Skip to content

CHANGELOG.md

Latest commit

 

History

History
63 lines (49 loc) · 5.03 KB

CHANGELOG.md

File metadata and controls

63 lines (49 loc) · 5.03 KB
Raw

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

1.0.0 - 2026-03-07

Added

  • Harness API — init(), run(), @agent() for zero-change observability, scoped budget runs, and decorated agent policy. Three modes: off, observe, enforce.
  • SDK auto-instrumentation — Patches OpenAI and Anthropic Python SDKs to intercept every LLM call for cost tracking, budget enforcement, compliance gating, and decision tracing.
  • Six-dimension optimization — Cost, latency, quality, budget, compliance, and energy tracked across every model call.
  • KPI-weighted routing — Inject business priorities (quality, cost, latency, energy) as weights into model selection decisions.
  • Compliance gating — GDPR, HIPAA, PCI, and strict model allowlists; block non-compliant models before execution.
  • Energy tracking — Deterministic compute-intensity coefficients for carbon-aware AI operations.
  • Decision traces — Full per-step audit trail: action, reason, model, cost, budget state, enforcement status.
  • Budget enforcement — Per-run and per-user budget caps with automatic stop actions when limits are exceeded.
  • Framework integrations — LangChain (Python + TypeScript), OpenAI Agents SDK, CrewAI, Google ADK, n8n, Vercel AI SDK.
  • TypeScript SDK — @cascadeflow/core, @cascadeflow/langchain, @cascadeflow/vercel-ai, @cascadeflow/ml, @cascadeflow/n8n-nodes-cascadeflow published on npm.
  • Proxy Gateway — Drop-in OpenAI/Anthropic-compatible HTTP server with mock and agent modes, streaming, tool calling, and embeddings support.
  • OpenClaw Server — Standalone OpenAI-compatible server for OpenClaw deployments with semantic routing.
  • Paygentic integration — Usage reporting and billing proxy for Paygentic platform.
  • Tool risk classification — ToolRiskClassifier for per-tool-call routing based on risk level.
  • Circuit breaker — Per-provider circuit breaker with configurable thresholds and recovery.
  • Dynamic configuration — Runtime config updates via file watcher with change events.
  • Rule engine — RuleEngine for declarative routing and policy rules.
  • Agent loops — Multi-turn tool execution with automatic tool call, result, re-prompt cycles.
  • Semantic quality validation — Optional ML-based quality scoring via FastEmbed embeddings.
  • 15-domain auto-detection — Code, math, medical, legal, finance, data, and more with per-domain routing pipelines.
  • Complexity detection — 500+ technical terms, mathematical notation detection, density-aware scoring for long documents.

Changed

  • Lazy imports — import cascadeflow no longer eagerly loads all providers, numpy, or heavyweight submodules. Import time reduced from ~1900ms to ~20ms via PEP 562 lazy loading.
  • __all__ reduced — From 127 to ~20 essential public symbols. Non-essential exports remain accessible but are not star-exported.
  • rich moved to optional — No longer a core dependency; falls back to stdlib logging when not installed. Install with pip install cascadeflow[rich].
  • Integration import errors — Failed optional integration imports now return proxy objects that raise ImportError with install hints on use, instead of silently returning None.
  • Proxy CORS default — cors_allow_origin changed from "*" to None (opt-in) for secure-by-default deployments.

Removed

  • Deprecated CascadeAgent parameters — config, tiers, workflows, enable_caching, cache_size, enable_callbacks removed from constructor. Use HarnessConfig or dedicated APIs instead.
  • Submodule __version__ strings — Removed from quality, streaming, telemetry, ml, tools, routing, interface subpackages. Use cascadeflow.__version__ instead.
  • Benchmark infrastructure — tests/benchmarks/, benchmark_results/, and related docs removed (moved to separate benchmark repo).

Fixed

  • Thread safety — Added threading.Lock around SDK patch/unpatch state. HarnessRunContext counters guarded with lock for multi-threaded use.
  • Trace buffer — _trace changed from list with manual slicing to collections.deque(maxlen=1000) for bounded memory.
  • Regex pre-compilation — ComplexityDetector now pre-compiles all regex patterns in __init__() instead of per-detect() call.
  • Proxy body limit — Added max_body_bytes (default 10MB) to ProxyConfig; returns 413 for oversized requests.
  • Proxy auth — Added optional auth_token to ProxyConfig; returns 401 for unauthenticated requests when set.

Security

  • Proxy gateway CORS tightened to opt-in (None default).
  • Request body size limit prevents memory exhaustion attacks.
  • Bearer token authentication for proxy gateway endpoints.
  • Updated SECURITY.md supported version to 1.0.x.