v0.3.0

What's new

• OpenTelemetry metrics: Prometheus metrics endpoint on :9464/metrics, enabled by default (disable with --no-otel)
• Grafana + Prometheus monitoring stack: docker compose up -d in monitoring/ for a pre-built dashboard with token savings, request rates, compression ratio, pipeline latency, and per-transform breakdowns
• OTLP trace export: traces exported to localhost:4317 for Jaeger/Tempo if available

v0.2.0

What's new

• Per-request transform control: Send X-Kompact-Disable: toon,code_compressor header to disable specific transforms for a single request without affecting other clients
• Updated README with SDK examples and cost savings table

v0.1.0

Initial release of Kompact — multi-layer context optimization proxy for LLM agents.

Features

• 8 transforms: TOON, JSON Crusher, Code Compressor, Log Compressor, Content Compressor (TF-IDF extractive), Schema Optimizer (TF-IDF tool selection), Observation Masker, Cache Aligner
• Transparent HTTP proxy: works with any Anthropic/OpenAI-compatible agent — just change base_url
• Adaptive pipeline: scales compression parameters based on context size
• Artifact index: tracks compressed content for re-fetching
• Live dashboard: real-time metrics at /dashboard
• Benchmark suite: context-bench integration with NIAH, answer recall, effective ratio, cost-of-pass metrics

Benchmarks

47.7% compression with 100% needle preservation on search-heavy scenarios. See README for full comparison against Headroom and LLMLingua-2.

Install

pip install kompact