/usage suggests ways to save usage and can queue fixes for them

What

/usage shows session cost, plan limits and activity. It now looks through your local transcripts over a 24-hour window and a 7-day window and works out which habits cost the most. The work runs only when the server-controlled switch tengu_tidy_abacus allows it; that switch falls back to on when the server sends nothing.

• Behaviours it measures include max effort, workflow agents, always-on context (memory, skills, MCP servers, agents), large tool results, cache misses, long context, subagents, skills, plugins and MCP servers.
• It turns them into ranked suggestions, for example "Switch from max to ... effort", "Compact before you step away" (/compact), "Start fresh when you switch tasks" (/clear), "See what fills your context" (/context) and "Trim ...".
• The summary text now lists these server-provided recommendations, each as a percentage, a title and an optional detail. It used to list behaviours above a fixed percentage.
• The usage report schema gains the optional recommendations list (key, pct, title, detail) and an optional fix object holding keys and messages. The older agents, skills, plugins and mcp_servers lists stay.
• A fix is sent as one or more of your own messages, queued at a low priority to run after the current work, with any /effort commands first. The fix prompt begins "Fix what /usage flagged. Show me each change before you make it."
• A review prompt can gain a section, ## Check 10 — what /usage flagged, listing what /usage measured from the last 7 days of local sessions and asking the model to report each item with its share. It is empty unless several conditions hold.

Why

/usage moves from a plain report to advice you can act on, and a chosen fix asks Claude to show each change before making it.