ShiZai · Shenzhen · Open source
AI runtime reliability, Kubernetes scheduling, and the awkward boundaries where sessions, processes, queues, and real cluster state stop agreeing.
专注 Agent 运行时可靠性、Kubernetes 调度与可复现的系统边界问题。
Selected patches
Each entry links to the upstream discussion, code, tests, and maintainer review. No star counts, no invented scale—just the change and what it proved.
A fair-sharing preemption path reduced a LocalQueue to its short name, so identically named queues in different namespaces could be mixed up.
Made topology-aware node replacement distinguish a real domain boundary from a string that merely shares the same prefix.
Global-session fallback could route concurrent session diagnostics to the wrong transcript. The fix makes the active session context authoritative.
Validated remote fetch targets before marketplace refresh and plugin installation could reach protected network locations.
Working method
Static suspicion is a lead, not a finding. The useful work starts when the behavior can be made observable and repeatable.
Start from the event, API, controller, or process boundary a user can actually reach.
Use a narrow state sequence or A/B reproduction so timing and coincidence cannot explain the result.
Land the smallest behavior change with a regression that fails for the original reason.
Field notes
Short write-ups about identity, lifecycle, scheduling, and the tests that separate a plausible bug from a proven one.
How a missing namespace crossed a fair-sharing boundary, and how a small regression test made the bug undeniable.
Read field note →A process-global fallback looks harmless until two agent sessions run at the same time.
Read field note →Why cancellation, bind, and retry windows deserve executable proofs rather than source-level suspicion.
Read field note →Open channel
Open an issue, send a minimal reproduction, or point me at the state transition nobody trusts.