mirror of
https://github.com/NeuralNomadsAI/CodeNomad.git
synced 2026-11-17 13:59:29 +00:00
## Problem A process keeps the numeric ID of a parent that exited, and Windows reuses that ID (Raymond Chen, [*Why is CreateToolhelp32Snapshot returning incorrect parent process IDs*](https://devblogs.microsoft.com/oldnewthing/20150403-00/?p=44313)). A tree walk over parent IDs alone adopts unrelated processes. An application an agent launched through a launcher that exited names the launcher's PID; once a process being stopped receives that PID, the application and its children are terminated with it. Observed: a Houdini session died with exit code 1 (`TerminateProcess(…, 1)`), its `cmd` parent 0.2 s earlier, two seconds after two `powershell.exe` started. - `processTreeFromRows` (`electron/main/process-stop.ts`) walked parent IDs only. The native and CIM Windows captures already filter children older than their parent, but the shared walk relied on that, and the Linux `/proc` scan and macOS `ps` listing are not atomic. - `taskkill /T` walks recorded parent IDs the same way (see e.g. openai/codex#36258, #58999). It was used by the server's foreign OpenCode `--version` probe on timeout, by `ui/src/lib/server-events-lifecycle.test.ts` and by two Windows desktop fixtures. ## Change - `packages/server/src/process-tree.ts` (Node built-ins only, shared by Electron main, the server, tests and tsx-loaded scripts): - `processTreeRows(rootPid, rows)`: a row is a child only when it started **no earlier** than the process holding its parent ID; unknown start times are left out; the first row of a PID wins. - `stopWindowsProcessTree(rootPid, spawnedBy)`: one CIM listing, the same walk, refusal when the root ID was reused after the spawn, then `Stop-Process` through `Process` objects whose creation time is checked after their handle is bound. - Electron's `processTreeFromRows` uses `processTreeRows` with Windows creation ticks, Linux boot clock ticks and macOS `lstart` seconds (same-second children are kept). - The foreign probe, the UI test and both desktop fixtures use `stopWindowsProcessTree` instead of `taskkill /T`. Checked and unchanged: Tauri (Windows Job Object with kill-on-close, POSIX process groups: membership inherited at creation, no walk), Electron POSIX `kill(-pid)` in `shell-environment.ts` (process group), Electron force termination (already checks each member's creation identity before `TerminateProcess`). ## Tests - `process-stop.test.ts`: a recycled-parent table on Windows, Linux and macOS leaves the orphan and its child out of the capture. **Fails on `dev`** (`[100, 300, 301, 200, 201]` instead of `[100, 200, 201]`), passes here; the 34 other tests pass. - `server/src/process-tree.test.ts`: recycled parent, a cycle through a recycled ID, unknown start times, duplicate rows, a missing root, reused root ID refusal, CIM row parsing, a real Windows wrapper and its descendant stopped. - `legacy-installation.test.ts` (foreign probe kills the wrapper's descendant on timeout) and `server-events-lifecycle.test.ts` pass; server, Electron and UI `tsc` pass. ## Docs `AGENTS.md` (never `taskkill /T` or a parent-ID-only walk), the desktop conventions reference and `OPENCODE_RUNTIME_SETUP.md`. No user-visible behaviour, so no help change. |
||
|---|---|---|
| .. | ||
| agent | ||
| commands | ||
| skills | ||