cold-start

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Cold Start: Memory Bank + Agents bootstrap (greenfield & brownfield)

冷启动：Memory Bank + Agent 搭建（全新项目greenfield & 遗留系统brownfield）

Summary

概述

Note:
cold-start
is the package skill for all-in-one bootstrap. Do not confuse it with the generated project command
/cold-start
: that command is a lightweight router used after skeleton creation inside a target repo. For modular usage, prefer individual skills:
mb-init
(skeleton),
mb-from-prd
(greenfield),
mb-map-codebase
(brownfield),
mb-review
(review),
mb-execute
(implementation),
mb-verify
(UAT).

What it does: creates the Memory Bank skeleton, writes agent entry points, and routes the repo into the right workflow.
Use it when: you want one entry point for either a new project with a PRD or an existing codebase that needs mapping first.
Input: repository root plus either
```
prd.md
```
/ requirements text or an existing codebase.
Output:
```
.memory-bank/
```
,
```
.tasks/
```
,
```
.protocols/
```
, agent entry files, and the next step for greenfield or brownfield work.

Supported scenarios:

Greenfield: repo has
```
prd.md
```
or requirements text, but no code yet.
Brownfield: repo already contains code and needs as-is documentation before change planning.

注意：
cold-start
是一站式搭建的包技能。请勿将其与生成的项目命令
/cold-start
混淆：该命令是目标仓库内骨架创建完成后使用的轻量级路由工具。如需模块化使用，建议选用独立技能：
mb-init
（骨架搭建）、
mb-from-prd
（全新项目greenfield）、
mb-map-codebase
（遗留系统brownfield）、
mb-review
（评审）、
mb-execute
（实施）、
mb-verify
（用户验收测试UAT）。

功能：创建Memory Bank骨架，编写Agent入口点，并将仓库接入对应的工作流。
适用场景：你希望通过单一入口点处理两种情况——一是带有PRD的新项目，二是需要先进行映射的现有代码库。
输入：仓库根目录，加上
```
prd.md
```
/需求文本，或现有代码库。
输出：
```
.memory-bank/
```
、
```
.tasks/
```
、
```
.protocols/
```
、Agent入口文件，以及全新项目或遗留系统工作的下一步指引。

支持的场景：

全新项目（Greenfield）：仓库包含
```
prd.md
```
或需求文本，但尚无代码。
遗留系统（Brownfield）：仓库已包含代码，在变更规划前需要如实完成文档记录。

Non-negotiable principles

不可妥协的原则

1) Orchestrator vs workers

1) 编排者 vs 执行者

You (main agent) are the orchestrator. You plan and coordinate.
Workers (subagents) do the heavy lifting (scanning many files, producing long reports).
Max depth = 2: orchestrator → subagent. No sub-sub-agents.

你（主Agent）是编排者：负责规划与协调工作。
执行者（子Agent）负责繁重工作（扫描大量文件、生成冗长报告）。
最大层级=2：编排者 → 子Agent。不允许存在子子Agent。

.tasks/

is runtime memory (НЕ Memory Bank)

.tasks/

是运行时内存（而非Memory Bank）

Every agent process gets a unique task folder:
```
.tasks/TASK-XXX/
```
.
Subagents write detailed outputs there, and return only short summaries to the orchestrator.

每个Agent进程都会获得唯一的任务文件夹：
```
.tasks/TASK-XXX/
```
。
子Agent在其中写入详细输出，仅向编排者返回简短摘要。

3) Avoid conflicts by design

3) 从设计层面避免冲突

Split work by non-overlapping file sets.
If overlap is unavoidable, use git branches/worktrees per agent and merge later.

按无重叠的文件集拆分工作。
若无法避免重叠，为每个Agent使用Git分支/工作区，后续再合并。

4) Parallelism limits

4) 并行性限制

Run up to 5–7 subagents in parallel.
If your runtime supports multi-tool calls: spawn parallel subagents in a single orchestrator turn.

最多并行运行5–7个子Agent。
若运行时支持多工具调用：在单次编排者轮次中启动并行子Agent。

5) Docs First

5) 文档优先

After completing any meaningful unit of work:

Update Memory Bank while context is fresh.
Then change code / commit.

完成任何有意义的工作单元后：

在上下文清晰时更新Memory Bank。
再进行代码修改/提交。

Output you MUST create/maintain

你必须创建/维护的输出

Repo root

仓库根目录

```
AGENTS.md
```
(canonical, short, ~100 lines)
```
CLAUDE.md
```
→ symlink (or copy) to
```
AGENTS.md
```
(optional)
```
GEMINI.md
```
→ symlink (or copy) to
```
AGENTS.md
```

```
AGENTS.md
```
（标准规范，简短，约100行）
```
CLAUDE.md
```
→ 符号链接（或副本）指向
```
AGENTS.md
```
(可选)
```
GEMINI.md
```
→ 符号链接（或副本）指向
```
AGENTS.md
```

Repo folders

仓库文件夹

```
.memory-bank/
```
— durable knowledge base
```
.tasks/
```
— operational runtime memory
```
.protocols/
```
— file-based protocols (plans / decision logs / resume)

```
.memory-bank/
```
— 持久化知识库
```
.tasks/
```
— 运行时操作内存
```
.protocols/
```
— 基于文件的协议（计划/决策日志/进展记录）

Step 0 — Detect environment (don’t guess)

步骤0 — 检测环境（请勿猜测）

Check what tools you have:
- Codex available? (Codex CLI / MCP tool)
- Claude Code available? (CLI
```
claude
```
  , subagents)
- Browser tools? (Playwright / agent-browser / CDP MCP)
Decide primary engine:
- If working inside Codex: use
```
gpt-5.2-high
```
  for implementation and review (prefer fresh sessions for critical reviews).
- If working inside Claude Code: use subagents for analysis + use Opus for synthesis; optionally call Codex via shell for second opinion or structured review.
- If both Codex and Claude are available (dual-engine): use Claude for wide-context scanning and synthesis, Codex for structured review and implementation. Cross-validate critical outputs (MB compliance, architecture) by running the review step on the other engine. See Step 4 for the hybrid workflow.

If you don’t control Codex model selection via UI, create a project
.codex/config.toml
profile set (see Step 1.5).

检查可用工具：
- 是否有Codex？（Codex CLI / MCP工具）
- 是否有Claude Code？（CLI
```
claude
```
  、子Agent）
- 是否有浏览器工具？（Playwright / agent-browser / CDP MCP）
确定主引擎：
- 若在Codex内工作：使用
```
gpt-5.2-high
```
  进行实施与评审（关键评审建议使用全新会话）。
- 若在Claude Code内工作：使用子Agent进行分析 + 使用Opus进行合成；可通过Shell调用Codex获取第二意见或结构化评审。
- 若同时拥有Codex和Claude（双引擎）：使用Claude进行宽上下文扫描与合成，使用Codex进行结构化评审与实施。通过在另一引擎上运行评审步骤，交叉验证关键输出（MB合规性、架构）。混合工作流详见步骤4。

若无法通过UI选择Codex模型，创建项目
.codex/config.toml
配置文件（见步骤1.5）。

Step 1 — Bootstrap the skeleton

步骤1 — 搭建骨架

1.1 Create directories

1.1 创建目录

Create (if missing):

```
.memory-bank/
```
- ```
mbb/
```
- ```
architecture/
```
- ```
guides/
```
- ```
adrs/
```
- ```
tech-specs/
```
- ```
domains/
```
  (optional but recommended)
- ```
contracts/
```
  (optional but recommended)
- ```
runbooks/
```
  (optional but recommended)
- ```
workflows/
```
  (optional; can keep
  wfs/
  if you already use it)
- ```
quality/
```
  (optional but recommended)
- ```
testing/
```
- ```
skills/
```
- ```
epics/
```
- ```
features/
```
- ```
tasks/
```
  (Memory Bank backlog & plans)
- ```
commands/
```
  (slash-command specs used by humans/agents)
- ```
agents/
```
  (subagent prompt library)
- ```
archive/
```
- ```
bugs/
```
```
.tasks/
```
```
.protocols/
```

创建（若不存在）：

```
.memory-bank/
```
- ```
mbb/
```
- ```
architecture/
```
- ```
guides/
```
- ```
adrs/
```
- ```
tech-specs/
```
- ```
domains/
```
  (可选但推荐)
- ```
contracts/
```
  (可选但推荐)
- ```
runbooks/
```
  (可选但推荐)
- ```
workflows/
```
  (可选；若已使用
  wfs/
  可保留)
- ```
quality/
```
  (可选但推荐)
- ```
testing/
```
- ```
skills/
```
- ```
epics/
```
- ```
features/
```
- ```
tasks/
```
  (Memory Bank 待办事项与计划)
- ```
commands/
```
  (人类/Agent使用的斜杠命令规范)
- ```
agents/
```
  (子Agent提示词库)
- ```
archive/
```
- ```
bugs/
```
```
.tasks/
```
```
.protocols/
```

1.2 Create core files (use the templates)

1.2 创建核心文件（使用模板）

Use the templates in:

./references/shared-structure-template.md

At minimum you must create:

```
AGENTS.md
```
```
CLAUDE.md
```
symlink/copy
```
.memory-bank/index.md
```
```
.memory-bank/mbb/index.md
```
```
.memory-bank/product.md
```
```
.memory-bank/requirements.md
```
```
.memory-bank/testing/index.md
```
```
.memory-bank/tasks/backlog.md
```

Also create the command specs under

.memory-bank/commands/

(use

references/commands/*

使用以下路径中的模板：

./references/shared-structure-template.md

至少必须创建：

```
AGENTS.md
```
```
CLAUDE.md
```
符号链接/副本
```
.memory-bank/index.md
```
```
.memory-bank/mbb/index.md
```
```
.memory-bank/product.md
```
```
.memory-bank/requirements.md
```
```
.memory-bank/testing/index.md
```
```
.memory-bank/tasks/backlog.md
```

同时在

.memory-bank/commands/

下创建命令规范（使用

references/commands/*

）。

1.2.1 Create native skills (proxy commands)

1.2.1 创建原生技能（代理命令）

Create thin proxy skills so commands work natively in each runtime:

```
.claude/skills/<name>/SKILL.md
```
→ Claude Code + OpenCode
```
.agents/skills/<name>/SKILL.md
```
→ Codex CLI + OpenCode

Each proxy just says:

Read and follow the instructions in .memory-bank/commands/<name>.md

. This makes commands available natively (

/mb

/prd

/execute

, etc.) in all three tools.

The

init-mb.js

script creates both sets automatically.

创建轻量代理技能，使命令可在各运行时原生工作：

```
.claude/skills/<name>/SKILL.md
```
→ Claude Code + OpenCode
```
.agents/skills/<name>/SKILL.md
```
→ Codex CLI + OpenCode

每个代理仅需包含：

Read and follow the instructions in .memory-bank/commands/<name>.md

（读取并遵循

.memory-bank/commands/<name>.md

中的指令）。这样可使命令在三个工具中原生可用（

/mb

、

/prd

、

/execute

等）。

init-mb.js

脚本会自动创建这两组文件。

1.3 Enforce frontmatter rule

1.3 强制实施前置元数据规则

Every markdown file inside

.memory-bank/

must include YAML frontmatter with at least:

```
description: ...
```

status: draft|active|deprecated|archived

.memory-bank/

内的每个Markdown文件必须包含YAML前置元数据，至少包括：

```
description: ...
```

status: draft|active|deprecated|archived

1.4 Create

.tasks/

protocol

1.4 创建

.tasks/

协议

Create a task folder for this run (pick a new id):

```
.tasks/TASK-MB-INIT/
```

Inside it, create:

```
TASK-MB-INIT-S-00-orchestrator-plan.md
```
— what you will do + which subagents

为此运行创建任务文件夹（选择新ID）：

```
.tasks/TASK-MB-INIT/
```

在其中创建：

```
TASK-MB-INIT-S-00-orchestrator-plan.md
```
— 你的工作计划 + 子Agent分配

1.5 Optional: Codex profile presets

1.5 可选：Codex配置预设

If Codex is used, create

.codex/config.toml

with profiles:

default: coding (gpt-5.2, high)
profile
```
deep-review
```
: review (gpt-5.2, xhigh)

(If your repo is shared, consider keeping it local or documenting it in

.memory-bank/runbooks/

若使用Codex，创建

.codex/config.toml

配置文件，包含以下配置：

default: coding (gpt-5.2, high)
profile
```
deep-review
```
: review (gpt-5.2, xhigh)

（若仓库为共享仓库，建议将其设为本地文件，或在

.memory-bank/runbooks/

中记录）。

Step 2 — Choose scenario

步骤2 — 选择场景

Decision rule

决策规则

If repo has substantial code (
```
src/
```
,
```
package.json
```
,
```
go.mod
```
,
```
Cargo.toml
```
, etc.) → Brownfield (Step 3B).
If repo is mostly empty and you have
```
prd.md
```
→ Greenfield (Step 3A).
If both exist: treat as Brownfield + PRD delta (Step 3B).
If repo is empty/new and no
prd.md
→ Skeleton-only (Step 3C).

Record the scenario in:

.tasks/TASK-MB-INIT/TASK-MB-INIT-S-00-orchestrator-plan.md

若仓库包含大量代码（
```
src/
```
、
```
package.json
```
、
```
go.mod
```
、
```
Cargo.toml
```
等）→ 遗留系统（Brownfield）（步骤3B）。
若仓库基本为空且包含
```
prd.md
```
→ 全新项目（Greenfield）（步骤3A）。
若两者都存在：按遗留系统 + PRD增量处理（步骤3B）。
若仓库为新创建/空仓库且无
prd.md
→ 仅骨架（步骤3C）。

将场景记录在：

.tasks/TASK-MB-INIT/TASK-MB-INIT-S-00-orchestrator-plan.md

Step 3A — Greenfield workflow (PRD → Memory Bank)

步骤3A — 全新项目工作流（PRD → Memory Bank）

3A.1 Read PRD and do Deep Questioning

3A.1 读取PRD并进行深度提问

Read
```
prd.md
```
.
If gaps exist, run deep questioning in rounds (3–5 questions each). Use
```
./references/shared-deep-questioning.md
```
.
If PRD mentions “use skills/tools/CLIs” — run
```
/find-skills
```
first (project-installed → marketplace).

If user is temporarily unavailable (“запуск и ушёл”):

Record

Open questions

.protocols/PRD-BOOTSTRAP/decision-log.md

Stop and wait (do not invent facts; do not proceed to EP/FT/backlog generation without answers).

If the user explicitly wants autonomous mode:

record non-blocking gaps as
```
Assumptions
```
halt only on blocking gaps (security/compliance/external contract/data-loss risks)
after L1–L3 + review gate, continue with
```
/autonomous
```

读取
```
prd.md
```
。
若存在信息缺口，分轮次进行深度提问（每轮3–5个问题）。使用
```
./references/shared-deep-questioning.md
```
。
若PRD提及“使用技能/工具/CLI” — 先运行
```
/find-skills
```
（优先项目已安装技能 → 市场技能）。

若用户暂时不可用（“启动后离开”）：

将

未解决问题

记录在

.protocols/PRD-BOOTSTRAP/decision-log.md

中。

停止并等待（请勿编造信息；未得到答案前，请勿进行史诗/功能/待办事项生成）。

若用户明确要求自主模式：

将非阻塞缺口记录为
```
假设
```
仅在遇到阻塞缺口（安全/合规/外部合同/数据丢失风险）时暂停
完成L1–L3 + 评审关卡后，使用
```
/autonomous
```
继续

3A.2 Write product brief (L1)

3A.2 编写产品简介（L1）

Update

.memory-bank/product.md

using user’s wording.

使用用户的表述更新

.memory-bank/product.md

。

3A.3 Requirements + RTM

3A.3 需求 + 需求跟踪矩阵（RTM）

Update

.memory-bank/requirements.md

REQ-IDs
RTM table mapping REQ → Epic → Feature → Test

更新

.memory-bank/requirements.md

：

需求ID（REQ-IDs）
RTM表格，映射需求 → 史诗 → 功能 → 测试

3A.4 Create epics and features

3A.4 创建史诗与功能

Create:

```
.memory-bank/epics/EP-001-<slug>.md
```
```
.memory-bank/features/FT-001-<slug>.md
```

Each feature MUST include:

use cases
acceptance criteria
failure modes / edge cases
test strategy pointers

Status policy:

Default EP/FT frontmatter to
```
status: draft
```
until
```
Open questions
```
are resolved.
Promote to
```
status: active
```
only when acceptance criteria + verification plan are stable.

创建：

```
.memory-bank/epics/EP-001-<slug>.md
```
```
.memory-bank/features/FT-001-<slug>.md
```

每个功能必须包含：

用例
验收标准
失败模式/边缘情况
测试策略指引

状态规则：

默认史诗/功能前置元数据为
```
status: draft
```
，直至
```
未解决问题
```
得到解决。
仅当验收标准 + 验证计划稳定后，才升级为
```
status: active
```
。

3A.5 Tasks planning (per-feature, no “everything at once”)

3A.5 任务规划（按功能拆分，禁止“一次性全部生成”）

Do not generate a full task backlog for all features in one pass.

Instead:

Create/refresh
```
.memory-bank/tasks/backlog.md
```
as a skeleton (waves + placeholders).

For each selected feature, run

/prd-to-tasks FT-<NNN>

to produce:

.memory-bank/tasks/plans/IMPL-FT-<NNN>.md

atomic
```
TASK-*
```
items grouped by waves in
```
backlog.md
```

请勿一次性为所有功能生成完整的任务待办事项。

而是：

创建/更新
```
.memory-bank/tasks/backlog.md
```
作为骨架（阶段 + 占位符）。

对每个选定的功能，运行

/prd-to-tasks FT-<NNN>

生成：

.memory-bank/tasks/plans/IMPL-FT-<NNN>.md

在
```
backlog.md
```
中按阶段分组的原子化
```
TASK-*
```
项

3A.6 Identify key concepts and create duo docs

3A.6 识别关键概念并创建配套文档

For every non-trivial concept, create a duo:

```
.memory-bank/architecture/<concept>.md
```
(WHAT/WHY)
```
.memory-bank/guides/<concept>.md
```
(HOW)

对于每个非平凡概念，创建一组配套文档：

```
.memory-bank/architecture/<concept>.md
```
（是什么/为什么）
```
.memory-bank/guides/<concept>.md
```
（怎么做）

3A.7 Update index

3A.7 更新索引

Update

.memory-bank/index.md

with annotated links to all created docs.

使用带注释的链接更新

.memory-bank/index.md

，指向所有已创建的文档。

Step 3B — Brownfield workflow (Repo → Memory Bank)

步骤3B — 遗留系统工作流（仓库 → Memory Bank）

3B.1 Spawn repo-scanning subagents (parallel)

3B.1 启动仓库扫描子Agent（并行）

Create a new task folder:

```
.tasks/TASK-MB-MAP/
```

Spawn up to 5 subagents in parallel with non-overlapping scopes:

```
S-01
```
: build/tooling (package managers, scripts, CI)
```
S-02
```
: backend/services
```
S-03
```
: frontend/UI
```
S-04
```
: data layer (DB, migrations, schema)
```
S-05
```
: tests + quality gates

Each subagent MUST:

verify its file glob targets exist ("smart calling")

write a detailed report into

.tasks/TASK-MB-MAP/

using naming:

TASK-MB-MAP-S-0X-final-report-<code|docs>-YY.md

return a 5–10 line summary + file list

Use

./agents/shared-repo-scanner.md

as baseline prompt, but scope it.

创建新的任务文件夹：

```
.tasks/TASK-MB-MAP/
```

并行启动最多5个具有无重叠范围的子Agent：

```
S-01
```
：构建/工具链（包管理器、脚本、CI）
```
S-02
```
：后端/服务
```
S-03
```
：前端/UI
```
S-04
```
：数据层（数据库、迁移、 schema）
```
S-05
```
：测试 + 质量关卡

每个子Agent必须：

验证其文件匹配目标是否存在（“智能调用”）

将详细报告写入

.tasks/TASK-MB-MAP/

，命名格式：

TASK-MB-MAP-S-0X-final-report-<code|docs>-YY.md

返回5–10行摘要 + 文件列表

使用

./agents/shared-repo-scanner.md

作为基础提示词，但需限定范围。

3B.2 Synthesize Memory Bank from reports

3B.2 从报告合成Memory Bank

Using the

.tasks/TASK-MB-MAP/

reports, fill:

```
.memory-bank/product.md
```
— what the system is today
```
.memory-bank/architecture/
```
— C4 L1–L3 overview + key invariants
```
.memory-bank/runbooks/
```
— setup, dev, test, deploy
```
.memory-bank/contracts/
```
— API/event contracts
```
.memory-bank/testing/index.md
```
— canonical gates + verification notes
```
.memory-bank/index.md
```
— annotated links

PRD-less rule (non-negotiable): if there is no
prd.md
, you MUST NOT create or populate:
.memory-bank/epics/*
.memory-bank/features/*
.memory-bank/tasks/backlog.md
with waves/tasks
Empty skeleton files/folders are allowed if they were created by bootstrap.

Repo mapping is as-is documentation, not roadmap planning. If something cannot be derived from code/logs/tests, record it as a hypothesis or an open question.

使用

.tasks/TASK-MB-MAP/

中的报告，填充：

```
.memory-bank/product.md
```
— 当前系统的实际情况
```
.memory-bank/architecture/
```
— C4 L1–L3概览 + 关键不变量
```
.memory-bank/runbooks/
```
— 搭建、开发、测试、部署指南
```
.memory-bank/contracts/
```
— API/事件契约
```
.memory-bank/testing/index.md
```
— 标准质量关卡 + 验证说明
```
.memory-bank/index.md
```
— 带注释的链接

无PRD规则（不可妥协）：若无
prd.md
，禁止创建或填充：
.memory-bank/epics/*
.memory-bank/features/*
包含阶段/任务的
.memory-bank/tasks/backlog.md
允许保留由搭建过程创建的空骨架文件/文件夹。

仓库映射是如实记录文档，而非路线图规划。若无法从代码/日志/测试中推导信息，将其记录为假设或未解决问题。

3B.3 Ask user for PRD delta

3B.3 向用户请求PRD增量

After baseline MB exists:

ask the user for
```
prd.md
```
describing what to change/add
run
```
/prd
```
and
```
/prd-to-tasks
```
style decomposition against the existing baseline

基准Memory Bank创建完成后：

向用户请求
```
prd.md
```
，描述需要修改/添加的内容
针对现有基准运行
```
/prd
```
和
```
/prd-to-tasks
```
风格的分解流程

Step 3C — Skeleton-only workflow (no PRD, no code)

步骤3C — 仅骨架工作流（无PRD，无代码）

When the repo is new/empty and no

prd.md

is available:

当仓库为新创建/空仓库且无

prd.md

时：

3C.1 Create skeleton only

3C.1 仅创建骨架

Run Step 1 as usual — create all directories, core files from templates,

AGENTS.md

CLAUDE.md

symlink.

The skeleton provides a ready-to-fill structure:

product.md

requirements.md

backlog.md

, etc. remain as draft stubs.

按步骤1正常执行 — 创建所有目录、模板核心文件、

AGENTS.md

、

CLAUDE.md

符号链接。

骨架提供了可直接填充的结构：

product.md

、

requirements.md

、

backlog.md

等保留为草稿占位符。

3C.2 Ask for PRD

3C.2 请求PRD

After skeleton is created, ask the user to provide a PRD:

"Memory Bank skeleton created. To fill it with product details, epics, features, and a backlog, please provide a
prd.md
file (or paste requirements text). You can do this now or later — run
/prd
when ready."

骨架创建完成后，向用户请求提供PRD：

"Memory Bank骨架已创建。如需填充产品详情、史诗、功能与待办事项，请提供
prd.md
文件（或粘贴需求文本）。你可现在提供或稍后提供 — 准备就绪后运行
/prd
命令即可。"

3C.3 Wait or proceed

3C.3 等待或继续

If user provides PRD now → continue to Step 3A (Greenfield workflow).
If user defers → stop here. The skeleton is valid and usable. The user can invoke
```
$mb-from-prd
```
or
```
/prd
```
later to fill the Memory Bank.
If user provides partial info → run deep questioning (
```
./references/shared-deep-questioning.md
```
) to extract enough for product.md, then stop and wait for full PRD.

Note: The skeleton-only state is a valid stopping point.
AGENTS.md
+
.memory-bank/index.md
+ MBB rules are enough for agents to start navigating the repo.

若用户现在提供PRD → 继续执行步骤3A（全新项目工作流）。
若用户推迟提供 → 在此停止。骨架已可用，用户可稍后调用
```
$mb-from-prd
```
或
```
/prd
```
填充Memory Bank。
若用户提供部分信息 → 运行深度提问（
```
./references/shared-deep-questioning.md
```
）以提取足够信息用于
```
product.md
```
，然后停止并等待完整PRD。

注意：仅骨架状态是有效的停止点。
AGENTS.md
+
.memory-bank/index.md
+ MBB规则已足够Agent开始导航仓库。

Step 4 — Hybrid mode (Claude ↔ Codex)

步骤4 — 混合模式（Claude ↔ Codex）

If you’re in Claude Code but want Codex quality or long-horizon autonomy:

Use Claude subagents to produce the scan reports into
```
.tasks/TASK-MB-MAP/
```
.
Call Codex via shell to synthesize Memory Bank:

bash

codex exec --ephemeral --full-auto -m gpt-5.2-high \
  'Read .tasks/TASK-MB-MAP and build/refresh .memory-bank per MBB. Keep AGENTS.md short. Produce a summary and run a self-review.'

Then run a Codex deep review profile (or a fresh Claude session) for 5-expert review.

若在Claude Code内工作，但希望获得Codex的质量或长周期自主性：

使用Claude子Agent将扫描报告生成到
```
.tasks/TASK-MB-MAP/
```
。
通过Shell调用Codex合成Memory Bank：

bash

codex exec --ephemeral --full-auto -m gpt-5.2-high \
  'Read .tasks/TASK-MB-MAP and build/refresh .memory-bank per MBB. Keep AGENTS.md short. Produce a summary and run a self-review.'

然后运行Codex深度评审配置（或全新Claude会话）进行5专家评审。

Step 5 — Multi-expert review loop (fresh context)

步骤5 — 多专家评审循环（全新上下文）

Run fresh-context reviewers (do not reuse the writer context):

Architect (C4 + dependencies)
Scope analyst (REQ → Epic → Feature → Task coverage)
MBB compliance reviewer (frontmatter, links, atomics, duo)
Plan reviewer (backlog quality, waves, gates)
Security reviewer (auth, sensitive data, OWASP risks)
Code quality reviewer (conditional: if code exists — quality gates, conventions, hotspots)

Use prompts in

./agents/shared-review-*.md

and

./agents/shared-mb-reviewer.md

Rules:

If any reviewer returns REJECT → fix MB and repeat review.
Persist reviewer reports into
```
.tasks/TASK-MB-REVIEW/
```
.

运行全新上下文评审者（请勿复用编写者上下文）：

架构师（C4 + 依赖关系）
范围分析师（需求 → 史诗 → 功能 → 任务覆盖）
MBB合规性评审者（前置元数据、链接、原子化、配套文档）
计划评审者（待办事项质量、阶段、关卡）
安全评审者（认证、敏感数据、OWASP风险）
代码质量评审者（条件：若存在代码 — 质量关卡、规范、热点）

使用

./agents/shared-review-*.md

和

./agents/shared-mb-reviewer.md

中的提示词。

规则：

若任何评审者返回REJECT → 修复Memory Bank并重复评审。
将评审报告持久化到
```
.tasks/TASK-MB-REVIEW/
```
。

Step 6 — Start executing tasks

步骤6 — 开始执行任务

After review gate passes (APPROVE):

Pick the highest-priority task from
```
.memory-bank/tasks/backlog.md
```
.
Run
```
mb-execute
```
for the task (plan → implement → quality gates → MB-SYNC).
Run
```
mb-verify
```
to check acceptance criteria and record evidence.
Repeat until the wave is complete or user stops.

If the intended mode is unattended end-to-end:

do not stay in manual loop here
switch to generated project command
```
/autonomous
```

If
mb-execute
or
mb-verify
are not installed, follow their SKILL.md manually.

评审关卡通过（APPROVE）后：

从
```
.memory-bank/tasks/backlog.md
```
中选择最高优先级任务。
对该任务运行
```
mb-execute
```
（计划 → 实施 → 质量关卡 → MB同步）。
运行
```
mb-verify
```
检查验收标准并记录证据。
重复直至阶段完成或用户停止。

若目标模式为无人值守端到端执行：

请勿停留在手动循环
切换为生成的项目命令
```
/autonomous
```

若未安装
mb-execute
或
mb-verify
，手动遵循其SKILL.md中的指引。

Definition of done

完成定义

You are done when:

```
AGENTS.md
```
exists, short, points to
```
.memory-bank/index.md
```
.
```
CLAUDE.md
```
is a symlink/copy of
```
AGENTS.md
```
.
```
.memory-bank/
```
contains at minimum: index + MBB + product + testing (requirements/backlog can remain as stubs until PRD exists).
```
.tasks/
```
contains scan/review artifacts with naming + stage ids.
Greenfield: epics/features/backlog created from PRD.
Brownfield: repo mapped as-is into MB and user asked for PRD delta (no roadmap entities invented without PRD).
Skeleton-only: skeleton created, user asked for PRD (valid stopping point).
Multi-expert review passes (APPROVE) — for Greenfield/Brownfield; skip for Skeleton-only.
Execution loop is available (mb-execute + mb-verify reachable or documented).
Autonomous loop is available (
```
/autonomous
```
+
```
/autopilot
```
documented).

满足以下条件即完成：

```
AGENTS.md
```
已存在，内容简短，指向
```
.memory-bank/index.md
```
。
```
CLAUDE.md
```
是
```
AGENTS.md
```
的符号链接/副本。
```
.memory-bank/
```
至少包含：索引 + MBB + 产品 + 测试文档（需求/待办事项可保留为占位符，直至PRD存在）。
```
.tasks/
```
包含符合命名规范 + 阶段ID的扫描/评审产物。
全新项目场景：已从PRD创建史诗/功能/待办事项。
遗留系统场景：仓库已如实映射到Memory Bank，且已向用户请求PRD增量（无PRD时禁止创建路线图实体）。
仅骨架场景：骨架已创建，已向用户请求PRD（有效停止点）。
多专家评审通过（APPROVE） — 适用于全新项目/遗留系统场景；仅骨架场景可跳过。
执行循环可用（
```
mb-execute
```
+
```
mb-verify
```
可访问或已文档化）。
自主循环可用（
```
/autonomous
```
+
```
/autopilot
```
已文档化）。

References in this skill

本技能中的参考文件

./references/shared-structure-template.md

```
./references/shared-deep-questioning.md
```
```
./agents/shared-repo-scanner.md
```
```
./agents/shared-mb-reviewer.md
```
```
./agents/shared-review-architect.md
```
```
./agents/shared-review-scope.md
```
```
./agents/shared-review-code.md
```
```
./agents/shared-review-plan.md
```
```
./agents/shared-review-security.md
```
```
./scripts/shared-init-mb.js
```
(optional helper)

./references/shared-structure-template.md

```
./references/shared-deep-questioning.md
```
```
./agents/shared-repo-scanner.md
```
```
./agents/shared-mb-reviewer.md
```
```
./agents/shared-review-architect.md
```
```
./agents/shared-review-scope.md
```
```
./agents/shared-review-code.md
```
```
./agents/shared-review-plan.md
```
```
./agents/shared-review-security.md
```
```
./scripts/shared-init-mb.js
```
（可选辅助工具）

cold-start

Original

Translation

Cold Start: Memory Bank + Agents bootstrap (greenfield & brownfield)

冷启动：Memory Bank + Agent 搭建（全新项目greenfield & 遗留系统brownfield）

Summary

概述

Non-negotiable principles

不可妥协的原则

1) Orchestrator vs workers

1) 编排者 vs 执行者

2) .tasks/ is runtime memory (НЕ Memory Bank)

2) .tasks/是运行时内存（而非Memory Bank）

3) Avoid conflicts by design

3) 从设计层面避免冲突

4) Parallelism limits

4) 并行性限制

5) Docs First

5) 文档优先

Output you MUST create/maintain

你必须创建/维护的输出

Repo root

仓库根目录

Repo folders

仓库文件夹

Step 0 — Detect environment (don’t guess)

步骤0 — 检测环境（请勿猜测）

Step 1 — Bootstrap the skeleton

步骤1 — 搭建骨架

1.1 Create directories

1.1 创建目录

1.2 Create core files (use the templates)

1.2 创建核心文件（使用模板）

1.2.1 Create native skills (proxy commands)

1.2.1 创建原生技能（代理命令）

1.3 Enforce frontmatter rule

1.3 强制实施前置元数据规则

1.4 Create .tasks/ protocol

1.4 创建.tasks/协议

1.5 Optional: Codex profile presets

1.5 可选：Codex配置预设

Step 2 — Choose scenario

步骤2 — 选择场景

Decision rule

决策规则

Step 3A — Greenfield workflow (PRD → Memory Bank)

步骤3A — 全新项目工作流（PRD → Memory Bank）

3A.1 Read PRD and do Deep Questioning

3A.1 读取PRD并进行深度提问

3A.2 Write product brief (L1)

3A.2 编写产品简介（L1）

3A.3 Requirements + RTM

3A.3 需求 + 需求跟踪矩阵（RTM）

3A.4 Create epics and features

3A.4 创建史诗与功能

3A.5 Tasks planning (per-feature, no “everything at once”)

3A.5 任务规划（按功能拆分，禁止“一次性全部生成”）

3A.6 Identify key concepts and create duo docs

3A.6 识别关键概念并创建配套文档

3A.7 Update index

3A.7 更新索引

Step 3B — Brownfield workflow (Repo → Memory Bank)

步骤3B — 遗留系统工作流（仓库 → Memory Bank）

3B.1 Spawn repo-scanning subagents (parallel)

3B.1 启动仓库扫描子Agent（并行）

3B.2 Synthesize Memory Bank from reports

3B.2 从报告合成Memory Bank

3B.3 Ask user for PRD delta

3B.3 向用户请求PRD增量

Step 3C — Skeleton-only workflow (no PRD, no code)

步骤3C — 仅骨架工作流（无PRD，无代码）

3C.1 Create skeleton only

3C.1 仅创建骨架

3C.2 Ask for PRD

3C.2 请求PRD

3C.3 Wait or proceed

3C.3 等待或继续

Step 4 — Hybrid mode (Claude ↔ Codex)

步骤4 — 混合模式（Claude ↔ Codex）

Step 5 — Multi-expert review loop (fresh context)

2)
`.tasks/`
is runtime memory (НЕ Memory Bank)

2)
`.tasks/`
是运行时内存（而非Memory Bank）

1.4 Create
`.tasks/`
protocol

1.4 创建
`.tasks/`
协议