The Advisor, new providers, and more

1 October 2026 · by @dubsthedev

I added Grok and Muse to Mercury after seeing a lot of people talking about Grok 4.7. I loaded up some credits and gave it a shot myself. It’s actually better than I expected, and I’m looking forward to trying Muse too.

The advisor now works on intervals in minutes rather than turns. That felt more natural and less awkward to me. Each chat has its own switch, with more deliberate control over how often the advisor checks in and how much usage goes towards it.

There were also compaction fixes following problems I encountered in long GPT sessions. One fault was in the timeout detection: the summary request’s watchdog watched for text or reasoning output, but missed lower-level activity arriving on the connection. It could stop a live request after two minutes, leaving the conversation too large to continue. It now tracks network activity, and the initial waiting allowance scales with the size of the conversation while retaining an overall deadline.

Automatic compaction also starts earlier, leaving room for another reply or tool result before the context limit. Turning off early compaction no longer disables emergency recovery when a conversation overflows.

I’m interested in experimenting with the advisor next, particularly a stronger Qwen model running locally with Astra or Fable advising it.