runward gate

August 7, 2026

Version updated for https://github.com/stranxik/runward to version v0.33.0.

This action is used across all versions by 0 repositories.

Action Type

This is a Composite action.

Go to the GitHub Marketplace to find the latest changes.

Action Summary

Runward is an open-source tool designed to verify the engineering decisions behind AI-generated code by using plain code. It automates the verification process of full engineering missions from framing through handover, ensuring that the load-bearing decisions are followed and documented effectively. Runward provides a deterministic gate that checks against predefined expectations, preventing unintended outcomes and ensuring code integrity.

What’s Changed

0.32.0 fixed what an audit found. This one measures what the net that guards it would catch, and publishes the answer with everything in it that counts against us.

No behaviour of the gate changes. 24 golden outputs across four missions and six flag combinations are byte-identical to 0.32.0, exit codes included.

How much does this test suite actually detect

A full mutation pass (Stryker 9.6.1) on the seven library modules the verdict is computed from. Mutation score 60.78 % — 2 973 mutants, 1 769 killed, 38 timeout, 1 166 survived, in 2 h 35.

A survivor count is not a defect count, and treating it as one produces a day of false findings. So: 433 survivors carry mutators able to flip a decision and were re-run against the whole net (unit suite, self-gate, OSCAL validation, end-to-end smoke). 53 died there, 380 survived everything. 246 of those were then instructed one function at a time, each by applying the mutant to a real mission and reading the verdict rather than reasoning about the code. 181 now die, measured centrally rather than claimed.

Three mechanisms were correct in every shipped release and guarded by no test:

Mechanism	What one mutation does
Seal tamper detection	One field returned false in `verifyEvidenceLock` takes a sealed, tampered mission from exit 1 to exit 0. `check.ts` gates the whole seal section on it, so the violations were neither printed nor counted. Reproduced by hand.
The ReDoS screen (ADR-0020)	The nested-group collapse loop was entered by no fixture. It could be deleted with the suite still green.
Pointer containment	The repository fallback was dead code under test: every containment test ran in a bare temp directory.

One correction the pass forced on its own reading, and the reason ADR-0046 exists: a surviving mutant is not automatically a false green. Forcing artifactState to call every ADR directory filled survives the unit suite, the self-gate, the smoke run and the audit corpus — and the mission is still refused, because a typed pointer does not resolve. Defence in depth. What it corrupts is the printed line, and for this tool a proof surface that lies under a correct verdict is a defect of its own.

Mutation testing is adopted as an instrument, never a gate. No score is ever a crossing condition: a number in a manifest is a verdict satisfied by a figure nobody re-derived, which is exactly what ADR-0045 forbids, and runward does not do to itself what it refuses from an operator.

The verdict is now computed where a test can reach it

The largest absence of that measurement was the one that mattered: src/commands/check.ts, where the verdict is assembled and the exit code chosen, sat at 8.70 % line and 0 % function coverage, no unit test imported it, and the mutation pass could not reach it at all.

src/lib/verdict.ts is now a pure reading of the mission: it prints nothing, never touches process.exitCode, runs no hook. 97.79 % line, 100 % function, and inside the measured perimeter from now on. 12 hand-written mutants, 11 killed; the survivor is argued rather than assumed (ADR-0047).

What we publish about ourselves

docs/compliance/known-defects.md — 20 entries in four classes, each with the version range it affects and the command that verifies it. It lists both directions: the undue passes and the five undue refusals that shipped in 0.31.x. A register that only published false greens describes half a campaign.
docs/compliance/regulated-adoption.md gains the axis it was missing. It was written on the shape of a SaaS questionnaire; the axis that decides for a tool that renders a verdict — whether the verdict is right — was absent from all 82 lines. New section 8 works the tool-confidence analysis through for medical device, automotive, rail and airborne software, adverse case first, and asserts no level and no class for runward in any scheme.
Its footer no longer claims every verifiable claim on the page is enforced in CI. That was a reassuring count over a set the author chose.
One published falsehood corrected: GATE_NON_SCOPE is not printed in every compliance pack. It appears only in the ISO/IEC 42001 draft — and the pack it is missing from is the one that goes to a high-risk provider.

Guards that had a broken perimeter

Four instances in one week, the recurring defect of this repository:

The reproducible-build job had never once compared two tarballs — npm pack --pack-destination does not create the directory, so it exited ENOENT on its own output path. The claim it guards is true; the guard never reached the comparison.
The overclaim guard saw neither TQL, nor TCL2/3, nor the rail classes. Widened before the section that produces that risk was written.
CITATION.cff was pinned at 0.21.0 for eleven releases.
docs/compliance/eu-ai-act.md contradicted itself two lines apart on the high-risk binding date.

Also

test/audit-corpus.js — the adversarial campaign behind ADR-0045 becomes a corpus anyone can replay against the real CLI.
regimes/eu-ai-act@2026-1744.json — the expired sheet had missed an amending regulation. The dated-facts watch now detects that the text moved instead of waiting for a date we guessed.
Unit suite 209 → 342. Whole-project line coverage 74.90 → 79.70 %.

Full changelog: https://github.com/stranxik/runward/blob/main/CHANGELOG.md