AI Engines Overview

Codemetry can optionally use AI to generate human-readable explanations of analysis results. AI summarization is entirely opt-in and privacy-conscious.

Key Principles

Opt-In Only

AI features are disabled by default. You must explicitly enable them:

php artisan codemetry:analyze --days=7 --ai=1

Or in configuration:

'ai' => [
    'enabled' => true,
    // ...
],

Metrics Only

AI engines never receive code or diffs. They only receive:

• Aggregated metrics (churn, scatter, fix counts)
• Normalized values (z-scores, percentiles)
• Computed scores and labels
• Confounders (flags like large_refactor_suspected)

This design ensures your source code stays private.

Graceful Degradation

If AI is enabled but unavailable:

• Analysis continues normally
• Heuristic scores are unaffected
• ai_unavailable confounder is added
• No AI summary in output

Supported Engines

OpenAI

GPT-6 Luna (default), Sol, Astra. Best balance of quality and cost.

Anthropic

Claude Sonnet 5.5 (default), Haiku 5.5, Opus 5.5. Excellent reasoning quality.

DeepSeek

DeepSeek Flash (default), V4 Pro. Cost-effective alternative.

Google

Gemini 3.8 Flash (default), 3.5 Flash-Lite, 4 Argon. Google Cloud integration.

Quick Setup

1. Choose an Engine

# Set in environment
export CODEMETRY_AI_ENGINE=openai  # or: anthropic, deepseek, google

2. Provide API Key

# Engine-specific key
export OPENAI_API_KEY=sk-...
# Or unified key
export CODEMETRY_AI_API_KEY=sk-...

3. Run with AI

php artisan codemetry:analyze --days=7 --ai=1

Configuration Options

config/codemetry.php

'ai' => [
    'enabled' => false,  // Set true to enable by default
    'engine' => env('CODEMETRY_AI_ENGINE', 'openai'),
    'model' => env('CODEMETRY_AI_MODEL', null),  // null = engine default
    'api_key' => env('CODEMETRY_AI_API_KEY', null),
    'batch_size' => env('CODEMETRY_AI_BATCH_SIZE', 10),  // Days per API call
],

Option	Description
enabled	Enable AI by default (can override with --ai=0/1)
engine	Which AI provider to use
model	Specific model (null = use engine’s default)
api_key	API key (falls back to engine-specific env vars)
batch_size	Number of days to process per API call (default: 10)

What AI Adds

When AI is enabled, each analysis window includes:

{
  "ai_summary": {
    "explanation_bullets": [
      "High churn (95th percentile) indicates major development activity.",
      "The presence of reverts suggests some changes needed rollback.",
      "Despite the bad score, the large_refactor_suspected flag hints at planned restructuring.",
      "Consider reviewing commit history to confirm this was intentional work."
    ],
    "score_delta": 0,
    "confidence_delta": 0.0,
    "label_override": null
  }
}

explanation_bullets

Human-readable insights about the day’s metrics. Useful for:

• Team reports
• Executive summaries
• Understanding why a day scored poorly

score_delta / confidence_delta

AI can suggest score adjustments (V1: always 0). Future versions may allow AI to fine-tune scores based on context.

label_override

AI can suggest overriding the label (V1: always null). Example: changing “bad” to “medium” if metrics suggest false positive.

Note

In V1, AI provides explanations only. Score modifications are reserved for future versions with more sophisticated prompts.

Batch Processing

Codemetry batches multiple days into a single API call for efficiency. By default, 10 days are processed per request.

'ai' => [
    'batch_size' => 10,  // Process 10 days per API call
],

Benefits:

• Fewer API calls: A 30-day analysis requires only 3 API calls instead of 30
• Lower latency: Fewer round-trips to the AI provider
• Cost efficiency: Batch requests are more token-efficient

You can adjust batch_size based on your needs:

• Higher values (15-20): Fewer API calls, but larger request/response sizes
• Lower values (1-5): More granular error handling, but more API calls

Note

If a batch fails, Codemetry marks remaining days with the ai_unavailable confounder and continues without AI for those days.

Cost Considerations

AI costs are per API call, and batching significantly reduces total calls. Codemetry’s requests are small: roughly 500-1000 input tokens per day analyzed and 200-500 output tokens per summary.

Engine	Default model	Cheaper option	Pricing
OpenAI	gpt-6-luna	—	openai.com/api/pricing
Anthropic	claude-sonnet-5-5	claude-haiku-5-5	Anthropic pricing
DeepSeek	deepseek-flash	—	DeepSeek pricing
Google	gemini-3.8-flash	gemini-3.5-flash-lite	Gemini API pricing

Tip

Use lighter models (gpt-6-luna, claude-haiku-5-5, gemini-3.5-flash-lite, deepseek-flash) for routine analyses. Reserve heavier models for deep investigations.

Without AI

Codemetry is fully functional without AI:

• Heuristic scores are deterministic and reliable
• Reasons explain score contributors
• Confounders provide context flags
• JSON output includes all raw data

AI adds explanatory convenience, not core functionality.