Skip to content

Testing

The Norm Editor is two languages in one repository — a Quasar/Vue frontend and a Node backend on Vitest, three Flask services on pytest — and its CI runs on GitLab, not GitHub Actions. Both facts make it the odd one out among the IOU components, and both shape what follows.

The suite arrived in one release, v2026.07.3, together with the pipeline that runs it. Before that the repository had no automated tests at all.

Figures on this page are measured, not estimated

Every count below was produced by running each suite the way .gitlab-ci.yml runs it, against 2026.09.1 on 7 September 2026, on main at 46e44c8. One exception is marked in the table and explained under What could not be measured here.

At a glance: 14 test files · 110 measured tests, all passing · plus 9 more in a suite this machine cannot run.

Service Language Runner Files Tests
gui JavaScript (Quasar/Vue) Vitest 6 38 measured
backend JavaScript (Fastify) Vitest 2 11 measured
wrap_up_api Python (Flask) pytest 3 34 measured
unwrap_api Python (Flask) pytest 2 27 measured
nlp_api Python (Flask) pytest 1 9 † not run here

† Counted from def test_ declarations rather than from the runner. The file contains no parameterised cases, so the static count and the runner's should agree — but that is an inference, not a measurement.

Running the tests

The commands below are exactly what CI runs, from the repository root.

Service Command
gui cd gui && npm i && npm test
backend cd backend && npm i && npm test
wrap_up_api cd wrap_up_api && pip install -r requirements.txt -r requirements-dev.txt && pytest
unwrap_api cd unwrap_api && pip install -r requirements.txt -r requirements-dev.txt && pytest
nlp_api cd nlp_api/API_NLP && pip install -r requirements.txt -r requirements-dev.txt && pytest

Both npm test scripts are vitest run — the non-watching form, so they exit.

Use a virtual environment for the Python services

Each of the three has its own requirements.txt and pins versions independently, so installing them into one shared environment invites conflicts. Create a venv per service.

What could not be measured here

nlp_api's suite did not run on this machine, and the reason is the runtime, not the code. Its requirements.txt pins networkx==3.6.1, which declares requires_python !=3.14.1,>=3.11. This machine has Python 3.10.12, so pip resolves no candidate and the install fails before pytest is reached:

ERROR: Could not find a version that satisfies the requirement networkx==3.6.1
ERROR: No matching distribution found for networkx==3.6.1

CI runs these jobs on python:3.14.6, where the pin resolves. This is an environment limitation on the measuring machine, not a defect in the repository — stated here so the 9 in the table is not mistaken for a measured figure, and so nobody re-investigates it as a bug.

The other two Python services install and run cleanly on 3.10, because neither depends on the transformer stack that pulls networkx in.

What CI actually gates

.gitlab-ci.yml defines two stages, and the test stage runs first:

Stage Jobs
test test_gui, test_backend on node:24; test_wrap_up_api, test_unwrap_api, test_nlp_api on python:3.14.6
build One job per service — backend, gui, nlp_api, unwrap_api, wrap_up_api, nginx — each building a Docker image and pushing it to Azure Container Registry

Every service with a test suite has a job, and the build stage runs after the test stage, so a failing test blocks the image build. The build jobs are restricted to main, develop and tags; the test jobs are not, so they run on every branch.

Images are tagged twice, with the commit SHA and with latest, and pushed to the registry named by ACR_REGISTRY_USERNAME.

This is the only IOU component on GitLab CI

The CPSV Editor, RONL Business API and Linked Data Explorer all run GitHub Actions, and the supply-chain policy documented in Supply-Chain Pinning is written against that shape — digest-pinned uses: references, a zizmor audit job, an acc branch ruleset. None of it applies here, because none of those mechanisms exists in GitLab CI in the same form. The Norm Editor's images are also built and pushed by its own pipeline rather than by a vendor action, so the container exception that dominates the other three does not arise either.

Test inventory

gui — 6 files, 38 tests

Domain logic only. The tests exercise the editor's in-memory model and its helper functions, not Vue components.

File Covers
test/unit/model/sentence.test.js Sentence structure
test/unit/model/snippet.test.js Source snippets
test/unit/model/booleanConstruct.test.js Boolean fact construction
test/unit/helpers/dateTimeFunctions.test.js Date and time handling
test/unit/helpers/sourceFormatting.test.js Source text formatting
test/unit/helpers/utilities.test.js Shared utilities

backend — 2 files, 11 tests

test/unit/routes.test.js and test/unit/helpers.test.js. Since v2026.07.4 the route tests use light-my-request, which drives Fastify's routing in-process rather than over a socket — faster, and with no port to collide on.

wrap_up_api — 3 files, 34 tests

test_wrap_up.py, test_routes.py and test_helpers.py. The largest suite in the repository, covering the service that assembles a finished interpretation.

unwrap_api — 2 files, 27 tests

test_routes.py and test_helpers.py, covering the inverse transformation.

nlp_api — 1 file, 9 tests

test_labeling.py, and it is worth reading even though it does not run here: every test targets the token-merging logic rather than the model. It covers dropping [CLS]/[SEP] specials, renaming raw labels to entity names, merging ## continuation tokens into the preceding word, a continuation inheriting the previous token's label even when its own differs, a leading continuation with nothing before it, chained continuations, a realistic mixed sentence, and empty input.

That is the right thing to test. The model's predictions are not deterministic across versions and are not this repository's code; the reassembly of WordPiece tokens into labelled words is both, and it is where an off-by-one silently mislabels a word.

What is not covered

  • No component tests. The gui suite covers domain logic and helpers; nothing renders a Vue component.
  • No end-to-end tests. There is no Playwright or Cypress suite, so no test exercises the editor through a browser.
  • No coverage measurement. Neither Vitest run passes --coverage, no pytest run passes --cov, and no threshold is configured anywhere. Coverage percentages for this component do not exist — which is why this page reports counts only.
  • The NLP model itself is untested, deliberately. See the nlp_api note above.