Cut Grok Bot roster context token tax

On a multi-Bot roster, weekly usage often dies from context reload, not from the work you asked for. Cursor staff (deanrie) on the prune/compact thread confirmed what fleet operators report: each turn still sends the full transcript to the model, long chats only auto-summarize near the context limit, and there is still no Compact, same-Bot New session with archive, or working-set meter.

On Sep 5, 2026 staff also accepted the fleet-wide angle: when every specialist Bot carries the same tax, Duplicate → Hide is a weak workaround, and dumping the skills catalog every turn is a separate cost signal they will pass to the team.

What burns the weekly pool

1. Full chat transcript on every reply (staff: not intended end-state; working-set primitives are missing).
2. Long-lived Chief-of-Staff threads that stay open for weeks (~200–250k input tokens per reply in one reported roster).
3. Scheduled routines that wake a fat chat and re-send the thread.
4. Skills / catalog reload on every turn (staff: skills-on-demand feedback received; not shipped yet).

Auto-summarize near the limit is not Compact. It does not give you a thin working set or a meter.

What you can do today

1. Reset chat on the heavy Bot when the thread is the problem. Durable Identity (name, description, memory, routines, skills, connectors, shared-computer files) survives Reset chat. The transcript does not — export anything you must keep first. Step-by-step: Reset a Grok Bot chat without cloning the Bot.
2. Put standing instructions in the Bot description and durable memory, and put long reference material in files on the shared computer.
3. Pause or delete routines that re-open a fat conversation on a schedule.
4. Prefer one Bot per role with a thin transcript over Duplicate → Hide clones of the same obese thread. Cloning does not fix the missing Compact control.
5. Watch Settings → Usage / weekly pool so context tax does not surprise you into On-Demand.

What is still missing

Staff say these are on the team’s radar with no timeline yet:

• Same-Bot new session that archives (does not delete) the old thread
• Manual or agent-triggered Compact that drops obsolete turns from the working set
• Visible working-set token meter vs the cap
• Skills loaded on demand, so the full catalog is not dumped every turn

Pitfalls

• Hide / Delete in the sidebar is not Compact.
• Spawning a clone Bot to “clear context” multiplies roster noise and still shares one computer.
• Bot-to-bot review loops are a separate burn path: Stop bot-to-bot loops burning weekly usage.