Automatic compaction keeps timing out ("summarizer did not settle within 300s") #5054

Describe the bug

Title: Automatic compaction keeps timing out ("summarizer did not settle within 300s") on large contexts with xhigh reasoning, and retries almost every turn

Environment

• Copilot CLI on Windows (current machine 1.0.89). Model gpt-6-astra, token limit 1,050,000, reasoning effort xhigh.
• One long-running session (id d3a5658e-386a-4d79-97f2-b5d70a58e7aa).

What happens

• 131 compactions so far; 110 failed:
  • 100x "automatic compaction summarizer did not settle within 300s of starting; pending summary discarded"
  • 5x "compaction model call failed: RetriesExhausted { retries: 5, ... api_error ... }"
  • 3x "Compaction Cancelled"
  • 2x "background compaction summarizer did not settle within Ns; proceeding without the pending summary"
• The threshold triggers at about 840K tokens. The summarizer runs at effort xhigh over about 840K-920K input tokens. The last successful compaction took 209 s, close to the 300 s cap.
• After a failure, the CLI retries on almost every turn while the context keeps growing, so each retry is slower. Example: on 2026-10-04 it failed 15 times in a row between 05:34 and 09:15 (local), as the context grew from 842K to 924K tokens.
• To the user, compaction appears to run after nearly every message. Each attempt wastes a large model call and adds latency.

Suggestions

• Run the summarizer at lower reasoning effort, or with a faster model, regardless of the session's effort setting.
• Scale the timeout with the input size, or keep a summary that is late instead of discarding it.
• Trigger compaction earlier, or summarize incrementally, so that input stays well under the time budget.
• After a failure, back off instead of retrying every turn, and tell the user that compaction failed.

Affected version

No response

Steps to reproduce the behavior

No response

Expected behavior

No response

Additional context

No response