每日版本追踪
Anthropic 的终端原生编程 Agent 与开发工作流工具。
This release focuses on stability by addressing bug fixes and reliability improvements, ensuring a more consistent experience for developers using Claude Code. Operators can expect fewer unexpected errors during automated workflows and code generation tasks.
本次更新主要修复了已知问题并提升了系统稳定性，为开发者提供更可靠的编码辅助体验。对于依赖自动化工作流的团队而言，这有助于减少运行时的异常中断，提升整体操作效率。
This update introduces Claude Opus 5 as the default model with a 1M context window and adds a strict sandbox allowlist for enhanced security. It also improves operational reliability through better self-hosted runner deregistration, structured failure categorization, and fixes for mid-stream API errors and permission drops.
本次更新将Claude Opus 5设为默认模型，支持1M上下文窗口，并新增沙箱网络的严格白名单设置以强化安全性。同时优化了自托管运行器的状态清理、结构化故障分类及API中断恢复，提升了开发者的操作稳定性和安全性。
This update enhances workflow stability by offloading code reviews to background subagents and fixes critical Windows path corruption and terminal encoding issues. It also improves accessibility with screen-reader support and resolves edge cases in MCP connectivity and token usage reporting.
本次更新通过将代码审查移至后台子代理来优化工作流，并修复了Windows路径损坏及终端编码等关键问题。同时增强了屏幕阅读器无障碍支持，并解决了MCP连接和令牌使用报告中的边缘情况。
This update resolves critical stability issues including a memory leak in MCP tool outputs and Windows auto-update failures that could corrupt the executable. It also enhances operational reliability by fixing corporate proxy/mTLS configurations and adding explicit warnings for transcript write failures.
该版本修复了MCP工具输出的内存泄漏及Windows自动更新导致的可执行文件损坏等关键稳定性问题。同时，修正了企业级代理和mTLS配置被忽略的缺陷，并增加了转录写入失败的显式警告，提升了运维可靠性。
This update resolves critical performance regressions in long sessions and OAuth token rotation errors in auto mode, while introducing a new sandbox setting to bypass filesystem isolation for specific use cases. It also fixes significant UX bugs regarding session resumption, background agent state restoration, and command execution permissions.
本次更新修复了长会话中的性能瓶颈及自动模式下的OAuth令牌过期问题，并新增沙箱文件系统隔离绕过设置。同时，解决了会话恢复、后台代理状态还原及命令执行权限检查等多个影响开发者体验的关键缺陷。
Claude Code v2.1.215 shifts the /verify and /code-review skills from automatic execution to explicit user invocation, requiring developers to manually trigger these commands when needed. This change gives teams greater control over workflow timing and prevents unintended automated reviews.
Claude Code v2.1.215 将 /verify 和 /code-review 技能从自动执行改为需用户手动调用，开发者需按需主动触发这些命令。此举赋予开发团队对工作流程节奏的更强控制权，避免不必要的自动化审查干扰。
This patch resolves critical permission-check bypasses in Bash and PowerShell, ensuring strict fail-closed behavior for complex redirects and long commands to prevent unauthorized file writes. It also introduces the EndConversation tool for handling abusive users and enhances OpenTelemetry logging with message-level correlation for better operational observability.
该版本修复了Bash和PowerShell中的权限检查绕过漏洞，对复杂重定向和超长命令实施严格的失败关闭策略，防止未经授权的写入操作。同时新增EndConversation工具以应对恶意用户，并增强OpenTelemetry日志的消息级关联能力，提升了AI代理的安全性与运维可观测性。
This update introduces critical stability controls by capping WebSearch and subagent spawns to prevent runaway loops, while refining session management with /fork backgrounding and MCP auto-backgrounding. It also resolves significant edge-case bugs in plan mode permissions, shell mode execution, and Windows daemon compatibility.
本次更新通过限制WebSearch和子代理调用次数来防止无限循环，增强了会话稳定性；同时优化了/fork后台处理和MCP超时机制。此外，修复了计划模式权限、Shell模式执行及Windows环境下的多项关键Bug，提升了开发体验。