Scientific Compression Evolution
Six algorithms from information theory, graph theory, and statistical mechanics now power lean-ctx's compression pipeline — all automatic, all local, zero configuration.
New Algorithms
| Algorithm | What it decides | Based on |
|---|---|---|
| Predictive Surprise | Which lines to keep | BPE cross-entropy (Zipfian prior) |
| Spectral Relevance | Which files matter | Heat diffusion + PageRank on dependency graph |
| Boltzmann Allocation | Token budget per file | Statistical mechanics temperature from task specificity |
| Semantic Chunking | How to order output | AST boundaries + attention bridges + tail anchors |
| MMR Deduplication | Redundant lines | Bigram Jaccard similarity (Maximum Marginal Relevance) |
| BPE Optimization | Final text encoding | Token-boundary-aligned string replacements |
Also in this release
- Auto-build graph index —
load_or_build()builds the project graph automatically on first use. No manualctx_graph buildneeded. - Fish shell doctor check —
lean-ctx doctornow detects aliases in~/.config/fish/config.fish - Codex hook refresh —
lean-ctx updatenow refreshes Codex PreToolUse hooks alongside Claude/Cursor/Gemini
Benchmark Results
Session simulation (30-min coding):
Raw (no compression): 449.2K tokens $1.12
lean-ctx (with CCP): 64.5K tokens $0.16 85.7% savings
Mode performance:
map: 96.0% savings, 94.4% quality
signatures: 96.0% savings, 88.1% quality
cache re-read: 99.3% savings (13 tokens)
git commands: 95.0% savings across all patterns
Scientific verification: 12/12 checks passed
Intensive benchmarks: 39/39 tests passed
Install / Update
# New install
curl -fsSL https://leanctx.com/install.sh | sh
lean-ctx setup
# Update existing
lean-ctx updateFull changelog: https://github.com/yvgude/lean-ctx/blob/main/CHANGELOG.md
Full Changelog: v2.14.5...v2.15.0