Changelog

0.4.0 (Draft)

• Package the kit as an agent skill under skills/scaffold-kit/: SKILL.md entry, the Conductor and Claude addendum under references/, the four templates under assets/templates/ without numeric prefixes, a Claude Code plugin manifest and single-entry marketplace, and an Agent Plugins manifest.
• Engineering Document section 11: bound verification effort by the consequence of the requirement it proves; a harness, campaign or observer is a build item with its own requirement ID. State that decision statuses and claim confidence are two vocabularies. Section 16: anti-dark-code unknowns and coverage rows land here as U- rows.
• Slice Brief: a SLICE STATE block; a Gate column and an unlisted-inputs row in the acceptance table; an evidence-bound checkbox; a step-level-planner line in the guardrails; the approval footer says that idea approval is not brief approval.
• Decision Log: rule 6, evidence stays out of entries and the index stays a one-page view.
• Conductor: the brownfield procedure points at the anti-dark-code artifacts; a spoken build gate after slice selection; a mechanical-audit line in the Phase 6 checklist; three failure modes to refuse (sunk cost, unrequested harnesses, "too obvious" briefs).
• New reference working-with-anti-dark-code.md: map artifacts and where they land, the vocabulary mapping, acceptance-to-gate binding, proportionality, one protected-areas list, hygiene and the audit as a gate.
• New scripts/kit_audit.py with tests: banned punctuation and words, decision-block and log statuses, index and entry agreement, superseded links both ways, section cross-references, state block while in interview.
• Claude addendum: current documentation URL, an unattended-agents-never-touch list in the CLAUDE.md stub, packaging text replaced by the skill layout.
• License: FSL-1.1-MIT, in the repository root and inside the skill directory.

0.3 (Field-Tested), 2026-07-30

• Imported as the initial commit of this repository. See the README revision notes.