Windows: taskkill /PID is mangled by MSYS path conversion when SHELL is set (Git Bash under Claude Code) — cancel and SessionEnd cleanup never kill the process tree

[danscMax]

Summary

On Windows, terminateProcessTree() invokes taskkill /PID <pid> /T /F through a shell because runCommand defaults to shell: process.env.SHELL || true on win32 (plugins/codex/scripts/lib/process.mjs). When SHELL points to Git Bash — which is exactly the environment Claude Code provides on Windows (SHELL=/bin/bash.exe) — MSYS argument conversion rewrites the /PID switch into a filesystem path, so taskkill fails and the process tree is never killed.

Reproduction

Windows 11, plugin 1.0.6, Node 24.16.0, SHELL=/bin/bash.exe (default under Claude Code's bash tool):

const { spawnSync } = require("child_process");
const r = spawnSync("taskkill", ["/PID", "999999", "/T", "/F"],
  { shell: process.env.SHELL || true, encoding: "utf8", windowsHide: true });
console.log(r.status, r.stderr || r.stdout);

Output (status 1, cp866 mojibake from the Russian locale):

Ошибка: неправильный параметр или аргумент - 'C:/Program Files/Git/PID'.

MSYS has translated /PID → C:/Program Files/Git/PID. With shell: false the same call works correctly.

Impact

• /codex:cancel marks a background job cancelled, but the detached task-worker tree survives (the soft turn/interrupt path still works when a turnId is already known, but the hard-kill never lands).
• The SessionEnd hook's job cleanup silently fails — cleanupSessionJobs swallows the error, leaking worker processes on every session end.
• SpawnedCodexAppServerClient.close() on Windows relies on the same terminateProcessTree, so the shell-wrapped codex app-server grandchild can be leaked too.
• Independently of MSYS: looksLikeMissingProcessMessage() matches English-only taskkill output (not found|no running instance|...). On non-English Windows locales taskkill prints localized (OEM-codepage) text, so the "process already gone" detection never matches and dead-process cases surface as thrown errors.

This also shows up when running the repo's own test suite on Windows: the cancel/session-end/status tests fail with the mangled-/PID assertion above (CI currently runs ubuntu-latest only, so it goes unnoticed).

Suggested fix

1. Run taskkill with shell: false — its arguments are fixed strings, so the shell adds nothing. This mirrors what Remove shell expansion for git commands #447 already did for git ("repository-derived arguments must never pass through a shell"); the same reasoning applies to the win32 default shell: process.env.SHELL || true, which makes behavior depend on whatever SHELL happens to be.
2. Prefer exit-code-based detection over locale-dependent stderr parsing: taskkill exits with code 128 when the process does not exist.