Tokf – Stop wasting LLM context on verbose build output

If you use Claude Code (or similar AI coding tools), you've probably noticed that a single `cargo test` or `npm run build` can dump thousands of tokens of noise into your context window. Passing tests, repetitive warnings, full stack traces — stuff that's useful for humans scrolling through a terminal, but wasteful for an LLM that just needs to know what failed and why.

tokf is a Rust CLI that intercepts command output and compresses it using TOML-defined filters. You can strip passing tests, collapse repetitive errors, trim stack traces, etc.

The tricky part was making it work with build orchestrators. Hooking `just test` doesn't help if the noisy output comes from the `cargo test` that just runs internally. I ended up solving this three different ways depending on the tool:

- make/just: use their --shell flag to route execution through tokf - git hooks / other tools: prepend a shim directory to PATH

Wrote up the full rabbit hole: https://pecan.si/posts/getting-into-make-just-and-hooks/

Built in Rust, ~1,800 daily executions in my workflow. Filters are TOML with a Lua escape hatch. Open source: https://tokf.net