Changelog
v2.8.0April 15, 2025

FORG Atlas dashboard and AI insights

The intelligence layer arrives. FORG Atlas lets you query your usage data in plain English. The AI Insights tab surfaces waste patterns automatically. Model recommendations in beta.


New

FORG Atlas alpha: natural language usage queries

FORG Atlas is now available in alpha for Business+ plans. Query your usage data in plain English: 'Which developer has the highest cost per session?' or 'What is our cache hit rate trend over the last 30 days?' Powered by pgvector semantic search over nightly signal chunk embeddings, answered via Claude Sonnet. See the FORG Atlas documentation for example queries and limitations.

New

AI Insights dashboard tab with automated waste detection

The new Optimize tab in the dashboard surfaces automatically detected waste patterns: idle sessions, oversized model usage, low cache hit rates, and duplicate API calls. Each pattern includes an estimated monthly cost impact and a one-click link to create the relevant rule. No configuration required — waste detection runs continuously on your signal data.

New

Model recommendation engine beta

For teams with 30+ days of data, the model recommendation engine analyzes your usage patterns and suggests model substitutions based on task type and cost. Beta feature — recommendations require manual review before acting on them. Available under Optimize → Model Recommendations.

Improved

Signal ingestion throughput increased to 10,000/sec per Worker

The Rule Engine Worker now batches incoming signals and processes them in parallel using Cloudflare Durable Objects for coordination. Peak throughput increased from ~3,000 signals/sec to ~10,000 signals/sec per Worker instance. This headroom supports large Enterprise customers with many simultaneous active developers.

Fixed

Cache hit rate calculation now excludes cache_write tokens

The cache efficiency metric in the dashboard was incorrectly including cache_write tokens in the denominator, which underreported cache hit rates for sessions that were establishing the cache for the first time. The corrected formula: cache_hit_rate = cache_read / (input_tokens - cache_write). Teams will see their reported cache hit rates increase slightly after this fix.