New gdd-auditor sub-agent — independent reviewer that audits a draft Game Design Document against a 9-category checklist (state/lifecycle, failure recovery, win/loss specifics, onboarding, balance numbers, feedback gaps, mechanic interactions, asset scope, skipped sections) and returns 5-8 high-impact follow-up questions per pass. Read-only, MUST NOT modify project files.
tests/test_agents.py — parametrized validation of agents/*.md frontmatter (parses, required name/description/model, name matches filename stem, model is a valid alias or claude-* id). 21 cases; protects future agent additions from drifting out of contract.
game-planner skill now runs two fixed audit rounds (Rounds 6-7) after synthesizing the GDD draft and before showing it to the user. Each round dispatches the new gdd-auditor with a fresh context; questions are batched 5-8 at a time. Round 7 explicitly populates the auditor's Previously Asked field to avoid repeats. Pattern follows GSD's Generator-Verifier loop with fresh-context audit rather than self-critique.
Wiki (EN + zh) updated to document the new gdd-auditor agent and the two-pass audit loop in game-planner — core-skills.md (game-planner row) and codebase-guide.md (new agents/ section enumerating all 5 sub-agents with role + dispatcher).