Measured per round in a pro session: the person's MCP tool schemas were 10k tokens in a bare folder and 58k inside the repo (294 tools), sent every round before the model did anything. They are now two small tools, mcp_tools (list, or one tool's schema on request) and mcp_call. A context map — the piece lifecycle and the common calls with their real signatures, generated from the bundled API map (~750 tokens) — rides in the system prompt, so a simple piece needs no reading rounds; the model is told not to browse the repo for examples unless asked. Tool output is capped lower (16k chars, 200-line reads), and old results are elided only once a history passes 240k characters: eliding early broke the provider's prompt cache and saved nothing. Same flower prompt: finished in 3¼ min on ~284k tokens, against a turn that ran 5½ min on ~767k and was cut off. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>