You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Teach the decree-ddd agent skill to classify the work mode, name an
authoritative decision and prefer intent-check --under, and interpret intent-check exit-1 findings by severity (blocking vs advisory vs corpus
maintenance) instead of resolving every finding first.
decree health gained two advisory governance-quality signals: lifecycle_drift (a decision at 100% primary acceptance criteria with
commits attached but still non-terminal, or a terminal decision whose
governance has gone stale/dead) and broad_governance (governs-count,
exact/directory split, commits ratio, and shared-hot-file overlap). Both are
advisory and never affect the health exit code.
decree intent-check gained a first-class --under view — additive owned_files, contextual_overlaps, and contradictions keys that frame
multi-governed paths around the active decision — plus an advisory directory_overlaps list that surfaces co-governance via directory governs: entries the exact-path conflict query cannot see. Both are
additive and never change the exit code.
decree intent-check now classifies planned files as source_changes / corpus_changes / generated_artifact_changes (a decree-document self-edit
no longer earns add_governance) and classes findings as blocking_findings
/ advisory_findings / corpus_hygiene_findings, with the human report
leading with a "Block now / Clean later" summary and a recommended next
command. The new keys are additive and the exit code is unchanged
(ADR-01KWXMRRB44CE78H0659D9WDY7).
Bug Fixes
decree sprint drop now works on backlog and draft-pool items, not only
active-sprint items, so a backlog item that shipped and reached a terminal
status can be removed from live membership (previously it errored and could
only be resolved by editing the store by hand).
Documentation
Add the Agentkith research backlog
(docs/dogfooding-feedback/06-research-backlog.md) grading the dogfooding
proposals against the real code and external precedent, with prioritized
buildable items B1–B15.
Document Agentkith dogfooding feedback for improving decree command output
and agent skill usage.
Internal Changes
Guard that the two decree-ddd SKILL.md copies (dev-facing and packaged
template) stay byte-identical so shipped agent guidance cannot silently
drift.