hyacinth-ai 0.9.6

package/dist/world-engine/types.d.ts

@@ -0,0 +1,202 @@
+/** 环境 —— WorldTicker 独占写 */
+export interface Ambient {
+    /** 世界内时间（ISO 字符串） */
+    time: string;
+    /** 时段（清晨/上午/下午/晚上…），由 time 推导。世界时间经 timeScale 折算、与现实时钟脱钩，模型无法凭数字判断昼夜，故显式给出 */
+    period: string;
+    weather: string;
+    season: string;
+    temperature: number;
+}
+/** 物品的空间相对位置：object 在 anchor 的 relation 处（如 花瓶 在 柜子 上） */
+export interface ObjectPlacement {
+    object: string;
+    /** 关系：上/里/下/旁边/附近 等 */
+    relation: string;
+    /** 参照物（另一件物品或家具，如"柜子""床"） */
+    anchor: string;
+}
+/** 地点 —— WorldAgent 增长/精修 */
+export interface WorldLocation {
+    desc: string;
+    /** 地点内的物品 */
+    objects: string[];
+    /** 物品的相对摆放（花瓶在柜子上、柜子在床旁），由 WorldAgent 维护 */
+    layout?: ObjectPlacement[];
+    /** 临近可见地点 id（读取时只带一层摘要，不再展开——防无限递归） */
+    visible: string[];
+    /** 可移动到的地点 id */
+    connects: string[];
+    /** 主人（角色名，如"柔柔"）。为空表示公共场所。主人在此地是主人不是客人 */
+    owner?: string;
+}
+/**
+ * NPC —— desc/persona/home/roaming/mobility 由 WorldAgent 创建后只读；
+ * location 由 WorldTicker 概率放置；state 由 WorldAgent 更新。
+ */
+export interface WorldNpc {
+    desc: string;
+    /** 别名/其他称呼（昵称、头衔、全名等），如 徐青 的 aliases=[小青, 蛋糕店老板]。由 WorldAgent 维护 */
+    aliases?: string[];
+    /** 种族/物种（如"兔子""天马"）——属性，非身份；同族不同个体各建各的记录。由 WorldAgent 维护 */
+    race?: string;
+    /** 认知留痕：这个 NPC 的身份是如何被逐渐了解的（初见→得知名字/性格/住处…）。由 WorldAgent 维护 */
+    note?: string;
+    /** 人设，喂给 NPC 回应生成 */
+    persona: string;
+    /** 常驻地点 id（可多个） */
+    home: string[];
+    /** 0..1，离开 home 去邻近地点的概率 */
+    roaming: number;
+    /** anchored=基本不动 / free=会到处逛 */
+    mobility: 'anchored' | 'free';
+    /** 当前实际位置（Ticker 写） */
+    location: string;
+    /** 当前状态描述（Agent 写） */
+    state: string;
+}
+/**
+ * 社会关系（有向边）。可连接任意角色：主角 / 陪伴角色 / NPC。
+ * 语义："to 是 from 的 <type>"。对称关系（朋友/恋人）记一条即可；
+ * 有方向的（宠物：安吉尔是柔柔的宠物 → from=柔柔 to=安吉尔 type=宠物）方向有意义。
+ * 由 observe agent 依对话判断更新——不随时间自动变化。
+ */
+export interface Relationship {
+    from: string;
+    to: string;
+    /** 关系类型：陌生人 / 朋友 / 好友 / 恋人 / 家人 / 宠物 / 主人 / 同事 ... */
+    type: string;
+    /** 可选补充说明（如"刚认识，还有些拘谨"） */
+    note?: string;
+}
+/**
+ * 角色身份的动态补充（旁路随对话逐渐"认识"一个角色时积累的信息）。
+ * 以角色【规范名】为键（个体），与配置身份叠加。由 observe agent 依对话演变。
+ *
+ * 关键：race（种族）是"属性"，不是"身份键"——同种族的不同个体各有各的记录，
+ * 绝不因同族而合并。别名（aliases）只统一"同一个体的多个称呼"，从不跨个体。
+ */
+export interface CharacterInfo {
+    /** 种族/物种（如"天马""独角兽"）。属性，非身份键 */
+    race?: string;
+    /** 旁路学到的其他称呼（名字/昵称/外号/全名）；与配置 aliases 取并集，都指同一个体 */
+    aliases?: string[];
+    /** 认知留痕：这个角色的身份是如何被逐渐了解的（初见只知种族 → 第N轮得知名字 → 后来知道昵称…） */
+    note?: string;
+}
+/**
+ * 声明式自定义动态类型：让 LLM 用受限的规则（而非代码）定义新的动态对象种类，
+ * 框架按规则演化。用于内置 snow/plant/timer 覆盖不了的东西（融化的冰淇淋、燃烧的蜡烛…）。
+ */
+export interface KindSpec {
+    kind: string;
+    desc?: string;
+    /** amount 每"世界小时"的基础变化率（正=增长/积累，负=衰减/消耗） */
+    ratePerHour: number;
+    /** 温度影响：有效速率 += tempFactor × 当前温度（如冰淇淋 tempFactor<0，越热化越快） */
+    tempFactor?: number;
+    /** amount 上限，默认 100 */
+    max?: number;
+    /** amount ≤ 此值则移除该对象（如融化殆尽消失）；不填=不自动移除 */
+    removeAtOrBelow?: number;
+    /** 初始 amount；不填时：衰减型默认满值，增长型默认 0 */
+    start?: number;
+    /** 阶段标签，按 amount 阈值从高到低匹配 */
+    phases?: Array<{
+        atOrAbove: number;
+        label: string;
+    }>;
+}
+/**
+ * 框架维护的动态对象：由 Ticker 按世界时间+环境自动演化（雪/植物/计时器等）。
+ * LLM 负责"造出来"（world_spawn_sim），框架负责演化。见 sim.ts。
+ */
+export interface SimObject {
+    id: string;
+    /** 行为类型：snow / plant / timer / …（决定演化规则） */
+    kind: string;
+    /** 所在地点 id */
+    location: string;
+    desc?: string;
+    /** 通用动态量 0-100：雪量 / 植物成熟度 / 计时进度 */
+    amount: number;
+    /** 当前阶段标签：积雪/融化中；生长中/成熟；进行中/完成 */
+    phase?: string;
+    /** kind 相关参数（如 plant 的 multiHarvest/growthRate、timer 的 durationHours） */
+    params?: Record<string, unknown>;
+}
+/** 动态状态 —— WorldAgent 独占写 */
+export interface WorldState {
+    /** 用户/主角当前所在地点 id */
+    location: string;
+    /** 陪伴角色（如柔柔）当前所在地点 id。narrate 服务于主 LLM，故以此作为"她的场景"；多数时候与 location 相同（在一起） */
+    companionLocation: string;
+    /** 对基础环境的临时变更（如 "灯关了"） */
+    sceneOverrides: Record<string, string>;
+    /** 最近发生的事（有上限，见 WorldStore.MAX_EVENTS） */
+    recentEvents: string[];
+    /** 主角/陪伴角色的动态身份补充（旁路随对话学到的种族/别名/认知留痕），以规范名为键，与配置身份叠加 */
+    characters: Record<string, CharacterInfo>;
+}
+export interface WorldMeta {
+    id: string;
+    name: string;
+    createdAt: string;
+}
+/** 一个完整的世界 */
+export interface World {
+    meta: WorldMeta;
+    ambient: Ambient;
+    locations: Record<string, WorldLocation>;
+    npcs: Record<string, WorldNpc>;
+    /** 社会关系图（有向边），由 observe agent 依对话演变 */
+    relationships: Relationship[];
+    /** 框架维护的动态对象（雪/植物/计时器等） */
+    simObjects: Record<string, SimObject>;
+    /** LLM 声明的自定义动态类型（内置类型之外） */
+    customKinds: Record<string, KindSpec>;
+    state: WorldState;
+}
+/** readEnvironment() 的返回：当前地点的环境快照（供旁白化） */
+export interface EnvironmentSnapshot {
+    /** 场景地点 = 陪伴角色所在地（narrate 服务于她） */
+    location: string;
+    /** 用户/主角当前所在地点（与 location 相同则两人同场） */
+    userLocation: string;
+    /** 用户/主角是否也在此场景（与陪伴角色同地） */
+    userPresent: boolean;
+    desc: string;
+    /** 地点主人（角色名）；为空表示公共场所 */
+    owner?: string;
+    objects: string[];
+    /** 物品相对摆放（花瓶在柜子上…） */
+    layout: ObjectPlacement[];
+    /** 对基础环境的临时变更 */
+    sceneOverrides: Record<string, string>;
+    ambient: Ambient;
+    /** 临近可见地点（单层摘要） */
+    visible: Array<{
+        id: string;
+        desc: string;
+    }>;
+    /** 当前在场的 NPC（含其主标识 id） */
+    npcs: Array<WorldNpc & {
+        id: string;
+    }>;
+    /** 与当前场景相关的社会关系（涉及主角/陪伴/在场 NPC） */
+    relationships: Relationship[];
+    /** 主角/陪伴角色的动态身份补充（种族/别名/留痕），供旁白按规范名叠加渲染 */
+    characters: Record<string, CharacterInfo>;
+    /** 当前地点的动态对象（雪/植物/计时器等） */
+    simObjects: SimObject[];
+    /** 最近事件（截断） */
+    recentEvents: string[];
+}
+/**
+ * 由小时（0-23）推导"时段"标签。世界时间经 timeScale 折算、与现实时钟脱钩，
+ * 模型无法凭 "09:50" 这样的数字判断是上午还是晚上，故显式给出语义标签。
+ */
+export declare function periodOf(hour: number): string;
+/** 创建一个空世界（world 从对话中生长，初始无地点/NPC） */
+export declare function createEmptyWorld(id: string, name: string, now: Date): World;
+//# sourceMappingURL=types.d.ts.map

package/dist/world-engine/types.js

@@ -0,0 +1,56 @@
+// ============================================================
+// world-engine / types — 世界数据结构
+// ============================================================
+//
+// world.json 的 schema。世界从对话中"长出来"：初始为空，
+// 后置 WorldAgent 根据对话逐渐创建地点/物品/NPC。
+//
+// 写入者所有权（关键约束，由工具层强制）：
+//   ambient          → WorldTicker 独占
+//   locations        → WorldAgent 创建/精修
+//   npcs 静态字段     → WorldAgent 创建后只读（desc/persona/home/roaming/mobility）
+//   npcs.location    → WorldTicker 放置
+//   state.*          → WorldAgent 独占（当前位置 / 场景变更 / 最近事件 / 社会关系等）
+// ============================================================
+/**
+ * 由小时（0-23）推导"时段"标签。世界时间经 timeScale 折算、与现实时钟脱钩，
+ * 模型无法凭 "09:50" 这样的数字判断是上午还是晚上，故显式给出语义标签。
+ */
+export function periodOf(hour) {
+    if (hour < 5)
+        return '凌晨';
+    if (hour < 8)
+        return '清晨';
+    if (hour < 11)
+        return '上午';
+    if (hour < 13)
+        return '中午';
+    if (hour < 17)
+        return '下午';
+    if (hour < 19)
+        return '傍晚';
+    if (hour < 23)
+        return '晚上';
+    return '深夜';
+}
+/** 创建一个空世界（world 从对话中生长，初始无地点/NPC） */
+export function createEmptyWorld(id, name, now) {
+    const iso = now.toISOString();
+    return {
+        meta: { id, name, createdAt: iso },
+        ambient: { time: iso, period: periodOf(now.getHours()), weather: '晴', season: '春', temperature: 20 },
+        locations: {},
+        npcs: {},
+        relationships: [],
+        simObjects: {},
+        customKinds: {},
+        state: {
+            location: '',
+            companionLocation: '',
+            sceneOverrides: {},
+            recentEvents: [],
+            characters: {},
+        },
+    };
+}
+//# sourceMappingURL=types.js.map

package/docs/api.md

@@ -0,0 +1,180 @@
+# HTTP API Reference
+
+The Agent framework provides a REST API server via `agent serve`.
+
+## Starting the Server
+
+```bash
+agent serve --port 3000 --provider anthropic --model claude-sonnet-4-20250514
+```
+
+Options:
+
+| Flag | Default | Description |
+|------|---------|-------------|
+| `--port` | `3000` | Server port |
+| `--provider` | auto | Provider type (anthropic/openai/deepseek/local/...) |
+| `--model` | auto | Model name |
+| `--max-turns` | `40` | Max turns per chat request |
+| `--max-context` | `200000` | Max context tokens |
+
+All endpoints are prefixed with `/api`.
+
+---
+
+## Health Check
+
+```
+GET /api/health
+```
+
+**Response:**
+
+```json
+{
+  "status": "ok",
+  "version": "1.0.0"
+}
+```
+
+---
+
+## Chat Completion
+
+```
+POST /api/chat
+```
+
+Execute a single prompt against the agent.
+
+**Request Body:**
+
+```json
+{
+  "message": "Refactor the database connection pool",
+  "sessionId": "20260524-182358-c11f"
+}
+```
+
+| Field | Type | Required | Description |
+|-------|------|----------|-------------|
+| `message` | string | yes | The prompt to execute |
+| `sessionId` | string | no | Resume existing session; omit to create new |
+
+**Response:**
+
+```json
+{
+  "sessionId": "20260524-192358-a1b2",
+  "content": "Here's the refactored connection pool...",
+  "turns": 3,
+  "toolCalls": [
+    { "name": "read", "input": "src/db/pool.ts" },
+    { "name": "edit", "input": "src/db/pool.ts" }
+  ]
+}
+```
+
+---
+
+## Sessions
+
+### List Sessions
+
+```
+GET /api/sessions
+```
+
+Returns all sessions for the current project, sorted by creation time descending.
+
+### Create Session
+
+```
+POST /api/sessions
+```
+
+Creates a new session and returns its ID.
+
+### Get Session
+
+```
+GET /api/sessions/:id
+```
+
+Returns session metadata.
+
+### Delete Session
+
+```
+DELETE /api/sessions/:id
+```
+
+Permanently deletes a session and all its data.
+
+---
+
+## Tools
+
+```
+GET /api/tools
+```
+
+List all registered tools (built-in + MCP + plugin).
+
+**Response:**
+
+```json
+[
+  { "name": "read", "description": "Read file contents..." },
+  { "name": "bash", "description": "Execute shell commands..." },
+  { "name": "mcp__my-server__search", "description": "[MCP:my-server] Search documentation" }
+]
+```
+
+---
+
+## Skills
+
+```
+GET /api/skills
+```
+
+List all registered skills.
+
+**Response:**
+
+```json
+[
+  {
+    "name": "code-review",
+    "description": "Review code for bugs, style issues, and best practices",
+    "source": "builtin"
+  },
+  {
+    "name": "my-custom-skill",
+    "description": "A file-based skill",
+    "source": "file"
+  }
+]
+```
+
+---
+
+## Error Responses
+
+All endpoints return standard HTTP status codes:
+
+| Status | Meaning |
+|--------|---------|
+| 200 | Success |
+| 400 | Bad request (e.g., missing `message` field) |
+| 404 | Resource not found (e.g., invalid session ID) |
+| 500 | Internal server error |
+
+Error bodies:
+
+```json
+{
+  "error": "message is required"
+}
+```

package/docs/architecture.md

@@ -0,0 +1,206 @@
+# System Architecture
+
+## Overview
+
+Agent is a modular AI agent framework with multi-provider LLM support, hierarchical context management, tool execution, plugin/skill systems, and multi-agent collaboration. The core loop is: **compose context -> LLM inference -> parse output -> execute tools -> repeat**.
+
+```
+                   ┌─────────────────────────────┐
+                   │      Gateway Layer           │
+                   │  CLI / TUI / HTTP Server     │
+                   └──────────┬──────────────────┘
+                              │
+                   ┌──────────▼──────────────────┐
+                   │      AgentLoop               │
+                   │  compose → LLM → parse → tool│
+                   └──────────┬──────────────────┘
+                              │
+         ┌────────────────────┼────────────────────┐
+         │                    │                    │
+   ┌─────▼─────┐      ┌──────▼──────┐     ┌──────▼──────┐
+   │  Context   │      │   Provider   │     │    Tools     │
+   │  Composer  │      │   (LLM)     │     │  Read/Write  │
+   │  L0-L4     │      │ 6 backends  │     │  Edit/Bash   │
+   └───────────┘      └─────────────┘     │  Glob/Skill  │
+                                          │  MCP/Agent   │
+                                          └─────────────┘
+```
+
+## Module Map
+
+### 1. Gateway Layer
+Entry points that initialize the system and handle user interaction.
+
+| Module | File | Role |
+|--------|------|------|
+| CLI | `gateway/cli.ts` | Commander-based CLI, interactive/repl mode, session management |
+| TUI | `gateway/tui.ts` | Neo-blessed full-screen terminal UI |
+| HTTP API | `gateway/server.ts` | Fastify HTTP server, 8 REST endpoints |
+| Factory | `gateway/factory.ts` | Shared `createAgent()` init, eliminates duplication across gateways |
+
+### 2. Orchestrator
+The core agent loop and planning system.
+
+- **AgentLoop** (`orchestrator/loop.ts`): Main loop — compose context, call LLM, parse stream events, execute tools, repeat. Supports `normal` and `precise` scheduling modes.
+- **LLMOrchestrator** (`orchestrator/planner.ts`): Plan-based orchestration — assesses task, creates plan, tracks progress.
+- **PlanStore** (`orchestrator/plan-store.ts`): Plan CRUD with status tracking per step.
+
+### 3. Context System (L0-L4)
+
+```
+Zone 1: Persistent identity/system prompts (always included)
+Zone 2: Manifest index of available tools/skills/agents (always included)
+Zone 3: Lazy-expand definitions (tools/skills/agents on demand)
+Zone 4: Conversation history (most recent messages)
+```
+
+| Component | File | Role |
+|-----------|------|------|
+| LayeredContextComposer | `context/composer.ts` | Assembles L0-L4 zones, supports normal/precise modes |
+| CompressorOrchestrator | `context/compressor.ts` | Context compression via summarization |
+| StructuredSummarizer | `context/compressor.ts` | LLM-powered structured summarization |
+| TokenCounter | `context/tokenizer.ts` | Token counting via js-tiktoken |
+| PromptBuilder | `context/prompt-builder.ts` | Assembles final prompt array from zones |
+| ContextModes | `context/modes.ts` | Normal vs precise scheduling mode logic |
+
+### 4. Provider Layer
+Unified LLM interface with 6 backend implementations.
+
+- **Interface** (`provider/interface.ts`): `Provider` abstract with `createStream()`
+- **AnthropicProvider**: Claude models via Anthropic SDK
+- **OpenAIProvider**: GPT models via OpenAI SDK
+- **DeepSeekProvider**: DeepSeek models
+- **LocalProvider**: Local models (llama.cpp, ollama, etc.)
+- **OpenAICompatibleProvider**: Groq, xAI, Mistral, OpenRouter, Moonshot
+- **GeminiProvider**: Google Gemini via @google/genai SDK
+- **ResilientProvider**: Retry (exponential backoff) + Circuit breaker (3-state) wrapper
+- **FallbackProviderChain**: Sequential fallback across multiple providers
+- **ProviderManager**: Auto-detection, config file loading, env var resolution
+- **ModelCatalog**: Model metadata, capabilities, cost info
+
+### 5. Tool System
+Registry-based tool execution with sandbox support.
+
+| Tool | File | Description |
+|------|------|-------------|
+| ReadTool | `tools/read.ts` | Read files with line numbers, offset/limit |
+| WriteTool | `tools/write.ts` | Write/create files, auto-create dirs |
+| EditTool | `tools/edit.ts` | String replacement in files |
+| BashTool | `tools/bash.ts` | Shell command execution with optional sandbox |
+| GlobTool | `tools/glob.ts` | File pattern matching |
+| SkillTool | `skills/skill-tool.ts` | Invoke registered skills |
+| DelegateToAgentTool | `agents/delegate-tool.ts` | Delegate tasks to sub-agents |
+
+- **ToolRegistry**: Register/lookup tools
+- **FilteredToolRegistry**: Tool whitelisting for sub-agents
+- **ToolExecutor**: Execute tool calls with error handling
+
+### 6. Memory & Persistence
+
+| Component | File | Storage |
+|-----------|------|---------|
+| ConversationStore | `memory/conversation.ts` | JSONL, auto-truncation |
+| EventStore | `memory/events.ts` | JSONL |
+| StatsManager | `memory/stats.ts` | JSON |
+| SummaryStore | `memory/summary.ts` | Markdown |
+| SessionManager | `memory/session.ts` | Directory-based under `~/.agent/sessions/` |
+
+### 7. Sub-Agent System
+
+- **AgentRegistry**: Register/lookup agent definitions
+- **Built-in agents**: code-reviewer, security-auditor, test-writer
+- **DelegateToAgentTool**: 3 collaboration modes
+  - `delegate`: Single sub-agent, single session
+  - `adversarial`: Two agents independently review, results merged
+  - `parallel`: Multiple agents via Promise.all
+- **FilteredToolRegistry**: Per-agent tool whitelisting
+
+### 8. MCP (Model Context Protocol)
+
+| Component | File | Description |
+|-----------|------|-------------|
+| MCPClient | `mcp/client.ts` | stdio + SSE transport, tool discovery |
+| MCPBridge | `mcp/bridge.ts` | Multi-client bridge to ToolRegistry |
+| MCPConfigLoader | `mcp/config.ts` | Load from `.agent/mcp.json` or `.mcp.json` |
+| MCPServerManager | `mcp/lifecycle.ts` | Process lifecycle, auto-restart, health check |
+
+### 9. Plugin System
+
+- **PluginManager**: discover -> load -> activate -> connect MCP -> register skills
+- **PluginLoader**: Scan directories, load manifests, dynamic import
+- **PluginApi**: registerTool/registerSkill/registerContextSource/registerMcpServer
+
+### 10. Skills
+
+- **SkillRegistry**: Register/lookup, index/full-definitions generation
+- **File-based skills**: Load from `~/.agent/skills/*.md` and `.agent/skills/*.md`
+- **Hot-reload**: `fs.watch` with 500ms debounce on skill directories
+- **Built-in skills**: code-review, debug, refactor
+
+### 11. Schedule System
+
+- **HeartbeatScheduler**: Configurable heartbeat loop
+- **Schedule types**: interval, cron (5-field), daily (HH:mm), fixed-time (ISO)
+- **Persistence**: JSON storage at `~/.agent/scheduler/tasks.json`
+
+### 12. Lifecycle Management
+
+- **ProcessManager**: spawn/stop/restart, health check (HTTP/TCP), crash auto-recovery
+- **LocalModelManager**: Auto-detect local model backends (llama.cpp, ollama, etc.)
+- **LifecycleSupervisor**: Signal handling, managed entity registration, graceful shutdown
+
+### 13. Other Modules
+
+| Module | File | Description |
+|--------|------|-------------|
+| Parser | `parser/router.ts` | Routes stream events (text/thinking/tool_use) |
+| Dependency | `dependency/` | Import graph analysis, change impact assessment |
+| Logging | `logging/logger.ts` | ConsoleLogger, 5 levels, msg-first, JSON to stderr |
+| Setup | `setup/` | ConfigManager + SetupWizard for first-run |
+| Prompts | `prompts/` | System prompt templates (identity, safety, tools, orchestration) |
+
+## Data Flow
+
+```
+User Input
+    │
+    ▼
+AgentLoop.run()
+    │
+    ├─► ContextComposer.compose() → [Zone1..Zone4]
+    │      Zone1: Identity + safety prompts
+    │      Zone2: Tool/skill/agent index
+    │      Zone3: Expanded definitions
+    │      Zone4: Conversation history
+    │
+    ├─► Provider.createStream() → StreamEvent[]
+    │      ├─ TEXT → OutputHandler.onText()
+    │      ├─ THINKING → OutputHandler.onThinking()
+    │      ├─ TOOL_USE → OutputHandler.onToolUse()
+    │      │      └─ ToolExecutor.execute() → ToolResult
+    │      │             └─ OutputHandler.onToolResult()
+    │      ├─ USAGE → StatsManager.update()
+    │      └─ STOP → turn complete
+    │
+    ├─► ConversationStore.append() (persist)
+    │
+    └─► Repeat until maxTurns or STOP
+```
+
+## Error Recovery
+
+```
+Provider chain:
+  Primary Provider
+    └─ ResilientProvider (retry 4×, exponential backoff)
+         └─ Circuit Breaker (5 failures → open 30s)
+              └─ FallbackProviderChain → next provider
+```
+
+## Key Design Decisions
+
+1. **msg-first logging**: `logger.info('message', { context })` — human-readable first, structured second
+2. **Async generator streams**: `async *createStream()` for unified streaming across all providers
+3. **Context zones over sliding window**: L0-L4 enables precise control over what stays vs. what compresses
+4. **Sub-agent isolation**: Each sub-agent gets its own session directory under `sub-agents/{name}-{id}/`
+5. **Plugin-host co-location**: Plugins load in-process via dynamic import (no sandbox isolation yet)