truereward-harwell

The recorded-state reader for truereward, over the Harwell engine.

It grades a change to a codebase that has no unit test suite. A case records what a job leaves behind when the code is right, and claims the parts of it the case is about: data sets, spool files and tables. Grading runs the same job against the changed codebase and reads those claims against the approved baseline.

• A case is a starting state, a run and claims, and nothing unclaimed is compared. A case passes when every claim of it holds, fails when one does not, and is ungradable under a declared code when one could not be evaluated. A claim that names an observable and reads no field of it asserts that observable whole, byte for byte; an observable no claim names is not read at all, so a correct change is never failed on something nobody was asking about.
• A case is data. A directory with case.json, a start/ snapshot and an expected/ snapshot, validated against the JSON Schema shipped in the package. Every field a case reads is resolved to bytes when the case is recorded, so grading never opens the candidate's source to decide what is compared.
• grade_case(case, codebase) runs the case twice in a contained run, reads the claims over the first end state and the baseline, and answers with a truereward.CaseResult. The two runs are compared with each other, over the same claimed observables, under the author's own mask, which answers the determinism check.
• record.record(...) records a case: refuse a case that names no claim, run the job twice, refuse a run that did not complete or that varied in something the case claims, keep what it left as the approved baseline, freeze every field map, and read every claim back over the baseline it just approved.
• record_again(case, out_dir, where) and moved(old, new) are the reader's half of python -m truereward rerecord -p truereward_harwell CASES OUT --where CODEBASE: each case recorded again from its own declaration, frozen policy and all, on the engine installed now, and what moved read over the claimed observables only, each at its first differing record and byte with the old and new values there, beside every declaration field that moved.
• selftest.run(...) asks whether a case set would catch a bug: the reference change must score 1.0, the unchanged codebase 0.0, and every seeded defect less than 1.0.

A task's test file:

from functools import partial
from pathlib import Path

import pytest
from truereward_harwell import cases, grade_case

CASES = Path(__file__).parent / "cases"


@pytest.mark.parametrize("case", cases(CASES), ids=lambda c: c.id)
def test_case(case, grade):
    grade(case, partial(grade_case, codebase="/grade"))

Run it with pytest -p truereward_harwell, which brings the truereward plugin with it.

Licensed under Apache-2.0.