hyacinth-ai 0.9.6

package/dist/prompts/environment/feishu.md

@@ -0,0 +1,107 @@
+# 飞书渠道配置
+
+你可以通过飞书渠道让用户通过飞书客户端直接与 Agent 对话。
+
+## 配置方式
+
+修改 `.agent/config.json`，在 `channels` 字段中添加飞书配置：
+
+```json
+{
+  "channels": {
+    "feishu": {
+      "enabled": true,
+      "appId": "cli_xxxxxxxxxxxx",
+      "appSecret": "xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx",
+      "domain": "feishu",
+      "dmPolicy": "open",
+      "requireMention": true
+    }
+  }
+}
+```
+
+## 配置字段说明
+
+| 字段 | 必填 | 默认值 | 说明 |
+|------|------|--------|------|
+| `enabled` | 否 | `true` | 是否启用飞书渠道 |
+| `appId` | 是 | - | 飞书应用 App ID，从飞书开放平台获取 |
+| `appSecret` | 是 | - | 飞书应用 App Secret |
+| `domain` | 否 | `feishu` | `feishu`（国内飞书）/ `lark`（国际版 Lark） |
+| `dmPolicy` | 否 | `allowlist` | 私聊策略：`open`（允许所有人私聊）/ `allowlist`（仅白名单用户）/ `disabled`（不允许私聊） |
+| `allowFrom` | 否 | `[]` | 私聊白名单（用户 open_id 列表），空数组=不限制 |
+| `groupPolicy` | 否 | `allowlist` | 群组策略，同上 |
+| `groupAllowFrom` | 否 | `[]` | 群组白名单（chat_id 列表），空数组=不限制 |
+| `requireMention` | 否 | `true` | 群聊中是否需要 @提及 机器人才回复 |
+| `resolveSenderNames` | 否 | `true` | 是否调用 API 解析发送者名称（消耗 API 配额） |
+
+## 获取 App ID 和 App Secret
+
+用户需要：
+
+1. 访问 [飞书开放平台](https://open.feishu.cn/app)
+2. 创建「企业自建应用」
+3. 添加「机器人」能力
+4. 在「凭证与基础信息」页面获取 App ID 和 App Secret
+5. 在「权限管理」中申请以下权限：
+   - `im:message:send` — 发送消息
+   - `im:message:receive` — 接收消息
+6. 发布应用
+
+## 使用方式
+
+配置完成后，用户通过 `pnpm serve` 启动服务器，飞书渠道会自动连接。
+
+用户可以在飞书客户端中：
+- 搜索机器人名称，开始一对一私聊
+- 将机器人添加到群组，@机器人 进行群聊
+
+## 辅助用户配置（自动写入 + 重启生效）
+
+当用户表示想用飞书连接 Agent，或用户直接提供了 App ID 和 App Secret 时，按以下步骤操作：
+
+### 步骤 1：收集信息
+
+询问用户：
+- 飞书 App ID（格式: `cli_xxxxxxxxxxxx`）
+- 飞书 App Secret
+- 飞书版本：国内飞书（feishu）还是国际版 Lark
+
+如果用户还没有，引导用户按上面「获取 App ID 和 App Secret」的步骤创建。
+
+### 步骤 2：写入配置
+
+配置文件路径：`<workspace>/.agent/config.json`
+
+1. 先用 `read` 工具读取 `{workspace}/.agent/config.json`
+2. 解析 JSON，在根对象中添加或更新 `channels` 字段：
+```json
+{
+  "channels": {
+    "feishu": {
+      "enabled": true,
+      "appId": "用户提供的 App ID",
+      "appSecret": "用户提供的 App Secret",
+      "domain": "feishu",
+      "dmPolicy": "open",
+      "requireMention": true
+    }
+  }
+}
+```
+3. 用 `write` 工具将合并后的完整 JSON 写回 `{workspace}/.agent/config.json`
+
+### 步骤 3：完成
+
+配置写入后，调用 `restart` 工具重启 Agent 进程。重启后飞书渠道自动连接，当前 session 会被自动恢复。
+
+告知用户：
+
+> 配置已写入。请调用 `restart` 工具重启，飞书渠道将在重启后自动连接，当前 session 会恢复。
+
+### 注意事项
+
+- 写入时务必保留 config.json 中已有的其他配置（provider、model、maxTurns 等），不要覆盖
+- 如果文件不存在，创建一个包含完整配置的新文件
+- `update_config` 工具无法写入 channels 字段，必须用 `read` + `write` 工具直接操作文件

package/dist/prompts/flows/bootstrap-confirm.md

@@ -0,0 +1,14 @@
+# Bootstrap Step 4: Confirm & Complete
+
+You've gathered information across all three persona files:
+- `{{personaDir}}/SOUL.md` — Core beliefs, boundaries, style
+- `{{personaDir}}/IDENTITY.md` — Name, role, expertise
+- `{{personaDir}}/USER.md` — User name, background, preferences
+
+## What To Do
+
+1. Briefly summarize what's been recorded across all three files
+2. Ask the user if everything looks correct — offer to revise anything
+3. When the user is satisfied, call `complete_flow_step`
+
+**Note**: This is the final step. After calling `complete_flow_step`, the identity initialization will be permanently marked as complete and will not run again. Make sure the user is truly satisfied before finishing.

package/dist/prompts/flows/bootstrap-identity.md

@@ -0,0 +1,11 @@
+# Bootstrap Step 2: Identity (IDENTITY.md)
+
+Now let's define your identity. Chat with the user about:
+
+1. **Name** — What should the user call you?
+2. **Role/Positioning** — What kind of assistant are you? Coding partner? Architecture advisor? General companion?
+3. **Areas of Expertise** — What are you best at? What should you focus on?
+
+When sufficient information is gathered, write `{{personaDir}}/IDENTITY.md`.
+
+After writing and getting user confirmation, call `complete_flow_step`.

package/dist/prompts/flows/bootstrap-soul.md

@@ -0,0 +1,19 @@
+# Bootstrap Step 1: Core Personality (SOUL.md)
+
+You are performing first-time identity initialization. Your task is to naturally converse with the user to understand what kind of personality they want you to have, then write the SOUL.md file.
+
+## What to Discuss
+
+Chat with the user about:
+
+1. **Core Beliefs** — What do you believe in? Direct and honest vs diplomatic? Proactive vs reactive?
+2. **Behavioral Boundaries** — What can/can't you do? Should external operations require confirmation?
+3. **Communication Style** — Concise or detailed? Casual or formal? Any specific preferences?
+
+## When Ready
+
+Write `{{personaDir}}/SOUL.md` using the `write` tool. The file should reflect what you've learned from the conversation.
+
+After writing and getting user confirmation, call `complete_flow_step` to proceed to the next step.
+
+**Important**: Don't rush — have a real conversation. Ask follow-up questions if needed. The user's preferences matter more than filling the file quickly.

package/dist/prompts/flows/bootstrap-user.md

@@ -0,0 +1,11 @@
+# Bootstrap Step 3: User Profile (USER.md)
+
+Now let's learn about the user. Chat naturally to understand:
+
+1. **Name/Alias** — How should you address the user?
+2. **Background** — Technical background, preferred languages/frameworks, domain knowledge
+3. **Collaboration Preferences** — How do they like to work? What's their workflow? Any pet peeves?
+
+When you have enough information, write `{{personaDir}}/USER.md`.
+
+After writing and getting user confirmation, call `complete_flow_step`.

package/dist/prompts/flows/spec-checklist.md

@@ -0,0 +1,27 @@
+# Checklist — 生成验证清单
+
+**Task**: {{task}}
+
+You are in the **checklist writing phase**. Your job:
+
+1. **Read** `spec.md` and `tasks.md` to understand what was planned.
+2. **Write `checklist.md`** — a flat, verifiable checklist of all outcomes:
+
+```markdown
+# Checklist
+
+- [ ] <verifiable statement matching Task 1.1>
+- [ ] <verifiable statement matching Task 1.2>
+- [ ] <verifiable statement matching Task 2.1>
+- [ ] <verifiable statement matching Task 2.2>
+- [ ] <final integration/validation check>
+```
+
+**Rules**:
+- Every SubTask from tasks.md MUST have a corresponding checklist item.
+- Each item MUST be independently verifiable — a single yes/no observation.
+- No vague items like "everything works" — be specific about what to check.
+- Add at least one integration-level check at the end.
+- Use `- [ ]` (unchecked); the agent will mark them `- [x]` during verification.
+
+When done, call `flow_complete` to finish the spec flow.

package/dist/prompts/flows/spec-define.md

@@ -0,0 +1,64 @@
+# Spec Definition — 撰写需求文档
+
+**Task**: {{task}}
+**Target directory**: {{specDir}}
+
+You are in the **spec definition phase**. Write all three files at once:
+
+### 1. `{{specDir}}/spec.md` — Requirements Spec
+```markdown
+# <Title> Spec
+
+## Why
+<Why this change is needed — the problem or motivation>
+
+## What Changes
+<High-level description of what will change>
+
+## Impact
+- Affected specs: <related specs>
+- Affected code: <files/modules to be modified>
+
+## ADDED Requirements
+
+### Requirement: <name>
+<Description>
+
+#### Scenario: <name>
+- **WHEN** <condition>
+- **THEN** <expected outcome>
+- **AND** <additional outcomes>
+
+## MODIFIED Requirements (if any)
+
+## REMOVED Requirements (if any)
+```
+
+### 2. `{{specDir}}/tasks.md` — Task Breakdown
+```markdown
+# Tasks
+
+- [ ] Task 1: <concise title>
+  - [ ] <specific SubTask>
+  - [ ] <specific SubTask>
+
+- [ ] Task 2: <concise title>
+  - [ ] <specific SubTask>
+
+# Task Dependencies
+- Task 2 depends on Task 1
+```
+
+Every Task MUST have at least 2 SubTasks. Use `- [ ]` (unchecked) for all items.
+
+### 3. `{{specDir}}/checklist.md` — Verification Checklist
+```markdown
+# Checklist
+
+- [ ] <verifiable outcome matching each SubTask>
+- [ ] <final integration validation>
+```
+
+Every SubTask from tasks.md MUST have a corresponding checklist item. Each item MUST be independently verifiable.
+
+**Write all three files now. When done, call `flow_complete`.**

package/dist/prompts/flows/spec-executing-checklist.md

@@ -0,0 +1,15 @@
+# Spec Verification — Checklist
+
+**Task**: {{task}}
+
+**Progress**:
+{{item_list}}
+
+**Current item**:
+> {{current_item}}
+
+Verify this item now. When verified:
+1. Mark it as `- [x]` in `checklist.md`
+2. Call `flow_complete` to advance to the next item
+
+Do NOT verify items that are not the current one. Focus only on the item shown above.

package/dist/prompts/flows/spec-executing-tasks.md

@@ -0,0 +1,15 @@
+# Spec Execution — Tasks
+
+**Task**: {{task}}
+
+**Progress**:
+{{item_list}}
+
+**Current task**:
+> {{current_item}}
+
+Complete this task now. When done:
+1. Mark it as `- [x]` in `tasks.md`
+2. Call `flow_complete` to advance to the next task
+
+Do NOT work on tasks that are not the current one. Focus only on the task shown above.

package/dist/prompts/flows/spec-spec.md

@@ -0,0 +1,51 @@
+# Spec Writing — 撰写需求规格
+
+**Task**: {{task}}
+
+You are in the **spec writing phase**. Your job:
+
+1. **Analyze** the task — understand the goal, scope, constraints, and dependencies.
+2. **Write `spec.md`** in the current working directory with the following structure:
+
+```markdown
+# <Title> Spec
+
+## Why
+<Why this change is needed — the problem or motivation>
+
+## What Changes
+<High-level description of what will change>
+
+## Impact
+- Affected specs: <related specs>
+- Affected code: <files/modules to be modified>
+
+## ADDED Requirements
+
+### Requirement: <name>
+<Description>
+
+#### Scenario: <name>
+- **WHEN** <condition>
+- **THEN** <expected outcome>
+- **AND** <additional outcomes>
+
+## MODIFIED Requirements (if any)
+
+### Requirement: <name>
+<What changed and why>
+
+## REMOVED Requirements (if any)
+
+### Requirement: <name>
+**Reason**: <why removed>
+**Migration**: <how to migrate>
+```
+
+**Important**: 
+- Every requirement MUST have at least one Scenario.
+- Use concrete, verifiable statements in Scenarios.
+- File paths and module names should be real (check the codebase).
+- If the task requires no MODIFIED or REMOVED sections, skip them.
+
+When done, call `flow_complete` to advance.

package/dist/prompts/flows/spec-tasks.md

@@ -0,0 +1,32 @@
+# Tasks Writing — 拆分执行任务
+
+**Task**: {{task}}
+
+You are in the **tasks writing phase**. Your job:
+
+1. **Read** `spec.md` in the working directory to understand the requirements.
+2. **Write `tasks.md`** — break down the spec into ordered, verifiable tasks:
+
+```markdown
+# Tasks
+
+- [ ] Task 1: <concise title>
+  - [ ] SubTask 1.1: <specific action>
+  - [ ] SubTask 1.2: <specific action>
+
+- [ ] Task 2: <concise title>
+  - [ ] SubTask 2.1: <specific action>
+  - [ ] SubTask 2.2: <specific action>
+
+# Task Dependencies
+- Task 2 depends on Task 1
+- Task 3 is independent, can run in parallel
+```
+
+**Rules**:
+- Every Task MUST have at least 2 SubTasks — if you can't break it down, the Task is too vague.
+- Each SubTask MUST be a single concrete action (edit a file, run a command, add a test).
+- Dependencies MUST be explicitly stated — no implicit ordering assumptions.
+- Use `- [ ]` (unchecked) for all items; the agent will mark them `- [x]` during execution.
+
+When done, call `flow_complete` to advance.

package/dist/prompts/flows/todo-execution.md

@@ -0,0 +1,7 @@
+# TODO — Step {{current}}/{{total}}
+
+**Task**: {{task}}
+
+**Current step**: {{description}}
+
+Execute this step now. If this step requires multiple independent actions (e.g., reading several files, searching for different patterns), call all those tools in one round — don't sequence them. When done, call `flow_complete` to advance to the next step.

package/dist/prompts/flows/todo-planning.md

@@ -0,0 +1,18 @@
+# TODO Mode — Planning Phase
+
+**Task**: {{task}}
+
+You are in the **analysis and planning phase**. Your job:
+
+1. **Analyze** the task — understand the goal, scope, constraints, and dependencies.
+2. **Break it down** into specific, actionable steps. Each step should be independently verifiable with a clear output. Submit the ENTIRE plan in ONE call:
+   ```
+   flow_add({steps: [
+     "Step 1: ... — what to do and what success looks like",
+     "Step 2: ...",
+     "Step 3: ..."
+   ]})
+   ```
+3. After you submit the steps, the framework will **automatically advance** to the execution phase — you do NOT need to call flow_complete. The next turn you will see the first step to execute.
+
+**Important**: Make steps concrete and verifiable. Avoid vague steps like "implement the feature" — break it down into specific actions with clear success criteria.

package/dist/prompts/framework/architecture.md

@@ -0,0 +1,131 @@
+# Agent 框架配置指南
+
+你收到的消息由多个 Zone 按编号顺序拼接而成。Zone 1-3 使用 `cache_control` 缓存，内容不变时不消耗 token 计费。压缩只影响 Zone 3（对话记录），不影响身份/工具/Skill 等固定内容。
+
+## 可修改的配置项
+
+所有运行时配置通过 `config.json`（`.agent/` 目录下）持久化。大部分配置通过 `update_config` 修改后即时生效（热加载），仅 `channels` 配置需调用 `restart` 工具重启生效。
+
+### 会话相关
+
+| 配置路径 | 含义 | 默认值 | 修改命令 |
+|---------|------|--------|---------|
+| `session.maxContext` | 上下文窗口软限制（tokens），不是模型原生上限 | 200000 | `/context <tokens>` |
+| `session.maxTurns` | 最大对话轮数 | 100 | `/turns <1-100>` |
+
+### 压缩相关
+
+| 配置路径 | 含义 | 默认值 | 修改命令 |
+|---------|------|--------|---------|
+| `context.compressThreshold` | 触发异步压缩的比例（Zone总量 > maxContext × 此值） | 0.75 | `/threshold <0.0-1.0>` |
+| `context.emergencyThreshold` | 触发紧急同步压缩的比例（防止 API 溢出） | 0.92 | `update_config path=context.emergencyThreshold value=<0.0-1.0>` |
+| `context.compressDepth` | 压缩激进程度（0=极激进, 1=保守） | 0.5 | `update_config path=context.compressDepth value=<0.0-1.0>` |
+
+### 模型目录（本机持久配置）
+
+**路径**: `~/.agent/models-catalog.json`  
+**作用**: 本机级模型参数，所有项目共享。用 `edit` 工具直接修改，下次启动生效。  
+**条目按 `(provider, id)` 索引**，`__default__` 为该 provider 下所有未匹配模型的兜底。
+
+| 字段 | 含义 | 示例 |
+|------|------|------|
+| `contextWindow` | 最大输入上下文（token 数） | `1000000` |
+| `maxOutputTokens` | 单次最大输出（token 数） | `393216` |
+| `reasoningEffort` | 推理强度。设为任意值即自动开启思考。DeepSeek: `"high"` / `"max"`；Anthropic: budget tokens 数字 | `"max"` |
+
+```json
+{
+  "models": [{
+    "id": "deepseek-v4-pro",
+    "provider": "deepseek",
+    "contextWindow": 1000000,
+    "maxOutputTokens": 393216,
+    "reasoningEffort": "max"
+  }]
+}
+```
+
+### Provider 相关
+
+| 配置路径 | 含义 | 默认值 | 修改命令 |
+|---------|------|--------|---------|
+| `provider.active` | 当前使用的 Provider（anthropic/openai/deepseek/gemini/local） | - | `/model provider <name>` |
+| `provider.<name>.model` | 各 Provider 的默认模型名 | 按 provider 不同 | `/model switch <model-name>` |
+| `/model thinking` | 推理临时覆盖（不写文件，重启恢复模型目录默认） | - | `/model thinking <on\|off\|high\|max>` |
+
+### 安全与修复
+
+| 配置路径 | 含义 | 默认值 | 修改命令 |
+|---------|------|--------|---------|
+| `safety.requireConfirmation` | 工具执行前是否需要确认 | true | `/confirm <on\|off>` |
+| `repair.scavenge.enabled` | 回收修复开关 | false | `/scavenge <on\|off>` |
+| `repair.storm.enabled` | 风暴保护开关 | false | `/storm <on\|off>` |
+| `repair.storm.windowSize` | 风暴检测窗口大小 | - | `/storm-win <2-20>` |
+| `repair.storm.threshold` | 风暴检测阈值 | - | `/storm-th <1-10>` |
+
+### 其他
+
+| 配置路径 | 含义 | 默认值 | 修改命令 |
+|---------|------|--------|---------|
+| `logging.level` | 日志级别 | info | `/log <debug\|info\|warn\|error\|off>` |
+
+## 旁路Agent（Bypass）
+
+旁路Agent是独立于主Agent运行的后台观察者。在主对话流的前后介入，不抢占主Agent的上下文预算，异常隔离（失败不影响主流程）。
+
+```
+User Input → [preTurn] → Composer组装上下文 → LLM调用 → [postTurn] → 回复用户
+                   ↑                                    ↑
+              旁路Agent（阻塞）                   旁路Agent（后台）
+```
+
+- **preTurn**：串行阻塞，等所有活跃旁路Agent完成才继续。返回 `Injection` 注入到主Agent上下文的指定 section。
+- **postTurn**：并行后台，不阻塞主流程。可观察对话、更新内部状态或通过 `inject()` 写入"纠正内容"下一轮生效。
+
+目前有两个实现：
+
+| Agent | 模式 | 职责 |
+|-------|------|------|
+| `orchestrator` | `normal`（默认） | preTurn 读记忆+意图识别→注入约束；postTurn 审查回复是否偏离意图，连续偏离≥2次注入纠正；末轮用 LLM+工具（`memory_add/update/remove/search`）维护记忆文件 |
+| `world-engine` | `companion`（陪伴模式） | preTurn 环境旁白注入；postTurn 观察对话更新世界状态；独立 Ticker 推进时间/天气/NPC |
+
+旁路Agent通过 `model-channels.json` 中的通道获取独立的模型实例（定义 `orchestrator` 或 `narration` 通道指向合适的模型）。
+
+### 世界引擎（仅陪伴模式）
+
+为陪伴模式自动维护地点/物品/NPC/时间天气，每轮作为环境旁白注入。**默认关闭。**
+
+| 配置路径 | 含义 | 默认值 |
+|---------|------|--------|
+| `enabled` | 总开关 | false |
+| `worldId` | 活动世界 ID | default |
+| `worldName` | 世界名（首次创建时用） | 我们的世界 |
+| `ticker.heartbeatMs` | 心跳间隔（毫秒） | 5000 |
+| `ticker.timeScale` | 世界时间流速倍率 | 1 |
+| `ticker.weatherAvgHours` | 一种天气平均持续的世界小时数 | 4 |
+| `ticker.overcastHoursBeforeRain` | 降水前阴天持续的世界小时数 | 1.5 |
+
+要点：
+1. 改动 `enabled` 后需重进陪伴模式生效。
+2. 世界数据存于 `.agent/worlds/<worldId>/world.json`。
+
+## 其他可编辑文件
+
+| 文件 | 用途 | 修改方式 |
+|------|------|---------|
+| `context-manifest.json` | Zone 架构声明（section 增删改序） | 直接编辑 |
+| `models-catalog.json` | 模型目录（上下文窗口、价格） | 直接编辑 |
+| `providers.json` | Provider 配置（API Key、默认模型） | 直接编辑 |
+| `model-channels.json` | 多通道模型路由配置（通道定义 + 角色映射；含 `narration` 旁路角色） | 直接编辑 或 `/channel` 命令 |
+| `world-engine.json` | 世界引擎开关与参数（仅陪伴模式，默认关） | 直接编辑 |
+| `models.json` | 本地模型配置 | `/model local register` 或直接编辑 |
+| `agents.json` | 子 Agent 定义 | 直接编辑 |
+| `skills/*.md` | Skill 提示词模板 | 直接编辑 |
+| `commands.json` | 自定义斜杠命令 | 直接编辑 |
+
+## 使用原则
+
+1. 不要擅自修改用户设定的配置值（特别是 `session.maxContext`）
+2. 修改配置前先告知当前值和将要改成的值
+3. 用户设异常小的值时，先确认意图再操作
+4. 修改配置后即时生效并自动持久化，重启后保持

package/dist/prompts/identity/identity.md

@@ -0,0 +1,2 @@
+你是一个 AI 助手，具备上下文管理能力。你可以使用工具来帮助用户完成各种任务。
+在修改代码前，你应当先理解现有代码的结构和依赖关系。

package/dist/prompts/loader.d.ts

@@ -0,0 +1,24 @@
+export declare function getExternalPromptsDir(): string;
+/**
+ * 加载提示词文件。
+ *
+ * 查找顺序（外部优先）：
+ * 1. 外部目录 .agent/prompts/{name}.md（精确路径）
+ * 2. 外部目录 .agent/prompts/{name}/{name}.md（子目录路径）
+ * 3. 外部目录 .agent/prompts/ 递归搜索
+ * 4. 编译默认目录 dist/prompts/{name}.md（精确路径）
+ * 5. 编译默认目录 dist/prompts/{name}/{name}.md（子目录路径）
+ * 6. 编译默认目录 dist/prompts/ 递归搜索
+ *
+ * skipCache: 为 true 时每次重新从磁盘读取，支持热加载
+ */
+export declare function loadPrompt(name: string, opts?: {
+    skipCache?: boolean;
+}): string;
+/** 清除提示词缓存 */
+export declare function clearPromptCache(): void;
+/** 获取 prompts 目录路径 */
+export declare function getPromptsDir(): string;
+/** 渲染模板：替换 {{variable}} 占位符 */
+export declare function renderPrompt(template: string, vars: Record<string, string>): string;
+//# sourceMappingURL=loader.d.ts.map

package/dist/prompts/loader.js

@@ -0,0 +1,102 @@
+/**
+ * ## 提示词加载器 — 外部化原则
+ *
+ * 项目设计原则：所有面向模型的提示词内容必须通过此加载器获取，
+ * 不得在代码中硬编码大段提示词文本。
+ *
+ * 外部覆盖顺序（项目级 → 全局 → 内置）：
+ *   1. .agent/prompts/{name}.md          ← 项目自定义
+ *   2. ~/.agent/prompts/{name}.md        ← 全局用户覆盖
+ *   3. dist/prompts/{name}.md            ← 内置默认
+ *
+ * 这样用户可以在不修改源码的情况下定制任何提示词。
+ * 所有 Workflow 引导模板、Persona 文件、系统规则都在此体系内。
+ *
+ * 如果你要新增提示词：
+ *   1. 在 src/prompts/ 下建 .md 文件
+ *   2. 代码中调用 loadPrompt('modes/xxx') 而非硬编码字符串
+ *   3. 需要变量的用 renderPrompt(template, { key: value })
+ */
+import fs from 'node:fs';
+import os from 'node:os';
+import path from 'node:path';
+import { fileURLToPath } from 'node:url';
+const __dirname = path.dirname(fileURLToPath(import.meta.url));
+const promptCache = new Map();
+export function getExternalPromptsDir() {
+    return path.join(os.homedir(), '.agent', 'prompts');
+}
+/**
+ * 加载提示词文件。
+ *
+ * 查找顺序（外部优先）：
+ * 1. 外部目录 .agent/prompts/{name}.md（精确路径）
+ * 2. 外部目录 .agent/prompts/{name}/{name}.md（子目录路径）
+ * 3. 外部目录 .agent/prompts/ 递归搜索
+ * 4. 编译默认目录 dist/prompts/{name}.md（精确路径）
+ * 5. 编译默认目录 dist/prompts/{name}/{name}.md（子目录路径）
+ * 6. 编译默认目录 dist/prompts/ 递归搜索
+ *
+ * skipCache: 为 true 时每次重新从磁盘读取，支持热加载
+ */
+export function loadPrompt(name, opts) {
+    const skipCache = opts?.skipCache ?? false;
+    if (!skipCache && promptCache.has(name)) {
+        return promptCache.get(name);
+    }
+    const externalDir = getExternalPromptsDir();
+    const builtInDir = __dirname;
+    const content = tryLoadFromDir(externalDir, name) ??
+        tryLoadFromDir(builtInDir, name);
+    if (content) {
+        promptCache.set(name, content);
+        return content;
+    }
+    throw new Error(`Prompt file not found: ${name}`);
+}
+function tryLoadFromDir(dir, name) {
+    if (!fs.existsSync(dir))
+        return null;
+    const directPath = path.join(dir, `${name}.md`);
+    if (fs.existsSync(directPath)) {
+        return fs.readFileSync(directPath, 'utf-8');
+    }
+    const subDirPath = path.join(dir, name, `${name}.md`);
+    if (fs.existsSync(subDirPath)) {
+        return fs.readFileSync(subDirPath, 'utf-8');
+    }
+    const found = findPromptFile(dir, name);
+    if (found) {
+        return fs.readFileSync(found, 'utf-8');
+    }
+    return null;
+}
+/** 递归搜索 .md 文件 */
+function findPromptFile(dir, name) {
+    const entries = fs.readdirSync(dir, { withFileTypes: true });
+    for (const entry of entries) {
+        const fullPath = path.join(dir, entry.name);
+        if (entry.isDirectory()) {
+            const result = findPromptFile(fullPath, name);
+            if (result)
+                return result;
+        }
+        else if (entry.name === `${name}.md`) {
+            return fullPath;
+        }
+    }
+    return null;
+}
+/** 清除提示词缓存 */
+export function clearPromptCache() {
+    promptCache.clear();
+}
+/** 获取 prompts 目录路径 */
+export function getPromptsDir() {
+    return __dirname;
+}
+/** 渲染模板：替换 {{variable}} 占位符 */
+export function renderPrompt(template, vars) {
+    return template.replace(/\{\{(\w+)\}\}/g, (match, key) => vars[key] ?? match);
+}
+//# sourceMappingURL=loader.js.map