失败测试分流 Skill：只读日志，不改生产密钥

---
name: failing-tests
description: Triage failing tests from runner logs only. Do not change production secrets or skip tests to go green.
---

# Failing-test triage

Follow this order. Read logs first. Do not touch production credentials.

这不是 Roo Mode YAML。

## Read

- Identify the failing node ids / test names from the runner output already in the conversation or in CI logs in the repo. Use whatever this repo already runs (`pytest`, `vitest`, `go test`, `cargo test`).
- Open the failing test file and the nearest production module it asserts against.
- Quote a short, redacted log excerpt. Strip tokens, cookies, private keys, and `.env` values before quoting.

## Classify

- Assertion mismatch in product code.
- Fixture / testdata drift.
- Missing environment *name* (the variable is unset). That is not a reason to paste a live value.
- Flake (timing, order, network). Say so; do not hide it with a skip.

## Do not

- Do not open `.env`, `.env.*`, or secret stores to "just make CI pass."
- Do not rotate, rewrite, or inject production keys, PATs, or session cookies.
- Do not add `|| true`, unconditional skips, or skipped tests unless the request explicitly asks to skip and the skip message names the ticket.
- Do not add a second test runner.
- Do not patch the Cursor workbench, unlock Free Cursor, or inject tokens into the IDE.

## Fix path

- Prefer a failing assertion that matches real behavior, or a fixture that uses fake data.
- If a credential is required, document the environment variable *name* only and keep the value out of the repo.
- Re-run the nearest existing command for the touched package.

## Output

Write: failing names, redacted log excerpt, classification, and a read-only next step. If a secret was needed, say which *name* — never the value.