Skip to content

CHANGELOG.md

Latest commit

 

History

History
62 lines (49 loc) · 5.53 KB

CHANGELOG.md

File metadata and controls

62 lines (49 loc) · 5.53 KB

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.

[Unreleased]

Added

  • SDK Interceptors (@context-engineering/sdk-interceptors): Drop-in context management wrappers for OpenAI and Anthropic SDKs — intercept API calls and automatically pack messages within budget
  • Adaptive Learning (@context-engineering/adaptive): Observes which context items correlate with good model outputs and adjusts scoring weights over time via EMA-based feedback loops
  • Framework Middleware (@context-engineering/frameworks): Duck-typed middleware for LangChain, LlamaIndex, and CrewAI — zero framework dependencies
  • Context Debugger (@context-engineering/debugger): Diagnoses bad model outputs by tracing them to context quality problems (missing context, redundancy, stale items, budget waste)
  • Context-Aware RAG (@context-engineering/rag): Retrieves chunks based on information gain relative to existing context, not just similarity — supports hybrid vector + BM25 retrieval
  • Model Router (@context-engineering/router): Analyses context complexity across six dimensions and routes to the cheapest capable model, with adaptive learning and quality-based fallback
  • Context Replay: Record pack decisions and replay with different strategies for A/B testing
  • Context Inspector: Web UI for debugging context windows (internal)
  • Council of Experts (@context-engineering/council): Multi-model deliberation with 4 strategies (parallel, debate, stepladder, delphi), 8 role presets, Jaccard convergence detection
  • Adversarial Tester (@context-engineering/adversarial): Red-team context pipelines with 6 attack types (contradiction, noise-flood, subtle-error, authority-spoof, temporal-poison, relevance-dilution)
  • Context Time Travel (@context-engineering/time-travel): Git-like checkpoint/rewind/fork/merge for context states with 5 merge strategies (union, intersection, best-quality, highest-priority, manual)
  • Drift Detector (@context-engineering/drift): Continuous monitoring across 6 dimensions (relevance, redundancy, diversity, density, freshness, utilisation) with trend detection and alerting
  • Context Immune System (@context-engineering/immune): Learns from context failures via fingerprint-based similarity matching, generates antibodies to screen future packs
  • Context Compiler (@context-engineering/compiler): Declarative context programs with slot-based allocation, constraint validation, and per-model optimisation (Claude, GPT-5.4, Gemini 2.5)
  • Context Entanglement (@context-engineering/entangle): Multi-agent context sharing via scoped pub/sub mesh with propagation policies (immediate, next-pack, on-demand)
  • Python parity: All packages ported to Python with full API surface

Changed

  • Pipeline: Eliminated sync/async code duplication using the MaybeAsync pattern — build() and buildAsync() now share a single implementation
  • Provider adapters: Use properly typed SDK params (ChatCompletionCreateParamsNonStreaming, MessageCreateParamsNonStreaming) instead of Record<string, unknown> casts — fixes type checking against latest OpenAI/Anthropic SDK versions
  • Python context_framework: Domain runtime modules now use lazy imports via __getattr__ — importing the package no longer eagerly loads all 12 runtime modules

Fixed

  • Type errors in ce-providers caused by OpenAI/Anthropic SDK union return types (Stream | ChatCompletion)
  • Harmonised vitest versions across all packages (^4.1.0)

0.1.0 - 2026-02-27

Added

  • Core: pack(), tracePack(), diff(), estimateTokens(), createContextItem(), createScorer() — the six core functions
  • Cache topology: packWithCacheTopology() for prefix cache optimisation with static/session/request volatility classification
  • Allocation: packWithAllocation() for kind-aware budget splits with min/max/target constraints
  • Sessions: createSession() for stateful context with differential tracking
  • Pipeline: Fluent builder chaining .add().allocate().cacheTopology().place().qualityGate().build()
  • Placement: placeItems() with attention profile-aware ordering (claude, gpt4, uniform)
  • Quality: analyzeContext() returning density, diversity, freshness, redundancy scores
  • Cost: estimateCost() and projectCosts() with prefix cache savings for Claude and GPT models
  • BEADS handoff: createHandoff() / pickupHandoff() for agent-to-agent context serialisation
  • Compaction: createContextManager() for auto-summarisation across conversation turns
  • Stream: packStream() async generator variant
  • Memory stores (@context-engineering/memory): InMemory, File (JSONL with atomic writes), SQLite
  • Provider adapters (@context-engineering/providers): OpenAI (tiktoken) and Anthropic token estimators
  • CLI (@context-engineering/cli): 11 commands — pack, trace, diff, budget, lint, place, quality, effective-budget, handoff, pickup, cost
  • Python SDK: Full API parity with TypeScript plus advanced features (negation/supersession, hierarchical inclusion, semantic redundancy, AgentContextManager, segmenters)
  • Error hierarchy: ValidationError, BudgetExceededError, EstimationError with structured details
  • Shared JSON Schemas for cross-language validation