neuron-explorer
Qwen3.5‑4B 전용 조향형 뉴런 탐색기, 모든 MLP 뉴런을 밀어 결과를 보관·탐색
코딩 에이전트를 위한 로컬 우선 저장소 컨텍스트: 코드그래프·인용 프로젝트 지식·변경 영향·증거 기반 워크플로우 제공
Local-first repository context for coding agents: codegraph, cited project knowledge, change impact, and evidence-backed workflows.
개발 도구Rust · PythonGitHub @xcrft직접 설치운영자 미확인
써 보셨어요? 한 번만 눌러 주세요. 로그인은 필요 없어요.
README 전체를 불러오는 중입니다. 아래는 저장된 발췌입니다.
Repository context and evidence-backed workflows for coding agents.
Mastermind gives coding agents local code and documentation search, project
knowledge, optional personal preferences, and a controller that checks task
evidence before completion.
Current-task mining requires Mastermind 3.2.1 or newer. See the
source build instructions for local development.
Requires Node.js 24+. The npm manifest lists
prebuilt binaries for macOS, Linux, and Windows. Run these commands inside your repository:README 전문 보기 ↗Mastermind gives coding agents local code and documentation search, project knowledge, optional personal preferences, and a controller that checks task evidence before completion.
Current-task mining requires Mastermind 3.2.1 or newer. See the source build instructions for local development.
Requires Node.js 24+. The npm manifest lists prebuilt binaries for macOS, Linux, and Windows. Run these commands inside your repository:
npm install -g @xcraftmind/mastermind
mastermind init
The first run detects the active client, or the installed Claude Code and Codex
clients. It creates project guidance, indexes code and documentation, connects
the clients, and enables profile delivery and --mining task. Preferences are
proposed by the current agent and saved locally after complete capture. No
separate model is launched. Use --client none for local indexing only.
Saved choices survive repeated init.
| Command | Result |
|---|---|
mastermind init | Apply saved project choices from .mastermind/setup.json |
mastermind status --json | Inspect the index, client configuration, session observations, and miner |
mastermind update | Update the package in its existing npm scope and refresh installed workflows |
mastermind ui --since main | Open the read-only Lens UI against your chosen Git baseline |
Restart the selected client and review the generated definitions in /hooks.
Registration, session
observations, and client trust are separate states. Keep .mastermind/ working
data out of version control. See Getting started for
mining choices, Windows support, and other installation paths.
| Need | Result |
|---|---|
| Review a change | Changed symbols, callers, boundaries, and candidate tests |
| Find context | Code search, cited Markdown, decisions, and task history |
| Apply working preferences | Reviewed habits selected by project, role, and workflow |
| Delegate a task | Approved scope, observed checks, audit, and semantic review |
| Share evidence | Lens, offline HTML, SARIF, and audit envelopes |
| Import analysis | SCIP, scanner reports, coverage, tests, traces, and facts |
init --client claude|codex|all connects Claude Code, Codex, or both.
Selected clients receive relevant Git observations and reviewed preferences
through prompt hooks and MCP. Git observations refresh locally at init and task
boundaries. macOS and Linux support native hooks.
Other clients use MCP setup.
| Context | Scope |
|---|---|
| Person | Advisory habits and preferences shared across repositories |
| Project | Repository decisions, constraints, and knowledge |
| Code and documentation | Indexed structure and cited source material |
| Role | Duties for this invocation |
| Workflow | Planning, execution, verification, and review |
| Task evidence | Results, blockers, and the reviewed revision |
Profile access is scoped to the selected project and clients. Use
--profile-access off to disable it, or --mining capture to collect without
semantic analysis. Preferences grant no action permissions. Context previews expose revisions
and omissions, not proof of model use. See Architecture
and Persona hooks.
mastermind new-spec "Add account recovery"
Fill the goal, scope, acceptance criteria, and checks, then follow the Workflow guide:
criteria → implementation → checks → audit → review → complete or feedback
Any client can implement the spec. Recorded native execution and review use Claude Code. Completion requires current evidence and resolved review.
| Boundary | Meaning |
|---|---|
| Static codegraph | Dynamic dispatch, reflection, generated code, and cross-language calls may be incomplete |
| Personal profile | Reviewed preferences remain advisory |
| Task records | Local evidence does not independently establish correctness |
| Native process policy | Uses client restrictions, not an OS sandbox |
| Task mining | Uses the current agent's allowance for added context and tools. Drafts require review |
| Separate model requests | Background mining, refinement, drafting, execution and review require their own options. See provider contracts |
Language coverage includes Python,
TypeScript/TSX, JavaScript/JSX, Vue SFC, Rust, C#, Go, Java, PHP, and C/C++.
Native hooks, managed mining, and supervised execution require macOS or Linux.
Windows can use local indexing and client setup with --mining off. See
runtime support.
| Start | Reference |
|---|---|
| Getting started | CLI and MCP |
| Architecture | Benchmarks |
| Workflow | Changelog |
| All guides | Security policy |
Source builds require Rust 1.96+, declared in Cargo.toml.
cargo install --path mcp/servers/mmcg --locked
just check
Cargo installs mmcg. npm exposes mastermind and mmcg.
See Contributing, GitHub Issues,
and the security reporting policy.
MIT.