OpenClaw v2026.6.5-beta.1

## 2026.6.5 ### 亮点 - QQBot 现在会在原生投递前剥离模型推理/思考框架，防止原始 `<thinking>` 内容泄露到频道回复中。(#89913, #90132) 感谢 @openperf。 - MCP 工具结果现在会在物化边界强制转换 `resource_link`、`resource`、`audio`、格式错误的图像以及未来的非文本/图像块，防止工具返回更丰富的 MCP 内容后出现 Anthropic 400 错误和中毒的会话历史。(#90710, #90728) 感谢 @RanSHammer 和 @849261680。 - Anthropic 扩展思考会话在提示缓存过期或网关重启后能够恢复，因为流启动事件等待 `message_start`，使得预生成签名错误能够触发现有的恢复重试。(#90667, #90697) 感谢 @openperf。 - Parallel 现在是一个捆绑的 `web_search` 提供商，支持 `PARALLEL_API_KEY` 发现、受保护的端点处理、缓存安全的会话 ID、入门选择器支持和文档。(#85158) 感谢 @NormallyGaussian。 - Google Vertex ADC 用户再次获得静态目录行和运行时模型解析，同时单提供商冷却恢复和内存适配器状态检查更加可靠。(#90506, #90609, #90717, #90816) 感谢 @849261680。 - Matrix 现在可以在提及门控之前预检语音笔记，通过 Matrix 关系分页保留线程读取/回复，并为语音和线程流程提供 QA 覆盖。(#78016, #90415) - 认证和插件安装状态更加持久：认证配置文件现在存储在 SQLite 中，官方 npm 插件安装记录保留其受信任的固定版本，预发布回退完整性检查避免携带过时的完整性信息。(#89102, #88585) - macOS 节点模式不再静默地从健康的直接网关会话自动重新连接，减少了意外的配套应用会话切换。(#90668, #90815) 感谢 @vrurg。 - 升级和服务路径更加安全：cron 遗留 JSON 存储在 doctor 预检期间迁移，服务环境占位符不再掩盖状态目录机密，WhatsApp 启动等待有界，禁用的 WhatsApp 账户在配置重载时拆除。(#90072, #90208, #90277, #90488, #90486, #87951, #87965) 感谢 @MonkeyLeeT、@sallyom、@mcaxtr 和 @MukundaKatta。 ### 变更 - 搜索/提供商：新增 Parallel 捆绑的网络搜索插件、实时提供商测试、注册合约、入门/文档连接，以及受保护的 `api.parallel.ai/v1/search` 支持。(#85158) 感谢 @NormallyGaussian。 - Matrix/频道：新增语音消息预检和线程感知的读取/回复行为，包括 Matrix QA 场景连接和语音消息行为文档。(#78016, #90415) - 技能/ClawHub：通过解析后的安装 API 安装由 GitHub 仓库支持的 ClawHub 技能，下载固定的 GitHub 提交，保留安装策略检查，并在成功后报告安装遥测。(#90478) 感谢 @Patrick-Erichsen。 - Google Chat/频道：新增原生审批卡片操作和点击处理，使 Google Chat 审批使用平台原生卡片而非通用消息流程。 - 移动端：Android 提供商/模型屏幕现在更清晰地显示过期、不可用、未解析和注意状态，而 iOS 设置和 Talk 标签页保持诊断、网关行、附件标签和不可用的 Talk 控件可访问。 - 内存：QMD 搜索可以使用新的重排开关，内存适配器状态在检查普通状态时使用已解析的默认模型标识。(#61834) - 文档/工具：新增 Parallel 搜索文档，刷新天气技能指南转向 `web_fetch`，澄清旧版 `openai-codex` 认证，记录发布/测试辅助脚本，并收紧变更测试路由文档用于 CI/调试工作。(#90028, #90250) 感谢 @fuller-stack-dev。 - 平台维护：为此版本更新了 Android、Swift/macOS、Docker、CodeQL、Buildx、Docker 构建/推送和 Codex Action 依赖。(#74980, #81757, #86481, #86483, #90601) ### 修复 - 频道内容边界：QQBot 现在在发送前剥离推理/思考标签，保留最终答案同时向用户隐藏内部模型叙述。(#89913, #90132) 感谢 @openperf。 - 代理/MCP/提供商：在非文本/图像 MCP 工具结果到达提供商转换器之前进行强制转换，保留有效图像，将更丰富的 MCP 内容转换为文本而非格式错误的图像块。(#90710, #90728) 感谢 @RanSHammer 和 @849261680。 - Anthropic/Codex/ACP/代理恢复：延迟 Anthropic 流启动事件直到 `message_start`，在 Anthropic 重放前剥离过时的压缩思考签名，检测未签名的纯思考停滞，在压缩写入后刷新提示围栏，拒绝空完成移交，保留父流覆盖/共享进度评论，将心跳元数据转发到上下文引擎钩子，并覆盖 Codex 会话/线程迁移边缘情况。(#90667, #90697, #90163, #90108, #89874, #89505, #90632, #89302, #90729, #90317, #90319) 感谢 @openperf、@100yenadmin 和 @ooiuuii。 - 提供商/模型解析：在生成的目录中保留 Google Vertex ADC 认证标记，在冷却后重新探测单一提供商主节点，共享 Codex 模型可见性，对未知模型认证进行失败关闭，保留 Codex 别名可用性，保持未解析的配置文件引用为未知，并在列出模型时避免解析认证。(#90506, #90609, #90717, #90702) 感谢 @849261680。 - 网关/macOS/移动端：通过身份避免重复的网关探测警告，限制节点配对请求速率同时保留已配对节点重新连接，保持 macOS 节点模式在健康的直接网关会话中，保持 iOS 诊断和网关行可访问，避免 Android 构建期间 Linux ARM Gradle 资源任务。(#85791, #90147, #90668, #90815) 感谢 @giodl73-repo 和 @vrurg。 - TUI/聊天/工作板/自动回复：乐观用户消息在过时历史重载、runId 重新分配和中止窗口期间保持稳定，不会消失、跳跃或作为幽灵行滞留；工作板过时生命周期批量更新不再覆盖较新的状态/来源；消息工具发送现在计入投递。(#86205, #89600, #88592, #90123) 感谢 @RomneyDa。 - Cron/更新/服务环境：doctor 配置预检现在在运行时读取之前将遗留 cron JSON 存储迁移到 SQLite，服务环境规划跳过会掩盖状态目录 `.env` 值的未解析占位符，会话转录重写保持注册表标记/判别器一致。(#90072, #90208, #90277, #90488) 感谢 @MonkeyLeeT 和 @sallyom。 - 安全/配置/工具：保护 MCP HTTP 重定向，保护全局代理配置默认值，并使发布/测试/工具证明失败有界且明确。(#89732, #90145) - 频道：WhatsApp 在按账户配置变更时重启，限制后台启动等待，关闭失败套接字，并保留重新连接行为；Mattermost 斜杠命令在 `globalThis` 上保留状态；飞书流式卡片保留完整合并内容；语音通话在连接后跟踪 Twilio 流；ClickClack 回复工具尊重 `toolsAllow`。(#87951, #87965, #90486, #68113, #90534, #90181, #90607, #89500) 感谢 @MukundaKatta、@mcaxtr、@infoanton、@mushuiyu886 和 @sahibzada-allahyar。 - 发布/CI/E2E：主 CI 防止漂移，PR 合并差异范围限定，实时 Docker 凭证暂存，基础镜像鉴定，安装程序 Docker 分类，Playwright 依赖安装恢复，Codex 实时 Docker 通道的 API 密钥认证，Parallels 选项终止符，JSON 模式进度处理更加严格，使发布证明更干净地失败。(#90532, #90287, #90058) 感谢 @RomneyDa、@hxy91819 和 @mrunalp。 - 测试/状态隔离：提供商、媒体、认证、cron、任务、会话、沙箱、网关和 Codex 超时装置现在每个测试范围内包含更多 home/state/env 数据，减少跨测试泄漏，使发布验证失败更少噪声。(#90027, #89974)

## 2026.6.5 ### Highlights - QQBot now strips model reasoning/thinking scaffolding before native delivery, preventing raw `<thinking>` content from leaking into channel replies. (#89913, #90132) Thanks @openperf. - MCP tool results now coerce `resource_link`, `resource`, `audio`, malformed image, and future non-text/image blocks at the materialize boundary, preventing Anthropic 400s and poisoned session history after a tool returns richer MCP content. (#90710, #90728) Thanks @RanSHammer and @849261680. - Anthropic extended-thinking sessions recover after prompt-cache expiry or Gateway restart because stream start events wait for `message_start`, letting pre-generation signature errors trigger the existing recovery retry. (#90667, #90697) Thanks @openperf. - Parallel is now a bundled `web_search` provider with `PARALLEL_API_KEY` discovery, guarded endpoint handling, cache-safe session ids, onboarding picker support, and docs. (#85158) Thanks @NormallyGaussian. - Google Vertex ADC users get static catalog rows and runtime model resolution again, while single-provider cooldown recovery and memory adapter status checks are more reliable. (#90506, #90609, #90717, #90816) Thanks @849261680. - Matrix can preflight voice notes before mention gating, preserve thread reads/replies through Matrix relations pagination, and carry QA coverage for voice and thread flows. (#78016, #90415) - Auth and plugin install state is more durable: auth profiles now live in SQLite, official npm plugin install records keep their trusted pins, and prerelease fallback integrity checks avoid carrying stale integrity forward. (#89102, #88585) - macOS node mode no longer silently self-reconnects away from a healthy direct Gateway session, reducing unexpected companion app session churn. (#90668, #90815) Thanks @vrurg. - Upgrade and service paths are safer: cron legacy JSON stores migrate during doctor preflight, service env placeholders no longer mask state-dir secrets, WhatsApp startup waits are bounded, and disabled WhatsApp accounts tear down on config reload. (#90072, #90208, #90277, #90488, #90486, #87951, #87965) Thanks @MonkeyLeeT, @sallyom, @mcaxtr, and @MukundaKatta. ### Changes - Search/providers: add the Parallel bundled web-search plugin, live provider tests, registration contracts, onboarding/docs wiring, and guarded `api.parallel.ai/v1/search` support. (#85158) Thanks @NormallyGaussian. - Matrix/channels: add voice-message preflight and thread-aware read/reply behavior, including Matrix QA scenario wiring and docs for voice-message behavior. (#78016, #90415) - Skills/ClawHub: install ClawHub skills backed by GitHub repositories through the resolved install API, download the pinned GitHub commit, keep install-policy checks, and report install telemetry after success. (#90478) Thanks @Patrick-Erichsen. - Google Chat/channels: add native approval card actions and click handling so Google Chat approvals use platform-native cards instead of generic message flow. - Mobile: Android provider/model screens now surface expiring, unavailable, unresolved, and attention states more clearly, while iOS settings and Talk tabs keep diagnostics, gateway rows, attachment labels, and unavailable Talk controls reachable. - Memory: QMD search can use the new rerank toggle, and memory adapter status uses the resolved default model identity when checking plain status. (#61834) - Docs/tooling: add Parallel search docs, refresh weather-skill guidance toward `web_fetch`, clarify legacy `openai-codex` auth, document release/test helper scripts, and tighten changed-test routing docs for CI/debugging work. (#90028, #90250) Thanks @fuller-stack-dev. - Platform maintenance: refresh Android, Swift/macOS, Docker, CodeQL, Buildx, Docker build/push, and Codex Action dependencies for this release train. (#74980, #81757, #86481, #86483, #90601) ### Fixes - Channel content boundaries: QQBot now strips reasoning/thinking tags before sending, preserving final answers while hiding internal model narration from users. (#89913, #90132) Thanks @openperf. - Agents/MCP/providers: coerce non-text/image MCP tool-result blocks before they reach provider converters, preserving valid images and turning richer MCP content into text instead of malformed image blocks. (#90710, #90728) Thanks @RanSHammer and @849261680. - Anthropic/Codex/ACP/agent recovery: defer Anthropic stream start events until `message_start`, strip stale compaction thinking signatures before Anthropic replay, detect unsigned thinking-only stalls, refresh prompt fences after compaction writes, reject empty completion handoffs, preserve parent streaming-off overrides/shared progress commentary, forward heartbeat metadata to context-engine hooks, and cover Codex session/thread migration edge cases. (#90667, #90697, #90163, #90108, #89874, #89505, #90632, #89302, #90729, #90317, #90319) Thanks @openperf, @100yenadmin, and @ooiuuii. - Provider/model resolution: preserve Google Vertex ADC auth markers in generated catalogs, re-probe a single-provider primary after cooldown, share Codex model visibility, fail closed for unknown model auth, preserve Codex alias availability, keep unresolved profile refs unknown, and avoid resolving auth while listing models. (#90506, #90609, #90717, #90702) Thanks @849261680. - Gateway/macOS/mobile: avoid duplicate Gateway probe warnings by identity, rate-limit node pairing requests while preserving paired-node reconnects, keep macOS node mode on a healthy direct Gateway session, keep iOS diagnostics and gateway rows reachable, and avoid Linux ARM Gradle resource tasks during Android builds. (#85791, #90147, #90668, #90815) Thanks @giodl73-repo and @vrurg. - TUI/chat/Workboard/auto-reply: optimistic user messages stay stable across stale history reloads, runId reassignment, and abort windows instead of disappearing, jumping, or lingering as ghost rows; Workboard stale lifecycle bulk updates no longer overwrite newer status/provenance; message-tool sends now count as delivery. (#86205, #89600, #88592, #90123) Thanks @RomneyDa. - Cron/update/service env: doctor config preflight now migrates legacy cron JSON stores into SQLite before runtime reads, service env planning skips unresolved placeholders that would mask state-dir `.env` values, and session transcript rewrites keep registry markers/discriminants consistent. (#90072, #90208, #90277, #90488) Thanks @MonkeyLeeT and @sallyom. - Security/config/tooling: guard MCP HTTP redirects, protect global agent config defaults, and keep release/test/tooling proof failures bounded and explicit. (#89732, #90145) - Channels: WhatsApp restarts when per-account config changes, bounds background startup waits, closes failed sockets, and preserves reconnect behavior; Mattermost slash commands keep their state on `globalThis`; Feishu streaming cards preserve full merged content; voice-call tracks Twilio streams after connect; ClickClack reply tools respect `toolsAllow`. (#87951, #87965, #90486, #68113, #90534, #90181, #90607, #89500) Thanks @MukundaKatta, @mcaxtr, @infoanton, @mushuiyu886, and @sahibzada-allahyar. - Release/CI/E2E: main CI guard drift, PR merge diff scoping, live Docker credential staging, base-image qualification, installer Docker classification, Playwright dependency install recovery, API-key auth for Codex live Docker lanes, Parallels option terminators, and JSON-mode progress handling are tighter so release proof fails cleaner. (#90532, #90287, #90058) Thanks @RomneyDa, @hxy91819, and @mrunalp. - Tests/state isolation: provider, media, auth, cron, task, session, sandbox, Gateway, and Codex timeout fixtures now scope more home/state/env data per test, reducing cross-test leakage and making release validation failures less noisy. (#90027, #89974)

OpenClaw v2026.6.5-beta.1 预发布