Two ideas about where output goes, and the bugs in between

For most of huck's life, the answer to "where does this command's output go?" depended on who you asked.

There was the real answer — the operating system's file descriptor table, the thing >, 2>&1 and pipes actually manipulate. And there was huck's own answer: a software sink passed down through the interpreter, saying whether output should go to the terminal, or be captured into a buffer for $(...), or be merged with stderr.

Two answers to one question is fine right up until they disagree:

# huck, 30 July
$ huck -c 'echo "[$(exec 2>&1; echo to-err >&2)]"'
to-err          <- leaked to the terminal
[]              <- and the capture came back empty
 
# bash
[to-err]

Inside $( ), exec 2>&1 is a real dup on the real fd table: stderr now points wherever stdout points, which is the capture. huck performed that dup faithfully — and the software sink knew nothing about it, so the interpreter kept writing stderr to the terminal and the capture stayed empty. The output didn't just go to the wrong place; it went to the wrong place and was missing from the right one.

That bug had siblings, months of them: output lost from a background pipeline, stderr interleaved in the wrong order inside a captured group, 2>&1 honoured by an external command but not by a builtin. Each got fixed where it was found. None of the fixes made the next one less likely, because every one of them was the same shape — a place where the two models had to be reconciled by hand, and weren't.

Deleting a model

So the August arc did the thing that had been avoidable for months: get rid of the software sink entirely and keep only the real fd table. If the interpreter never gets to have an opinion about where output goes, it can't hold a wrong one.

That meant command substitution had to become what it is in bash — a real forked subshell writing into a real pipe — rather than an in-process execution writing to a buffer. With that in place, the sinks came out: the capture types deleted, and the parameter threading them through roughly sixty call sites deleted with them.

# huck, 2 August
$ huck -c 'echo "[$(exec 2>&1; echo to-err >&2)]"'
[to-err]

The general lesson

The reason this is worth a post isn't the fix, it's the diagnosis. A bug you fix is a bug. Bugs that keep arriving in different code with the same shape are a design telling you something, and no amount of care at the individual call site will out-run it. The tell here was in the fix history: each patch reconciled two representations of one fact, which is an admission that one of them shouldn't exist.

Removing a model is a strange kind of progress — nothing new works that didn't work before, and the diff is mostly red. What changed is the class of bug that can be written at all.