Implementation planning with optional adversarial cross-review (Claude Code ⇄ Codex). Tasks are sequential vertical slices, each carrying its own testable behaviors. Pass --visual to also render the plan as a self-contained HTML page (PLAN.html). Pass --skills <skills> to preload skills that shape the plan and that execute loads before coding
Execute all fixes from a REVIEW.md — address warnings, blocking issues, and gaps found during review
Test-driven execution loop — one task at a time, red→green→refactor, atomic commits
Plan vs implementation verification with goal-backward analysis and test-quality audit
Move completed features from wip to archive
Socratic interview that captures domain language, decisions, and ADRs inside .neural/wip/<feature>/
Harvest knowledge from archived features into .neural/knowledge/ — run after archiving a feature, or invoke manually to rebuild the project knowledge base
Systematic root-cause investigation for bugs and failures