Field notes现场笔记
Field notes on coding agents, leftover cloud infrastructure, and the bills that outlive the chat.关于编程代理、会话留下的云资源，以及对话结束后仍在计费的现场笔记。
Latest最新
All articles全部文章Treat each AI agent as its own principal—scoped tokens, delegated permissions, one-agent revoke. Compare Mintive, Okta Govern AI Agent Identity, and WorkOS Agent Auth; draw the boundary vs an MCP gateway.把每个 AI agent 当作它自己的主体：范围受限的令牌、委托权限、单 agent 吊销。对比 Mintive、Okta Govern AI Agent Identity 与 WorkOS Agent Auth，并划清其与 MCP gateway 的边界。
Read article阅读文章
Stop guessing from viral 'digital snitch' posts. Decide by account type whether self-assessment runs, what retention applies, how Cyber Verification handles research false positives, and why Claude Code telemetry is not server-side assessment.别再凭疯传的「数字告密者」帖子去猜。按账户类型判断 self-assessment 是否运行、适用何种保留、Cyber Verification 如何处理研究误报，以及为什么 Claude Code 遥测不是服务端评估。
Decide which file owns a third-party hook, which official unload removes it, and which managed hooks a user disable cannot turn off after an orchestrator or desktop installer rewrites Claude, Codex, Cursor, or Copilot config.在编排器或桌面安装程序改写 Claude、Codex、Cursor 或 Copilot 配置之后，判断哪个文件拥有第三方钩子、哪一次官方卸除能移除它，以及哪些受管钩子无法被用户禁用关掉。
Decide which parallel-agent conflict class you are solving before spinning N worktrees. Compare deterministic intent claims (scope+operation), contract-first merges, ownership orchestration, and LLM mutual review—with Foremerge as a worked case, not a launch post.在拉起 N 个工作树之前，先判断你要解决的是哪一类并行智能体冲突。比较确定性意图声明（范围+操作）、契约优先的合并、所有权编排，以及 LLM 互审——以 Foremerge 作为实例，而不是发布帖。
Chrome ships fixes; Playwright pins Chromium later. Decide by workload—own-app E2E, agent untrusted browsing, Playwright MCP—using chromiumSandbox, channel chrome, non-root hosts, and egress allowlists. A decision matrix for Claude Code, Cursor, and MCP.Chrome 发布修复；Playwright 稍后才固定 Chromium。按工作负载做决定——自有应用 E2E、代理不受信任浏览、Playwright MCP——使用 chromiumSandbox、channel chrome、非 root 主机与出站允许列表。一份面向 Claude Code、Cursor 与 MCP 的决策矩阵。
Public GitHub repos block pull_request_target by default on December 2, 2026, with brownouts on November 9 and 16. Inventory workflows, switch CI to pull_request, and allow the trigger only for jobs that never execute fork code.公开的 GitHub 仓库将于 2026 年 12 月 2 日默认拦截 pull_request_target，并在 11 月 9 日与 16 日进行限时中断。盘点工作流，将 CI 切换到 pull_request，并仅对从不执行 fork 代码的作业放行该触发器。
Decide when a coding agent should share your real Firefox logins versus driving a separate automation browser. Compare Kitsune, firefox-bridge, Playwright MCP, and Claude in Chrome on session model, profiles, and approval gates—without pretending modes stop prompt injection.判断编程智能体应共享你真实的 Firefox 登录态，还是驱动一个独立的自动化浏览器。在会话模型、配置文件与审批关卡上比较 Kitsune、firefox-bridge、Playwright MCP 与 Claude in Chrome，同时不要假装模式能阻止提示注入。
Coding agents create VMs, databases, and deploys mid-task—then sessions end without teardown. A practical FinOps playbook for read-only audit, human-confirmed cleanup, Claude Code/Codex/MCP blind spots, and what --verify can (and cannot) prove.编程智能体会在任务中途创建虚拟机、数据库和部署，然后会话结束却没有拆除。一份实用的 FinOps 实操手册，涵盖只读审计、经人工确认的清理、Claude Code/Codex/MCP 盲区，以及 --verify 能够证明什么、不能证明什么。
A stub so the site has a second category. The FinOps essay is the real article on windtask.com.一个占位页，用来让站点有第二个分类。Windtask 上真正的文章是那篇 FinOps 随笔。
2 articles2 篇
Writing filed under Developer Tools.归在开发工具下的文章。
1 article1 篇
Cloud cost, orphan resources, and the habits that keep agent-created infrastructure from billing after the session ends.云成本、无人认领的资源，以及避免代理创建的基础设施在会话结束后继续计费的习惯。
1 article1 篇
Short site notes. This shelf exists so a second category has a real route.站点短记。这个分类用来保证第二个分类有一条真实路径。
1 article1 篇
Account-type decisions for AI retention and monitoring: what self-assessment actually runs, what stays, and what you can turn off.按账户类型决定 AI 的留存与监控：自评实际会运行什么、留下什么，以及你能关掉什么。
4 articles4 篇
Browser freshness, containment, and the trust decisions coding agents make when they open pages you did not write.浏览器新鲜度、隔离，以及编程代理打开并非你所写的页面时做出的信任判断。