OpenAI Dots 全自动修 bug 流程拆解

6AWORON0k @6AW0RON0k原推发布于 2026.10.03

在 X 上打开原推

A bug gets filed. A tested pull request shows up. Nobody woke up. 5 steps. 0 humans. The blueprint behind it: 328,922 views. 505 bookmarks. There is one step the agent refuses to take. The night, as the blueprint draws it: A new bug lands in GitHub or Linear. The owner is asleep. An OpenAI Dots agent spots it. STEP 1. The task goes to a cloud coding sandbox. Repo and dependencies are already loaded. STEP 2. The failure gets recreated inside a sealed container. A tiny unit test gets written. It has to fail first. STEP 3. One narrow fix, aimed at that failure only. STEP 4. The whole test suite runs. Red turns green. STEP 5. A pull request gets assembled. 5 of 5. So who is doing all this? Four roles. - DOTS watches the trackers and hands off the job - SANDBOX reproduces, tests and patches - AUDITOR wakes on every PR or a weekly scan - MAINTAINER is the human with the merge button The fourth one is the guard. Merges by the agent: 0. Patch without a maintainer's approval: blocked. Then the ladder goes quiet. The agent sits at the top. It looks stuck. It's waiting. Merge is the step it won't take. The PR in its hands carries 3 things. The diff. The test evidence. A list of what it still isn't sure about. The agent writes down its own doubts and asks a person to check. 6 steps from detection to review. A human touches exactly 1. That's 17% of the chain. The claimed payoff is about 20 minutes per bug. That's an estimate from the write-up. Nobody timed it. At 5 bugs a week, 100 minutes come back. Over a 50-week year, about 83 hours. More than two 40-hour work weeks of pulling branches and rebuilding environments. Bookmark this one. Follow for the next teardown.

这条是引用着另一条写的：x.com/mojunel/status/2106042827679408577