Report where model spend and context went, and what to change

Run node ~/.pi/agent/extensions/pi-extensions/scripts/context-report.mjs and read it as the source of truth; quote its numbers rather than estimating.

Report, briefly:

1. Total spend, spend per turn, and the share that is cache reads versus fresh input and output.
2. Cost concentration: any model taking a large share of spend on a small share of calls. Check whether ~/.pi/agent/router.json still routes to it, and whether a cheaper model of the same family is in pi --list-models.
3. Compaction: where it fires (p50/p90) and whether large contexts are being re-sent.

End with at most three concrete changes (a router.json pair, a setting, a habit), each with the number that justifies it. If nothing stands out, say so.