fix(shell): bound post-exit pipe draining on all platforms

[Hona]

Issue for this PR

A command can exit while a background child still has its output pipes open. We keep waiting for EOF, so the shell tool never finishes.

Related issue reports

This fixes the V2 implementation of the post-exit hangs reported below.

• Fixes bash tool hangs when command spawns background child processes #20902
• Fixes Long-running shell commands (e.g. Gradle build) hang even after “BUILD SUCCESSFUL” #25038
• Fixes Agent hangs indefinitely after "BUILD SUCCESSFUL" when running assembleDebug #28697
• Fixes Bash tool hangs until timeout on Windows when spawned process keeps a stdout/stderr pipe open (dev servers, daemons) #32504
• Fixes Stucked when calling the gradle build android #36342
• Fixes [Bug] opencode hangs during playwright cli execution and other commands #36384
• Fixes After updating to version 1.17.18, it often gets stuck when using the command "Start-Process -NoNewWindow". #36799
• Fixes Bash tool hangs indefinitely when PowerShell uses Start-Process with -RedirectStandardOutput/-RedirectStandardError on Windows #37838
• Fixes [Bug] Shell timeout does not settle a running tool part, leaving parent subagent blocked indefinitely #38276
• Fixes [Bug] Shell timeout kill chain deadlocks: handle.kill() never resolves, tool part stays running forever #38291
• Fixes [Bug] Gradle build success but session keep going forever #42044
• Fixes Bash tool hangs indefinitely when executing commands that launch daemon processes with continuous output #42524
• Fixes Tool call suceeds but OC waits indefinitely #43315
• Fixes shell: detached Windows process leaves parent session stuck running #43910

Earlier reports, already closed: #19252, #22012, #22154, #24731, #24784, #24926, #25938, #29294, #29822, #33363.

Previous fix attempts: #20901, #28336, #29831, #42756, #43511, #44601.

Type of change

• Bug fix

What does this PR do?

Give stdout and stderr one shared second to drain after the parent exits. Then finish capture and let the shell return instead of waiting on the background child.

The important part is keeping that separate from process cleanup. Shell timeouts still escalate to SIGKILL, and closing an MCP server still cleans up its descendants. No public API changes.

The tradeoff is that we can lose trailing output. Later bytes are discarded, and writers can get EPIPE/SIGPIPE once their readers close.

How did you verify your code works?

Added four real-process regression cases: held pipes after exit, SIGKILL completion, MCP descendant cleanup, and shell timeout escalation. Each test also fails when its corresponding fix is disabled.

Validation and remaining limits

• Typechecks pass. Linux MCP, spawner, and available shell-tool tests pass.
• Windows checks pass apart from the existing echo quoting failure; Unix-only cases are skipped.
• Native macOS and the reported browser/.NET workloads have not been tested directly.
• Parallel scope cleanup can still fall back to the old EOF wait. No affected production caller was established.
• Normal Windows MCP shutdown now makes an extra taskkill attempt, taking roughly 0.35-0.73 seconds in our checks.

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

[christianhola]

1

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[gterminator]

the problem still exists on 1.18.27