yvgude/lean-ctx v2.14.0

v2.14.0 -- Intelligence Layer Architecture

on GitHub

Intelligence Layer Architecture

lean-ctx transforms from a pure compressor into an Intelligence Layer between user, AI tool, and LLM. Based on empirical attention analysis (L-curve discovery), all output formats, filters, and encoding strategies are now calibrated to how LLMs actually process context.

Added

• ctx_preload MCP tool — Proactive context orchestration based on task description + import graph. Analyzes project files, extracts task-critical lines, key signatures, and imports, then delivers a compact L-curve-optimized context snapshot
• L-Curve Context Reorder Engine — Classifies lines into 7 categories and reorders output so the most task-relevant lines occupy position 0 (where LLMs allocate 20x more attention)

Changed

• Output-Format Reordering — ctx_read output now places file content first, metadata last. Task-critical information at position 0.0
• IB-Filter 2.0 — Upgraded with empirical L-curve attention weights, score-descending output, error-handling prioritization
• LLM-Native Encoding — 15+ new token optimization rules: generic simplification, lifetime elision, path shortening, brace collapsing
• System-Prompt Cleanup — Removed duplicate decoder block (~200 wasted tokens), consolidated redundant warnings, ASCII symbols replace multi-token Unicode
• CRP/TDD Response Shaping — Dynamic response budgets by complexity

Fixed

• Shell hook compression broken — exec_buffered() bypassed compression when stdout was piped (always the case for agent hooks)
• Shell hook stats lost — std::process::exit() skipped stats flush, discarding all CLI stats silently

Full changelog: https://github.com/yvgude/lean-ctx/blob/main/CHANGELOG.md

Full Changelog: v2.13.1...v2.14.0