senoldogann/llm-context-manager
Developer toolsこのページは AI により自動翻訳されています。技術的な詳細は原文をご確認ください。 原文を見る
Bridge the gap between your codebase and your AI editor. CCM transforms static source code into a dynamic, queryable Knowledge Graph, enabling AI agents to navigate, understand, and reason about your...
概要
Bridge the gap between your codebase and your AI editor. CCM transforms static source code into a dynamic, queryable Knowledge Graph, enabling AI agents to navigate, understand, and reason about your project with surgical precision. An external benchmark on real repositories (serde, flask, express) with 35 hand-verified golden tasks, honest Recall@K/MRR/latency metrics, and hybrid scoring at 82.9% vs 80.0% semantic-only — plus the v0.3.12 durable background semantic upgrade and the 180/180 offline semantic gate. Modern AI coding assistants (Claude, Cursor, Windsurf) are powerful but suffer from : CCM turns raw code into a queryable graph so agents answer connected logic questions instead of guessing from similar words. Unlike tools that dump raw code, CCM injects : - - Explains why code was retrieved - - Maps how files talk to each other - - Shows certainty in results Many AI-context tools stop at semantic search: they embed files and return chunks that merely look similar.
README
以下は原文 README の抜粋です。原文を見る。
Cognitive Codebase Matrix (CCM)
English | Turkce
A code graph for AI coding agents that keeps up with your edits — and tells you when it is not watching.
CCM parses your project with tree-sitter (13 languages), keeps a call and import graph plus an optional semantic index on disk, and serves them to agents such as Claude Code, Codex and Cursor through MCP. The recommended agent entry point is
context: give it a symbol,path:line, or a natural- language coding task and CCM chooses structural graph lookup or hybrid search server-side. On Django 5.1 (6,670 files) a saved change shows up in graph results after a median of about 0.6 s (measurement: release build, embedder off). Query responses start with the index state (fresh,stale · …while a registered change is pending, orauto-refresh off · indexed Xs ago), and a failed re-index never replaces the last good graph. In the pre-registered L1 freshness benchmark (one-function edits in Flask 3.0.3 and Django 5.1, probes at fixed times), every graph answer from 0.25 s after a save on matched the edit; in the sub-250 ms window before the watcher applies the event, 6 of 126 probes returned the pre-edit answer still labeledfresh, and 4 more a partial or empty one (benchmarks/).
Status: search quality has a 35-task pilot benchmark. In an LLM-free benchmark of 24 fixed questions on Flask and Django, the 0.4.0 tools answer with 74% fewer response bytes and a third fewer calls than the 0.3 tools (token cost). A 2026-10-02 Claude Code pilot then showed the product-level gap: on three small/medium tasks the agent never called CCM, while merely exposing CCM increased input tokens. The active work therefore moves tool selection into the new one-call
contextentry point. That change has not yet been shown to reduce end-to-end agent tokens; the pilot and its negative result remain published underbenchmarks/results/agent/pilot-claude/.
Why CCM?
Coding agents find code well with grep. What grep does not give them is
推奨ツール
別のキーワードを試すか、フィルタを外してください。
インストール
npx skillfish add senoldogann/llm-context-manager