Token-Waste

Mine today's Claude sessions for token waste: context that was loaded but didn't earn its place. Where /dream asks "did the agent act effectively?", this asks "did the agent pay for tokens it didn't need?"

Two phases, mirroring /dream:

• Phase 1 — Analyze (cheap, high-volume): a deterministic Python detector sizes every tool result and shortlists waste; a haiku waste-analyst confirms and scores. Output: /tmp/session-waste-examples-<YYYYMMDD>.json.
• Phase 2 — Adapt (one capable model, one edit): the opus waste-adapter reads that tally, picks the single highest-impact recurring pattern, and makes exactly one surgical change to an agent / skill / tool / rule so it stops recurring. One improvement per day — attributable and reversible.

Phase 1 flags, per session, in three families.