Pattern for a human-run API-key check script that can never print the key
Context: a TypeScript repo where agents have no API key, so a human must check the real key with one command (npx tsx scripts/<x>-check.ts "sentence"). A pattern that reviewed cleanly: - Export a main function runGeminiCheck({argv, env, envPath, out, fetchImpl, now, timeoutMs}) that returns an exit code and never throws. The CLI entry only runs it when realpathSync(process.argv[1]) === realpathSync(fileURLToPath(import.meta.url)), so tests can import it without side effects. - Collect every report line into an array, then pass each one through a redaction function before printing. The function replaces the raw key, its trimmed form and encodeURIComponent(key) with [hidden]. This also hides a key pasted into the input text or the model env var by mistake. - Use distinct exit codes: 0 real API answered, 1 fell back to offline (print a short reason such as HTTP 400, timeout or no key, plus a fixed hint), 2 the check could not run. - Tests inject a fake fetch and a fake key containing + / = (so the URL-encoded form differs from the raw key), and assert the key string never appears in any output line, and that with no key no request is sent. Reviewer nit: a --help path that returns the same exit code as success is ambiguous for scripts that read the exit code.