@tienne/gestalt 0.5.1 → 0.7.0

package/README.md

@@ -3,349 +3,402 @@
 </p>
 
 <p align="center">
-  <strong>Gestalt psychology-driven AI development harness</strong><br/>
-  Transforms scattered requirements into structured, validated specifications through interactive interviews.
+  <strong>Gestalt — AI Development Harness</strong><br/>
+  Turn vague requirements into structured, executable plans — right inside Claude Code.
 </p>
 
 <p align="center">
-  <a href="https://github.com/tienne/gestalt/actions/workflows/ci.yml"><img src="https://github.com/tienne/gestalt/actions/workflows/ci.yml/badge.svg" alt="CI" /></a>
   <a href="https://www.npmjs.com/package/@tienne/gestalt"><img src="https://img.shields.io/npm/v/@tienne/gestalt" alt="npm version" /></a>
+  <a href="LICENSE"><img src="https://img.shields.io/badge/license-MIT-blue.svg" alt="MIT License" /></a>
+  <a href="https://github.com/tienne/gestalt/actions/workflows/ci.yml"><img src="https://github.com/tienne/gestalt/actions/workflows/ci.yml/badge.svg" alt="CI" /></a>
 </p>
 
+<p align="center">
+  <a href="./README.ko.md">한국어</a>
+</p>
+
+---
+
+## What is Gestalt?
+
+Gestalt is an MCP (Model Context Protocol) server that runs inside Claude Code. It conducts structured requirement interviews, generates a validated **Spec** (a JSON document capturing your goal, constraints, and acceptance criteria), and transforms that Spec into a dependency-aware execution plan — all without a separate API key.
+
+> **Prerequisites:** Node.js >= 20.0.0. Use `nvm install 22 && nvm use 22` if needed.
+
 ---
 
-## Prerequisites
+## See it in action
 
-- **Node.js >= 20.0.0** (22.x recommended)
+> **30 seconds to a structured execution plan** — no API key needed.
 
 ```bash
-node --version  # v20.x.x 이상이어야 합니다
+# Start a Gestalt-driven requirements interview
+/interview "user authentication system"
+
+# Once the interview is done, generate a structured Spec
+/spec
+
+# Transform the Spec into a validated execution plan and run it
+/execute
 ```
 
-Node 18 이하에서는 MCP SDK의 ES2024 기능 호환성 문제로 실행이 실패합니다. [nvm](https://github.com/nvm-sh/nvm) 사용 시:
+_(Demo coming soon.)_
 
-```bash
-nvm install 22
-nvm use 22
+---
+
+## Why Gestalt?
+
+Vague requirements are the primary source of implementation drift. When the goal isn't precise, Claude fills in gaps with assumptions — and those assumptions diverge from intent as the project grows.
+
+Gestalt addresses this at the source. Before any code is written, it runs a structured interview guided by **Gestalt psychology principles** to reduce ambiguity to a measurable threshold (≤ 0.2). The result is a **Spec**: a validated JSON document that serves as the single source of truth for planning and execution.
+
+### The Five Gestalt Principles
+
+- **Closure** — Finds what's missing; fills in implicit requirements
+- **Proximity** — Groups related features and tasks by domain
+- **Similarity** — Identifies repeating patterns across requirements
+- **Figure-Ground** — Separates MVP (figure) from nice-to-have (ground)
+- **Continuity** — Validates dependency chains; detects contradictions
+
+> "The whole is greater than the sum of its parts." — Aristotle
+
+### How does Passthrough Mode work?
+
+Gestalt runs as an **MCP server**. Claude Code acts as the LLM: Gestalt returns prompts and context, and Claude Code does the reasoning. The server makes no API calls.
+
+```
+You (in Claude Code)
+       │
+       ▼ /interview "topic"
+  Gestalt MCP Server
+  (returns context + prompts)
+       │
+       ▼
+  Claude Code executes the prompts
+  (generates questions, scores, plans)
+       │
+       ▼
+  Gestalt MCP Server
+  (validates, stores state, advances)
+       │
+       ▼ repeat until ambiguity ≤ 0.2
+  Final Spec → Execution Plan
 ```
 
+`ANTHROPIC_API_KEY` is not required. All LLM work is handled by Claude Code.
+
+---
+
 ## Installation
 
-### Claude Plugin (Recommended)
+### Option 1: Claude Code Plugin (Recommended)
 
-Claude Code 마켓플레이스를 통해 설치하면 MCP 서버, Skill, Agent, CLAUDE.md가 자동 구성된다.
+Bundles the MCP server, slash-command skills, Gestalt agents, and project context — pre-configured in a single install.
 
 ```bash
-# 1. 마켓플레이스 등록 (최초 1회)
+# Step 1: Add to marketplace (one-time setup)
 /plugin marketplace add tienne/gestalt
 
-# 2. 플러그인 설치
+# Step 2: Install the plugin
 /plugin install gestalt@gestalt
 ```
 
-플러그인 설치 시 포함되는 항목:
+What you get out of the box:
 
-| 항목 | 내용 |
-|------|------|
-| **MCP Tools** | `ges_interview`, `ges_generate_spec`, `ges_execute`, `ges_create_agent`, `ges_status` |
-| **Skills** | `/interview`, `/spec`, `/execute` (slash command) |
-| **Agents** | 5개 Gestalt Agent + 8개 Role Agent |
-| **CLAUDE.md** | 프로젝트 컨텍스트 및 MCP 사용 가이드 자동 주입 |
+| Item | Details |
+|------|---------|
+| **MCP Tools** | `ges_interview`, `ges_generate_spec`, `ges_execute`, `ges_create_agent`, `ges_agent`, `ges_status` |
+| **Slash Commands** | `/interview`, `/spec`, `/execute`, `/agent` |
+| **Agents** | 5 Gestalt pipeline agents + 9 Role agents + 3 Review agents |
+| **CLAUDE.md** | Project context and MCP usage guide auto-injected |
 
-설치 후 바로 사용 가능:
+> **Requires Node.js >= 20.0.0** — use [nvm](https://github.com/nvm-sh/nvm) if needed: `nvm install 22 && nvm use 22`
 
-```bash
-/interview "사용자 인증 시스템"    # 요구사항 인터뷰 시작
-/spec                             # 인터뷰 완료 후 Spec 생성
-/execute                          # Spec 기반 실행 계획 수립 및 실행
+---
+
+### Option 2: Claude Code Desktop
+
+Open your Claude Code Desktop settings and add to `settings.json` (or `claude_desktop_config.json`):
+
+```json
+{
+  "mcpServers": {
+    "gestalt": {
+      "command": "npx",
+      "args": ["-y", "@tienne/gestalt"]
+    }
+  }
+}
 ```
 
-### Claude Code MCP
+Restart Claude Code Desktop. The MCP tools become available immediately. Slash commands require the plugin or manual skills setup.
+
+---
+
+### Option 3: Claude Code CLI
 
 ```bash
+# Via the claude CLI
 claude mcp add gestalt -- npx -y @tienne/gestalt
 ```
 
-### npx (No Install)
+Or edit `~/.claude/settings.json` directly:
 
-```bash
-npx -y @tienne/gestalt
+```json
+{
+  "mcpServers": {
+    "gestalt": {
+      "command": "npx",
+      "args": ["-y", "@tienne/gestalt"]
+    }
+  }
+}
 ```
 
-### Global Install
+---
+
+## Usage: Full Pipeline
+
+### Step 1 — Interview
+
+Start with any topic. A single rough sentence is enough.
 
 ```bash
-npm install -g @tienne/gestalt
-gestalt
+/interview "I want to build a checkout flow with Stripe"
 ```
 
-## How It Works
+Gestalt conducts a multi-round interview. Each round targets a specific ambiguity dimension:
+
+- **Closure** — What's missing? What did you assume but not say?
+- **Proximity** — Which features belong together?
+- **Similarity** — Are there repeating patterns across requirements?
+- **Figure-Ground** — What's the core MVP vs. what's optional?
+- **Continuity** — Any contradictions or conflicts?
+
+The interview continues until the **ambiguity score reaches ≤ 0.2**:
 
 ```
-Interview → Spec → Execute → Evaluate → Evolve
-                                           ↓
-                              Lateral Thinking (resilience)
-                                           ↓
-                              Human Escalation (if exhausted)
+Round 1 → ambiguity: 0.72  (lots of unknowns)
+Round 4 → ambiguity: 0.45  (getting clearer)
+Round 8 → ambiguity: 0.19  ✓ ready for Spec
 ```
 
-### 1. Interview
+---
 
-Gestalt 원리가 각 라운드의 질문 방향을 가이드하여 **모호성 점수(ambiguity score)가 0.2 이하**로 떨어질 때까지 요구사항을 명확화한다.
+### Step 2 — Spec Generation
 
-- 프로젝트 유형 자동 감지 (greenfield / brownfield)
-- 5가지 Gestalt 원리 기반 동적 질문 선택 알고리즘
-- 차원별 가중치 기반 모호성 점수 산출
-- 모순 감지 시 Continuity 원리 자동 개입 + 페널티 부여
+When the ambiguity score reaches ≤ 0.2, run:
 
-### 2. Spec Generation
+```bash
+/spec
+```
 
-완료된 인터뷰에서 **구조화된 프로젝트 스펙**을 생성한다. 2-Call Passthrough 패턴으로 동작.
+Generates a structured **Spec** — a validated JSON document that drives the rest of the pipeline:
 
 ```
-Call 1: specContext 요청 → systemPrompt + specPrompt + 전체 인터뷰 라운드 반환
-Call 2: caller가 생성한 Spec JSON 제출 → Zod 검증 후 저장
+goal                → Clear, unambiguous project objective
+constraints         → Technical and business constraints
+acceptanceCriteria  → Measurable, verifiable success conditions
+ontologySchema      → Entity-relationship model (entities + relations)
+gestaltAnalysis     → Key findings per Gestalt principle
 ```
 
-Spec 구성요소: `goal`, `constraints`, `acceptanceCriteria`, `ontologySchema` (entities + relations), `gestaltAnalysis`
+---
+
+### Step 3 — Execute (Planning + Execution)
 
-### 3. Execute (Planning)
+Transforms the Spec into a dependency-aware execution plan and runs it:
 
-Gestalt 4원리를 **고정 순서**로 적용하여 실행 계획을 수립한다.
+```bash
+/execute
+```
 
-| Step | Principle | 역할 |
-|:---:|---|---|
-| 1 | **Figure-Ground** | AC를 essential(figure) / supplementary(ground)로 분류 |
-| 2 | **Closure** | AC를 원자적 태스크로 분해 (암묵적 하위 태스크 포함) |
-| 3 | **Proximity** | 관련 태스크를 도메인/기능 영역별로 그룹핑 |
-| 4 | **Continuity** | DAG 검증 — 순환 의존, 그룹 간 충돌 확인, 위상 정렬 |
+**Planning phase** applies 4 Gestalt principles in a fixed sequence:
 
-결과물: `ExecutionPlan` (classifiedACs, atomicTasks, taskGroups, dagValidation, criticalPath)
+| Step | Principle | What it does |
+|:---:|-----------|-------------|
+| 1 | **Figure-Ground** | Classifies acceptance criteria (ACs) as critical (figure) vs. supplementary (ground) |
+| 2 | **Closure** | Decomposes ACs into atomic tasks, including implicit ones |
+| 3 | **Proximity** | Groups related tasks by domain into logical task groups |
+| 4 | **Continuity** | Validates the dependency DAG — no cycles, clear topological order |
 
-### 4. Execute (Execution) + Drift Detection
+**Execution phase** runs tasks in topological order. After each task, **drift detection** checks alignment with the Spec:
 
-위상 정렬 순서대로 태스크를 실행하며, **Similarity 원리**를 활용하여 유사 패턴의 완료된 태스크를 참조 컨텍스트로 제공한다.
+- 3-dimensional score: Goal (50%) + Constraint (30%) + Ontology (20%)
+- Jaccard similarity-based measurement
+- Auto-triggers a retrospective when drift exceeds threshold
 
-매 태스크 완료 시 **Drift Detection**이 자동 실행된다:
+---
 
-- **3차원 가중합**: Goal(50%) + Constraint(30%) + Ontology(20%)
-- Jaccard 기반 키워드 유사도 측정
-- Threshold(기본 0.3) 초과 시 `retrospectiveContext` 반환
+### Step 4 — Evaluate
 
-### 5. Evaluate (2-Stage)
+Execution triggers a 2-stage evaluation:
 
-| Stage | 이름 | 방식 | 실패 시 |
-|:---:|---|---|---|
-| 1 | **Structural** | lint → build → test 자동 검증 | Contextual 스킵 (short-circuit) |
-| 2 | **Contextual** | LLM 기반 완료 조건(AC) 검증 + goalAlignment | Evolution Loop 진입 |
+| Stage | Method | On failure |
+|:---:|-------|-----------|
+| 1 | **Structural** — runs lint → build → test | Short-circuits; skips Stage 2 |
+| 2 | **Contextual** — LLM validates each AC + goal alignment | Enters Evolution Loop |
 
-성공 조건: `overallScore ≥ 0.85` && `goalAlignment ≥ 0.80`
+**Success condition:** `score ≥ 0.85` AND `goalAlignment ≥ 0.80`
 
-### 6. Evolution Loop
+---
+
+### Step 5 — Evolve
 
-Evaluate 실패 시 3가지 경로로 자동 복구를 시도한다:
+When evaluation fails, the Evolution Loop engages. Three recovery flows are available:
 
-**Flow A — Structural Fix** (structural 실패 시)
+**Flow A — Structural Fix** (when lint/build/test fails)
 ```
-evolve_fix → fixTasks 제출 → re-evaluate
+evolve_fix → submit fix tasks → re-evaluate
 ```
 
-**Flow B — Contextual Evolution** (contextual 점수 미달 시)
+**Flow B — Contextual Evolution** (when AC score is too low)
 ```
-evolve → evolve_patch(specPatch) → evolve_re_execute(impacted tasks) → re-evaluate
+evolve → patch Spec (ACs/constraints) → re-execute impacted tasks → re-evaluate
 ```
 
-Spec Patch 범위: 완료 조건(AC) + constraints 자유 수정, ontology 추가/변경만, **goal 변경 금지**
-
-**Flow C — Lateral Thinking** (stagnation 감지 시 자동 분기)
+Spec patch scope: ACs and constraints are freely editable; ontology can be extended; **goal is immutable**.
 
-evolve 호출 시 stagnation이 감지되면 즉시 종료하는 대신 **Lateral Thinking Persona**로 자동 분기한다. (아래 섹션 참조)
+**Flow C — Lateral Thinking** (when stagnation is detected)
 
-**종료 조건:**
+Gestalt rotates through lateral thinking personas rather than terminating:
 
-| 조건 | 설명 |
-|---|---|
-| `success` | score ≥ 0.85, goalAlignment ≥ 0.80 |
-| `stagnation` | 2회 연속 delta < 0.05 |
-| `oscillation` | 2회 연속 점수 진동 (up↔down) |
-| `hard_cap` | structural 3회 + contextual 3회 실패 |
-| `caller` | 사용자 수동 종료 |
-| `human_escalation` | Lateral Thinking 4개 persona 모두 실패 |
-
-## Gestalt Principles
+| Stagnation Pattern | Persona | Strategy |
+|--------------------|---------|---------|
+| Hard cap hit | **Multistability** | See from a different angle |
+| Oscillating scores | **Simplicity** | Strip down and converge |
+| No progress (no drift) | **Reification** | Fill in what's missing |
+| Diminishing returns | **Invariance** | Replicate what worked |
 
-Each letter in the logo represents a core Gestalt principle used throughout the system:
+When all 4 personas are exhausted, the session terminates with **Human Escalation** — a structured list of actionable suggestions for manual resolution.
 
-| Principle | Role in Gestalt |
-|-----------|----------------|
-| **Closure** | Finds implicit requirements that aren't explicitly stated |
-| **Proximity** | Groups related requirements and tasks by domain |
-| **Continuation** | Validates dependency chains and execution order (DAG) |
-| **Similarity** | Identifies repeating patterns across requirements |
-| **Figure & Ground** | Separates core (figure) from supporting (ground) requirements |
+**Termination conditions:**
 
-> For a deep dive into how each principle maps to interview strategies, ambiguity scoring, and the dynamic principle selection algorithm, see [Gestalt 5원리 상세 문서](docs/gestalt-principles.md).
+| Condition | Trigger |
+|-----------|---------|
+| `success` | score ≥ 0.85 AND goalAlignment ≥ 0.80 |
+| `stagnation` | 2 consecutive rounds with delta < 0.05 |
+| `oscillation` | 2 consecutive score reversals |
+| `hard_cap` | 3 structural + 3 contextual failures |
+| `caller` | Manual termination |
+| `human_escalation` | All 4 lateral personas exhausted |
 
-## Lateral Thinking Personas
+---
 
-Evolution Loop에서 stagnation이 감지되면 패턴을 분류하고 매칭되는 persona를 활성화한다:
+### Step 6 — Code Review
 
-| Stagnation Pattern | Persona | Strategy |
-|--------------------|---------|----------|
-| Spinning (hard cap) | **Multistability** | See from a different angle |
-| Oscillation | **Simplicity** | Simplify and converge |
-| No drift | **Reification** | Fill in missing pieces |
-| Diminishing returns | **Invariance** | Replicate success patterns |
+When evolution finishes, code review starts automatically:
 
-4개 persona를 순차적으로 시도한다. 모두 실패하면 **Human Escalation**을 트리거하여 사용자에게 actionable suggestions를 제공한다.
+```
+review_start → agents submit perspectives → consensus → auto-fix
+```
 
-## Role Agent System
+9 built-in **Role Agents** provide multi-perspective review:
 
-**8개의 내장 Role Agent**가 Execute Phase에서 다중 관점을 제공한다.
+| Agent | Domain |
+|-------|--------|
+| `architect` | System design, scalability |
+| `frontend-developer` | UI, React, accessibility |
+| `backend-developer` | API, database, server |
+| `devops-engineer` | CI/CD, infrastructure, monitoring |
+| `qa-engineer` | Testing, quality, automation |
+| `designer` | UX/UI, design systems |
+| `product-planner` | Roadmap, user stories, metrics |
+| `researcher` | Analysis, data, benchmarks |
+| `technical-writer` | Documentation, API docs, guides, README |
 
-| Agent | Domain | Description |
-|---|---|---|
-| `architect` | system design, scalability | 시스템 아키텍처 전문가 |
-| `backend-developer` | api, database, server | 백엔드 개발 전문가 |
-| `frontend-developer` | ui, react, css, accessibility | 프론트엔드 개발 전문가 |
-| `designer` | ux, ui, design system | UX/UI 디자인 전문가 |
-| `devops-engineer` | ci/cd, infra, monitoring | DevOps/인프라 전문가 |
-| `product-planner` | roadmap, user story, metrics | 제품 기획 전문가 |
-| `qa-engineer` | testing, quality, automation | QA/테스트 전문가 |
-| `researcher` | analysis, data, benchmarks | 리서치 및 분석 전문가 |
+3 built-in **Review Agents** run focused code analysis:
 
-### 동작 방식
+| Agent | Focus |
+|-------|-------|
+| `security-reviewer` | Injection, XSS, auth vulnerabilities, secrets |
+| `performance-reviewer` | Memory leaks, N+1 queries, bundle size, async |
+| `quality-reviewer` | Readability, SOLID, error handling, DRY |
 
-Execute Phase에서 각 태스크마다 2단계로 진행된다:
+Use any agent outside the pipeline with `/agent`:
 
-1. **Role Match** — 태스크의 도메인 키워드와 매칭되는 Role Agent를 동적으로 선택
-2. **Role Consensus** — 선택된 Agent들의 관점을 수집하고, 충돌을 해소하여 통합 가이던스 생성
+```bash
+# List all available agents
+/agent
 
-### Custom Agent 생성
+# Run a specific agent on any task
+/agent architect "review the module boundaries in this codebase"
+/agent security-reviewer "check this authentication code for vulnerabilities"
+/agent technical-writer "write a README for this module"
+```
 
-`ges_create_agent` 도구로 **인터뷰 결과를 기반으로 커스텀 Role Agent를 생성**할 수 있다:
+Generate custom Role Agents from interview results:
 
 ```
-# 1. 인터뷰 완료 후 agent creation context 요청
-ges_create_agent({ action: "start", sessionId: "<id>" })
-→ systemPrompt + creationPrompt + agentMdSchema 반환
+# Step 1: Get agent creation context
+ges_create_agent  →  action: "start", sessionId: "<id>"
+                  →  returns agentContext (systemPrompt, creationPrompt, schema)
 
-# 2. AGENT.md 콘텐츠 생성 후 제출
-ges_create_agent({ action: "submit", sessionId: "<id>", agentContent: "---\nname: ..." })
-→ agents/{name}/AGENT.md 파일 생성
+# Step 2: Submit the generated AGENT.md content
+ges_create_agent  →  action: "submit", sessionId: "<id>", agentContent: "..."
+                  →  creates agents/{name}/AGENT.md
 ```
 
-커스텀 Agent는 `agents/` 디렉토리에 저장되며, 동일 이름의 내장 Agent를 override할 수 있다.
-
-### AGENT.md 포맷
-
-```yaml
----
-name: security-expert        # 에이전트 이름 (kebab-case)
-tier: standard                # frugal | standard | frontier
-pipeline: execute             # interview | spec | execute | evaluate
-role: true                    # Role Agent 필수
-domain: ["oauth", "jwt"]      # 도메인 키워드 목록
-description: "보안 전문가"     # 한줄 설명
 ---
 
-System prompt 내용 (Markdown)
-```
-
-## MCP Tools
-
-5 MCP tools exposed in passthrough mode (no API key required):
+### Alternative: CLI Direct Mode
 
-| Tool | Description |
-|---|---|
-| `ges_interview` | Gestalt-driven requirement interview (start, respond, score, complete) |
-| `ges_generate_spec` | Generate a structured Spec from completed interview (2-Call) |
-| `ges_execute` | Execute Spec via Gestalt pipeline (plan, execute, evaluate, evolve, lateral thinking, role agents) |
-| `ges_create_agent` | Create custom Role Agent from completed interview (2-Call: start, submit) |
-| `ges_status` | Check session status |
+Requires `ANTHROPIC_API_KEY`. Runs Gestalt directly from the terminal without Claude Code:
 
-## Skill System
+```bash
+# Start an interactive interview
+npx @tienne/gestalt interview "my topic"
 
-`skills/` 디렉토리의 SKILL.md 파일로 각 파이프라인 단계의 동작을 정의한다.
+# Record the session as a GIF
+npx @tienne/gestalt interview "my topic" --record
 
-| Skill | Description |
-|---|---|
-| `interview` | 인터뷰 프로세스 — 4단계 (start → iterate → score → complete) |
-| `spec` | Spec 생성 프로세스 — 2-Call Passthrough 패턴 |
-| `execute` | 실행 계획 수립 — 4단계 Planning + Execution + Evaluation |
+# Generate Spec from a completed session
+npx @tienne/gestalt spec <session-id>
 
-### SKILL.md 포맷
+# List all sessions
+npx @tienne/gestalt status
 
-```yaml
----
-name: interview
-version: "1.0.0"
-description: "Gestalt-driven interview"
-triggers: ["interview", "clarify requirements"]
-inputs:
-  topic: { type: string, required: true }
-outputs: [session, ambiguityScore]
----
+# Generate gestalt.json config
+npx @tienne/gestalt setup
 
-Skill documentation (Markdown)
+# Start the MCP server manually
+npx @tienne/gestalt serve
 ```
 
-커스텀 Skill은 `skills/` 디렉토리에 추가하면 **chokidar hot-reload**로 자동 로드된다.
+#### Recording an interview session
 
-## Multi-provider LLM
+Add `--record` (or `-r`) to capture the terminal session as a GIF:
 
-Agent System의 **FiguralRouter**가 에이전트의 tier에 따라 다른 LLM provider를 라우팅한다:
+```bash
+npx @tienne/gestalt interview "my topic" --record
+```
 
-| Tier | Provider | Model Example | Use Case |
-|---|---|---|---|
-| `frugal` | OpenAI | gpt-4o-mini | 반복적/저비용 태스크 (proximity-worker) |
-| `standard` | Anthropic | claude-sonnet | 일반 태스크 (closure-completer, ground-mapper) |
-| `frontier` | Anthropic | claude-opus | 고품질 판단 (continuity-judge) |
+When the interview completes, a GIF is written to the current directory:
 
-이를 위해 `@anthropic-ai/sdk`와 `openai` 두 SDK가 모두 포함되어 있다. Passthrough 모드에서는 caller가 LLM 호출을 하므로 이 라우팅이 적용되지 않는다.
+```
+user-auth-interview-20260327.gif
+```
 
-## Event Sourcing
+The filename is generated by the LLM from the interview topic (kebab-case) plus a `YYYYMMDD` date stamp. No external binaries are required — recording uses `gifencoder` and `jimp` only.
 
-모든 세션 활동은 **SQLite WAL 모드 이벤트 스토어**에 기록된다.
+**Resuming an interrupted session:** If a session is interrupted mid-recording, restarting the same session automatically detects the `.frames` file and continues where it left off. Temporary frame data is stored at `.gestalt/recordings/{sessionId}.frames` and deleted once the GIF is written.
 
-- Interview 이벤트: session started/completed, question asked, response recorded, ambiguity scored
-- Execute 이벤트: planning steps, task completed, drift measured
-- Evaluate 이벤트: structural/contextual results, short-circuit
-- Evolution 이벤트: spec patched, lateral thinking, human escalation
-- Agent 이벤트: role match/consensus, agent created
+---
 
-완전한 audit trail로 세션 히스토리 재생 및 사후 분석이 가능하다.
+## Configuration
 
-## CLI Commands
+Generate a `gestalt.json` with IDE autocompletion support:
 
 ```bash
-gestalt                    # Start MCP server (default)
-gestalt serve              # Start MCP server (explicit)
-gestalt interview "topic"  # Interactive interview
-gestalt spec <session-id>  # Generate Spec from interview
-gestalt status             # List all sessions
-gestalt setup              # Generate gestalt.json config file
-gestalt monitor            # TUI dashboard (real-time session monitoring)
+npx @tienne/gestalt setup
 ```
 
-### TUI Dashboard
-
-`gestalt monitor` 명령으로 터미널 기반 대시보드를 실행한다:
-
-- **Session List** — 진행 중인 세션 목록 및 상태
-- **Interview Monitor** — 실시간 인터뷰 진행 상황 + Gestalt 원리 표시
-- **Evolution Tracker** — Evolution Loop 진행, Drift 수치 게이지
-- **Task DAG Tree** — 태스크 의존성 관계 시각화
-- **Spec Viewer** — 생성된 Spec 조회
-- **Event Log** — 실시간 이벤트 로그
-
-## Configuration
-
-Run `gestalt setup` to generate a `gestalt.json` configuration file with IDE autocompletion support.
-
 ```json
 {
   "$schema": "./node_modules/@tienne/gestalt/schemas/gestalt.schema.json",
+  "llm": {
+    "model": "claude-sonnet-4-20250514"
+  },
   "interview": {
     "ambiguityThreshold": 0.2,
     "maxRounds": 10
@@ -358,85 +411,70 @@ Run `gestalt setup` to generate a `gestalt.json` configuration file with IDE aut
 }
 ```
 
-Configuration is loaded with the following priority (highest → lowest):
-
-1. Code overrides (`loadConfig(overrides)`)
-2. Shell environment variables (`GESTALT_*`)
-3. `.env` file
-4. `gestalt.json`
-5. Built-in defaults
+**Config priority** (highest → lowest): code overrides → shell env vars → `.env` → `gestalt.json` → built-in defaults
 
-Invalid values trigger a warning and fall back to defaults.
+Invalid values emit a warning and fall back to defaults.
 
 ### Environment Variables
 
-| Variable | Config Path | Default |
-|---|---|---|
-| `ANTHROPIC_API_KEY` | `llm.apiKey` | `""` |
-| `GESTALT_MODEL` | `llm.model` | `claude-sonnet-4-20250514` |
-| `GESTALT_AMBIGUITY_THRESHOLD` | `interview.ambiguityThreshold` | `0.2` |
-| `GESTALT_MAX_ROUNDS` | `interview.maxRounds` | `10` |
-| `GESTALT_DRIFT_THRESHOLD` | `execute.driftThreshold` | `0.3` |
-| `GESTALT_EVOLVE_SUCCESS_THRESHOLD` | `execute.successThreshold` | `0.85` |
-| `GESTALT_EVOLVE_GOAL_ALIGNMENT_THRESHOLD` | `execute.goalAlignmentThreshold` | `0.80` |
-| `GESTALT_DB_PATH` | `dbPath` | `~/.gestalt/events.db` |
-| `GESTALT_SKILLS_DIR` | `skillsDir` | `skills` |
-| `GESTALT_AGENTS_DIR` | `agentsDir` | `agents` |
-| `GESTALT_LOG_LEVEL` | `logLevel` | `info` |
-
-## Operating Modes
-
-Gestalt supports two operating modes:
-
-| Mode | API Key | LLM Calls | Use Case |
-|---|---|---|---|
-| **Passthrough** (default) | Not needed | Caller (e.g., Claude Code) | MCP 통합 환경에서 권장 |
-| **Provider-backed** | `ANTHROPIC_API_KEY` 필요 | Gestalt가 직접 호출 | 독립 실행 (CLI 등) |
+| Variable | Config Path | Default | Description |
+|----------|-------------|---------|-------------|
+| `ANTHROPIC_API_KEY` | `llm.apiKey` | `""` | Required only for CLI direct mode |
+| `GESTALT_MODEL` | `llm.model` | `claude-sonnet-4-20250514` | LLM model (provider-backed mode) |
+| `GESTALT_AMBIGUITY_THRESHOLD` | `interview.ambiguityThreshold` | `0.2` | Interview completion threshold |
+| `GESTALT_MAX_ROUNDS` | `interview.maxRounds` | `10` | Max interview rounds |
+| `GESTALT_DRIFT_THRESHOLD` | `execute.driftThreshold` | `0.3` | Task drift detection threshold |
+| `GESTALT_EVOLVE_SUCCESS_THRESHOLD` | `execute.successThreshold` | `0.85` | Evolution success score |
+| `GESTALT_EVOLVE_GOAL_ALIGNMENT_THRESHOLD` | `execute.goalAlignmentThreshold` | `0.80` | Goal alignment threshold |
+| `GESTALT_DB_PATH` | `dbPath` | `~/.gestalt/events.db` | SQLite event store path |
+| `GESTALT_SKILLS_DIR` | `skillsDir` | `skills` | Custom skills directory |
+| `GESTALT_AGENTS_DIR` | `agentsDir` | `agents` | Custom agents directory |
+| `GESTALT_LOG_LEVEL` | `logLevel` | `info` | Log level (`debug`/`info`/`warn`/`error`) |
 
-### Passthrough Mode
-
-Gestalt runs in passthrough mode by default: it returns prompts and context to the calling LLM (e.g., Claude Code) instead of making its own API calls. No `ANTHROPIC_API_KEY` needed.
-
-### Provider-backed Mode
-
-`ANTHROPIC_API_KEY`를 설정하면 Gestalt가 직접 LLM 호출을 수행한다. CLI 명령어(`gestalt interview`, `gestalt spec`)는 이 모드에서 동작한다.
+---
 
-### N-Call Passthrough Pattern
+## Architecture
 
-All MCP tools follow a consistent **N-Call Passthrough Pattern**:
+![Gestalt Architecture](./docs/architecture.png)
+_(Diagram coming soon.)_
 
 ```
-Call 1: Context 요청 → systemPrompt + actionPrompt 반환
-Call N: Caller가 LLM으로 결과 생성 → 제출 및 검증
+Claude Code (you)
+     │
+     ▼  MCP / stdio transport
+┌──────────────────────────────────┐
+│        Gestalt MCP Server        │
+│                                  │
+│  Interview Engine                │
+│  ├─ GestaltPrincipleSelector     │
+│  ├─ AmbiguityScorer              │
+│  └─ SessionManager               │
+│                                  │
+│  Spec Generator                  │
+│  └─ PassthroughSpecGenerator     │
+│                                  │
+│  Execute Engine                  │
+│  ├─ DAG Validator                │
+│  ├─ DriftDetector                │
+│  ├─ EvaluationEngine             │
+│  └─ ExecuteSessionManager        │
+│                                  │
+│  Resilience Engine               │
+│  ├─ StagnationDetector           │
+│  ├─ LateralThinkingPersonas      │
+│  └─ HumanEscalation              │
+│                                  │
+│  Agent System                    │
+│  ├─ RoleAgentRegistry            │
+│  ├─ RoleMatchEngine              │
+│  └─ RoleConsensusEngine          │
+│                                  │
+│  EventStore (SQLite WAL)         │
+└──────────────────────────────────┘
 ```
 
-| Tool | Pattern | Calls |
-|---|---|---|
-| `ges_interview` | Multi-Call | start → (respond + score) × N → complete |
-| `ges_generate_spec` | 2-Call | specContext 요청 → spec JSON 제출 |
-| `ges_execute` | Multi-Call | plan (4-step) → execute → evaluate (2-stage) → evolve |
-| `ges_create_agent` | 2-Call | agentContext 요청 → agentContent 제출 |
-
-## Benchmarks
-
-`pnpm bench`로 파이프라인 벤치마크를 실행할 수 있다. 3개의 시나리오가 Interview → Spec → Execute → Evaluate 전체 파이프라인을 통과한다.
-
-```bash
-pnpm bench                          # 전체 시나리오 실행
-pnpm bench -s auth-system           # 특정 시나리오만 실행
-pnpm bench -o benchmarks/results    # 결과 출력 디렉토리 지정
-```
-
-### Latest Results
-
-| Scenario | Tasks | Completion | Score | Goal Alignment | ACs |
-|----------|-------|------------|-------|----------------|-----|
-| auth-system | 7 | 100% | 0.92 | 0.90 | 5/5 |
-| dashboard | 6 | 100% | 0.88 | 0.85 | 4/4 |
-| api-gateway | 8 | 100% | 0.95 | 0.92 | 5/5 |
-
-**Average Score: 0.92 | Average Goal Alignment: 0.89**
+---
 
 ## License
 
-MIT
+MIT © [tienne](https://github.com/tienne)