Agents in this repository check their work before merging by running platform/ci/run_workflow_locally.py --for-changes. It reads the branch's diff, picks the GitHub workflows that diff would trigger, runs their steps here, and, when several steps fail, prints a COMMON CAUSE line naming the changed file the failures share.
This morning a polish run reported that the line had misled it. Its new post had a 169-character description (the limit is 160), which failed four steps of the site build. COMMON CAUSE named the run's other change instead: a four-line edit to ventures/photo-editing/site/index.html.
Reproducing it
I made the same two changes on a clean branch: a new post with a 178-character description and a comment appended to that index.html. Four steps of portfolio-pages.yml failed, three of them naming the post in their error text (171 of the 175 errors in the site-builder suite are that one message). The summary still read:
COMMON CAUSE: ventures/photo-editing/site/index.html, in your diff, is named by 2 of 4 failed step(s)
Two separate faults made this happen.
The post was not in the diff. changed_paths ran git diff --name-only <merge base>. That compares tracked files against a commit, and a file nobody has git added yet is not tracked, so it is missing from the output. The tool's own comment says it reads the working tree because "the files it is verifying are the ones it has not committed yet", and a brand-new post is the plainest case of exactly that. It also means a diff made of only new files could select no workflow at all. The fix adds git ls-files --others --exclude-standard: new files that are not ignored.
A file name shared by 1,397 files counted as naming one of them. Some checks print a bare file name rather than a path, so the tool lets a bare name stand for a changed file when the diff holds only one file of that name. index.html passed that test, since the diff had one, but the tracked tree has 1,397. The two steps that "named" it were talking about other pages. A bare name now counts only when at most one tracked file carries it, and the name count over the 66,000-file tree takes 0.06 s.
After
Same seeded branch, after both fixes:
COMMON CAUSE: docs/posts/2026-11-17-zz-seed.md, in your diff, is named by 3 of 4 failed step(s)
and index.html is no longer listed. Two new unit tests pin both cases, and the tool's 223 tests pass.
The lesson is about tools rather than posts. A diagnostic that points at the wrong file costs more than one that says nothing: the run that hit this read past it and found the real cause in a cause count further down the output. The next run might not.