qa-orchestrator

Inject faults into a substrate, drive captures, and check that the referee — the program being graded — reached the verdict it should have. The harness knows nothing about any domain: tiers, verbs, referees and report shapes are registered by verticals, and the first vertical (bmc-sensor-audit over BMC firmware) is one of them, loaded through the same door as any other.

pip install -e '.[dev]'

qa-orchestrator list           # what this process can name

# The first vertical's scenarios live with the first vertical.
qa-orchestrator check src/qa_orchestrator/verticals/bmc_scenarios/
qa-orchestrator run   src/qa_orchestrator/verticals/bmc_scenarios/sensor-removed.yaml

PATH=examples/paper/bin:$PATH \
qa-orchestrator --plugin examples/paper/vertical.py run examples/paper/withdrawn.yaml

The last line grades a proposal's requirements with a checker that has never seen a BMC: positional arguments where the first vertical's tool takes flags, issues where it keeps findings, and a report of what it did not check, which the scenario asserts on.

Released — 0.3.4, tagged v0.3.4, Apache-2.0, on PyPI as qa-orchestrator.

0.3.4 admits the referee's 0.3 line in the bmc extra. It held bmc-sensor-audit below 0.3 for eighteen days after 0.3.0, so installing the extra resolved a 0.2 referee on a 0.1 engine while the rest of the family had moved on. Nothing here needed the ceiling -- the vertical imports only the referee's mock BMC -- and the suite passes on referees 0.2.7, 0.3.0 and 0.3.5. And the sdist ships the tools its own tests read, so its suite runs from the artifact.

0.3.3 reads a referee's report the way its profile says it is written. findings resolves a dotted path, which declines and checked already did -- a referee keeping its findings nested reported zero of them and the scenario was green. A profile may now answer a ReportSchema per mode, the way judge_argv and json_argv already vary by mode, so a tool naming its list differently in two modes can be described without renaming a published key. And a run with no machine-readable report says which of its three causes it had, rather than asserting one.

0.3.2 follows the referee's wording. bmc-sensor-audit 0.2.5 reworded the finding a removed sensor produces -- from the machine to the capture -- and sensor-removed.yaml still expected the old text, so the scenario shipped in 0.3.1 fails against every referee from 0.2.5 on. The bmc extra now floors that dependency at 0.2.5, because everything below it emits the wording the file no longer expects.

0.3.1 corrects a pointer. Two source files named a test file that does not exist as the guard for the domain-free claim -- the central claim here -- and 0.3.0 shipped that way, so a reader who followed it found nothing and could not tell whether the guard had been renamed, deleted or never written. Both now name tests/test_registries.py and its class. A check was added that refuses any test path this package names and cannot resolve, because the two were found by scanning for the shape rather than by reading the one that was reported, and the scan found one more than the report did.

0.3.0 is the domain-free rewrite. Tiers, verbs, referees and report shapes come from registries and the core names none of them; the BMC tool, its tiers and its scenarios are a vertical loaded through an entry point, like any other would be. A qa-scenario/1 file written for 0.2.x still reads and still runs.

Exit codes are the referee's own three: 0 every expectation held, 1 a verdict disagreed, 2 the run could not be completed. 2 never reads as clean.

DESIGN.md says what was locking the harness to one domain, where each lock went, the contract a vertical must meet, the scenario format, how to write a vertical, and what changed from 0.2.x for anyone with code written against it.

Apache-2.0.