Hello, huck

huck is a bash-compatible shell written in Rust. That one sentence hides three separate reasons the project exists — and this blog is where I'll track how it goes.

Understanding a shell, in detail

Everyone uses a shell. Almost no one knows what it actually does. Word splitting, the order expansions happen in, how $IFS interacts with globbing, what job control really requires from the terminal, how programmable completion is wired — these are the parts you never see until you try to reproduce them exactly. The surest way to understand a system is to rebuild it and diff your output against the original, byte for byte. That's the bar huck holds itself to: every feature ships with a spec, a plan, tests, and a harness that runs the same fragment through huck and real bash and asserts identical output.

diff <(huck -c 'echo ${x:-hi}') <(bash -c 'echo ${x:-hi}') && echo identical

Learning Rust by building something demanding

A shell is a great forcing function for a language. It touches parsing, process management, signals, file descriptors, terminals, and performance all at once. huck is where I'm learning Rust in the deep end — ownership across a copy-on-write shell state, careful unsafe at the OS boundary, and a parser/lexer architecture that had to be rebuilt more than once.

Watching how Claude handles long-term development

The third reason is the most interesting to me. huck is built almost entirely with Claude, one numbered iteration at a time: brainstorm a design, write a spec, write an implementation plan, execute it with a fresh agent per task, review, and merge. It's now more than 270 iterations deep. The open question isn't "can an AI write a function" — it's whether that workflow holds up across months of real, cumulative software development: regressions, architecture changes, tech debt, and all. This blog is partly a logbook of that experiment.

More soon.