Files
aigc-home/scripts/hel840

HEL-840 isolated verification

  • Product build: cd calliope-web && npm ci && npm run check && npm run lint && npm test && npm run build.
  • Backend: Python >=3.11, install existing project [dev] dependencies in calliope-backend/.venv; run calliope-backend/.venv/bin/python -m pytest calliope-backend/tests/test_prompt_tools.py -q.
  • Capture: run python3 scripts/hel840/capture.py --output ./evidence/hel840/final. The orchestrating Python needs the existing Playwright test runtime; the server uses calliope-backend/.venv/bin/python.
  • Chromium runtime libraries may be supplied via HEL840_BROWSER_LIBS; these are local test prerequisites, not product dependencies.
  • serve_fixture.py creates only the explicitly named fixture directory. Never point it at existing user data. It serves the built product, real FastAPI router and real LLMClient, with an HTTP fake provider. Browser responses are not mocked.
  • Fake images and project rows are deterministic QA data and never compiled into product. Approved original samples are separate and untouched.
  • capture.py owns and cleans its exact child process after comparing the workspace daemon PID. It must finish in foreground; no service handoff is provided.

Additional verification:

  • python3 scripts/hel840/capture.py --output ./evidence/hel840/states --states-only runs the full interaction assertions without the 60-image matrix.
  • python3 scripts/hel840/capture.py --output ./evidence/hel840/motion --motion-only records the nine viewport/motion combinations and both rings at 1440×900 video resolution.
  • python3 scripts/hel840/restart_config.py --output ./evidence/hel840/restart stops and restarts two exact owned server processes and checks persisted assignments plus actual transform routing.
  • python3 scripts/hel840/compare_reference.py --reference /path/to/approved-v2.html --output ./evidence/hel840/tokens.json compares token values; it does not claim screenshot pixel equality.

The approved original HTML and PNGs must be supplied unchanged from the approved attachment. Keep originals, mapping and actual application captures separate. For cancellation/timeout screenshots the isolated fixture accelerates the total timeout to 1.5 seconds; the product uses 120 seconds. Complete recognized weighted expressions are preserved verbatim so their weights cannot move to another operand. Closing a request does not promise that a supplier stops charging.

Review packaging supports --matrix-run finalized --states-run finalized when the final full capture was recorded under that directory. This keeps a single source capture for the delivered matrix and states.