v0.10.0-insider.1: Tool execution hangs after preToolUse hook completes (no tool.execution_complete) in multi-MCP setup #1353

Summary

In a fresh squad init'd folder running Squad CLI v0.10.0-insider.1 on GitHub Copilot CLI v1.0.63 (Windows), the first tool the assistant tries to invoke (powershell reading .squad/team.md) hangs indefinitely after the preToolUse hook completes successfully. The actual tool process is never spawned, and no tool.execution_complete event is ever emitted.

Reproduced twice in the same session (after Ctrl+C abort, retry hit the identical hang on the very next tool call).

Environment

Reproduction

1. cd C:\temp\perception-squad
2. copilot to start a fresh session
3. Wait through MCP startup (all 6 user-level MCPs + squad_state emit "taking longer than expected to connect" warnings — squad_state starts last, ~45s after session start)
4. Prompt the assistant with anything that should trigger a tool call. In our case: "hi" → assistant ran if (Test-Path .squad/team.md) { Write-Host "===TEAM_EXISTS==="; Get-Content .squad/team.md }

Observed behavior

From events.jsonl:

12:31:27  assistant.turn_start
12:31:27  assistant.message    (Squad banner)
12:31:27  tool.execution_start (toolName="powershell", reading team.md)
12:31:27  hook.start  (preToolUse)
12:31:32  hook.end    (preToolUse, success=true)
... silence for ~10 minutes ...
12:41:10  abort       (reason="user_initiated"  ← Ctrl+C)

No tool.execution_complete event ever fired. The session was retried after abort and hit the identical hang on the very next tool call (same pattern: tool.execution_start → hook.end → silence).

Event tally over 28 minutes of session lifetime:

Process-level diagnostic

The hung copilot.exe (PID 63572) was alive (1465s CPU, 188 MB RSS) but no powershell.exe / pwsh.exe child process was ever spawned. All 8 MCP server child processes were alive and idle. The hang is in the CLI's tool dispatch path between hook.end and child_process.spawn(powershell).

Hypothesis

Race condition between Squad CLI's state-mcp initialization (which connects ~45s into the session) and Copilot CLI's tool dispatcher. When many MCPs are still settling, the CLI's post-hook dispatch path may be blocked on an unfulfilled IPC promise.

Possibly a Copilot CLI bug, but the Squad state-mcp is the late-connecting MCP that pushes the session over the edge.

Workaround tested

Reducing the global mcp-config.json from 7 MCPs to 3 (kept: enghub, teams, mail; dropped: aspire, azure-devops, playwright, nano-banana) — pending verification by user.

Asks

1. Could state-mcp startup be made non-blocking / lazy, only initializing on first tool call?
2. Is there a known interaction between preToolUse hooks and parallel MCP initialization in 0.10.0-insider.1?
3. Worth a --no-state-mcp flag for repro / debugging?

Happy to attach the full events.jsonl (~285 KB) if useful — it's redacted-safe (just protocol events, no message content). Just say the word.