The past month I spent 40% of my token budget on bash command outputs... It's expected that models frequently use bash to edit or read files, but keeping all outputs in my context feels like throwaway usage. There are a few tools that compress command outputs e.g. headroom or beacon, and achieve impressive savings. I did it myself with Thinker and cut down input token usage by 18% on 14 open-source PR benchmarks run 3x each. I only caught this by looking deeper into my token usage. Visibility & controls on your coding agents are super important imo - you might have a rogue tool or messy AGENT.md burning a bunch of your tokens.

I'm building Thinker, a knowledge cache for coding agents. On every fresh session, agents rebuild their understanding of your codebase from scratch, then throw it away. Thinker keeps it. It distills past agent sessions and PRs into short notes about how your code works, e.g. all DB writes use `/sql-utils.py`. A small local model ranks the notes and injects the relevant ones into new sessions, so the agent spends less time exploring. Notes are rescored and updated as PRs merge, so stale context gets demoted. On 50 PRs from open-source repos, agents using Thinker (Fable 5.1, GPT-6 Astra, Gemini Flash 3.8) used 18% fewer tokens and 10% faster on average with equal correctness using LLM-as-a-judge. Per-model results and the eval harness: github.com/yoavshmariahu/thi…. Repo: github.com/yoavshmariahu/thi…. Install: `npx @zerotimedev/thinker` Unlike AGENTS.md or built-in memory, notes are generated and maintained automatically. Next I want to work on rigor, not just efficiency.