yvgude/lean-ctx v2.12.7

v2.12.7 — MCP Tool Enforcement + Token Optimization + 19 IDE Coverage

on GitHub

Highlights

MCP Tool Enforcement — LLMs now reliably use lean-ctx MCP tools instead of native tools across all 19 supported IDEs/agents through a 3-layer strategy:

1. PreToolUse Hooks — Soft-redirect hooks for Claude Code, Cursor, and Gemini CLI intercept native Read, Grep, ListFiles calls and guide LLMs to ctx_read/ctx_search/ctx_tree. Safety fallback ensures native tools work if MCP server is unreachable.
2. LITM-Optimized Rules v5 — Critical instructions placed at beginning AND end of rules blocks (exploiting LLM attention patterns).
3. Project-Level Reinforcement — Auto-creates AGENTS.md and .cursorrules on lean-ctx init --agent.

Token Optimization Deep Dive — Performance improvements across all phases:

• Config caching with OnceLock, stats batching, deferred writes
• 20+ new compression patterns (cargo, git, docker, npm, uv/conda/poetry)
• Trimmed tool descriptions, strengthened CRP/TDD output constraints

Added

• 6 new IDE/agent rules targets: Qwen Code, Trae, Amazon Q Developer, JetBrains, Antigravity, Pi Coding Agent
• 19 total supported IDEs/agents (up from 13)
• Auto-refresh hooks + rules on lean-ctx update and MCP server start
• Complete uninstall cleanup for rules files (13 locations), hook scripts (7 files), and configs

Changed

• Removed double savings footer, compressed entropy/stale headers
• Stats batching with in-memory accumulator + periodic flush
• Fixed npm test heuristic (proper Jest/Vitest parsing)
• Hook installation refactored for reusability

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

Full Changelog: v2.12.6...v2.12.7