Permission prompts auto-cancelled with "Session host did not acknowledge 'permission.requested' delivery within 5 seconds"; session gets stuck until exit and resume

[mandyallstars]

Describe the bug

In an interactive session, tool calls that need a permission prompt sometimes fail instantly and no prompt is shown. The tool result is:

native tool permission flow failed: GenericFailure, permission host call failed: GenericFailure,
Session host did not acknowledge 'permission.requested' delivery within 5 seconds

The matching permission.completed event is:

{"result": {"kind": "cancelled", "reason": "permission_prompt_delivery_failed"}, "decisionSource": null}

When this starts, it keeps happening for the rest of the burst. Tool calls that don't need a prompt (reads, edits of already-approved paths, pre-approved shell commands) keep working. Any call that needs a prompt fails. The burst ends with a session.error:

Execution failed: Error: Request session.send failed with message: session event delivery failed:
Session host did not acknowledge 'permission.requested' delivery within 5 seconds

At that point the session doesn't accept any more input. The only way to recover is to exit and run copilot --resume. The failure is fail-closed, so it's safe, but work in progress is interrupted and the agent may be mid-task.

Affected version

1.0.88 (all failures observed). Now on 1.0.89; not yet re-observed.

Steps to reproduce

Not deterministic. The common factor we saw is several interactive CLI sessions open at the same time:

1. Open 4–5 copilot sessions in separate VS Code integrated terminals (Windows, PowerShell 7).
2. Work in one session for a long time (hours, several auto-compactions) with default permissions, so shell commands regularly prompt.
3. Eventually a prompt fails as described above, and the failures repeat until session.error.

Evidence (from local events.jsonl, 3 affected sessions)

Session	Prompts requested	Auto-cancelled (permission_prompt_delivery_failed)	Bursts	Other CLI sessions active during each burst
A	137	39	2	4
B	114	13	2	4
C	218	13	2	3

• In every burst, 3–4 other CLI sessions were active on the same machine within the preceding 10 minutes. Isolated one-off failures also happened when 0–2 other sessions were active.
• Every burst ended with session.error (errorType: "query", then session.send failed), followed by a manual session.resume.
• open-sessions-state.json still showed old sessions with "working": true days after they ended. This may be related if the host sends prompt delivery to the wrong or a stale session.
• In the failure windows, calls that didn't need a prompt succeeded normally, e.g. 70 edit and 64 powershell calls in session A. That points to the prompt delivery/ack path, not the tool runtime.
• Some bursts came right after an auto-compaction, but not all of them.

Expected behavior

• The permission prompt is shown to the user, or
• if delivery to the UI really fails, the CLI retries or re-attaches instead of permanently breaking session.send, so the session keeps working without an exit and resume.
• A 5-second ack timeout seems short for a busy host with several sessions; making it configurable or retrying it would help.

Environment

• OS: Windows 11 (10.0.26200)
• Shell: PowerShell 7.6.6
• Terminal: VS Code integrated terminal
• Model: claude-opus-5.5 (also seen with claude-sonnet-5)
• Multiple concurrent interactive sessions, same user and machine

Related

• --resume replays orphaned permission.requested events (no matching permission.completed) after mid-prompt process death #4259 and Orphaned  permission.requested  event replays on every session resume, causing repeated directory-access prompts that can't be dismissed #4469 (orphaned permission.requested events replayed on resume). Different symptom, same event path.