Waystone
An audit plugin for Claude Code that catches what AI-generated code gets wrong.
The Problem
AI writes code fast. It also:
- Writes 1000-line files and never hooks them up
- Assumes how APIs work instead of reading docs
- Patches failures with sledgehammers instead of understanding
- Writes tests that cover structure but not behavior
- Creates code nobody asked for
Rules and workflows help, but they describe process. They don’t verify outcomes.
The Approach
Waystone treats quality as verifiable criteria, not just steps to follow.
Universal quality (applies everywhere):
- Error paths are logged
- Public functions have tests
- APIs are used as documented, not guessed
Project quality (configurable):
- Coverage thresholds
- File and function size limits
- Domain-specific invariants
Key Components
Agents
- Structural Auditor: Checks metrics (size, coverage, logging)
- Semantic Auditor: Does the code do what it claims?
- API Contract Auditor: Were docs read, or did AI guess and patch?
- Spec Tracer: Can this code trace back to a requirement?
Commands
/waystone:audit-init: Build a checklist of files to audit/waystone:audit-run: Process the checklist with agents/waystone:audit-dead-code: Find orphaned, unreachable code/waystone:audit-connectedness: Trace files to entry points
Status
Completed. See the implementation report for the full build process and final deliverables.
Why “Waystone”?
A waystone marks a path. This plugin marks where AI went wrong, and lights the way back.