What problem does it solve?
When a project opts into multi-model auditing, a single reviewer's verdict may miss defects that another model would catch. This Skill defines how plan-auditor and sync-auditor agents fan a code review out across the Codex and GLM backends in parallel, converge their verdicts with the in-session Claude verdict, and fold the per-backend results and disagreement flag into the final audit output.
Core Features & Use Cases
- Single MCP entry point: Documents the
mcp__moai__audit_multi tool, its input parameters (claude_verdict, target, focus, gates, session_id, project_root), and its structured ConvergenceResult output.
- Independence enforcement: Explains the load-bearing rule that secondary backends receive only a scope, focus area, and directory — never Claude's analysis — so their verdicts remain uncorrelated second opinions.
- Convergence policy: Defines the 4-case table deriving overall_verdict from required/advisory gates, including fail-open behavior when optional backends are inconclusive.
- Use Case: An auditor reviewing uncommitted changes in a project configured with
audit_model: multi calls audit_multi with its Claude verdict, receives pass/fail verdicts from Codex and GLM, and surfaces a residual-risk note when the backends disagree.
Quick Start
With the project configured to audit_model multi, ask the auditor agent to run a cross-model audit of the uncommitted changes and report the converged verdict with any backend disagreements.
