🤖 Claude Code 更新 v2.1.76 发布日期: 2026-03-17 📋 版本维护 ✨ 新增功能 Added MCP elicitation support — MCP servers can now request structured input mid-task via an interactive dialog (form fields or browser URL) Added new Elicitation and ElicitationResult hooks to intercept and override responses before they’re sent back Added -n / --name <name> CLI flag to set a display name for the session at startup Added worktree.sparsePaths setting for claude --worktree in large monorepos to check out only the directories you need via git sparse-checkout Added PostCompact hook that fires after compaction completes Added /effort slash command to set model effort level Added session quality survey — enterprise admins can configure the sample rate via the feedbackSurveyRate setting 🐛 重要修复 Fixed deferred tools (loaded via ToolSearch) losing their input schemas after conversation compaction, causing array and number parameters to be rejected with type errors Fixed slash commands showing “Unknown skill” Fixed plan mode asking for re-approval after the plan was already accepted Fixed voice mode swallowing keypresses while a permission dialog or plan editor was open Fixed /voice not working on Windows when installed via npm Fixed spurious “Context limit reached” when invoking a skill with model: frontmatter on a 1M-context session Fixed “adaptive thinking is not supported on this model” error when using non-standard model strings Fixed Bash(cmd:*) permission rules not matching when a quoted argument contains # Fixed “don’t ask again” in the Bash permission dialog showing the full raw command for pipes and compound commands Fixed auto-compaction retrying indefinitely after consecutive failures — a circuit breaker now stops after 3 attempts ⚡ 优化改进 Improved --worktree startup performance by reading git refs directly and skipping redundant git fetch when the remote branch is already available locally Improved background agent behavior — killing a background agent now preserves its partial results in the conversation context Improved model fallback notifications — now always visible instead of hidden behind verbose mode, with human-friendly model names Improved blockquote readability on dark terminal themes — text is now italic with a left bar instead of dim Improved stale worktree cleanup — worktrees left behind after an interrupted parallel run are now automatically cleaned up Improved Remote Control session titles — now derived from your first prompt instead of showing “Interactive session” Improved /voice to show your dictation language on enable and warn when your language setting isn’t supported for voice input Updated --plugin-dir to only accept one path to support subcommands — use repeated --plugin-dir for multiple directories [VSCode] Fixed gitignore patterns containing commas silently excluding entire filetypes from the @-mention file picker 💡 深度点评 核心亮点 MCP Elicitation 支持：MCP server 现在可以在任务执行过程中通过交互式对话（如表单字段或浏览器 URL）请求结构化输入，增强了复杂任务中的人机协作能力。 Monorepo 性能优化：引入 worktree.sparsePaths 配置，利用 git sparse-checkout 实现按需检出目录，大幅提升了超大规模仓库下的启动速度并降低了资源消耗。 任务控制增强：新增 /effort 命令支持手动调节模型的推理力度，配合 -n 启动标志自定义会话显示名称，进一步细化了开发者的操作权限。 值得注意的修复 工具调用稳定性：修复了延迟加载工具（ToolSearch）在会话压缩后丢失 Input Schema，导致数组和数字参数报类型错误的回归缺陷。 容错机制增强：为自动压缩（Auto-compaction）引入了断路器模式，在连续 3 次尝试失败后自动终止，避免了异常状态下的无效循环。 个人评价 v2.1.76 版本是一个侧重于“工程健壮性”的迭代。它不仅解决了大规模代码库下的性能瓶颈，还通过完善 MCP 协议和会话压缩逻辑，显著提升了长时会话的可靠性。这些针对底层架构和边缘场景的打磨，标志着 Claude Code 正在从简单的 AI 助手演进为更专业、可预测的生产力工具。 ...

🔧 Codex 更新 rust-v0.114.0 发布日期: 2026-03-16 📋 版本维护 New Features Added an experimental code mode for more isolated coding workflows. (#13418) Added an experimental hooks engine with SessionStart and Stop hook events. (#13276) WebSocket app-server deployments now expose GET /readyz and GET /healthz on the same listener for easier health checks. (#13782) Added a config switch to disable bundled system skills entirely. (#13792) Handoffs now carry realtime transcript context, which improves continuity when work is transferred between turns. (#14132) Improved the $ mention picker by clearly labeling Skills, Apps, and Plugins, and by surfacing plugins first. (#14147, #14163) Bug Fixes Fixed a Linux tmux crash caused by concurrent user-shell lookups. (#13900) Fixed apps being enabled in unsupported sessions by tightening the enablement check. (#14011) Fixed reopened threads getting stuck as in-progress after quitting mid-run and then resuming later. (#14125) Fixed permission handling so legacy workspace-write behavior is preserved and newer permission profiles degrade more safely on older builds. (#13957, #14107) Fixed approval flows so granted permissions persist across turns, work with reject-style configs, and are honored by apply_patch. (#14009, #14055, #14118, #14165) Chores Laid the groundwork for the Python SDK’s generated v2 schema types and pinned platform-specific runtime binaries. (#13953) Changelog #14009 feat(core) Persist request_permission data across turns @dylan-hurd-oai #14136 fix(core): use dedicated types for responsesapi web search tool config @owenlin0 #13782 codex-rs/app-server: add health endpoints for –listen websocket server @maxj-oai #14055 fix(core) RequestPermissions + ApplyPatch @dylan-hurd-oai #14118 feat(approvals) RejectConfig for request_permissions @dylan-hurd-oai #13957 fix(protocol): preserve legacy workspace-write semantics @viyatb-oai #14107 fix: keep permissions profiles forward compatible @viyatb-oai #14147 make dollar-mention always clarify item category (skill, app, plugin) @sayan-oai #14152 Refactor tool output into trait implementations @pakrym-oai #14163 sort plugins first in menu @sayan-oai #13418 Add code_mode experimental feature @pakrym-oai #13276 start of hooks engine @eternal-openai 💡 深度点评 作为 Codex Rust 版本的又一次快速迭代，v0.114.0 在工程化体验和底层架构上做了多项关键调整。以下是针对该版本的深度分析： ...

🔧 Openclaw 更新 2026.3.13 发布日期: 2026-03-16 📋 版本维护 Changes Android/chat settings: redesign the chat settings sheet with grouped device and media sections, refresh the Connect and Voice tabs, and tighten the chat composer/session header for a denser mobile layout. (#44894) Thanks @obviyus. iOS/onboarding: add a first-run welcome pager before gateway setup, stop auto-opening the QR scanner, and show /pair qr instructions on the connect step. (#45054) Thanks @ngutman. Browser/existing-session: add an official Chrome DevTools MCP attach mode for signed-in live Chrome sessions, with docs for chrome://inspect/#remote-debugging enablement and direct backlinks to Chrome’s own setup guides. Browser/agents: add built-in profile="user" for the logged-in host browser and profile="chrome-relay" for the extension relay, so agent browser calls can prefer the real signed-in browser without the extra browserSession selector. Browser/act automation: add batched actions, selector targeting, and delayed clicks for browser act requests with normalized batch dispatch. Thanks @vincentkoc. Docker/timezone override: add OPENCLAW_TZ so docker-setup.sh can pin gateway and CLI containers to a chosen IANA timezone instead of inheriting the daemon default. (#34119) Thanks @Lanfei. Dependencies/pi: bump @mariozechner/pi-agent-core, @mariozechner/pi-ai, @mariozechner/pi-coding-agent, and @mariozechner/pi-tui to 0.58.0. Cron/sessions: add sessionTarget: "current" and session:<id> support so cron jobs can bind to the creating session or a persistent named session instead of only main or isolated. Thanks @kkhomej33-netizen and @ImLukeF. Telegram/message send: add --force-document so Telegram image and GIF sends can upload as documents without compression. (#45111) Thanks @thepagent. Breaking BREAKING: Agents now load at most one root memory bootstrap file. MEMORY.md wins; memory.md is only used when MEMORY.md is absent. If you intentionally kept both files and depended on both being injected, merge them before upgrade. This also fixes duplicate memory injection on case-insensitive Docker mounts. (#26054) Thanks @Lanfei. Fixes Dashboard/chat UI: stop reloading full chat history on every live tool result in dashboard v2 so tool-heavy runs no longer trigger UI freeze/re-render storms while the final event still refreshes persisted history. (#45541) Thanks @BunsDev. Gateway/client requests: reject unanswered gateway RPC calls after a bounded timeout and clear their pending state, so stalled connections no longer leak hanging GatewayClient.request() promises indefinitely. Build/plugin-sdk bundling: bundle plugin-sdk subpath entries in one shared build pass so published packages stop duplicating shared chunks and avoid the recent plugin-sdk memory blow-up. (#45426) Thanks @TarasShyn. Ollama/reasoning visibility: stop promoting native thinking and reasoning fields into final assistant text so local reasoning models no longer leak internal thoughts in normal replies. (#45330) Thanks @xi7ang. Android/onboarding QR scan: switch setup QR scanning to Google Code Scanner so onboarding uses a more reliable scanner instead of the legacy embedded ZXing flow. (#45021) Thanks @obviyus. Browser/existing-session: harden driver validation and session lifecycle so transport errors trigger reconnects while tool-level errors preserve the session, and extract shared ARIA role sets to deduplicate Playwright and Chrome MCP snapshot paths. (#45682) Thanks @odysseus0. Browser/existing-session: accept text-only list_pages and new_page responses from Chrome DevTools MCP so live-session tab discovery and new-tab open flows keep working when the server omits structured page metadata. Control UI/insecure auth: preserve explicit shared token and password auth on plain-HTTP Control UI connects so LAN and reverse-proxy sessions no longer drop shared auth before the first WebSocket handshake. (#45088) Thanks @velvet-shark. Gateway/session reset: preserve lastAccountId and lastThreadId across gateway session resets so replies keep routing back to the same account and thread after /reset. (#44773) Thanks @Lanfei. macOS/onboarding: avoid self-restarting freshly bootstrapped launchd gateways and give new daemon installs longer to become healthy, so openclaw onboard --install-daemon no longer false-fails on slower Macs and fresh VM snapshots. Gateway/status: add openclaw gateway status --require-rpc and clearer Linux non-interactive daemon-install failure reporting so automation can fail hard on probe misses instead of treating a printed RPC error as green. macOS/exec approvals: respect per-agent exec approval settings in the gateway prompter, including allowlist fallback when the native prompt cannot be shown, so gateway-triggered system.run requests follow configured policy instead of always prompting or denying unexpectedly. (#13707) Thanks @sliekens. 💡 深度点评 核心亮点 浏览器原生会话挂载（Chrome DevTools MCP）：新增官方 Chrome DevTools MCP 挂载模式，支持通过 chrome://inspect 直接接入已登录的实时 Chrome 会话。配合内置的 profile="user" 属性，Agent 现在可以直接在用户的主浏览器环境内执行任务，无需额外的 browserSession 选择器，极大简化了复杂 Web 任务的上下文继承。 浏览器自动化指令集增强：针对 browser/act 引入了批量操作（batched actions）、选择器定位优化以及延迟点击功能。通过标准化的批处理调度，Agent 在处理动态 Web 页面时的响应速度和动作准确度有了显著提升。 Cron 任务会话绑定：sessionTarget 现在支持 current 或特定 session:<id>。这意味着定时任务不再局限于孤立环境或主会话，而是可以灵活绑定到当前活跃会话或持久化的命名会话中，实现了自动化流与人工交互流的深度协同。 值得注意的修复 Dashboard 渲染性能优化：修复了在执行工具密集型（tool-heavy）任务时，UI 频繁重载全量历史记录导致的重绘风暴和界面冻结问题。现在只有最终状态会触发持久化历史刷新，大幅提升了长链条任务的交互流畅度。 本地推理模型「思维泄露」修复：针对 Ollama 等具备思考能力的模型，优化了 thinking 和 reasoning 字段的处理逻辑，防止内部推理过程泄露到最终的助理回复正文中，保证了输出结果的纯净性。 多平台安全审批加固：针对 macOS 和 Windows 的 exec 审批逻辑进行了深度重构，能够识别并拆解 pnpm、PowerShell 以及 env 包装器下的真实可执行路径。同时，MEMORY.md 权重的明确化解决了 Docker 挂载时大小写不敏感带来的内存注入重复问题。 个人评价 openclaw 2026.3.13 是一个典型的「工程化落地」版本，核心价值在于消弭了 Agent 与用户真实工作环境（如已登录浏览器、活跃会话）之间的鸿沟。通过对浏览器自动化指令的批量化改版和 Dashboard 渲染瓶颈的修复，开发者能够明显感受到在大规模工具调用场景下的稳定性提升。整体演进方向正从「对话式助理」快速向「深度嵌入工作流的系统级插件」转变。 ...

🤖 Claude Code 更新 v2.1.76 发布日期: 2026-03-16 📋 版本维护 ✨ 新增功能 Added MCP elicitation support — MCP servers can now request structured input mid-task via an interactive dialog (form fields or browser URL) Added new Elicitation and ElicitationResult hooks to intercept and override responses before they’re sent back Added -n / --name <name> CLI flag to set a display name for the session at startup Added worktree.sparsePaths setting for claude --worktree in large monorepos to check out only the directories you need via git sparse-checkout Added PostCompact hook that fires after compaction completes Added /effort slash command to set model effort level Added session quality survey — enterprise admins can configure the sample rate via the feedbackSurveyRate setting 🐛 重要修复 Fixed deferred tools (loaded via ToolSearch) losing their input schemas after conversation compaction, causing array and number parameters to be rejected with type errors Fixed slash commands showing “Unknown skill” Fixed plan mode asking for re-approval after the plan was already accepted Fixed voice mode swallowing keypresses while a permission dialog or plan editor was open Fixed /voice not working on Windows when installed via npm Fixed spurious “Context limit reached” when invoking a skill with model: frontmatter on a 1M-context session Fixed “adaptive thinking is not supported on this model” error when using non-standard model strings Fixed Bash(cmd:*) permission rules not matching when a quoted argument contains # Fixed “don’t ask again” in the Bash permission dialog showing the full raw command for pipes and compound commands Fixed auto-compaction retrying indefinitely after consecutive failures — a circuit breaker now stops after 3 attempts ⚡ 优化改进 Improved --worktree startup performance by reading git refs directly and skipping redundant git fetch when the remote branch is already available locally Improved background agent behavior — killing a background agent now preserves its partial results in the conversation context Improved model fallback notifications — now always visible instead of hidden behind verbose mode, with human-friendly model names Improved blockquote readability on dark terminal themes — text is now italic with a left bar instead of dim Improved stale worktree cleanup — worktrees left behind after an interrupted parallel run are now automatically cleaned up Improved Remote Control session titles — now derived from your first prompt instead of showing “Interactive session” Improved /voice to show your dictation language on enable and warn when your language setting isn’t supported for voice input Updated --plugin-dir to only accept one path to support subcommands — use repeated --plugin-dir for multiple directories [VSCode] Fixed gitignore patterns containing commas silently excluding entire filetypes from the @-mention file picker 💡 深度点评 大家好，我是技术博主。今天我们来深度梳理一下 Claude Code v2.1.76 的最新更新。 ...

🔧 Codex 更新 rust-v0.114.0 发布日期: 2026-03-15 📋 版本维护 New Features Added an experimental code mode for more isolated coding workflows. (#13418) Added an experimental hooks engine with SessionStart and Stop hook events. (#13276) WebSocket app-server deployments now expose GET /readyz and GET /healthz on the same listener for easier health checks. (#13782) Added a config switch to disable bundled system skills entirely. (#13792) Handoffs now carry realtime transcript context, which improves continuity when work is transferred between turns. (#14132) Improved the $ mention picker by clearly labeling Skills, Apps, and Plugins, and by surfacing plugins first. (#14147, #14163) Bug Fixes Fixed a Linux tmux crash caused by concurrent user-shell lookups. (#13900) Fixed apps being enabled in unsupported sessions by tightening the enablement check. (#14011) Fixed reopened threads getting stuck as in-progress after quitting mid-run and then resuming later. (#14125) Fixed permission handling so legacy workspace-write behavior is preserved and newer permission profiles degrade more safely on older builds. (#13957, #14107) Fixed approval flows so granted permissions persist across turns, work with reject-style configs, and are honored by apply_patch. (#14009, #14055, #14118, #14165) Chores Laid the groundwork for the Python SDK’s generated v2 schema types and pinned platform-specific runtime binaries. (#13953) Changelog #14009 feat(core) Persist request_permission data across turns @dylan-hurd-oai #14136 fix(core): use dedicated types for responsesapi web search tool config @owenlin0 #13782 codex-rs/app-server: add health endpoints for –listen websocket server @maxj-oai #14055 fix(core) RequestPermissions + ApplyPatch @dylan-hurd-oai #14118 feat(approvals) RejectConfig for request_permissions @dylan-hurd-oai #13957 fix(protocol): preserve legacy workspace-write semantics @viyatb-oai #14107 fix: keep permissions profiles forward compatible @viyatb-oai #14147 make dollar-mention always clarify item category (skill, app, plugin) @sayan-oai #14152 Refactor tool output into trait implementations @pakrym-oai #14163 sort plugins first in menu @sayan-oai #13418 Add code_mode experimental feature @pakrym-oai #13276 start of hooks engine @eternal-openai 💡 深度点评 核心亮点 实验性 Hooks 引擎与 Code Mode 落地：版本引入了 SessionStart 和 Stop 钩子事件，允许开发者在会话生命周期内注入自定义逻辑。配合新增的实验性 Code Mode，Codex 正在尝试提供更纯净、隔离的编码专用工作流，这对于追求高确定性输出的任务至关重要。 会话移交（Handoff）上下文连贯性优化：现在的任务移交能够携带实时转录（Realtime Transcript）上下文。这一改进解决了多轮对话或复杂任务切换时的“断层”感，确保后继步骤能精准继承前序状态，减少了重复指令的输入。 工具发现机制与交互优化：$ 提及菜单（Mention Picker）进行了重构，通过明确标注 Skill、App 和 Plugin 分类，并优先置顶插件，显著提升了开发者在复杂环境下的工具调用效率。此外，允许完全禁用系统内置技能，为构建高度定制化的垂直领域 Agent 提供了可能。 值得注意的修复 权限持久化与审批流修复：修复了权限请求无法跨回合持久化的问题，并确保 apply_patch 等关键操作能正确遵循权限配置（如拒绝逻辑），解决了长会话中频繁授权的体验痛点。 会话恢复状态同步：解决了线程在中断并重新打开后可能卡死在“进行中”状态的 Bug，提升了在不稳定网络或手动中断场景下的任务恢复可靠性。 Linux 环境稳定性：修复了 Linux 下 tmux 因并发用户 Shell 查询导致的崩溃（Segfault），增强了该工具在重度终端用户环境中的健壮性。 个人评价 Codex Rust-v0.114.0 是一个更偏向“生产力基建”的迭代版本。Hooks 引擎的加入标志着它正从一个被动工具向可编程的任务引擎演进。重点优化权限处理和移交上下文，显示出开发团队正在深挖长任务、复杂流场景下的用户痛点。整体价值取向非常清晰：在保持交互灵活性的同时，通过底层重构提供更严谨的权限控制和更稳定的执行环境。 ...

🔧 Openclaw 更新 2026.3.13 发布日期: 2026-03-15 ⚠️ 新版本发布 Changes Android/chat settings: redesign the chat settings sheet with grouped device and media sections, refresh the Connect and Voice tabs, and tighten the chat composer/session header for a denser mobile layout. (#44894) Thanks @obviyus. iOS/onboarding: add a first-run welcome pager before gateway setup, stop auto-opening the QR scanner, and show /pair qr instructions on the connect step. (#45054) Thanks @ngutman. Browser/existing-session: add an official Chrome DevTools MCP attach mode for signed-in live Chrome sessions, with docs for chrome://inspect/#remote-debugging enablement and direct backlinks to Chrome’s own setup guides. Browser/agents: add built-in profile="user" for the logged-in host browser and profile="chrome-relay" for the extension relay, so agent browser calls can prefer the real signed-in browser without the extra browserSession selector. Browser/act automation: add batched actions, selector targeting, and delayed clicks for browser act requests with normalized batch dispatch. Thanks @vincentkoc. Docker/timezone override: add OPENCLAW_TZ so docker-setup.sh can pin gateway and CLI containers to a chosen IANA timezone instead of inheriting the daemon default. (#34119) Thanks @Lanfei. Dependencies/pi: bump @mariozechner/pi-agent-core, @mariozechner/pi-ai, @mariozechner/pi-coding-agent, and @mariozechner/pi-tui to 0.58.0. Cron/sessions: add sessionTarget: "current" and session:<id> support so cron jobs can bind to the creating session or a persistent named session instead of only main or isolated. Thanks @kkhomej33-netizen and @ImLukeF. Telegram/message send: add --force-document so Telegram image and GIF sends can upload as documents without compression. (#45111) Thanks @thepagent. Breaking BREAKING: Agents now load at most one root memory bootstrap file. MEMORY.md wins; memory.md is only used when MEMORY.md is absent. If you intentionally kept both files and depended on both being injected, merge them before upgrade. This also fixes duplicate memory injection on case-insensitive Docker mounts. (#26054) Thanks @Lanfei. Fixes Dashboard/chat UI: stop reloading full chat history on every live tool result in dashboard v2 so tool-heavy runs no longer trigger UI freeze/re-render storms while the final event still refreshes persisted history. (#45541) Thanks @BunsDev. Gateway/client requests: reject unanswered gateway RPC calls after a bounded timeout and clear their pending state, so stalled connections no longer leak hanging GatewayClient.request() promises indefinitely. Build/plugin-sdk bundling: bundle plugin-sdk subpath entries in one shared build pass so published packages stop duplicating shared chunks and avoid the recent plugin-sdk memory blow-up. (#45426) Thanks @TarasShyn. Ollama/reasoning visibility: stop promoting native thinking and reasoning fields into final assistant text so local reasoning models no longer leak internal thoughts in normal replies. (#45330) Thanks @xi7ang. Android/onboarding QR scan: switch setup QR scanning to Google Code Scanner so onboarding uses a more reliable scanner instead of the legacy embedded ZXing flow. (#45021) Thanks @obviyus. Browser/existing-session: harden driver validation and session lifecycle so transport errors trigger reconnects while tool-level errors preserve the session, and extract shared ARIA role sets to deduplicate Playwright and Chrome MCP snapshot paths. (#45682) Thanks @odysseus0. Browser/existing-session: accept text-only list_pages and new_page responses from Chrome DevTools MCP so live-session tab discovery and new-tab open flows keep working when the server omits structured page metadata. Control UI/insecure auth: preserve explicit shared token and password auth on plain-HTTP Control UI connects so LAN and reverse-proxy sessions no longer drop shared auth before the first WebSocket handshake. (#45088) Thanks @velvet-shark. Gateway/session reset: preserve lastAccountId and lastThreadId across gateway session resets so replies keep routing back to the same account and thread after /reset. (#44773) Thanks @Lanfei. macOS/onboarding: avoid self-restarting freshly bootstrapped launchd gateways and give new daemon installs longer to become healthy, so openclaw onboard --install-daemon no longer false-fails on slower Macs and fresh VM snapshots. Gateway/status: add openclaw gateway status --require-rpc and clearer Linux non-interactive daemon-install failure reporting so automation can fail hard on probe misses instead of treating a printed RPC error as green. macOS/exec approvals: respect per-agent exec approval settings in the gateway prompter, including allowlist fallback when the native prompt cannot be shown, so gateway-triggered system.run requests follow configured policy instead of always prompting or denying unexpectedly. (#13707) Thanks @sliekens. 💡 深度点评 这份关于 openclaw 2026.3.13 的更新说明显示，该版本在浏览器深度集成、自动化任务调度以及系统安全审计方面迈出了实质性的一步。以下是本次更新的深度点评： ...

🤖 Claude Code 更新 v2.1.76 发布日期: 2026-03-15 📋 版本维护 ✨ 新增功能 Added MCP elicitation support — MCP servers can now request structured input mid-task via an interactive dialog (form fields or browser URL) Added new Elicitation and ElicitationResult hooks to intercept and override responses before they’re sent back Added -n / --name <name> CLI flag to set a display name for the session at startup Added worktree.sparsePaths setting for claude --worktree in large monorepos to check out only the directories you need via git sparse-checkout Added PostCompact hook that fires after compaction completes Added /effort slash command to set model effort level Added session quality survey — enterprise admins can configure the sample rate via the feedbackSurveyRate setting 🐛 重要修复 Fixed deferred tools (loaded via ToolSearch) losing their input schemas after conversation compaction, causing array and number parameters to be rejected with type errors Fixed slash commands showing “Unknown skill” Fixed plan mode asking for re-approval after the plan was already accepted Fixed voice mode swallowing keypresses while a permission dialog or plan editor was open Fixed /voice not working on Windows when installed via npm Fixed spurious “Context limit reached” when invoking a skill with model: frontmatter on a 1M-context session Fixed “adaptive thinking is not supported on this model” error when using non-standard model strings Fixed Bash(cmd:*) permission rules not matching when a quoted argument contains # Fixed “don’t ask again” in the Bash permission dialog showing the full raw command for pipes and compound commands Fixed auto-compaction retrying indefinitely after consecutive failures — a circuit breaker now stops after 3 attempts ⚡ 优化改进 Improved --worktree startup performance by reading git refs directly and skipping redundant git fetch when the remote branch is already available locally Improved background agent behavior — killing a background agent now preserves its partial results in the conversation context Improved model fallback notifications — now always visible instead of hidden behind verbose mode, with human-friendly model names Improved blockquote readability on dark terminal themes — text is now italic with a left bar instead of dim Improved stale worktree cleanup — worktrees left behind after an interrupted parallel run are now automatically cleaned up Improved Remote Control session titles — now derived from your first prompt instead of showing “Interactive session” Improved /voice to show your dictation language on enable and warn when your language setting isn’t supported for voice input Updated --plugin-dir to only accept one path to support subcommands — use repeated --plugin-dir for multiple directories [VSCode] Fixed gitignore patterns containing commas silently excluding entire filetypes from the @-mention file picker 💡 深度点评 核心亮点 MCP Elicitation 支持：MCP 服务端现在可以在任务执行过程中通过交互式对话框（表单或浏览器 URL）主动请求结构化输入。配合新增的 Elicitation 钩子，这为复杂工具链的「人机协作」提供了更标准化的干预机制。 Monorepo 性能优化：引入 worktree.sparsePaths 配置，支持通过 git sparse-checkout 仅检出必要的目录。配合 --worktree 启动性能的改进，大幅降低了在大型超大规模仓库（Monorepo）中使用时的资源开销和等待时间。 精细化控制增强：新增 /effort 命令允许用户手动设定模型的工作强度（Effort Level），同时支持通过 -n 参数为 Session 命名。这些改动提升了开发者在处理不同复杂度任务时的掌控力。 值得注意的修复 长对话稳定性修复：解决了延迟加载工具（ToolSearch）在上下文压缩（Compaction）后丢失输入 Schema 的问题，避免了长 Session 中频繁出现的参数类型错误。 远程控制与连接优化：修复了 Remote Control 模式下 Session 意外挂起、消息积压以及 WebSocket 断连后无法恢复等稳定性缺陷，显著提升了远程协作场景的可靠性。 个人评价 Claude Code v2.1.76 是一个聚焦于「生产环境可靠性」与「大规模工程适配」的迭代版本。通过引入 sparse-checkout 和完善上下文压缩机制，它进一步解决了在复杂工程场景下的性能瓶颈与长效对话的逻辑断层。此外，MCP 交互能力的增强意味着它正从单纯的指令执行器向更智能的协同平台演进。对于重度依赖远程环境或处于大型单体仓库的企业级开发者而言，这一版本的稳定性改进极具实战价值。 ...

everything-claude-code（ECC）、BMAD-METHOD 和 Superpowers 代表了 AI 编程工具增强的三种不同路径：ECC 是工具链性能层，旨在压榨 Agent 的执行上限；BMAD 是项目管理方法论，将敏捷开发引入 AI 时代；Superpowers 是最小化工作流约束系统，通过强制 TDD 纪律减少随意性。三者解决的问题维度不同，不存在绝对优劣。 三条路，三种答案 当前 AI 编程工具的扩展正处于爆炸期，开发者们不仅在追求更强的模型，更在思考如何通过 skills、agents、hooks 和 workflows 构建更高质量的工程闭环。在这场竞赛中，三个开源项目各代表了不同的演进方向。 ECC (everything-claude-code)：Anthropic Hackathon 获奖项目，76k stars。它通过一套极其复杂的工程化配置，解决了 AI Agent 在长上下文管理、Token 效率和跨 session 记忆方面的痛点。 BMAD-METHOD：40.7k stars，专注于敏捷 AI 开发的方法论。它不只是工具增强，而是一套完整的项目生命周期管理框架，包含从需求分析到部署的标准化流程。 Superpowers：由 Jesse (obra) 发起，已上架官方 Claude 插件市场。主张"流程即法律"，通过自动触发的 Skills 强制执行 TDD 和 Git 工作流。 本文不是评选"冠军"，而是通过系统视角剖析它们的设计哲学，帮助工程师根据自己的项目规模和协作习惯建立选择框架。 三个项目的系统定位 这三个框架在 AI 编程的生态栈中处于不同的层次。 ECC：工具链性能优化层。 它不关心你写的是什么业务，它关心的是"如何让 AI 执行得更稳"。通过内置的 instincts、memory hooks 和 security scan，它为 Claude Code 或 Codex 等 Harness 提供了一个高性能的运行环境。核心价值在于 Token 优化和记忆持久化，解决"Agent 越用越笨"的问题。 ...

Kubernetes 官方 Dashboard 在 K8s 1.32 版本上存在兼容性问题（如 CronJob 页面 404）。如果你只需要一个简单的 Web UI 来管理本地 K8s 集群，kube-explorer 是一个轻量、无依赖的替代方案。 为什么选择 kube-explorer 特性 kube-explorer 官方 Dashboard 依赖 单二进制，无依赖 需部署多个组件 K8s 兼容性 广泛兼容 新版本存在兼容问题 安装方式 本地运行 / 集群部署 仅集群部署 认证 自动读取 kubeconfig 需要 Token 方式一：本地运行 适合 Docker Desktop / kind / minikube 等本地 K8s 环境。 下载并安装 # macOS (ARM64) curl -sL -o ~/bin/kube-explorer "https://github.com/cnrancher/kube-explorer/releases/download/v0.3.0/kube-explorer-darwin-arm64" chmod +x ~/bin/kube-explorer # macOS (Intel) curl -sL -o ~/bin/kube-explorer "https://github.com/cnrancher/kube-explorer/releases/download/v0.3.0/kube-explorer-darwin-amd64" chmod +x ~/bin/kube-explorer # Linux (amd64) curl -sL -o /usr/local/bin/kube-explorer "https://github.com/cnrancher/kube-explorer/releases/download/v0.3.0/kube-explorer-linux-amd64" chmod +x /usr/local/bin/kube-explorer 启动服务 kube-explorer --http-listen-port=9090 --https-listen-port=0 访问 http://localhost:9090 即可使用。 ...

🤖 Claude Code 更新 v2.1.75 发布日期: 2026-03-14 ⚠️ 重大更新 ✨ 新增功能 Added 1M context window for Opus 4.6 by default for Max, Team, and Enterprise plans (previously required extra usage) Added /color command for all users to set a prompt-bar color for your session Added session name display on the prompt bar when using /rename Added last-modified timestamps to memory files, helping Claude reason about which memories are fresh vs. stale Added hook source display (settings/plugin/skill) in permission prompts when a hook requires confirmation Added actionable suggestions to /context command — identifies context-heavy tools, memory bloat, and capacity warnings with specific optimization tips Added autoMemoryDirectory setting to configure a custom directory for auto-memory storage Added modelOverrides setting to map model picker entries to custom provider model IDs (e.g. Bedrock inference profile ARNs) 🐛 重要修复 Fixed voice mode not activating correctly on fresh installs without toggling /voice twice Fixed the Claude Code header not updating the displayed model name after switching models with /model or Option+P Fixed session crash when an attachment message computation returns undefined values Fixed Bash tool mangling ! in piped commands (e.g., jq 'select(.x != .y)' now works correctly) Fixed managed-disabled plugins showing up in the /plugin Installed tab — plugins force-disabled by your organization are now hidden Fixed token estimation over-counting for thinking and tool_use blocks, preventing premature context compaction Fixed corrupted marketplace config path handling Fixed /resume losing session names after resuming a forked or continued session ⚡ 优化改进 Improved startup performance on macOS non-MDM machines by skipping unnecessary subprocess spawns [VSCode] Fixed delete button not working for Untitled sessions [VSCode] Improved scroll wheel responsiveness in the integrated terminal with terminal-aware acceleration Improved Up arrow after interrupting Claude — now restores the interrupted prompt and rewinds the conversation in one step Improved IDE detection speed at startup Improved clipboard image pasting performance on macOS Improved /effort to work while Claude is responding, matching /model behavior Improved voice mode to automatically retry transient connection failures during rapid push-to-talk re-press 💡 深度点评 🎯 核心亮点 1. 定时调度工具 — 新增 cron 调度能力，支持在会话内设置周期性任务。 ...

🤖 Claude Code 更新 v2.1.74 发布日期: 2026-03-13 ⚠️ 重大更新 ✨ 新增功能 Added actionable suggestions to /context command — identifies context-heavy tools, memory bloat, and capacity warnings with specific optimization tips Added autoMemoryDirectory setting to configure a custom directory for auto-memory storage Added modelOverrides setting to map model picker entries to custom provider model IDs (e.g. Bedrock inference profile ARNs) Added actionable guidance when OAuth login or connectivity checks fail due to SSL certificate errors (corporate proxies, NODE_EXTRA_CA_CERTS) Added w key in /copy to write the focused selection directly to a file, bypassing the clipboard (useful over SSH) Added optional description argument to /plan (e.g., /plan fix the auth bug) that enters plan mode and immediately starts Added ExitWorktree tool to leave an EnterWorktree session Added CLAUDE_CODE_DISABLE_CRON environment variable to immediately stop scheduled cron jobs mid-session 🐛 重要修复 Fixed memory leak where streaming API response buffers were not released when the generator was terminated early, causing unbounded RSS growth on the Node.js/npm code path Fixed managed policy ask rules being bypassed by user allow rules or skill allowed-tools Fixed full model IDs (e.g., claude-opus-4-5) being silently ignored in agent frontmatter model: field and --agents JSON config — agents now accept the same model values as --model Fixed MCP OAuth authentication hanging when the callback port is already in use Fixed MCP OAuth refresh never prompting for re-auth after the refresh token expires, for OAuth servers that return errors with HTTP 200 (e.g. Slack) Fixed voice mode silently failing on the macOS native binary for users whose terminal had never been granted microphone permission — the binary now includes the audio-input entitlement so macOS prompts correctly Fixed SessionEnd hooks being killed after 1.5 s on exit regardless of hook.timeout — now configurable via CLAUDE_CODE_SESSIONEND_HOOKS_TIMEOUT_MS Fixed /plugin install failing inside the REPL for marketplace plugins with local sources ⚡ 优化改进 [VSCode] Fixed delete button not working for Untitled sessions [VSCode] Improved scroll wheel responsiveness in the integrated terminal with terminal-aware acceleration Improved Up arrow after interrupting Claude — now restores the interrupted prompt and rewinds the conversation in one step Improved IDE detection speed at startup Improved clipboard image pasting performance on macOS Improved /effort to work while Claude is responding, matching /model behavior Improved voice mode to automatically retry transient connection failures during rapid push-to-talk re-press Improved the Remote Control spawn mode selection prompt with better context 💡 深度点评 🎯 核心亮点 1. 定时调度工具 — 新增 cron 调度能力，支持在会话内设置周期性任务。 ...

🤖 Claude Code 更新 v2.1.73 发布日期: 2026-03-12 ⚠️ 重大更新 ✨ 新增功能 Added modelOverrides setting to map model picker entries to custom provider model IDs (e.g. Bedrock inference profile ARNs) Added actionable guidance when OAuth login or connectivity checks fail due to SSL certificate errors (corporate proxies, NODE_EXTRA_CA_CERTS) Added w key in /copy to write the focused selection directly to a file, bypassing the clipboard (useful over SSH) Added optional description argument to /plan (e.g., /plan fix the auth bug) that enters plan mode and immediately starts Added ExitWorktree tool to leave an EnterWorktree session Added CLAUDE_CODE_DISABLE_CRON environment variable to immediately stop scheduled cron jobs mid-session Added lsof, pgrep, tput, ss, fd, and fdfind to the bash auto-approval allowlist, reducing permission prompts for common read-only operations Added support for marketplace git URLs without .git suffix (Azure DevOps, AWS CodeCommit) 🐛 重要修复 Fixed freezes and 100% CPU loops triggered by permission prompts for complex bash commands Fixed a deadlock that could freeze Claude Code when many skill files changed at once (e.g. during git pull in a repo with a large .claude/skills/ directory) Fixed Bash tool output being lost when running multiple Claude Code sessions in the same project directory Fixed subagents with model: opus/sonnet/haiku being silently downgraded to older model versions on Bedrock, Vertex, and Microsoft Foundry Fixed background bash processes spawned by subagents not being cleaned up when the agent exits Fixed /resume showing the current session in the picker Fixed /ide crashing with onInstall is not defined when auto-installing the extension Fixed /loop not being available on Bedrock/Vertex/Foundry and when telemetry was disabled ⚡ 优化改进 Improved Up arrow after interrupting Claude — now restores the interrupted prompt and rewinds the conversation in one step Improved IDE detection speed at startup Improved clipboard image pasting performance on macOS Improved /effort to work while Claude is responding, matching /model behavior Improved voice mode to automatically retry transient connection failures during rapid push-to-talk re-press Improved the Remote Control spawn mode selection prompt with better context Improved /config — Escape now cancels changes, Enter saves and closes, Space toggles settings Improved up-arrow history to show current session’s messages first when running multiple concurrent sessions 💡 深度点评 🎯 核心亮点 1. 定时调度工具 — 新增 cron 调度能力，支持在会话内设置周期性任务。 ...

Andrej Karpathy 最近发布的 autoresearch 仓库在机器学习社区引发了关注。表面上看，这是一个自动搜索 Transformer 最优超参和架构的工具，但从系统设计视角来看，它更像是一个关于"如何设计真正自主的 Agent 系统"的完整示范。 大多数 AI Agent 开发者仍受困于"如何让 LLM 稳定执行复杂任务"的泥潭，而 autoresearch 通过极其克制的代码结构，回答了 Agent 自主性的三个核心问题：目标如何定义、边界如何划定、反馈如何保证客观。 这个系统仅由三个核心文件支撑：prepare.py、train.py 和 program.md。这种三层分离的架构，而非复杂的框架嵌套，才是其设计哲学的精髓。 三文件架构：冻结层、可变层与元层 在 autoresearch 中，Agent 的操作空间被严格限定。 prepare.py（冻结层）：定义评估标准、数据集和强制性的时间预算，Agent 无权修改。这确保了实验的基座稳固，Agent 无法通过修改数据或放宽测试标准来"作弊"。 train.py（可变层）：这是 Agent 唯一的操作对象。Agent 通过读取、修改并运行这个文件来验证自己的猜想。 program.md（元层）：人类写给 Agent 的"系统宪法"。它不包含具体的编程步骤，而是规定了运行逻辑、错误处理策略、结果记录格式以及"永不停止"的指令。 这种三层设计让 Agent 拥有了在受控环境下的完全自主权。它不需要询问用户"下一步做什么"，因为它在宪法约束下，在冻结层基准上，不断迭代可变层。 固定预算：从限制到设计 autoresearch 强制要求每个实验只能运行 5 分钟。在追求极致性能的 ML 领域，这看起来像是妥协，实则是高明的设计。 固定时间预算提供了绝对公平的比较前提。无论 Agent 把模型改得更深还是更宽，引入了怎样的优化算法，性能提升必须在 5 分钟内体现。这迫使 Agent 寻找的是"计算效率最优解"，而非堆砌算力的暴力解。 系统采用 val_bpb（验证集每字节位数）作为评价指标，与词表大小和分词方式无关。这保证了评估函数不可被 Agent 通过架构选择来"游戏化"。有了客观且难以作弊的反馈，Agent 的 keep/discard 决策才具备可信赖的意义。 Agent 发现了什么：时间约束下的帕累托最优 MLX 社区（Apple Silicon 移植版）的公开实验结果很能说明问题。Agent 将默认的 8 层深度砍到 4 层，val_bpb 从 2.667 降到了 1.808，单次实验降幅接近 28%。 ...

传统的 Java 压测诊断流程通常是这样的：打开监控大盘、发现异常、SSH 进跳板机、kubectl exec 进 Pod、attach Arthas、手工敲命令、拼结论。这个流程的问题不是步骤多，而是每一步都依赖人工，压测窗口很短，诊断常常跟不上问题出现的节奏。 Arthas MCP 把这件事往前推了一步——让 AI Agent 可以直接调用 Arthas 的诊断工具，把"人工执行诊断命令"这件事自动化掉。 但在远端 K8s 场景下，开始动手才会发现：真正的挑战不在 MCP 的几个参数配置，而在于如何构建一条从本地 Agent 穿透到 Pod 内部的安全链路。 本文的答案是：把链路复杂度封装在连接层，让 Agent 始终看到一个稳定的 localhost 地址。 远端 K8s 的核心障碍 在本地环境里，Arthas MCP 的接入几乎没有门槛——启动 Arthas、指定 httpPort 和 mcpEndpoint，Agent 直接连。 但在远端 K8s 里，有三件事会让这条路走不通： 网络隔离：Pod 处于 K8s 内网，本地无法直接访问 Pod IP。 安全边界：把 Arthas HTTP 端口通过 LoadBalancer 或公网 Ingress 暴露出来，等于把高权限诊断入口开放给外部，风险极高。 动态性：Pod 名称和 IP 随时可能变化，Agent 侧配置无法写死。 所以问题的本质是：如何用一套对 K8s 改动最小、安全边界清晰的方式，把 Pod 内部的 Arthas MCP 端口安全地带到本地。 ...

🤖 Claude Code 更新 v2.1.72 发布日期: 2026-03-11 ⚠️ 重大更新 ✨ 新增功能 Added w key in /copy to write the focused selection directly to a file, bypassing the clipboard (useful over SSH) Added optional description argument to /plan (e.g., /plan fix the auth bug) that enters plan mode and immediately starts Added ExitWorktree tool to leave an EnterWorktree session Added CLAUDE_CODE_DISABLE_CRON environment variable to immediately stop scheduled cron jobs mid-session Added lsof, pgrep, tput, ss, fd, and fdfind to the bash auto-approval allowlist, reducing permission prompts for common read-only operations Added support for marketplace git URLs without .git suffix (Azure DevOps, AWS CodeCommit) Added /loop command to run a prompt or slash command on a recurring interval (e.g. /loop 5m check the deploy) Added cron scheduling tools for recurring prompts within a session 🐛 重要修复 Fixed tool search to activate even with ANTHROPIC_BASE_URL as long as ENABLE_TOOL_SEARCH is set. Fixed slow exits when background tasks or hooks were slow to respond Fixed agent task progress stuck on “Initializing…” Fixed skill hooks firing twice per event when a hooks-enabled skill is invoked by the model Fixed several voice mode issues: occasional input lag, false “No speech detected” errors after releasing push-to-talk, and stale transcripts re-filling the prompt after submission Fixed --continue not resuming from the most recent point after --compact Fixed bash security parsing edge cases Fixed several plugin issues: installation failing on Windows with EEXIST error in OneDrive folders, marketplace blocking user-scope installs when a project-scope install exists, CLAUDE_CODE_PLUGIN_CACHE_DIR creating literal ~ directories, and plugin.json with marketplace-only fields failing to load ⚡ 优化改进 Improved /config — Escape now cancels changes, Enter saves and closes, Space toggles settings Improved up-arrow history to show current session’s messages first when running multiple concurrent sessions Improved voice input transcription accuracy for repo names and common dev terms (regex, OAuth, JSON) Improved bash command parsing by switching to a native module — faster initialization and no memory leak Improved marketplace clone failure messages to show diagnostic info even when git produces no stderr Improved CPU utilization in long sessions Improved startup time by deferring native image processor loading to first use Improved bridge session reconnection to complete within seconds after laptop wake from sleep, instead of waiting up to 10 minutes 💡 深度点评 🎯 核心亮点 1. 定时调度工具 — 新增 cron 调度能力，支持在会话内设置周期性任务。 ...

上一篇我已经把核心结论讲清楚了： 远端 K8s 里的 Arthas MCP，不应该直接暴露给本地 Agent，而应该通过 SSH + kubectl port-forward，在本地制造一个安全、稳定、看起来像本地的 MCP 地址。 但光有这个结论还不够。 因为只要你真的开始用，就会立刻遇到一堆工程层问题： 每次都手工敲一堆命令，太容易出错 Pod 名称会变 压测前要快速切换实例 链路断了之后，很难判断是 SSH 问题、kubectl 问题，还是 Arthas 根本没起来 本地 Agent 需要的不只是“知道原理”，而是一个稳定可复用的本地 MCP 地址 所以这一篇不再讲抽象架构，而是专注一件事： 如何把本地 Agent → 远端 K8s Pod 内 Arthas MCP 的接入流程，做成一键脚本，并真正用于压测实战。 本文配套脚本： ~/.openclaw/workspace/scripts/connect-arthas-mcp.sh 一、脚本到底解决了什么问题 先看没有脚本时，你通常要做哪些事。 手工方式 登录远端 master / bastion 找到目标 namespace 和 Pod 执行 kubectl port-forward pod/<pod> 8563:8563 在本地再执行 ssh -L 18563:127.0.0.1:8563 user@host 手工 curl 检查 /mcp 再复制一段 MCP 配置到本地客户端 这套流程最大的问题不是“步骤多”，而是： ...

Arthas 最近放出了 MCP Server 能力。这个变化的意义，不只是“Arthas 多了一个新入口”，而是： Java 诊断能力，第一次可以被本地 Agent 以标准工具协议直接调用。 以前我们做性能测试，常见流程是这样的： 压测工具把流量打上去 我们盯着监控看 CPU、RT、GC 发现异常后，再 SSH 到机器里 attach Arthas 手工跑 thread、dashboard、trace、watch 最后把零散信息拼成结论 这个流程不是不能用，但它有两个问题： 慢：压测窗口很短，手工分析经常跟不上问题出现的节奏 碎：数据、判断和命令是割裂的，很难形成可复用流程 Arthas MCP 把这件事往前推进了一步：AI Agent 可以直接把 Arthas 当工具调用。 于是问题就变成了： 如果我的 Java 服务跑在远端 K8s 集群里，本地 Agent 怎么安全地连上远端 Arthas MCP，并在压测时参与分析？ 这篇文章我想把这件事讲透，重点回答四个问题： Arthas MCP 到底解决了什么问题 远端 K8s 场景下，链路应该怎么设计 最推荐的接入方式是什么 我如何把这件事做成一键脚本，并真正用于压测分析 一、Arthas MCP 到底是什么 根据 Arthas 官方文档，Arthas MCP Server 是 Arthas 的实验性模块，实现了基于 MCP（Model Context Protocol） 的服务端能力，通过 HTTP / Netty + JSON-RPC 2.0 对外暴露工具接口。 ...

最近看到一篇公众号文章在聊 OpenAI 刚开源的 Symphony。如果只看标题，很容易把它理解成“又一个 AI 编程神器”。但我把它的 GitHub 仓库、SPEC.md 和 Elixir 参考实现说明都过了一遍之后，感觉这项目真正值得看的地方，不是“AI 会不会写代码”，而是： OpenAI 正在尝试把 AI 编程从“对话式辅助”推进到“任务级执行”。 一句话总结： Symphony 不是 IDE 里的代码助手，而是一个围绕 issue、workspace、coding agent、PR、CI 和人工审批构建的任务编排器。 项目地址： GitHub：https://github.com/openai/symphony 先说结论：Symphony 解决的是“管理工作”，不是“补全代码” 今天大多数 AI 编程工具的典型使用方式还是这样： 你在 IDE 或终端里提问 AI 给你一段实现 你审查、纠偏、继续补上下文 AI 再继续写 这种模式当然有效，但本质上还是 人工实时驾驶。 你虽然在用 AI，但注意力并没有真正被解放。你只是把“自己敲代码”变成了“盯着 AI 敲代码”。 Symphony 想解决的正是这个问题。它的目标不是让 AI 把某个函数写得更快，而是把一项工程工作变成一个 可调度、可隔离、可观察、可审批 的自动化执行单元。 也就是说，工程师关注的对象从： “这几行代码怎么写” 转向： “这个任务有没有被正确完成” “交付结果是否可信” “现在是否应该批准合并” 这背后其实是一种非常明确的范式变化： 从管理代码生成过程，转向管理任务完成结果。 Symphony 是怎么工作的 根据 README.md、SPEC.md 和 elixir/README.md，Symphony 的大致工作流程是这样的： 持续轮询任务系统（当前规范版本主要是 Linear） 找出符合条件的 issue 为每个 issue 创建一个独立 workspace 在 workspace 里启动 coding agent 按仓库中的 WORKFLOW.md 指令推进实现 让 agent 生成 PR、CI 状态、review 反馈等结果 把任务交给人工审批或流转到下一状态 如果任务状态改变，比如变成： ...

🤖 Claude Code 更新 v2.1.71 发布日期: 2026-03-10 **📋 版本维护 ✨ 新增功能 Added /loop command to run a prompt or slash command on a recurring interval (e.g. /loop 5m check the deploy) Added cron scheduling tools for recurring prompts within a session Added voice:pushToTalk keybinding to make the voice activation key rebindable in keybindings.json (default: space) — modifier+letter combos like meta+k have zero typing interference Added fmt, comm, cmp, numfmt, expr, test, printf, getconf, seq, tsort, and pr to the bash auto-approval allowlist Added the /claude-api skill for building applications with the Claude API and Anthropic SDK Added Ctrl+U on an empty bash prompt (!) to exit bash mode, matching escape and backspace Added numeric keypad support for selecting options in Claude’s interview questions (previously only the number row above QWERTY worked) Added optional name argument to /remote-control and claude remote-control (/remote-control My Project or --name "My Project") to set a custom session title visible in claude.ai/code 🐛 重要修复 Fixed stdin freeze in long-running sessions where keystrokes stop being processed but the process stays alive Fixed a 5–8 second startup freeze for users with voice mode enabled, caused by CoreAudio initialization blocking the main thread after system wake Fixed startup UI freeze when many claude.ai proxy connectors refresh an expired OAuth token simultaneously Fixed forked conversations (/fork) sharing the same plan file, which caused plan edits in one fork to overwrite the other Fixed the Read tool putting oversized images into context when image processing failed, breaking subsequent turns in long image-heavy sessions Fixed false-positive permission prompts for compound bash commands containing heredoc commit messages Fixed plugin installations being lost when running multiple Claude Code instances Fixed claude.ai connectors failing to reconnect after OAuth token refresh ⚡ 优化改进 Improved startup time by deferring native image processor loading to first use Improved bridge session reconnection to complete within seconds after laptop wake from sleep, instead of waiting up to 10 minutes Improved /plugin uninstall to disable project-scoped plugins in .claude/settings.local.json instead of modifying .claude/settings.json, so changes don’t affect teammates Improved plugin-provided MCP server deduplication — servers that duplicate a manually-configured server (same command/URL) are now skipped, preventing duplicate connections and tool sets. Suppressions are shown in the /plugin menu. Updated /debug to toggle debug logging on mid-session, since debug logs are no longer written by default Removed startup notification noise for unauthenticated org-registered claude.ai connectors Improved error message when microphone captures silence to distinguish from “no speech detected” Improved compaction to preserve images in the summarizer request, allowing prompt cache reuse for faster and cheaper compaction 💡 深度点评 🎯 核心亮点 1. 定时调度工具 — 新增 cron 调度能力，支持在会话内设置周期性任务。 ...

🤖 Claude Code 更新 v2.1.71 发布日期: 2026-03-09 **📋 版本维护 ✨ 新增功能 Added /loop command to run a prompt or slash command on a recurring interval (e.g. /loop 5m check the deploy) Added cron scheduling tools for recurring prompts within a session Added voice:pushToTalk keybinding to make the voice activation key rebindable in keybindings.json (default: space) — modifier+letter combos like meta+k have zero typing interference Added fmt, comm, cmp, numfmt, expr, test, printf, getconf, seq, tsort, and pr to the bash auto-approval allowlist Added the /claude-api skill for building applications with the Claude API and Anthropic SDK Added Ctrl+U on an empty bash prompt (!) to exit bash mode, matching escape and backspace Added numeric keypad support for selecting options in Claude’s interview questions (previously only the number row above QWERTY worked) Added optional name argument to /remote-control and claude remote-control (/remote-control My Project or --name "My Project") to set a custom session title visible in claude.ai/code 🐛 重要修复 Fixed stdin freeze in long-running sessions where keystrokes stop being processed but the process stays alive Fixed a 5–8 second startup freeze for users with voice mode enabled, caused by CoreAudio initialization blocking the main thread after system wake Fixed startup UI freeze when many claude.ai proxy connectors refresh an expired OAuth token simultaneously Fixed forked conversations (/fork) sharing the same plan file, which caused plan edits in one fork to overwrite the other Fixed the Read tool putting oversized images into context when image processing failed, breaking subsequent turns in long image-heavy sessions Fixed false-positive permission prompts for compound bash commands containing heredoc commit messages Fixed plugin installations being lost when running multiple Claude Code instances Fixed claude.ai connectors failing to reconnect after OAuth token refresh ⚡ 优化改进 Improved startup time by deferring native image processor loading to first use Improved bridge session reconnection to complete within seconds after laptop wake from sleep, instead of waiting up to 10 minutes Improved /plugin uninstall to disable project-scoped plugins in .claude/settings.local.json instead of modifying .claude/settings.json, so changes don’t affect teammates Improved plugin-provided MCP server deduplication — servers that duplicate a manually-configured server (same command/URL) are now skipped, preventing duplicate connections and tool sets. Suppressions are shown in the /plugin menu. Updated /debug to toggle debug logging on mid-session, since debug logs are no longer written by default Removed startup notification noise for unauthenticated org-registered claude.ai connectors Improved error message when microphone captures silence to distinguish from “no speech detected” Improved compaction to preserve images in the summarizer request, allowing prompt cache reuse for faster and cheaper compaction 💡 深度点评 🎯 核心亮点 1. 定时调度工具 — 新增 cron 调度能力，支持在会话内设置周期性任务。 ...