github-featured / yvgude--lean-ctx

发布

0 发布 257 Git 标签

RSS 订阅源
  • v2.6.0 65b3580222
    比较

    release: v2.6.0 — Scientific optimization: adaptive entropy, attention model, task relevance, codebook, feedback loop
    Release / Build aarch64-apple-darwin (push) Has been cancelled
    详情
    Release / Build x86_64-apple-darwin (push) Has been cancelled
    详情
    Release / Build x86_64-unknown-linux-musl (push) Has been cancelled
    详情
    Release / Build aarch64-unknown-linux-gnu (push) Has been cancelled
    详情
    Release / Build x86_64-unknown-linux-gnu (push) Has been cancelled
    详情
    Release / Build x86_64-pc-windows-msvc (push) Has been cancelled
    详情
    Release / Create Release (push) Has been cancelled
    详情

    frostbyte_neo 发布于 2026-03-27 20:29:46 +00:00

    New core modules:

    • adaptive_thresholds: per-language BPE entropy thresholds with Kolmogorov complexity adjustment and feedback-loop learning
    • task_relevance: BFS-based relevance scoring through project dependency graph with keyword matching and Information Bottleneck filter
    • attention_model: heuristic U-shaped attention prediction (alpha/beta/gamma) with structural importance scoring
    • codebook: cross-file TF-IDF codebook for semantic deduplication with cosine similarity proxy
    • feedback: compression outcome tracking with threshold learning across sessions

    New MCP tool:

    • ctx_overview: multi-resolution project map with task-conditioned relevance scoring

    Integration:

    • entropy mode uses adaptive per-language thresholds (Rust 0.85, Python 1.2, JSON 0.6, etc.)
    • ctx_dedup includes TF-IDF cosine similarity analysis for semantic duplicate detection
    • LITM module uses content-aware attention efficiency via structural importance analysis
    • CEP includes output token budget guidance (Mechanical: 50 tok, Standard: 200, Architectural: full)
    • System prompt prefix-cache aligned (stable instructions before variable session state)
    • Feedback loop feeds learned thresholds back into adaptive compression

    Made-with: Cursor

    下载附件