Skip to content

Measured Windows child cleanup comparison (35f613545)

Pre-release
Pre-release

Choose a tag to compare

@Krarilotus Krarilotus released this 07 Oct 14:39
· 310 commits to fix/windows-child-cleanup since this release

Matched Windows child-process cleanup evidence

This is a synthetic source-level comparison, not a running installed T3 app or whole-machine measurement.
Versions and complete per-run numbers are in measured-comparison.json. No private app state or ETL is included.

Reproduce on Windows 11 with Node 24 and Python 3:

  1. In an empty parent folder, clone pingdotgg/t3code into t3code-windows-cleanup-sol and check out 35f6135.
  2. Create sibling t3-cleanup-evidence-sol and put the supplied scripts there.
  3. In that sibling: npm init -y; npm install --ignore-scripts --no-audit --no-fund @effect/[email protected] [email protected] @effect/[email protected] [email protected] [email protected] [email protected] @types/[email protected]
  4. Run python prepare.py. It copies the exact published package into independent baseline/fixed directories and reapplies the already-committed patch without changing its contents.
  5. Run python measure.py baseline behavior, then python measure.py fixed behavior. Only owned fixture descendants are killed; other processes are untouched.
  6. For four matched windows, sequentially run python measure.py baseline measure matched-A1; python measure.py fixed measure matched-B1; python measure.py fixed measure matched-B2; python measure.py baseline measure matched-A2.
  7. Run python report.py to summarize/render the matched results. It expects the committed fix as HEAD and its baseline as HEAD^.

The spawner is supplied to the unchanged T3 ProcessRunner from source. Imports and two warmup probes are excluded. Each 22-second window schedules the same 20 one-second ticks with two Git probes. Direct spawn calls are counted through temporary node:child_process hooks. Python opens/retains native handles immediately after each spawn, reads final GetProcessTimes and closes the handles. Runner CPU is a process.cpuUsage delta. All measured children closed before the end; no CPU handles were missed.

CPU excludes WMI, Defender, conhost, secondary subprocesses and unrelated workloads. Wall-clock command duration sums are sensitive to scheduling. The two matched variants had identical Git volume; this experiment makes no claim about application polling cadence or total PC CPU. The orphan case uses detached/ignored-stdio descendants to isolate leader-exit behavior.

Existing T3 tests were run under an external minimal Vitest configuration to avoid installing/building the full app. processRunner.test.ts and VcsProcess.test.ts: 43 passed. Selected existing dependency tests (buildSpawnOptions, unread stdout, nonzero exit, invalid command/cwd): 6 passed, 72 unselected. No new tests are committed. Strict source typechecking with repo-pinned @types/node 24.12.4 reports the same pre-existing ExecException/ExecFileException callback mismatch on original and patched source. Frozen lockfile validation and git apply --check passed.