Audit coding-agent instruction drift and verifier health with Waza
Run a bounded, report-only audit of coding-agent settings, instruction reachability, hooks, MCP wiring, and verification evidence with actionable findings.
npx skills add agentskillexchange/skills --skill audit-coding-agent-instruction-drift-and-verifier-health-with-waza
Use Waza’s health skill when a coding agent ignores repository instructions, hooks or MCP integrations appear broken, or verifier output no longer establishes the behavior it claims to check. Waza contains eight engineering skills; this entry covers the health audit of a particular project and its agent setup. It does not install an entire development methodology or perform an automatic repair.
What this skill actually does
Invoke the installed health skill by name in a supported agent, or use /waza:health through the host plugin. Name the repository and request an instruction-and-configuration audit for a static-only review. That scope passes instructions to the bundled collector and excludes conversation history. Otherwise upstream’s summary audit can inspect bounded recent sessions for the current project; broader history requires the corresponding scope. Begin with summary collection and reserve deep inspection for explicitly requested depth or unresolved evidence-backed risks.
Collect the current agent configuration, relevant project instructions, hook and MCP executable references, existing verifier output, and generated-document relationships. Separate configuration drift from maintainability concerns. Distinguish configured integrations from actually connected ones, unavailable evidence from a failure, and stale logs from current verification. Return prioritized findings identifying the affected layer, concrete evidence, impact, coverage limits, and a diagnostic command or next action. A harmless read-only probe may clarify an already connected integration; do not launch an unknown executable merely to test it.
This audit is report-only: it does not authorize configuration edits, dependency installation, project tests, builds, generators, commits, publication, or repair. Keep credentials out of findings. A syntax linter such as agnix checks supported file rules; this workflow examines how instructions, runtime settings and existing verifier evidence fit together in the current project. General Superpowers development loops do not replace that bounded operational diagnosis.
Install with npx skills add tw93/Waza -a claude-code codex cursor -g -y, or use the documented host plugin. Multi-Framework reflects upstream’s explicit Claude Code, Codex and Cursor skill distribution and shared Markdown playbooks. The exact repository had 7,307 stars and 424 forks; release v3.40.0 was published on October 15, 2026. Metadata and instructions were inspected; the collector was not executed and this entry has not received a security review. Upstream health workflow.