fix(core): graceful process tree shutdown for continuous tasks

[agcty]

Summary

This builds on top of Leo's nxc-3480 branch (PR nrwl#33655) to fix the graceful shutdown regression that was blocking the PR from landing.

Problem: PR nrwl#33655 added killProcessTree using sysinfo to kill all descendants on Ctrl+C, but it was fire-and-forget — processes were killed so fast that cleanup handlers (e.g. server.close()) never had a chance to run.

Solution: Added killProcessTreeGraceful() — an async Rust function that:

1. Snapshots the entire process tree via BFS before sending any signals
2. Sends SIGTERM to all descendants (leaves first)
3. Polls every 100ms, waiting up to 5s (configurable) for processes to exit
4. Force-kills survivors with SIGKILL after the grace period
5. Detects reparented descendants (when root exits quickly, children reparented to init)

All TypeScript kill() methods across the codebase now use the graceful variant, and signal handlers properly await the kill promises before calling process.exit().

Key changes

• process_killer/mod.rs — Added killProcessTreeGraceful, collect_descendants_of, and helper functions
• running-tasks.ts — Per-child SIGINT handlers no longer call process.exit() (prevents first-child-exits race); parent-level handler awaits all kills
• forked-process-task-runner.ts — cleanup() is now async, SIGINT handler awaits it
• task-orchestrator.ts — Awaits forked runner cleanup alongside other running tasks
• pseudo-terminal.ts — kill() uses graceful tree kill directly; shutdown() uses sync killProcessTree for exit handler
• node-child-process.ts — Both kill methods use killProcessTreeGraceful
• run-script.impl.ts — Exit handler uses graceful kill with .finally()

Test results

Tested against the reproduction repo (agcty/nx-32438-continuous-tasks-repro) with 3 concurrent services (service-a:3001, service-b:3002, frontend):

• ✅ No orphaned processes after Ctrl+C / SIGTERM
• ✅ All ports freed (servers closed gracefully)
• ✅ Cleanup handlers confirmed running (log files written by onCleanup callbacks)
• ✅ Shutdown completes in ~1s (well within 5s grace period)

Added 6 integration tests in kill_process_tree.spec.ts covering tree killing, graceful SIGTERM with cleanup verification, slow cleanup within grace period, and already-dead process fast path.

Review by OpenAI Codex (gpt-5.3-codex, xhigh reasoning)

The fix was reviewed by Codex CLI which found and helped address:

1. Unawaited async kills in forked-process-task-runner → Fixed
2. Per-child SIGINT listener race condition → Fixed
3. PTY root-exits-quickly before tree snapshot → Fixed
4. TaskOrchestrator not awaiting forked cleanup → Fixed

Note: The nx.json change (removing @nx/gradle) was needed to build locally without Java and should be dropped before merging upstream.

Related

• Fixes Node processes not closing after ending NX session nrwl/nx#32438
• Builds on fix(core): use native graceful process tree shutdown nrwl/nx#33655

Test plan

• Automated: npx jest packages/nx/src/native/tests/kill_process_tree.spec.ts (6/6 pass)
• Manual: Multi-service reproduction with cleanup handler verification
• nx affected -t build,test,lint on nrwl/nx monorepo
• e2e tests

🤖 Generated with Claude Code