jvishnefske-agent-plugins

Swiss Cheese Plugin

A Claude Code plugin providing iterative Rust software development lifecycle tooling with multi-layer defect escape reduction methodology.

Features

• Single Orchestrator Architect: All coordination through one top-level agent
• Upfront Design Review: ALL questions asked before any work begins
• Gate Validation by Exit Code: Each layer validated with exit 0/1/2/3
• Iterative Loop (ralph-wiggum): Automatic retry until all gates pass
• Layer Skip with Proof: Skip only with proven inapplicability

Installation

# Add Anthropic marketplace
/plugin marketplace add anthropics/claude-code

# Install plugin
/plugin install swiss-cheese

Quick Start

# Start with design review
/swiss-cheese "CAN-based motor controller"

# Answer all design review questions...

# Orchestrator runs 9-layer verification automatically
# Gates validate each layer by exit code
# Loop continues until ALL_GATES_PASS

Commands

Command	Description
/swiss-cheese	Start new verification with design review
/swiss-cheese:gate N	Run gate N validation (exit 0=pass)
/swiss-cheese:loop	Start iterative loop until completion
/swiss-cheese:status	Show verification status
/swiss-cheese:skip-layer N	Request layer skip (requires proof)
/swiss-cheese:cancel	Cancel active loop

9-Layer Swiss Cheese Model

┌─────────────────────────────────────────────────────────┐
│ Layer 9: Release Analysis   → Gate 9 (exit 0/1)        │
├─────────────────────────────────────────────────────────┤
│ Layer 8: Independent Review → Gate 8 (exit 0/1)        │
├─────────────────────────────────────────────────────────┤
│ Layer 7: Dynamic Analysis   → Gate 7 (exit 0/1)        │
│          Miri, Fuzz, Coverage, Timing                   │
├─────────────────────────────────────────────────────────┤
│ Layer 6: Formal Verification → Gate 6 (exit 0/1/3)     │
│          Kani, Prusti, Creusot                          │
├─────────────────────────────────────────────────────────┤
│ Layer 5: Static Analysis    → Gate 5 (exit 0/1)        │
│          Clippy, audit, deny, geiger                    │
├─────────────────────────────────────────────────────────┤
│ Layer 4: Implementation     → Gate 4 (exit 0/1)        │
│          Safe Rust, all tests pass                      │
├─────────────────────────────────────────────────────────┤
│ Layer 3: TDD Tests          → Gate 3 (exit 0/1)        │
│          Tests MUST FAIL (Red phase)                    │
├─────────────────────────────────────────────────────────┤
│ Layer 2: Architecture       → Gate 2 (exit 0/1)        │
│          Type-state, newtypes, ownership                │
├─────────────────────────────────────────────────────────┤
│ Layer 1: Requirements       → Gate 1 (exit 0/1)        │
│          Formalized FR/SR/RC/TR                         │
└─────────────────────────────────────────────────────────┘

Exit Codes

Code	Meaning	Action
0	PASS	Advance to next layer
1	FAIL	Route to root cause layer
2	BLOCKED	Resolve prerequisites
3	SKIP	Approved skip

Design Review

The Orchestrator asks ALL questions upfront:

1. Functional: What, inputs, outputs, timing
2. Risk: Failure modes, mitigations, fallback behavior
3. Rust: no_std, panic policy, target
4. Verification: Tools, coverage, timing
5. Skips: Pre-approved layer skips

Iterative Loop

Using ralph-wiggum pattern:

/swiss-cheese:loop
  → Execute Layer N
  → Run Gate N
  → If PASS: advance
  → If FAIL: route to root cause
  → Stop hook blocks exit
  → Re-inject prompt
  → Continue until ALL_GATES_PASS

Layer Skip Policy

Convenience is NOT a valid reason.

Valid: "Pure lookup table with no arithmetic" Invalid: "We don't have time"

Project Structure

.swiss-cheese/
├── design-spec.yaml      # Complete design specification
├── state.json            # Verification state
├── loop-state.json       # Loop iteration state
├── gates/                # Gate validation results
├── artifacts/
│   ├── layer-1/          # Requirements
│   ├── layer-2/          # Architecture
│   ├── layer-3/          # Tests
│   ├── layer-4/          # Implementation (src/)
│   ├── layer-5/          # Static analysis reports
│   ├── layer-6/          # Formal verification
│   ├── layer-7/          # Dynamic analysis
│   ├── layer-8/          # Review findings
│   └── layer-9/          # Release case
└── release/              # Release package

Hooks

• Stop: Blocks exit during loop, re-injects prompt
• PostToolUse: Invalidates layers when files change
• SessionStart: Restores state, shows paused loops

Skills

• design-review: Upfront question methodology
• gate-validation: Exit code gate implementation
• swiss-cheese-patterns: Type-state, newtypes, no-panic

Agents