Multi-Model Verification

Parallel code verification using multiple AI models for comprehensive quality assurance. Each model focuses on a different aspect, and results are aggregated with severity classification.

Background

Inspired by Pi Coding Agent Workflow Extension's multi-model verification pattern. Instead of a single reviewer, multiple models verify code simultaneously, each with specialized focus areas.

Verification Roles

Model	Role	Focus Areas
opus	Architecture Reviewer	Design patterns, separation of concerns, extensibility, security architecture
sonnet	Code Quality Reviewer	Logic correctness, error handling, edge cases, performance patterns
haiku	Style & Convention Reviewer	Naming conventions, formatting, documentation, code organization

Severity Classification

Severity	Description	Action Required
CRITICAL	Bugs, security vulnerabilities, data loss risks	Must fix before merge
WARNING	Code smells, suboptimal patterns, missing error handling	Should fix, justify if skipped
INFO	Style suggestions, minor improvements, alternative approaches	Optional improvement

Workflow

Prerequisites

• None. This skill runs fully without Agent Teams — the three reviewers spawn as parallel Agent tool calls (see "Agent Tool Fallback" below).
• Agent Teams mode is used only when available: CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 AND TeamCreate present in the tool list. The env var alone is NOT sufficient, and SendMessage presence is not evidence of Teams (see R018 Detection).

Execution Flow

1. Input: File path(s) or diff to verify
2. Spawn Parallel Reviewers:
   • Agent(opus) → Architecture & design review
   • Agent(sonnet) → Code quality & correctness review
   • Agent(haiku) → Style & convention review
3. Collect Results: Each reviewer returns findings with severity
4. Aggregate: Merge and deduplicate findings
5. Report: Unified report sorted by severity

Agent Teams Mode

When Agent Teams is available, create a verification team:

TeamCreate("verification-team")
├── architect-reviewer (opus) → Architecture review
├── quality-reviewer (sonnet) → Code quality review
└── style-reviewer (haiku) → Style review

Members communicate findings via SendMessage for cross-cutting concerns.

Agent Tool Fallback

When Agent Teams is unavailable, spawn parallel agents:

[1] Agent(general-purpose):opus → Architecture review
[2] Agent(general-purpose):sonnet → Code quality review
[3] Agent(general-purpose):haiku → Style & convention review

Output Format

## Multi-Model Verification Report

### Summary
- Files reviewed: {count}
- Findings: {critical} CRITICAL, {warning} WARNING, {info} INFO
- Reviewers: opus (architecture), sonnet (quality), haiku (style)

### CRITICAL
[opus] {file}:{line} — {description}
[sonnet] {file}:{line} — {description}

### WARNING
[sonnet] {file}:{line} — {description}
[haiku] {file}:{line} — {description}

### INFO
[haiku] {file}:{line} — {description}

### Consensus
Issues flagged by 2+ reviewers:
- {file}:{line} — {description} (flagged by: opus, sonnet)

Integration

• Works with dev-review skill as an enhanced verification mode
• Integrates with structured-dev-cycle skill at the "Verify Implementation" stage
• Compatible with R009 (parallel execution) and R018 (Agent Teams) rules

Configuration

Parameter	Default	Description
models	[opus, sonnet, haiku]	Models to use for verification
severity_filter	all	Minimum severity to report (critical, warning, all)
consensus_threshold	2	Number of reviewers needed for consensus flag
include_suggestions	true	Include INFO-level suggestions

When to Use

Scenario	Recommended
Pre-merge review of critical code	Yes
Architecture changes	Yes
Security-sensitive code	Yes
Simple formatting changes	No (use single reviewer)
Rapid prototyping	No (overhead too high)

Cost Awareness

Multi-model verification uses 3x the tokens of a single review. Reserve for:

• Critical path code changes
• Security-sensitive modifications
• Architecture decisions
• Pre-release verification

When spawning agents via the Agent tool during this skill's execution, always pass mode: "bypassPermissions". The Agent tool default (acceptEdits) overrides agent frontmatter permissionMode, causing permission prompts during unattended execution.