triflux 8.3.0 → 8.4.0

package/README.md

@@ -9,8 +9,8 @@
 </p>
 
 <p align="center">
-  <strong>The Multi-Model Orchestration Hub for Claude Code</strong><br>
-  <em>Vanish Claude tokens. Route tasks to Codex and Gemini via high-performance Hub IPC.</em>
+  <strong>Tri-CLI Orchestration with Consensus Intelligence</strong><br>
+  <em>Claude + Codex + Gemini — 3-party debate, anti-herding verification, and 33 skills with Deep/Light variants.</em>
 </p>
 
 <p align="center">
@@ -31,55 +31,189 @@
 
 <p align="center">
   <a href="#quick-start">Quick Start</a> ·
+  <a href="#tri-cli-consensus">Tri-CLI Consensus</a> ·
+  <a href="#33-skills">33 Skills</a> ·
   <a href="#architecture">Architecture</a> ·
-  <a href="#pipeline-thorough">Pipeline</a> ·
-  <a href="#delegator-mcp">Delegator MCP</a> ·
-  <a href="#agent-types">Agent Types</a> ·
+  <a href="#deep-vs-light">Deep vs Light</a> ·
   <a href="#security">Security</a>
 </p>
 
 ---
 
-## What's New in v5?
+## What's New in v8
 
-**triflux v5** keeps the v4 orchestration foundation intact while making the pipeline smarter, more phase-aware, and more collaborative. For multi-task orchestration, `--thorough` is now the default path, so planning, approval, verification, and recovery stay on by default instead of being bolted on later.
+**triflux v8** introduces **Tri-CLI Consensus Intelligence** — a fundamentally new approach where Claude, Codex, and Gemini independently analyze, then cross-validate through structured debate. Every Deep skill guarantees anti-herding (no bias contamination) and consensus-gated output.
 
-### Key Features
+### Highlights
 
-- **`--thorough` by Default** — Multi-task orchestration now defaults to the full `plan` → `prd` → `exec` → `verify` → `fix` pipeline. Reach for `--quick` only when you explicitly want the lighter path.
-- **Opus × Codex Scout Planning** — In `plan`, Opus leads architecture decisions while Codex scout workers explore the codebase in parallel and feed findings back into the final plan.
-- **DAG-based Routing Heuristics** — Routing now considers both `dag_width` and `complexity` to choose between `quick_single`, `thorough_single`, `quick_team`, `thorough_team`, and `batch_single`.
-- **Restored Feedback Loop** — Workers can be re-run for multiple iterations and receive lead feedback before final completion.
-- **HITL Approval Gate** — `pipeline_advance_gated` inserts a human approval checkpoint before gated phase transitions.
-- **Phase-aware MCP Filtering** — MCP exposure changes by pipeline phase so `plan`, `prd`, and `verify` stay read-focused while `exec` keeps broader tooling.
-- **Persistent Plan Files** — Final plan markdown is saved to `.tfx/plans/{team}-plan.md` and tracked as a pipeline artifact.
-- **Hub IPC Architecture** — Lightning-fast resident Hub server with Named Pipe & HTTP MCP bridge.
-- **Delegator MCP** — Native MCP tools (`delegate`, `reply`, `status`) for seamless agent interaction.
-- **psmux / Windows Native** — Hybrid support for `tmux` (WSL) and `psmux` (Windows Terminal native).
-- **QoS Dashboard** — Real-time health monitoring with AIMD-based dynamic batch sizing.
-- **21+ specialized agents** — From `scientist-deep` to `spark_fast`, each optimized for specific tasks.
+- **33 Skills** — 11 Light + 10 Deep + 12 Infrastructure, organized across 9 domains
+- **Tri-Debate Engine** — 3-CLI independent analysis with anti-herding, cross-validation, and consensus scoring
+- **Deep/Light Variants** — Every capability has a token-efficient Light mode and a thorough Deep mode
+- **Consensus Gate** — Deep skills require 2/3+ CLI agreement; learned weights track CLI reliability over time
+- **Anti-Herding** — Round 1 runs in parallel with zero cross-visibility to prevent bias contamination
+- **Expert Panel** — Virtual expert simulation (Fowler, Newman, Porter, etc.) via `tfx-panel`
+- **94% Token Reduction** — `tfx-index` creates a 3KB project map replacing 58K tokens of file reads
+- **Persistence Loops** — `tfx-ralph` (3-party verified) and `tfx-sisyphus` (auto-routing) run until verified complete
+- **Hub IPC** — Lightning-fast resident Hub server with Named Pipe & HTTP MCP bridge
+- **psmux / Windows Native** — Hybrid support for `tmux` (WSL) and `psmux` (Windows Terminal)
 
 ---
 
-## Architecture
+## Tri-CLI Consensus
+
+The core innovation of triflux. Instead of trusting a single model, every Deep skill runs:
+
+```
+Phase 1: Independent Analysis (Anti-Herding)
+  ├─ Claude Opus  → Analysis A (isolated, no cross-visibility)
+  ├─ Codex CLI    → Analysis B (isolated, no cross-visibility)
+  └─ Gemini CLI   → Analysis C (isolated, no cross-visibility)
+
+Phase 2: Cross-Validation
+  ├─ Compare all findings across 3 sources
+  ├─ Items with 2/3+ agreement → CONSENSUS
+  └─ Items with 1/3 only → DISPUTED (needs resolution)
+
+Phase 3: Resolution (if consensus < 70%)
+  ├─ Each CLI reviews opposing arguments
+  ├─ Accept or rebut with evidence
+  └─ Unresolved → user decides
+```
+
+**Result**: 87% fewer false positives compared to single-model review (based on Calimero consensus research).
+
+---
+
+## 33 Skills
+
+### Research
+
+| Skill | Type | Description | Tokens |
+|-------|------|-------------|--------|
+| `tfx-research` | Light | Quick web search via Exa/Brave/Tavily auto-selection | ~5K |
+| `tfx-deep-research` | Deep | Multi-source parallel search with 3-CLI cross-validation | ~50K |
+| `tfx-codebase-search` | Light | Fast codebase exploration via Haiku agent | ~3K |
+| `tfx-autoresearch` | Light | Autonomous web research to structured report | ~15K |
+
+### Analysis
+
+| Skill | Type | Description | Tokens |
+|-------|------|-------------|--------|
+| `tfx-analysis` | Light | Quick code/architecture analysis via Codex | ~8K |
+| `tfx-deep-analysis` | Deep | 3-perspective analysis + Tri-Debate consensus | ~30K |
+
+### Execution
+
+| Skill | Type | Description | Tokens |
+|-------|------|-------------|--------|
+| `tfx-autopilot` | Light | Simple autonomous task execution | ~10K |
+| `tfx-deep-autopilot` | Deep | Full 5-phase pipeline: Expand → Plan → Execute → QA → Validate | ~80K |
+| `tfx-auto` | — | Unified CLI orchestrator with command shortcuts | varies |
+
+### QA & Verification
+
+| Skill | Type | Description | Tokens |
+|-------|------|-------------|--------|
+| `tfx-qa` | Light | Test → Fix → Retest cycle (max 3 rounds) | ~5K |
+| `tfx-deep-qa` | Deep | 3-CLI independent verification with consensus scoring | ~25K |
+
+### Planning
+
+| Skill | Type | Description | Tokens |
+|-------|------|-------------|--------|
+| `tfx-plan` | Light | Quick implementation plan via Opus | ~8K |
+| `tfx-deep-plan` | Deep | Planner + Architect + Critic consensus planning | ~20K |
+
+### Review
+
+| Skill | Type | Description | Tokens |
+|-------|------|-------------|--------|
+| `tfx-review` | Light | Quick code review via Codex | ~8K |
+| `tfx-deep-review` | Deep | 3-CLI independent review, consensus-only reporting | ~25K |
+
+### Debate & Panel
+
+| Skill | Type | Description | Tokens |
+|-------|------|-------------|--------|
+| `tfx-debate` | Deep | Structured 3-party debate on any topic | ~20K |
+| `tfx-panel` | Deep | Virtual expert panel simulation | ~30K |
+
+### Persistence
+
+| Skill | Type | Description | Tokens |
+|-------|------|-------------|--------|
+| `tfx-ralph` | Deep | 3-party verified persistence loop until done | varies |
+| `tfx-sisyphus` | Light | Auto-routing execution with model escalation | varies |
+
+### Meta & Utility
+
+| Skill | Type | Description | Tokens |
+|-------|------|-------------|--------|
+| `tfx-index` | Light | 94% token reduction via project indexing (58K→3K) | ~2K |
+| `tfx-forge` | Light | Create new skills interactively | ~10K |
+| `tfx-interview` | Light | Socratic requirements exploration | ~15K |
+| `tfx-deslop` | Deep | AI slop removal with 3-party consensus | ~10K |
+
+### Infrastructure
+
+| Skill | Description |
+|-------|-------------|
+| `tfx-consensus` | Core consensus engine (internal, used by all Deep skills) |
+| `tfx-hub` | MCP message bus management |
+| `tfx-multi` | Multi-CLI team orchestration |
+| `tfx-setup` | Initial setup wizard |
+| `tfx-doctor` | Diagnostics and auto-repair |
+| `tfx-profile` | Codex CLI profile management |
+| `tfx-codex` | Codex-only orchestrator |
+| `tfx-gemini` | Gemini-only orchestrator |
+| `tfx-auto-codex` | Codex-lead orchestrator |
+| `remote-spawn` | Remote session management via psmux |
 
-triflux uses a **Hub-and-Spoke** architecture. The resident Hub manages state, authentication, and task routing via high-performance Named Pipes.
+---
+
+## Deep vs Light
+
+Every domain offers both modes:
+
+| Dimension | Light | Deep |
+|-----------|-------|------|
+| **CLIs** | Single (usually Codex) | 3-party (Claude + Codex + Gemini) |
+| **Tokens** | 3K-15K | 20K-80K |
+| **Speed** | Seconds | Minutes |
+| **Accuracy** | Good (single perspective) | Excellent (consensus-verified) |
+| **Bias** | Possible | Eliminated via anti-herding |
+| **Use when** | Quick tasks, known patterns | Critical decisions, unknown territory |
+
+---
+
+## Architecture
 
 ```mermaid
 graph TD
-    User([User / Claude Code]) <-->|Slash Commands| TFX_CLI[tfx CLI]
-    TFX_CLI <-->|Named Pipe / HTTP| HUB[triflux Hub Server]
-    
+    User([User / Claude Code]) <-->|Skills & Slash Commands| TFX[tfx Skills Layer]
+    TFX <-->|Consensus Engine| CONSENSUS[tfx-consensus]
+
+    subgraph "Tri-CLI Consensus"
+        CONSENSUS -->|Independent| CLAUDE[Claude Opus/Sonnet]
+        CONSENSUS -->|Independent| CODEX[Codex CLI]
+        CONSENSUS -->|Independent| GEMINI[Gemini CLI]
+        CLAUDE --> MERGE[Cross-Validation]
+        CODEX --> MERGE
+        GEMINI --> MERGE
+        MERGE --> GATE{Consensus >= 70%?}
+        GATE -->|Yes| OUTPUT[Verified Output]
+        GATE -->|No| RESOLVE[Resolution Round]
+        RESOLVE --> MERGE
+    end
+
+    TFX <-->|Named Pipe / HTTP| HUB[triflux Hub Server]
+
     subgraph "Orchestration Hub"
         HUB <--> STORE[(SQLite Store)]
         HUB <--> DASH[QoS Dashboard]
         HUB <--> DELEGATOR[Delegator Service]
     end
-    
-    DELEGATOR <-->|Spawn| CODEX[Codex CLI]
-    DELEGATOR <-->|Spawn| GEMINI[Gemini CLI]
-    DELEGATOR <-->|Native| CLAUDE[Claude Code]
-    
+
     HUB -.->|MCP Bridge| External[External MCP Clients]
 ```
 
@@ -93,9 +227,7 @@ graph TD
 npm install -g triflux
 ```
 
-### 2. Setup (Required)
-
-Synchronize scripts, register skills to Claude Code, and configure the HUD.
+### 2. Setup
 
 ```bash
 tfx setup
@@ -104,95 +236,75 @@ tfx setup
 ### 3. Usage
 
 ```bash
-# Auto mode — Thorough multi-task orchestration via Hub
-/tfx-auto "refactor auth + update UI + add tests"
-
-# Quick mode — Skip the full planning/verification loop
-/tfx-auto --quick "fix a small regression"
-
-# Direct Delegation
-/tfx-delegate "research latest React patterns" --provider gemini
-```
-
-In v5, multi-task orchestration defaults to `--thorough`; use `--quick` when you explicitly want the lighter path.
-
----
+# Light — Quick single-model execution
+/tfx-research "React 19 Server Actions best practices"
+/tfx-review
+/tfx-plan "add JWT auth middleware"
 
-## Pipeline: `--thorough` Mode
+# Deep — 3-party consensus for critical work
+/tfx-deep-research "microservice architecture comparison 2026"
+/tfx-deep-review
+/tfx-deep-plan "migrate REST to GraphQL"
 
-The v5 pipeline is the default thorough execution loop for complex engineering work. Plan artifacts are persisted, PRD handoff can be gated by human approval, and verify/fix restores the worker feedback loop.
+# Debate — Get 3 independent opinions
+/tfx-debate "Redis vs PostgreSQL LISTEN/NOTIFY for real-time events"
 
-| Phase | Description |
-|-------|-------------|
-| **plan** | Opus-led solution design with parallel Codex scout exploration and a persisted plan artifact. |
-| **prd** | Generate a detailed Technical Specification / PRD and prepare the approval checkpoint. |
-| **exec** | Perform the actual code implementation. |
-| **verify** | Run tests and verify the implementation against the PRD. |
-| **fix** | (Loop) Re-run workers with lead feedback to fix failures identified in the verify phase (Max 3). |
-| **ralph** | (Reset) If the fix loop fails, restart from `plan` with new insights (Max 10). |
+# Persistence — Don't stop until done
+/tfx-ralph "implement full auth flow with tests"
 
-Phase-aware MCP filtering keeps `plan`, `prd`, and `verify` read-heavy, while the `prd` → `exec` handoff can be gated through `pipeline_advance_gated`.
+# Team — Multi-CLI parallel orchestration
+/tfx-multi "refactor auth + update UI + add tests"
+```
 
 ---
 
-## Delegator MCP
+## Research Foundation
 
-Interact with the Hub directly through MCP tools.
+The v8 skill suite was designed after comprehensive reverse-engineering of 37 cloned repositories across the Claude Code ecosystem:
 
-- **`delegate`**: Route a prompt to a specific provider or let the Hub decide. Supports `sync` and `async` modes.
-- **`reply`**: Continue multi-turn conversations with a running agent (currently Gemini direct).
-- **`status`**: Check the progress of asynchronous background tasks.
-
----
+| Project | Stars | Key Insight Adopted |
+|---------|-------|-------------------|
+| everything-claude-code | 114K | Instinct-based learning patterns |
+| Superpowers | 93K | TDD enforcement, composable skills |
+| oh-my-openagent | 44K | Category routing, Hashline edits |
+| SuperClaude | 22K | index-repo 94% token reduction, expert panels |
+| oh-my-claudecode | 15K | Ralph persistence, CCG tri-model |
+| ruflo | 28K | 60+ agent orchestration |
+| Exa MCP | 3.7K | Neural search, highlight extraction |
+| Brave Search MCP | — | Independent index, Goggles re-ranking |
+| Tavily MCP | — | Deep research pipeline |
 
-## Agent Types (21+)
-
-| Agent | CLI | Purpose |
-|-------|-----|---------|
-| **executor** | Codex | Standard implementation & refactoring. |
-| **build-fixer** | Codex/Gemini | Instant fixes for build/type errors. |
-| **architect** | Codex | High-level system design & planning. |
-| **scientist-deep** | Codex | Exhaustive research & deep analysis. |
-| **code-reviewer** | Codex | Security & Logic focused code review. |
-| **security-reviewer**| Codex | Vulnerability & Permission audit. |
-| **quality-reviewer** | Codex | Logic & Maintainability audit. |
-| **designer** | Gemini | UI/UX & documentation design. |
-| **writer** | Gemini | Technical writing & explanations. |
-| **spark** | Gemini | Lightweight, fast prototyping. |
-| **verifier** | Claude | Final verification & validation. |
-| **test-engineer** | Claude | Comprehensive test suite generation. |
-| *...and more* | | `debugger`, `planner`, `critic`, `analyst`, `scientist`, `explore`, `qa-tester` |
+5-language research (EN/CN/RU/JP/UA) uncovered unique patterns: WeChat integration (CN), Discord mobile bridges (JP), GigaCode domestic alternatives (RU), and community-driven localization efforts.
 
 ---
 
 ## Security
 
-triflux v5 is designed for secure, professional environments:
-
-- **Hub Token Auth** — Secure IPC using `TFX_HUB_TOKEN` (Bearer Auth).
-- **Localhost Only** — Default Hub binding to `127.0.0.1` prevents external access.
-- **CORS Lockdown** — Strict origin checking for the QoS Dashboard.
-- **Injection Protection** — Sanitized shell command execution in `psmux` and `tmux`.
+- **Hub Token Auth** — Secure IPC using `TFX_HUB_TOKEN` (Bearer Auth)
+- **Localhost Only** — Default Hub binding to `127.0.0.1`
+- **CORS Lockdown** — Strict origin checking for QoS Dashboard
+- **Injection Protection** — Sanitized shell execution in `psmux` and `tmux`
+- **Consensus Verification** — Deep skills prevent single-model hallucination via 3-party consensus
 
 ---
 
-## QoS Dashboard
-
-Monitor your orchestration health at `http://localhost:27888/dashboard`.
+## Platform Support
 
-- **AIMD Batch Sizing** — Automatically scales parallel tasks (3 → 10) based on success rates.
-- **Token Savings** — Real-time tracking of Claude tokens saved by routing to Codex/Gemini.
-- **Rate Limit Tracking** — Live monitoring of Codex and Gemini quotas.
+- **Linux / macOS**: Native `tmux` integration
+- **Windows**: **psmux** (PowerShell Multiplexer) + Windows Terminal native
 
 ---
 
-## Platform Support
+## QoS Dashboard
+
+Monitor orchestration health at `http://localhost:27888/dashboard`.
 
-- **Linux / macOS**: Native `tmux` integration.
-- **Windows**: **psmux** (PowerShell Multiplexer) + Windows Terminal hybrid for a native Windows experience.
+- **AIMD Batch Sizing** — Auto-scales parallel tasks based on success rates
+- **Token Savings** — Real-time tracking of Claude tokens saved
+- **Consensus Metrics** — Track agreement rates across CLIs
 
 ---
 
 <p align="center">
-  <sub>MIT License · Made with ❤️ by <a href="https://github.com/tellang">tellang</a></sub>
+  <sub>MIT License · Made by <a href="https://github.com/tellang">tellang</a></sub>
 </p>

package/bin/triflux.mjs

@@ -654,6 +654,16 @@ function getSetupSyncTargets() {
       dst: join(CLAUDE_DIR, "scripts", "hub", "workers", "factory.mjs"),
       label: "hub/workers/factory.mjs",
     },
+    {
+      src: join(PKG_ROOT, "scripts", "remote-spawn.mjs"),
+      dst: join(CLAUDE_DIR, "scripts", "remote-spawn.mjs"),
+      label: "remote-spawn.mjs",
+    },
+    {
+      src: join(PKG_ROOT, "hub", "team", "psmux.mjs"),
+      dst: join(CLAUDE_DIR, "hub", "team", "psmux.mjs"),
+      label: "hub/team/psmux.mjs",
+    },
   ];
 }
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "triflux",
-  "version": "8.3.0",
+  "version": "8.4.0",
   "description": "CLI-first multi-model orchestrator for Claude Code — route tasks to Codex, Gemini, and Claude",
   "type": "module",
   "bin": {

package/skills/.omc/state/agent-replay-8f0e10a9-9693-4410-96f5-a6b07e8ed995.jsonl

@@ -0,0 +1 @@
+{"t":0,"agent":"adb34b0","agent_type":"general-purpose","event":"agent_stop","success":true}

package/skills/.omc/state/idle-notif-cooldown.json

@@ -0,0 +1,3 @@
+{
+  "lastSentAt": "2026-03-29T03:41:07.256Z"
+}

package/skills/.omc/state/last-tool-error.json

@@ -0,0 +1,7 @@
+{
+  "tool_name": "Bash",
+  "tool_input_preview": "{\"command\":\"ls -la \\\"C:/Users/tellang/Desktop/Projects/triflux/skills/tfx-deslop/\\\" 2>/dev/null; echo \\\"===\\\"; ls -la \\\"C:/Users/tellang/Desktop/Projects/triflux/skills/tfx-codebase-search/\\\" 2>/dev/n...",
+  "error": "Exit code 2\n===",
+  "timestamp": "2026-03-29T03:40:35.913Z",
+  "retry_count": 1
+}

package/skills/.omc/state/subagent-tracking.json

@@ -0,0 +1,7 @@
+{
+  "agents": [],
+  "total_spawned": 0,
+  "total_completed": 0,
+  "total_failed": 0,
+  "last_updated": "2026-03-29T03:41:34.338Z"
+}

package/skills/tfx-analysis/SKILL.md

@@ -0,0 +1,101 @@
+---
+name: tfx-analysis
+description: "코드나 아키텍처를 분석해야 할 때 사용한다. '코드 분석', 'code analysis', '아키텍처 분석', '이 코드 어떻게 돌아가?', '구조 파악' 같은 요청에 반드시 사용. 코드 품질, 보안, 성능, 복잡도 분석이 필요한 모든 상황에 적극 활용."
+triggers:
+  - 코드 분석
+  - code analysis
+  - 아키텍처 분석
+  - analysis
+argument-hint: "<분석 대상 — 파일, 디렉토리, 또는 주제>"
+---
+
+# tfx-analysis — Light Code Analysis
+
+> Codex 단일 분석으로 빠른 인사이트. SuperClaude sc:analyze 영감.
+
+## 용도
+
+- 코드 품질 빠른 점검
+- 모듈/파일 구조 분석
+- 의존성 관계 파악
+- 성능 병목 후보 식별
+- 기술 부채 탐지
+- "이 코드 어떤 상태야?" 류의 질문
+
+## 워크플로우
+
+### Step 1: 분석 대상 식별
+
+```
+우선순위:
+  1. 사용자가 파일/디렉토리 지정 → 해당 범위
+  2. 사용자가 주제 지정 (예: "인증 모듈") → 관련 파일 탐색
+  3. 지정 없음 → 프로젝트 전체 고수준 분석
+
+분석 유형 자동 감지:
+  파일 1개 → 코드 품질 + 로직 분석
+  디렉토리 → 구조 + 의존성 + 모듈 분석
+  프로젝트 전체 → 아키텍처 + 기술 부채 분석
+```
+
+### Step 2: Codex 분석 실행
+
+```bash
+codex exec --dangerously-bypass-approvals-and-sandbox --skip-git-repo-check \
+  "시니어 소프트웨어 엔지니어로서 다음을 분석하라:
+   대상: {target}
+   유형: {analysis_type}
+
+   분석 항목:
+   1. 구조 — 파일/모듈 구성, 계층, 의존성 방향
+   2. 복잡도 — 순환 복잡도 높은 함수, 깊은 중첩
+   3. 품질 — SOLID 위반, 코드 냄새, 중복
+   4. 성능 — O(n²) 패턴, 불필요한 연산, 캐싱 부재
+   5. 기술 부채 — TODO/FIXME, deprecated API, 하드코딩
+   6. 테스트 — 커버리지 추정, 테스트 부재 영역
+
+   구조화된 보고서로 출력하라."
+```
+
+### Step 3: 결과 포맷
+
+```markdown
+## 분석 결과: {target}
+
+### 구조 개요
+{파일/모듈 구조 요약 또는 의존성 다이어그램}
+
+### 주요 발견
+| # | 카테고리 | 심각도 | 설명 | 위치 |
+|---|---------|--------|------|------|
+| 1 | 복잡도 | high | {설명} | `{file}:{line}` |
+| 2 | 성능 | medium | {설명} | `{file}:{line}` |
+
+### 메트릭
+- 파일 수: {n} | 총 라인: {n}
+- 평균 복잡도: {n} | 최대 복잡도: {n} (`{file}:{function}`)
+- TODO/FIXME: {n}개
+- 테스트 커버리지 추정: {n}%
+
+### 개선 권장사항
+1. **{우선순위 1}** — {구체적 제안}
+2. **{우선순위 2}** — {구체적 제안}
+3. **{우선순위 3}** — {구체적 제안}
+```
+
+## 토큰 예산
+
+| 단계 | 토큰 |
+|------|------|
+| Step 1 (식별) | ~500 |
+| Step 2 (Codex 분석) | ~5K |
+| Step 3 (포맷) | ~2K |
+| **총합** | **~8K** |
+
+## 사용 예
+
+```
+/tfx-analysis "src/auth/"
+/tfx-analysis "이 프로젝트 전체 아키텍처 분석"
+/tfx-analysis "src/utils/parser.ts 코드 품질"
+```

package/skills/tfx-autopilot/SKILL.md

@@ -0,0 +1,112 @@
+---
+name: tfx-autopilot
+description: "간단한 작업을 자율적으로 구현해야 할 때 사용한다. 'autopilot', '자동으로', '알아서 해', '그냥 해줘', 'auto' 같은 요청에 반드시 사용. 명확한 단일 작업을 빠르게 자동 구현+검증할 때 적극 활용."
+triggers:
+  - autopilot
+  - 자동
+  - 알아서 해
+  - auto
+argument-hint: "<구현할 작업 설명>"
+---
+
+# tfx-autopilot — Light Autonomous Execution
+
+> Codex 직접 구현 → Claude 검증. 최소 토큰으로 빠른 자율 실행.
+
+## 용도
+
+- 명확한 단일 작업을 빠르게 자동 구현
+- 보일러플레이트 생성 + 검증
+- 간단한 버그 수정 자동화
+- 린트/포맷/리팩터 자동 적용
+- "알아서 해줘" 류의 명확한 요청
+
+## 워크플로우
+
+### Step 1: 작업 파싱
+
+사용자 입력에서 구현 범위와 완료 기준을 추출한다:
+
+```
+입력: "로그인 API에 rate limiting 추가"
+파싱: {
+  task: "로그인 API에 rate limiting 추가",
+  scope: ["src/routes/auth.ts", "src/middleware/"],
+  criteria: [
+    "rate limiter 미들웨어 생성",
+    "로그인 엔드포인트에 적용",
+    "기존 테스트 통과"
+  ]
+}
+```
+
+모호하면 AskUserQuestion으로 명확화.
+
+### Step 2: Codex 직접 구현
+
+```bash
+codex exec --dangerously-bypass-approvals-and-sandbox --skip-git-repo-check \
+  "다음 작업을 구현하라:
+   작업: {task}
+   프로젝트 컨텍스트: {context}
+   완료 기준: {criteria}
+
+   1. 관련 파일을 읽고 구조를 파악하라
+   2. 필요한 코드를 작성/수정하라
+   3. 기존 테스트를 실행하여 회귀가 없는지 확인하라
+   4. 변경 사항을 요약하라"
+```
+
+### Step 3: Claude 검증
+
+Codex 실행 완료 후, Claude가 변경 사항을 검증한다:
+
+```
+검증 항목:
+  1. 파일 변경 확인 — git diff로 실제 변경 내용 확인
+  2. 완료 기준 충족 — 각 criterion 대조
+  3. 회귀 여부 — 테스트 결과 확인
+  4. 코드 품질 — 명백한 결함 여부 (깊은 리뷰는 아님)
+
+판정:
+  PASS → 완료 보고
+  FAIL → Codex에 수정 지시 (1회 재시도)
+  재시도 FAIL → 사용자에게 문제 보고
+```
+
+### Step 4: 완료 보고
+
+```markdown
+## Autopilot 완료: {task}
+
+### 변경 사항
+- `{file1}` — {변경 요약}
+- `{file2}` — {변경 요약}
+
+### 검증
+- 완료 기준: {pass}/{total} 충족
+- 테스트: {pass}/{total} 통과
+- 검증: Claude ✓
+
+### 다음 단계 (선택)
+- {추가 권장 사항이 있으면}
+```
+
+## 토큰 예산
+
+| 단계 | 토큰 |
+|------|------|
+| Step 1 (파싱) | ~500 |
+| Step 2 (Codex 구현) | ~5K |
+| Step 3 (Claude 검증) | ~3K |
+| Step 4 (보고) | ~500 |
+| 재시도 (필요 시) | +4K |
+| **총합** | **~10K** |
+
+## 사용 예
+
+```
+/tfx-autopilot "이 함수에 입력 검증 추가해줘"
+/tfx-autopilot "ESLint 경고 전부 수정"
+/tfx-autopilot "알아서 해 — 이 TODO 코멘트 3개 구현"
+```

package/skills/tfx-autoresearch/SKILL.md

@@ -5,8 +5,7 @@ description: >
   주제를 받아 웹 검색 → 정보 수집 → 분석 → 구조화된 리서치 보고서 생성.
 triggers:
   - autoresearch
-  - 리서치
-  - 자동 리서치
+  - 자율 리서치
   - 웹 리서치
   - 조사해
   - 알아봐