Fix ToolTask hang when grandchild processes inherit pipe handles

[YuliiaKovalova]

Context

ToolTask can hang indefinitely when the tool it spawns creates grandchild processes that inherit stdout/stderr pipe handles. This is a long-standing issue reported in #2981 (opened 2018), with a previous fix attempt in #10297 that was reverted (#10395) because it caused output loss (#10378).

Root Cause

On .NET Framework (and .NET Core), the parameterless Process.WaitForExit() internally calls AsyncStreamReader.WaitUtilEOF(), which blocks until all write handles to the stdout/stderr pipes are closed. When a tool like cl.exe spawns a grandchild process (e.g., mspdbsrv.exe) that inherits the pipe handles, EOF is never reached even though the tool itself has exited - causing an infinite hang.

Why the previous fix (#10297) failed

PR #10297 changed proc.WaitForExit() to proc.WaitForExit(int.MaxValue). The int overload skips WaitUtilEOF(), preventing the hang. But it also skips waiting for all DataReceived callbacks to be delivered. For fast tools like command -v ls, the AsyncStreamReader hadn't delivered its output before the drain calls ran, causing ConsoleOutput to be empty.

Changes Made

The fix (behind ChangeWave 18.6) uses the Data==null EOF sentinel that AsyncStreamReader sends via DataReceived when each pipe reaches EOF:

Step 1: proc.WaitForExit(int.MaxValue)

Waits for the process handle only, not pipe EOF. Since _toolExited already fired before WaitForProcessExit is called, the process is dead and this returns immediately.

Step 2: WaitHandle.WaitAll(eofEvents, 2000)

Waits for our own _standardOutputEOF / _standardErrorEOF events, which are set by ReceiveStandardErrorOrOutputData when Data==null arrives from the AsyncStreamReader.

• Normal case (no grandchild): Pipe closes immediately after tool exits, EOF arrives within milliseconds, events fire, all data including final partial line is guaranteed delivered. This provides identical guarantees to the original proc.WaitForExit().
• Grandchild case: Grandchild holds pipe open, EOF never arrives, events time out after 2 seconds, proceed. The tool's line-by-line output was already delivered during the HandleToolNotifications loop.

Why this doesn't lose data (unlike #10297)

Our _standardOutputEOF fires inside AsyncStreamReader.FlushMessageQueue(), which is called before the internal eofEvent.Set(). By the time our event fires, the AsyncStreamReader has already:

1. Read all remaining bytes from the pipe buffer
2. Decoded them into characters
3. Flushed the final partial line from its StringBuilder
4. Delivered every line (including the final one) via DataReceived callbacks

This is functionally equivalent to WaitUtilEOF - just observed from our callback instead of the internal event.

Testing

ToolTaskDoesNotHangWhenGrandchildInheritsPipeHandles

Spawns cmd.exe /c echo hello & start /b ping -n 120 127.0.0.1 > nul:

• cmd.exe writes "hello" and exits immediately
• ping inherits pipe handles and runs for 120 seconds
• With fix: Test completes in ~2 seconds, "hello" is captured
• Without fix (MSBUILDDISABLEFEATURESFROMVERSION=18.6): Test hangs until 30s timeout

ToolTaskCapturesAllOutputWithFix

Spawns cmd.exe /c echo line1 & echo line2 & echo line3 (no grandchild):

• Verifies all three lines are captured - regression test for [Bug]: Recent ToolTask change causes Exec task to miss output #10378

[Copilot]

Pull request overview

Addresses a long-standing ToolTask hang caused by Process.WaitForExit() waiting for stdout/stderr pipe EOF when grandchild processes inherit the redirected pipe handles, by switching to a bounded EOF-drain wait observed via DataReceived EOF sentinels (behind ChangeWave 18.6).

Changes:

• Add stdout/stderr EOF ManualResetEvents and use them to wait (bounded) for async stream-drain completion after waiting for the process handle.
• Update the DataReceived handler to signal EOF events when Data == null.
• Add regression tests for the hang scenario and for preserving tool output; introduce ChangeWave 18.6 and document it.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 4 comments.

File	Description
src/Utilities/ToolTask.cs	Implements Wave18.6 guarded hang fix by decoupling process-exit wait from pipe-EOF wait, using EOF callbacks with a bounded timeout.
src/Utilities.UnitTests/ToolTask_Tests.cs	Adds regression tests for the grandchild-inherited-pipe hang and for output capture behavior.
src/Framework/ChangeWaves.cs	Introduces Wave18_6 and adds it to AllWaves.
documentation/wiki/ChangeWaves.md	Documents the new 18.6 change wave and its associated feature.

[adamsitnik]

LGTM @YuliiaKovalova ! I left some comments, but they are all just nits.

[baronfel]

This looks targeted, tested, and protected by a change wave. Nice work!