Repository navigation
Testing
sameernagar-hub edited this page Oct 9, 2026
·
2 revisions
Every number in the docs can be reproduced, and every behaviour that matters has a test.
| File | What it proves |
|---|---|
test_core.py |
discovery filters, board harvesting, fit gate, tailoring and validation, answer rules, option picking |
test_v08.py |
scheduler (overlap, preemption, failure isolation, quantum maths), drafts (only open questions, only Fact Bank text, honest closings), event log (3 processes × 200 large lines never tear; incremental, tolerant reads), the long-term sponsorship rule and its airbag, no control characters in the rules, experience and "select all" answers, code matching |
test_api.py |
every API endpoint over a seeded log; the answer write path is off by default, localhost only, refuses legal/EEO/sensitive questions, replaces instead of duplicating, re-queues the job |
test_forms.py |
the real filler in headless Chromium against local look-alikes of Greenhouse, Ashby and Lever: drafted "why" answer from the Fact Bank, sponsorship from presets, EEO declined, CSS-only required field caught, follow-up questions, proof screenshot on disk, dry runs never submit |
Tests use a fictional persona (tests/fixtures/presets.json, profile.example/) and a throwaway workspace; they never
read real data. The form tests intercept requests to the ATS hosts and answer with the fixture page, so nothing
leaves the machine.
python -m pytest -q --cov=engine --cov=apps/api --cov-report=term-missingCI runs the same command on every pull request (the form tests are skipped there if Chromium isn't installed).
| Script | Claim it checks |
|---|---|
scripts/bench_scheduler.py [jobs] [tabs] [seed] |
round-robin speed-up and browser utilization (deterministic clock) |
scripts/bench_tailor.py <old tailor.py> |
the new tailoring returns identical output on every saved JD, and how much faster |
-
python -m engine report: only submissions whose proof screenshot exists. -
scheduleevents: real wall time, active time and switches for each run. -
workspace/drafts_review.md: every drafted answer with its fact ids.
Related: Home · Architecture · Discovery · Fit gate · Tailoring · Apply engine · Email codes · Outcomes · Platform · Metrics
REGEN · source-first discovery, proof-backed tailoring, receipts for every submit · Home · Architecture · Metrics · pages live in docs/wiki/ and are reviewed like code