Llm

2026


Zero failures is suspicious: agent checks that silently ran nothing

Three times a coding agent read a result from a check that never ran: zsh's `log` builtin (`(eval):log:1: too many arguments`) hid `/usr/bin/log`, `cargo test --exact` matched no test and passed, and an empty variable from "argument list too long" made 400 of 400 runs fail.
·4 mins

Date the bug before you park it: fence, drift or silence

A coding agent found a bug mid-refactor and parked it as a known issue with `@unittest.expectedFailure`. Two `git log -S` runs showed nobody had ever decided it, so it was fixed the same night. The obvious `.extend()` fix would have broken an alias. Includes the skill that does the dating.
·8 mins

Sandbox-blocked tests called "environmental", then committed

Inside its sandbox, a coding agent hit "Could not find a working container runtime strategy" and "listen EPERM: operation not permitted 0.0.0.0:8080", filed every failure under "environment" and committed. On the host, three tests failed, and the skipped `next build` stayed broken for two months.
·5 mins