Mode A

Single-Shot Structured Reviewer

One Responses API call with Structured Outputs emits {file, line_range, comment, severity}. No tools. Great for 80% of routine PRs.

Lowest

Cost

~2-5s

Latency

Good

Quality

80%

Coverage

How It Works

PR Webhook

Context Pack

Single LLM Call

JSON Output

Post Comments

Structured Output Schema

Deterministic JSON response via Structured Outputs

{
  "findings": [
    {
      "file": "string",
      "line_range": [number, number],
      "severity": "critical" | "warning" | "info",
      "title": "string",
      "comment": "string",
      "confidence": number
    }
  ]
}

Key Implementation

Core configuration for single-shot reviews

Model

gpt-4.1-mini or gpt-4.1

Response Format

response_format: json_schema

Tools

None (read-only)

Trade-offs

Advantages

Lowest cost per PR review
Fastest time-to-comment (single round-trip)
Deterministic output via Structured Outputs
Simple to implement and maintain
Scales linearly with throughput

Limitations

No tool execution for verification
May miss issues requiring deeper context
Fixed depth - no escalation path
Best suited for routine, low-risk PRs

Best For

Routine feature PRs with small diffsStyle and formatting checksBasic code smell detectionDocumentation updatesDependency version bumps

Agentic Scale