neuron-explorer
Qwen3.5‑4B 전용 조향형 뉴런 탐색기, 모든 MLP 뉴런을 밀어 결과를 보관·탐색
개인 계정의 공개 저장소 소유자는 GitHub로 로그인해 운영자를 확인하고 소개를 직접 관리할 수 있습니다. 관리자 승인은 필요하지 않습니다.
로그인한 뒤 확인 파일을 만들어 이 저장소의 기본 브랜치에 30분 안에 커밋해 주세요. 다른 브랜치나 포크의 파일은 인정하지 않습니다. 파일을 올린 뒤 아래에서 확인하면 관리자 승인 없이 관리할 수 있습니다. 자격을 유지하려면 파일을 남겨 두세요.
로그인이 필요하면 로그인 화면으로 이동합니다. 로그인 후 이곳으로 돌아와 원하는 확인 버튼을 다시 눌러 주세요.
프로젝트 폴더에서 /nomorevibe verify를 실행하면 소유권을 확인하고 소개를 직접 관리할 수 있습니다.
1분 만에 확인하기Lightning Code Index — sub-millisecond semantic code search with an MCP server for AI assistants. 13 languages, CLI + HTTP + Unix socket.
저장소 README에서 가져옴
README 전체를 불러오는 중입니다. 아래는 저장된 발췌입니다.
lci aims to make semantic code search fast enough for interactive use and
to cut the context an AI assistant has to read compared with grep. Tree-sitter
parsing across 13 languages. An MCP server for AI assistants, plus CLI, HTTP,
and Unix-socket interfaces.
Version tracks the VERSION field of project(lci ...) in CMakeLists.txt (currently 0.10.1); check lci --version for what you have installed.README 전문 보기 ↗lci aims to make semantic code search fast enough for interactive use and to cut the context an AI assistant has to read compared with grep. Tree-sitter parsing across 13 languages. An MCP server for AI assistants, plus CLI, HTTP, and Unix-socket interfaces.
Version tracks the VERSION field of project(lci ...) in CMakeLists.txt (currently 0.10.1); check lci --version for what you have installed.
Platform availability: prebuilt releases ship for Linux x86_64 (
.tar.gz,.deb,.rpm), Windows x86_64 (.tar.gz), and macOS arm64 / Apple Silicon (.tar.gz). Intel (x86_64) macs are not yet served by a prebuilt binary — build from source on those. The installers fail fast with a clear message on any unsupported platform/arch.
# Linux / macOS
curl -fsSL https://raw.githubusercontent.com/standardbeagle/lci-cpp/main/install.sh | sh
# Windows (PowerShell)
irm https://raw.githubusercontent.com/standardbeagle/lci-cpp/main/install.ps1 | iex
Both detect your OS/arch, download the matching release binary, and install
it (/usr/local/bin or ~/.local/bin on Unix; %LOCALAPPDATA%\Programs\lci
on Windows). Override with LCI_PREFIX; pin a version with LCI_VERSION=0.10.1.
lci update # self-update to the latest release
lci update --check # report current vs latest without installing
lci update --version 0.10.1 # install a specific release
lci update works regardless of how lci was installed.
# Linux: tarball or Debian package
curl -fLO https://github.com/standardbeagle/lci-cpp/releases/download/v0.10.1/lci-0.10.1-Linux.tar.gz
tar -xzf lci-0.10.1-Linux.tar.gz
sudo install lci-0.10.1-Linux/bin/lci /usr/local/bin/
# or
sudo dpkg -i lci-0.10.1-Linux.deb
lci --version # 0.10.1
lci search "myFunction"
# Toolchain: GCC 13+ or Clang 17+, CMake 3.25+, vcpkg, Ninja
cmake --preset release \
-DCMAKE_TOOLCHAIN_FILE=$VCPKG_ROOT/scripts/buildsystems/vcpkg.cmake
cmake --build build/release --parallel
./build/release/src/lci --version
# Install + package
cd build/release
sudo cmake --install .
cpack -G TGZ # produces lci-0.10.1-Linux.tar.gz
cpack -G DEB # produces lci-0.10.1-Linux.deb
The CPack output contains only bin/lci. Tree-sitter grammars are
statically linked; the binary dynamically links libc, libstdc++, libssl
(via libcrypto), zlib, and brotli — the Debian package declares only
libc6 (>= 2.35) since the rest ship on modern Linux by default.
| Surface | Use |
|---|---|
lci search PATTERN | substring + symbol-aware ranked search across the corpus |
lci grep PATTERN | literal text search over the corpus, grep-style output |
lci tree SYMBOL | function call hierarchy with annotations |
lci def SYMBOL / lci refs SYMBOL | definition + reference lookup |
lci git-analyze | analyze working-set changes for duplicates, naming, complexity |
lci mcp | MCP (Model Context Protocol) server over stdio for AI assistants |
lci server | long-running HTTP server over a Unix socket; per-corpus daemon |
lci servers | list every index server you are running, and the root each serves |
lci shutdown [--all] | stop the server for this root, or every one of them |
lci status | index health + runtime metrics |
Full CLI reference: lci --help or lci <command> --help.
lci mcp exposes 15 tools to AI assistants: search, find_files,
get_context, context, list_symbols, inspect_symbol, browse_file,
callers, semantic_annotations, side_effects, code_insight,
git_analysis, index_stats, debug_info, and info. code_insight
emits its report in LCF (LCI Compact Format), a token-dense, agent-oriented
layout; the other tools return structured JSON.
code_insight is the session-startup workhorse: repository map, health
dashboard, entry points, complexity/coupling/cohesion statistics, module
and feature breakdowns, naming vocabulary, and git change/hotspot
analysis.
Full per-tool reference — parameters, modes, output shapes, errors — is in
docs/TOOLS.md.
Any command that needs an index auto-starts a background server for that root
and leaves it running, because re-indexing per command would dominate every
query. Left unbounded that leaks: a caller that indexes a temp directory and
deletes it strands a multi-GB server per root (175 were once found on one
host). Three policies bound it, all under a server block in .lci.kdl:
| Setting | Default | Effect |
|---|---|---|
idle_timeout_sec | 1800 | exit after this long with no real request (/ping doesn't count); the next command respawns |
max_instances | 8 | a starting server asks the least-recently-active servers beyond this cap to stop |
max_rss_mb | 4096 | self-cap; the server sheds cache and, if still over, exits rather than grow past this RSS (0 disables) |
| — | always on | exit within ~500 ms of the indexed root being deleted |
lci servers shows what is running and which root each one serves — the socket
name hashes the root, so the filesystem alone cannot tell you. lci shutdown
stops the server for the current root; lci shutdown --all stops every one of
them.
Both read the instance registry, which only standalone servers publish into.
The index server embedded in lci mcp stays out of it deliberately: it belongs
to the MCP session, and evicting it would break that session while leaving the
process alive. Stop those by ending the MCP client. Servers from before 0.8.0
predate the registry entirely — they answer neither command, and must be killed
by signal.
.claude/rules/karpathy-principles.md
for the performance discipline.13 languages via vendored tree-sitter grammars: Go, Python, JavaScript, TypeScript (incl. TSX), Rust, C, C++, Java, C#, PHP, Kotlin, Zig, and Ruby. Cross-file import resolution is supported for Go, JavaScript, Python, Rust, C#, PHP, Zig, and C/C++.
cmake --build build/release --parallel
ctest --preset release --parallel $(nproc)
The preset excludes the separately gated benchmark target. Run
./build/release/tests/lci_benchmarks explicitly for performance checks.
CI gates the synthetic suite against
tests/benchmarks/baseline/linux-x64.json via scripts/bench_gate.py
(fails on >1.5x regression); regenerate the baseline on the benchmark
runner with scripts/bench_gate.py --update when benchmarks change.
The suite covers unit tests, integration tests (bundled per-binary for in-process cache amortization), real-project tests (chi, fastapi, trpc, pocketbase), regression snapshot tests, and google-benchmark performance gates.
Real-project tests fetch corpora via tests/parity/corpora/prep_real.sh
and skip cleanly when corpora are absent.
src/ binary + library
include/lci/ public-ish headers
tests/
integration/ bundled per-process integration suite
benchmarks/ google-benchmark perf gates
cmake/ custom CMake modules
.github/workflows/ CI + release
| Surface | State |
|---|---|
| C++ binary install | TGZ / DEB / RPM via CPack |
| MCP server | 15 tools, all engine-backed |
| Performance | synthetic benchmark suite gated in CI (fails on >1.5x regression vs stored baseline); real-project benches run locally |
MIT. See LICENSE.