Repository navigation
fix(cli-agent-bridge): recover verified children of zombie parents - #9
Conversation
A parent can retain its original /proc identity as a zombie after its child has been reparented. Recover this remaining diagnostic-tracker race only after stable run-marker evidence and a final immutable PID identity check. Keep upstream's early child binding, sticky uncertainty, and Windows-only production support boundary. Add six event-driven regressions covering recovery, wrong or absent markers, PID reuse, child exit, and uncertainty. Do not change integration budgets or workflow configuration.
|
Independent follow-up verification — 2026-10-15 Rechecked unchanged head
The recovery still requires stable exact run-marker evidence plus a final same-generation identity check. It does not replace a recorded identity, clear prior uncertainty, or expand the Windows-only production containment support boundary. No live provider invocation was used. Marked ready for review. The target intentionally remains |
Scope
Own-fork maintenance review against upstream main
6481e4ae1e0375a4ddb3d5e9b29845354ab1cead. The base branch is an unchanged snapshot of that commit. This isolates the useful remainder offix/bridge-stability-20260911after upstream MiniMax-AI#43 and MiniMax-AI#50; it does not reopen the old test-timing changes or port larger Windows smoke budgets.Remaining bug
On Linux, an exited parent can remain in
/procas a zombie with its original start identity after the child has already been reparented. The current early-child-binding logic cannot establish the old PPID relationship in this case. Later queue processing permanently marks the child uncertain and omits it from the snapshot even when its run marker independently proves ownership.Recover that child only after two consecutive stable run-marker scans and a final same-generation PID identity check. Preserve early child binding, immutable recorded identities, sticky prior uncertainty, and rejection of wrong/absent markers. This improves the retained POSIX diagnostic tracker; it does not enable POSIX production delegation or change the Windows Job Object support boundary.
Validation
git diff --checkpassede8974b8cf7598661556873cd69122cce599e42ed; the CI merge commit and head have the identical tree6dc99bec177a4e114d6cf80388348aa530064568Only
process-tree.mjsandtests/process-tree.test.mjsare changed. No workflow edits, deadline-budget changes, new dependencies, releases, or deployment.