Skip to content

CHANGELOG.md

Latest commit

 

History

History
58 lines (40 loc) · 3.04 KB

CHANGELOG.md

File metadata and controls

58 lines (40 loc) · 3.04 KB

Changelog

v0.7.0

ContextFlow natively introduces the Engine Stabilization Layer, rectifying core caching invariants and standardizing external distribution hooks globally.

Added:

  • PluginRegistry.discover() enabling automated parsing of entry_points registered inside third-party application modules conforming to the contextflow.plugins.* format securely.

Fixed:

  • Semantic Caching Stability: NativeCache upgraded to dynamically encapsulate compressor properties during hash generation, structurally guaranteeing independent key indexing for deterministic text inputs traversing separate cognitive pipelines.

v0.6.0

ContextFlow natively introduces the Intelligence Layer to filter and compress context intelligently utilizing Semantic and Distillation logic.

Added:

  • SemanticMode capable of structurally dropping uncorrelated chat history via goal-similarity.
  • DistillationCompressor replacing rigid deterministic heuristics with natively awaited asynchronous LLM chunk summarization.
  • acompress() hook dynamically inserted into the base Compressor interface explicitly enabling intelligent LLM pipeline compression securely.

Refactored:

  • ContextPipeline.arun() and NativeCache.aget_or_set() natively await and cascade async implementations cleanly.
  • Provider interface corrected to strictly parse flat message arrays natively, standardizing Prompt injection.

v0.5.0

ContextFlow natively introduces Plugin Registries enabling infinite extensibility without modifying core architecture.

Added:

  • PluginRegistry foundational pattern inside core/registry.py.
  • ModeRegistry, CompressorRegistry, ProviderRegistry, SourceRegistry, and ScorerRegistry natively exporting decorators for dynamic logic insertion.
  • Exhaustive test suite (test_registry.py) covering type-safe constraints and instantiation mechanisms natively.

Refactored:

  • All built-in classes (MinimalMode, OpenAIProvider, NativeCache etc) register implicitly on import seamlessly.

v0.4.0

ContextFlow natively transitions from a stateless pipeline array into a stateful Context Engine.

Added:

  • ContextSession architectural wrapper capturing multi-turn integrations.
  • NativeCache generating cryptographic hashes to safely bypass redundant execution.
  • ContextRanker enabling mathematical arrays (via TimeDecayScorer).
  • Native debug tracing telemetry (pipeline = ContextPipeline(debug=True)).
  • Synthetic array benchmarks (benchmarks/benchmark.py).
  • Standalone execution loops (examples/agent_session.py).

Refactored:

  • Migrated flat Python module architecture into extensible nested sub-packages (mode/, compression/, provider/).
  • Safely isolated shared schemas and interfaces inside a protected core/ package.

Improved:

  • Async Provider execution pathways ensuring asynchronous graph compatibility.
  • TokenBudget enforcing structural retention limits using Priority-slicing metrics over standard deletion.
  • Deterministic Compression safety guarding LangGraph JSON nodes from scraper mutations exactly.