mirror of
https://github.com/alibaba/open-code-review.git
synced 2026-10-02 09:24:52 +08:00
Hao Guo
14b84f08a3
Closing MCP clients in the review teardown was serial and unbounded: each Close hit the SDK's three-stage escalation (wait after stdin close, wait after SIGTERM, wait after SIGKILL) at the default 5s per stage, so one unresponsive stdio server could hold shutdown for ~15s and N configured servers multiplied that. The overruns defeated docker stop's 10s grace period and the VS Code extension's 3s SIGKILL escalation, orphaning the subprocesses the graceful path exists to reclaim (#1141). - Set an explicit 800ms TerminateDuration on CommandTransport, keeping a single server's worst case at ~2.4s instead of ~15s. - Add mcp.CloseAll: closes clients concurrently with per-server errors joined into one report, so the phase costs the slowest server rather than the sum of all servers. - Cap the whole close phase with a 2.8s deadline in review_cmd, chosen on context.Background() so an already-cancelled run keeps its budget; on timeout the warning reports the servers still shutting down and keeps any errors already collected. The SDK still escalates to SIGKILL in the meantime; whatever outlives the process is left to the OS. Tests cover an unresponsive stdio server (SIGTERM-ignoring child, full stdin-close -> SIGTERM -> SIGKILL path, < 3s), unchanged fast error-free close for responsive servers, joined per-server error messages, and the deadline abandonment. Child fixtures run this test binary in three modes; under -race the widened child shutdown latency is absorbed by a build-tag budget so the suite stays deterministic.