v0.42.0

What's new

/build-rule rebuild

• Step 0 primitive routing gate — redirects to hook/skill/CLAUDE.md before intake
• Step 2 rule type classification — 8-category taxonomy sets fix-safety default, severity, and framing before drafting
• 5-component Intent requirement — failure cost and exception policy are load-bearing; rule is refused without them
• 12-criteria validation checklist — extended from 9, adding Intent completeness, single canonical example, example realism
• Co-located .tests.md written at save time — Gate 1 (3+3 min) enforced at creation

/check-rule rebuild

• 6-element LLM prompt anatomy — criterion statement, PASS/FAIL scale, anchors, CoT, evidence-before-verdict output, default-closed instruction
• Dim 6 (Intent Completeness) added — checks all five Intent components per rule
• Every finding requires a Recommendation: line with a specific actionable repair (from repair-playbook.md)
• New repair-playbook.md reference — CHANGE/FROM/TO/REASON entries per failure mode
• Evaluation and repair are now explicitly separate concerns

Shared primitive routing

• skills/_shared/references/primitive-routing.md — conversational guide to the two reliability tiers, what each primitive was designed for, and wrong-primitive failure mode mechanisms
• All four build-* skills updated to reference the shared guide and redirect appropriately

Project-wide audit fixes

• Created .claude/hooks/skill-lint.sh (was referenced in settings but missing)
• Added ## Key Instructions with explicit won't-do scope to 8 skills: audit, distill, research, build-subagent, retrospective, setup, ingest, finish-work
• finish-work — disable-model-invocation: true (performs merges, branch deletes, PR creation)
• ingest — Handoff Produces corrected (wiki/ not docs/context/); confirmation gate added before writes
• retrospective — concrete example added; approval gate promoted to first Key Rule
• setup — argument-hint corrected (was empty string)
• refine-prompt — persona framing replaced with directive framing