rc-context-budgetlisted

# Context Budget The context window is a budget, not free space. Every always-loaded agent description, skill, rule, MCP tool schema, and `CLAUDE.md` line is spent on *every* turn — a 200k window can effectively start at 70k once enough tooling is enabled, forcing early compaction and degrading reasoning. This skill inventories that fixed overhead, estimates its token cost, and tells you what to trim. Read-only: recommend, do not disable servers, edit configs, or delete skills yourself. ## What to inventory Discover and measure each contributor (skip what does not exist): - **Always-loaded prose**: `CLAUDE.md`, `AGENTS.md`, and any always-on rules/`.cursor/rules` injected globally. - **Agents**: `.claude/agents/**`, `.opencode/agent/**`, and plugin-bundled `agents/**` — every agent's `description` is loaded whenever the orchestrator considers delegating (Task tool). Long descriptions cost on every invocation. - **Skills**: bundled + installed `*/SKILL.md`. The `description` frontmatter is what loads for routing; the body loads only when the skill activates. - **MCP servers**: each enabled server's tool schemas load up front. Estimate **~400–600 tokens per tool**; a server exposing 20 tools is ~10k tokens always-on. - **Slash commands**: their `description`/`argument-hint`. ## Estimating tokens - Prose/markdown: `tokens ≈ words × 1.3`. Count words with `wc -w`. - MCP: `tokens ≈ tool_count × 500`. Count tools from the server's manifest/inspect output, or estimate from its