sentix
v2.47.0
Published
Autonomous multi-agent DevSecOps pipeline CLI
Maintainers
Readme
Sentix
Adaptive Guardrail for AI Coding — A framework that enforces boundaries when AI writes code, learns from mistakes, and improves itself over time.
What is Sentix?
| | | |---|---| | Definition | Adaptive Guardrail for AI Coding | | Identity | Not a harness (direction), not an agent (autonomy) — boundary enforcement + verification + learning | | Core Principle | Instead of telling AI "do it this way," mechanically enforce "never do this" | | Differentiator | Not just blocking — adaptive: learns from failures to prevent the same mistake next time |
Features
| Feature | Description | |---|---| | Destruction Prevention | Physically blocks deletion of existing features, tests, and out-of-scope files | | Deterministic Verification | Catches eval/syntax errors/test regression/security vulnerabilities — 100% by machine | | Input Enrichment | Vague request → interactive selection menu → rich specification | | Auto-Learning | Verification failure → constraint auto-added → same mistake prevented next time | | Multi-Generation | Same request implemented 3 ways → best one selected by Quality Gate score | | Cross-Model Review | Independent review by different AI model + adversarial prompting | | Pattern Learning | Usage pattern analysis → behavioral directives auto-generated | | Integrity Monitoring | Protected file tampering/deletion detected → auto-restored from git |
Feature Availability by Environment
Each cell below was verified against the current code path. sentix run is split by execution context (terminal TTY vs Claude Code's Bash tool) because the two differ for prompts that require stdin.
| Feature | Terminal sentix run | sentix run from Claude Code Bash | Claude Code chat (sentix init done) | claude.ai Web | API |
|---|:---:|:---:|:---:|:---:|:---:|
| 6 Hard Rules injected as context | ✅ via phase prompts | ✅ via phase prompts | ✅ via SessionStart hook | ⚠️ paste manually | ⚠️ paste manually |
| PreToolUse hook blocks Write/Edit | — pass-through (state=in_progress) | — pass-through | ✅ blocks when no active ticket | ❌ no hook support | ❌ no hook support |
| Interactive input enrichment (readline) | ✅ | ⚠️ stdin not TTY → prompts shown but answers cannot be captured | ❌ (path doesn't go through sentix run) | ❌ | ❌ |
| Quality Gate (5 deterministic checks) | ✅ | ✅ | ❌ | ❌ | ❌ |
| Feedback Loop (constraints auto-appended) | ✅ | ✅ | ❌ | ❌ | ❌ |
| Multi-Gen (--multi-gen [N]) | ✅ | ✅ | ❌ | ❌ | ❌ |
| Cross-Review (--cross-review <model>) | ✅ | ✅ | ❌ | ❌ | ❌ |
| Pattern analysis + directives | ✅ | ✅ | ❌ | ❌ | ❌ |
| Integrity verify + auto-restore from git | ❌ (verify is not invoked) | ❌ | ✅ via SessionStart hook | ❌ | ❌ |
| Integrity snapshot refresh | ✅ in finalize | ✅ in finalize | ❌ | ❌ | ❌ |
Install / Update / Commands
# Install (choose one)
npm install sentix # Local install (for Claude Code app/web)
npm install -g sentix # Global install (for terminal CLI)
sentix init # or: npx sentix init
# Update
npm install sentix@latest # Local update
npm install -g sentix@latest # Global update
sentix update # Sync framework filesWhich install do I need?
| Environment | Install | Command |
|---|---|---|
| Claude Code App / Web | Local (npm install sentix) | npx sentix run "request" |
| Terminal CLI | Global (npm install -g sentix) | sentix run "request" |
| Both | Both (no conflict) | Either works |
Important: Global install (
-g) is NOT accessible from Claude Code app/web. If you use Claude Code app/web, you MUST install locally (npm install sentix).
How to use in Claude Code chat
One-time setup — make Claude route through sentix run automatically
Run sentix init once in your project root:
npm install sentix
npx sentix initThis installs three things that keep Claude on the Governor pipeline without you re-typing the rule each request:
| Installed by sentix init | What it does |
|---|---|
| CLAUDE.md / .claude/rules/* | Loaded at every session start — tells Claude it is already the Sentix Governor and must route code changes through sentix run. |
| .claude/settings.json hooks | SessionStart injects the Governor role, UserPromptSubmit re-states the rule on every turn, PreToolUse blocks raw Write/Edit without an active ticket. |
| tasks/ + .sentix/ scaffolding | Ticket index, lessons, constraints, integrity snapshot — all the state Claude reads/writes during a cycle. |
| docs/system-prompt-template.md | A ready-to-paste system prompt for Claude API / Desktop / Project Instructions, so external sessions (no CLAUDE.md auto-load) follow the same Governor rules. Preserved if you've already customized it. sentix doctor warns if missing. |
After this, a fresh chat session automatically starts in Governor mode. You can still ask in plain language:
You: "Fix the login bug"
Claude: (Governor routes this to) sentix run "Fix the login bug"
→ Pipeline: PLAN → DEV → GATE → REVIEW → FINALIZEExplicit fallback
If Claude drifts or you want to force the pipeline explicitly:
You: "Run npx sentix run 'Make a login page' in terminal"App / Web specifics
Global install is not visible to Claude Code app/web. You must install locally (
npm install sentix) sonpx sentixresolves inside the project'snode_modules/.bin.
Memory files (persistent preferences across sessions) live in
~/.claude/projects/<project>/memory/. Sentix'srequire-tickethook treats paths outside the project root as out-of-scope and lets them through, so memory writes are never blocked.
Sentix란?
| 항목 | 설명 | |---|---| | 정의 | AI 코딩 가드레일 (Adaptive Guardrail) | | 정체성 | 하네스(방향 지시)도, 에이전트(자율 행동)도 아닌 — 경계 강제 + 검증 + 학습 | | 핵심 원리 | AI에게 "이렇게 해라"가 아니라 "이것만은 하지 마라"를 기계적으로 강제 | | 차별점 | 단순 차단이 아닌 adaptive — 실패에서 배워서 다음에 같은 실수를 방지 |
기능 요약
| 기능 | 설명 | |---|---| | 파괴 방지 | 기존 기능 삭제, 테스트 파괴, 범위 초과를 물리적으로 차단 | | 결정론적 검증 | eval/구문오류/테스트회귀/보안취약점을 기계가 100% 잡음 | | 입력 구체화 | 모호한 요청 → 인터랙티브 선택지 → 풍성한 스펙 | | 자동 학습 | 검증 실패 → 제약 자동 추가 → 다음엔 같은 실수 방지 | | 다중 생성 | 같은 요청을 3가지 접근으로 구현 → 최선 선택 | | 이종 검증 | 다른 AI 모델로 독립 리뷰 + 적대적 프롬프트 | | 패턴 학습 | 사용 패턴 분석 → 행동 지시문 자동 생성 | | 무결성 감시 | 보호 파일 변조/삭제 감지 → git에서 자동 복원 |
환경별 기능 지원
아래 표의 모든 셀은 현재 코드 동작과 1:1 검증되었다. sentix run 은 stdin TTY 유무에 따라 동작이 달라지므로 터미널 직접 실행 과 Claude Code 의 Bash 도구로 실행 을 분리해 표기했다.
| 기능 | 터미널 sentix run | Claude Code Bash 로 sentix run | Claude Code 대화 (sentix init 완료) | claude.ai 웹 | API |
|---|:---:|:---:|:---:|:---:|:---:|
| 하드 룰 6개 컨텍스트 주입 | ✅ phase 프롬프트 경유 | ✅ phase 프롬프트 경유 | ✅ SessionStart 훅 | ⚠️ 수동 복붙 | ⚠️ 수동 복붙 |
| PreToolUse 훅의 Write/Edit 차단 | — pass-through (state=in_progress) | — pass-through | ✅ 활성 티켓 없으면 차단 | ❌ 훅 미지원 | ❌ 훅 미지원 |
| 인터랙티브 입력 구체화 (readline) | ✅ | ⚠️ stdin 비-TTY → 질문은 출력되나 응답 수집 불가 | ❌ (sentix run 경로 아님) | ❌ | ❌ |
| Quality Gate (결정론적 5검사) | ✅ | ✅ | ❌ | ❌ | ❌ |
| Feedback Loop (제약 자동 추가) | ✅ | ✅ | ❌ | ❌ | ❌ |
| Multi-Gen (--multi-gen [N]) | ✅ | ✅ | ❌ | ❌ | ❌ |
| Cross-Review (--cross-review <모델>) | ✅ | ✅ | ❌ | ❌ | ❌ |
| 패턴 분석 + 행동 지시 | ✅ | ✅ | ❌ | ❌ | ❌ |
| Integrity verify + git 자동 복원 | ❌ (verify 호출 안 함) | ❌ | ✅ SessionStart 훅 | ❌ | ❌ |
| Integrity snapshot 갱신 | ✅ finalize 단계 | ✅ finalize 단계 | ❌ | ❌ | ❌ |
설치 / 업데이트 / 명령어
# 설치 (둘 중 선택)
npm install sentix # 로컬 설치 (Claude Code 앱/웹용)
npm install -g sentix # 글로벌 설치 (터미널 CLI용)
sentix init # 또는: npx sentix init
# 업데이트
npm install sentix@latest # 로컬 업데이트
npm install -g sentix@latest # 글로벌 업데이트
sentix update # 프레임워크 파일 동기화어떤 설치가 필요한가?
| 환경 | 설치 방법 | 명령어 |
|---|---|---|
| Claude Code 앱 / 웹 | 로컬 (npm install sentix) | npx sentix run "요청" |
| 터미널 CLI | 글로벌 (npm install -g sentix) | sentix run "요청" |
| 둘 다 | 둘 다 설치 (충돌 없음) | 어느 쪽이든 작동 |
중요: 글로벌 설치(
-g)는 Claude Code 앱/웹에서 접근 불가합니다. Claude Code 앱/웹을 사용한다면 반드시 로컬 설치(npm install sentix)가 필요합니다.
Claude Code 채팅에서 사용하는 법
한 번만 설정 — 채팅에서 Claude가 자동으로 sentix run 을 거치게 하기
프로젝트 루트에서 sentix init 을 한 번 실행하면, 이후 모든 세션이 자동으로 Governor 모드로 시작합니다:
npm install sentix
npx sentix initsentix init 이 자동 설치하는 3가지:
| 설치 항목 | 역할 |
|---|---|
| CLAUDE.md / .claude/rules/* | 세션 시작 시 자동 로드되어, Claude가 이미 Sentix Governor임을 인지하고 코드 변경은 반드시 sentix run 경로로 돌리도록 지시 |
| .claude/settings.json 훅 3종 | SessionStart 가 Governor 역할 주입, UserPromptSubmit 가 매 요청마다 규칙 리마인드, PreToolUse 가 활성 티켓 없는 Write/Edit 차단 |
| tasks/ + .sentix/ 구조 | 티켓 인덱스, 교훈, 제약, integrity snapshot — 사이클 중 Claude가 읽고 쓰는 상태 저장소 |
| docs/system-prompt-template.md | Claude API / Desktop / Project Instructions 에 그대로 붙여넣기 가능한 system prompt. CLAUDE.md 자동 로드가 안 되는 외부 세션도 같은 Governor 규칙을 따르게 한다. 이미 커스터마이즈된 파일은 보존. sentix doctor 가 누락 시 경고. |
설정 후에는 자연어로 요청해도 자동으로 파이프라인을 탑니다:
나: "로그인 버그 고쳐줘"
Claude: (Governor 판단) sentix run "로그인 버그 고쳐줘"
→ 파이프라인: PLAN → DEV → GATE → REVIEW → FINALIZE명시적 실행 (fallback)
Claude가 경로를 벗어나거나 강제로 파이프라인을 타게 하고 싶으면:
나: "터미널에서 npx sentix run '로그인 페이지 만들어줘' 실행해"앱 / 웹 사용 시 참고
글로벌 설치는 Claude Code 앱/웹에서 보이지 않습니다. 반드시 프로젝트에 로컬 설치(
npm install sentix) 해야npx sentix가node_modules/.bin에서 해결됩니다.
세션 간 기억(memory 파일)은
~/.claude/projects/<프로젝트>/memory/에 저장됩니다. Sentix의require-ticket훅은 프로젝트 루트 밖 경로를 보호 범위 밖으로 간주해 통과시키므로, memory 쓰기가 막히지 않습니다.
Sentixとは?
| 項目 | 説明 | |---|---| | 定義 | AIコーディング・ガードレール(Adaptive Guardrail) | | アイデンティティ | ハーネス(方向指示)でもエージェント(自律行動)でもない — 境界強制 + 検証 + 学習 | | 基本原則 | AIに「こうしろ」ではなく「これだけはするな」を機械的に強制 | | 差別化要素 | 単なるブロックではなくadaptive — 失敗から学び、次回同じミスを防止 |
機能一覧
| 機能 | 説明 | |---|---| | 破壊防止 | 既存機能の削除、テスト破壊、スコープ外の変更を物理的にブロック | | 決定論的検証 | eval/構文エラー/テスト回帰/セキュリティ脆弱性をマシンが100%検出 | | 入力具体化 | 曖昧なリクエスト → インタラクティブ選択肢 → 豊富な仕様 | | 自動学習 | 検証失敗 → 制約自動追加 → 次回は同じミスを防止 | | 多重生成 | 同じリクエストを3つのアプローチで実装 → 最善を選択 | | 異種検証 | 異なるAIモデルによる独立レビュー + 敵対的プロンプト | | パターン学習 | 使用パターン分析 → 行動指示を自動生成 | | 整合性監視 | 保護ファイルの改ざん/削除を検知 → gitから自動復元 |
環境別機能サポート
下表のすべてのセルは現在のコード動作と 1:1 検証済み。sentix run は stdin の TTY 有無で動作が変わるため、ターミナル直接実行 と Claude Code の Bash ツールから実行 を分けて記載している。
| 機能 | ターミナル sentix run | Claude Code Bash から sentix run | Claude Code チャット (sentix init 済) | claude.ai Web | API |
|---|:---:|:---:|:---:|:---:|:---:|
| ハードルール6個のコンテキスト注入 | ✅ phase プロンプト経由 | ✅ phase プロンプト経由 | ✅ SessionStart フック | ⚠️ 手動貼付 | ⚠️ 手動貼付 |
| PreToolUse フックの Write/Edit ブロック | — pass-through (state=in_progress) | — pass-through | ✅ アクティブチケットなしならブロック | ❌ フック非対応 | ❌ フック非対応 |
| インタラクティブ入力具体化 (readline) | ✅ | ⚠️ stdin が TTY でない → 質問は出るが回答取得不可 | ❌ (sentix run 経路ではない) | ❌ | ❌ |
| Quality Gate (決定論的5検査) | ✅ | ✅ | ❌ | ❌ | ❌ |
| Feedback Loop (制約自動追加) | ✅ | ✅ | ❌ | ❌ | ❌ |
| Multi-Gen (--multi-gen [N]) | ✅ | ✅ | ❌ | ❌ | ❌ |
| Cross-Review (--cross-review <モデル>) | ✅ | ✅ | ❌ | ❌ | ❌ |
| パターン分析 + 行動指示 | ✅ | ✅ | ❌ | ❌ | ❌ |
| Integrity verify + git からの自動復元 | ❌ (verify は呼ばれない) | ❌ | ✅ SessionStart フック | ❌ | ❌ |
| Integrity スナップショット更新 | ✅ finalize 段階 | ✅ finalize 段階 | ❌ | ❌ | ❌ |
インストール / アップデート / コマンド
# インストール (どちらか選択)
npm install sentix # ローカル (Claude Code アプリ/ウェブ用)
npm install -g sentix # グローバル (ターミナルCLI用)
sentix init # または: npx sentix init
# アップデート
npm install sentix@latest # ローカル更新
sentix update # フレームワークファイル同期
# 使用方法
sentix run "リクエスト" # パイプライン実行
sentix run "リクエスト" --multi-gen # 多重生成モード
sentix run "リクエスト" --cross-review # 異種モデルレビュー
sentix doctor # インストール診断重要: グローバルインストール (
-g) は Claude Code アプリ/ウェブから見えません。アプリ/ウェブを使う場合はローカルインストール (npm install sentix) が必須です。詳細な自動ルーティング設定は上の英語/韓国語セクションを参照してください。
Claude Codeチャットでの使い方
チャットでnpx sentix runを明示的に指定する必要があります:
あなた:「ターミナルで npx sentix run 'ログインページを作って' を実行して」
Claude:(Bashツールで実行)
→ パイプライン開始:PLAN → DEV → GATE → REVIEW → FINALIZE重要:「ログインページを作って」だけ入力すると、Claudeはsentixなしで直接コーディングします。 必ず
npx sentix runをリクエストに含めてください。
Sentix是什么?
| 项目 | 说明 | |---|---| | 定义 | AI编程护栏(Adaptive Guardrail) | | 定位 | 既不是缰绳(方向指引),也不是代理(自主行为)— 边界强制 + 验证 + 学习 | | 核心原则 | 不是告诉AI"这样做",而是机械性地强制"绝对不能做这个" | | 差异化 | 不仅仅是阻断 — 具有自适应性:从失败中学习,防止下次犯同样的错误 |
功能概述
| 功能 | 说明 | |---|---| | 破坏防护 | 物理阻断现有功能删除、测试破坏、超出范围的修改 | | 确定性验证 | eval/语法错误/测试回归/安全漏洞 — 机器100%捕获 | | 输入具体化 | 模糊请求 → 交互式选择菜单 → 丰富的规格说明 | | 自动学习 | 验证失败 → 自动添加约束 → 下次防止同样的错误 | | 多重生成 | 同一请求用3种方法实现 → 选择最佳方案 | | 异构验证 | 不同AI模型独立审查 + 对抗性提示 | | 模式学习 | 使用模式分析 → 自动生成行为指令 | | 完整性监控 | 受保护文件篡改/删除检测 → 从git自动恢复 |
各环境功能支持
下表所有单元格均与当前代码行为 1:1 验证。sentix run 因 stdin 是否为 TTY 而行为不同,故将 终端直接执行 与 从 Claude Code 的 Bash 工具执行 分开列出。
| 功能 | 终端 sentix run | 从 Claude Code Bash 执行 sentix run | Claude Code 对话 (已完成 sentix init) | claude.ai Web | API |
|---|:---:|:---:|:---:|:---:|:---:|
| 6条硬规则的上下文注入 | ✅ 经 phase 提示 | ✅ 经 phase 提示 | ✅ SessionStart 钩子 | ⚠️ 手动粘贴 | ⚠️ 手动粘贴 |
| PreToolUse 钩子阻断 Write/Edit | — pass-through (state=in_progress) | — pass-through | ✅ 无活动票据时阻断 | ❌ 不支持钩子 | ❌ 不支持钩子 |
| 交互式输入具体化 (readline) | ✅ | ⚠️ stdin 非 TTY → 显示问题但无法收集回答 | ❌ (不经 sentix run) | ❌ | ❌ |
| Quality Gate (确定性5项检查) | ✅ | ✅ | ❌ | ❌ | ❌ |
| Feedback Loop (约束自动追加) | ✅ | ✅ | ❌ | ❌ | ❌ |
| Multi-Gen (--multi-gen [N]) | ✅ | ✅ | ❌ | ❌ | ❌ |
| Cross-Review (--cross-review <模型>) | ✅ | ✅ | ❌ | ❌ | ❌ |
| 模式分析 + 行为指令 | ✅ | ✅ | ❌ | ❌ | ❌ |
| Integrity verify + 从 git 自动恢复 | ❌ (不调用 verify) | ❌ | ✅ SessionStart 钩子 | ❌ | ❌ |
| Integrity 快照刷新 | ✅ finalize 阶段 | ✅ finalize 阶段 | ❌ | ❌ | ❌ |
安装 / 更新 / 命令
# 安装 (二选一)
npm install sentix # 本地安装 (供 Claude Code 应用/网页使用)
npm install -g sentix # 全局安装 (供终端 CLI 使用)
sentix init # 或: npx sentix init
# 更新
npm install sentix@latest # 本地更新
sentix update # 框架文件同步
# 使用
sentix run "请求" # 管道执行
sentix run "请求" --multi-gen # 多重生成模式
sentix run "请求" --cross-review # 异构模型审查
sentix doctor # 安装诊断重要:全局安装 (
-g) 在 Claude Code 应用/网页中不可见。如使用应用/网页,必须本地安装 (npm install sentix)。自动路由详细设置请参考上方英文/韩文部分。
在Claude Code聊天中如何使用
聊天中必须明确指定npx sentix run:
你:"在终端执行 npx sentix run '做一个登录页面'"
Claude:(通过Bash工具执行)
→ 管道启动:PLAN → DEV → GATE → REVIEW → FINALIZE重要:如果只输入"做一个登录页面",Claude会不经过sentix直接编码。 请务必在请求中包含
npx sentix run。
Layer Status (v2.6.0)
The 5-layer architecture's current implementation status is recorded honestly so the framework's claims match the codebase:
- L1 — Governor + Agents ✅
- L2 — Learning Pipeline ✅
- L3 — Pattern Engine ✅
- L4 — Visual Perception ⏳ Planned (toggle and
sentix statuslabel exist; no consuming code yet) - L5 — Self-Evolution ✅ (depth TBD — pending audit)
Pipeline Architecture
User Request
↓
[Spec Enricher] — Load .sentix/constraints.md + past lessons
↓
[Spec Questions] — Structured clarifying questions for vague requests
↓
[Pattern Directive] — Generated from tasks/pattern-log.jsonl
↓
planner → dev ── or ──→ dev-swarm (parallel worktrees)
└─ or ─→ dev × N [Multi-Gen, --multi-gen [N]]
↓
[Quality Gate] — 5 deterministic checks
↓
[verify-gates] — Scope / export / test / net-deletion
↓
[Feedback Loop] — Failures → constraints.md auto-appended
↓
pr-review (adversarial — reviewer must find ≥ 2 issues)
↓
[Cross-Review] — Independent review by external model (opt-in, --cross-review)
↓
finalize → integrity snapshot + lesson promotion → DoneHard Rules
| # | Rule | Enforcement | |---|---|---| | 1 | Test snapshot required before changes | Pre-execution gate (auto-creates if missing) | | 2 | No out-of-scope file modifications | git diff analysis (only when planner defines SCOPE) | | 3 | No export/API deletion | git diff analysis (signature extensions are excluded) | | 4 | No test deletion | git diff analysis | | 5 | Max 50 lines net deletion | git diff analysis | | 6 | No existing feature deletion | Dev/review agent prompts (LLM-enforced, not a hard hook) |
Quality Gate — 5 Deterministic Checks
| Check | Catches | Why AI Misses It |
|---|---|---|
| Banned patterns | eval(), innerHTML, hardcoded secrets | AI assumes "it's intentional" |
| Debug artifacts | console.log in src/ | AI adds them and forgets to remove |
| Syntax check | .js syntax errors | AI is "sure the brackets match" |
| npm audit | Security vulnerabilities | AI can't access CVE DB in real-time |
| Test regression | Decreased test count | AI assumes "tests probably pass" |
Multi-Gen
Runs dev N times (default 3; override with -mg N or --multi-gen N), scores each generation via the Quality Gate, and applies the highest-scoring patch.
Gen 1 [Simplest] → score 85 (issues: 1)
Gen 2 [Robust] → score 95 (issues: 0) ★ Selected
Gen 3 [Elegant] → score 80 (issues: 2)Cross-Review
sentix run "request" --cross-review openai # External model reviewReviewer is required to find at least 2 issues (adversarial prompting).
| Command | Description |
|---|---|
| sentix | Status + recommended next action |
| sentix run "request" | Pipeline execution |
| sentix status | Governor dashboard |
| sentix doctor | Installation diagnostics |
| sentix ticket create "desc" | Bug ticket |
| sentix ticket close <id> [--force] | Close a resolved ticket (or force-close any state) |
| sentix feature add "desc" | Feature ticket |
| sentix version bump patch | Version bump |
| sentix metrics | AI success rate stats |
| sentix init | Project setup |
| sentix update | Framework file sync |
| sentix config | Settings |
| sentix safety set <word> | Anti-injection safety word |
| sentix evolve | Self-analysis |
my-project/
├── CLAUDE.md ← AI execution rules (core)
├── FRAMEWORK.md ← Detailed design document
├── .sentix/
│ ├── config.toml ← Layer enable/disable
│ ├── constraints.md ← Project constraints (auto-learning)
│ ├── providers/ ← AI selection (Claude, OpenAI, Ollama)
│ └── rules/ ← Hard rules + auto-generated rules
├── .claude/
│ ├── settings.json ← Hook registration
│ ├── agents/ ← planner, dev, pr-review agents
│ └── rules/ ← Conditional rules
├── scripts/hooks/ ← SessionStart, UserPromptSubmit, PreToolUse hooks
├── tasks/
│ ├── lessons.md ← Lessons from failures
│ ├── patterns.md ← Usage patterns
│ └── tickets/ ← Work tickets
├── docs/ ← SOP, agent scopes, severity
└── (existing project files unchanged)Q: Does it change existing code?
A: sentix init adds framework files (CLAUDE.md, .claude/, .sentix/, scripts/hooks/, tasks/) and, if a CLAUDE.md already exists, merges its rules. Your source tree is not modified — only the scaffolding around it.
Q: Is sentix run required?
A: Inside Claude Code with sentix init done, the installed hooks + CLAUDE.md strongly bias Claude toward routing code changes through sentix run. It is not a hard guarantee — if Claude still tries a direct Write/Edit, the PreToolUse hook blocks it when no active ticket exists.
Q: Without Claude Code? A: claude.ai web/mobile: guidance mode only (you must manually paste the hard rules and run the pipeline yourself). API: same — no hooks, no auto-enforcement.
Q: Is it free? A: Sentix itself is MIT-licensed and has zero external dependencies. AI costs depend on your provider (Claude/OpenAI/Ollama).
Q: Which languages?
A: The project scanner auto-recognizes .js, .ts, .jsx, .tsx, .py, .go, .rs, .java, .rb files. Hard rules and pipeline are language-agnostic; per-language conventions go in CLAUDE.md.
Test
npm test # 212 tests, Node.js built-in test runner (node --test)Contributing
- providers: Add new AI (Gemini, Mistral, etc.)
- plugins: Custom plugins (
sentix plugin create)
License
MIT — by JANUS
