openclaw v2026.6.6-beta.2 發佈說明:新功能與變更日誌
openclaw v2026.6.6-beta.2 發佈說明:新功能與變更日誌
openclaw v2026.6.6-beta.2 提供了實質性的安全邊界強化、擴展的模型供應商支援,以及針對 Control UI 和記憶體索引的顯著效能優化。
關鍵變更
安全強化
安全邊界現在在多個關鍵層面上都變得更加嚴密:
- 主機環境: 主機 exec sanitizer 現在會阻擋請求範圍內的 Rustup 環境覆寫,並擴展了不安全主機環境的黑名單,以拒絕額外的解釋器啟動和搜尋路徑變數。Git 協定控制環境變數也被阻擋,以防止放寬 Git 傳輸政策。
- 沙盒與 Codex: 沙盒 bind 來源現在會進行雙向驗證,以拒絕覆蓋已封鎖 Descendant 的父級來源。Codex 沙盒 exec-server HTTP 橋接器現在會拒絕私有或內部 HTTP 目標,並固定已驗證的 DNS 結果,以防止驗證/連線漂移。
- 存取控制: Discord 審核動作現在需要受信任的請求者發送者。此外,已刪除代理(deleted-agent)守衛現在需要 ACP metadata 用於繞過,以確保單憑 ACP 形式的 session keys 並不被視為 ACP 執行時期的 session 證明。
- 資料保護: 逐字稿圖像遮蔽現在會保留已驗證的模糊圖像負載,以防止機密模式損壞 base64 位元組。瀏覽器工具輸出現在會中和行首的
MEDIA:指令,以防止它們出現在面向模型的輸出中。
供應商與模型支援
- Anthropic: 增加了對 Claude Fable 5 的支援,涵蓋直接 API、Vertex、Bedrock 以及與 Anthropic 相容的部署別名,包括適應性思考預設值和拒絕處理。
- OpenRouter: 將 OpenRouter PKCE OAuth 登入整合至頂層 CLI onboarding 流程中。
- OpenAI: 即時語音現在需要 OpenAI Platform API-key 憑證,而非 OAuth bearer tokens。此外,對於 Gemma 4 openai-completions 模型,現在會保留推理內容重播。
- Ollama: SDK session factory 現在會遵循供應商的思考預設值(例如,Ollama 的 "off"),以防止模型在非預期情況下將其輸出預算花費在推理上。
效能與 UI 優化
- Control UI: 透過快取模型 metadata,移除啟動目錄等待,以及延遲載入 slash-command,降低了啟動和首次回應的延遲。新增的
first-assistant-event伺服器計時階段允許進行更好的延遲診斷。 - 記憶體索引: 對於選擇加入全源批次提交的供應商(例如, OpenAI),記憶體 embedding 運作現在會跨檔案進行批次處理,顯著減少了供應商批次作業的數量。
- TUI: TUI 現在會在首次發送前預熱執行時插件,執行時插件的冷啟動尖峰會因此消除,且頁尾現在會顯示用於 URL 支援的 Gateway 連線的連線主機名稱。
通訊管道改善
- Telegram: 在工具呼叫之間保留串流文本塊,使傳遞更具連貫性。帳號範圍內的 topic agents 現在會被正確路由,且在通用訊息傳入時會恢復了
/compact指令。 - iMessage: 入站接收與去重功能現在為常駐開啟,取代了原先的選用式 catchup 子系統。出站發送傳輸現在透過新的
sendTransport配置和獨立的發送客戶端進行強化,以防止卡住的 watch 訂閱會阻擋發送。 - Discord: 當存在
message_reference時,入站訊息現在會透過 REST 補全缺失的引用訊息。回覆 context metadata 現在透過插件 hook contracts 揭露。 - WhatsApp: 在重啟後,擷取的回答會透過繼任者控制器進行路由,以防止暫時性斷線後的靜默回應失敗。
影響
使用者體驗
使用者會注意到 Control UI 載入時間更短,且 TUI 互動更具響應性。Telegram 和 iMessage 使用者會體驗到更可靠的訊息傳遞,以及對串流內容更好的處理。
安全姿勢
此發佈版本透過在 exec 批准通過時間限制到期時採取「失敗關閉」策略,並嚴格過濾主機環境變數,顯著降低了 unattended execution(無人值守執行)和環境逃逸的風險。
營運變更
使用本地記憶體 embedding 的營運人員可能需要安裝新的 @openclaw/llama-cpp-provider 插件,因為 node-llama-cpp 依賴項已從核心包中移至此供應商插件。
升級指南
破壞性變更與遷移
- 本地記憶體 Embedding:
localembedding 供應商不再是內建的。使用者必須安裝@openclaw/llama-cpp-provider插件。執行openclaw doctor --fix可自動將memorySearch.provider: "local"映射至官方插件安裝。 - Exec 批准: 未回答的 exec 批准請求現在在配置的超時時間到期後,預設為拒絕。需要進行超時自動批准的營運人員必須在 exec 批准政策中明確設定
askFallback。 - OpenAI Realtime Voice: 僅使用 OAuth 的設定不再適用於 Realtime 語音。使用者必須配置 OpenAI Platform API key 或
openaiAPI-key auth profile。 - Android Nodes: 持續性節點的前台服務類型已從
dataSync變改為connectedDevice,以符合 Android 15 的時間預算。
FAQ
v2026.6.6-beta.2 有什麼新功能? 此發佈版本為主機和沙盒環境引入了更嚴密的安全邊界,增加了對 Claude Fable 5 和 OpenRouter OAuth 的支援,優化了 Control UI 和 TUI 效能,並提升了 Telegram, iMessage, 和 WhatsApp 的傳遞可靠性。
是否有任何破壞性變更? 是的。本地記憶體 embedding 供應商現在需要獨立的插件插件,exec 批准超時後會預設為拒絕,且 OpenAI Realtime 語音現在需要 API-key 驗證。
如何升級?
使用者應更新套件,然後執行 openclaw doctor --fix 進行處理本地記憶體 embedding 供應商的遷移與其他狀態修復。
Sources
- Releasev2026.6.6-beta.2 — openclaw 2026.6.6-beta.2
- PRfix(browser): honor cdpUrl for user default profile
- PRfix(imessage): skip idle approval discovery scans
- PRfix(codex): avoid guardian review for local models
- PRPreserve Telegram streamed text blocks between tool calls
- PRfix(codex): normalize dynamic tool progress results
- PRfix(imessage): persist echo markers before send
- PRfix(telegram): restore /compact on generic message ingress
- PRfeat(cli): emit commentary progress events from Claude CLI parser
- PRfix(gateway): support Streamable HTTP MCP transport on loopback server
- PRFail closed on exec approval timeout
- PRfix(agents): deliver native /compact replies through source suppression
- PRfix(cli): bridge inter-tool commentary events to channel progress
- PRfix(telegram): route account-scoped topic agents
- PRfeat(diagnostics-otel): capture tool input/output content via trusted channel
- PRfix(imessage): always-on inbound recovery and dedupe
- PRfix(browser): neutralize media directives in browser output [AI]
- PRfix(imessage): honor block streaming config
- PRfix(mcp): repair OAuth redirect, errors, and unicode schema patterns
- PRblock unauthorized Telegram DM text from prompt context
- PRFix transcript image redaction
- PRperf(control-ui): reuse startup model metadata
- PRperf(control-ui): avoid startup catalog wait
- PRperf(control-ui): trace first assistant event
- PRfeat: dogfood reusable ClawHub package publish
- PRperf(control-ui): warn on slow first replies
- PRFix context-engine compaction ownership for Codex sessions
- PRperf(control-ui): lazy load slash commands
- PRfix: block rustup toolchain env overrides [AI]
- PRfix: expand unsafe host env denylist
- PRfix: block git protocol env controls [AI]
- PRfix(agents): preserve reasoning_content replay for Gemma 4 openai-completions models
- PRSupport existing-session browser CDP endpoints
- PRValidate sandbox bind parent paths [AI]
- PRfix(discord): require sender for moderation actions [AI]
- PRfix(msteams): require admin for group actions
- PRfix(browser): validate discovered CDP websocket URLs
- PRfix(elevated): reject group ids as senders
- PRfix(gateway): restrict non-owner loopback tools
- PRfix(search): enforce native web search tool policy
- PRfix(mcp): harden stdio env filtering
- PRfix(codex): guard sandbox http requests
- PRfix: require ACP metadata for deleted-agent bypass
- PRfix(imessage): harden outbound send transport
- PRfix(imessage): surface inbound startup diagnostics
- PRfeat: add OpenRouter OAuth to onboarding
- PRShare channel draft chunking resolver
- PRFix Telegram callback API handling
- PRfeat(anthropic): support Claude Fable 5 adaptive thinking
- PRfix(telegram): use SDK dispatch dedupe
- PRfix(telegram): audit follow-ups — block-mode chunk config, dedupe bucket cleanup, grammy contract trust
- PRfix(CLI): validate gateway-facing timeout input
- PRfix(cli): validate gateway-rpc --timeout
- PRfix(agents): drop stale exec approval followups after session rebind
- PRfix(whatsapp): route captured replies through successor controller after restart
- PRfix(gateway): share approval runtime socket token
- PRfix #88009: [Feature]: batched memory embedding should batch over files
- PRfix(delivery): suppress Codex/Harmony internal protocol artifacts from user-facing channels
- PRfix(feishu): retry on send rate-limit errors (230020/230006)
- PRfix(tui): show connection host in footer
- PRperf(config): dedupe plugin auto-enable fanout work
- PR[plugin sdk] Allow declared installed trusted hooks
- PRfix(discord): hydrate reply context metadata
- PRfix(talk): show OpenAI Realtime WebRTC assistant transcripts
- PRfix(models): persist agent catalog cache
- PRfix(cron): cancel active cron task runs
- PRfix(cron): recover no-deliver tool warnings
- PRperf(tui): prewarm runtime plugins before first send
- PRfix(config): warn for retired skill-workshop plugin entry instead of failing validation (#90244)
- PRfix(doctor): report managed plugin version drift
- PRfix(gateway): skip deleted-agent guard for ACP harness session keys
- PRfix(memory): move local llama.cpp runtime to provider plugin
- PRfix(compaction): lower default timeout from 900s to 180s, preserve explicit config
- PRfix: clarify provider quota errors
- PRfeat(qqbot): add /bot-group-allways command to toggle mention requirement
- PRfix(memory-lancedb): guard memory recall output [AI]
- PRfix(reply-queue): remove the drained item by reference instead of front index
- PRfeat: canonicalize Codex protocol JSON asset ordering
- PRFix config.patch explicit array replacement
- PRImprove iPad and iPhone control surfaces
- PRFix stale main session startup recovery
- PRfix(openai): require api-key auth for realtime voice
- PRfix(agents): trim dense text delta snapshots
- PRfix(update): recover package gateway restart after refresh failure
- PRfix: let ClawHub dry runs skip publish approval
- PRfix(update): expose plugin convergence repair
- PRfix(gateway): surface headless LaunchAgent state
- PRfix(line): canonicalize trailing-slash webhook paths
- PRfix(ollama): use provider thinking default in SDK session factory
- PRfix(mattermost): keep default replies in existing threads
- PRfix(discord): restore runtime timeout compatibility exports
- PRfix(status): restore Codex synthetic usage line
- PRfix(auth): verify SQLite auth migration before cleanup
- PRfix(memory-core): keep QMD JSON search one-shot
- PRFix stale visible reply recovery
- PRfix(memory-core): filter stale recall entries in REM harness preview
- PRfix(webchat): finalize provider failure lifecycle
- PRfix(agents): retry same model across short rate-limit windows
- PRfix(config): stop config.patch replacePaths index suffix from widening array consent
- PRfix(gateway): arm qmd startup maintenance
- PRfix(plugins): rescan storm in "/models" call (regression shipped since v2026.5.18)
- PRfix(android): avoid dataSync FGS for persistent node
- PRfix(ui): require user intent for chat sessions
- PRFix Docker store seed target packages
- PRfix: bound native hook relay lifetime