fix: remove hardcoded developer paths from shipped code

[abrichr]

This is a public repo, and a path under one developer's home directory was baked into library code, two tests, an example script, and a checked-in HTML artifact. PR #13 removed the two in the screenshot pipeline. This gets the rest.

The one that actually broke things

real_data_loader.py defaulted to /Users/abrichr/oa/src/openadapt-capture/turn-off-nightshift, so openadapt-viewer benchmark with no --data raised FileNotFoundError for everyone who isn't that developer.

Re-pointing that default at a sibling checkout would have looked like a fix and not been one. turn-off-nightshift/ is gitignored in openadapt-capture (.gitignore:56), so it's local data nobody else can obtain. There's no location the default could point to that would work. So the default is gone. Name a recording with --data, or set $OPENADAPT_CAPTURE_RECORDING. Missing config now prints one line and exits 1 instead of dumping a traceback.

The one that was quieter and worse

Key frames in episodes.json are stored relative to the checkout root, like ../openadapt-capture/<rec>/screenshots/step_0.png. The loader resolved them by prepending a literal /Users/abrichr/oa/src. Copy a recording anywhere else and all six screenshot paths came back pointing at a machine you don't have, with no error, because nothing checks that a screenshot path exists.

It now re-anchors on the capture directory the caller passed. I checked both directions:

• Against the real recording in place: 6 frames, 0 missing, byte-identical paths to the old behaviour.
• Against a copy under /private/tmp/.../fakeroot/: 6 frames, 0 missing. The old code returned /Users/abrichr/... here.

Deleting test_benchmark_refactored.html

I removed it rather than editing the paths inside it. It's generated by openadapt-viewer benchmark --output test_benchmark_refactored.html, no code reads it, and it embedded both the absolute paths and the step text of that private gitignored recording. .gitignore already excluded *_viewer.html; this one only survived because its name doesn't match the glob. It's ignored by name now.

verify_real_data.py is the only script that mentions it, and it already handles the file being absent without failing.

Say the word if you'd rather keep it and I'll regenerate it with portable paths instead.

Keeping it out

tests/test_no_hardcoded_paths.py fails the build if /Users/ or /home/ shows up in tracked code. I verified it fails on a planted violation, not just that it passes.

#13 merged while I was working on this, so the branch is rebased on it and the allowlist is now empty: the guard covers every tracked file with no exceptions. Re-verified by planting a violation in scripts/generate_for_web.sh, which the allowlist used to hide.

Markdown is out of scope here. About 50 status documents at the repo root still carry these paths. #15 sweeps them and drops that exemption.

Two variables, on purpose

$OPENADAPT_CAPTURE_DIR (PR #13, and tests/test_screenshot_generation.py here) names the openadapt-capture checkout that holds many recordings. $OPENADAPT_CAPTURE_RECORDING names one recording directory inside it. Overloading a single name with both meanings would have been a bug waiting to happen.

Found but not fixed

The loader reads capture.db with SELECT * FROM capture. openadapt-capture's recorder writes recording.db now (recorder.py:1292, recorder.py:2368), so this path loads pre-PR-#28 recordings only. That's a schema change, not a path fix, so I left it alone. Worth an issue.

Tests

CI is green on all five checks. pytest tests/ gives 145 passed, 46 skipped, and ruff check . passes. CLI verified end to end both with and without a recording configured.

Running pytest over the whole tree instead of tests/ turns up one more failure, test_minimal_viewer.py::test_viewer_html. It asserts on viewers/benchmark/minimal_viewer.html, which .gitignore:15 excludes, so it fails on any fresh clone. I reproduced it on a clean clone of main. CI only runs tests/, so it never sees it. Untouched here.

One thing worth calling out, since it's the kind of bug this PR is nominally about. The guard failed CI on first push because it flagged its own source, which has to spell out the prefixes it bans. It had passed locally only because the file was still untracked when I ran it, so git ls-files never returned it. Fixed in the second commit, and re-verified with planted violations in both src/ and tests/ with the file tracked.

[github-actions]

📸 Generated Screenshots Preview

Screenshots have been generated. Download the artifacts to preview them.

Generated Files:

• demo_new_controls.png (0.07 MB)
• demo_new_events.png (0.04 MB)
• demo_new_full.png (0.11 MB)
• turn-off-nightshift_controls.png (0.15 MB)
• turn-off-nightshift_events.png (0.13 MB)
• turn-off-nightshift_full.png (0.23 MB)

---

🔗 Download screenshots artifact