Code Review Swarm

The code review swarm dispatches multiple specialized sub-agents to review a pull request in parallel. Each agent focuses on a single dimension — security, performance, style, or logic — then a root agent synthesizes the findings into a single structured report.

How it works

Parallel sub-agents — When a review is triggered, the root agent spawns one sub-agent per enabled dimension. All sub-agents receive the diff and any relevant context (README, test files, dependency manifest) simultaneously and run in parallel
Dimension analysis — Each sub-agent applies its specialized evaluation criteria to the diff and produces a list of findings with file, line, severity, and a plain-language explanation
Synthesis — After all sub-agents complete, the root agent deduplicates overlapping findings, resolves conflicting assessments, assigns an overall risk score, and renders the final structured report

Invoking

bash

# Trigger a code review swarm on a pull request
POST /agents/code-review

{
  "repo": "owner/repository",
  "pr_number": 42,
  "dimensions": ["security", "performance", "style", "logic"]
}

# Retrieve the completed report
GET /agents/code-review/reports/:reportId

Review dimensions

Dimension	Focus area	Example findings
Security	Vulnerabilities, injection risks, secrets in code	SQL injection, hardcoded API key, insecure deserialization
Performance	Algorithmic complexity, unnecessary allocations, N+1 queries	O(n²) loop, missing index, unbounded fetch
Style	Naming conventions, formatting, idiomatic patterns	Non-standard casing, deep nesting, magic numbers
Logic	Correctness, edge cases, off-by-one errors, null handling	Missing null guard, incorrect boundary condition, unreachable branch
Test coverage	Uncovered paths, missing assertions, test quality	New branch without test, assertion on mutable state, empty test body

Output format

The synthesized report is returned as JSON and rendered as a structured document. Each finding includes:

Severity — critical, high, medium, low, or info
Dimension — Which sub-agent raised the finding
Location — File path and line range within the diff
Description — Plain-language explanation of the issue
Suggestion — Concrete recommended fix or refactor

The report also includes a top-level riskScore (0–100) and a recommendation field with one of: approve, approve_with_suggestions, or request_changes.

Configuration

yaml

codeReview:
  enabled: true
  maxAgents: 5               # Maximum parallel sub-agents
  dimensions:
    - security
    - performance
    - style
    - logic
    - testCoverage
  severityThreshold: medium  # Only surface findings at this severity or above
  postComment: true          # Automatically post report as a PR comment

ℹSet postComment: true together with a configured GitLab integration to have the swarm post its report directly as a pull request review comment, including inline annotations at the relevant diff lines.