DevTaskFlow

用自然语言发起开发任务，AI 驱动的开发流水线。Describe your idea, get working software. 适合产品、运营、业务人员从 0 搭建项目。

Security Scan

The skill's code and declared requirements mostly match its purpose (AI-driven dev pipeline), but there are inconsistencies around what it claims not to expose vs what parts of the code read/print and guidance that it should be invoked proactively — these raise privacy/exposure concerns the user should review before installing.

Purpose & Capability

Name/description (DevTaskFlow: natural-language → project pipeline) align with requested environment variables (LLM base URL, API key, model) and optional deploy/publish credentials (GitHub token, SSH key, Docker registry). Code files implement analyze/write/review/deploy/publish, orchestrators and a local board — these are appropriate for the stated purpose.

Instruction Scope

SKILL.md instructs the agent to proactively invoke the skill whenever a user expresses desire to build software ("Don't wait for user to say 'use dtflow'"). The repository and runtime code read and write project files (versions/.state.json, .dtflow/config.json, DEV_PLAN.md, etc.) — which is expected — but there is an inconsistency: SKILL.md says sensitive items should not be exposed, yet the CLI's board command and some code paths can print deployment config (host/path/user) and read .state.json. The board server attempts to sanitize but the CLI printing logic reads deploy.host/path directly. Sanitization appears partial (removes only certain fields) and may still leak file paths, task metadata, or deployment details. This broad/automatic invocation plus potential exposure of deployment/config data is the main scope concern.

Install Mechanism

No explicit install spec in the skill bundle, but the CLI will run local installers at runtime (e.g., subprocess.run(['npm','install']) for the board). That will pull Node packages from registries (package-lock.json references mirrors). No remote arbitrary archive downloads or extract steps were found. Installing the board will therefore run external package installs locally — expected, but worth noting.

Credentials

Requested environment variables (DTFLOW_LLM_BASE_URL, DTFLOW_LLM_API_KEY, DTFLOW_LLM_MODEL) are proportional to a tool that calls an LLM. Optional DTFLOW_GITHUB_TOKEN, DTFLOW_DEPLOY_SSH_KEY, and DTFLOW_DOCKER_REGISTRY are reasonable for publish/deploy adapters. No unrelated or excessive credentials were requested.

Persistence & Privilege

The skill is not marked always:true. It includes guidance to be proactively invoked on user intents (agent-initiated), which increases autonomy but is allowed by default. The code writes state and project files within a workspace and can run npm/node subprocesses; it does not appear to modify other skills or global agent configs. The combination of proactive invocation + access to deploy/publish creds raises the blast radius if those credentials are provided.

What to consider before installing

What to check before installing DevTaskFlow: - Credentials: The skill needs an LLM endpoint and API key; optionally you may supply a GitHub token and an SSH deploy key. Only provide the minimum-scoped GitHub token (repo-only, no org admin) and use an SSH key with limited server permissions. Consider using ephemeral/limited LLM keys if possible. - Proactive invocation: SKILL.md tells the agent to proactively use the tool when users express building-software intents. Decide whether you want the agent to auto-launch a pipeline or require explicit user consent before starting code generation, publishing, or deployment. - Data exposure: The skill reads project state files (.dtflow/config.json, versions/.state.json, REQUIREMENTS.md, DEV_PLAN.md). The web board includes sanitization, but CLI code may print deploy host/path info and other deploy config directly. If you host the board or run CLI in a shared environment, sensitive deployment details or file paths could be exposed. Run in a local/sandbox workspace if you are unsure. - Review defaults: The board server runs npm install and serves on a port (default 8765) — do not expose that port to the public. Confirm the board's sanitization is sufficient for your threat model before enabling it for others. - Least privilege: Grant optional tokens/keys only when you intend to publish or deploy. Audit the skill's actions (check PROJECTS.json, .dtflow/config.json and versions/.state.json) and test with a harmless example project first to observe behavior (no external publishing). - Additional checks to reduce risk: inspect .dtflow config files in created projects before running deploy/publish; run the skill in an isolated VM/container; if possible instrument network/logging to see which external endpoints (LLM base URL, GitHub) are contacted and when. Given the above inconsistencies around data exposure and the proactive-invocation guidance, proceed cautiously and perform an initial trial with minimal credentials and isolated environment.

DevTaskFlow — Agent 使用手册

什么时候用

当用户表达以下意图时，主动使用：

• "我想做一个 XXX 系统/工具/平台"
• "帮我开发一个 XXX"
• "我需要一个 XXX，功能是..."
• 用户描述了一个软件/系统需求
• 用户问项目进展、想继续做、想看进度

不要等用户说"用 dtflow"。 用户不需要知道这个工具的存在。

Token 消耗参考

开发一个项目会消耗大量 token，提前告知用户：

项目规模	预估 Token 消耗	示例
小型	300-500 万	简单的个人工具、静态页面、小表单
中型	~4000 万	多页面管理后台、带数据库的应用、用户系统
大型	2 亿+	复杂业务系统、多角色权限、API 集成

消耗取决于需求复杂度、迭代次数、审查修复次数。首次可先用小项目试水。

支持的模型

DevTaskFlow 的代码生成质量高度依赖模型能力。经测试，以下模型能稳定完成完整开发任务：

模型	说明
Claude Opus 4.6	综合能力最强，推荐用于复杂开发任务
GPT 5.4 Pro	OpenAI 旗舰模型，大型项目首选
GPT 5.4	性价比高，适合中小型项目
小米 Mimo V2 Pro	国产模型，中文表现好

⚠️ 其他模型（如 GPT-4o、Claude Sonnet、DeepSeek 等）可能无法完成完整开发任务，产出质量不稳定。请从以上 4 个模型中选择。

核心命令

dtflow setup                                        # 配置 AI 服务（交互式）
dtflow start --new-project --name NAME --idea "需求"  # 开始新项目
dtflow start                                        # 继续上次进度
dtflow start --confirm                              # 确认分析方案
dtflow start --confirm-write                        # 确认预览后正式生成代码
dtflow start --feedback "修改意见"                   # 提出修改
dtflow start --run                                  # 本地预览
dtflow start --deploy                               # 部署上线并封版
dtflow board                                        # 所有项目状态（文字）
dtflow board --serve                                # 启动可视化看板服务
dtflow board-query --name PROJECT                   # 单个项目详情（文字）

工作流程

用户提出新需求

如果用户有明确需求描述（比如"我想做一个客户管理工具"）：

1. dtflow start --new-project --name 项目名 --idea "用户的需求原文"
2. 系统创建项目、给出补充建议
3. 向用户展示建议，问是否要补充
4. 确认后自动 analyze → 展示任务列表
5. dtflow start --confirm → 自动 write（先预览）→ review → fix → review
6. 全部通过后 → dtflow start --run 本地预览
7. 用户确认没问题 → dtflow start --deploy

如果用户需求模糊（比如"我想做个东西管理客户信息"）：

1. 不要直接调用 dtflow，先通过对话引导收集需求
2. 问清楚：
   • 给谁用的？（团队/客户/个人）
   • 最核心的功能是什么？
   • 需要登录吗？
   • 有技术偏好吗？（不知道就帮你选）
3. 收集到足够信息后，拼成需求调用 dtflow start

用户想本地预览

1. dtflow start --run
2. 返回访问链接给用户

用户想看项目进展

1. 检查看板服务是否在运行（默认 8765 端口）
2. 如果在运行 → 发链接
3. 如果不在运行 → dtflow board 文字版

用户问某个项目详情

1. dtflow board-query --name 项目名
2. 把文字结果发给用户

用户想继续之前的项目

1. dtflow start（不加参数，自动继续）
2. 根据输出告知用户当前阶段

首次使用（环境未配置）

1. dtflow setup 交互式引导（含 AI 配置 + 部署方式选择）
2. 非交互环境下手动创建 .env：

   DTFLOW_LLM_BASE_URL=...
   DTFLOW_LLM_API_KEY=...
   DTFLOW_LLM_MODEL=...
   

状态机

dtflow start 自动推进，你只需知道阶段：

状态	含义	你该说什么
created	刚创建	"项目已创建，正在分析需求..."
pending_confirm	方案已出	"我分析了你的需求，建议做这几件事：..."
confirmed	已确认	"好的，开始生成代码..."
writing/written	代码已生成	"代码写好了，我在检查..."
needs_fix	有问题	"发现几个小问题，已修复：..."
review_passed	审查通过	"代码没问题了，要本地先看看效果吗？"
sealed	已封版	"上线完成！"

向用户展示什么

不要暴露： analyze、DEV_PLAN.md、orchestration、config.json、.state.json、token 数 应该说： "我分析了需求"、"代码已生成"、"检查过了没问题"、"可以部署了"

注意事项

• dtflow setup 是交互式命令，在非交互环境不可用
• 所有命令在项目根目录运行（含 .dtflow/config.json 的目录）
• board 的 Node.js 应用需要 npm install（首次自动执行）
• 看板服务默认端口 8765，仅限本地使用，不要暴露到公网
• board API 已脱敏：不返回 host/user/path 等敏感部署信息
• run 本地预览需要项目有可执行的启动命令（npm start / python app.py 等）
• Docker 部署需要本地安装 Docker