Function runPlantedBugEval

test/skill-e2e.test.ts:707–826

* Shared planted-bug eval runner. * Gives the agent concise bug-finding instructions (not the full QA workflow), * then scores the report with an LLM outcome judge.

(fixture: string, groundTruthFile: string, label: string)

   * then scores the report with an LLM outcome judge.
   */
  async function runPlantedBugEval(fixture: string, groundTruthFile: string, label: string) {
    // Each test gets its own isolated working directory to prevent cross-contamination
    // (agents reading previous tests' reports and hallucinating those bugs)
    const testWorkDir = fs.mkdtempSync(path.join(os.tmpdir(), `skill-e2e-${label}-`));
    setupBrowseShims(testWorkDir);
    const reportDir = path.join(testWorkDir, 'reports');
    fs.mkdirSync(path.join(reportDir, 'screenshots'), { recursive: true });
    const reportPath = path.join(reportDir, 'qa-report.md');

    // Direct bug-finding with browse. Keep prompt concise — no reading long SKILL.md docs.
    // "Write early, update later" pattern ensures report exists even if agent hits max turns.
    const targetUrl = `${testServer.url}/${fixture}`;
    const result = await runSkillTest({
      prompt: `Find bugs on this page: ${targetUrl}

Browser binary: B="${browseBin}"

PHASE 1 — Quick scan (5 commands max):
$B goto ${targetUrl}
$B console --errors
$B snapshot -i
$B snapshot -c
$B accessibility

PHASE 2 — Write initial report to ${reportPath}:
Write every bug you found so far. Format each as:
- Category: functional / visual / accessibility / console
- Severity: high / medium / low
- Evidence: what you observed

PHASE 3 — Interactive testing (targeted — max 15 commands):
- Test email: type "user@" (no domain) and blur — does it validate?
- Test quantity: clear the field entirely — check the total display
- Test credit card: type a 25-character string — check for overflow
- Submit the form with zip code empty — does it require zip?
- Submit a valid form and run $B console --errors
- After finding more bugs, UPDATE ${reportPath} with new findings

PHASE 4 — Finalize report:
- UPDATE ${reportPath} with ALL bugs found across all phases
- Include console errors, form validation issues, visual overflow, missing attributes

CRITICAL RULES:
- ONLY test the page at ${targetUrl} — do not navigate to other sites
- Write the report file in PHASE 2 before doing interactive testing
- The report MUST exist at ${reportPath} when you finish`,
      workingDirectory: testWorkDir,
      maxTurns: 50,
      timeout: 300_000,
      testName: `qa-${label}`,
      runId,
    });

    logCost(`/qa ${label}`, result);

    // Phase 1: browse mechanics. Accept error_max_turns — agent may have written
    // a partial report before running out of turns. What matters is detection rate.
    if (result.browseErrors.length > 0) {