skills.rest
taewook486
taewook486COMMUNITY
Explore all skills by this author →
WebsiteGitHub

moai-ref-cross-model-audit

Converge Claude, Codex, and GLM code review verdicts via the audit_multi MCP tool.

Updated Mar 22, 2026
Search Tags:#code review#codex#claude#glm#mcp tool#verdict convergence#multi-model audit
One-click install
npx skills add https://github.com/taewook486/Masters_degree --skill moai-ref-cross-model-audit-taewook486
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: moai-ref-cross-model-audit
Source: https://github.com/taewook486/Masters_degree/tree/main/.claude/skills/moai-ref-cross-model-audit
Command: npx skills add https://github.com/taewook486/Masters_degree --skill moai-ref-cross-model-audit-taewook486

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve? When a project opts into multi-model auditing, an auditor needs a consistent way to collect independent second opinions from Codex and GLM backends and merge them with the in-session Claude verdict without corrupting their independence or misreading disagreements. ## Core Features & Use Cases - Single convergence entry point: Documents the mcp__moai__audit_multi MCP tool that fans a review out to Codex and GLM in parallel and returns a structured ConvergenceResult. - Independence enforcement: Explains the load-bearing rule that secondary backends receive only the target and focus, never Claude's full analysis, preserving them as uncorrelated super-reviews. - Deterministic verdict policy: Provides the 4-case convergence table mapping required/advisory gate outcomes to overall pass/fail plus a disagreement flag, including fail-open behavior when optional backends are inconclusive. - Use Case: A plan-auditor agent with audit_model: multi finishes its Claude review, calls audit_multi with its synthesized verdict, and folds the returned per-backend verdicts and residual-risk note into the audit report. ## Quick Start Ask the auditor agent to run a multi-model audit on the uncommitted changes with focus on concurrency and fold the convergence result into the audit report.

Frequently Asked Questions about moai-ref-cross-model-audit

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I run a multi-model code audit across Claude, Codex, and GLM?▼

Call the mcp__moai__audit_multi MCP tool with your in-session Claude verdict as the claude_verdict object, plus optional target, focus, gates, and session_id parameters. The tool fans the review out to the Codex and GLM backends in parallel and returns a ConvergenceResult with per-backend verdicts and an overall verdict.

When should the cross-model audit skill be loaded?▼

Load it only when the project's audit_model setting is multi and the auditor needs a cross-backend second opinion. Single-backend paths (claude-only, codex-only, or glm-only) do not load this skill because no convergence is needed.

How is the overall verdict derived from multiple backend verdicts?▼

A 4-case table applies: all required backends passing yields pass, any required failure yields fail, a required split yields fail with a disagreement flag, and an advisory-only conflict yields pass with a disagreement flag. Advisory backends can never flip the overall verdict to fail.

What happens if the Codex or GLM backend is unavailable during an audit?▼

The missing, unauthenticated, or erroring backend is marked inconclusive in its per-backend verdict slot and convergence continues over the remaining backends. If all non-Claude backends are inconclusive, the result fails open to the in-session Claude verdict.

Why can't I pass Claude's full analysis to the secondary review backends?▼

Codex and GLM act as uncorrelated super-reviews whose value collapses if they see Claude's reasoning. The convergence engine structurally enforces this by sending only target and focus to secondary backends, so the focus field must carry a short area name, not analysis text.

Does a disagreement between audit backends block the workflow?▼

No. A disagreement_flag of true is surfaced as a residual-risk advisory in the audit report and never hard-blocks the flow on its own. Only a required-gate backend failure produces a blocking overall fail verdict.

Related Skills

Explore More Software Engineering Skills→
Seung-zeddSeung-zedd

moai-ref-cross-model-audit

Converges Claude, codex, and GLM audit verdicts via the audit_multi MCP tool.

Community
Advanced
modu-aimodu-ai
1.2k

moai-ref-cross-model-audit

Converges Claude, Codex, and GLM code review verdicts via the audit_multi MCP tool.

Official
Advanced
taewook486taewook486

moai-ref-cross-model-audit

Converge Claude, Codex, and GLM code review verdicts via the audit_multi MCP tool.

Community
Intermediate
zenobi-uszenobi-us
68

sharing-skills

Automate GitHub PR review resolution by fixing comments and failing CI checks.

Official
Intermediate