context-auditlisted

# Context Audit Find what a project pays for on **every single turn** and cut it. The expensive thing in a long Claude Code session is rarely the user's prompt — it's the **always-on context** (system prompt + CLAUDE.md + loaded tools/skills + conversation) that gets re-sent each turn and billed as cache-read. Shave 5k tokens off something that loads every turn and you save it across the whole session. ## The model (read first) - **Output** costs ~5× input — but it's usually a small share of a long session's spend. - **Cache-read** is cheap per token (~1/10th input) but you pay it on the **whole context, every turn** — so a 50-turn session re-reads its context ~50×. That multiplier is where the money goes. - Therefore: the highest-leverage cut is anything **fixed and re-sent every turn**. In order of typical size: the conversation itself (→ `/compact`), CLAUDE.md files, enabled **MCP servers** (their tool schemas), available **skills** (name + description listed each turn), memory, agents, hooks. ## Step 1 — Get the live signal (if Gander is running) Ground the audit in real spend before guessing. The bridge runs on `http://localhost:3131`: ```bash curl -s http://localhost:3131/api/usage ``` In `bySession`, find this project's sessions and read: `cacheHit`, `outShare`, `ctxPct`, `costUSD`. - **High `cacheHit` (>90%) + low `outShare` (<15%)** → the lever is **context size**, not verbosity. Focus on CLAUDE.md / MCP / `/compact`. - **High `outShare` (>25%)** → answers are v