Tools whose bugs are features, or at least have been promoted to features for the purposes of this page.

A pedagogical fork bomb with a hard ceiling. It doubles the process count every 200 ms until it hits RLIMIT_NPROC / 2, then stops and prints a histogram of how long each generation took to spawn. Used exactly once, at a workshop, to demonstrate why ulimit exists. It has been sitting in the archive ever since, and I refuse to delete it because someone might ask me to teach that workshop again.

Broken behaviour: will deadlock under some combinations of cgroup limits. Fix: unknown. Accepted: yes.

A drop-in replacement for cat(1) that is bytewise identical in output, but reads one byte at a time using read(fd, buf, 1). Exists to give junior engineers a hands-on understanding of why the standard cat is ~2000× faster on real files. I have, on two occasions, watched someone have the "oh" moment in real time. Worth its weight.

Broken behaviour: is, itself, the joke. Not actually broken. Deliberately stupid.

A UTF-8 encoder that, once in every 256 characters on average, emits a syntactically valid but semantically incorrect byte sequence. Useful for stress-testing decoders that claim to be strict. Every decoder I have pointed this at has eventually lost. Every single one.

Broken behaviour: intentional corruption. Fix: turn it off. Accepted: yes, behind a loud flag.

Like yes(1), but it gets tired. Emits "y" at a rate that decays on a negative exponential with a half-life of seven minutes. Written in response to a specific argument about whether a certain load-testing harness would ever notice if a client's throughput curve was not, in fact, flat. The harness did not notice. The tool stayed.

Broken behaviour: pretending to be a benchmark fixture. Fix: don't use it for that. Accepted: reluctantly.

This one also appears in the gallery. It is here too, because on three specific kernel revisions it panics when /proc/stupid/self is read twice within the same jiffy. I know which revisions. I have chosen not to fix it, because the behaviour is extremely funny and no one is running it in production (see front-page warning).

Broken behaviour: real kernel panic on real hardware. Fix: known, four lines. Accepted: for now.

A one-line wrapper around tee(1) that prints, to stderr, a running estimate of how much data it has seen so that you know whether the pipeline is stuck or merely slow. Violates the principle that a tool in a pipe should shut up. Violates it happily.

Broken behaviour: chatty on stderr. Fix: none planned.