jaap-janandClaude Opus 5 208aca1191 Make a failing test run say what went wrong ...

Two suites fail on the runner and pass everywhere else, and every attempt to work out why
has been an inference from a filename. The runner prints the path of a log written to a
disk nobody has a shell on, and the log is where the exception type, the message and the
stack all live — so a red build has been a guess, and the last guess was wrong: 69 layout
failures looked like missing fonts and were a missing shared library instead.

This prints the log, and three facts about the machine that no log will ever carry: which
distribution it is and who the job runs as, whether docker answers, and — the one that
matters for the layout suite — ldd against the libSkiaSharp.so the test project carries,
filtered to its unresolved rows. A managed TypeInitializationException on SKImageInfo is a
symptom several missing libraries share; ldd names the library. The fontconfig step ahead
of this exits early when ldconfig already reports one, so if that is present and Skia still
will not load, the answer is a different dependency and this is what says which.

head rather than tail on the log, which is the whole trick. A suite that fails wholesale
writes one stack per test and they are the same stack; the first explains it and the last
two hundred lines are that sentence repeated.

if: failure() and exit 0, so it runs only on a red build and reports without becoming a
second failure on top of the first.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

2026-08-01 14:32:32 +02:00

..

workflows

Make a failing test run say what went wrong

2026-08-01 14:32:32 +02:00