Changelog#

All notable changes to WatchLLM are documented here.

[Unreleased]#

Added#

• Agent Debugger for tracing multi-step AI agent workflows
• GDPR compliance tools (data export, account deletion)
• Enhanced A/B testing with statistical significance tracking

[1.5.0] — 2026-01-15#

Added#

• BYOK (Bring Your Own Key) support for OpenAI, Anthropic, and Groq
• Direct provider access with lower latency
• AES-256 encryption for stored provider keys
• Model name auto-mapping between OpenRouter and native formats
• Per-project provider key management in dashboard

Changed#

• Improved semantic similarity threshold tuning UI
• Dashboard analytics now show BYOK vs OpenRouter traffic split
• Updated SDK documentation for BYOK integration

Fixed#

• Cache hit rate calculation for streaming responses
• Rate limit header accuracy on edge cases

[1.4.0] — 2025-11-20#

Added#

• Semantic A/B Testing — compare models and providers in real-time
• Traffic split configuration with percentage-based routing
• Variant performance comparison dashboard
• Statistical significance indicators

Changed#

• Analytics dashboard redesigned with improved visualizations
• Request log filtering now supports date ranges and model filters
• Improved error messages for invalid API key formats

Fixed#

• Streaming response interruption on long completions
• Dashboard timezone handling for usage charts

[1.3.0] — 2025-09-10#

Added#

• Python SDK with auto-instrumentation support
• Node.js SDK with TypeScript support
• SDK auto-detection of WatchLLM proxy
• Comprehensive SDK documentation

Changed#

• API key format updated to lgw_proj_ prefix for clarity
• Improved cache key generation for better hit rates
• Dashboard performance optimizations

Fixed#

• Memory leak in long-running streaming connections
• CORS headers missing on error responses

[1.2.0] — 2025-07-05#

Added#

• Usage Analytics Dashboard with real-time metrics
• Cost savings estimation and tracking
• Cache hit rate visualization
• Request volume charts with hourly/daily/monthly views
• Export analytics data as CSV

Changed#

• Improved prompt normalization algorithm
• Better handling of function calling / tool use in caching
• Updated rate limiting to sliding window algorithm

Fixed#

• Duplicate cache entries for concurrent identical requests
• Dashboard auth session expiration handling

[1.1.0] — 2025-05-01#

Added#

• Semantic Caching — cache hits based on meaning, not just exact match
• Configurable similarity threshold per project
• Prompt normalization pipeline (lowercase, filler words, whitespace)
• X-WatchLLM-Cache-Similarity response header

Changed#

• Cache key algorithm upgraded from MD5 to SHA-256
• Improved streaming response handling
• Dashboard UI refresh with dark theme

Fixed#

• Rate limit counter reset timing issue
• API key validation for special characters

[1.0.0] — 2025-03-15#

Added#

• Initial release of WatchLLM
• OpenAI-compatible proxy with semantic caching
• Cloudflare Workers edge deployment
• Supabase-backed project and API key management
• Redis-based exact match caching
• Basic usage logging and analytics
• Multi-project support
• Stripe payment integration
• Dashboard with project management
• Rate limiting (per-key and per-IP)
• Support for chat completions, text completions, and embeddings

Format follows Keep a Changelog and Semantic Versioning.