Written by an AI agent working in this repository.
Sessions here check their work before pushing with python3 platform/ci/run_workflow_locally.py --for-changes. It reads the paths: filters of every workflow, works out which ones the diff would trigger, and runs their steps on this machine.
It works out the diff from the working tree: committed changes, edits not yet committed, and new files not yet added. That is deliberate. A session usually runs it before committing.
The steps it then runs did not all agree. ventures-tests.yml runs run_venture_tests.py --since "$BASE", and that script listed the changed ventures with git diff BASE HEAD, which sees commits only. So an edit to a venture that had not been committed yet:
- selected
ventures-tests.yml(the working tree said a venture changed), - reached
run_venture_tests.py --since, which saw no changed venture, and - printed
0 suites run, 0 failedand exited 0.
A pass, for a suite that never ran. I reproduced it with a one-line edit to ventures/gage-rr/engine.mjs: 0 suites run, 0 failed. The same was true of clauses.mjs --changed-since, the check that every touched venture page carries the no-warranty clause.
The fix
Both now diff against the working tree (git diff BASE, no HEAD), and, when not on GitHub Actions, add untracked files. On a CI checkout the tree is clean, so the two forms give the same answer there; untracked files stay out on CI because a build step's leftovers are not a change the pull request made. The same edit now gives ok gage-rr (0.2s, 43 tests).
One step is still commit-only: the tsc step in ventures-tests.yml diffs "$BASE" HEAD in shell, and that file is under owner review. So --for-changes now says so up front when the diff has uncommitted files:
note: 2 changed file(s) are not committed (...); steps that diff "$BASE" HEAD
(the tsc step in ventures-tests.yml) will not see them. Commit first for a full verdict.
Why it matters
A check that reports a pass it did not run is worse than no check at all. The session stops looking. Two run records on 2026-11-18 describe venture pull requests that passed the local run and then went red in CI. The records don't say whether their work was committed at the time, so I can't tell whether this gap played a part. It could have produced that outcome on its own. Tests for both scripts pin the new behaviour: an edited venture and a brand-new one are both found locally, and only the edited one on Actions.