Windows / npm Copilot CLI 1.0.105: intermittent multi-minute stalls mid-turn; cancellation and continue required #5356

Description

Using the npm CLI via npx copilot --yolo interactively on Windows, some turns remain busy for many minutes without a final reply. Cancelling and sending continue sometimes obtains an answer but sometimes produces another long no-reply interval. This has occurred across unrelated tasks and sessions, including while investigating this report. These observations establish prolonged apparent stalls, not a proven infinite loop or a known root cause.

Environment and version

• Windows, Node.js 25.8.1, npm 11.11.0
• npx --no-install copilot --version: 1.0.105
• npm view @github/copilot version: 1.0.105 at the time of this report; the installed CLI is current
• Interactive invocation: npx copilot --yolo
• Historical events used normal logging; an isolated synthetic probe used --log-level debug

Redacted examples (UTC, 2026)

These are differences between recorded user turns, not precise measurements of how long the process was actively stuck. In the live investigation on Nov 4, the event journal also had intervals with no events for 9.8 minutes (17:45:18–17:55:04 UTC) and 12.7 minutes (18:04:48–18:17:27 UTC); each ended at an interrupted turn boundary after the user had to intervene.

Controlled check on the current release

I also started a separate noninteractive synthetic run on 1.0.105 using npx --no-install copilot --yolo --fleet -p <synthetic prompt> --log-level debug, with custom instructions and built-in MCPs disabled. It had not produced a final reply after roughly 2½ minutes, so I stopped it rather than leave another process running. Its journal recorded three completed sql tool calls and no subagent.started event. This is a bounded no-final-answer observation, not proof of an infinite hang or a reproduction involving subagents; stopping the run prevents knowing whether it would eventually have completed. No private code or prompts were used in that probe.

Event-journal observations and uncertainty

The sampled historical sessions record abort when cancelled. In the sampled cancellations, tool starts/completions matched, with no tool call still open at cancellation; this does not establish which component was waiting. No subagent.started events appeared in those sampled histories. I cannot presently attribute this issue to subagent scheduling, an MCP server, a particular model, the renderer, or a network timeout. Raw debug logs and session transcripts are not attached because logs may contain private context.

Steps to reproduce (intermittent)

1. On Windows, run npx copilot --yolo (1.0.105) and request a normal multi-step task in an interactive session.
2. If the turn remains busy without a final answer for several minutes, cancel it and type continue.
3. An answer may arrive after continuation, or another prolonged no-reply interval may follow. This is an observed pattern, not a deterministic minimal repro.

Expected: The turn completes or surfaces an actionable timeout/error, and cancellation reliably restores the prompt. Actual: Prolonged no-reply intervals can require repeated cancellation and continuation, without showing what is awaited.

Related issues: #4850 (background subagent stall), #4533 (parallel subagent/UI hang), #2703 (post-work hang/cancellation). Shared causality is not established.