Codex Can Loop on Tests; Try a Two-Attempt Stop Rule
A Juejin author reports that Codex repeatedly modified code and reran failing tests, consuming substantial token usage without completing the task. The article recommends pausing after two unsuccessful attempts on the same issue, preserving evidence, avoiding redundant checks and scope expansion, and recording goals, test results, and remaining work for longer sessions.
The article provides a practical guardrail for coding-agent workflows, emphasizing that repeated automated activity should yield to evidence-based checkpoints when failures continue without new information.
最近用Codex，老金就一直在等。
测试没过，它改一下，再测，还是没过，那就接着改。
我看着它忙，心想这一轮总该交活了吧，结果它自己又续上了。
真是屡试不爽，屡试的是它，不爽的是我。
等太久了，感觉很不对劲儿，我就让它统计了一下Token用量，看看都花哪儿了。
#Codex CLI#AGENTS.md#Claude#Gemini CLI#OpenCode#Anthropic#OpenAI#Google