Skip to content

Multi-agent audit pattern for configuration health checks #189

@dean0x

Description

@dean0x

Summary

A configuration/health-check audit workflow using parallel specialized agents with tiered model assignment. Each agent audits a specific area; a synthesis agent merges findings into actionable tiers.

Key Features

  • 6 parallel audit agents, each focused on a specific config area (CLAUDE.md, MEMORY.md, Skills, MCP, Commands, Settings)
  • Tiered model assignment: Haiku for counting/listing, Sonnet for judgment, Opus for synthesis
  • Missing-file tolerance: agents note absence and continue rather than failing
  • Structured output: Quick Wins / Medium Effort / Deep Changes / Behavioral Adjustments

Competitive Reference

Ref: SIEVE — agent-prompts.md, SKILL.md:192-213. Six specialized agents + Opus synthesis. Explicit model tiers.

Implementation Notes

  • Follows existing review:orch pattern (parallel reviewers + synthesizer)
  • Could become /audit command or devflow:audit:orch skill
  • Audit targets: CLAUDE.md, MEMORY.md, skills, MCP servers, commands, settings
  • Consider: partially overlaps with existing /audit-claude plugin — evaluate whether to extend that or create a broader audit
  • The tiered model pattern (Haiku counting, Sonnet judging, Opus synthesizing) could become a reusable convention

Milestone

post-v2

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions