---
name: feedback_session_token_hygiene
description: Proactively suggest /clear and /compact — token cost is dominated by marathon sessions
metadata: 
  node_type: memory
  type: feedback
  originSessionId: 139f6ae6-9ebf-43ae-95f9-e67b336fa662
---

Shane's token spend is driven almost entirely by a few extremely long sessions, not by config overhead (no MCP/CLAUDE.md/hooks; permission allowlists are harness-side and free). Across all sessions: ~576M cache-read + ~60M cache-creation tokens, with cache-read:real-input ≈ 13,000:1 — i.e. paying to re-read accumulated context every turn. The top 4 sessions (one at 694 turns / 778K-token context) were ~95% of usage. None ever compacted.

**Why:** cache-read cost grows ~quadratically with session length (turns × context size); a 694-turn session re-reads its whole context 694 times. Small tool outputs (287KB of Reads in the worst session) are not the problem — accumulation is.

**How to apply:**
- Suggest `/clear` when switching to a distinct task/ticket (his work is naturally per-ticket: PO clone, chainlink wizard, upload audit). Default to one session per ticket.
- Suggest `/compact` (or rely on auto-compact, now enabled) once context passes ~40-50%.
- Push wide grep/read exploration into subagents (Explore agent) so bulky intermediate output never lands in main context and get re-read every turn.

Related: [[evolution_workspace_mapping]]