17 个月，
346 次更新 17 Months,
346 Updates

"当「用户」取代「全局」成为最高作用域，产品传递的信号是：工具服务于人，而非项目。" "When 'user' replaces 'global' as the top scope, the signal is clear: tools serve people, not projects."

"Task 工具从这天起有了「手」——Claude Code 第一次可以自己写文件、跑命令。" "From this release on, the Task tool has hands — Claude Code can write files and run commands on its own."

"这不是一次功能上新，而是 v1.0 之前最后一周的管道活。" "Not a feature drop — the last week of plumbing before v1.0."

"v1.0 前一天，旧的默认和旧的习惯一起被清掉。" "The day before v1.0, the old defaults and the old habits get swept away together."

"三个月，从 research preview 到 v1.0——以及说出「一般可用」这四个字的勇气。" "Three months from research preview to v1.0 — and the nerve to finally say "generally available.""

"/context 终于走进 Bedrock 和 Vertex，企业部署的最后一道墙拆掉。" "/context finally reaches Bedrock and Vertex — the last wall around enterprise deploys comes down."

"原生 VS Code 扩展、/rewind、SDK 改名同日落地——一次塞进去的三件大事。" "Native VS Code, /rewind, and an SDK rename — three big moves crammed into one day."

"v2.0 三天后，IME 误发、OAuth 401、SDK 重放——三个最尴尬的洞先补上。" "Three days post-v2.0, the three most embarrassing holes — IME mis-submits, OAuth 401s, SDK message replays — get plugged first."

"旧 .claude.json 里的四个老字段一次性下架，迁移没有缓冲期。" "Four legacy fields evicted from .claude.json in one shot — no grace period."

"改名三周后，旧 SDK 入口被直接拆走，迁移再无退路。" "Three weeks after the rename, the old SDK entrypoint is yanked — migration, no more delays."

"Nx 事件之后，dangerouslyDisableSandbox 这扇逃生门第一次能从策略层焊死。" "Post-Nx, the dangerouslyDisableSandbox escape hatch can finally be welded shut at the policy layer."

"脚本跑完了不知道该不该退？现在有一个环境变量替你做决定。" "Scripts that don't know when to leave finally get an env var to decide for them."

"这是 Claude Code 从命令行工具走向全平台编程体验的关键一步，也是 Opus 级模型能力首次完整注入开发者日常工作流。" "This is the moment Claude Code steps beyond the terminal into a full-platform experience, with Opus-class intelligence fully woven into the developer's daily workflow."

"长会话的内存占用一夜砍掉三分之二，每天被打字清空一次的输入框也终于不再清空。" "Long sessions cost a third of the memory they used to, and your half-typed prompt stops vanishing every time a queued command lands."

"2.1 的开门第一刀打在自己身上——debug 日志会泄露 OAuth token。" "2.1 opens by patching itself — the debug log that could spill OAuth tokens, API keys, and passwords is finally sealed."

"长输出不再被截断——它们落到磁盘上，由 Claude 自己回头读完整版。" "Long outputs stop being chopped at the edge; they land on disk and Claude reads the whole thing back."

"在 home 目录点「信任」却没真的解锁 hooks 的那条沉默漏洞，今天被修掉。" "The trust dialog that said yes from your home directory but never actually unlocked hooks finally tells the truth."

"`Bash(npm *)` 原本能放过 `npm install && rm -rf .`——今天的第二个补丁堵的是这个。" "`Bash(npm *)` was quietly waving through `npm install && rm -rf .`. The day's second security patch closes that door."

"一个反斜杠加换行就能绕过权限的旧把戏，今天被堵掉——而它的孪生补丁同天就到。" "A backslash and a newline used to be enough to walk around the permission system. Today that trick dies — and its twin patch ships hours later."

"并行工具调用甩出孤儿 tool_result 的那个 bug，今天被堵上了。" "The bug where parallel tool calls spat out an orphan tool_result finally gets sealed off."

"/rewind 这次面向所有 VSCode 用户开门，对话可以拉回任意一步重走。" "/rewind opens up to every VSCode user — pull the conversation back to any step and try again."

"这不是一次换代，而是推理透明度的第一次悄然让渡。" "Not just a model bump — the first quiet handoff of reasoning transparency."

"新模型登场的同一天，团队选择先把 Agent Teams 跑不起来的洞补上。" "On the day a new model arrives, the team spends it patching the holes Agent Teams kept falling through."

"模型曾经能用一句虚构的「这条命令匹配了规则 X」给自己开闸——这条路今天被堵上。" "The model used to wave commands through by fabricating which rule they matched. That door gets nailed shut today."

"后台代理第一次有了自己的工作树，主线程不再被它们踩。" "Background agents get their own working tree for the first time — the main thread stops getting trampled."

"符号链接路径下，会话恢复时整段历史会凭空消失——今天这条路被堵上。" "Under a symlinked working directory, resumed sessions used to disappear into thin air. Not anymore."

"外部系统第一次有了官方的「钩进 Claude Code」入口。" "External systems get their first official way to reach into a Claude Code session."

"Claude Code 第一次把「用户代理决定权」让渡给模型的自动判断。" "The first time Claude Code traded user agency for model-side auto-tuning."

"云上的 `model: opus` 一直在被悄悄降级到老版本——这条路今天截住了。" "On the cloud providers, `model: opus` had been silently downgraded to older versions. Cut off today."

"1M 上下文从「按用量解锁」改成订阅默认，门槛被一脚踢平。" "1M context shifts from a metered unlock to the subscription default — the gate gets kicked over."

"Remote Control 的四个静默故障一次性按掉——这次它真的能挂在长会话上了。" "Four silent failures in Remote Control all get pinned down at once — long sessions can actually live on it now."

"一个 PreToolUse 钩子曾经能用 `"allow"` 一句话推翻企业 deny 规则——这扇门今天被焊上。" "A PreToolUse hook used to be able to override an enterprise deny rule with a single "allow". Welded shut today."

"源代码被翻了个底朝天的那一周，团队选择把每一处可滑出的边界各打一颗钉。" "The week the source got picked apart in public, the team spent it driving a nail into every boundary that could slip."

"这是 Claude Code 第一次遭遇「用户信任撤回」这类无法靠补丁收尾的问题。" "The first Claude Code incident that couldn't be resolved by a patch alone — once trust leaks, it doesn't stop on command."

"Linux sandbox 一度漏发了 seccomp helper，unix socket 拦截整个失效——今天补回来。" "The Linux sandbox had been shipping without its seccomp helper, leaving unix-socket blocking wide open. Restored today."

"当代理能执行任意命令时，权限模型的严谨程度决定了产品能否被企业真正采用。" "When an agent can execute arbitrary commands, the rigor of the permission model determines whether enterprises will actually adopt it."

"当工具开始处理真实代码库时，可靠性不再是加分项，而是生存前提。" "Once a tool touches real repositories, reliability isn't a bonus—it's a survival requirement."

"真正的企业级产品不是靠功能清单赢得信任，而是靠修补那些用户从未察觉却日夜承受的隐性缺陷。" "Enterprise trust isn't earned by feature lists but by fixing the invisible defects users never notice yet silently endure."

"企业级采用暴露的不是功能缺口，而是并发凭证管理这一基础设施级信任问题。" "Enterprise adoption exposes not feature gaps, but an infrastructure-grade trust problem in concurrent credential management."

"交互模式从「逐轮对话」转向「目标驱动执行」，Claude Code 真正成为可以委派任务的自主代理。" "Interaction shifts from turn-by-turn dialogue to goal-driven execution, making Claude Code a delegate-worthy autonomous agent."

"当工具能在防火墙、VPN 和权限沙箱中依然流畅运行，开发者才真正敢于将其嵌入任意工作流。" "When a tool runs smoothly behind firewalls, VPNs, and permission sandboxes, developers finally trust it inside any workflow."

"Claude Code 从代码生成工具进化为代码质量把关者，直接嵌入 Pull Request 审查流程。" "Claude Code evolves from a code generator into a code-quality gatekeeper, embedding itself directly into the pull request review workflow."

"当审查与修复合为一体、自治模式不再需要许可，Claude Code 跨过了从顾问到维护者的最后一道门槛。" "By merging review with remediation and removing the consent gate on autonomous mode, Claude Code crosses the last threshold from advisor to maintainer."

"多代理系统的可靠性不取决于最强的一环，而取决于最弱的一环——此版本将安全与配置治理的边界推进到了此前被忽略的子代理深处。" "A multi-agent system is only as reliable as its weakest link — this release pushes security and configuration governance into the subagent depths that were previously overlooked."

"Claude Code 不再只是一个编程助手，而是一个能自我扩展的代理调度平台。" "Claude Code ceases to be a coding assistant and becomes a self-scaling agent orchestration platform."

"代理的自主权越大，每一次写入越需要用户的明确授权——这是从"能做"到"被允许做"的治理转向。" "The greater the agent's autonomy, the more each write requires explicit user authorization — a governance shift from "can do" to "permitted to do.""

"第一个 Mythos 级、可公开使用的模型落地——能力天花板被整体抬高了一档。" "The first Mythos-class model cleared for general use — the capability ceiling moves up a whole tier."

"大幅提升默认模型的上下文处理能力和推理表现，影响所有用户。" "Significantly improves default context capacity and reasoning for all users."

"当代理能自主走完从编码到提 PR 的全流程，人类评审者才真正成为唯一的人类环节。" "When agents autonomously own the path from code to pull request, human review becomes the only remaining human step."

"当默认行为从信任转向确认，产品宣告了其成熟阶段的核心立场：能力越强，越需要人类在场。" "When the default shifts from trust to confirmation, the product declares its mature-stage stance: the greater the capability, the more essential the human presence."

"当代理的自主性开始带来不可预期的开销，显式的努力校准成为信任的基石。" "When agent autonomy starts incurring unpredictable overhead, explicit effort calibration becomes the foundation of trust."

"当后台会话不再需要用户时刻盯着才能活下去，编程代理才真正具备了无人值守的资格。" "When background sessions no longer need a human watching to stay alive, coding agents earn the right to run unattended."

"当自主执行成为默认能力时，安全边界必须先于功能一步到位。" "When autonomous execution becomes the default, security boundaries must arrive one step ahead of the feature."

"当代理能力足够强时，真正的产品挑战从"能做什么"转向"如何安全地做"。" "When an agent is powerful enough, the real product challenge shifts from "what it can do" to "how it does so safely.""