architect-mode · v1.0.0 · 2026-08-28 · sha256 38d567508e2279bd
architect-mode v1.0.0A
Immutable. This exact content is served forever at /api/v1/blob/38d567508e2279bd.
---
name: architect-mode
display-name: Architect Mode
description: 双阶段架构模式 - 先规划后编辑，分离推理与代码修改，提升多文件编辑准确率
version: 1.0.0
category: development
user-invocable: true
tags: [architect, planning, multi-file, edit, reasoning, two-phase]
capabilities:
[architecture-planning, edit-generation, plan-review, multi-file-coordination]
tools:
- file_reader
- file_writer
- code_analyzer
- project_navigator
execution-capabilities: [data:result, data:task, filesystem:read, filesystem:write, host:filesystem, host:logger, process:cwd, runtime:crypto, runtime:time]
handler: ./handler.js
instructions: |
Use this skill when the user needs to make complex, multi-file changes. First phase: analyze
the codebase and generate a structured plan (which files to modify, what changes, in what order).
Second phase: translate the plan into precise file edits. The user can review and approve/modify
the plan before execution. This improves accuracy by separating "what to do" from "how to do it".
examples:
- input: "/architect-mode 'add user authentication with JWT'"
output: "Plan: 1) Create auth middleware (src/main/auth/jwt-middleware.js) 2) Add login/register routes 3) Update database schema 4) Add frontend login page. Approve? [Y/n]"
- input: "/architect-mode --plan-only 'refactor database layer to use repository pattern'"
output: "Architecture Plan: 5 files to create, 8 files to modify. Estimated: 450 lines added, 120 lines modified."
- input: "/architect-mode --execute plan-001"
output: "Executing plan-001: 13/13 edits applied successfully. 0 conflicts."
input-schema:
type: object
properties:
task:
type: string
description: High-level description of the architectural change
plan-only:
type: boolean
description: Only generate plan, don't execute
execute:
type: string
description: Execute a previously generated plan by ID
output-schema:
type: object
properties:
plan:
type: object
description: Structured plan with phases, files, and edits
edits:
type: array
description: List of file edits to apply
status:
type: string
description: Current status (planning, awaiting-review, executing, completed)
model-hints:
planning: [claude-opus-4-6, gpt-4-turbo]
editing: [claude-sonnet-4-5-20250929, gpt-4o-mini]
cost: medium
os: [win32, darwin, linux]
author: ChainlessChain
license: MIT
homepage: https://github.com/nicekwell/chainlesschain
repository: https://github.com/nicekwell/chainlesschain
---
# Architect Mode 技能
## 描述
双阶段架构模式，将复杂的多文件编辑任务拆分为"规划"与"执行"两个独立阶段。通过分离推理（reasoning）和代码修改（editing），显著提升多文件变更的准确率和可控性。灵感来源于 Aider 的 architect mode 和 Claude Code 的 plan mode。
## 核心理念
传统 AI 编码助手在处理多文件变更时容易出现：
- 遗漏需要修改的文件
- 修改顺序导致中间状态不一致
- "迷失在上下文中"忘记原始目标
Architect Mode 解决这些问题的方式：
1. **Phase 1 (规划)**: 使用推理能力强的模型分析代码库，生成结构化变更计划
2. **Phase 2 (执行)**: 按照计划逐个文件生成精确的编辑指令
## 使用方法
```
/architect-mode <任务描述> # 完整流程：规划 → 审核 → 执行
/architect-mode --plan-only <任务描述> # 仅生成计划，不执行
/architect-mode --execute <plan-id> # 执行已有计划
```
## 操作模式
### 默认模式（规划 + 执行）
```
/architect-mode 'add WebSocket support for real-time notifications'
```
流程：
1. 扫描代码库，提取与任务相关的文件
2. 生成结构化变更计划（分阶段、按依赖排序）
3. 输出计划摘要，等待用户审核
4. 用户确认后，逐个文件生成编辑指令
5. 输出执行结果和摘要
### 仅规划模式
```
/architect-mode --plan-only 'migrate from Express to Fastify'
```
只生成变更计划，不自动执行。适用于：
- 大型重构前的影响评估
- 团队讨论变更方案
- 多方案对比
### 执行已有计划
```
/architect-mode --execute plan-abc123
```
执行之前生成并保存的计划。支持断点续传。
## 计划结构
```json
{
"id": "plan-abc123",
"task": "add user authentication with JWT",
"createdAt": "2026-02-17T10:30:00Z",
"phases": [
{
"order": 1,
"description": "Create authentication infrastructure",
"files": [
{
"path": "src/main/auth/jwt-middleware.js",
"action": "create",
"reason": "JWT token verification middleware",
"estimatedLines": 85
},
{
"path": "src/main/database.js",
"action": "modify",
"reason": "Add users table schema",
"estimatedLines": 15
}
]
},
{
"order": 2,
"description": "Add API routes and frontend",
"files": [
{
"path": "src/main/ipc/auth-ipc.js",
"action": "create",
"reason": "IPC handlers for login/register",
"estimatedLines": 120
}
]
}
],
"estimate": {
"totalFiles": 6,
"filesToCreate": 3,
"filesToModify": 3,
"filesToDelete": 0,
"linesAdded": 450,
"linesModified": 120,
"linesDeleted": 0
}
}
```
## 编辑指令格式
每个文件的编辑指令包含：
```json
{
"file": "src/main/auth/jwt-middleware.js",
"action": "create",
"description": "JWT token verification middleware",
"edits": [
{
"type": "insert",
"after": 45,
"content": "// new code to insert"
},
{
"type": "replace",
"startLine": 100,
"endLine": 105,
"content": "// replacement code"
}
]
}
```
## 输出格式
### 计划阶段输出
```
Architecture Plan: plan-abc123
════════════════════════════════════════════
Task: add user authentication with JWT
Phase 1: Create authentication infrastructure (3 files)
[CREATE] src/main/auth/jwt-middleware.js — JWT verification (~85 lines)
[MODIFY] src/main/database.js — Add users table (+15 lines)
[CREATE] src/main/auth/password-utils.js — Password hashing (~40 lines)
Phase 2: Add API routes and frontend (3 files)
[CREATE] src/main/ipc/auth-ipc.js — IPC handlers (~120 lines)
[MODIFY] src/renderer/router.js — Add auth routes (+8 lines)
[CREATE] src/renderer/pages/LoginPage.vue — Login UI (~150 lines)
Estimate: 6 files (3 new, 3 modified), +418 lines, ~120 lines modified
Approve this plan? [Y/n/edit]
```
### 执行阶段输出
```
Executing plan-abc123...
[1/6] Created src/main/auth/jwt-middleware.js (85 lines)
[2/6] Modified src/main/database.js (+15 lines)
[3/6] Created src/main/auth/password-utils.js (40 lines)
[4/6] Created src/main/ipc/auth-ipc.js (120 lines)
[5/6] Modified src/renderer/router.js (+8 lines)
[6/6] Created src/renderer/pages/LoginPage.vue (150 lines)
Result: 6/6 edits applied successfully. 0 conflicts.
```
## 与其他技能的区别
| 特性 | architect-mode | code-review | refactor |
| ---- | -------------------- | ------------ | ------------------ |
| 目标 | 多文件变更规划与执行 | 代码质量分析 | 代码异味检测与重构 |
| 模式 | 双阶段（规划+执行） | 只读 | 单阶段执行 |
| 范围 | 跨模块、跨层级 | 单文件或目录 | 单文件或目录 |
| 审核 | 内置计划审核流程 | 无需 | Diff 预览 |
| 用途 | 新功能、大型重构 | 代码审查 | 小型重构 |
## 最佳实践
1. **任务描述越具体越好**: "add JWT auth with refresh tokens, bcrypt hashing, and rate limiting" 比 "add auth" 更好
2. **先 --plan-only**: 大型变更先生成计划，确认无误后再执行
3. **分阶段执行**: 对于超过 10 个文件的变更，建议分批执行
4. **结合 code-review**: 执行后运行 code-review 验证代码质量
5. **结合 test-generator**: 为新增代码自动生成测试