fix: kill entire CLI process tree on stop/forceStop (Windows)

[rinceyuan]

Summary

Fix process tree leak on Windows when stop()/forceStop() terminates the CLI.

On Windows, ChildProcess.kill() (Node.js) and Popen.terminate()/kill() (Python) only terminate the immediate process via TerminateProcess(), leaving grandchildren orphaned. Each create_session()/stop() cycle leaks a full copilot process tree (~3-5 processes) that survives until idle timeout or OOM.

Changes

SDK	Fix
Node.js (client.ts)	New killProcessTree() helper: uses taskkill /T /F /PID on Windows, falls back to standard .kill(signal) on Unix
Python (client.py)	New _kill_process_tree() helper: uses taskkill /T on Windows, os.killpg() on Unix, with fallback to proc.kill()

Validation

• TypeScript compiles clean (tsc --noEmit)
• Python syntax validated (ast.parse)
• Manual Windows test: spawned a 2-level node process tree, confirmed taskkill /T kills both parent and grandchild (see test output below)

parent=47632 child=27936 pre:parent alive pre:child alive taskkill OK post:parent DEAD(OK) post:child DEAD(OK)

Notes

• Go, .NET, Rust, and Java SDKs may have similar issues but use different process management; those are left for follow-up.
• The Unix path uses os.killpg() in Python (requires the child to be a process group leader); falls back gracefully if not applicable.

Closes #1804

[rinceyuan]

@microsoft-github-policy-service agree company=Microsoft

[rinceyuan]

@stephentoub This fixes the Windows process tree leak reported in #1804. Affects both Node.js and Python SDKs — each stop()/forceStop() cycle was orphaning the CLI's child processes. The fix uses \ askkill /T\ on Windows. Manually verified on Windows 11. Happy to add the Go/.NET fixes in a follow-up if desired.

[SteveSandersonMS]

The way this is implemented in the PR currently won’t work because:

• Python calls killpg() without starting the CLI in a separate process group, which can kill the host and sibling processes.
• Node’s process-group behavior is opt-in and defaults to false, so normal POSIX clients retain the leak.
• Python does not check whether taskkill succeeded.
• The change covers only Node and Python, despite the same lifecycle requirement applying across SDKs.
• There are no real process-tree tests, and the existing Node lifecycle test fails.

There is a small, coherent cross-language change we could accept: add one private “terminate owned runtime tree” operation per SDK, called from the existing owned-process termination point.

Its behavior should be:

Windows: taskkill /T /F /PID <root>
POSIX:   signal the runtime’s private process group

POSIX also needs one small spawn-time change to place the runtime in its own process group/session. Otherwise group termination could kill the host. No public API is needed.

Per language, this is approximately:

SDK	Spawn-time isolation	Teardown
Node	detached: true	process.kill(-pid, signal)
Python	start_new_session=True	os.killpg(pid, signal)
Go	SysProcAttr.Setpgid = true	syscall.Kill(-pid, signal)
Rust	process_group(0)	signal negative PID
Java	Use ProcessHandle.descendants()	snapshot descendants, kill them, then root
.NET	None needed	existing Kill(entireProcessTree: true)

Each SDK should call that helper from the process-termination section used by stop() and forceStop(). External-server and in-proc paths must not call it.

This is about the smallest useful implementation across all languages and OSes:

• one private helper per SDK;
• one POSIX spawn flag per applicable SDK;
• no public options;
• no Job Objects;
• no crash-cleanup guarantee;
• no redesign of graceful shutdown.

The tests can also be narrow: start a helper process that starts one long-lived child, then verify both disappear after stop() and forceStop() on Windows and POSIX. Also verify external and in-proc modes do not enter tree termination.