Copilot 1.0.20 Does Not Close Child Pwsh.exe Earlier Then On Exit

[znamenap]

Describe the bug

I've let run the github speckit over the night and in the mornig have found 197 child processes made by copilot.exe.
I'm assuminig it is a bug while handling the results.
I'm assuming also the amount of the private bytes signifiies for a bug also.

This is the command line of the pwsh.exe
pwsh.exe -NoProfile -NoLogo -NoExit -Command "try { Set-PSReadLineOption -PredictionSource None -BellStyle None -HistorySaveStyle SaveNothing -MaximumHistoryCount 0 } catch { Set-PSReadLineOption -HistorySaveStyle SaveNothing }"

I did not observe this with 1.0.17

Affected version

GitHub Copilot CLI 1.0.20.

Steps to reproduce the behavior

I'm not certain if this is reproducible, I'll keep en eye on it during the day if it happen again.

Expected behavior

Copilot should end the process just after it finished collected results and ensure the child process is gone.

Additional context

Windows 11 / Windows Terminal Preview / pwsh.exe 7.5.5

[examon]

This should be resolved as of v1.0.21, which automatically shuts down shell sessions that are no longer needed, preventing child pwsh.exe processes from accumulating over time. Updating from v1.0.20 to v1.0.21 should fix this. If the problem persists after updating, feel free to reopen this issue.

[znamenap]

In the copilot's output have found following:

● Read (Speckit.implement agent — Implement feature 010)
└ Running

◐ The agent seems stuck - it's been at 52 tool calls for a while with no progress. It asked a question about incomplete checklists, the user (me) answered "yes" earlier based on the context summary, but it seems the agent might be stuck in its implementation phase. Let me check
if it's still making progress by waiting a bit more, or if I should send it a nudge.

Actually, looking at the output more carefully, the agent has completed turns 0 and 1, and turn 1's response was asking about proceeding despite incomplete checklists. It seems like it may have gotten a response already (since in the context summary it says "asked for
confirmation twice (approved design? proceed despite incomplete checklist?)" and I approved both). But it's been at 52 tool calls for quite a while now.

I should probably send a message to check what's happening or give it a push forward.

● The implementation agent appears stuck at 52 tool calls for several minutes. Let me nudge it.

● Write Speckit.implement agent (Implement feature 010)
└ Yes, proceed with implementation despite incomplete checklist. Begin with Pha…

● Read (Speckit.implement agent — Implement feature 010)
└ Running

◐ The agent seems truly stuck - it hasn't progressed beyond 52 tool calls even after I sent a message. It's been at this state for over 22 minutes now with no new turns. The intent hasn't changed from "Reviewing implementation plan" and no new tool calls have been made.

Let me check the list of all agents to see their statuses. Then I should probably take over implementation myself, since the agent seems to be in an infinite loop or stuck state.

which perhaps means it tried to nudge the child agent multiple times and did not clean the tools

● Listing background agents

◐ The implement-010 agent is stuck. I also notice there's a "contacts-rubberduck" rubber-duck agent running for 391 seconds - the implement agent may have launched it as part of its review. But the implement agent itself is stuck at 52 tool calls.

Given that the agent appears genuinely stuck (no progress in tool calls over several minutes despite receiving my message), I should take over implementation myself. Let me:

1. Record the current state/timing
2. Read the tasks.md to understand what needs to be done
3. Start implementing the feature myself following the tasks.md phase order

But first, let me try one more time to see if the agent makes progress in the next minute. Actually, no - it's been stuck for too long. Let me just take over.

Let me check what the rubber-duck agent is currently working on.

● The implementation agent appears stuck. There's also a rubber-duck it launched that's still running. Let me check its status, then take over implementation myself.

● Read (Rubber-duck agent — Validate implementation approach)
└ Running

◐ Both agents are stuck. Let me take over implementation. Let me:

1. Read the tasks.md to know the implementation order
2. Read the current state of the codebase (key files I'll need to modify)
3. Start implementing phase by phase

First, let me read the key planning documents to remind myself of what needs to be done.

● Both agents are stuck. I'll take over implementation directly. Let me first review the tasks and key design docs.

which actually started to cost me Claude Opus High premium requests on implementation that I did not want to, the implementation agent was configured to use gpt-5.3-codex.

then after some implementation steps made by the parent agent, I found the following.

◐ OK so there are actually two different implement agents:

1. implement-010 - still running (stuck at 52 tool calls)
2. contacts-implementer - just finished, asking a clarification question

It seems like in the previous turn before the context was summarized, another implementation agent was launched. Let me respond to this one and tell it to proceed with auto-commit, and also continue with my own implementation work. Actually, let me just focus on one path - let
me respond to this agent and have it do the implementation.

Actually wait, I was about to implement things myself. But now this new agent is asking about committing. Let me respond to it to proceed, and let it handle the implementation. That's more efficient.

I should probably abandon the stuck implement-010 agent and just work with contacts-implementer instead.

● Two implementation agents exist — the original implement-010 is stuck, but contacts-implementer is alive and asking a question. Let me work with it.

● Write Speckit.implement agent (Execute contacts tasks)
└ Yes, perform the auto-commit in T059 with the Co-authored-by trailer. Proceed…

● Compacting conversation history...

● Now let me also start implementing Phase 1 myself in parallel since the agent may take a long time. I'll create the three value objects:

I'm not sure about the relevancy of the above content to investigate this issue.
Then copilot followed the expected flow of implementing the prompt.