OpenAI 调试 Agent 更新：从崩溃点回溯到错误决策

wast3 @0xWast3原推发布于 2026.10.06

在 X 上打开原推

OpenAI shipped an agent update that stops reporting what crashed and starts naming the decision that made the crash possible. A stack trace points at the line that failed. That line is almost never where the problem was introduced. > WALK - the failure traced backwards through every call that led into it > BLAME - the commit that introduced each assumption on that path identified > ECHO - other places in the codebase running on the same assumption flagged > RANK - the paths sorted by how much live traffic crosses them Debugging usually ends at the first thing that looks wrong. You patch it, the error disappears, and the cause stays in the repo. That is why the same bug comes back wearing a different message three weeks later. The echo step is the uncomfortable one. One bad assumption rarely lives in a single file. It returns a list of places that have not broken yet, which is a different thing from a list of places that are fine. A fix without that list is a guess that happened to work. This turns an incident into a map of everything standing on the same floor. The trace format and the ranking rules are laid out in the breakdown below.

这条是引用着另一条写的：x.com/0xwast3/status/2106328851705987398