Reachability before severity
A dramatic code path is irrelevant if production cannot reach it. Start with the caller and the state transition.
About / 关于
Reliability is what remains when the demo is over.
I’m ShiZai, also known as Tomsen. I work on AI agent runtimes and Kubernetes scheduling—systems where a tiny identity, lifecycle, or ordering mistake can survive every happy-path test and still break real work.
我专注于 Agent 运行时可靠性、Kubernetes 调度，以及那些必须构造真实状态序列才能证明的边界问题。 我更在意问题是否可达、能否稳定复现，以及测试究竟保护了什么，而不是把代码看起来“像有问题”当成结论。
BASEShenzhen, China
FOCUSRuntime · Scheduler · State
MODEOpen source, evidence first
House rules
How I decide a bug is real.
A patch starts after the uncertainty has been reduced enough that another person can inspect the same evidence.
One changed variable
A controlled A/B or exact state sequence makes the causal claim reviewable instead of rhetorical.
Tests are explanations
The best regression names the behavior, creates the old failure, and stays quiet for the right reason after the fix.
Upstream is the record
Claims about impact link to maintainer discussion, merged code, and backports—not a private summary.
Current terrain
Kubernetes controllersKueue / TASGoAgent run lifecycleTypeScriptPythonProcess cleanupSession isolationFailure injection
Show me the state transition.
The fastest way to start a useful conversation is a minimal reproduction, an event sequence, or a link to the code path.
GitHub ↗