assertion: an LLM judges the scenario output (score 0-100).\n * Prompt can use {{output}}, {{cwd}}, {{changedFiles}}, {{modifiedFiles}}, {{newFiles}}, {{trace}}.\n * Passes if judge score >= minScore.\n */\nexport const LlmJudgeAssertionSchema = z.object({\n  type: z.literal(\"llm_judge\"),\n  /** Prompt template; placeholders: {{output}}, {{cwd}}, {{changedFiles}}, {{modifiedFiles}}, {{newFiles}}, {{trace}} */\n  prompt: z.string(),\n  /** Optional system prompt for the judge (default asks for JSON with score) */\n  systemPrompt: z.string().optional(),\n  /** Minimum score to pass (0-100, default 70) */\n  minScore: z.number().int().min(0).max(100).optional(),\n  /** Model for the judge (e.g. claude-3-5-haiku) */\n  model: z.string().optional(),\n  maxTokens: z.number().int().optional(),\n  temperature: z.number().min(0).max(1).optional(),\n});\n\nexport type LlmJudgeAssertion = z.infer<typeof LlmJudgeAssertionSchema>;\n\n/**\n * Union of all assertion types.\n * Each assertion has a type and type-specific data.\n * Uses z.union (not z.discriminatedUnion) for Zod v4 compatibility when used as array element.\n */\nexport const AssertionSchema = z.union([\n  SkillWasCalledAssertionSchema,\n  BuildPassedAssertionSchema,\n  LlmJudgeAssertionSchema,\n]);\n\nexport type Assertion = z.infer<typeof AssertionSchema>;\n", "import { z } from \"zod\";\n\n/**\n * Token usage schema.\n */\nexport const TokenUsageSchema = z.object({\n  prompt: z.number(),\n  completion: z.number(),\n  total: z.number(),\n});\n\nexport type TokenUsage = z.infer<typeof TokenUsageSchema>;\n\n/**\n * LLM step type enum.\n */\nexport enum LLMStepType {\n  COMPLETION = \"completion\",\n  TOOL_USE = \"tool_use\",\n  TOOL_RESULT = \"tool_result\",\n  THINKING = \"thinking\",\n}\n\n/**\n * LLM trace step schema.\n */\nexport const LLMTraceStepSchema = z.object({\n  id: z.string(),\n  stepNumber: z.number(),\n  type: z.enum(LLMStepType),\n  model: z.string(),\n  provider: z.string(),\n  startedAt: z.string(),\n  durationMs: z.number(),\n  tokenUsage: TokenUsageSchema,\n  costUsd: z.number(),\n  toolName: z.string().optional(),\n  toolArguments: z.string().optional(),\n  inputPreview: z.string().optional(),\n  outputPreview: z.string().optional(),\n  success: z.boolean(),\n  error: z.string().optional(),\n});\n\nexport type LLMTraceStep = z.infer<typeof LLMTraceStepSchema>;\n\n/**\n * LLM breakdown stats schema.\n */\nexport const LLMBreakdownStatsSchema = z.object({\n  count: z.number(),\n  durationMs: z.number(),\n  tokens: z.number(),\n  costUsd: z.number(),\n});\n\nexport type LLMBreakdownStats = z.infer<typeof LLMBreakdownStatsSchema>;\n\n/**\n * LLM trace summary schema.\n */\nexport const LLMTraceSummarySchema = z.object({\n  totalSteps: z.number(),\n  totalDurationMs: z.number(),\n  totalTokens: TokenUsageSchema,\n  totalCostUsd: z.number(),\n  stepTypeBreakdown: z.record(z.string(), LLMBreakdownStatsSchema).optional(),\n  modelBreakdown: z.record(z.string(), LLMBreakdownStatsSchema),\n  modelsUsed: z.array(z.string()),\n});\n\nexport type LLMTraceSummary = z.infer<typeof LLMTraceSummarySchema>;\n\n/**\n * LLM trace schema.\n */\nexport const LLMTraceSchema = z.object({\n  id: z.string(),\n  steps: z.array(LLMTraceStepSchema),\n  summary: LLMTraceSummarySchema,\n});\n\nexport type LLMTrace = z.infer<typeof LLMTraceSchema>;\n", "import { z } from \"zod\";\nimport { LLMTraceStepSchema } from \"./trace.js\";\n\n/**\n * Assertion result status enum.\n */\nexport enum AssertionResultStatus {\n  PASSED = \"passed\",\n  FAILED = \"failed\",\n  SKIPPED = \"skipped\",\n  ERROR = \"error\",\n}\n\n/**\n * Assertion result schema.\n */\nexport const AssertionResultSchema = z.object({\n  id: z.string(),\n  assertionId: z.string(),\n  assertionType: z.string(),\n  assertionName: z.string(),\n  status: z.enum(AssertionResultStatus),\n  message: z.string().optional(),\n  expected: z.string().optional(),\n  actual: z.string().optional(),\n  duration: z.number().optional(),\n  details: z.record(z.string(), z.unknown()).optional(),\n  llmTraceSteps: z.array(LLMTraceStepSchema).optional(),\n});\n\nexport type AssertionResult = z.infer<typeof AssertionResultSchema>;\n", "import type { Assertion, AssertionResult } from \"../types/index.js\";\nimport { AssertionResultStatus } from \"../types/index.js\";\nimport { randomUUID } from \"crypto\";\nimport type { AssertionContext } from \"./assertion-evaluator.js\";\nimport { AssertionEvaluator } from \"./assertion-evaluator.js\";\nimport { SkillWasCalledEvaluator } from \"./skill-was-called-evaluator.js\";\nimport { BuildPassedEvaluator } from \"./build-passed-evaluator.js\";\nimport { LlmJudgeEvaluator } from \"./llm-judge-evaluator.js\";\nimport type { EvaluationInput } from \"../types/index.js\";\n\nconst llmJudgeEvaluator = new LlmJudgeEvaluator();\n\nconst evaluators: Record<string, AssertionEvaluator> = {\n  skill_was_called: new SkillWasCalledEvaluator(),\n  build_passed: new BuildPassedEvaluator(),\n  llm_judge: llmJudgeEvaluator,\n  // Custom assertions use the same LLM-based evaluation as llm_judge\n  custom: llmJudgeEvaluator,\n};\n\n/**\n * Register a custom assertion evaluator.\n *\n * @param type - The assertion type identifier\n * @param evaluator - The evaluator instance\n */\nexport function registerEvaluator(\n  type: string,\n  evaluator: AssertionEvaluator,\n): void {\n  evaluators[type] = evaluator;\n}\n\n/**\n * Get a registered evaluator by type.\n *\n * @param type - The assertion type identifier\n * @returns The evaluator or undefined if not found\n */\nexport function getEvaluator(type: string): AssertionEvaluator | undefined {\n  return evaluators[type];\n}\n\n/**\n * Evaluate all assertions against the input.\n *\n * @param input - Evaluation input (includes outputText, llmTrace, fileDiffs)\n * @param assertions - List of assertions to evaluate\n * @param context - Optional context (e.g. workDir for build_passed, llmConfig for llm_judge)\n * @returns Array of assertion results; empty if no assertions\n */\nexport async function evaluateAssertions(\n  input: EvaluationInput,\n  assertions: Assertion[],\n  context?: AssertionContext,\n): Promise<AssertionResult[]> {\n  if (assertions.length === 0) {\n    return [];\n  }\n  return Promise.all(\n    assertions.map(async (assertion) => {\n      const evaluator = evaluators[assertion.type];\n      if (!evaluator) {\n        return {\n          id: randomUUID(),\n          assertionId: randomUUID(),\n          assertionType: assertion.type,\n          assertionName: \"Unknown assertion\",\n          status: AssertionResultStatus.ERROR,\n          message: `Unsupported assertion type: ${assertion.type}`,\n          duration: 0,\n        };\n      }\n      const startMs = Date.now();\n      const result = await evaluator.evaluate(assertion, input, context);\n      const durationMs = Date.now() - startMs;\n      return { ...result, duration: durationMs };\n    }),\n  );\n}\n\n// Re-export evaluator classes and types\nexport { AssertionEvaluator } from \"./assertion-evaluator.js\";\nexport type {\n  AssertionContext,\n  LlmConfig,\n  LlmJudgeGenerateTextOptions,\n} from \"./assertion-evaluator.js\";\nexport { SkillWasCalledEvaluator } from \"./skill-was-called-evaluator.js\";\nexport { BuildPassedEvaluator } from \"./build-passed-evaluator.js\";\nexport {\n  LlmJudgeEvaluator,\n  formatTraceForJudge,\n  replacePlaceholders,\n  validateJudgeResult,\n  type JudgeResult,\n} from \"./llm-judge-evaluator.js\";\n", "import type {\n  SkillWasCalledAssertion,\n  AssertionResult,\n  LLMTrace,\n  EvaluationInput,\n} from \"../types/index.js\";\nimport { AssertionResultStatus } from \"../types/index.js\";\nimport { randomUUID } from \"crypto\";\nimport type { AssertionContext } from \"./assertion-evaluator.js\";\nimport { AssertionEvaluator } from \"./assertion-evaluator.js\";\n\n/**\n * Check if parsed Skill tool args reference the expected skill by name.\n */\nfunction toolArgsReferenceSkillByName(\n  args: unknown,\n  expectedSkillName: string,\n): boolean {\n  if (args === null || typeof args !== \"object\") {\n    return false;\n  }\n  const obj = args as Record<string, unknown>;\n  const name = obj.skill;\n  return typeof name === \"string\" && name === expectedSkillName;\n}\n\n/**\n * Evaluator for \"skill_was_called\" assertion: the LLM trace must contain a step\n * where the \"Skill\" tool was used with the expected skill (by name).\n */\nexport class SkillWasCalledEvaluator extends AssertionEvaluator<SkillWasCalledAssertion> {\n  readonly type = \"skill_was_called\" as const;\n\n  evaluate(\n    assertion: SkillWasCalledAssertion,\n    input: EvaluationInput,\n    // eslint-disable-next-line @typescript-eslint/no-unused-vars -- context not used for skill_was_called\n    _context?: AssertionContext,\n  ): AssertionResult {\n    const assertionId = randomUUID();\n\n    const llmTrace: LLMTrace | undefined = input.llmTrace;\n    if (!llmTrace?.steps?.length) {\n      return {\n        id: randomUUID(),\n        assertionId,\n        assertionType: \"skill_was_called\",\n        assertionName: \"Skill was called\",\n        status: AssertionResultStatus.FAILED,\n        message: \"No LLM trace steps to check for skill invocation\",\n        expected: assertion.skillName,\n      };\n    }\n\n    for (const step of llmTrace.steps) {\n      if (step.toolName !== \"Skill\") {\n        continue;\n      }\n      let args: unknown;\n      try {\n        args = step.toolArguments\n          ? (JSON.parse(step.toolArguments) as unknown)\n          : undefined;\n      } catch {\n        continue;\n      }\n      if (toolArgsReferenceSkillByName(args, assertion.skillName)) {\n        return {\n          id: randomUUID(),\n          assertionId,\n          assertionType: \"skill_was_called\",\n          assertionName: \"Skill was called\",\n          status: AssertionResultStatus.PASSED,\n          message: `Skill \"${assertion.skillName}\" was called`,\n          expected: assertion.skillName,\n        };\n      }\n    }\n\n    return {\n      id: randomUUID(),\n      assertionId,\n      assertionType: \"skill_was_called\",\n      assertionName: \"Skill was called\",\n      status: AssertionResultStatus.FAILED,\n      message: `Skill \"${assertion.skillName}\" was not called`,\n      expected: assertion.skillName,\n    };\n  }\n}\n", "import type {\n  Assertion,\n  AssertionResult,\n  EvaluationInput,\n} from \"../types/index.js\";\n\n/**\n * Options passed to the LLM for llm_judge. Used by the optional stub for testing.\n */\nexport interface LlmJudgeGenerateTextOptions {\n  prompt: string;\n  system: string;\n  maxOutputTokens: number;\n  temperature: number;\n}\n\n/**\n * Configuration for LLM calls (used by llm_judge assertion).\n */\nexport interface LlmConfig {\n  /** Base URL for the AI API (e.g., 'https://api.anthropic.com') */\n  baseUrl: string;\n  /** Headers to include in API requests (e.g., API key) */\n  headers: Record<string, string>;\n}\n\n/**\n * Optional context passed when evaluating assertions.\n */\nexport interface AssertionContext {\n  /** Working directory for the scenario (used by build_passed) */\n  workDir?: string;\n  /** LLM configuration (used by llm_judge) */\n  llmConfig?: LlmConfig;\n  /**\n   * Optional stub for llm_judge: when set, the evaluator uses this instead of the real AI call.\n   * Used only in tests to avoid hitting the API.\n   */\n  generateTextForLlmJudge?: (\n    options: LlmJudgeGenerateTextOptions,\n  ) => Promise<{ text: string }>;\n}\n\n/**\n * Abstract base for assertion evaluators.\n * Each assertion type has a concrete class that implements evaluate().\n * evaluate() may return a Promise for async assertions (e.g. llm_judge).\n */\nexport abstract class AssertionEvaluator<T extends Assertion = Assertion> {\n  abstract readonly type: T[\"type\"];\n\n  abstract evaluate(\n    assertion: T,\n    input: EvaluationInput,\n    context?: AssertionContext,\n  ): AssertionResult | Promise<AssertionResult>;\n}\n", "import type {\n  BuildPassedAssertion,\n  AssertionResult,\n  EvaluationInput,\n} from \"../types/index.js\";\nimport { AssertionResultStatus } from \"../types/index.js\";\nimport { randomUUID } from \"crypto\";\nimport { execSync } from \"child_process\";\nimport type { AssertionContext } from \"./assertion-evaluator.js\";\nimport { AssertionEvaluator } from \"./assertion-evaluator.js\";\n\nconst DEFAULT_COMMAND = \"yarn build\";\nconst DEFAULT_EXIT_CODE = 0;\n\n/**\n * Evaluator for \"build_passed\" assertion: runs a build command in the scenario\n * working directory and passes if the command exits with the expected code (default 0).\n */\nexport class BuildPassedEvaluator extends AssertionEvaluator<BuildPassedAssertion> {\n  readonly type = \"build_passed\" as const;\n\n  evaluate(\n    assertion: BuildPassedAssertion,\n    _input: EvaluationInput,\n    context?: AssertionContext,\n  ): AssertionResult {\n    const assertionId = randomUUID();\n    const workDir = context?.workDir;\n    const command = assertion.command ?? DEFAULT_COMMAND;\n    const expectedExitCode = assertion.expectedExitCode ?? DEFAULT_EXIT_CODE;\n\n    if (!workDir) {\n      return this.createResult(assertionId, {\n        status: AssertionResultStatus.FAILED,\n        message: \"No working directory provided for build_passed assertion\",\n        expected: String(expectedExitCode),\n      });\n    }\n\n    let exitCode: number | null = null;\n    let errorMessage: string | null = null;\n    let stdout: string | undefined;\n    let stderr: string | undefined;\n\n    console.log(`[build_passed] Running \"${command}\" in: ${workDir}`);\n\n    try {\n      execSync(command, {\n        cwd: workDir,\n        encoding: \"utf-8\",\n        stdio: [\"ignore\", \"pipe\", \"pipe\"],\n      });\n      exitCode = 0;\n    } catch (err) {\n      const error = err as Error & {\n        status?: number;\n        code?: number;\n        stdout?: string | Buffer;\n        stderr?: string | Buffer;\n      };\n      exitCode =\n        typeof error.status === \"number\"\n          ? error.status\n          : typeof error.code === \"number\"\n            ? error.code\n            : null;\n      errorMessage = error.message;\n      stdout = this.bufferToString(error.stdout);\n      stderr = this.bufferToString(error.stderr);\n    }\n\n    const passed = exitCode !== null && exitCode === expectedExitCode;\n\n    const details: Record<string, unknown> = { workDir, command };\n    if (stdout !== undefined && stdout !== \"\") {\n      details.stdout = stdout;\n    }\n    if (stderr !== undefined && stderr !== \"\") {\n      details.stderr = stderr;\n    }\n\n    return this.createResult(assertionId, {\n      status: passed\n        ? AssertionResultStatus.PASSED\n        : AssertionResultStatus.FAILED,\n      message: this.formatMessage(exitCode, expectedExitCode, errorMessage),\n      expected: String(expectedExitCode),\n      actual: exitCode !== null ? String(exitCode) : undefined,\n      details,\n    });\n  }\n\n  private createResult(\n    assertionId: string,\n    fields: Partial<AssertionResult>,\n  ): AssertionResult {\n    return {\n      id: randomUUID(),\n      assertionId,\n      assertionType: \"build_passed\",\n      assertionName: \"Build passed\",\n      status: AssertionResultStatus.FAILED,\n      ...fields,\n    };\n  }\n\n  private bufferToString(\n    value: string | Buffer | undefined,\n  ): string | undefined {\n    if (value === undefined || value === null) return undefined;\n    if (typeof value === \"string\") return value;\n    return value.toString(\"utf-8\");\n  }\n\n  private formatMessage(\n    exitCode: number | null,\n    expectedExitCode: number,\n    errorMessage: string | null,\n  ): string {\n    if (exitCode === null) {\n      return `Build failed: ${errorMessage}`;\n    }\n    if (exitCode === expectedExitCode) {\n      return `Build passed (exit code ${exitCode})`;\n    }\n    return `Build exited with ${exitCode}, expected ${expectedExitCode}`;\n  }\n}\n", "import type {\n  LlmJudgeAssertion,\n  AssertionResult,\n  LLMTrace,\n  EvaluationInput,\n} from \"../types/index.js\";\nimport { AssertionResultStatus } from \"../types/index.js\";\nimport { randomUUID } from \"crypto\";\nimport { createAnthropic } from \"@ai-sdk/anthropic\";\nimport { generateText } from \"ai\";\nimport type { AssertionContext, LlmConfig } from \"./assertion-evaluator.js\";\nimport { AssertionEvaluator } from \"./assertion-evaluator.js\";\n\nexport interface JudgeResult {\n  text: string;\n  score: number;\n  scoreReasoning: string;\n}\n\n/**\n * Format LLM trace as readable text for the judge (step number, type, tool name/args, output preview).\n */\nexport function formatTraceForJudge(llmTrace: LLMTrace | undefined): string {\n  if (!llmTrace?.steps?.length) {\n    return \"No trace available.\";\n  }\n  const lines: string[] = [];\n  for (const step of llmTrace.steps) {\n    const parts: string[] = [\n      `Step ${step.stepNumber}`,\n      `type: ${step.type}`,\n      `duration: ${step.durationMs}ms`,\n    ];\n    if (step.toolName) {\n      parts.push(`tool: ${step.toolName}`);\n      if (step.toolArguments) {\n        parts.push(`args: ${step.toolArguments}`);\n      }\n    }\n    if (step.outputPreview) {\n      parts.push(`output: ${step.outputPreview}`);\n    }\n    if (step.error) {\n      parts.push(`error: ${step.error}`);\n    }\n    lines.push(parts.join(\", \"));\n  }\n  return lines.join(\"\\n\");\n}\n\n/**\n * Context object for placeholder replacement.\n */\nexport interface PlaceholderContext {\n  output: string;\n  cwd: string;\n  changedFiles: string;\n  modifiedFiles: string;\n  newFiles: string;\n  trace: string;\n}\n\nexport function replacePlaceholders(\n  str: string,\n  ctx: PlaceholderContext,\n): string {\n  return str\n    .replace(/\\{\\{output\\}\\}/g, ctx.output)\n    .replace(/\\{\\{cwd\\}\\}/g, ctx.cwd)\n    .replace(/\\{\\{changedFiles\\}\\}/g, ctx.changedFiles)\n    .replace(/\\{\\{modifiedFiles\\}\\}/g, ctx.modifiedFiles)\n    .replace(/\\{\\{newFiles\\}\\}/g, ctx.newFiles)\n    .replace(/\\{\\{trace\\}\\}/g, ctx.trace);\n}\n\nexport function validateJudgeResult(parsed: unknown): JudgeResult {\n  if (parsed === null || typeof parsed !== \"object\") {\n    throw new Error(\"Judge result is not an object\");\n  }\n  const obj = parsed as Record<string, unknown>;\n  if (typeof obj.text !== \"string\") {\n    throw new Error(\"Judge result does not contain a valid text field\");\n  }\n  if (typeof obj.score !== \"number\") {\n    throw new Error(\"Judge result does not contain a valid score field\");\n  }\n  if (obj.score < 0 || obj.score > 100) {\n    throw new Error(\"Judge result score is not between 0 and 100\");\n  }\n  if (typeof obj.scoreReasoning !== \"string\") {\n    throw new Error(\n      \"Judge result does not contain a valid scoreReasoning field\",\n    );\n  }\n  return {\n    text: obj.text,\n    score: obj.score,\n    scoreReasoning: obj.scoreReasoning,\n  };\n}\n\nconst DEFAULT_MIN_SCORE = 70;\nconst DEFAULT_MODEL = \"claude-3-5-haiku-20241022\";\n\n/** Default system prompt context with placeholders; used when assertion.systemPrompt is empty. */\nconst DEFAULT_SYSTEM_PROMPT_CONTEXT = `You are judging a scenario run. The ACTUAL run data is provided below \u2014 use it to verify facts:\n\n- {{output}}: the agent's final output\n- {{cwd}}: working directory\n- {{changedFiles}}: list of all files changed (or \"No files were changed\")\n- {{modifiedFiles}}: list of existing files that were modified (or \"No files were modified\")\n- {{newFiles}}: list of new files that were created (or \"No new files were created\")\n- {{trace}}: step-by-step trace (tool calls, completions) so you can check e.g. which tools were called and how many times\n\nCRITICAL: When the user asks you to verify a specific fact, compare it strictly against the actual data above. If the expected outcome does NOT match the actual outcome, you MUST give a score of 0 or near 0. Do not be lenient \u2014 factual mismatches are failures.`;\n\nconst DEFAULT_SYSTEM_PROMPT = `You must respond only with a valid JSON object that conforms exactly to the following structure:\n\n{\n  \"text\": string,\n  \"score\": number (0-100),\n  \"scoreReasoning\": string\n}\n\n- text: A brief textual verdict of the test result.\n- score: A number from 0 to 100 that reflects how well the answer meets the acceptance criteria.\n- scoreReasoning: A concise explanation justifying the assigned score.\n\nYour response must:\n- Contain only the JSON object above \u2014 no introductory text, no code formatting (e.g., no triple backticks), and no trailing comments.\n- Be valid and parseable by \\`JSON.parse\\`.\n- Use only double quotes for all keys and strings, as required by JSON.\n\nAny response that includes extra content or deviates from the specified format will cause parsing to fail. Follow these instructions exactly.`;\n\n/**\n * Evaluator for \"llm_judge\" assertion: an LLM judges the scenario output\n * (prompt with {{output}}, {{cwd}}, {{changedFiles}}, {{trace}}) and returns a score 0-100.\n * Passes if score >= minScore.\n */\nexport class LlmJudgeEvaluator extends AssertionEvaluator<LlmJudgeAssertion> {\n  readonly type = \"llm_judge\" as const;\n\n  async evaluate(\n    assertion: LlmJudgeAssertion,\n    input: EvaluationInput,\n    context?: AssertionContext,\n  ): Promise<AssertionResult> {\n    const assertionId = randomUUID();\n    const llmConfig = context?.llmConfig;\n    const workDir = context?.workDir ?? \"\";\n    const generateTextStub = context?.generateTextForLlmJudge;\n\n    const output = input.outputText ?? \"\";\n    const fileDiffs = input.fileDiffs ?? [];\n\n    // Compute file lists by status\n    const changedPaths = fileDiffs.map((d) => d.path);\n    const modifiedPaths = fileDiffs\n      .filter((d) => d.status === \"modified\")\n      .map((d) => d.path);\n    const newPaths = fileDiffs\n      .filter((d) => d.status === \"new\")\n      .map((d) => d.path);\n\n    const changedFiles =\n      changedPaths.length > 0\n        ? changedPaths.map((path: string) => `- ${path}`).join(\"\\n\")\n        : \"No files were changed\";\n    const modifiedFiles =\n      modifiedPaths.length > 0\n        ? modifiedPaths.map((path: string) => `- ${path}`).join(\"\\n\")\n        : \"No files were modified\";\n    const newFiles =\n      newPaths.length > 0\n        ? newPaths.map((path: string) => `- ${path}`).join(\"\\n\")\n        : \"No new files were created\";\n\n    const trace = formatTraceForJudge(input.llmTrace);\n    const ctx: PlaceholderContext = {\n      output,\n      cwd: workDir,\n      changedFiles,\n      modifiedFiles,\n      newFiles,\n      trace,\n    };\n    const replace = (s: string) => replacePlaceholders(s, ctx);\n\n    const finalPrompt = replace(assertion.prompt);\n    const systemPrompt =\n      assertion.systemPrompt != null && assertion.systemPrompt !== \"\"\n        ? replace(assertion.systemPrompt) + \"\\n\\n\" + DEFAULT_SYSTEM_PROMPT\n        : replace(DEFAULT_SYSTEM_PROMPT_CONTEXT) +\n          \"\\n\\n\" +\n          DEFAULT_SYSTEM_PROMPT;\n\n    const minScore = assertion.minScore ?? DEFAULT_MIN_SCORE;\n    const maxOutputTokens = assertion.maxTokens ?? 1024;\n    const temperature = assertion.temperature ?? 0;\n\n    if (!generateTextStub && !llmConfig) {\n      return {\n        id: randomUUID(),\n        assertionId,\n        assertionType: \"llm_judge\",\n        assertionName: \"LLM judge\",\n        status: AssertionResultStatus.FAILED,\n        message: \"No llmConfig for llm_judge assertion (AI gateway required)\",\n        expected: String(minScore),\n      };\n    }\n\n    const maxParseAttempts = 3;\n    let lastParseError: Error | undefined;\n    let lastRawText: string | undefined;\n\n    try {\n      for (let attempt = 1; attempt <= maxParseAttempts; attempt++) {\n        const result = generateTextStub\n          ? await generateTextStub({\n              prompt: finalPrompt,\n              system: systemPrompt,\n              maxOutputTokens,\n              temperature,\n            })\n          : await this.callGenerateText(\n              llmConfig!,\n              assertion.model ?? DEFAULT_MODEL,\n              finalPrompt,\n              systemPrompt,\n              maxOutputTokens,\n              temperature,\n            );\n\n        lastRawText = result.text;\n        try {\n          const parsed = JSON.parse(result.text);\n          const judgeResult = validateJudgeResult(parsed);\n          const passed = judgeResult.score >= minScore;\n          return {\n            id: randomUUID(),\n            assertionId,\n            assertionType: \"llm_judge\",\n            assertionName: \"LLM judge\",\n            status: passed\n              ? AssertionResultStatus.PASSED\n              : AssertionResultStatus.FAILED,\n            message: passed\n              ? `Judge score ${judgeResult.score} >= ${minScore}: ${judgeResult.text}`\n              : `Judge score ${judgeResult.score} < ${minScore}: ${judgeResult.text}`,\n            expected: String(minScore),\n            actual: String(judgeResult.score),\n            details: {\n              score: judgeResult.score,\n              scoreReasoning: judgeResult.scoreReasoning,\n              text: judgeResult.text,\n            },\n          };\n        } catch (parseErr) {\n          lastParseError =\n            parseErr instanceof Error ? parseErr : new Error(String(parseErr));\n        }\n      }\n\n      return {\n        id: randomUUID(),\n        assertionId,\n        assertionType: \"llm_judge\",\n        assertionName: \"LLM judge\",\n        status: AssertionResultStatus.FAILED,\n        message: `Failed to parse judge response after ${maxParseAttempts} attempts: ${lastParseError?.message ?? \"unknown\"}`,\n        expected: String(minScore),\n        actual: undefined,\n        details: { rawText: lastRawText?.slice(0, 500) },\n      };\n    } catch (err) {\n      const message = err instanceof Error ? err.message : String(err);\n      return {\n        id: randomUUID(),\n        assertionId,\n        assertionType: \"llm_judge\",\n        assertionName: \"LLM judge\",\n        status: AssertionResultStatus.FAILED,\n        message: `LLM judge call failed: ${message}`,\n        expected: String(minScore),\n        details: { error: message },\n      };\n    }\n  }\n\n  private async callGenerateText(\n    llmConfig: LlmConfig,\n    modelId: string,\n    prompt: string,\n    system: string,\n    maxOutputTokens: number,\n    temperature: number,\n  ): Promise<{ text: string }> {\n    const anthropic = createAnthropic({\n      baseURL: llmConfig.baseUrl,\n      apiKey: \"dummy\",\n      headers: llmConfig.headers,\n    });\n    const result = await generateText({\n      model: anthropic(modelId),\n      prompt,\n      system,\n      maxOutputTokens,\n      temperature,\n    });\n    return { text: result.text };\n  }\n}\n"],
+  "mappings": ";;;;;;;;;;;;;;;;;;;;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;;;ACAA,iBAAkB;AAOX,IAAM,gCAAgC,aAAE,OAAO;AAAA,EACpD,MAAM,aAAE,QAAQ,kBAAkB;AAAA;AAAA,EAElC,WAAW,aAAE,OAAO;AACtB,CAAC;AAUM,IAAM,6BAA6B,aAAE,OAAO;AAAA,EACjD,MAAM,aAAE,QAAQ,cAAc;AAAA;AAAA,EAE9B,SAAS,aAAE,OAAO,EAAE,SAAS;AAAA;AAAA,EAE7B,kBAAkB,aAAE,OAAO,EAAE,IAAI,EAAE,SAAS;AAC9C,CAAC;AASM,IAAM,0BAA0B,aAAE,OAAO;AAAA,EAC9C,MAAM,aAAE,QAAQ,WAAW;AAAA;AAAA,EAE3B,QAAQ,aAAE,OAAO;AAAA;AAAA,EAEjB,cAAc,aAAE,OAAO,EAAE,SAAS;AAAA;AAAA,EAElC,UAAU,aAAE,OAAO,EAAE,IAAI,EAAE,IAAI,CAAC,EAAE,IAAI,GAAG,EAAE,SAAS;AAAA;AAAA,EAEpD,OAAO,aAAE,OAAO,EAAE,SAAS;AAAA,EAC3B,WAAW,aAAE,OAAO,EAAE,IAAI,EAAE,SAAS;AAAA,EACrC,aAAa,aAAE,OAAO,EAAE,IAAI,CAAC,EAAE,IAAI,CAAC,EAAE,SAAS;AACjD,CAAC;AASM,IAAM,kBAAkB,aAAE,MAAM;AAAA,EACrC;AAAA,EACA;AAAA,EACA;AACF,CAAC;;;AC7DD,IAAAA,cAAkB;AAKX,IAAM,mBAAmB,cAAE,OAAO;AAAA,EACvC,QAAQ,cAAE,OAAO;AAAA,EACjB,YAAY,cAAE,OAAO;AAAA,EACrB,OAAO,cAAE,OAAO;AAClB,CAAC;AAOM,IAAK,cAAL,kBAAKC,iBAAL;AACL,EAAAA,aAAA,gBAAa;AACb,EAAAA,aAAA,cAAW;AACX,EAAAA,aAAA,iBAAc;AACd,EAAAA,aAAA,cAAW;AAJD,SAAAA;AAAA,GAAA;AAUL,IAAM,qBAAqB,cAAE,OAAO;AAAA,EACzC,IAAI,cAAE,OAAO;AAAA,EACb,YAAY,cAAE,OAAO;AAAA,EACrB,MAAM,cAAE,KAAK,WAAW;AAAA,EACxB,OAAO,cAAE,OAAO;AAAA,EAChB,UAAU,cAAE,OAAO;AAAA,EACnB,WAAW,cAAE,OAAO;AAAA,EACpB,YAAY,cAAE,OAAO;AAAA,EACrB,YAAY;AAAA,EACZ,SAAS,cAAE,OAAO;AAAA,EAClB,UAAU,cAAE,OAAO,EAAE,SAAS;AAAA,EAC9B,eAAe,cAAE,OAAO,EAAE,SAAS;AAAA,EACnC,cAAc,cAAE,OAAO,EAAE,SAAS;AAAA,EAClC,eAAe,cAAE,OAAO,EAAE,SAAS;AAAA,EACnC,SAAS,cAAE,QAAQ;AAAA,EACnB,OAAO,cAAE,OAAO,EAAE,SAAS;AAC7B,CAAC;AAOM,IAAM,0BAA0B,cAAE,OAAO;AAAA,EAC9C,OAAO,cAAE,OAAO;AAAA,EAChB,YAAY,cAAE,OAAO;AAAA,EACrB,QAAQ,cAAE,OAAO;AAAA,EACjB,SAAS,cAAE,OAAO;AACpB,CAAC;AAOM,IAAM,wBAAwB,cAAE,OAAO;AAAA,EAC5C,YAAY,cAAE,OAAO;AAAA,EACrB,iBAAiB,cAAE,OAAO;AAAA,EAC1B,aAAa;AAAA,EACb,cAAc,cAAE,OAAO;AAAA,EACvB,mBAAmB,cAAE,OAAO,cAAE,OAAO,GAAG,uBAAuB,EAAE,SAAS;AAAA,EAC1E,gBAAgB,cAAE,OAAO,cAAE,OAAO,GAAG,uBAAuB;AAAA,EAC5D,YAAY,cAAE,MAAM,cAAE,OAAO,CAAC;AAChC,CAAC;AAOM,IAAM,iBAAiB,cAAE,OAAO;AAAA,EACrC,IAAI,cAAE,OAAO;AAAA,EACb,OAAO,cAAE,MAAM,kBAAkB;AAAA,EACjC,SAAS;AACX,CAAC;;;AChFD,IAAAC,cAAkB;AAMX,IAAK,wBAAL,kBAAKC,2BAAL;AACL,EAAAA,uBAAA,YAAS;AACT,EAAAA,uBAAA,YAAS;AACT,EAAAA,uBAAA,aAAU;AACV,EAAAA,uBAAA,WAAQ;AAJE,SAAAA;AAAA,GAAA;AAUL,IAAM,wBAAwB,cAAE,OAAO;AAAA,EAC5C,IAAI,cAAE,OAAO;AAAA,EACb,aAAa,cAAE,OAAO;AAAA,EACtB,eAAe,cAAE,OAAO;AAAA,EACxB,eAAe,cAAE,OAAO;AAAA,EACxB,QAAQ,cAAE,KAAK,qBAAqB;AAAA,EACpC,SAAS,cAAE,OAAO,EAAE,SAAS;AAAA,EAC7B,UAAU,cAAE,OAAO,EAAE,SAAS;AAAA,EAC9B,QAAQ,cAAE,OAAO,EAAE,SAAS;AAAA,EAC5B,UAAU,cAAE,OAAO,EAAE,SAAS;AAAA,EAC9B,SAAS,cAAE,OAAO,cAAE,OAAO,GAAG,cAAE,QAAQ,CAAC,EAAE,SAAS;AAAA,EACpD,eAAe,cAAE,MAAM,kBAAkB,EAAE,SAAS;AACtD,CAAC;;;AC1BD,IAAAC,iBAA2B;;;ACK3B,oBAA2B;;;ACyCpB,IAAe,qBAAf,MAAmE;AAQ1E;;;AD1CA,SAAS,6BACP,MACA,mBACS;AACT,MAAI,SAAS,QAAQ,OAAO,SAAS,UAAU;AAC7C,WAAO;AAAA,EACT;AACA,QAAM,MAAM;AACZ,QAAM,OAAO,IAAI;AACjB,SAAO,OAAO,SAAS,YAAY,SAAS;AAC9C;AAMO,IAAM,0BAAN,cAAsC,mBAA4C;AAAA,EAC9E,OAAO;AAAA,EAEhB,SACE,WACA,OAEA,UACiB;AACjB,UAAM,kBAAc,0BAAW;AAE/B,UAAM,WAAiC,MAAM;AAC7C,QAAI,CAAC,UAAU,OAAO,QAAQ;AAC5B,aAAO;AAAA,QACL,QAAI,0BAAW;AAAA,QACf;AAAA,QACA,eAAe;AAAA,QACf,eAAe;AAAA,QACf;AAAA,QACA,SAAS;AAAA,QACT,UAAU,UAAU;AAAA,MACtB;AAAA,IACF;AAEA,eAAW,QAAQ,SAAS,OAAO;AACjC,UAAI,KAAK,aAAa,SAAS;AAC7B;AAAA,MACF;AACA,UAAI;AACJ,UAAI;AACF,eAAO,KAAK,gBACP,KAAK,MAAM,KAAK,aAAa,IAC9B;AAAA,MACN,QAAQ;AACN;AAAA,MACF;AACA,UAAI,6BAA6B,MAAM,UAAU,SAAS,GAAG;AAC3D,eAAO;AAAA,UACL,QAAI,0BAAW;AAAA,UACf;AAAA,UACA,eAAe;AAAA,UACf,eAAe;AAAA,UACf;AAAA,UACA,SAAS,UAAU,UAAU,SAAS;AAAA,UACtC,UAAU,UAAU;AAAA,QACtB;AAAA,MACF;AAAA,IACF;AAEA,WAAO;AAAA,MACL,QAAI,0BAAW;AAAA,MACf;AAAA,MACA,eAAe;AAAA,MACf,eAAe;AAAA,MACf;AAAA,MACA,SAAS,UAAU,UAAU,SAAS;AAAA,MACtC,UAAU,UAAU;AAAA,IACtB;AAAA,EACF;AACF;;;AEnFA,IAAAC,iBAA2B;AAC3B,2BAAyB;AAIzB,IAAM,kBAAkB;AACxB,IAAM,oBAAoB;AAMnB,IAAM,uBAAN,cAAmC,mBAAyC;AAAA,EACxE,OAAO;AAAA,EAEhB,SACE,WACA,QACA,SACiB;AACjB,UAAM,kBAAc,2BAAW;AAC/B,UAAM,UAAU,SAAS;AACzB,UAAM,UAAU,UAAU,WAAW;AACrC,UAAM,mBAAmB,UAAU,oBAAoB;AAEvD,QAAI,CAAC,SAAS;AACZ,aAAO,KAAK,aAAa,aAAa;AAAA,QACpC;AAAA,QACA,SAAS;AAAA,QACT,UAAU,OAAO,gBAAgB;AAAA,MACnC,CAAC;AAAA,IACH;AAEA,QAAI,WAA0B;AAC9B,QAAI,eAA8B;AAClC,QAAI;AACJ,QAAI;AAEJ,YAAQ,IAAI,2BAA2B,OAAO,SAAS,OAAO,EAAE;AAEhE,QAAI;AACF,yCAAS,SAAS;AAAA,QAChB,KAAK;AAAA,QACL,UAAU;AAAA,QACV,OAAO,CAAC,UAAU,QAAQ,MAAM;AAAA,MAClC,CAAC;AACD,iBAAW;AAAA,IACb,SAAS,KAAK;AACZ,YAAM,QAAQ;AAMd,iBACE,OAAO,MAAM,WAAW,WACpB,MAAM,SACN,OAAO,MAAM,SAAS,WACpB,MAAM,OACN;AACR,qBAAe,MAAM;AACrB,eAAS,KAAK,eAAe,MAAM,MAAM;AACzC,eAAS,KAAK,eAAe,MAAM,MAAM;AAAA,IAC3C;AAEA,UAAM,SAAS,aAAa,QAAQ,aAAa;AAEjD,UAAM,UAAmC,EAAE,SAAS,QAAQ;AAC5D,QAAI,WAAW,UAAa,WAAW,IAAI;AACzC,cAAQ,SAAS;AAAA,IACnB;AACA,QAAI,WAAW,UAAa,WAAW,IAAI;AACzC,cAAQ,SAAS;AAAA,IACnB;AAEA,WAAO,KAAK,aAAa,aAAa;AAAA,MACpC,QAAQ;AAAA,MAGR,SAAS,KAAK,cAAc,UAAU,kBAAkB,YAAY;AAAA,MACpE,UAAU,OAAO,gBAAgB;AAAA,MACjC,QAAQ,aAAa,OAAO,OAAO,QAAQ,IAAI;AAAA,MAC/C;AAAA,IACF,CAAC;AAAA,EACH;AAAA,EAEQ,aACN,aACA,QACiB;AACjB,WAAO;AAAA,MACL,QAAI,2BAAW;AAAA,MACf;AAAA,MACA,eAAe;AAAA,MACf,eAAe;AAAA,MACf;AAAA,MACA,GAAG;AAAA,IACL;AAAA,EACF;AAAA,EAEQ,eACN,OACoB;AACpB,QAAI,UAAU,UAAa,UAAU,KAAM,QAAO;AAClD,QAAI,OAAO,UAAU,SAAU,QAAO;AACtC,WAAO,MAAM,SAAS,OAAO;AAAA,EAC/B;AAAA,EAEQ,cACN,UACA,kBACA,cACQ;AACR,QAAI,aAAa,MAAM;AACrB,aAAO,iBAAiB,YAAY;AAAA,IACtC;AACA,QAAI,aAAa,kBAAkB;AACjC,aAAO,2BAA2B,QAAQ;AAAA,IAC5C;AACA,WAAO,qBAAqB,QAAQ,cAAc,gBAAgB;AAAA,EACpE;AACF;;;ACxHA,IAAAC,iBAA2B;AAC3B,uBAAgC;AAChC,gBAA6B;AAatB,SAAS,oBAAoB,UAAwC;AAC1E,MAAI,CAAC,UAAU,OAAO,QAAQ;AAC5B,WAAO;AAAA,EACT;AACA,QAAM,QAAkB,CAAC;AACzB,aAAW,QAAQ,SAAS,OAAO;AACjC,UAAM,QAAkB;AAAA,MACtB,QAAQ,KAAK,UAAU;AAAA,MACvB,SAAS,KAAK,IAAI;AAAA,MAClB,aAAa,KAAK,UAAU;AAAA,IAC9B;AACA,QAAI,KAAK,UAAU;AACjB,YAAM,KAAK,SAAS,KAAK,QAAQ,EAAE;AACnC,UAAI,KAAK,eAAe;AACtB,cAAM,KAAK,SAAS,KAAK,aAAa,EAAE;AAAA,MAC1C;AAAA,IACF;AACA,QAAI,KAAK,eAAe;AACtB,YAAM,KAAK,WAAW,KAAK,aAAa,EAAE;AAAA,IAC5C;AACA,QAAI,KAAK,OAAO;AACd,YAAM,KAAK,UAAU,KAAK,KAAK,EAAE;AAAA,IACnC;AACA,UAAM,KAAK,MAAM,KAAK,IAAI,CAAC;AAAA,EAC7B;AACA,SAAO,MAAM,KAAK,IAAI;AACxB;AAcO,SAAS,oBACd,KACA,KACQ;AACR,SAAO,IACJ,QAAQ,mBAAmB,IAAI,MAAM,EACrC,QAAQ,gBAAgB,IAAI,GAAG,EAC/B,QAAQ,yBAAyB,IAAI,YAAY,EACjD,QAAQ,0BAA0B,IAAI,aAAa,EACnD,QAAQ,qBAAqB,IAAI,QAAQ,EACzC,QAAQ,kBAAkB,IAAI,KAAK;AACxC;AAEO,SAAS,oBAAoB,QAA8B;AAChE,MAAI,WAAW,QAAQ,OAAO,WAAW,UAAU;AACjD,UAAM,IAAI,MAAM,+BAA+B;AAAA,EACjD;AACA,QAAM,MAAM;AACZ,MAAI,OAAO,IAAI,SAAS,UAAU;AAChC,UAAM,IAAI,MAAM,kDAAkD;AAAA,EACpE;AACA,MAAI,OAAO,IAAI,UAAU,UAAU;AACjC,UAAM,IAAI,MAAM,mDAAmD;AAAA,EACrE;AACA,MAAI,IAAI,QAAQ,KAAK,IAAI,QAAQ,KAAK;AACpC,UAAM,IAAI,MAAM,6CAA6C;AAAA,EAC/D;AACA,MAAI,OAAO,IAAI,mBAAmB,UAAU;AAC1C,UAAM,IAAI;AAAA,MACR;AAAA,IACF;AAAA,EACF;AACA,SAAO;AAAA,IACL,MAAM,IAAI;AAAA,IACV,OAAO,IAAI;AAAA,IACX,gBAAgB,IAAI;AAAA,EACtB;AACF;AAEA,IAAM,oBAAoB;AAC1B,IAAM,gBAAgB;AAGtB,IAAM,gCAAgC;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAWtC,IAAM,wBAAwB;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAwBvB,IAAM,oBAAN,cAAgC,mBAAsC;AAAA,EAClE,OAAO;AAAA,EAEhB,MAAM,SACJ,WACA,OACA,SAC0B;AAC1B,UAAM,kBAAc,2BAAW;AAC/B,UAAM,YAAY,SAAS;AAC3B,UAAM,UAAU,SAAS,WAAW;AACpC,UAAM,mBAAmB,SAAS;AAElC,UAAM,SAAS,MAAM,cAAc;AACnC,UAAM,YAAY,MAAM,aAAa,CAAC;AAGtC,UAAM,eAAe,UAAU,IAAI,CAAC,MAAM,EAAE,IAAI;AAChD,UAAM,gBAAgB,UACnB,OAAO,CAAC,MAAM,EAAE,WAAW,UAAU,EACrC,IAAI,CAAC,MAAM,EAAE,IAAI;AACpB,UAAM,WAAW,UACd,OAAO,CAAC,MAAM,EAAE,WAAW,KAAK,EAChC,IAAI,CAAC,MAAM,EAAE,IAAI;AAEpB,UAAM,eACJ,aAAa,SAAS,IAClB,aAAa,IAAI,CAAC,SAAiB,KAAK,IAAI,EAAE,EAAE,KAAK,IAAI,IACzD;AACN,UAAM,gBACJ,cAAc,SAAS,IACnB,cAAc,IAAI,CAAC,SAAiB,KAAK,IAAI,EAAE,EAAE,KAAK,IAAI,IAC1D;AACN,UAAM,WACJ,SAAS,SAAS,IACd,SAAS,IAAI,CAAC,SAAiB,KAAK,IAAI,EAAE,EAAE,KAAK,IAAI,IACrD;AAEN,UAAM,QAAQ,oBAAoB,MAAM,QAAQ;AAChD,UAAM,MAA0B;AAAA,MAC9B;AAAA,MACA,KAAK;AAAA,MACL;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF;AACA,UAAM,UAAU,CAAC,MAAc,oBAAoB,GAAG,GAAG;AAEzD,UAAM,cAAc,QAAQ,UAAU,MAAM;AAC5C,UAAM,eACJ,UAAU,gBAAgB,QAAQ,UAAU,iBAAiB,KACzD,QAAQ,UAAU,YAAY,IAAI,SAAS,wBAC3C,QAAQ,6BAA6B,IACrC,SACA;AAEN,UAAM,WAAW,UAAU,YAAY;AACvC,UAAM,kBAAkB,UAAU,aAAa;AAC/C,UAAM,cAAc,UAAU,eAAe;AAE7C,QAAI,CAAC,oBAAoB,CAAC,WAAW;AACnC,aAAO;AAAA,QACL,QAAI,2BAAW;AAAA,QACf;AAAA,QACA,eAAe;AAAA,QACf,eAAe;AAAA,QACf;AAAA,QACA,SAAS;AAAA,QACT,UAAU,OAAO,QAAQ;AAAA,MAC3B;AAAA,IACF;AAEA,UAAM,mBAAmB;AACzB,QAAI;AACJ,QAAI;AAEJ,QAAI;AACF,eAAS,UAAU,GAAG,WAAW,kBAAkB,WAAW;AAC5D,cAAM,SAAS,mBACX,MAAM,iBAAiB;AAAA,UACrB,QAAQ;AAAA,UACR,QAAQ;AAAA,UACR;AAAA,UACA;AAAA,QACF,CAAC,IACD,MAAM,KAAK;AAAA,UACT;AAAA,UACA,UAAU,SAAS;AAAA,UACnB;AAAA,UACA;AAAA,UACA;AAAA,UACA;AAAA,QACF;AAEJ,sBAAc,OAAO;AACrB,YAAI;AACF,gBAAM,SAAS,KAAK,MAAM,OAAO,IAAI;AACrC,gBAAM,cAAc,oBAAoB,MAAM;AAC9C,gBAAM,SAAS,YAAY,SAAS;AACpC,iBAAO;AAAA,YACL,QAAI,2BAAW;AAAA,YACf;AAAA,YACA,eAAe;AAAA,YACf,eAAe;AAAA,YACf,QAAQ;AAAA,YAGR,SAAS,SACL,eAAe,YAAY,KAAK,OAAO,QAAQ,KAAK,YAAY,IAAI,KACpE,eAAe,YAAY,KAAK,MAAM,QAAQ,KAAK,YAAY,IAAI;AAAA,YACvE,UAAU,OAAO,QAAQ;AAAA,YACzB,QAAQ,OAAO,YAAY,KAAK;AAAA,YAChC,SAAS;AAAA,cACP,OAAO,YAAY;AAAA,cACnB,gBAAgB,YAAY;AAAA,cAC5B,MAAM,YAAY;AAAA,YACpB;AAAA,UACF;AAAA,QACF,SAAS,UAAU;AACjB,2BACE,oBAAoB,QAAQ,WAAW,IAAI,MAAM,OAAO,QAAQ,CAAC;AAAA,QACrE;AAAA,MACF;AAEA,aAAO;AAAA,QACL,QAAI,2BAAW;AAAA,QACf;AAAA,QACA,eAAe;AAAA,QACf,eAAe;AAAA,QACf;AAAA,QACA,SAAS,wCAAwC,gBAAgB,cAAc,gBAAgB,WAAW,SAAS;AAAA,QACnH,UAAU,OAAO,QAAQ;AAAA,QACzB,QAAQ;AAAA,QACR,SAAS,EAAE,SAAS,aAAa,MAAM,GAAG,GAAG,EAAE;AAAA,MACjD;AAAA,IACF,SAAS,KAAK;AACZ,YAAM,UAAU,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAC/D,aAAO;AAAA,QACL,QAAI,2BAAW;AAAA,QACf;AAAA,QACA,eAAe;AAAA,QACf,eAAe;AAAA,QACf;AAAA,QACA,SAAS,0BAA0B,OAAO;AAAA,QAC1C,UAAU,OAAO,QAAQ;AAAA,QACzB,SAAS,EAAE,OAAO,QAAQ;AAAA,MAC5B;AAAA,IACF;AAAA,EACF;AAAA,EAEA,MAAc,iBACZ,WACA,SACA,QACA,QACA,iBACA,aAC2B;AAC3B,UAAM,gBAAY,kCAAgB;AAAA,MAChC,SAAS,UAAU;AAAA,MACnB,QAAQ;AAAA,MACR,SAAS,UAAU;AAAA,IACrB,CAAC;AACD,UAAM,SAAS,UAAM,wBAAa;AAAA,MAChC,OAAO,UAAU,OAAO;AAAA,MACxB;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF,CAAC;AACD,WAAO,EAAE,MAAM,OAAO,KAAK;AAAA,EAC7B;AACF;;;AJ/SA,IAAM,oBAAoB,IAAI,kBAAkB;AAEhD,IAAM,aAAiD;AAAA,EACrD,kBAAkB,IAAI,wBAAwB;AAAA,EAC9C,cAAc,IAAI,qBAAqB;AAAA,EACvC,WAAW;AAAA;AAAA,EAEX,QAAQ;AACV;AAQO,SAAS,kBACd,MACA,WACM;AACN,aAAW,IAAI,IAAI;AACrB;AAQO,SAAS,aAAa,MAA8C;AACzE,SAAO,WAAW,IAAI;AACxB;AAUA,eAAsB,mBACpB,OACA,YACA,SAC4B;AAC5B,MAAI,WAAW,WAAW,GAAG;AAC3B,WAAO,CAAC;AAAA,EACV;AACA,SAAO,QAAQ;AAAA,IACb,WAAW,IAAI,OAAO,cAAc;AAClC,YAAM,YAAY,WAAW,UAAU,IAAI;AAC3C,UAAI,CAAC,WAAW;AACd,eAAO;AAAA,UACL,QAAI,2BAAW;AAAA,UACf,iBAAa,2BAAW;AAAA,UACxB,eAAe,UAAU;AAAA,UACzB,eAAe;AAAA,UACf;AAAA,UACA,SAAS,+BAA+B,UAAU,IAAI;AAAA,UACtD,UAAU;AAAA,QACZ;AAAA,MACF;AACA,YAAM,UAAU,KAAK,IAAI;AACzB,YAAM,SAAS,MAAM,UAAU,SAAS,WAAW,OAAO,OAAO;AACjE,YAAM,aAAa,KAAK,IAAI,IAAI;AAChC,aAAO,EAAE,GAAG,QAAQ,UAAU,WAAW;AAAA,IAC3C,CAAC;AAAA,EACH;AACF;",
   "names": ["import_zod", "LLMStepType", "import_zod", "AssertionResultStatus", "import_crypto", "import_crypto", "import_crypto"]
 }

package/build/index.mjs

@@ -14,7 +14,7 @@ var BuildPassedAssertionSchema = z.object({
 });
 var LlmJudgeAssertionSchema = z.object({
   type: z.literal("llm_judge"),
-  /** Prompt template; placeholders: {{output}}, {{cwd}}, {{changedFiles}}, {{trace}} */
+  /** Prompt template; placeholders: {{output}}, {{cwd}}, {{changedFiles}}, {{modifiedFiles}}, {{newFiles}}, {{trace}} */
   prompt: z.string(),
   /** Optional system prompt for the judge (default asks for JSON with score) */
   systemPrompt: z.string().optional(),
@@ -316,7 +316,8 @@ function validateJudgeResult(parsed) {
 }
 var DEFAULT_MIN_SCORE = 70;
 var DEFAULT_MODEL = "claude-3-5-haiku-20241022";
-var DEFAULT_SYSTEM_PROMPT_CONTEXT = `You are judging a scenario run. Use these values:
+var DEFAULT_SYSTEM_PROMPT_CONTEXT = `You are judging a scenario run. The ACTUAL run data is provided below \u2014 use it to verify facts:
+
 - {{output}}: the agent's final output
 - {{cwd}}: working directory
 - {{changedFiles}}: list of all files changed (or "No files were changed")
@@ -324,7 +325,7 @@ var DEFAULT_SYSTEM_PROMPT_CONTEXT = `You are judging a scenario run. Use these v
 - {{newFiles}}: list of new files that were created (or "No new files were created")
 - {{trace}}: step-by-step trace (tool calls, completions) so you can check e.g. which tools were called and how many times
 
-Judge how well the output meets the acceptance criteria stated in the user prompt.`;
+CRITICAL: When the user asks you to verify a specific fact, compare it strictly against the actual data above. If the expected outcome does NOT match the actual outcome, you MUST give a score of 0 or near 0. Do not be lenient \u2014 factual mismatches are failures.`;
 var DEFAULT_SYSTEM_PROMPT = `You must respond only with a valid JSON object that conforms exactly to the following structure:
 
 {
@@ -384,52 +385,58 @@ var LlmJudgeEvaluator = class extends AssertionEvaluator {
         expected: String(minScore)
       };
     }
+    const maxParseAttempts = 3;
+    let lastParseError;
+    let lastRawText;
     try {
-      const result = generateTextStub ? await generateTextStub({
-        prompt: finalPrompt,
-        system: systemPrompt,
-        maxOutputTokens,
-        temperature
-      }) : await this.callGenerateText(
-        llmConfig,
-        assertion.model ?? DEFAULT_MODEL,
-        finalPrompt,
-        systemPrompt,
-        maxOutputTokens,
-        temperature
-      );
-      let judgeResult;
-      try {
-        const parsed = JSON.parse(result.text);
-        judgeResult = validateJudgeResult(parsed);
-      } catch (parseErr) {
-        return {
-          id: randomUUID3(),
-          assertionId,
-          assertionType: "llm_judge",
-          assertionName: "LLM judge",
-          status: "failed" /* FAILED */,
-          message: `Failed to parse judge response: ${parseErr instanceof Error ? parseErr.message : String(parseErr)}`,
-          expected: String(minScore),
-          actual: void 0,
-          details: { rawText: result.text.slice(0, 500) }
-        };
+      for (let attempt = 1; attempt <= maxParseAttempts; attempt++) {
+        const result = generateTextStub ? await generateTextStub({
+          prompt: finalPrompt,
+          system: systemPrompt,
+          maxOutputTokens,
+          temperature
+        }) : await this.callGenerateText(
+          llmConfig,
+          assertion.model ?? DEFAULT_MODEL,
+          finalPrompt,
+          systemPrompt,
+          maxOutputTokens,
+          temperature
+        );
+        lastRawText = result.text;
+        try {
+          const parsed = JSON.parse(result.text);
+          const judgeResult = validateJudgeResult(parsed);
+          const passed = judgeResult.score >= minScore;
+          return {
+            id: randomUUID3(),
+            assertionId,
+            assertionType: "llm_judge",
+            assertionName: "LLM judge",
+            status: passed ? "passed" /* PASSED */ : "failed" /* FAILED */,
+            message: passed ? `Judge score ${judgeResult.score} >= ${minScore}: ${judgeResult.text}` : `Judge score ${judgeResult.score} < ${minScore}: ${judgeResult.text}`,
+            expected: String(minScore),
+            actual: String(judgeResult.score),
+            details: {
+              score: judgeResult.score,
+              scoreReasoning: judgeResult.scoreReasoning,
+              text: judgeResult.text
+            }
+          };
+        } catch (parseErr) {
+          lastParseError = parseErr instanceof Error ? parseErr : new Error(String(parseErr));
+        }
       }
-      const passed = judgeResult.score >= minScore;
       return {
         id: randomUUID3(),
         assertionId,
         assertionType: "llm_judge",
         assertionName: "LLM judge",
-        status: passed ? "passed" /* PASSED */ : "failed" /* FAILED */,
-        message: passed ? `Judge score ${judgeResult.score} >= ${minScore}: ${judgeResult.text}` : `Judge score ${judgeResult.score} < ${minScore}: ${judgeResult.text}`,
+        status: "failed" /* FAILED */,
+        message: `Failed to parse judge response after ${maxParseAttempts} attempts: ${lastParseError?.message ?? "unknown"}`,
         expected: String(minScore),
-        actual: String(judgeResult.score),
-        details: {
-          score: judgeResult.score,
-          scoreReasoning: judgeResult.scoreReasoning,
-          text: judgeResult.text
-        }
+        actual: void 0,
+        details: { rawText: lastRawText?.slice(0, 500) }
       };
     } catch (err) {
       const message = err instanceof Error ? err.message : String(err);

package/build/index.mjs.map

@@ -1,7 +1,7 @@
 {
   "version": 3,
   "sources": ["../src/types/assertions.ts", "../src/types/trace.ts", "../src/types/result.ts", "../src/evaluators/index.ts", "../src/evaluators/skill-was-called-evaluator.ts", "../src/evaluators/assertion-evaluator.ts", "../src/evaluators/build-passed-evaluator.ts", "../src/evaluators/llm-judge-evaluator.ts"],
-  "sourcesContent": ["import { z } from \"zod\";\n\n/**\n * Assertion: the agent must have invoked a specific skill during the run.\n * Checked by inspecting the LLM trace for a \"Skill\" tool use with the given skill.\n * Data: skillName (the skill that must have been called).\n */\nexport const SkillWasCalledAssertionSchema = z.object({\n  type: z.literal(\"skill_was_called\"),\n  /** Name of the skill that must have been called (matched against trace Skill tool args) */\n  skillName: z.string(),\n});\n\nexport type SkillWasCalledAssertion = z.infer<\n  typeof SkillWasCalledAssertionSchema\n>;\n\n/**\n * Assertion: a build command must exit with the expected code (default 0).\n * Runs the command in the scenario working directory.\n */\nexport const BuildPassedAssertionSchema = z.object({\n  type: z.literal(\"build_passed\"),\n  /** Command to run (default: \"yarn build\") */\n  command: z.string().optional(),\n  /** Expected exit code (default: 0) */\n  expectedExitCode: z.number().int().optional(),\n});\n\nexport type BuildPassedAssertion = z.infer<typeof BuildPassedAssertionSchema>;\n\n/**\n * Assertion: an LLM judges the scenario output (score 0-100).\n * Prompt can use {{output}}, {{cwd}}, {{changedFiles}}, {{trace}}.\n * Passes if judge score >= minScore.\n */\nexport const LlmJudgeAssertionSchema = z.object({\n  type: z.literal(\"llm_judge\"),\n  /** Prompt template; placeholders: {{output}}, {{cwd}}, {{changedFiles}}, {{trace}} */\n  prompt: z.string(),\n  /** Optional system prompt for the judge (default asks for JSON with score) */\n  systemPrompt: z.string().optional(),\n  /** Minimum score to pass (0-100, default 70) */\n  minScore: z.number().int().min(0).max(100).optional(),\n  /** Model for the judge (e.g. claude-3-5-haiku) */\n  model: z.string().optional(),\n  maxTokens: z.number().int().optional(),\n  temperature: z.number().min(0).max(1).optional(),\n});\n\nexport type LlmJudgeAssertion = z.infer<typeof LlmJudgeAssertionSchema>;\n\n/**\n * Union of all assertion types.\n * Each assertion has a type and type-specific data.\n * Uses z.union (not z.discriminatedUnion) for Zod v4 compatibility when used as array element.\n */\nexport const AssertionSchema = z.union([\n  SkillWasCalledAssertionSchema,\n  BuildPassedAssertionSchema,\n  LlmJudgeAssertionSchema,\n]);\n\nexport type Assertion = z.infer<typeof AssertionSchema>;\n", "import { z } from \"zod\";\n\n/**\n * Token usage schema.\n */\nexport const TokenUsageSchema = z.object({\n  prompt: z.number(),\n  completion: z.number(),\n  total: z.number(),\n});\n\nexport type TokenUsage = z.infer<typeof TokenUsageSchema>;\n\n/**\n * LLM step type enum.\n */\nexport enum LLMStepType {\n  COMPLETION = \"completion\",\n  TOOL_USE = \"tool_use\",\n  TOOL_RESULT = \"tool_result\",\n  THINKING = \"thinking\",\n}\n\n/**\n * LLM trace step schema.\n */\nexport const LLMTraceStepSchema = z.object({\n  id: z.string(),\n  stepNumber: z.number(),\n  type: z.enum(LLMStepType),\n  model: z.string(),\n  provider: z.string(),\n  startedAt: z.string(),\n  durationMs: z.number(),\n  tokenUsage: TokenUsageSchema,\n  costUsd: z.number(),\n  toolName: z.string().optional(),\n  toolArguments: z.string().optional(),\n  inputPreview: z.string().optional(),\n  outputPreview: z.string().optional(),\n  success: z.boolean(),\n  error: z.string().optional(),\n});\n\nexport type LLMTraceStep = z.infer<typeof LLMTraceStepSchema>;\n\n/**\n * LLM breakdown stats schema.\n */\nexport const LLMBreakdownStatsSchema = z.object({\n  count: z.number(),\n  durationMs: z.number(),\n  tokens: z.number(),\n  costUsd: z.number(),\n});\n\nexport type LLMBreakdownStats = z.infer<typeof LLMBreakdownStatsSchema>;\n\n/**\n * LLM trace summary schema.\n */\nexport const LLMTraceSummarySchema = z.object({\n  totalSteps: z.number(),\n  totalDurationMs: z.number(),\n  totalTokens: TokenUsageSchema,\n  totalCostUsd: z.number(),\n  stepTypeBreakdown: z.record(z.string(), LLMBreakdownStatsSchema).optional(),\n  modelBreakdown: z.record(z.string(), LLMBreakdownStatsSchema),\n  modelsUsed: z.array(z.string()),\n});\n\nexport type LLMTraceSummary = z.infer<typeof LLMTraceSummarySchema>;\n\n/**\n * LLM trace schema.\n */\nexport const LLMTraceSchema = z.object({\n  id: z.string(),\n  steps: z.array(LLMTraceStepSchema),\n  summary: LLMTraceSummarySchema,\n});\n\nexport type LLMTrace = z.infer<typeof LLMTraceSchema>;\n", "import { z } from \"zod\";\nimport { LLMTraceStepSchema } from \"./trace.js\";\n\n/**\n * Assertion result status enum.\n */\nexport enum AssertionResultStatus {\n  PASSED = \"passed\",\n  FAILED = \"failed\",\n  SKIPPED = \"skipped\",\n  ERROR = \"error\",\n}\n\n/**\n * Assertion result schema.\n */\nexport const AssertionResultSchema = z.object({\n  id: z.string(),\n  assertionId: z.string(),\n  assertionType: z.string(),\n  assertionName: z.string(),\n  status: z.enum(AssertionResultStatus),\n  message: z.string().optional(),\n  expected: z.string().optional(),\n  actual: z.string().optional(),\n  duration: z.number().optional(),\n  details: z.record(z.string(), z.unknown()).optional(),\n  llmTraceSteps: z.array(LLMTraceStepSchema).optional(),\n});\n\nexport type AssertionResult = z.infer<typeof AssertionResultSchema>;\n", "import type { Assertion, AssertionResult } from \"../types/index.js\";\nimport { AssertionResultStatus } from \"../types/index.js\";\nimport { randomUUID } from \"crypto\";\nimport type { AssertionContext } from \"./assertion-evaluator.js\";\nimport { AssertionEvaluator } from \"./assertion-evaluator.js\";\nimport { SkillWasCalledEvaluator } from \"./skill-was-called-evaluator.js\";\nimport { BuildPassedEvaluator } from \"./build-passed-evaluator.js\";\nimport { LlmJudgeEvaluator } from \"./llm-judge-evaluator.js\";\nimport type { EvaluationInput } from \"../types/index.js\";\n\nconst llmJudgeEvaluator = new LlmJudgeEvaluator();\n\nconst evaluators: Record<string, AssertionEvaluator> = {\n  skill_was_called: new SkillWasCalledEvaluator(),\n  build_passed: new BuildPassedEvaluator(),\n  llm_judge: llmJudgeEvaluator,\n  // Custom assertions use the same LLM-based evaluation as llm_judge\n  custom: llmJudgeEvaluator,\n};\n\n/**\n * Register a custom assertion evaluator.\n *\n * @param type - The assertion type identifier\n * @param evaluator - The evaluator instance\n */\nexport function registerEvaluator(\n  type: string,\n  evaluator: AssertionEvaluator,\n): void {\n  evaluators[type] = evaluator;\n}\n\n/**\n * Get a registered evaluator by type.\n *\n * @param type - The assertion type identifier\n * @returns The evaluator or undefined if not found\n */\nexport function getEvaluator(type: string): AssertionEvaluator | undefined {\n  return evaluators[type];\n}\n\n/**\n * Evaluate all assertions against the input.\n *\n * @param input - Evaluation input (includes outputText, llmTrace, fileDiffs)\n * @param assertions - List of assertions to evaluate\n * @param context - Optional context (e.g. workDir for build_passed, llmConfig for llm_judge)\n * @returns Array of assertion results; empty if no assertions\n */\nexport async function evaluateAssertions(\n  input: EvaluationInput,\n  assertions: Assertion[],\n  context?: AssertionContext,\n): Promise<AssertionResult[]> {\n  if (assertions.length === 0) {\n    return [];\n  }\n  return Promise.all(\n    assertions.map(async (assertion) => {\n      const evaluator = evaluators[assertion.type];\n      if (!evaluator) {\n        return {\n          id: randomUUID(),\n          assertionId: randomUUID(),\n          assertionType: assertion.type,\n          assertionName: \"Unknown assertion\",\n          status: AssertionResultStatus.ERROR,\n          message: `Unsupported assertion type: ${assertion.type}`,\n          duration: 0,\n        };\n      }\n      const startMs = Date.now();\n      const result = await evaluator.evaluate(assertion, input, context);\n      const durationMs = Date.now() - startMs;\n      return { ...result, duration: durationMs };\n    }),\n  );\n}\n\n// Re-export evaluator classes and types\nexport { AssertionEvaluator } from \"./assertion-evaluator.js\";\nexport type {\n  AssertionContext,\n  LlmConfig,\n  LlmJudgeGenerateTextOptions,\n} from \"./assertion-evaluator.js\";\nexport { SkillWasCalledEvaluator } from \"./skill-was-called-evaluator.js\";\nexport { BuildPassedEvaluator } from \"./build-passed-evaluator.js\";\nexport {\n  LlmJudgeEvaluator,\n  formatTraceForJudge,\n  replacePlaceholders,\n  validateJudgeResult,\n  type JudgeResult,\n} from \"./llm-judge-evaluator.js\";\n", "import type {\n  SkillWasCalledAssertion,\n  AssertionResult,\n  LLMTrace,\n  EvaluationInput,\n} from \"../types/index.js\";\nimport { AssertionResultStatus } from \"../types/index.js\";\nimport { randomUUID } from \"crypto\";\nimport type { AssertionContext } from \"./assertion-evaluator.js\";\nimport { AssertionEvaluator } from \"./assertion-evaluator.js\";\n\n/**\n * Check if parsed Skill tool args reference the expected skill by name.\n */\nfunction toolArgsReferenceSkillByName(\n  args: unknown,\n  expectedSkillName: string,\n): boolean {\n  if (args === null || typeof args !== \"object\") {\n    return false;\n  }\n  const obj = args as Record<string, unknown>;\n  const name = obj.skill;\n  return typeof name === \"string\" && name === expectedSkillName;\n}\n\n/**\n * Evaluator for \"skill_was_called\" assertion: the LLM trace must contain a step\n * where the \"Skill\" tool was used with the expected skill (by name).\n */\nexport class SkillWasCalledEvaluator extends AssertionEvaluator<SkillWasCalledAssertion> {\n  readonly type = \"skill_was_called\" as const;\n\n  evaluate(\n    assertion: SkillWasCalledAssertion,\n    input: EvaluationInput,\n    // eslint-disable-next-line @typescript-eslint/no-unused-vars -- context not used for skill_was_called\n    _context?: AssertionContext,\n  ): AssertionResult {\n    const assertionId = randomUUID();\n\n    const llmTrace: LLMTrace | undefined = input.llmTrace;\n    if (!llmTrace?.steps?.length) {\n      return {\n        id: randomUUID(),\n        assertionId,\n        assertionType: \"skill_was_called\",\n        assertionName: \"Skill was called\",\n        status: AssertionResultStatus.FAILED,\n        message: \"No LLM trace steps to check for skill invocation\",\n        expected: assertion.skillName,\n      };\n    }\n\n    for (const step of llmTrace.steps) {\n      if (step.toolName !== \"Skill\") {\n        continue;\n      }\n      let args: unknown;\n      try {\n        args = step.toolArguments\n          ? (JSON.parse(step.toolArguments) as unknown)\n          : undefined;\n      } catch {\n        continue;\n      }\n      if (toolArgsReferenceSkillByName(args, assertion.skillName)) {\n        return {\n          id: randomUUID(),\n          assertionId,\n          assertionType: \"skill_was_called\",\n          assertionName: \"Skill was called\",\n          status: AssertionResultStatus.PASSED,\n          message: `Skill \"${assertion.skillName}\" was called`,\n          expected: assertion.skillName,\n        };\n      }\n    }\n\n    return {\n      id: randomUUID(),\n      assertionId,\n      assertionType: \"skill_was_called\",\n      assertionName: \"Skill was called\",\n      status: AssertionResultStatus.FAILED,\n      message: `Skill \"${assertion.skillName}\" was not called`,\n      expected: assertion.skillName,\n    };\n  }\n}\n", "import type {\n  Assertion,\n  AssertionResult,\n  EvaluationInput,\n} from \"../types/index.js\";\n\n/**\n * Options passed to the LLM for llm_judge. Used by the optional stub for testing.\n */\nexport interface LlmJudgeGenerateTextOptions {\n  prompt: string;\n  system: string;\n  maxOutputTokens: number;\n  temperature: number;\n}\n\n/**\n * Configuration for LLM calls (used by llm_judge assertion).\n */\nexport interface LlmConfig {\n  /** Base URL for the AI API (e.g., 'https://api.anthropic.com') */\n  baseUrl: string;\n  /** Headers to include in API requests (e.g., API key) */\n  headers: Record<string, string>;\n}\n\n/**\n * Optional context passed when evaluating assertions.\n */\nexport interface AssertionContext {\n  /** Working directory for the scenario (used by build_passed) */\n  workDir?: string;\n  /** LLM configuration (used by llm_judge) */\n  llmConfig?: LlmConfig;\n  /**\n   * Optional stub for llm_judge: when set, the evaluator uses this instead of the real AI call.\n   * Used only in tests to avoid hitting the API.\n   */\n  generateTextForLlmJudge?: (\n    options: LlmJudgeGenerateTextOptions,\n  ) => Promise<{ text: string }>;\n}\n\n/**\n * Abstract base for assertion evaluators.\n * Each assertion type has a concrete class that implements evaluate().\n * evaluate() may return a Promise for async assertions (e.g. llm_judge).\n */\nexport abstract class AssertionEvaluator<T extends Assertion = Assertion> {\n  abstract readonly type: T[\"type\"];\n\n  abstract evaluate(\n    assertion: T,\n    input: EvaluationInput,\n    context?: AssertionContext,\n  ): AssertionResult | Promise<AssertionResult>;\n}\n", "import type {\n  BuildPassedAssertion,\n  AssertionResult,\n  EvaluationInput,\n} from \"../types/index.js\";\nimport { AssertionResultStatus } from \"../types/index.js\";\nimport { randomUUID } from \"crypto\";\nimport { execSync } from \"child_process\";\nimport type { AssertionContext } from \"./assertion-evaluator.js\";\nimport { AssertionEvaluator } from \"./assertion-evaluator.js\";\n\nconst DEFAULT_COMMAND = \"yarn build\";\nconst DEFAULT_EXIT_CODE = 0;\n\n/**\n * Evaluator for \"build_passed\" assertion: runs a build command in the scenario\n * working directory and passes if the command exits with the expected code (default 0).\n */\nexport class BuildPassedEvaluator extends AssertionEvaluator<BuildPassedAssertion> {\n  readonly type = \"build_passed\" as const;\n\n  evaluate(\n    assertion: BuildPassedAssertion,\n    _input: EvaluationInput,\n    context?: AssertionContext,\n  ): AssertionResult {\n    const assertionId = randomUUID();\n    const workDir = context?.workDir;\n    const command = assertion.command ?? DEFAULT_COMMAND;\n    const expectedExitCode = assertion.expectedExitCode ?? DEFAULT_EXIT_CODE;\n\n    if (!workDir) {\n      return this.createResult(assertionId, {\n        status: AssertionResultStatus.FAILED,\n        message: \"No working directory provided for build_passed assertion\",\n        expected: String(expectedExitCode),\n      });\n    }\n\n    let exitCode: number | null = null;\n    let errorMessage: string | null = null;\n    let stdout: string | undefined;\n    let stderr: string | undefined;\n\n    console.log(`[build_passed] Running \"${command}\" in: ${workDir}`);\n\n    try {\n      execSync(command, {\n        cwd: workDir,\n        encoding: \"utf-8\",\n        stdio: [\"ignore\", \"pipe\", \"pipe\"],\n      });\n      exitCode = 0;\n    } catch (err) {\n      const error = err as Error & {\n        status?: number;\n        code?: number;\n        stdout?: string | Buffer;\n        stderr?: string | Buffer;\n      };\n      exitCode =\n        typeof error.status === \"number\"\n          ? error.status\n          : typeof error.code === \"number\"\n            ? error.code\n            : null;\n      errorMessage = error.message;\n      stdout = this.bufferToString(error.stdout);\n      stderr = this.bufferToString(error.stderr);\n    }\n\n    const passed = exitCode !== null && exitCode === expectedExitCode;\n\n    const details: Record<string, unknown> = { workDir, command };\n    if (stdout !== undefined && stdout !== \"\") {\n      details.stdout = stdout;\n    }\n    if (stderr !== undefined && stderr !== \"\") {\n      details.stderr = stderr;\n    }\n\n    return this.createResult(assertionId, {\n      status: passed\n        ? AssertionResultStatus.PASSED\n        : AssertionResultStatus.FAILED,\n      message: this.formatMessage(exitCode, expectedExitCode, errorMessage),\n      expected: String(expectedExitCode),\n      actual: exitCode !== null ? String(exitCode) : undefined,\n      details,\n    });\n  }\n\n  private createResult(\n    assertionId: string,\n    fields: Partial<AssertionResult>,\n  ): AssertionResult {\n    return {\n      id: randomUUID(),\n      assertionId,\n      assertionType: \"build_passed\",\n      assertionName: \"Build passed\",\n      status: AssertionResultStatus.FAILED,\n      ...fields,\n    };\n  }\n\n  private bufferToString(\n    value: string | Buffer | undefined,\n  ): string | undefined {\n    if (value === undefined || value === null) return undefined;\n    if (typeof value === \"string\") return value;\n    return value.toString(\"utf-8\");\n  }\n\n  private formatMessage(\n    exitCode: number | null,\n    expectedExitCode: number,\n    errorMessage: string | null,\n  ): string {\n    if (exitCode === null) {\n      return `Build failed: ${errorMessage}`;\n    }\n    if (exitCode === expectedExitCode) {\n      return `Build passed (exit code ${exitCode})`;\n    }\n    return `Build exited with ${exitCode}, expected ${expectedExitCode}`;\n  }\n}\n", "import type {\n  LlmJudgeAssertion,\n  AssertionResult,\n  LLMTrace,\n  EvaluationInput,\n} from \"../types/index.js\";\nimport { AssertionResultStatus } from \"../types/index.js\";\nimport { randomUUID } from \"crypto\";\nimport { createAnthropic } from \"@ai-sdk/anthropic\";\nimport { generateText } from \"ai\";\nimport type { AssertionContext, LlmConfig } from \"./assertion-evaluator.js\";\nimport { AssertionEvaluator } from \"./assertion-evaluator.js\";\n\nexport interface JudgeResult {\n  text: string;\n  score: number;\n  scoreReasoning: string;\n}\n\n/**\n * Format LLM trace as readable text for the judge (step number, type, tool name/args, output preview).\n */\nexport function formatTraceForJudge(llmTrace: LLMTrace | undefined): string {\n  if (!llmTrace?.steps?.length) {\n    return \"No trace available.\";\n  }\n  const lines: string[] = [];\n  for (const step of llmTrace.steps) {\n    const parts: string[] = [\n      `Step ${step.stepNumber}`,\n      `type: ${step.type}`,\n      `duration: ${step.durationMs}ms`,\n    ];\n    if (step.toolName) {\n      parts.push(`tool: ${step.toolName}`);\n      if (step.toolArguments) {\n        parts.push(`args: ${step.toolArguments}`);\n      }\n    }\n    if (step.outputPreview) {\n      parts.push(`output: ${step.outputPreview}`);\n    }\n    if (step.error) {\n      parts.push(`error: ${step.error}`);\n    }\n    lines.push(parts.join(\", \"));\n  }\n  return lines.join(\"\\n\");\n}\n\n/**\n * Context object for placeholder replacement.\n */\nexport interface PlaceholderContext {\n  output: string;\n  cwd: string;\n  changedFiles: string;\n  modifiedFiles: string;\n  newFiles: string;\n  trace: string;\n}\n\nexport function replacePlaceholders(\n  str: string,\n  ctx: PlaceholderContext,\n): string {\n  return str\n    .replace(/\\{\\{output\\}\\}/g, ctx.output)\n    .replace(/\\{\\{cwd\\}\\}/g, ctx.cwd)\n    .replace(/\\{\\{changedFiles\\}\\}/g, ctx.changedFiles)\n    .replace(/\\{\\{modifiedFiles\\}\\}/g, ctx.modifiedFiles)\n    .replace(/\\{\\{newFiles\\}\\}/g, ctx.newFiles)\n    .replace(/\\{\\{trace\\}\\}/g, ctx.trace);\n}\n\nexport function validateJudgeResult(parsed: unknown): JudgeResult {\n  if (parsed === null || typeof parsed !== \"object\") {\n    throw new Error(\"Judge result is not an object\");\n  }\n  const obj = parsed as Record<string, unknown>;\n  if (typeof obj.text !== \"string\") {\n    throw new Error(\"Judge result does not contain a valid text field\");\n  }\n  if (typeof obj.score !== \"number\") {\n    throw new Error(\"Judge result does not contain a valid score field\");\n  }\n  if (obj.score < 0 || obj.score > 100) {\n    throw new Error(\"Judge result score is not between 0 and 100\");\n  }\n  if (typeof obj.scoreReasoning !== \"string\") {\n    throw new Error(\n      \"Judge result does not contain a valid scoreReasoning field\",\n    );\n  }\n  return {\n    text: obj.text,\n    score: obj.score,\n    scoreReasoning: obj.scoreReasoning,\n  };\n}\n\nconst DEFAULT_MIN_SCORE = 70;\nconst DEFAULT_MODEL = \"claude-3-5-haiku-20241022\";\n\n/** Default system prompt context with placeholders; used when assertion.systemPrompt is empty. */\nconst DEFAULT_SYSTEM_PROMPT_CONTEXT = `You are judging a scenario run. Use these values:\n- {{output}}: the agent's final output\n- {{cwd}}: working directory\n- {{changedFiles}}: list of all files changed (or \"No files were changed\")\n- {{modifiedFiles}}: list of existing files that were modified (or \"No files were modified\")\n- {{newFiles}}: list of new files that were created (or \"No new files were created\")\n- {{trace}}: step-by-step trace (tool calls, completions) so you can check e.g. which tools were called and how many times\n\nJudge how well the output meets the acceptance criteria stated in the user prompt.`;\n\nconst DEFAULT_SYSTEM_PROMPT = `You must respond only with a valid JSON object that conforms exactly to the following structure:\n\n{\n  \"text\": string,\n  \"score\": number (0-100),\n  \"scoreReasoning\": string\n}\n\n- text: A brief textual verdict of the test result.\n- score: A number from 0 to 100 that reflects how well the answer meets the acceptance criteria.\n- scoreReasoning: A concise explanation justifying the assigned score.\n\nYour response must:\n- Contain only the JSON object above \u2014 no introductory text, no code formatting (e.g., no triple backticks), and no trailing comments.\n- Be valid and parseable by \\`JSON.parse\\`.\n- Use only double quotes for all keys and strings, as required by JSON.\n\nAny response that includes extra content or deviates from the specified format will cause parsing to fail. Follow these instructions exactly.`;\n\n/**\n * Evaluator for \"llm_judge\" assertion: an LLM judges the scenario output\n * (prompt with {{output}}, {{cwd}}, {{changedFiles}}, {{trace}}) and returns a score 0-100.\n * Passes if score >= minScore.\n */\nexport class LlmJudgeEvaluator extends AssertionEvaluator<LlmJudgeAssertion> {\n  readonly type = \"llm_judge\" as const;\n\n  async evaluate(\n    assertion: LlmJudgeAssertion,\n    input: EvaluationInput,\n    context?: AssertionContext,\n  ): Promise<AssertionResult> {\n    const assertionId = randomUUID();\n    const llmConfig = context?.llmConfig;\n    const workDir = context?.workDir ?? \"\";\n    const generateTextStub = context?.generateTextForLlmJudge;\n\n    const output = input.outputText ?? \"\";\n    const fileDiffs = input.fileDiffs ?? [];\n\n    // Compute file lists by status\n    const changedPaths = fileDiffs.map((d) => d.path);\n    const modifiedPaths = fileDiffs\n      .filter((d) => d.status === \"modified\")\n      .map((d) => d.path);\n    const newPaths = fileDiffs\n      .filter((d) => d.status === \"new\")\n      .map((d) => d.path);\n\n    const changedFiles =\n      changedPaths.length > 0\n        ? changedPaths.map((path: string) => `- ${path}`).join(\"\\n\")\n        : \"No files were changed\";\n    const modifiedFiles =\n      modifiedPaths.length > 0\n        ? modifiedPaths.map((path: string) => `- ${path}`).join(\"\\n\")\n        : \"No files were modified\";\n    const newFiles =\n      newPaths.length > 0\n        ? newPaths.map((path: string) => `- ${path}`).join(\"\\n\")\n        : \"No new files were created\";\n\n    const trace = formatTraceForJudge(input.llmTrace);\n    const ctx: PlaceholderContext = {\n      output,\n      cwd: workDir,\n      changedFiles,\n      modifiedFiles,\n      newFiles,\n      trace,\n    };\n    const replace = (s: string) => replacePlaceholders(s, ctx);\n\n    const finalPrompt = replace(assertion.prompt);\n    const systemPrompt =\n      assertion.systemPrompt != null && assertion.systemPrompt !== \"\"\n        ? replace(assertion.systemPrompt) + \"\\n\\n\" + DEFAULT_SYSTEM_PROMPT\n        : replace(DEFAULT_SYSTEM_PROMPT_CONTEXT) +\n          \"\\n\\n\" +\n          DEFAULT_SYSTEM_PROMPT;\n\n    const minScore = assertion.minScore ?? DEFAULT_MIN_SCORE;\n    const maxOutputTokens = assertion.maxTokens ?? 1024;\n    const temperature = assertion.temperature ?? 0;\n\n    if (!generateTextStub && !llmConfig) {\n      return {\n        id: randomUUID(),\n        assertionId,\n        assertionType: \"llm_judge\",\n        assertionName: \"LLM judge\",\n        status: AssertionResultStatus.FAILED,\n        message: \"No llmConfig for llm_judge assertion (AI gateway required)\",\n        expected: String(minScore),\n      };\n    }\n\n    try {\n      const result = generateTextStub\n        ? await generateTextStub({\n            prompt: finalPrompt,\n            system: systemPrompt,\n            maxOutputTokens,\n            temperature,\n          })\n        : await this.callGenerateText(\n            llmConfig!,\n            assertion.model ?? DEFAULT_MODEL,\n            finalPrompt,\n            systemPrompt,\n            maxOutputTokens,\n            temperature,\n          );\n\n      let judgeResult: JudgeResult;\n      try {\n        const parsed = JSON.parse(result.text);\n        judgeResult = validateJudgeResult(parsed);\n      } catch (parseErr) {\n        return {\n          id: randomUUID(),\n          assertionId,\n          assertionType: \"llm_judge\",\n          assertionName: \"LLM judge\",\n          status: AssertionResultStatus.FAILED,\n          message: `Failed to parse judge response: ${parseErr instanceof Error ? parseErr.message : String(parseErr)}`,\n          expected: String(minScore),\n          actual: undefined,\n          details: { rawText: result.text.slice(0, 500) },\n        };\n      }\n\n      const passed = judgeResult.score >= minScore;\n      return {\n        id: randomUUID(),\n        assertionId,\n        assertionType: \"llm_judge\",\n        assertionName: \"LLM judge\",\n        status: passed\n          ? AssertionResultStatus.PASSED\n          : AssertionResultStatus.FAILED,\n        message: passed\n          ? `Judge score ${judgeResult.score} >= ${minScore}: ${judgeResult.text}`\n          : `Judge score ${judgeResult.score} < ${minScore}: ${judgeResult.text}`,\n        expected: String(minScore),\n        actual: String(judgeResult.score),\n        details: {\n          score: judgeResult.score,\n          scoreReasoning: judgeResult.scoreReasoning,\n          text: judgeResult.text,\n        },\n      };\n    } catch (err) {\n      const message = err instanceof Error ? err.message : String(err);\n      return {\n        id: randomUUID(),\n        assertionId,\n        assertionType: \"llm_judge\",\n        assertionName: \"LLM judge\",\n        status: AssertionResultStatus.FAILED,\n        message: `LLM judge call failed: ${message}`,\n        expected: String(minScore),\n        details: { error: message },\n      };\n    }\n  }\n\n  private async callGenerateText(\n    llmConfig: LlmConfig,\n    modelId: string,\n    prompt: string,\n    system: string,\n    maxOutputTokens: number,\n    temperature: number,\n  ): Promise<{ text: string }> {\n    const anthropic = createAnthropic({\n      baseURL: llmConfig.baseUrl,\n      apiKey: \"dummy\",\n      headers: llmConfig.headers,\n    });\n    const result = await generateText({\n      model: anthropic(modelId),\n      prompt,\n      system,\n      maxOutputTokens,\n      temperature,\n    });\n    return { text: result.text };\n  }\n}\n"],
-  "mappings": ";AAAA,SAAS,SAAS;AAOX,IAAM,gCAAgC,EAAE,OAAO;AAAA,EACpD,MAAM,EAAE,QAAQ,kBAAkB;AAAA;AAAA,EAElC,WAAW,EAAE,OAAO;AACtB,CAAC;AAUM,IAAM,6BAA6B,EAAE,OAAO;AAAA,EACjD,MAAM,EAAE,QAAQ,cAAc;AAAA;AAAA,EAE9B,SAAS,EAAE,OAAO,EAAE,SAAS;AAAA;AAAA,EAE7B,kBAAkB,EAAE,OAAO,EAAE,IAAI,EAAE,SAAS;AAC9C,CAAC;AASM,IAAM,0BAA0B,EAAE,OAAO;AAAA,EAC9C,MAAM,EAAE,QAAQ,WAAW;AAAA;AAAA,EAE3B,QAAQ,EAAE,OAAO;AAAA;AAAA,EAEjB,cAAc,EAAE,OAAO,EAAE,SAAS;AAAA;AAAA,EAElC,UAAU,EAAE,OAAO,EAAE,IAAI,EAAE,IAAI,CAAC,EAAE,IAAI,GAAG,EAAE,SAAS;AAAA;AAAA,EAEpD,OAAO,EAAE,OAAO,EAAE,SAAS;AAAA,EAC3B,WAAW,EAAE,OAAO,EAAE,IAAI,EAAE,SAAS;AAAA,EACrC,aAAa,EAAE,OAAO,EAAE,IAAI,CAAC,EAAE,IAAI,CAAC,EAAE,SAAS;AACjD,CAAC;AASM,IAAM,kBAAkB,EAAE,MAAM;AAAA,EACrC;AAAA,EACA;AAAA,EACA;AACF,CAAC;;;AC7DD,SAAS,KAAAA,UAAS;AAKX,IAAM,mBAAmBA,GAAE,OAAO;AAAA,EACvC,QAAQA,GAAE,OAAO;AAAA,EACjB,YAAYA,GAAE,OAAO;AAAA,EACrB,OAAOA,GAAE,OAAO;AAClB,CAAC;AAOM,IAAK,cAAL,kBAAKC,iBAAL;AACL,EAAAA,aAAA,gBAAa;AACb,EAAAA,aAAA,cAAW;AACX,EAAAA,aAAA,iBAAc;AACd,EAAAA,aAAA,cAAW;AAJD,SAAAA;AAAA,GAAA;AAUL,IAAM,qBAAqBD,GAAE,OAAO;AAAA,EACzC,IAAIA,GAAE,OAAO;AAAA,EACb,YAAYA,GAAE,OAAO;AAAA,EACrB,MAAMA,GAAE,KAAK,WAAW;AAAA,EACxB,OAAOA,GAAE,OAAO;AAAA,EAChB,UAAUA,GAAE,OAAO;AAAA,EACnB,WAAWA,GAAE,OAAO;AAAA,EACpB,YAAYA,GAAE,OAAO;AAAA,EACrB,YAAY;AAAA,EACZ,SAASA,GAAE,OAAO;AAAA,EAClB,UAAUA,GAAE,OAAO,EAAE,SAAS;AAAA,EAC9B,eAAeA,GAAE,OAAO,EAAE,SAAS;AAAA,EACnC,cAAcA,GAAE,OAAO,EAAE,SAAS;AAAA,EAClC,eAAeA,GAAE,OAAO,EAAE,SAAS;AAAA,EACnC,SAASA,GAAE,QAAQ;AAAA,EACnB,OAAOA,GAAE,OAAO,EAAE,SAAS;AAC7B,CAAC;AAOM,IAAM,0BAA0BA,GAAE,OAAO;AAAA,EAC9C,OAAOA,GAAE,OAAO;AAAA,EAChB,YAAYA,GAAE,OAAO;AAAA,EACrB,QAAQA,GAAE,OAAO;AAAA,EACjB,SAASA,GAAE,OAAO;AACpB,CAAC;AAOM,IAAM,wBAAwBA,GAAE,OAAO;AAAA,EAC5C,YAAYA,GAAE,OAAO;AAAA,EACrB,iBAAiBA,GAAE,OAAO;AAAA,EAC1B,aAAa;AAAA,EACb,cAAcA,GAAE,OAAO;AAAA,EACvB,mBAAmBA,GAAE,OAAOA,GAAE,OAAO,GAAG,uBAAuB,EAAE,SAAS;AAAA,EAC1E,gBAAgBA,GAAE,OAAOA,GAAE,OAAO,GAAG,uBAAuB;AAAA,EAC5D,YAAYA,GAAE,MAAMA,GAAE,OAAO,CAAC;AAChC,CAAC;AAOM,IAAM,iBAAiBA,GAAE,OAAO;AAAA,EACrC,IAAIA,GAAE,OAAO;AAAA,EACb,OAAOA,GAAE,MAAM,kBAAkB;AAAA,EACjC,SAAS;AACX,CAAC;;;AChFD,SAAS,KAAAE,UAAS;AAMX,IAAK,wBAAL,kBAAKC,2BAAL;AACL,EAAAA,uBAAA,YAAS;AACT,EAAAA,uBAAA,YAAS;AACT,EAAAA,uBAAA,aAAU;AACV,EAAAA,uBAAA,WAAQ;AAJE,SAAAA;AAAA,GAAA;AAUL,IAAM,wBAAwBC,GAAE,OAAO;AAAA,EAC5C,IAAIA,GAAE,OAAO;AAAA,EACb,aAAaA,GAAE,OAAO;AAAA,EACtB,eAAeA,GAAE,OAAO;AAAA,EACxB,eAAeA,GAAE,OAAO;AAAA,EACxB,QAAQA,GAAE,KAAK,qBAAqB;AAAA,EACpC,SAASA,GAAE,OAAO,EAAE,SAAS;AAAA,EAC7B,UAAUA,GAAE,OAAO,EAAE,SAAS;AAAA,EAC9B,QAAQA,GAAE,OAAO,EAAE,SAAS;AAAA,EAC5B,UAAUA,GAAE,OAAO,EAAE,SAAS;AAAA,EAC9B,SAASA,GAAE,OAAOA,GAAE,OAAO,GAAGA,GAAE,QAAQ,CAAC,EAAE,SAAS;AAAA,EACpD,eAAeA,GAAE,MAAM,kBAAkB,EAAE,SAAS;AACtD,CAAC;;;AC1BD,SAAS,cAAAC,mBAAkB;;;ACK3B,SAAS,kBAAkB;;;ACyCpB,IAAe,qBAAf,MAAmE;AAQ1E;;;AD1CA,SAAS,6BACP,MACA,mBACS;AACT,MAAI,SAAS,QAAQ,OAAO,SAAS,UAAU;AAC7C,WAAO;AAAA,EACT;AACA,QAAM,MAAM;AACZ,QAAM,OAAO,IAAI;AACjB,SAAO,OAAO,SAAS,YAAY,SAAS;AAC9C;AAMO,IAAM,0BAAN,cAAsC,mBAA4C;AAAA,EAC9E,OAAO;AAAA,EAEhB,SACE,WACA,OAEA,UACiB;AACjB,UAAM,cAAc,WAAW;AAE/B,UAAM,WAAiC,MAAM;AAC7C,QAAI,CAAC,UAAU,OAAO,QAAQ;AAC5B,aAAO;AAAA,QACL,IAAI,WAAW;AAAA,QACf;AAAA,QACA,eAAe;AAAA,QACf,eAAe;AAAA,QACf;AAAA,QACA,SAAS;AAAA,QACT,UAAU,UAAU;AAAA,MACtB;AAAA,IACF;AAEA,eAAW,QAAQ,SAAS,OAAO;AACjC,UAAI,KAAK,aAAa,SAAS;AAC7B;AAAA,MACF;AACA,UAAI;AACJ,UAAI;AACF,eAAO,KAAK,gBACP,KAAK,MAAM,KAAK,aAAa,IAC9B;AAAA,MACN,QAAQ;AACN;AAAA,MACF;AACA,UAAI,6BAA6B,MAAM,UAAU,SAAS,GAAG;AAC3D,eAAO;AAAA,UACL,IAAI,WAAW;AAAA,UACf;AAAA,UACA,eAAe;AAAA,UACf,eAAe;AAAA,UACf;AAAA,UACA,SAAS,UAAU,UAAU,SAAS;AAAA,UACtC,UAAU,UAAU;AAAA,QACtB;AAAA,MACF;AAAA,IACF;AAEA,WAAO;AAAA,MACL,IAAI,WAAW;AAAA,MACf;AAAA,MACA,eAAe;AAAA,MACf,eAAe;AAAA,MACf;AAAA,MACA,SAAS,UAAU,UAAU,SAAS;AAAA,MACtC,UAAU,UAAU;AAAA,IACtB;AAAA,EACF;AACF;;;AEnFA,SAAS,cAAAC,mBAAkB;AAC3B,SAAS,gBAAgB;AAIzB,IAAM,kBAAkB;AACxB,IAAM,oBAAoB;AAMnB,IAAM,uBAAN,cAAmC,mBAAyC;AAAA,EACxE,OAAO;AAAA,EAEhB,SACE,WACA,QACA,SACiB;AACjB,UAAM,cAAcC,YAAW;AAC/B,UAAM,UAAU,SAAS;AACzB,UAAM,UAAU,UAAU,WAAW;AACrC,UAAM,mBAAmB,UAAU,oBAAoB;AAEvD,QAAI,CAAC,SAAS;AACZ,aAAO,KAAK,aAAa,aAAa;AAAA,QACpC;AAAA,QACA,SAAS;AAAA,QACT,UAAU,OAAO,gBAAgB;AAAA,MACnC,CAAC;AAAA,IACH;AAEA,QAAI,WAA0B;AAC9B,QAAI,eAA8B;AAClC,QAAI;AACJ,QAAI;AAEJ,YAAQ,IAAI,2BAA2B,OAAO,SAAS,OAAO,EAAE;AAEhE,QAAI;AACF,eAAS,SAAS;AAAA,QAChB,KAAK;AAAA,QACL,UAAU;AAAA,QACV,OAAO,CAAC,UAAU,QAAQ,MAAM;AAAA,MAClC,CAAC;AACD,iBAAW;AAAA,IACb,SAAS,KAAK;AACZ,YAAM,QAAQ;AAMd,iBACE,OAAO,MAAM,WAAW,WACpB,MAAM,SACN,OAAO,MAAM,SAAS,WACpB,MAAM,OACN;AACR,qBAAe,MAAM;AACrB,eAAS,KAAK,eAAe,MAAM,MAAM;AACzC,eAAS,KAAK,eAAe,MAAM,MAAM;AAAA,IAC3C;AAEA,UAAM,SAAS,aAAa,QAAQ,aAAa;AAEjD,UAAM,UAAmC,EAAE,SAAS,QAAQ;AAC5D,QAAI,WAAW,UAAa,WAAW,IAAI;AACzC,cAAQ,SAAS;AAAA,IACnB;AACA,QAAI,WAAW,UAAa,WAAW,IAAI;AACzC,cAAQ,SAAS;AAAA,IACnB;AAEA,WAAO,KAAK,aAAa,aAAa;AAAA,MACpC,QAAQ;AAAA,MAGR,SAAS,KAAK,cAAc,UAAU,kBAAkB,YAAY;AAAA,MACpE,UAAU,OAAO,gBAAgB;AAAA,MACjC,QAAQ,aAAa,OAAO,OAAO,QAAQ,IAAI;AAAA,MAC/C;AAAA,IACF,CAAC;AAAA,EACH;AAAA,EAEQ,aACN,aACA,QACiB;AACjB,WAAO;AAAA,MACL,IAAIA,YAAW;AAAA,MACf;AAAA,MACA,eAAe;AAAA,MACf,eAAe;AAAA,MACf;AAAA,MACA,GAAG;AAAA,IACL;AAAA,EACF;AAAA,EAEQ,eACN,OACoB;AACpB,QAAI,UAAU,UAAa,UAAU,KAAM,QAAO;AAClD,QAAI,OAAO,UAAU,SAAU,QAAO;AACtC,WAAO,MAAM,SAAS,OAAO;AAAA,EAC/B;AAAA,EAEQ,cACN,UACA,kBACA,cACQ;AACR,QAAI,aAAa,MAAM;AACrB,aAAO,iBAAiB,YAAY;AAAA,IACtC;AACA,QAAI,aAAa,kBAAkB;AACjC,aAAO,2BAA2B,QAAQ;AAAA,IAC5C;AACA,WAAO,qBAAqB,QAAQ,cAAc,gBAAgB;AAAA,EACpE;AACF;;;ACxHA,SAAS,cAAAC,mBAAkB;AAC3B,SAAS,uBAAuB;AAChC,SAAS,oBAAoB;AAatB,SAAS,oBAAoB,UAAwC;AAC1E,MAAI,CAAC,UAAU,OAAO,QAAQ;AAC5B,WAAO;AAAA,EACT;AACA,QAAM,QAAkB,CAAC;AACzB,aAAW,QAAQ,SAAS,OAAO;AACjC,UAAM,QAAkB;AAAA,MACtB,QAAQ,KAAK,UAAU;AAAA,MACvB,SAAS,KAAK,IAAI;AAAA,MAClB,aAAa,KAAK,UAAU;AAAA,IAC9B;AACA,QAAI,KAAK,UAAU;AACjB,YAAM,KAAK,SAAS,KAAK,QAAQ,EAAE;AACnC,UAAI,KAAK,eAAe;AACtB,cAAM,KAAK,SAAS,KAAK,aAAa,EAAE;AAAA,MAC1C;AAAA,IACF;AACA,QAAI,KAAK,eAAe;AACtB,YAAM,KAAK,WAAW,KAAK,aAAa,EAAE;AAAA,IAC5C;AACA,QAAI,KAAK,OAAO;AACd,YAAM,KAAK,UAAU,KAAK,KAAK,EAAE;AAAA,IACnC;AACA,UAAM,KAAK,MAAM,KAAK,IAAI,CAAC;AAAA,EAC7B;AACA,SAAO,MAAM,KAAK,IAAI;AACxB;AAcO,SAAS,oBACd,KACA,KACQ;AACR,SAAO,IACJ,QAAQ,mBAAmB,IAAI,MAAM,EACrC,QAAQ,gBAAgB,IAAI,GAAG,EAC/B,QAAQ,yBAAyB,IAAI,YAAY,EACjD,QAAQ,0BAA0B,IAAI,aAAa,EACnD,QAAQ,qBAAqB,IAAI,QAAQ,EACzC,QAAQ,kBAAkB,IAAI,KAAK;AACxC;AAEO,SAAS,oBAAoB,QAA8B;AAChE,MAAI,WAAW,QAAQ,OAAO,WAAW,UAAU;AACjD,UAAM,IAAI,MAAM,+BAA+B;AAAA,EACjD;AACA,QAAM,MAAM;AACZ,MAAI,OAAO,IAAI,SAAS,UAAU;AAChC,UAAM,IAAI,MAAM,kDAAkD;AAAA,EACpE;AACA,MAAI,OAAO,IAAI,UAAU,UAAU;AACjC,UAAM,IAAI,MAAM,mDAAmD;AAAA,EACrE;AACA,MAAI,IAAI,QAAQ,KAAK,IAAI,QAAQ,KAAK;AACpC,UAAM,IAAI,MAAM,6CAA6C;AAAA,EAC/D;AACA,MAAI,OAAO,IAAI,mBAAmB,UAAU;AAC1C,UAAM,IAAI;AAAA,MACR;AAAA,IACF;AAAA,EACF;AACA,SAAO;AAAA,IACL,MAAM,IAAI;AAAA,IACV,OAAO,IAAI;AAAA,IACX,gBAAgB,IAAI;AAAA,EACtB;AACF;AAEA,IAAM,oBAAoB;AAC1B,IAAM,gBAAgB;AAGtB,IAAM,gCAAgC;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAUtC,IAAM,wBAAwB;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAwBvB,IAAM,oBAAN,cAAgC,mBAAsC;AAAA,EAClE,OAAO;AAAA,EAEhB,MAAM,SACJ,WACA,OACA,SAC0B;AAC1B,UAAM,cAAcC,YAAW;AAC/B,UAAM,YAAY,SAAS;AAC3B,UAAM,UAAU,SAAS,WAAW;AACpC,UAAM,mBAAmB,SAAS;AAElC,UAAM,SAAS,MAAM,cAAc;AACnC,UAAM,YAAY,MAAM,aAAa,CAAC;AAGtC,UAAM,eAAe,UAAU,IAAI,CAAC,MAAM,EAAE,IAAI;AAChD,UAAM,gBAAgB,UACnB,OAAO,CAAC,MAAM,EAAE,WAAW,UAAU,EACrC,IAAI,CAAC,MAAM,EAAE,IAAI;AACpB,UAAM,WAAW,UACd,OAAO,CAAC,MAAM,EAAE,WAAW,KAAK,EAChC,IAAI,CAAC,MAAM,EAAE,IAAI;AAEpB,UAAM,eACJ,aAAa,SAAS,IAClB,aAAa,IAAI,CAAC,SAAiB,KAAK,IAAI,EAAE,EAAE,KAAK,IAAI,IACzD;AACN,UAAM,gBACJ,cAAc,SAAS,IACnB,cAAc,IAAI,CAAC,SAAiB,KAAK,IAAI,EAAE,EAAE,KAAK,IAAI,IAC1D;AACN,UAAM,WACJ,SAAS,SAAS,IACd,SAAS,IAAI,CAAC,SAAiB,KAAK,IAAI,EAAE,EAAE,KAAK,IAAI,IACrD;AAEN,UAAM,QAAQ,oBAAoB,MAAM,QAAQ;AAChD,UAAM,MAA0B;AAAA,MAC9B;AAAA,MACA,KAAK;AAAA,MACL;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF;AACA,UAAM,UAAU,CAAC,MAAc,oBAAoB,GAAG,GAAG;AAEzD,UAAM,cAAc,QAAQ,UAAU,MAAM;AAC5C,UAAM,eACJ,UAAU,gBAAgB,QAAQ,UAAU,iBAAiB,KACzD,QAAQ,UAAU,YAAY,IAAI,SAAS,wBAC3C,QAAQ,6BAA6B,IACrC,SACA;AAEN,UAAM,WAAW,UAAU,YAAY;AACvC,UAAM,kBAAkB,UAAU,aAAa;AAC/C,UAAM,cAAc,UAAU,eAAe;AAE7C,QAAI,CAAC,oBAAoB,CAAC,WAAW;AACnC,aAAO;AAAA,QACL,IAAIA,YAAW;AAAA,QACf;AAAA,QACA,eAAe;AAAA,QACf,eAAe;AAAA,QACf;AAAA,QACA,SAAS;AAAA,QACT,UAAU,OAAO,QAAQ;AAAA,MAC3B;AAAA,IACF;AAEA,QAAI;AACF,YAAM,SAAS,mBACX,MAAM,iBAAiB;AAAA,QACrB,QAAQ;AAAA,QACR,QAAQ;AAAA,QACR;AAAA,QACA;AAAA,MACF,CAAC,IACD,MAAM,KAAK;AAAA,QACT;AAAA,QACA,UAAU,SAAS;AAAA,QACnB;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,MACF;AAEJ,UAAI;AACJ,UAAI;AACF,cAAM,SAAS,KAAK,MAAM,OAAO,IAAI;AACrC,sBAAc,oBAAoB,MAAM;AAAA,MAC1C,SAAS,UAAU;AACjB,eAAO;AAAA,UACL,IAAIA,YAAW;AAAA,UACf;AAAA,UACA,eAAe;AAAA,UACf,eAAe;AAAA,UACf;AAAA,UACA,SAAS,mCAAmC,oBAAoB,QAAQ,SAAS,UAAU,OAAO,QAAQ,CAAC;AAAA,UAC3G,UAAU,OAAO,QAAQ;AAAA,UACzB,QAAQ;AAAA,UACR,SAAS,EAAE,SAAS,OAAO,KAAK,MAAM,GAAG,GAAG,EAAE;AAAA,QAChD;AAAA,MACF;AAEA,YAAM,SAAS,YAAY,SAAS;AACpC,aAAO;AAAA,QACL,IAAIA,YAAW;AAAA,QACf;AAAA,QACA,eAAe;AAAA,QACf,eAAe;AAAA,QACf,QAAQ;AAAA,QAGR,SAAS,SACL,eAAe,YAAY,KAAK,OAAO,QAAQ,KAAK,YAAY,IAAI,KACpE,eAAe,YAAY,KAAK,MAAM,QAAQ,KAAK,YAAY,IAAI;AAAA,QACvE,UAAU,OAAO,QAAQ;AAAA,QACzB,QAAQ,OAAO,YAAY,KAAK;AAAA,QAChC,SAAS;AAAA,UACP,OAAO,YAAY;AAAA,UACnB,gBAAgB,YAAY;AAAA,UAC5B,MAAM,YAAY;AAAA,QACpB;AAAA,MACF;AAAA,IACF,SAAS,KAAK;AACZ,YAAM,UAAU,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAC/D,aAAO;AAAA,QACL,IAAIA,YAAW;AAAA,QACf;AAAA,QACA,eAAe;AAAA,QACf,eAAe;AAAA,QACf;AAAA,QACA,SAAS,0BAA0B,OAAO;AAAA,QAC1C,UAAU,OAAO,QAAQ;AAAA,QACzB,SAAS,EAAE,OAAO,QAAQ;AAAA,MAC5B;AAAA,IACF;AAAA,EACF;AAAA,EAEA,MAAc,iBACZ,WACA,SACA,QACA,QACA,iBACA,aAC2B;AAC3B,UAAM,YAAY,gBAAgB;AAAA,MAChC,SAAS,UAAU;AAAA,MACnB,QAAQ;AAAA,MACR,SAAS,UAAU;AAAA,IACrB,CAAC;AACD,UAAM,SAAS,MAAM,aAAa;AAAA,MAChC,OAAO,UAAU,OAAO;AAAA,MACxB;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF,CAAC;AACD,WAAO,EAAE,MAAM,OAAO,KAAK;AAAA,EAC7B;AACF;;;AJtSA,IAAM,oBAAoB,IAAI,kBAAkB;AAEhD,IAAM,aAAiD;AAAA,EACrD,kBAAkB,IAAI,wBAAwB;AAAA,EAC9C,cAAc,IAAI,qBAAqB;AAAA,EACvC,WAAW;AAAA;AAAA,EAEX,QAAQ;AACV;AAQO,SAAS,kBACd,MACA,WACM;AACN,aAAW,IAAI,IAAI;AACrB;AAQO,SAAS,aAAa,MAA8C;AACzE,SAAO,WAAW,IAAI;AACxB;AAUA,eAAsB,mBACpB,OACA,YACA,SAC4B;AAC5B,MAAI,WAAW,WAAW,GAAG;AAC3B,WAAO,CAAC;AAAA,EACV;AACA,SAAO,QAAQ;AAAA,IACb,WAAW,IAAI,OAAO,cAAc;AAClC,YAAM,YAAY,WAAW,UAAU,IAAI;AAC3C,UAAI,CAAC,WAAW;AACd,eAAO;AAAA,UACL,IAAIC,YAAW;AAAA,UACf,aAAaA,YAAW;AAAA,UACxB,eAAe,UAAU;AAAA,UACzB,eAAe;AAAA,UACf;AAAA,UACA,SAAS,+BAA+B,UAAU,IAAI;AAAA,UACtD,UAAU;AAAA,QACZ;AAAA,MACF;AACA,YAAM,UAAU,KAAK,IAAI;AACzB,YAAM,SAAS,MAAM,UAAU,SAAS,WAAW,OAAO,OAAO;AACjE,YAAM,aAAa,KAAK,IAAI,IAAI;AAChC,aAAO,EAAE,GAAG,QAAQ,UAAU,WAAW;AAAA,IAC3C,CAAC;AAAA,EACH;AACF;",
+  "sourcesContent": ["import { z } from \"zod\";\n\n/**\n * Assertion: the agent must have invoked a specific skill during the run.\n * Checked by inspecting the LLM trace for a \"Skill\" tool use with the given skill.\n * Data: skillName (the skill that must have been called).\n */\nexport const SkillWasCalledAssertionSchema = z.object({\n  type: z.literal(\"skill_was_called\"),\n  /** Name of the skill that must have been called (matched against trace Skill tool args) */\n  skillName: z.string(),\n});\n\nexport type SkillWasCalledAssertion = z.infer<\n  typeof SkillWasCalledAssertionSchema\n>;\n\n/**\n * Assertion: a build command must exit with the expected code (default 0).\n * Runs the command in the scenario working directory.\n */\nexport const BuildPassedAssertionSchema = z.object({\n  type: z.literal(\"build_passed\"),\n  /** Command to run (default: \"yarn build\") */\n  command: z.string().optional(),\n  /** Expected exit code (default: 0) */\n  expectedExitCode: z.number().int().optional(),\n});\n\nexport type BuildPassedAssertion = z.infer<typeof BuildPassedAssertionSchema>;\n\n/**\n * Assertion: an LLM judges the scenario output (score 0-100).\n * Prompt can use {{output}}, {{cwd}}, {{changedFiles}}, {{modifiedFiles}}, {{newFiles}}, {{trace}}.\n * Passes if judge score >= minScore.\n */\nexport const LlmJudgeAssertionSchema = z.object({\n  type: z.literal(\"llm_judge\"),\n  /** Prompt template; placeholders: {{output}}, {{cwd}}, {{changedFiles}}, {{modifiedFiles}}, {{newFiles}}, {{trace}} */\n  prompt: z.string(),\n  /** Optional system prompt for the judge (default asks for JSON with score) */\n  systemPrompt: z.string().optional(),\n  /** Minimum score to pass (0-100, default 70) */\n  minScore: z.number().int().min(0).max(100).optional(),\n  /** Model for the judge (e.g. claude-3-5-haiku) */\n  model: z.string().optional(),\n  maxTokens: z.number().int().optional(),\n  temperature: z.number().min(0).max(1).optional(),\n});\n\nexport type LlmJudgeAssertion = z.infer<typeof LlmJudgeAssertionSchema>;\n\n/**\n * Union of all assertion types.\n * Each assertion has a type and type-specific data.\n * Uses z.union (not z.discriminatedUnion) for Zod v4 compatibility when used as array element.\n */\nexport const AssertionSchema = z.union([\n  SkillWasCalledAssertionSchema,\n  BuildPassedAssertionSchema,\n  LlmJudgeAssertionSchema,\n]);\n\nexport type Assertion = z.infer<typeof AssertionSchema>;\n", "import { z } from \"zod\";\n\n/**\n * Token usage schema.\n */\nexport const TokenUsageSchema = z.object({\n  prompt: z.number(),\n  completion: z.number(),\n  total: z.number(),\n});\n\nexport type TokenUsage = z.infer<typeof TokenUsageSchema>;\n\n/**\n * LLM step type enum.\n */\nexport enum LLMStepType {\n  COMPLETION = \"completion\",\n  TOOL_USE = \"tool_use\",\n  TOOL_RESULT = \"tool_result\",\n  THINKING = \"thinking\",\n}\n\n/**\n * LLM trace step schema.\n */\nexport const LLMTraceStepSchema = z.object({\n  id: z.string(),\n  stepNumber: z.number(),\n  type: z.enum(LLMStepType),\n  model: z.string(),\n  provider: z.string(),\n  startedAt: z.string(),\n  durationMs: z.number(),\n  tokenUsage: TokenUsageSchema,\n  costUsd: z.number(),\n  toolName: z.string().optional(),\n  toolArguments: z.string().optional(),\n  inputPreview: z.string().optional(),\n  outputPreview: z.string().optional(),\n  success: z.boolean(),\n  error: z.string().optional(),\n});\n\nexport type LLMTraceStep = z.infer<typeof LLMTraceStepSchema>;\n\n/**\n * LLM breakdown stats schema.\n */\nexport const LLMBreakdownStatsSchema = z.object({\n  count: z.number(),\n  durationMs: z.number(),\n  tokens: z.number(),\n  costUsd: z.number(),\n});\n\nexport type LLMBreakdownStats = z.infer<typeof LLMBreakdownStatsSchema>;\n\n/**\n * LLM trace summary schema.\n */\nexport const LLMTraceSummarySchema = z.object({\n  totalSteps: z.number(),\n  totalDurationMs: z.number(),\n  totalTokens: TokenUsageSchema,\n  totalCostUsd: z.number(),\n  stepTypeBreakdown: z.record(z.string(), LLMBreakdownStatsSchema).optional(),\n  modelBreakdown: z.record(z.string(), LLMBreakdownStatsSchema),\n  modelsUsed: z.array(z.string()),\n});\n\nexport type LLMTraceSummary = z.infer<typeof LLMTraceSummarySchema>;\n\n/**\n * LLM trace schema.\n */\nexport const LLMTraceSchema = z.object({\n  id: z.string(),\n  steps: z.array(LLMTraceStepSchema),\n  summary: LLMTraceSummarySchema,\n});\n\nexport type LLMTrace = z.infer<typeof LLMTraceSchema>;\n", "import { z } from \"zod\";\nimport { LLMTraceStepSchema } from \"./trace.js\";\n\n/**\n * Assertion result status enum.\n */\nexport enum AssertionResultStatus {\n  PASSED = \"passed\",\n  FAILED = \"failed\",\n  SKIPPED = \"skipped\",\n  ERROR = \"error\",\n}\n\n/**\n * Assertion result schema.\n */\nexport const AssertionResultSchema = z.object({\n  id: z.string(),\n  assertionId: z.string(),\n  assertionType: z.string(),\n  assertionName: z.string(),\n  status: z.enum(AssertionResultStatus),\n  message: z.string().optional(),\n  expected: z.string().optional(),\n  actual: z.string().optional(),\n  duration: z.number().optional(),\n  details: z.record(z.string(), z.unknown()).optional(),\n  llmTraceSteps: z.array(LLMTraceStepSchema).optional(),\n});\n\nexport type AssertionResult = z.infer<typeof AssertionResultSchema>;\n", "import type { Assertion, AssertionResult } from \"../types/index.js\";\nimport { AssertionResultStatus } from \"../types/index.js\";\nimport { randomUUID } from \"crypto\";\nimport type { AssertionContext } from \"./assertion-evaluator.js\";\nimport { AssertionEvaluator } from \"./assertion-evaluator.js\";\nimport { SkillWasCalledEvaluator } from \"./skill-was-called-evaluator.js\";\nimport { BuildPassedEvaluator } from \"./build-passed-evaluator.js\";\nimport { LlmJudgeEvaluator } from \"./llm-judge-evaluator.js\";\nimport type { EvaluationInput } from \"../types/index.js\";\n\nconst llmJudgeEvaluator = new LlmJudgeEvaluator();\n\nconst evaluators: Record<string, AssertionEvaluator> = {\n  skill_was_called: new SkillWasCalledEvaluator(),\n  build_passed: new BuildPassedEvaluator(),\n  llm_judge: llmJudgeEvaluator,\n  // Custom assertions use the same LLM-based evaluation as llm_judge\n  custom: llmJudgeEvaluator,\n};\n\n/**\n * Register a custom assertion evaluator.\n *\n * @param type - The assertion type identifier\n * @param evaluator - The evaluator instance\n */\nexport function registerEvaluator(\n  type: string,\n  evaluator: AssertionEvaluator,\n): void {\n  evaluators[type] = evaluator;\n}\n\n/**\n * Get a registered evaluator by type.\n *\n * @param type - The assertion type identifier\n * @returns The evaluator or undefined if not found\n */\nexport function getEvaluator(type: string): AssertionEvaluator | undefined {\n  return evaluators[type];\n}\n\n/**\n * Evaluate all assertions against the input.\n *\n * @param input - Evaluation input (includes outputText, llmTrace, fileDiffs)\n * @param assertions - List of assertions to evaluate\n * @param context - Optional context (e.g. workDir for build_passed, llmConfig for llm_judge)\n * @returns Array of assertion results; empty if no assertions\n */\nexport async function evaluateAssertions(\n  input: EvaluationInput,\n  assertions: Assertion[],\n  context?: AssertionContext,\n): Promise<AssertionResult[]> {\n  if (assertions.length === 0) {\n    return [];\n  }\n  return Promise.all(\n    assertions.map(async (assertion) => {\n      const evaluator = evaluators[assertion.type];\n      if (!evaluator) {\n        return {\n          id: randomUUID(),\n          assertionId: randomUUID(),\n          assertionType: assertion.type,\n          assertionName: \"Unknown assertion\",\n          status: AssertionResultStatus.ERROR,\n          message: `Unsupported assertion type: ${assertion.type}`,\n          duration: 0,\n        };\n      }\n      const startMs = Date.now();\n      const result = await evaluator.evaluate(assertion, input, context);\n      const durationMs = Date.now() - startMs;\n      return { ...result, duration: durationMs };\n    }),\n  );\n}\n\n// Re-export evaluator classes and types\nexport { AssertionEvaluator } from \"./assertion-evaluator.js\";\nexport type {\n  AssertionContext,\n  LlmConfig,\n  LlmJudgeGenerateTextOptions,\n} from \"./assertion-evaluator.js\";\nexport { SkillWasCalledEvaluator } from \"./skill-was-called-evaluator.js\";\nexport { BuildPassedEvaluator } from \"./build-passed-evaluator.js\";\nexport {\n  LlmJudgeEvaluator,\n  formatTraceForJudge,\n  replacePlaceholders,\n  validateJudgeResult,\n  type JudgeResult,\n} from \"./llm-judge-evaluator.js\";\n", "import type {\n  SkillWasCalledAssertion,\n  AssertionResult,\n  LLMTrace,\n  EvaluationInput,\n} from \"../types/index.js\";\nimport { AssertionResultStatus } from \"../types/index.js\";\nimport { randomUUID } from \"crypto\";\nimport type { AssertionContext } from \"./assertion-evaluator.js\";\nimport { AssertionEvaluator } from \"./assertion-evaluator.js\";\n\n/**\n * Check if parsed Skill tool args reference the expected skill by name.\n */\nfunction toolArgsReferenceSkillByName(\n  args: unknown,\n  expectedSkillName: string,\n): boolean {\n  if (args === null || typeof args !== \"object\") {\n    return false;\n  }\n  const obj = args as Record<string, unknown>;\n  const name = obj.skill;\n  return typeof name === \"string\" && name === expectedSkillName;\n}\n\n/**\n * Evaluator for \"skill_was_called\" assertion: the LLM trace must contain a step\n * where the \"Skill\" tool was used with the expected skill (by name).\n */\nexport class SkillWasCalledEvaluator extends AssertionEvaluator<SkillWasCalledAssertion> {\n  readonly type = \"skill_was_called\" as const;\n\n  evaluate(\n    assertion: SkillWasCalledAssertion,\n    input: EvaluationInput,\n    // eslint-disable-next-line @typescript-eslint/no-unused-vars -- context not used for skill_was_called\n    _context?: AssertionContext,\n  ): AssertionResult {\n    const assertionId = randomUUID();\n\n    const llmTrace: LLMTrace | undefined = input.llmTrace;\n    if (!llmTrace?.steps?.length) {\n      return {\n        id: randomUUID(),\n        assertionId,\n        assertionType: \"skill_was_called\",\n        assertionName: \"Skill was called\",\n        status: AssertionResultStatus.FAILED,\n        message: \"No LLM trace steps to check for skill invocation\",\n        expected: assertion.skillName,\n      };\n    }\n\n    for (const step of llmTrace.steps) {\n      if (step.toolName !== \"Skill\") {\n        continue;\n      }\n      let args: unknown;\n      try {\n        args = step.toolArguments\n          ? (JSON.parse(step.toolArguments) as unknown)\n          : undefined;\n      } catch {\n        continue;\n      }\n      if (toolArgsReferenceSkillByName(args, assertion.skillName)) {\n        return {\n          id: randomUUID(),\n          assertionId,\n          assertionType: \"skill_was_called\",\n          assertionName: \"Skill was called\",\n          status: AssertionResultStatus.PASSED,\n          message: `Skill \"${assertion.skillName}\" was called`,\n          expected: assertion.skillName,\n        };\n      }\n    }\n\n    return {\n      id: randomUUID(),\n      assertionId,\n      assertionType: \"skill_was_called\",\n      assertionName: \"Skill was called\",\n      status: AssertionResultStatus.FAILED,\n      message: `Skill \"${assertion.skillName}\" was not called`,\n      expected: assertion.skillName,\n    };\n  }\n}\n", "import type {\n  Assertion,\n  AssertionResult,\n  EvaluationInput,\n} from \"../types/index.js\";\n\n/**\n * Options passed to the LLM for llm_judge. Used by the optional stub for testing.\n */\nexport interface LlmJudgeGenerateTextOptions {\n  prompt: string;\n  system: string;\n  maxOutputTokens: number;\n  temperature: number;\n}\n\n/**\n * Configuration for LLM calls (used by llm_judge assertion).\n */\nexport interface LlmConfig {\n  /** Base URL for the AI API (e.g., 'https://api.anthropic.com') */\n  baseUrl: string;\n  /** Headers to include in API requests (e.g., API key) */\n  headers: Record<string, string>;\n}\n\n/**\n * Optional context passed when evaluating assertions.\n */\nexport interface AssertionContext {\n  /** Working directory for the scenario (used by build_passed) */\n  workDir?: string;\n  /** LLM configuration (used by llm_judge) */\n  llmConfig?: LlmConfig;\n  /**\n   * Optional stub for llm_judge: when set, the evaluator uses this instead of the real AI call.\n   * Used only in tests to avoid hitting the API.\n   */\n  generateTextForLlmJudge?: (\n    options: LlmJudgeGenerateTextOptions,\n  ) => Promise<{ text: string }>;\n}\n\n/**\n * Abstract base for assertion evaluators.\n * Each assertion type has a concrete class that implements evaluate().\n * evaluate() may return a Promise for async assertions (e.g. llm_judge).\n */\nexport abstract class AssertionEvaluator<T extends Assertion = Assertion> {\n  abstract readonly type: T[\"type\"];\n\n  abstract evaluate(\n    assertion: T,\n    input: EvaluationInput,\n    context?: AssertionContext,\n  ): AssertionResult | Promise<AssertionResult>;\n}\n", "import type {\n  BuildPassedAssertion,\n  AssertionResult,\n  EvaluationInput,\n} from \"../types/index.js\";\nimport { AssertionResultStatus } from \"../types/index.js\";\nimport { randomUUID } from \"crypto\";\nimport { execSync } from \"child_process\";\nimport type { AssertionContext } from \"./assertion-evaluator.js\";\nimport { AssertionEvaluator } from \"./assertion-evaluator.js\";\n\nconst DEFAULT_COMMAND = \"yarn build\";\nconst DEFAULT_EXIT_CODE = 0;\n\n/**\n * Evaluator for \"build_passed\" assertion: runs a build command in the scenario\n * working directory and passes if the command exits with the expected code (default 0).\n */\nexport class BuildPassedEvaluator extends AssertionEvaluator<BuildPassedAssertion> {\n  readonly type = \"build_passed\" as const;\n\n  evaluate(\n    assertion: BuildPassedAssertion,\n    _input: EvaluationInput,\n    context?: AssertionContext,\n  ): AssertionResult {\n    const assertionId = randomUUID();\n    const workDir = context?.workDir;\n    const command = assertion.command ?? DEFAULT_COMMAND;\n    const expectedExitCode = assertion.expectedExitCode ?? DEFAULT_EXIT_CODE;\n\n    if (!workDir) {\n      return this.createResult(assertionId, {\n        status: AssertionResultStatus.FAILED,\n        message: \"No working directory provided for build_passed assertion\",\n        expected: String(expectedExitCode),\n      });\n    }\n\n    let exitCode: number | null = null;\n    let errorMessage: string | null = null;\n    let stdout: string | undefined;\n    let stderr: string | undefined;\n\n    console.log(`[build_passed] Running \"${command}\" in: ${workDir}`);\n\n    try {\n      execSync(command, {\n        cwd: workDir,\n        encoding: \"utf-8\",\n        stdio: [\"ignore\", \"pipe\", \"pipe\"],\n      });\n      exitCode = 0;\n    } catch (err) {\n      const error = err as Error & {\n        status?: number;\n        code?: number;\n        stdout?: string | Buffer;\n        stderr?: string | Buffer;\n      };\n      exitCode =\n        typeof error.status === \"number\"\n          ? error.status\n          : typeof error.code === \"number\"\n            ? error.code\n            : null;\n      errorMessage = error.message;\n      stdout = this.bufferToString(error.stdout);\n      stderr = this.bufferToString(error.stderr);\n    }\n\n    const passed = exitCode !== null && exitCode === expectedExitCode;\n\n    const details: Record<string, unknown> = { workDir, command };\n    if (stdout !== undefined && stdout !== \"\") {\n      details.stdout = stdout;\n    }\n    if (stderr !== undefined && stderr !== \"\") {\n      details.stderr = stderr;\n    }\n\n    return this.createResult(assertionId, {\n      status: passed\n        ? AssertionResultStatus.PASSED\n        : AssertionResultStatus.FAILED,\n      message: this.formatMessage(exitCode, expectedExitCode, errorMessage),\n      expected: String(expectedExitCode),\n      actual: exitCode !== null ? String(exitCode) : undefined,\n      details,\n    });\n  }\n\n  private createResult(\n    assertionId: string,\n    fields: Partial<AssertionResult>,\n  ): AssertionResult {\n    return {\n      id: randomUUID(),\n      assertionId,\n      assertionType: \"build_passed\",\n      assertionName: \"Build passed\",\n      status: AssertionResultStatus.FAILED,\n      ...fields,\n    };\n  }\n\n  private bufferToString(\n    value: string | Buffer | undefined,\n  ): string | undefined {\n    if (value === undefined || value === null) return undefined;\n    if (typeof value === \"string\") return value;\n    return value.toString(\"utf-8\");\n  }\n\n  private formatMessage(\n    exitCode: number | null,\n    expectedExitCode: number,\n    errorMessage: string | null,\n  ): string {\n    if (exitCode === null) {\n      return `Build failed: ${errorMessage}`;\n    }\n    if (exitCode === expectedExitCode) {\n      return `Build passed (exit code ${exitCode})`;\n    }\n    return `Build exited with ${exitCode}, expected ${expectedExitCode}`;\n  }\n}\n", "import type {\n  LlmJudgeAssertion,\n  AssertionResult,\n  LLMTrace,\n  EvaluationInput,\n} from \"../types/index.js\";\nimport { AssertionResultStatus } from \"../types/index.js\";\nimport { randomUUID } from \"crypto\";\nimport { createAnthropic } from \"@ai-sdk/anthropic\";\nimport { generateText } from \"ai\";\nimport type { AssertionContext, LlmConfig } from \"./assertion-evaluator.js\";\nimport { AssertionEvaluator } from \"./assertion-evaluator.js\";\n\nexport interface JudgeResult {\n  text: string;\n  score: number;\n  scoreReasoning: string;\n}\n\n/**\n * Format LLM trace as readable text for the judge (step number, type, tool name/args, output preview).\n */\nexport function formatTraceForJudge(llmTrace: LLMTrace | undefined): string {\n  if (!llmTrace?.steps?.length) {\n    return \"No trace available.\";\n  }\n  const lines: string[] = [];\n  for (const step of llmTrace.steps) {\n    const parts: string[] = [\n      `Step ${step.stepNumber}`,\n      `type: ${step.type}`,\n      `duration: ${step.durationMs}ms`,\n    ];\n    if (step.toolName) {\n      parts.push(`tool: ${step.toolName}`);\n      if (step.toolArguments) {\n        parts.push(`args: ${step.toolArguments}`);\n      }\n    }\n    if (step.outputPreview) {\n      parts.push(`output: ${step.outputPreview}`);\n    }\n    if (step.error) {\n      parts.push(`error: ${step.error}`);\n    }\n    lines.push(parts.join(\", \"));\n  }\n  return lines.join(\"\\n\");\n}\n\n/**\n * Context object for placeholder replacement.\n */\nexport interface PlaceholderContext {\n  output: string;\n  cwd: string;\n  changedFiles: string;\n  modifiedFiles: string;\n  newFiles: string;\n  trace: string;\n}\n\nexport function replacePlaceholders(\n  str: string,\n  ctx: PlaceholderContext,\n): string {\n  return str\n    .replace(/\\{\\{output\\}\\}/g, ctx.output)\n    .replace(/\\{\\{cwd\\}\\}/g, ctx.cwd)\n    .replace(/\\{\\{changedFiles\\}\\}/g, ctx.changedFiles)\n    .replace(/\\{\\{modifiedFiles\\}\\}/g, ctx.modifiedFiles)\n    .replace(/\\{\\{newFiles\\}\\}/g, ctx.newFiles)\n    .replace(/\\{\\{trace\\}\\}/g, ctx.trace);\n}\n\nexport function validateJudgeResult(parsed: unknown): JudgeResult {\n  if (parsed === null || typeof parsed !== \"object\") {\n    throw new Error(\"Judge result is not an object\");\n  }\n  const obj = parsed as Record<string, unknown>;\n  if (typeof obj.text !== \"string\") {\n    throw new Error(\"Judge result does not contain a valid text field\");\n  }\n  if (typeof obj.score !== \"number\") {\n    throw new Error(\"Judge result does not contain a valid score field\");\n  }\n  if (obj.score < 0 || obj.score > 100) {\n    throw new Error(\"Judge result score is not between 0 and 100\");\n  }\n  if (typeof obj.scoreReasoning !== \"string\") {\n    throw new Error(\n      \"Judge result does not contain a valid scoreReasoning field\",\n    );\n  }\n  return {\n    text: obj.text,\n    score: obj.score,\n    scoreReasoning: obj.scoreReasoning,\n  };\n}\n\nconst DEFAULT_MIN_SCORE = 70;\nconst DEFAULT_MODEL = \"claude-3-5-haiku-20241022\";\n\n/** Default system prompt context with placeholders; used when assertion.systemPrompt is empty. */\nconst DEFAULT_SYSTEM_PROMPT_CONTEXT = `You are judging a scenario run. The ACTUAL run data is provided below \u2014 use it to verify facts:\n\n- {{output}}: the agent's final output\n- {{cwd}}: working directory\n- {{changedFiles}}: list of all files changed (or \"No files were changed\")\n- {{modifiedFiles}}: list of existing files that were modified (or \"No files were modified\")\n- {{newFiles}}: list of new files that were created (or \"No new files were created\")\n- {{trace}}: step-by-step trace (tool calls, completions) so you can check e.g. which tools were called and how many times\n\nCRITICAL: When the user asks you to verify a specific fact, compare it strictly against the actual data above. If the expected outcome does NOT match the actual outcome, you MUST give a score of 0 or near 0. Do not be lenient \u2014 factual mismatches are failures.`;\n\nconst DEFAULT_SYSTEM_PROMPT = `You must respond only with a valid JSON object that conforms exactly to the following structure:\n\n{\n  \"text\": string,\n  \"score\": number (0-100),\n  \"scoreReasoning\": string\n}\n\n- text: A brief textual verdict of the test result.\n- score: A number from 0 to 100 that reflects how well the answer meets the acceptance criteria.\n- scoreReasoning: A concise explanation justifying the assigned score.\n\nYour response must:\n- Contain only the JSON object above \u2014 no introductory text, no code formatting (e.g., no triple backticks), and no trailing comments.\n- Be valid and parseable by \\`JSON.parse\\`.\n- Use only double quotes for all keys and strings, as required by JSON.\n\nAny response that includes extra content or deviates from the specified format will cause parsing to fail. Follow these instructions exactly.`;\n\n/**\n * Evaluator for \"llm_judge\" assertion: an LLM judges the scenario output\n * (prompt with {{output}}, {{cwd}}, {{changedFiles}}, {{trace}}) and returns a score 0-100.\n * Passes if score >= minScore.\n */\nexport class LlmJudgeEvaluator extends AssertionEvaluator<LlmJudgeAssertion> {\n  readonly type = \"llm_judge\" as const;\n\n  async evaluate(\n    assertion: LlmJudgeAssertion,\n    input: EvaluationInput,\n    context?: AssertionContext,\n  ): Promise<AssertionResult> {\n    const assertionId = randomUUID();\n    const llmConfig = context?.llmConfig;\n    const workDir = context?.workDir ?? \"\";\n    const generateTextStub = context?.generateTextForLlmJudge;\n\n    const output = input.outputText ?? \"\";\n    const fileDiffs = input.fileDiffs ?? [];\n\n    // Compute file lists by status\n    const changedPaths = fileDiffs.map((d) => d.path);\n    const modifiedPaths = fileDiffs\n      .filter((d) => d.status === \"modified\")\n      .map((d) => d.path);\n    const newPaths = fileDiffs\n      .filter((d) => d.status === \"new\")\n      .map((d) => d.path);\n\n    const changedFiles =\n      changedPaths.length > 0\n        ? changedPaths.map((path: string) => `- ${path}`).join(\"\\n\")\n        : \"No files were changed\";\n    const modifiedFiles =\n      modifiedPaths.length > 0\n        ? modifiedPaths.map((path: string) => `- ${path}`).join(\"\\n\")\n        : \"No files were modified\";\n    const newFiles =\n      newPaths.length > 0\n        ? newPaths.map((path: string) => `- ${path}`).join(\"\\n\")\n        : \"No new files were created\";\n\n    const trace = formatTraceForJudge(input.llmTrace);\n    const ctx: PlaceholderContext = {\n      output,\n      cwd: workDir,\n      changedFiles,\n      modifiedFiles,\n      newFiles,\n      trace,\n    };\n    const replace = (s: string) => replacePlaceholders(s, ctx);\n\n    const finalPrompt = replace(assertion.prompt);\n    const systemPrompt =\n      assertion.systemPrompt != null && assertion.systemPrompt !== \"\"\n        ? replace(assertion.systemPrompt) + \"\\n\\n\" + DEFAULT_SYSTEM_PROMPT\n        : replace(DEFAULT_SYSTEM_PROMPT_CONTEXT) +\n          \"\\n\\n\" +\n          DEFAULT_SYSTEM_PROMPT;\n\n    const minScore = assertion.minScore ?? DEFAULT_MIN_SCORE;\n    const maxOutputTokens = assertion.maxTokens ?? 1024;\n    const temperature = assertion.temperature ?? 0;\n\n    if (!generateTextStub && !llmConfig) {\n      return {\n        id: randomUUID(),\n        assertionId,\n        assertionType: \"llm_judge\",\n        assertionName: \"LLM judge\",\n        status: AssertionResultStatus.FAILED,\n        message: \"No llmConfig for llm_judge assertion (AI gateway required)\",\n        expected: String(minScore),\n      };\n    }\n\n    const maxParseAttempts = 3;\n    let lastParseError: Error | undefined;\n    let lastRawText: string | undefined;\n\n    try {\n      for (let attempt = 1; attempt <= maxParseAttempts; attempt++) {\n        const result = generateTextStub\n          ? await generateTextStub({\n              prompt: finalPrompt,\n              system: systemPrompt,\n              maxOutputTokens,\n              temperature,\n            })\n          : await this.callGenerateText(\n              llmConfig!,\n              assertion.model ?? DEFAULT_MODEL,\n              finalPrompt,\n              systemPrompt,\n              maxOutputTokens,\n              temperature,\n            );\n\n        lastRawText = result.text;\n        try {\n          const parsed = JSON.parse(result.text);\n          const judgeResult = validateJudgeResult(parsed);\n          const passed = judgeResult.score >= minScore;\n          return {\n            id: randomUUID(),\n            assertionId,\n            assertionType: \"llm_judge\",\n            assertionName: \"LLM judge\",\n            status: passed\n              ? AssertionResultStatus.PASSED\n              : AssertionResultStatus.FAILED,\n            message: passed\n              ? `Judge score ${judgeResult.score} >= ${minScore}: ${judgeResult.text}`\n              : `Judge score ${judgeResult.score} < ${minScore}: ${judgeResult.text}`,\n            expected: String(minScore),\n            actual: String(judgeResult.score),\n            details: {\n              score: judgeResult.score,\n              scoreReasoning: judgeResult.scoreReasoning,\n              text: judgeResult.text,\n            },\n          };\n        } catch (parseErr) {\n          lastParseError =\n            parseErr instanceof Error ? parseErr : new Error(String(parseErr));\n        }\n      }\n\n      return {\n        id: randomUUID(),\n        assertionId,\n        assertionType: \"llm_judge\",\n        assertionName: \"LLM judge\",\n        status: AssertionResultStatus.FAILED,\n        message: `Failed to parse judge response after ${maxParseAttempts} attempts: ${lastParseError?.message ?? \"unknown\"}`,\n        expected: String(minScore),\n        actual: undefined,\n        details: { rawText: lastRawText?.slice(0, 500) },\n      };\n    } catch (err) {\n      const message = err instanceof Error ? err.message : String(err);\n      return {\n        id: randomUUID(),\n        assertionId,\n        assertionType: \"llm_judge\",\n        assertionName: \"LLM judge\",\n        status: AssertionResultStatus.FAILED,\n        message: `LLM judge call failed: ${message}`,\n        expected: String(minScore),\n        details: { error: message },\n      };\n    }\n  }\n\n  private async callGenerateText(\n    llmConfig: LlmConfig,\n    modelId: string,\n    prompt: string,\n    system: string,\n    maxOutputTokens: number,\n    temperature: number,\n  ): Promise<{ text: string }> {\n    const anthropic = createAnthropic({\n      baseURL: llmConfig.baseUrl,\n      apiKey: \"dummy\",\n      headers: llmConfig.headers,\n    });\n    const result = await generateText({\n      model: anthropic(modelId),\n      prompt,\n      system,\n      maxOutputTokens,\n      temperature,\n    });\n    return { text: result.text };\n  }\n}\n"],
+  "mappings": ";AAAA,SAAS,SAAS;AAOX,IAAM,gCAAgC,EAAE,OAAO;AAAA,EACpD,MAAM,EAAE,QAAQ,kBAAkB;AAAA;AAAA,EAElC,WAAW,EAAE,OAAO;AACtB,CAAC;AAUM,IAAM,6BAA6B,EAAE,OAAO;AAAA,EACjD,MAAM,EAAE,QAAQ,cAAc;AAAA;AAAA,EAE9B,SAAS,EAAE,OAAO,EAAE,SAAS;AAAA;AAAA,EAE7B,kBAAkB,EAAE,OAAO,EAAE,IAAI,EAAE,SAAS;AAC9C,CAAC;AASM,IAAM,0BAA0B,EAAE,OAAO;AAAA,EAC9C,MAAM,EAAE,QAAQ,WAAW;AAAA;AAAA,EAE3B,QAAQ,EAAE,OAAO;AAAA;AAAA,EAEjB,cAAc,EAAE,OAAO,EAAE,SAAS;AAAA;AAAA,EAElC,UAAU,EAAE,OAAO,EAAE,IAAI,EAAE,IAAI,CAAC,EAAE,IAAI,GAAG,EAAE,SAAS;AAAA;AAAA,EAEpD,OAAO,EAAE,OAAO,EAAE,SAAS;AAAA,EAC3B,WAAW,EAAE,OAAO,EAAE,IAAI,EAAE,SAAS;AAAA,EACrC,aAAa,EAAE,OAAO,EAAE,IAAI,CAAC,EAAE,IAAI,CAAC,EAAE,SAAS;AACjD,CAAC;AASM,IAAM,kBAAkB,EAAE,MAAM;AAAA,EACrC;AAAA,EACA;AAAA,EACA;AACF,CAAC;;;AC7DD,SAAS,KAAAA,UAAS;AAKX,IAAM,mBAAmBA,GAAE,OAAO;AAAA,EACvC,QAAQA,GAAE,OAAO;AAAA,EACjB,YAAYA,GAAE,OAAO;AAAA,EACrB,OAAOA,GAAE,OAAO;AAClB,CAAC;AAOM,IAAK,cAAL,kBAAKC,iBAAL;AACL,EAAAA,aAAA,gBAAa;AACb,EAAAA,aAAA,cAAW;AACX,EAAAA,aAAA,iBAAc;AACd,EAAAA,aAAA,cAAW;AAJD,SAAAA;AAAA,GAAA;AAUL,IAAM,qBAAqBD,GAAE,OAAO;AAAA,EACzC,IAAIA,GAAE,OAAO;AAAA,EACb,YAAYA,GAAE,OAAO;AAAA,EACrB,MAAMA,GAAE,KAAK,WAAW;AAAA,EACxB,OAAOA,GAAE,OAAO;AAAA,EAChB,UAAUA,GAAE,OAAO;AAAA,EACnB,WAAWA,GAAE,OAAO;AAAA,EACpB,YAAYA,GAAE,OAAO;AAAA,EACrB,YAAY;AAAA,EACZ,SAASA,GAAE,OAAO;AAAA,EAClB,UAAUA,GAAE,OAAO,EAAE,SAAS;AAAA,EAC9B,eAAeA,GAAE,OAAO,EAAE,SAAS;AAAA,EACnC,cAAcA,GAAE,OAAO,EAAE,SAAS;AAAA,EAClC,eAAeA,GAAE,OAAO,EAAE,SAAS;AAAA,EACnC,SAASA,GAAE,QAAQ;AAAA,EACnB,OAAOA,GAAE,OAAO,EAAE,SAAS;AAC7B,CAAC;AAOM,IAAM,0BAA0BA,GAAE,OAAO;AAAA,EAC9C,OAAOA,GAAE,OAAO;AAAA,EAChB,YAAYA,GAAE,OAAO;AAAA,EACrB,QAAQA,GAAE,OAAO;AAAA,EACjB,SAASA,GAAE,OAAO;AACpB,CAAC;AAOM,IAAM,wBAAwBA,GAAE,OAAO;AAAA,EAC5C,YAAYA,GAAE,OAAO;AAAA,EACrB,iBAAiBA,GAAE,OAAO;AAAA,EAC1B,aAAa;AAAA,EACb,cAAcA,GAAE,OAAO;AAAA,EACvB,mBAAmBA,GAAE,OAAOA,GAAE,OAAO,GAAG,uBAAuB,EAAE,SAAS;AAAA,EAC1E,gBAAgBA,GAAE,OAAOA,GAAE,OAAO,GAAG,uBAAuB;AAAA,EAC5D,YAAYA,GAAE,MAAMA,GAAE,OAAO,CAAC;AAChC,CAAC;AAOM,IAAM,iBAAiBA,GAAE,OAAO;AAAA,EACrC,IAAIA,GAAE,OAAO;AAAA,EACb,OAAOA,GAAE,MAAM,kBAAkB;AAAA,EACjC,SAAS;AACX,CAAC;;;AChFD,SAAS,KAAAE,UAAS;AAMX,IAAK,wBAAL,kBAAKC,2BAAL;AACL,EAAAA,uBAAA,YAAS;AACT,EAAAA,uBAAA,YAAS;AACT,EAAAA,uBAAA,aAAU;AACV,EAAAA,uBAAA,WAAQ;AAJE,SAAAA;AAAA,GAAA;AAUL,IAAM,wBAAwBC,GAAE,OAAO;AAAA,EAC5C,IAAIA,GAAE,OAAO;AAAA,EACb,aAAaA,GAAE,OAAO;AAAA,EACtB,eAAeA,GAAE,OAAO;AAAA,EACxB,eAAeA,GAAE,OAAO;AAAA,EACxB,QAAQA,GAAE,KAAK,qBAAqB;AAAA,EACpC,SAASA,GAAE,OAAO,EAAE,SAAS;AAAA,EAC7B,UAAUA,GAAE,OAAO,EAAE,SAAS;AAAA,EAC9B,QAAQA,GAAE,OAAO,EAAE,SAAS;AAAA,EAC5B,UAAUA,GAAE,OAAO,EAAE,SAAS;AAAA,EAC9B,SAASA,GAAE,OAAOA,GAAE,OAAO,GAAGA,GAAE,QAAQ,CAAC,EAAE,SAAS;AAAA,EACpD,eAAeA,GAAE,MAAM,kBAAkB,EAAE,SAAS;AACtD,CAAC;;;AC1BD,SAAS,cAAAC,mBAAkB;;;ACK3B,SAAS,kBAAkB;;;ACyCpB,IAAe,qBAAf,MAAmE;AAQ1E;;;AD1CA,SAAS,6BACP,MACA,mBACS;AACT,MAAI,SAAS,QAAQ,OAAO,SAAS,UAAU;AAC7C,WAAO;AAAA,EACT;AACA,QAAM,MAAM;AACZ,QAAM,OAAO,IAAI;AACjB,SAAO,OAAO,SAAS,YAAY,SAAS;AAC9C;AAMO,IAAM,0BAAN,cAAsC,mBAA4C;AAAA,EAC9E,OAAO;AAAA,EAEhB,SACE,WACA,OAEA,UACiB;AACjB,UAAM,cAAc,WAAW;AAE/B,UAAM,WAAiC,MAAM;AAC7C,QAAI,CAAC,UAAU,OAAO,QAAQ;AAC5B,aAAO;AAAA,QACL,IAAI,WAAW;AAAA,QACf;AAAA,QACA,eAAe;AAAA,QACf,eAAe;AAAA,QACf;AAAA,QACA,SAAS;AAAA,QACT,UAAU,UAAU;AAAA,MACtB;AAAA,IACF;AAEA,eAAW,QAAQ,SAAS,OAAO;AACjC,UAAI,KAAK,aAAa,SAAS;AAC7B;AAAA,MACF;AACA,UAAI;AACJ,UAAI;AACF,eAAO,KAAK,gBACP,KAAK,MAAM,KAAK,aAAa,IAC9B;AAAA,MACN,QAAQ;AACN;AAAA,MACF;AACA,UAAI,6BAA6B,MAAM,UAAU,SAAS,GAAG;AAC3D,eAAO;AAAA,UACL,IAAI,WAAW;AAAA,UACf;AAAA,UACA,eAAe;AAAA,UACf,eAAe;AAAA,UACf;AAAA,UACA,SAAS,UAAU,UAAU,SAAS;AAAA,UACtC,UAAU,UAAU;AAAA,QACtB;AAAA,MACF;AAAA,IACF;AAEA,WAAO;AAAA,MACL,IAAI,WAAW;AAAA,MACf;AAAA,MACA,eAAe;AAAA,MACf,eAAe;AAAA,MACf;AAAA,MACA,SAAS,UAAU,UAAU,SAAS;AAAA,MACtC,UAAU,UAAU;AAAA,IACtB;AAAA,EACF;AACF;;;AEnFA,SAAS,cAAAC,mBAAkB;AAC3B,SAAS,gBAAgB;AAIzB,IAAM,kBAAkB;AACxB,IAAM,oBAAoB;AAMnB,IAAM,uBAAN,cAAmC,mBAAyC;AAAA,EACxE,OAAO;AAAA,EAEhB,SACE,WACA,QACA,SACiB;AACjB,UAAM,cAAcC,YAAW;AAC/B,UAAM,UAAU,SAAS;AACzB,UAAM,UAAU,UAAU,WAAW;AACrC,UAAM,mBAAmB,UAAU,oBAAoB;AAEvD,QAAI,CAAC,SAAS;AACZ,aAAO,KAAK,aAAa,aAAa;AAAA,QACpC;AAAA,QACA,SAAS;AAAA,QACT,UAAU,OAAO,gBAAgB;AAAA,MACnC,CAAC;AAAA,IACH;AAEA,QAAI,WAA0B;AAC9B,QAAI,eAA8B;AAClC,QAAI;AACJ,QAAI;AAEJ,YAAQ,IAAI,2BAA2B,OAAO,SAAS,OAAO,EAAE;AAEhE,QAAI;AACF,eAAS,SAAS;AAAA,QAChB,KAAK;AAAA,QACL,UAAU;AAAA,QACV,OAAO,CAAC,UAAU,QAAQ,MAAM;AAAA,MAClC,CAAC;AACD,iBAAW;AAAA,IACb,SAAS,KAAK;AACZ,YAAM,QAAQ;AAMd,iBACE,OAAO,MAAM,WAAW,WACpB,MAAM,SACN,OAAO,MAAM,SAAS,WACpB,MAAM,OACN;AACR,qBAAe,MAAM;AACrB,eAAS,KAAK,eAAe,MAAM,MAAM;AACzC,eAAS,KAAK,eAAe,MAAM,MAAM;AAAA,IAC3C;AAEA,UAAM,SAAS,aAAa,QAAQ,aAAa;AAEjD,UAAM,UAAmC,EAAE,SAAS,QAAQ;AAC5D,QAAI,WAAW,UAAa,WAAW,IAAI;AACzC,cAAQ,SAAS;AAAA,IACnB;AACA,QAAI,WAAW,UAAa,WAAW,IAAI;AACzC,cAAQ,SAAS;AAAA,IACnB;AAEA,WAAO,KAAK,aAAa,aAAa;AAAA,MACpC,QAAQ;AAAA,MAGR,SAAS,KAAK,cAAc,UAAU,kBAAkB,YAAY;AAAA,MACpE,UAAU,OAAO,gBAAgB;AAAA,MACjC,QAAQ,aAAa,OAAO,OAAO,QAAQ,IAAI;AAAA,MAC/C;AAAA,IACF,CAAC;AAAA,EACH;AAAA,EAEQ,aACN,aACA,QACiB;AACjB,WAAO;AAAA,MACL,IAAIA,YAAW;AAAA,MACf;AAAA,MACA,eAAe;AAAA,MACf,eAAe;AAAA,MACf;AAAA,MACA,GAAG;AAAA,IACL;AAAA,EACF;AAAA,EAEQ,eACN,OACoB;AACpB,QAAI,UAAU,UAAa,UAAU,KAAM,QAAO;AAClD,QAAI,OAAO,UAAU,SAAU,QAAO;AACtC,WAAO,MAAM,SAAS,OAAO;AAAA,EAC/B;AAAA,EAEQ,cACN,UACA,kBACA,cACQ;AACR,QAAI,aAAa,MAAM;AACrB,aAAO,iBAAiB,YAAY;AAAA,IACtC;AACA,QAAI,aAAa,kBAAkB;AACjC,aAAO,2BAA2B,QAAQ;AAAA,IAC5C;AACA,WAAO,qBAAqB,QAAQ,cAAc,gBAAgB;AAAA,EACpE;AACF;;;ACxHA,SAAS,cAAAC,mBAAkB;AAC3B,SAAS,uBAAuB;AAChC,SAAS,oBAAoB;AAatB,SAAS,oBAAoB,UAAwC;AAC1E,MAAI,CAAC,UAAU,OAAO,QAAQ;AAC5B,WAAO;AAAA,EACT;AACA,QAAM,QAAkB,CAAC;AACzB,aAAW,QAAQ,SAAS,OAAO;AACjC,UAAM,QAAkB;AAAA,MACtB,QAAQ,KAAK,UAAU;AAAA,MACvB,SAAS,KAAK,IAAI;AAAA,MAClB,aAAa,KAAK,UAAU;AAAA,IAC9B;AACA,QAAI,KAAK,UAAU;AACjB,YAAM,KAAK,SAAS,KAAK,QAAQ,EAAE;AACnC,UAAI,KAAK,eAAe;AACtB,cAAM,KAAK,SAAS,KAAK,aAAa,EAAE;AAAA,MAC1C;AAAA,IACF;AACA,QAAI,KAAK,eAAe;AACtB,YAAM,KAAK,WAAW,KAAK,aAAa,EAAE;AAAA,IAC5C;AACA,QAAI,KAAK,OAAO;AACd,YAAM,KAAK,UAAU,KAAK,KAAK,EAAE;AAAA,IACnC;AACA,UAAM,KAAK,MAAM,KAAK,IAAI,CAAC;AAAA,EAC7B;AACA,SAAO,MAAM,KAAK,IAAI;AACxB;AAcO,SAAS,oBACd,KACA,KACQ;AACR,SAAO,IACJ,QAAQ,mBAAmB,IAAI,MAAM,EACrC,QAAQ,gBAAgB,IAAI,GAAG,EAC/B,QAAQ,yBAAyB,IAAI,YAAY,EACjD,QAAQ,0BAA0B,IAAI,aAAa,EACnD,QAAQ,qBAAqB,IAAI,QAAQ,EACzC,QAAQ,kBAAkB,IAAI,KAAK;AACxC;AAEO,SAAS,oBAAoB,QAA8B;AAChE,MAAI,WAAW,QAAQ,OAAO,WAAW,UAAU;AACjD,UAAM,IAAI,MAAM,+BAA+B;AAAA,EACjD;AACA,QAAM,MAAM;AACZ,MAAI,OAAO,IAAI,SAAS,UAAU;AAChC,UAAM,IAAI,MAAM,kDAAkD;AAAA,EACpE;AACA,MAAI,OAAO,IAAI,UAAU,UAAU;AACjC,UAAM,IAAI,MAAM,mDAAmD;AAAA,EACrE;AACA,MAAI,IAAI,QAAQ,KAAK,IAAI,QAAQ,KAAK;AACpC,UAAM,IAAI,MAAM,6CAA6C;AAAA,EAC/D;AACA,MAAI,OAAO,IAAI,mBAAmB,UAAU;AAC1C,UAAM,IAAI;AAAA,MACR;AAAA,IACF;AAAA,EACF;AACA,SAAO;AAAA,IACL,MAAM,IAAI;AAAA,IACV,OAAO,IAAI;AAAA,IACX,gBAAgB,IAAI;AAAA,EACtB;AACF;AAEA,IAAM,oBAAoB;AAC1B,IAAM,gBAAgB;AAGtB,IAAM,gCAAgC;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAWtC,IAAM,wBAAwB;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAwBvB,IAAM,oBAAN,cAAgC,mBAAsC;AAAA,EAClE,OAAO;AAAA,EAEhB,MAAM,SACJ,WACA,OACA,SAC0B;AAC1B,UAAM,cAAcC,YAAW;AAC/B,UAAM,YAAY,SAAS;AAC3B,UAAM,UAAU,SAAS,WAAW;AACpC,UAAM,mBAAmB,SAAS;AAElC,UAAM,SAAS,MAAM,cAAc;AACnC,UAAM,YAAY,MAAM,aAAa,CAAC;AAGtC,UAAM,eAAe,UAAU,IAAI,CAAC,MAAM,EAAE,IAAI;AAChD,UAAM,gBAAgB,UACnB,OAAO,CAAC,MAAM,EAAE,WAAW,UAAU,EACrC,IAAI,CAAC,MAAM,EAAE,IAAI;AACpB,UAAM,WAAW,UACd,OAAO,CAAC,MAAM,EAAE,WAAW,KAAK,EAChC,IAAI,CAAC,MAAM,EAAE,IAAI;AAEpB,UAAM,eACJ,aAAa,SAAS,IAClB,aAAa,IAAI,CAAC,SAAiB,KAAK,IAAI,EAAE,EAAE,KAAK,IAAI,IACzD;AACN,UAAM,gBACJ,cAAc,SAAS,IACnB,cAAc,IAAI,CAAC,SAAiB,KAAK,IAAI,EAAE,EAAE,KAAK,IAAI,IAC1D;AACN,UAAM,WACJ,SAAS,SAAS,IACd,SAAS,IAAI,CAAC,SAAiB,KAAK,IAAI,EAAE,EAAE,KAAK,IAAI,IACrD;AAEN,UAAM,QAAQ,oBAAoB,MAAM,QAAQ;AAChD,UAAM,MAA0B;AAAA,MAC9B;AAAA,MACA,KAAK;AAAA,MACL;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF;AACA,UAAM,UAAU,CAAC,MAAc,oBAAoB,GAAG,GAAG;AAEzD,UAAM,cAAc,QAAQ,UAAU,MAAM;AAC5C,UAAM,eACJ,UAAU,gBAAgB,QAAQ,UAAU,iBAAiB,KACzD,QAAQ,UAAU,YAAY,IAAI,SAAS,wBAC3C,QAAQ,6BAA6B,IACrC,SACA;AAEN,UAAM,WAAW,UAAU,YAAY;AACvC,UAAM,kBAAkB,UAAU,aAAa;AAC/C,UAAM,cAAc,UAAU,eAAe;AAE7C,QAAI,CAAC,oBAAoB,CAAC,WAAW;AACnC,aAAO;AAAA,QACL,IAAIA,YAAW;AAAA,QACf;AAAA,QACA,eAAe;AAAA,QACf,eAAe;AAAA,QACf;AAAA,QACA,SAAS;AAAA,QACT,UAAU,OAAO,QAAQ;AAAA,MAC3B;AAAA,IACF;AAEA,UAAM,mBAAmB;AACzB,QAAI;AACJ,QAAI;AAEJ,QAAI;AACF,eAAS,UAAU,GAAG,WAAW,kBAAkB,WAAW;AAC5D,cAAM,SAAS,mBACX,MAAM,iBAAiB;AAAA,UACrB,QAAQ;AAAA,UACR,QAAQ;AAAA,UACR;AAAA,UACA;AAAA,QACF,CAAC,IACD,MAAM,KAAK;AAAA,UACT;AAAA,UACA,UAAU,SAAS;AAAA,UACnB;AAAA,UACA;AAAA,UACA;AAAA,UACA;AAAA,QACF;AAEJ,sBAAc,OAAO;AACrB,YAAI;AACF,gBAAM,SAAS,KAAK,MAAM,OAAO,IAAI;AACrC,gBAAM,cAAc,oBAAoB,MAAM;AAC9C,gBAAM,SAAS,YAAY,SAAS;AACpC,iBAAO;AAAA,YACL,IAAIA,YAAW;AAAA,YACf;AAAA,YACA,eAAe;AAAA,YACf,eAAe;AAAA,YACf,QAAQ;AAAA,YAGR,SAAS,SACL,eAAe,YAAY,KAAK,OAAO,QAAQ,KAAK,YAAY,IAAI,KACpE,eAAe,YAAY,KAAK,MAAM,QAAQ,KAAK,YAAY,IAAI;AAAA,YACvE,UAAU,OAAO,QAAQ;AAAA,YACzB,QAAQ,OAAO,YAAY,KAAK;AAAA,YAChC,SAAS;AAAA,cACP,OAAO,YAAY;AAAA,cACnB,gBAAgB,YAAY;AAAA,cAC5B,MAAM,YAAY;AAAA,YACpB;AAAA,UACF;AAAA,QACF,SAAS,UAAU;AACjB,2BACE,oBAAoB,QAAQ,WAAW,IAAI,MAAM,OAAO,QAAQ,CAAC;AAAA,QACrE;AAAA,MACF;AAEA,aAAO;AAAA,QACL,IAAIA,YAAW;AAAA,QACf;AAAA,QACA,eAAe;AAAA,QACf,eAAe;AAAA,QACf;AAAA,QACA,SAAS,wCAAwC,gBAAgB,cAAc,gBAAgB,WAAW,SAAS;AAAA,QACnH,UAAU,OAAO,QAAQ;AAAA,QACzB,QAAQ;AAAA,QACR,SAAS,EAAE,SAAS,aAAa,MAAM,GAAG,GAAG,EAAE;AAAA,MACjD;AAAA,IACF,SAAS,KAAK;AACZ,YAAM,UAAU,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAC/D,aAAO;AAAA,QACL,IAAIA,YAAW;AAAA,QACf;AAAA,QACA,eAAe;AAAA,QACf,eAAe;AAAA,QACf;AAAA,QACA,SAAS,0BAA0B,OAAO;AAAA,QAC1C,UAAU,OAAO,QAAQ;AAAA,QACzB,SAAS,EAAE,OAAO,QAAQ;AAAA,MAC5B;AAAA,IACF;AAAA,EACF;AAAA,EAEA,MAAc,iBACZ,WACA,SACA,QACA,QACA,iBACA,aAC2B;AAC3B,UAAM,YAAY,gBAAgB;AAAA,MAChC,SAAS,UAAU;AAAA,MACnB,QAAQ;AAAA,MACR,SAAS,UAAU;AAAA,IACrB,CAAC;AACD,UAAM,SAAS,MAAM,aAAa;AAAA,MAChC,OAAO,UAAU,OAAO;AAAA,MACxB;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF,CAAC;AACD,WAAO,EAAE,MAAM,OAAO,KAAK;AAAA,EAC7B;AACF;;;AJ/SA,IAAM,oBAAoB,IAAI,kBAAkB;AAEhD,IAAM,aAAiD;AAAA,EACrD,kBAAkB,IAAI,wBAAwB;AAAA,EAC9C,cAAc,IAAI,qBAAqB;AAAA,EACvC,WAAW;AAAA;AAAA,EAEX,QAAQ;AACV;AAQO,SAAS,kBACd,MACA,WACM;AACN,aAAW,IAAI,IAAI;AACrB;AAQO,SAAS,aAAa,MAA8C;AACzE,SAAO,WAAW,IAAI;AACxB;AAUA,eAAsB,mBACpB,OACA,YACA,SAC4B;AAC5B,MAAI,WAAW,WAAW,GAAG;AAC3B,WAAO,CAAC;AAAA,EACV;AACA,SAAO,QAAQ;AAAA,IACb,WAAW,IAAI,OAAO,cAAc;AAClC,YAAM,YAAY,WAAW,UAAU,IAAI;AAC3C,UAAI,CAAC,WAAW;AACd,eAAO;AAAA,UACL,IAAIC,YAAW;AAAA,UACf,aAAaA,YAAW;AAAA,UACxB,eAAe,UAAU;AAAA,UACzB,eAAe;AAAA,UACf;AAAA,UACA,SAAS,+BAA+B,UAAU,IAAI;AAAA,UACtD,UAAU;AAAA,QACZ;AAAA,MACF;AACA,YAAM,UAAU,KAAK,IAAI;AACzB,YAAM,SAAS,MAAM,UAAU,SAAS,WAAW,OAAO,OAAO;AACjE,YAAM,aAAa,KAAK,IAAI,IAAI;AAChC,aAAO,EAAE,GAAG,QAAQ,UAAU,WAAW;AAAA,IAC3C,CAAC;AAAA,EACH;AACF;",
   "names": ["z", "LLMStepType", "z", "AssertionResultStatus", "z", "randomUUID", "randomUUID", "randomUUID", "randomUUID", "randomUUID", "randomUUID"]
 }

package/build/types/types/assertions.d.ts

@@ -21,7 +21,7 @@ export declare const BuildPassedAssertionSchema: z.ZodObject<{
 export type BuildPassedAssertion = z.infer<typeof BuildPassedAssertionSchema>;
 /**
  * Assertion: an LLM judges the scenario output (score 0-100).
- * Prompt can use {{output}}, {{cwd}}, {{changedFiles}}, {{trace}}.
+ * Prompt can use {{output}}, {{cwd}}, {{changedFiles}}, {{modifiedFiles}}, {{newFiles}}, {{trace}}.
  * Passes if judge score >= minScore.
  */
 export declare const LlmJudgeAssertionSchema: z.ZodObject<{

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@wix/eval-assertions",
-  "version": "0.4.0",
+  "version": "0.6.0",
   "description": "Assertion framework for AI agent evaluations - supports skill invocation checks, build validation, and LLM-based judging",
   "files": [
     "build"
@@ -60,5 +60,5 @@
   ],
   "license": "MIT",
   "author": "Wix",
-  "falconPackageHash": "693f93d9ed7e967f76f5d677ac8f80e91b7fbdfee7056f0de610e387"
+  "falconPackageHash": "e5ece20df4121359d4281425bfa2170554de6cca38afe2f5c4c90a2f"
 }