<?xml version="1.0" encoding="UTF-8"?><rss version="2.0"><channel><title>Zenty AI Guide — AI 编程与 Agent 开发实战指南</title><description>面向中文开发者的 AI 编程实战指南：Codex、Claude Code、MCP、Skills 与 Agent 自动化，从第一次真实任务到有检查点的流水线。</description><link>https://aiguide.zenty.top/</link><language>zh-cn</language><item><title>Codex CLI 安装与登录：把第一条命令跑成只读验收</title><link>https://aiguide.zenty.top/articles/codex-install-login/</link><guid isPermaLink="true">https://aiguide.zenty.top/articles/codex-install-login/</guid><description>从官方安装脚本或 npm 两条路径装上 Codex CLI，分清 ChatGPT 登录与 API Key 的计费边界，再用一条只读任务同时验证可执行文件、凭据和工作目录。</description><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><category>codex</category><category>Codex</category><category>CLI</category><category>安装</category><category>登录</category><category>认证</category></item><item><title>Windows 上的 Codex 环境：PowerShell、路径与换行符</title><link>https://aiguide.zenty.top/articles/codex-windows-setup/</link><guid isPermaLink="true">https://aiguide.zenty.top/articles/codex-windows-setup/</guid><description>在原生 Windows 上确认 Codex 命令来源、选对项目目录，分清 PowerShell 与 WSL 的路径和凭据，再用一条只读任务验证环境；不假设 Windows 必须依赖 WSL。</description><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><category>codex</category><category>Codex</category><category>Windows</category><category>PowerShell</category><category>WSL</category><category>环境配置</category></item><item><title>Codex 权限：沙箱和审批是两个开关</title><link>https://aiguide.zenty.top/articles/codex-permissions/</link><guid isPermaLink="true">https://aiguide.zenty.top/articles/codex-permissions/</guid><description>把 Codex 的沙箱边界和审批策略分开理解，用只读与工作区写入两组任务观察拒绝行为，并学会从路径、网络和保护目录三个方向诊断权限失败。</description><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><category>codex</category><category>Codex</category><category>权限</category><category>沙箱</category><category>审批</category><category>安全</category></item><item><title>Codex CLI 第一次任务：从一条指令到测试验收</title><link>https://aiguide.zenty.top/articles/codex-first-task/</link><guid isPermaLink="true">https://aiguide.zenty.top/articles/codex-first-task/</guid><description>用一个能跑测试的最小仓库走完 Codex CLI 完整链路：只读理解、下达约束、审阅 diff、跑 node --test 验收，并给出卡住时的排查顺序。</description><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><category>codex</category><category>Codex</category><category>CLI</category><category>工作流</category><category>代码审阅</category></item><item><title>让 Codex 修一个失败的测试：先复现，再限定改动</title><link>https://aiguide.zenty.top/articles/codex-debugging-tests/</link><guid isPermaLink="true">https://aiguide.zenty.top/articles/codex-debugging-tests/</guid><description>用一个 Node 内置测试的小例子，把失败测试当作验收合同：先独立复现，再让 Codex 只改实现，最后用回归结果和 diff 审查确认没有偷改测试。</description><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><category>codex</category><category>Codex</category><category>测试</category><category>调试</category><category>Node.js</category><category>代码审查</category></item><item><title>Codex 项目上下文：AGENTS.md 写什么、config.toml 边界在哪</title><link>https://aiguide.zenty.top/articles/codex-project-context/</link><guid isPermaLink="true">https://aiguide.zenty.top/articles/codex-project-context/</guid><description>Codex 每次运行只读一次指令链。讲清全局与项目 AGENTS.md 的发现顺序、config.toml 的信任边界，以及如何把上下文控制在真正有用的几百字里。</description><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><category>codex</category><category>Codex</category><category>AGENTS.md</category><category>上下文工程</category><category>配置</category></item><item><title>Codex 非交互模式：结果、事件流和退出码要分开看</title><link>https://aiguide.zenty.top/articles/codex-noninteractive/</link><guid isPermaLink="true">https://aiguide.zenty.top/articles/codex-noninteractive/</guid><description>用 codex exec 把 Codex 放进脚本：区分 stdout 最终消息与 stderr 进度，使用 -o、--json 和 --output-schema，并对结构化结果做独立校验。</description><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><category>codex</category><category>Codex</category><category>自动化</category><category>CLI</category><category>JSON</category><category>CI</category></item><item><title>用 Git worktree 隔离两个 Codex 任务</title><link>https://aiguide.zenty.top/articles/codex-worktrees/</link><guid isPermaLink="true">https://aiguide.zenty.top/articles/codex-worktrees/</guid><description>从同一个已提交基线创建两个工作树，让 Codex 在独立分支上改代码；合并前验证、移除前保留改动，并看清工作树共享仓库历史而不是安全沙箱。</description><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><category>codex</category><category>Codex</category><category>Git</category><category>worktree</category><category>并行开发</category><category>分支</category></item><item><title>Claude Code 安装：按平台用原生安装器，再验证登录</title><link>https://aiguide.zenty.top/articles/claude-code-install/</link><guid isPermaLink="true">https://aiguide.zenty.top/articles/claude-code-install/</guid><description>对照官方系统要求选择原生安装方式，用 claude --version 和 claude doctor 验证可执行文件，区分订阅与 API 认证，最后在一个练习仓库完成只读首任务。</description><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><category>claude-code</category><category>Claude Code</category><category>安装</category><category>登录</category><category>CLI</category><category>环境配置</category></item><item><title>CLAUDE.md 与规则文件：让 Claude Code 记住项目约定</title><link>https://aiguide.zenty.top/articles/claude-code-memory/</link><guid isPermaLink="true">https://aiguide.zenty.top/articles/claude-code-memory/</guid><description>用项目级 CLAUDE.md 和带 paths 的 .claude/rules 规则组织长期指导，检查实际加载了哪些文件，并分清“上下文建议”和“硬权限”的边界。</description><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><category>claude-code</category><category>Claude Code</category><category>CLAUDE.md</category><category>上下文</category><category>规则</category><category>项目配置</category></item><item><title>Claude Code 日常流程：/init、plan 模式与一次干净的提交</title><link>https://aiguide.zenty.top/articles/claude-code-workflow/</link><guid isPermaLink="true">https://aiguide.zenty.top/articles/claude-code-workflow/</guid><description>把 Claude Code 用成稳定工具而不是抽奖：先用 /init 建立项目记忆，用 plan 模式先看方案，按权限模式逐级放权，最后用 diff 和自己的一条验证命令收尾。</description><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><category>claude-code</category><category>Claude Code</category><category>工作流</category><category>权限模式</category><category>CLAUDE.md</category></item><item><title>Claude Code 权限：模式、规则和拒绝的优先级</title><link>https://aiguide.zenty.top/articles/claude-code-permissions/</link><guid isPermaLink="true">https://aiguide.zenty.top/articles/claude-code-permissions/</guid><description>分清权限模式、allow/ask/deny 规则和工具可用性，理解 deny 先于 ask 先于 allow 的匹配顺序，并用项目设置和只读验证排查“为什么被拒绝”。</description><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><category>claude-code</category><category>Claude Code</category><category>权限</category><category>安全</category><category>设置</category><category>工具</category></item><item><title>把本地 stdio MCP 服务器接到 Claude Code</title><link>https://aiguide.zenty.top/articles/claude-code-mcp/</link><guid isPermaLink="true">https://aiguide.zenty.top/articles/claude-code-mcp/</guid><description>复用自己构建的 wordcount MCP 服务器，选择 local、project 或 user 作用域接入 Claude Code，用 /mcp 检查连接和工具，并守住环境变量与凭据边界。</description><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><category>claude-code</category><category>Claude Code</category><category>MCP</category><category>stdio</category><category>工具集成</category><category>配置</category></item><item><title>Claude Code 代码评审：用 /review 和 hooks 做成独立工序</title><link>https://aiguide.zenty.top/articles/claude-code-review/</link><guid isPermaLink="true">https://aiguide.zenty.top/articles/claude-code-review/</guid><description>用 /review、/security-review 和无头模式把评审做成一道独立工序：限定输入范围、要求可定位的问题、用 hook 拦住危险动作，最后人工裁决。</description><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><category>claude-code</category><category>Claude Code</category><category>代码评审</category><category>Hooks</category><category>CI</category></item><item><title>给 Claude Code 加一个只读评审子代理</title><link>https://aiguide.zenty.top/articles/claude-code-subagents/</link><guid isPermaLink="true">https://aiguide.zenty.top/articles/claude-code-subagents/</guid><description>在 .claude/agents 中定义一个只拥有 Read、Grep、Glob 的评审员，把 diff 作为输入，要求带文件行号的证据输出，并看清共享工作树不是安全隔离。</description><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><category>claude-code</category><category>Claude Code</category><category>子代理</category><category>代码评审</category><category>权限</category><category>Agent</category></item><item><title>Claude Code 故障排查：先只读检查，再做最小修复</title><link>https://aiguide.zenty.top/articles/claude-code-troubleshooting/</link><guid isPermaLink="true">https://aiguide.zenty.top/articles/claude-code-troubleshooting/</guid><description>按“症状—只读检查—最小修复—验证”的顺序处理命令找不到、登录失败、记忆未加载、权限被拒和 MCP 连接失败，不打印密钥，也不删除用户状态。</description><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><category>claude-code</category><category>Claude Code</category><category>故障排查</category><category>诊断</category><category>MCP</category><category>权限</category></item><item><title>MCP 架构概念：Host、Client、Server 与一次工具调用</title><link>https://aiguide.zenty.top/articles/mcp-concepts/</link><guid isPermaLink="true">https://aiguide.zenty.top/articles/mcp-concepts/</guid><description>用一次 count_words 调用串起 MCP 的 host、client、server、三类原语与 stdio/远程传输，建立排错前的概念地图。</description><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><category>mcp</category><category>MCP</category><category>架构</category><category>协议</category></item><item><title>MCP 安全边界：接一个 server 前要确认的四类权限</title><link>https://aiguide.zenty.top/articles/mcp-boundaries/</link><guid isPermaLink="true">https://aiguide.zenty.top/articles/mcp-boundaries/</guid><description>MCP 把外部能力接进 Agent，也把外部风险接进来。逐个划清四道边界：进程权限、返回数据、凭据作用域、传输通道，并给出可落地的收紧配置。</description><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><category>mcp</category><category>MCP</category><category>安全</category><category>权限</category><category>prompt injection</category></item><item><title>写第一个 MCP Server：stdio 工具、Inspector 验证与注册</title><link>https://aiguide.zenty.top/articles/mcp-first-server/</link><guid isPermaLink="true">https://aiguide.zenty.top/articles/mcp-first-server/</guid><description>用当前官方 SDK 写一个只做一件事的 MCP server，用 Inspector 验证协议层，再分别注册到 Claude Code 和 Codex，最后确认工具真的被调用而不是只在列表里出现。</description><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><category>mcp</category><category>MCP</category><category>Node.js</category><category>stdio</category><category>SDK</category></item><item><title>MCP 工具设计：把 count_words 的契约写窄</title><link>https://aiguide.zenty.top/articles/mcp-tool-design/</link><guid isPermaLink="true">https://aiguide.zenty.top/articles/mcp-tool-design/</guid><description>以现有 count_words 为例，设计窄名称、输入输出 schema、校验与授权边界，并把业务失败变成稳定的结构化结果。</description><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><category>mcp</category><category>MCP</category><category>工具设计</category><category>Schema</category></item><item><title>MCP 协议排错：从进程启动到 tools/call 的四层定位</title><link>https://aiguide.zenty.top/articles/mcp-debugging/</link><guid isPermaLink="true">https://aiguide.zenty.top/articles/mcp-debugging/</guid><description>用 Inspector CLI 把 MCP 故障拆成进程、握手、tools/list、tools/call 四层，覆盖 stdout 污染、cwd、环境变量与 schema 问题。</description><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><category>mcp</category><category>MCP</category><category>Inspector</category><category>调试</category></item><item><title>第一个 Agent Skill：写一份可被发现的 change-summary</title><link>https://aiguide.zenty.top/articles/skills-first-skill/</link><guid isPermaLink="true">https://aiguide.zenty.top/articles/skills-first-skill/</guid><description>从目录、frontmatter 到完整 SKILL.md，一步步创建 change-summary，并用 Claude Code 的发现路径做手工验收。</description><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><category>skills</category><category>Agent Skills</category><category>SKILL.md</category><category>Claude Code</category></item><item><title>Agent Skill 设计：什么该写成 Skill，什么不该</title><link>https://aiguide.zenty.top/articles/skills-design/</link><guid isPermaLink="true">https://aiguide.zenty.top/articles/skills-design/</guid><description>Skill 不是提示词收藏夹。讲清它和项目规则、MCP 工具、斜杠命令的分工，怎么写可被正确触发的 description，以及为什么同一个 Skill 在不同工具里表现并不一致。</description><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><category>skills</category><category>Skills</category><category>SKILL.md</category><category>上下文工程</category><category>可移植性</category></item><item><title>Skill 资源拆分：references、scripts 与按需加载</title><link>https://aiguide.zenty.top/articles/skills-resources/</link><guid isPermaLink="true">https://aiguide.zenty.top/articles/skills-resources/</guid><description>把 change-summary 拆成 SKILL.md、references、assets 和离线脚本，理解渐进加载、路径与工作目录，并跑通自测。</description><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><category>skills</category><category>Agent Skills</category><category>渐进加载</category><category>Node.js</category></item><item><title>Skill 验证：格式、触发与内容三道闸门</title><link>https://aiguide.zenty.top/articles/skills-validation/</link><guid isPermaLink="true">https://aiguide.zenty.top/articles/skills-validation/</guid><description>Skill 的失败大多不是内容错，而是格式不合规范、触发条件太宽或太窄、跨客户端行为不一致。用格式校验、触发矩阵、内容演练三道闸门把它们提前暴露。</description><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><category>skills</category><category>Skills</category><category>校验</category><category>触发测试</category><category>质量</category></item><item><title>团队 Skill 分发：版本固定、兼容矩阵与可回滚交接</title><link>https://aiguide.zenty.top/articles/skills-team-workflow/</link><guid isPermaLink="true">https://aiguide.zenty.top/articles/skills-team-workflow/</guid><description>把团队 Skill 纳入版本控制，审查改动范围、记录不可变 revision、维护兼容矩阵，并用 revert 安全回滚。</description><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><category>skills</category><category>Agent Skills</category><category>团队协作</category><category>版本控制</category></item><item><title>固定工作流还是 Agent 循环：用发布说明分流做决策</title><link>https://aiguide.zenty.top/articles/agent-vs-workflow/</link><guid isPermaLink="true">https://aiguide.zenty.top/articles/agent-vs-workflow/</guid><description>以 release note 分流为例，比较确定性工作流与模型驱动工具循环的边界、成本和可观测性，给出可落地的混合方案。</description><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><category>agent</category><category>AI Agent</category><category>工作流</category><category>发布工程</category></item><item><title>Agent 工具契约：输入、授权、错误与输出校验</title><link>https://aiguide.zenty.top/articles/agent-tool-contracts/</link><guid isPermaLink="true">https://aiguide.zenty.top/articles/agent-tool-contracts/</guid><description>用一个可离线运行的工单查询工具，拆开输入校验、独立授权、结构化业务错误和输出契约，并跑通九条断言。</description><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><category>agent</category><category>AI Agent</category><category>工具调用</category><category>JSON Schema</category></item><item><title>Agent 工具循环：重复调用、超时与失败预算怎么停</title><link>https://aiguide.zenty.top/articles/agent-tool-loop/</link><guid isPermaLink="true">https://aiguide.zenty.top/articles/agent-tool-loop/</guid><description>Agent 的本质是“调工具—读结果—再决策”的循环。真正决定它能不能上线的不是模型多聪明，而是重复调用、无进展、超时和成本这四类失败有没有被硬性限制。</description><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><category>agent</category><category>Agent</category><category>工具调用</category><category>失败预算</category><category>可靠性</category></item><item><title>Agent 运行追踪：trace、run、step 与脱敏日志</title><link>https://aiguide.zenty.top/articles/agent-tracing/</link><guid isPermaLink="true">https://aiguide.zenty.top/articles/agent-tracing/</guid><description>用一个离线 Node 追踪器串起 trace/run/step、重试关联、耗时与结果，在写入前脱敏，并诚实划清玩具与生产追踪的边界。</description><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><category>agent</category><category>AI Agent</category><category>可观测性</category><category>日志脱敏</category></item><item><title>Agent 评测：用可复跑的状态断言代替“看起来不错”</title><link>https://aiguide.zenty.top/articles/agent-evaluation/</link><guid isPermaLink="true">https://aiguide.zenty.top/articles/agent-evaluation/</guid><description>给多轮 Agent 设计一套能复跑的评测：用状态检查代替话术检查，把确定性断言、模型评分和人工校准分开，并让每次失败都能定位到具体一步。</description><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><category>agent</category><category>Agent</category><category>评测</category><category>回归</category><category>指标</category></item><item><title>用 GitHub Actions 给静态站点加检查与构建门禁</title><link>https://aiguide.zenty.top/articles/automation-ci-checks/</link><guid isPermaLink="true">https://aiguide.zenty.top/articles/automation-ci-checks/</guid><description>为 Astro 静态站点写一条最小 CI：只读权限、固定 Node、显式工作目录，依次跑 check、build 和 smoke test，并避开 PR 密钥风险。</description><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><category>automation</category><category>GitHub Actions</category><category>CI</category><category>静态站点</category></item><item><title>Agent 自动化流水线：接入 CI 的四个环节</title><link>https://aiguide.zenty.top/articles/automation-pipeline/</link><guid isPermaLink="true">https://aiguide.zenty.top/articles/automation-pipeline/</guid><description>一条能上生产的自动化流水线不是“一键全自动”，而是触发、准备、执行、复盘的串联。给出可复现的流水线骨架、失败预算、幂等要求与成本控制。</description><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><category>automation</category><category>自动化</category><category>CI</category><category>Agent</category><category>流水线</category></item><item><title>GitHub Actions 密钥边界：注入、脱敏与产物卫生</title><link>https://aiguide.zenty.top/articles/automation-secrets/</link><guid isPermaLink="true">https://aiguide.zenty.top/articles/automation-secrets/</guid><description>说清仓库/环境/组织密钥的作用域、fork PR 权限、env 注入与脱敏的局限，并用只输出布尔结果的步骤验证配置。</description><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><category>automation</category><category>GitHub Actions</category><category>密钥管理</category><category>供应链安全</category></item><item><title>自动化幂等：重复事件、状态机与有界重试</title><link>https://aiguide.zenty.top/articles/automation-idempotency/</link><guid isPermaLink="true">https://aiguide.zenty.top/articles/automation-idempotency/</guid><description>用离线 Node 示例实现幂等键、in_progress/succeeded/failed 状态、瞬时与永久错误、指数退避，并说明内存演示为何不是 exactly-once。</description><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><category>automation</category><category>幂等性</category><category>Webhook</category><category>Node.js</category></item><item><title>自动化人工检查点：放在哪、怎么拦、失效会怎样</title><link>https://aiguide.zenty.top/articles/automation-checkpoints/</link><guid isPermaLink="true">https://aiguide.zenty.top/articles/automation-checkpoints/</guid><description>全自动不是目标，可控才是。讲清检查点该放在哪些决策点、钩子为什么只能当提示层、真正的门禁为什么必须在 Agent 之外，以及“默认放行”的检查点比没有更危险。</description><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><category>automation</category><category>自动化</category><category>人工审核</category><category>hooks</category><category>权限</category></item></channel></rss>