prompt-cache-economicslisted

# Prompt Cache Economics ## Overview Treat prompt cache as a real systems budget. Anything that changes the prefix, tool set, model parameters, or thinking configuration affects latency and token cost, so cache reuse has to be designed into the architecture from the start. ## Source Anchors - `src/constants/prompts.ts` - `src/services/compact/prompt.ts` - `src/tools/SkillTool/prompt.ts` ## Workflow 1. List every factor that participates in cache keys, such as system prompt prefix, tool set, model, and thinking config. 2. Classify prompt content by stability and keep the most stable material as far forward as possible. 3. Use an explicit boundary to split static and dynamic content so runtime bits do not contaminate the cached prefix. 4. Design summary, compact, and forked-agent prompts for cache-safe reuse, including text-only and no-tools modes where needed. 5. Put hard budgets on skill listings, tool descriptions, and server instructions so discovery text does not consume the main task budget. 6. Move late-changing data such as MCP connection state or experimental deltas into uncached tails or delta attachments. 7. When forks must share cache, keep tools, model, and thinking parameters exactly aligned with the parent request. 8. Keep summary scratchpads disposable and only persist the actual summary content back into context. ## Design Rules - Favor a stable static prefix and a narrow dynamic tail. - For summary agents, forbid tool calls when a single wasted turn wo