DocsChangelog
Changelog

What's new

Latest updates to the Aurex platform. Subscribe to releases on GitHub.

2026-07-18v0.4.0Platform

AI Runtime Layer & OpenAI-style Developer Documentation

Reframed Aurex around three core runtime capabilities - Behavioral Awareness, Adaptive Routing, and Provider Continuity. Streamlined the entire documentation suite to match OpenAI developer documentation standards.

  • •Three core runtime capabilities: Behavioral Awareness, Adaptive Routing, and Provider Continuity
  • •Interactive vertical Live Request Trace visualization in the hero section
  • •Complete rewrite of 29 documentation guides to OpenAI-style single-topic format
  • •Guaranteed under 2ms in-process decision overhead with zero network hop penalty
2026-06-04v0.3.0SDK

Behavioral Awareness - Hybrid Loop Detection (F1 0.72)

Upgraded loop detection from a structural-only heuristic (F1 ≈ 0.08) to hybrid structural + semantic detection (F1 ≈ 0.72). The new implementation adds three detection layers: exact hash matching for identical tool calls, context growth threshold for rapid expansion, and semantic fuzzy detection using cosine similarity for rephrased repetition. Based on techniques from arXiv:2511.10650.

  • •3-layer detection: exact hash, context growth, semantic similarity
  • •30-second cooldown after loop detection to prevent double-tripping
  • •Per-session mutable state - never persisted, no privacy impact
  • •Backward compatible - existing structural checks preserved
2026-06-04v0.3.0SDK

patch_all() multi-provider support

The existing patch_all() method now auto-detects and patches all installed LLM provider libraries with a single call. Previously only OpenAI was reliably auto-patched. Now supports OpenAI, Anthropic, Google Generative AI, LiteLLM, Ollama, and Hugging Face out of the box.

  • •Idempotent - safe to call multiple times
  • •Fail-open - patching failures log warnings but never break LLM calls
  • •Returns dict of {provider: was_patched} for introspection
  • •patch_openai() preserved as backward-compatible alias
2026-06-04v0.3.0SDK

Prompt optimization - caching opportunity detection

New prompt optimization signal detects calls where the system prompt is identical across sessions (matched by token count + SHA-256 hash) and the provider supports prompt caching but no cache header was sent. Calculates estimated monthly savings at the provider's cached rate.

  • •Supports Anthropic (90% savings), OpenAI (50% savings), Google (75% savings)
  • •Only flags prompts appearing 5+ times in 30 days
  • •Ignores trivial savings under $0.10/month
  • •Two additional signals: prompt drift tracking and CoT inflation detection

Command Palette

Search for a command to run...