Regression testing pra agents: evitar regredir por mudança

# tests/agent_regression/case_001_multi_tool.yaml
name: "User asks for weather + books — uses 2 tools"
input: "Qual é o tempo em SP agora e me indique 3 livros de sci-fi"

expected:
  tools_used:
    - { text: "get_weather", args_contains: { city: "São Paulo" } }
    - { text: "search_books", args_contains: { genre: "sci-fi", count: 3 } }
  output_must_contain: ["temperatura", "sci-fi"]
  output_judge_rubric: |
    Response should:
    1. State current weather in São Paulo
    2. Recommend 3 sci-fi books with titles
    3. Be in Portuguese
  max_steps: 5
  max_cost_usd: 0.05

import { runAgent } from './agent';
import { loadCases, judge } from './eval-lib';

describe('Agent regression', () => {
  const cases = loadCases('tests/agent_regression/*.yaml');

  it.each(cases)('$name', async (testCase) => {
    const result = await runAgent(testCase.input, {
      maxSteps: testCase.expected.max_steps,
      toolReplay: 'record',  // replay cached tool calls
    });

    // Check tools
    for (const expected of testCase.expected.tools_used) {
      const match = result.trace.find(s => s.tool === expected.name);
      expect(match).toBeDefined();
      // validate args subset
    }

    // Check output via LLM judge
    const verdict = await judge(result.output, testCase.expected.output_judge_rubric);
    expect(verdict.pass).toBe(true);

    // Cost cap
    expect(result.totalCost).toBeLessThanOrEqual(testCase.expected.max_cost_usd);
  });
});