Bound win-deskutils sessionDispatch child wait to 10s

[mikhailm-coder]

Why

A qa Windows device's mesh agent wedges seconds after every start: the core's idle-time poll (idle.getSecondsAllSessions(), present in the server-pushed core but not in the old frozen disk core) calls sessionDispatch() per active session, which spawns an agent.exe -b64exec helper and blocks in unbounded child.waitExit(). On that VM the helper hangs forever → the agent's single-threaded event machinery blocks → heartbeats starve → server closes the control channel (~18 min) → the agent's own reconnect dials fail at TLS while the box's network is fine. Killing the hung child just advances to the next spawn-and-wedge. Any Windows machine with a hang-prone session can hit this.

The identical fix (waitExit(10000)) has existed in the meshagent repo's modules/win-deskutils.js (and its binary-embedded blob) since April — but the server-pushed core's copy overrides the binary's at runtime, so the unbounded version is what executes post-migration.

What

One line in agents/modules_meshcore/win-deskutils.js: child.waitExit() → child.waitExit(10000). On timeout the exitCode != 0 path throws and the per-session try/catch in idle_getSecondsAllSessions skips that session — the poll resolves and the agent keeps running. Precedent for bounded waits: sysinfo.js uses waitExit(2000)/waitExit(5000).

Rollout

• Merge → latest image → dev/OSS auto-roll (keel); agents pick up the fixed core on reconnect via the standard push.
• qa/stage/prod pin immutable tags → needs a 0.0.28 release dispatch + values bumps (qa first — the affected device is there).
• The wedged qa VM may need its hung helper killed (or a reboot) once the qa server carries the fix, so the core push can be processed; it self-heals from there.

🤖 Generated with Claude Code