chain-of-verification
by verttiv1.2.0
Verify code, architecture, and bug fixes using Chain-of-Verification pattern with factored multi-agent verification
Commands
verifyVerify code, architecture, or bug fixes using Chain-of-Verification
Documentation
# SE-CoVe: Software Engineering Chain-of-Verification
A Claude Code plugin that catches bugs in AI-generated code by verifying against requirements, not implementation.
## Why This Works
When you ask an AI to review its own code, it's biased toward its solution. Meta's [Chain-of-Verification research](https://arxiv.org/abs/2309.11495) showed that **independent verification improves accuracy by 28%** - because the verifier checks against requirements, not the (possibly flawed) implementation.
SE-CoVe applies this to code: the verifier never sees the draft solution. It answers verification questions by checking docs, searching your codebase, and reasoning about requirements - then a synthesizer compares findings and corrects errors.
## Built for Software Engineering
Unlike general-purpose verification, SE-CoVe is designed specifically for code. Each agent understands **35+ software engineering claim categories** across three domains:
### Correctness
Behaviors, logic flow, boundary conditions, error handling, API usage patterns
### Security
Input validation, authentication, authorization, injection prevention, data exposure
### Performance
Time/space complexity, caching, memory management, async patterns, N+1 queries
The verification pipeline includes:
- **Confidence scoring** on all claims (High/Medium/Low)
- **Severity ranking** of issues (Critical/High/Medium/Low)
- **Parallel verification tracks** for thorough analysis
- **Domain-specific checklists** for security, performance, and error handling
- **Evidence quality scoring** with source authority ranking
## How It Works
```
Question → Baseline (generate + extract claims)
↓
Planner (create verification tasks)
↓
Executor (verify independently - never sees draft)
↓
Synthesizer (compare + correct)
↓
Verified Solution
```
## Installation
First add the marketplace
```
/plugin marketplace add vertti/se-cove-claude-plugin
```
then install the plugin
```
/plugin install chain-of-verification
```
## Usage
```
/chain-of-verification:verify <your question>
```
*Tip:* Type `/ver` and press Tab to autocomplete the command.
### Options
| Flag | Short | Description |
|------|-------|-------------|
| `--quick` | `-q` | Fast verification (haiku models, 1 executor) |
| `--thorough` | `-t` | Comprehensive verification (opus models, 4+ parallel executors) |
| `--focus=X` | `-f X` | Prioritize a specific area (see below) |
### Focus Areas
| Focus | What It Checks |
|-------|----------------|
| `security` | Auth, input validation, injection, data exposure |
| `performance` | Complexity, caching, memory, N+1 queries |
| `api` | Contracts, versioning, error responses, compatibility |
| `testing` | Coverage, edge cases, integration, mocks |
| `error-handling` | Exceptions, recovery, logging, user feedback |
| `style` | Naming, organization, documentation, types |
| `scalability` | Bottlenecks, resource limits, horizontal scaling |
### Examples
```bash
# Standard verification (default)
/chain-of-verification:verify How do I implement debounced search in React?
# Quick sanity check
/chain-of-verification:verify --quick Is this null check correct?
# Thorough security review (4 parallel executors)
/chain-of-verification:verify --thorough --focus=security Review the authentication flow
# API-focused verification
/chain-of-verification:verify -f api Does this endpoint handle errors correctly?
# Test coverage analysis
/chain-of-verification:verify --focus=testing Are these unit tests comprehensive?
# Error handling review
/chain-of-verification:verify -f error-handling How should this handle failures?
```
## When to Use
**Good for:** Complex code generation, architectural decisions, bug investigations, library/API questions
**Skip for:** Trivial changes, simple questions, exploratory coding
## References
Based on Meta's [Chain-of-Verification](https://arxiv.org/abs/2309.11495) research, which demonstrated 28% improvement in factual accuracy through independent verification.
## License
MIT