The default action scope is now pull-request; reports are version 4. Supply change-file, base-sha and an independently reviewed contract-sha256 using the PR coverage guide. All requested behavior and changed files must be covered, and named required jobs must pass. Missing inputs deny.
For intentionally narrow checks only, set scope: acceptance explicitly. Its approval covers supplied criteria, with coverage=not-assessed. Existing full-commit pins retain their behavior. No TV/SDK integration or branch rule is migrated automatically.
Version 0.1 judged whether completed CI operations were consistent with a policy. Version 0.2 asks whether supplied implementation and behavioral evidence establishes a task's acceptance criteria. Green checks no longer suffice.
This is a breaking change on main. Existing integrations pinned to an earlier commit remain on the earlier implementation. Update them deliberately; do not change a pinned reference without preparing the new inputs.
- Replace
policywith a reviewedtask-filecontaining an explicit task and required criteria. Arbitrary policy overrides are removed. - Replace
corpus-filewith the structuredevidence-filemanifest. A text log is now one named evidence item, not the entire evaluation contract. - Supply
head-shaexplicitly, using the PR head SHA. Evidence from another revision denies. - Keep
stepsorneedsfor ordinary CI results, and include all applicable dependencies. Workflowinputsandprior-inputsare no longer action inputs; include relevant context in named evidence items instead. - Replace aggregate
probabilityandconfidenceconsumption with the criterion-level report.decision,approved,reason,modelandduration-msremain, with new reason codes.head-sha,verified,totalandreport-fileare added. - Remove setup-only preflight evaluations. Acceptance verification belongs at the end, when implementation and behavioral evidence are available.
- Remove the TV-specific
collect-evidence.jscollector and old CI-consistency fixtures. The optionalscripts/evidence.jshelper packages explicit source files.
Publish a separate final job named REV. Start with it nonrequired, preserve its report on denial, and validate its judgment against known good and bad solutions before enforcing it. This release does not migrate TV or alter any repository's branch rules automatically.
The first acceptance prototype chose one primary citation per criterion. Equally useful sources split that choice's probability, creating false uncertainty about evidence selection. Report version 3 assesses each source independently and returns an evidence array, raw model verdict and all relationship assessments. It retains the existing thresholds and requires accepted implementation and behavioral support for satisfaction; a confident contradiction prevents approval. Consumers of version 2's single evidence object must update deliberately.