Distillation

Date: Sep 16, 2026

• 🟡 (15:42) Candidate tests verify private provider-failure details are redacted from public output and accounting completions, including assertions that JSON.stringify(completions) does not contain secret, "private-failure-message", "private principal query", or private terminal-model data; diagnostics included { detail: secret }.
• 🟡 (15:42) Candidate Responses-stream tests assert exactly one terminal event: response.failed frame via expect(out.match(/^event: response\.failed$/gm)).toHaveLength(1) and require public output to contain PUBLIC_GATEWAY_ERROR.
• 🟡 (15:42) Candidate failure-path tests track recall attempts with let recalls = 0 and assert expect(recalls).toBe(0), including follow-up-stream scenarios using runFollowUp.
• 🟡 (15:42) Candidate tests cover atomic handling of multiple output items using textItem(1, "second", "msg_atomic_second") and fixtures including response IDs "resp_atomic_reference" and "resp_visible_in_progress_failure" with model "gpt-5.6-terra".
• 🟡 (15:42) Candidate test added: test("redacts a sparse terminal-discovered recall before status validation", async () => {, covering redaction before terminal status validation.
• 🟡 (15:42) Candidate test instrumentation captures exceptions with captureException: (error) => capturedLogs.push(String(error)), drains streams using const out = await drain(client), and records completion success through successful: boolean.
• 🟡 (15:42) Candidate test code imports GatewayResponse from ../src/translate/types.