正文 Markdown
I want to share some unglamorous work I’m proud of. Starting in late March and landing around the Apr 7 OpenClaw release, we intentionally cut aggregate OpenRouter token usage by ~35%, down to ~400B tokens. This wasn’t one trick. It was good engineering across paths agents hit all day: oversized tool results, compaction/overflow recovery, background and subagent context, and prompt caching. My part was mostly the tool-result / overflow path. Vincent worked on cache boundaries and fingerprints. Boris fixed deterministic tool ordering and cache-preserving compaction. Ayaan helped with subagent light context and nearby context-shaping fixes. Relevant commits in reply.