What problem does it solve? When a project opts into multi-model audit (audit_model: multi), auditors need a second opinion from independent backends before reaching a verdict. This Skill documents how to fan a code review out across the codex and GLM (z.ai) backends in parallel, converge their verdicts with the in-session Claude verdict, and fold the per-backend results plus disagreement flag into the audit output. ## Core Features & Use Cases - Single tool surface: Invokes the mcp__moai__audit_multi MCP tool with the synthesized claude_verdict object, target, focus, gates, and session_id parameters. - Independence rule enforcement: Ensures secondary backends receive only (target, focus) so codex and GLM remain uncorrelated super-reviews rather than re-samples of Claude's reasoning. - Convergence policy: Applies a 4-case table deriving overall_verdict from required/advisory gates, with fail-open identity when optional backends are inconclusive. - Use Case: A plan-auditor finishes its Claude review of uncommitted changes with a concurrency focus, calls audit_multi, receives a required split (codex fails, Claude passes), and reports a conservative FAIL naming the failing backend in the residual-risk section. ## Quick Start Ask the auditor to run a multi-model audit of the uncommitted changes with focus on concurrency and fold the convergence result into the audit verdict.
