OpenClaw v2026.5.4-beta.3 预发布

## 2026.5.4 ### Highlights - Google Meet/Voice Call: make Twilio dial-in joins speak through the realtime Gemini voice bridge with paced audio streaming, backpressure-aware buffering, barge-in queue clearing, and no TwiML fallback during realtime speech, giving Meet participants a much snappier OpenClaw voice agent. (#77064) Thanks @scoootscooob. ### Changes - Gateway/Windows: bind the default loopback gateway listener only to `127.0.0.1` on Windows so libuv's dual-stack `::1` behavior cannot wedge localhost HTTP requests. (#69701, fixes #69674) Thanks @SARAMALI15792. - Plugins/migration: emit catalog-backed install hints when `plugins.entries` or `plugins.allow` references an official external plugin that is not installed, so upgraded configs point operators to `openclaw plugins install <spec>` instead of telling them to remove valid plugin config. (#77483) Thanks @hclsys. - OpenAI/Codex media: advertise Codex audio transcription in runtime and manifest metadata and route active Codex chat models to the OpenAI transcription default instead of sending chat model ids to audio transcription. Thanks @vincentkoc. - Dependencies: refresh runtime and provider packages including Pi 0.73.0, ACPX adapters, OpenAI, Anthropic, Slack, and TypeScript native preview, while keeping the Bedrock runtime installer override pinned below the Windows ARM Node 24 npm resolver failure. - Agents/performance: pass the resolved workspace through BTW, compaction, embedded-run model generation, and PDF model setup so explicit agent-dir model refreshes can reuse the current workspace-scoped plugin metadata snapshot instead of falling back to cold plugin metadata scans. (#77519, #77532) - Plugins/performance: let unscoped model catalog and manifest-contract readers reuse the current workspace-compatible plugin metadata snapshot, avoiding repeated cold plugin metadata scans on hot control-plane paths while preserving env/config/workspace compatibility checks. (#77519, #77532) - Config/plugin auto-enable: prefer the claiming plugin manifest id over a built-in channel alias when auto-allowlisting a configured channel, so WeCom/Yuanbao-style aliases resolve to the installed plugin id. Thanks @Beandon13. - Secrets/apply: preserve auth-profile `keyRef` and `tokenRef` fields when scrubbing provider-target secrets, so the canonical SecretRef metadata survives `secrets apply` without keeping plaintext values. Thanks @Beandon13. - Plugins/active-memory: skip session-store channel entries that contain `:` when resolving the recall subagent's channel, so QQ c2c agent IDs (e.g. `c2c:10D4F7C2…`) and other scoped conversation IDs do not reach bundled-plugin `dirName` validation and crash the recall run. The same guard already applied to explicit `channelId` params (#76704); this extends it to store-derived channels. (#77396) Thanks @hclsys. - Secrets/external channel contracts: also look in `<rootDir>/dist/` when resolving the `secret-contract-api` sidecar, so npm-published externalized channel plugins (e.g. `@openclaw/discord` since 2026.5.2) whose compiled artifacts live under `dist/` actually contribute their channel SecretRef contracts to the runtime snapshot. Without this, env-backed `channels.discord.token` SecretRefs silently failed to resolve at gateway start on 2026.5.3, leaving the channel `not configured` even though #76449 had landed the generic external-contract loader. Thanks @mogglemoss. - Models/auth: add `openclaw models auth list [--provider <id>] [--json]` so users can inspect saved per-agent auth profiles without dumping secrets or hitting the old “too many arguments” path. Thanks @vincentkoc. - Control UI/header: show the active agent name in dashboard breadcrumbs without adding the current session key, keeping non-chat views oriented without crowding the topbar. - Control UI/cron: make the New Job sidebar collapsible so the jobs list can reclaim space while keeping the form one click away. Thanks @BunsDev. - Gateway/startup: keep model-catalog test helpers, run-session lookup code, QR pairing helpers, and TypeBox memory-tool schema construction out of hot startup import paths, reducing default gateway benchmark plugin-load and memory pressure. - Control UI/performance: record browser long animation frame or long task entries in the debug event log when supported, making slow dashboard renders easier to attribute from the UI. - Slack/streaming: add `streaming.progress.render: "rich"` for Block Kit progress drafts backed by structured progress line data. - Slack/streaming: keep the newest rich progress lines when Block Kit limits trim long progress drafts. Thanks @vincentkoc. - Channels/streaming: cap progress-draft tool lines by default so edited progress boxes avoid jumpy reflow from long wrapped lines. - Agents/verbose: use compact explain-mode tool summaries for `/verbose` and progress drafts by default, with `agents.defaults.toolProgressDetail: "raw"` and per-agent overrides for debugging raw command/detail output. - Control UI/chat: add an agent-first filter to the chat session picker, keep chat controls/composer responsive across phone/tablet/desktop widths, keep desktop chat controls on one row, avoid duplicate avatar refreshes during initial chat load, and hide that row while scrolling down the transcript. Thanks @BunsDev. - Control UI/chat: collapse consecutive duplicate text messages into one bubble with a count so no-op heartbeat acknowledgements stay compact without hiding nearby context. - Agents/subagents: preserve every grouped child result when direct completion fallback has to bypass the requester-agent announce turn. Thanks @vincentkoc. - TTS/telephony: honor provider voice/model overrides in telephony synthesis providers so Google Meet agent speech logs match the backend that actually produced the audio. Thanks @vincentkoc. - Voice Call/realtime: bound the paced Twilio audio queue and close overloaded realtime streams before provider audio can pile up behind the websocket backpressure guard. Thanks @vincentkoc. - Docs: clarify that IRC uses raw TCP/TLS sockets outside operator-managed forward proxy routing, so direct IRC egress should be explicitly approved before enabling IRC. Thanks @jesse-merhi. - Gateway/performance: defer non-readiness sidecars until after the ready signal, avoid hot-path channel plugin barrel imports, and fast-path trusted bundled plugin metadata during Gateway startup. - Gateway/performance: avoid importing `jiti` on native-loadable plugin startup paths, so compiled bundled plugin surfaces do not pay source-transform loader cost unless fallback loading is actually needed. - Gateway/diagnostics: add startup phase spans, active work labels, stale terminal bridge markers, and default sync-I/O tracing in `pnpm gateway:watch` so slow Gateway turns are easier to attribute from logs and stability diagnostics. - Plugins/loader: preserve real compiled plugin module evaluation errors on the native fast path instead of treating every thrown `.js` module as a source-transform fallback miss. Thanks @vincentkoc. - QA/Mantis: add `pnpm openclaw qa mantis slack-desktop-smoke` to run Slack live QA inside a Crabbox VNC desktop, open Slack Web, and capture desktop screenshots beside the Slack QA artifacts. - QA/Mantis: pass the runtime env through desktop-browser Crabbox and artifact-copy child commands, so embedded Mantis callers can provide Crabbox credentials without mutating the parent process. Thanks @vincentkoc. - QA/Mantis: return the copied Slack desktop screenshot path even when remote Slack QA fails, so the CLI still prints the failure screenshot artifact. Thanks @vincentkoc. - QA/Mantis: accept Blacksmith Testbox `tbx_...` lease ids from desktop smoke warmup, so provider overrides do not fail before inspect/run. Thanks @vincentkoc. - QA/Codex harness: add targeted live Docker/Testbox diagnostics, auth preflight checks, cache mount fixes, and app-server protocol checkout discovery so maintainer harness failures are easier to reproduce. Thanks @vincentkoc. - Plugins/update: treat official externalized bundled npm migrations and ClawHub-to-npm fallbacks as trusted source-linked installs, so prerelease-only official plugin packages can migrate from bundled builds without being rejected as unsafe prerelease resolutions. Thanks @vincentkoc. - Plugins/update: move ClawHub-preferred externalized plugin installs back to ClawHub after an earlier npm fallback once the ClawHub package becomes available. Thanks @vincentkoc. - Plugins/update: clean stale bundled load paths for already-externalized pinned npm and ClawHub plugin installs, so release-channel sync does not leave removed bundled paths ahead of the installed external package. Thanks @vincentkoc. - Telegram: accept plugin-owned numeric forum-topic targets in the agent message tool and keep reply-dispatch provider chunks behind a real stable runtime alias during in-place package updates. Fixes #77137. Thanks @richardmqq. - Google Meet: preserve `realtime.introMessage: ""` so realtime Chrome joins can stay silent instead of restoring the default spoken intro. Thanks @vincentkoc. - Plugins/SDK: add bounded `before_agent_finalize` retry instructions so workflow plugins can request one more model pass. Thanks @100yenadmin. - Discord/status: add degraded Discord transport and gateway event-loop starvation signals to `openclaw channels status`, `openclaw status --deep`, and fetch-timeout logs so intermittent socket resets do not look like a healthy running channel. (#76327) Thanks @joshavant. - Providers/OpenRouter: add opt-in response caching params that send OpenRouter's `X-OpenRouter-Cache`, `X-OpenRouter-Cache-TTL`, and cache-clear headers only on verified OpenRouter routes. Thanks @vincentkoc. - Providers/OpenRouter: expand app-attribution categories so OpenClaw advertises coding, programming, writing, chat, and personal-agent usage on verified OpenRouter routes. Thanks @vincentkoc. - Plugins/update: make package upgrades swap pnpm/npm-prefix installs cleanly, keep legacy plugin install runtime chunks working, and on the beta channel fall back default-line npm plugins to default/latest when plugin beta releases are missing or fail install validation. Thanks @vincentkoc and @joshavant. - Channels/WhatsApp: support explicit WhatsApp Channel/Newsletter `@newsletter` outbound message targets with channel session metadata instead of DM routing. Fixes #13417; carries forward the narrow outbound target idea from #13424. Thanks @vincentkoc and @agentz-manfred. - Exec approvals: add a tree-sitter-backed shell command explainer for future approval and command-review surfaces. (#75004) Thanks @jesse-merhi. - Agents/sandbox: store sandbox container and browser registry entries as per-runtime shard files, reducing unrelated session lock contention while `openclaw doctor --fix` migrates legacy monolithic registry files. (#74831) Thanks @luckylhb90. - Plugins/ClawHub: annotate 429 errors from ClawHub with the reset window from `RateLimit-Reset`/`Retry-After` and append a `Sign in for higher rate limits.` hint when the request was unauthenticated, so users can see when downloads will recover and how to lift the cap. Thanks @romneyda. - Plugins/runtime state: add `registerIfAbsent` for atomic keyed-store dedupe claims that return whether a plugin successfully claimed a key without overwriting an existing live value. Thanks @amknight. - Plugin SDK: add plugin-owned `SessionEntry` slot projection and scoped trusted-policy session extension reads. (#75609; replaces part of #73384/#74483) Thanks @100yenadmin. - Sandbox/Windows: accept drive-absolute Docker bind sources while keeping sandbox blocked-path and allowed-root policy comparisons Windows-case-insensitive. (#42174) Thanks @6607changchun. ### Fixes - Plugins/install: honor the beta update channel for onboarding and doctor-managed plugin installs by requesting floating npm and ClawHub specs with `@beta` while keeping persistent install records on the catalog default. Thanks @vincentkoc. - WhatsApp/onboarding: canonicalize setup and pairing allowlist entries to WhatsApp's digit-only phone ids while still accepting E.164, JID, and `whatsapp:` inputs, so personal-phone allowlists match WhatsApp Web sender ids after setup. Thanks @vincentkoc. - Gateway/startup: load provider plugins that own explicitly configured image, video, or music generation defaults so generation tools become live after gateway restart instead of remaining catalog-only. Fixes #77244. Thanks @buyuangtampan, @Nikoxx99, and @vincentkoc. - Slack/subagents: keep resumed parent `message.send` calls in the originating Slack thread when ambient session thread context is present, and suppress successful silent child completion rows from follow-up findings. Thanks @bek91. - Slack/mentions: record thread participation for successful visible threaded Slack sends, including message-tool and media delivery paths, so unmentioned replies in bot-participated threads can bypass mention gating as documented. Fixes #77648. Thanks @bek91. - Infra/Windows: skip the POSIX `/tmp/openclaw` preferred path on Windows in `resolvePreferredOpenClawTmpDir` so log files, TTS temp files, and other writes land in `%TEMP%\openclaw-<uid>` instead of `C:\tmp\openclaw`. Fixes #60713. Thanks @juan-flores077. - Media/Windows: open saved attachment temp files read/write before fsync so Windows WebChat and `chat.send` media offloads no longer fail with EPERM during durability flush. (#76593) Thanks @qq230849622-a11y. - Agents/tools: honor narrow runtime tool allowlists when constructing embedded-runner tool families and bundled MCP/LSP runtimes, so cron/subagent runs that request tools such as `update_plan`, `browser`, `x_search`, channel login tools, or `group:plugins` no longer start with missing tools or unrelated bootstrap work. (#77519, #77532) - Codex plugin: mirror the experimental upstream app-server protocol and format generated TypeScript before drift checks, keeping OpenClaw's `experimentalApi` bridge compatible with latest Codex while preserving formatter gates. - Telegram/media: derive no-caption inbound media placeholders from saved MIME metadata instead of the Telegram `photo` shape, so non-image and mixed attachments no longer reach the model as `<media:image>`. Fixes #69793. Thanks @aspalagin. - Telegram/streaming: reuse the active preview as the first chunk for long text finals, so multi-chunk replies no longer create a transient extra bubble that appears and then disappears. Thanks @vincentkoc. - Agents/cache: keep per-turn runtime context out of ordinary chat system prompts while still delivering hidden current-turn context, restoring prompt-cache reuse on chat continuations. Fixes #77431. Thanks @Udjin79. - Gateway/startup: include resolved thinking and fast-mode defaults in the `agent model` startup log line, defaulting unset startup thinking to `medium` without mixing in reasoning visibility. - Gateway/update: resolve local gateway probe auth from the installed config during post-update restart verification, so token/device-authenticated VPS gateways are not misreported as unhealthy port conflicts after a package swap. Thanks @vincentkoc. - Agents/Tools: add post-compaction loop guard in `pi-embedded-runner` that arms after auto-compaction-retry and aborts the run with `compaction_loop_persisted` when the agent emits the same `(tool, args, result)` triple `windowSize` times (default 3) within that window. Disable via existing `tools.loopDetection.enabled`; tune via `tools.loopDetection.postCompactionGuard.windowSize`. Targets the failure mode where context-overflow + compaction does not break a tool-call loop. Refs #77474; carries forward #21597. Thanks @efpiva. - Gateway/watch: suppress sync-I/O trace output during `pnpm gateway:watch --benchmark` unless explicitly requested, so CPU profiling no longer floods the terminal with stack traces. - Gateway/watch: when benchmark sync-I/O tracing is explicitly enabled, tee trace blocks to the benchmark output log and filter them from the terminal pane while keeping normal Gateway logs visible. - Plugins/runtime-deps: include `json5` in the memory-core plugin runtime dependency set so packaged `memory_search` sandboxes can resolve generated OpenClaw runtime chunks that parse JSON5 config. Fixes #77461. - Plugins/Windows: show a Git install hint when npm plugin installation fails with `spawn git ENOENT`, and document the WhatsApp plugin's Git-on-PATH requirement for Baileys/libsignal installs. - Codex harness: preserve app-server usage-limit reset details and deliver OpenClaw-owned runtime failure notices through tool-only source-reply mode, so Telegram and other chat channels tell users when Codex subscription limits or API failures block a turn instead of going silent. (#77557) Thanks @pashpashpash. - Agents/OpenAI: default direct OpenAI Responses models to the SSE transport instead of WebSocket auto-selection, preventing pi runtime chat turns from hanging on servers where the WebSocket path stalls while the OpenAI HTTP stream works. Thanks @vincentkoc. - Plugins/update: repair missing plugin-local `openclaw` peer links before skipping unchanged npm plugin updates, so current external Codex installs can recover `openclaw/plugin-sdk/*` resolution during OTA repair. (#77544) Thanks @ProspectOre. - Discord/replies: treat failed final reply delivery as a failed turn instead of counting it as a delivered automatic visible reply, so guild/channel turns no longer show done when the final message was dropped. Fixes #77520. Thanks @Patrick-Erichsen. - Discord: prefer IPv4 for Discord REST and gateway WebSocket startup paths so IPv4-only networks no longer stall before Gateway READY and inbound message dispatch. Fixes #77398; refs #77526. Thanks @Beandon13. - Channels/plugins: key bundled package-state probes, env/config presence, and read-only command defaults by channel id instead of manifest plugin id, preserving setup and native-command detection for channel plugins whose package id differs from the channel alias. Thanks @vincentkoc. - Docker: prune package-excluded plugin dist directories from runtime images unless the build explicitly opts that plugin in, so official external plugins such as Feishu stay install-on-demand instead of shipping partial metadata without compiled runtime output. Fixes #77424. Thanks @vincentkoc. - Model switching: include the exact additive allowlist repair command when `/model ... --runtime ...` targets a blocked model, and make Telegram's model picker say that it changes only the session model while leaving the runtime unchanged. Thanks @vincentkoc. - Mattermost: clarify that the model picker only changes the session model and that runtime switches require `/oc_model <provider/model> --runtime <runtime>`. Thanks @vincentkoc. - Doctor/config: keep active `auth.profiles` metadata intact when `doctor --fix` strips stale secret fields from configs, repairing legacy `<provider>:default` API-key profile metadata when model fallbacks or explicit `model@profile` refs still depend on it. Fixes #77400. - Doctor/plugins: include `plugins.allow`-only official plugin ids in the release configured-plugin repair set, so `doctor --fix` installs official external plugins that are configured but not yet loaded instead of removing them as stale allow entries. Fixes #77155. Thanks @hclsys. - Doctor/sessions: clear auto-created stale session routing state from the sessions store when `doctor --fix` sees plugin-owned model/runtime/auth/session bindings outside the current configured route, while leaving explicit user model choices for manual review. Refs #68615. - CLI/update: disable and skip plugins that fail package-update plugin sync, so a broken npm/ClawHub/git/marketplace plugin cannot turn a successful OpenClaw package update into a failed update result. Thanks @vincentkoc. - CLI/update: use an absolute POSIX npm script shell during package-manager updates, so restricted PATH environments can still run dependency lifecycle scripts while updating from `--tag main`. Fixes #77530. Thanks @PeterTremonti. - Diagnostics: grant the internal diagnostics event bus to official installed diagnostics exporter plugins, so npm-installed `@openclaw/diagnostics-prometheus` can emit metrics without broadening the capability to arbitrary global plugins. Fixes #76628. Thanks @RayWoo. - Browser: enforce strict SSRF current-URL checks before existing-session screenshots, matching existing-session snapshot handling. Thanks @vincentkoc. - Active Memory: give timeout partial transcript recovery enough abort-settle headroom so temporary recall summaries are returned before cleanup. Thanks @vincentkoc. - Gateway/chat: clear the active reply-run guard before draining queued same-session follow-up turns, so sequential `chat.send` calls no longer trip `ReplyRunAlreadyActiveError` every other request. Fixes #77485. Thanks @bws14email. - Agents/media: avoid sending generated image, video, and music attachments twice when streamed reply text arrives before the final `MEDIA:` directive. - CLI/sessions: cap `openclaw sessions` output to the newest 100 rows by default and add `--limit <n|all>` plus JSON pagination metadata, so repeated machine polling of large session stores cannot fan out into unbounded per-row enrichment/output work. Fixes #77500. Thanks @Kaotic3. - Doctor/config: restore legacy group chat config migrations for `routing.allowFrom`, `routing.groupChat.*`, and `channels.telegram.requireMention` so upgrades keep WhatsApp, Telegram, and iMessage group mention gates and history settings instead of leaving configs invalid or silently blocked. Thanks @scoootscooob. - CLI/update: make package-update follow-up processes write completion results and exit explicitly, so Windows packaged upgrades do not hang after the new package finishes post-core plugin work. Thanks @vincentkoc. - Release validation: skip Slack live QA unless Slack credentials are explicitly configured, so release gates can keep proving non-Slack surfaces while Slack is still local and credential-gated. Thanks @vincentkoc. - Plugins/update: treat OpenClaw CalVer correction versions like `2026.5.3-1` as satisfying base plugin API ranges, so correction builds can install plugins that require the base runtime API. Fixes #77293. (#77450) Thanks @p3nchan. - Discord/Gateway startup: retry Discord READY waits with backoff, defer startup `sessions.list` and native approval readiness failures until sidecars recover, and preserve component-only Discord payloads when final reply scrubbing removes all text. (#77478) Thanks @NikolaFC. - CLI/launcher: forward termination signals to compile-cache respawn children, so killing a wrapper process no longer leaves the security audit worker orphaned. Fixes #77458. Thanks @jaikharbanda. - Plugins/registry: recover managed-npm external plugins from the owned npm root when a stale persisted registry would otherwise hide them after package-manager upgrades. Fixes #77266. Thanks @p3nchan. - fix(gateway): clamp unbound websocket auth scopes [AI]. (#77413) Thanks @pgondhi987. - Gate zalouser startup name matching [AI]. (#77411) Thanks @pgondhi987. - Active Memory: send a bounded latest-message search query to the recall worker so channel/runtime metadata does not become the memory search string. Fixes #65309. Thanks @joeykrug, @westley3601, @pimenov, and @tasi333. - fix(device-pair): require pairing scope for pair command [AI]. (#76377) Thanks @pgondhi987. - Providers/OpenRouter: keep DeepSeek V4 `reasoning_effort` on OpenRouter-supported values, mapping stale `max` thinking overrides to `xhigh` so `openrouter/deepseek/deepseek-v4-pro` no longer fails with OpenRouter's invalid-effort 400. Fixes #77350. (#77423) Thanks @krllagent, @mushuiyu886, and @sallyom. - fix(qqbot): keep private commands off framework surface [AI]. (#77212) Thanks @pgondhi987. - Claude CLI: honor non-off `/think` levels by passing Claude Code's session-scoped `--effort` flag through the CLI backend seam, so chat bridges no longer show an inert thinking control. Fixes #77303. Thanks @Petr1t. - Agents/subagents: refresh deferred final-delivery payloads when same-session completion output changes, so retried parent notifications use the final child summary instead of stale progress text. Thanks @vincentkoc. - Agents/media: route async music and video completion results back through the requester agent, preserving automatic replies while requiring the message tool only for message-tool-only group/channel delivery. - active-memory: skip the memory sub-agent gracefully instead of logging a confusing allowlist error when no memory plugin (`memory-core` or `memory-lancedb`) is loaded, so active-memory with no memory backend no longer produces misleading "No callable tools remain" warnings in the gateway log. Fixes #77506. Thanks @hclsys. - Memory/wiki: preserve representation from both corpora in `corpus=all` searches while backfilling unused result capacity, so memory hits are not starved by numerically higher wiki integer scores. Fixes #77337. Thanks @hclsys. - Docker/compose: pin container-side `OPENCLAW_CONFIG_DIR` and `OPENCLAW_WORKSPACE_DIR` on both gateway and CLI services so the host paths written into `.env` by `scripts/docker/setup.sh` (used as Compose bind-mount sources) cannot leak into runtime code via the `env_file` import. Fixes regressions on macOS Docker setups where the first agent reply died with `EACCES: permission denied, mkdir '/Users'` because the host-style workspace path got persisted into `agents.defaults.workspace`. Fixes #77436. Thanks @lonexreb. - Telegram: clean up tool-only draft previews after assistant message boundaries so transient `Surfacing...` tool-status bubbles do not linger when no matching final preview arrives. Thanks @BunsDev. - Slack: report `unknown error` instead of `undefined` in socket-mode startup retry logs and label the retry reason explicitly. - Telegram: let explicit forum-topic `requireMention` settings override persisted `/activate` and `/deactivate` state, so per-topic mention gates work consistently. Fixes #49864. Thanks @Panniantong. - Cron: surface failed isolated-run diagnostics in `cron show`, status, and run history when requested tools are unavailable, so blocked cron runs report the actual tool-policy failure instead of a misleading green result. Fixes #75763. Thanks @RyanSandoval. - TUI/escape abort: track the in-flight runId after `chat.send` resolves so pressing Esc during the gap before the first gateway event aborts the run instead of repeatedly printing `no active run`. Fixes #1296. Thanks @Lukavyi and @romneyda. - TUI/render: stop the long-token sanitizer from injecting literal spaces inside inline code spans, fenced code blocks, table borders, and bare hyphenated/dotted identifiers, so copied package names, entity IDs, and shell line-continuations stay byte-for-byte intact while narrow-terminal protection still chunks unidentifiable long prose tokens. Fixes #48432, #39505. Thanks @DocOellerson, @xeusoc, @CCcassiusdjs, @akramcodez, @brokemac79, @romneyda. - Plugin skills: publish plugin-declared skills through the generated plugin skills directory (`~/.openclaw/plugin-skills/`) while keeping direct prompt loading intact, so agent file-based discovery paths find plugin skill `SKILL.md` files and inactive plugin links are cleaned up. Fixes #77296. (#77328) Thanks @zhangguiping-xydt. - Gateway/status: label Linux managed gateway services as `systemd user`, making status output explicit about the user-service scope instead of implying a system-level unit. Thanks @vincentkoc. - Plugins/install: remove the previous managed plugin directory when a reinstall switches sources, so stale ClawHub and npm copies no longer keep duplicate plugin ids in discovery after the new install wins. Thanks @vincentkoc. - Plugins/install: let official plugin reinstall recovery repair source-only installed runtime shadows, so `openclaw plugins install npm:@openclaw/discord --force` can replace the bad package instead of stopping at stale config validation. Thanks @vincentkoc. - CLI/update: stage pnpm-detected npm-layout global package updates through a clean npm prefix swap, keep plugin install runtime imports behind a stable alias, and ship legacy install-runtime aliases back to `2026.3.22`, preventing stale overlay chunks from breaking plugin post-update sync. Thanks @vincentkoc. - Plugins/commands: allow the official ClawHub Codex plugin package to keep reserved `/codex` command ownership, matching the existing npm-managed Codex package behavior. Thanks @vincentkoc. - Auth/OpenAI Codex: rewrite invalidated per-agent Codex auth-order and session profile overrides toward a healthy relogin profile, so revoked OAuth accounts do not stay pinned after signing in again. Thanks @BunsDev. - Plugins/commands: scope QQBot framework slash commands to the QQBot channel so `/bot-*` command handlers and native specs do not leak onto unrelated chat surfaces. Thanks @vincentkoc. - fix: harden backend message action gateway routing [AI]. (#76374) Thanks @pgondhi987. - Gate QQBot streaming command auth [AI]. (#76375) Thanks @pgondhi987. - Plugins/discovery: ignore managed npm plugin packages that only expose TypeScript source entries without compiled runtime output, so stale/broken installs cannot hide a working bundled or reinstallable channel plugin during setup. Thanks @vincentkoc. - CLI/update: treat OpenClaw stable correction versions like `2026.5.3-1` as newer than their base stable release, so package updates no longer ask for downgrade confirmation. Thanks @vincentkoc. - Plugins/install: suppress dangerous-pattern scanner warnings for trusted official OpenClaw npm installs, so installing `@openclaw/discord` no longer prints credential-harvesting warnings for the official package. Thanks @vincentkoc. - Plugins/commands: suppress dangerous-pattern scanner warnings for trusted catalog npm installs from owner-gated `/plugins install` commands, so chat-driven installs match the CLI install trust path. Thanks @vincentkoc. - Plugins/release: make the published npm runtime verifier reject blank `openclaw.runtimeExtensions` entries instead of treating them as absent and passing via inferred outputs. Thanks @vincentkoc. - Plugins/security: ignore inline and block comments when matching source-rule context in plugin install scans, so comment-only `fetch`/`post` references near environment defaults do not block clean plugins. Thanks @vincentkoc. - Doctor/plugins: remove stale managed install records for bundled plugins even when the bundled plugin is not explicitly configured, so doctor cleanup cannot leave orphaned install metadata behind. Thanks @vincentkoc. - Web fetch: scope provider fallback cache entries by the selected fetch provider so config reloads cannot reuse another provider's cached fallback payload. Thanks @vincentkoc. - Web search: honor late-bound `tools.web.search.enabled: false` during tool execution so config reloads cannot leave an already-created `web_search` tool runnable. Thanks @vincentkoc. - Plugins/packages: reject inferred built runtime entries that exist but fail package-boundary checks instead of falling back to TypeScript source for installed packages. Thanks @vincentkoc. - Plugins/loader: do not retry native-loaded JavaScript plugin modules through the source transformer after native evaluation has already reached a missing dependency, avoiding duplicate top-level side effects. Thanks @vincentkoc. - Plugins/packages: reject blank `openclaw.runtimeExtensions` entries instead of silently ignoring them and falling back to inferred TypeScript runtime entries. Thanks @vincentkoc. - Doctor/plugins: remove stale managed npm plugin shadow entries from the managed package lock as well as `package.json` and `node_modules`, so future npm operations do not keep referencing repaired bundled-plugin shadows. Thanks @vincentkoc. - Plugins/runtime state: keep the key being registered when namespace eviction runs in the same millisecond as existing entries, so `register` and `registerIfAbsent` do not report success while evicting their own fresh value. Thanks @vincentkoc. - Plugins/providers: make bundled provider discovery honor restrictive `plugins.allow` by default for new configs, while doctor migrates legacy restrictive allowlist configs to `plugins.bundledDiscovery: "compat"` to preserve upgrade behavior. Thanks @dougbtv. - Control UI/Talk: make failed Talk startup errors dismissable and clear the stale Talk error state when dismissed, so missing realtime voice provider configuration does not leave a permanent chat banner. Fixes #77071. Thanks @ijoshdavis. - Control UI/Talk: stop and clear failed realtime Talk sessions when dismissing runtime error banners, so the next Talk click starts a fresh session instead of only stopping the stale one. Thanks @vincentkoc. - Control UI/Talk: retry from a failed realtime Talk session on the next Talk click instead of requiring a separate stale-session stop click first. Thanks @vincentkoc. - Canvas host: preserve the Gateway TLS scheme in browser canvas host URLs and startup mount logs, so direct HTTPS gateways do not advertise insecure canvas links. Thanks @vincentkoc. - WhatsApp/login: route login success and failure messages through the injected runtime, so setup/onboarding surfaces capture all login output instead of only the QR. Thanks @vincentkoc. - Google Chat: create an isolated Google auth transport per auth client, so google-auth-library interceptor mutations do not accumulate across webhook verification and access-token clients. Thanks @vincentkoc. - Doctor/plugins: remove orphaned or recovered managed npm copies of bundled `@openclaw/*` plugins during `doctor --fix`, so stale package manifests cannot shadow the current bundled plugin config schema. - Control UI/performance: cap long-task and long-animation-frame diagnostics in the shared event log, so slow-render telemetry does not evict gateway/plugin events from the Debug and Overview views. Thanks @vincentkoc. - Gateway/startup: log the canvas host mount only after the HTTP server has bound, so startup logs no longer report the canvas host as mounted before it can serve requests. - Control UI/i18n: render the Sessions active filter tooltip with the configured minute count in every locale and make the i18n check reject placeholder drift. Thanks @BunsDev. - Web fetch: late-bind `web_fetch` config and provider fallback metadata from the active runtime snapshot, matching `web_search` so long-lived tools do not use stale fetch provider settings. Thanks @vincentkoc. - Discord: clear stale startup probe bot/application status when the async bot probe throws, not just when it returns a degraded probe result. Thanks @vincentkoc. - Web search: scope explicit bundled `web_search` provider runtime loading through manifest ownership, so selecting DuckDuckGo/Gemini/etc. does not import unrelated bundled providers or log their optional dependency failures. Thanks @vincentkoc. - Plugins/discovery: demote the source-only TypeScript runtime check on already-installed `origin: "global"` plugin packages from a config-blocking error to a warning and let the runtime fall through to the TypeScript source via jiti, so a single broken installed package no longer blocks `plugins install` for unrelated plugins; install-time rejection of newly-installed source-only packages is unchanged. Thanks @romneyda. - Providers/OpenAI Codex: stop the OAuth progress spinner before showing the manual redirect paste prompt, so callback timeouts do not spam `Browser callback did not finish` across terminals. - Providers/OpenAI Codex: fail closed on malformed `/codex` control commands and diagnostics confirmations before changing bindings, permissions, model overrides, active turns, or feedback uploads. Thanks @vincentkoc. - Providers/OpenAI Codex: sanitize Codex app-server command readouts, failure replies, approval prompts, elicitation prompts, and `request_user_input` text before posting them back into chat. Thanks @vincentkoc. - Providers/OpenAI Codex: preserve local bound-turn image paths, reject stale same-thread turn notifications, enforce option-only user input prompts, and return failed dynamic tool results to Codex as unsuccessful tool calls. Thanks @vincentkoc. - Providers/DeepSeek: expose DeepSeek V4 `xhigh` and `max` thinking levels through the lightweight provider-policy surface, so Control UI `/think` pickers keep showing the max reasoning options when the runtime plugin registry is not active. Fixes #77139. Thanks @bittoby. - Release/beta smoke: resolve the dispatched Telegram beta E2E run from `gh run list` when `gh workflow run` returns no run URL, so the maintainer helper does not fail immediately after dispatch. Thanks @vincentkoc. - Media/images: keep HEIC/HEIF attachments fail-closed when optional Sharp conversion is unavailable instead of sending originals that still need conversion. Thanks @vincentkoc. - Google Meet: fork the caller's current agent transcript into agent-mode meeting consultant sessions, so Meet replies inherit the context from the tool call that joined the meeting. - iOS/mobile pairing: reject non-loopback `ws://` setup URLs before QR/setup-code issuance and let the iOS Gateway settings screen scan QR codes or paste full setup-code messages. Thanks @BunsDev. - Control UI: keep Gateway Access inputs and locale picker contained inside the card at narrow and tablet widths. - Agents/trajectory: bound runtime trajectory capture and yield queued sidecar writes so oversized traces stop recording instead of monopolizing Gateway cleanup. Fixes #77124. Thanks @loyur. - Telegram/streaming: sanitize tool-progress draft preview backticks before shared compaction, so long backtick-heavy progress text still renders inside the safe code-formatted preview instead of collapsing to an ellipsis. - UI/chat: remove the unsupported `line-clamp` declaration from the chat queue text rule to eliminate Firefox console noise without changing visible truncation behavior. Thanks @ZanderH-code. - Control UI: add explicit feedback for repeated actions by announcing session switches, flashing the active session selector, showing inline Save/Apply/Update progress, and distinguishing filtered-empty session lists from genuinely empty session stores. Thanks @BunsDev. - Agents/Pi: suppress persistence for synthetic mid-turn overflow continuation prompts, so transcript-retry recovery does not write the "continue from transcript" prompt as a new user turn. Thanks @vincentkoc. - Agents/tools: strip reasoning text from visible rich presentation titles, blocks, buttons, and select labels before message-tool sends, so structured channel payloads cannot leak hidden planning. Thanks @vincentkoc. - Telegram: keep reply-dispatch lazy provider runtime chunks behind stable dist names and delete `/reasoning stream` previews after final delivery so package updates and live reasoning drafts do not leave Telegram turns broken or noisy. Thanks @BunsDev. - Discord: start the gateway monitor without waiting for the startup bot/application probe, so WSL2 hosts with a slow `/users/@me` REST path still bring the channel online while status enrichment finishes asynchronously. Fixes #77103. Thanks @Suited78. - Exec approvals: detect `env -S` split-string command-carrier risks when `-S`/`-s` is combined with other env short options, so approval explanations do not miss split payloads hidden behind `env -iS...`. Thanks @vincentkoc. - Google Meet: log the concrete agent-mode TTS provider, model, voice, output format, and sample rate after speech synthesis, so Meet logs show which voice backend spoke each reply. - Voice Call: mark realtime calls completed when the realtime provider closes normally, so Twilio/OpenAI/Google realtime stop events do not leave active call records behind. Thanks @vincentkoc. - Gateway/update: keep the shutdown close path behind a stable runtime chunk and ship compatibility aliases for recent `server-close-*` hashes, so manual npm package replacement cannot leave an already-running Gateway unable to shut down cleanly. Fixes #77087. Thanks @westlife219. - Control UI/media: mint short-lived scoped tickets for assistant media fetches and render ticketed URLs instead of exposing long-lived auth tokens in chat image URLs. Fixes #70830 and #77097. Thanks @hclsys. - Exec approvals: treat POSIX `exec` as a command carrier for inline eval, shell-wrapper, and eval/source detection, so approval explanations and command-risk checks do not miss payloads hidden behind `exec`. Thanks @vincentkoc. - Google Meet: log the resolved audio provider model when starting Chrome and paired-node Meet talk-back bridges, so agent-mode joins show the STT model and bidi joins show the realtime voice model. - Diagnostics: handle missing session-tail files in cron recovery context without tripping extension test typecheck. Thanks @vincentkoc. - QA/Slack: update the Slack dispatch preview fallback test SDK mock for structured progress draft helpers, so the rich progress draft regression suite covers the new imports instead of failing before assertions run. Thanks @vincentkoc. - Release validation: allow focused QA live reruns to select Matrix and Telegram without running Slack, so known Slack credential-pool outages do not block non-Slack live proof. Thanks @vincentkoc. - Plugins/loader: keep bundled plugin package `test-api.js` aliases behind private QA mode, so source transforms do not expose test-only public surfaces during normal plugin loading. Thanks @vincentkoc. - Gateway/startup: start cron and record the post-ready memory trace even when deferred maintenance timers fail after readiness, so a non-fatal timer setup issue does not silently leave scheduled jobs idle. Thanks @vincentkoc. - Exec approvals: unwrap BSD/macOS `env -P <path>` carrier commands before approval-command and strict inline-eval checks, so `/approve` shell execution and inline interpreter payloads are still blocked behind that env form. - Agents/session status: keep semantic `session_status({ sessionKey: "current" })` on the live run session even before that run has a persisted session-store entry, instead of falling back to the sandbox policy key. Thanks @vincentkoc. - QA/Slack: resolve bundled official plugin public-surface package aliases during source-mode QA runs, so release Slack live validation can load `@openclaw/slack/api.js` without workspace symlinks. Thanks @vincentkoc. - Codex: pass the live run session key into app-server dynamic tools when sandbox policy uses a separate session key, so `session_status({ sessionKey: "current" })` reports the active run instead of the sandbox policy key. Thanks @vincentkoc. - Web search: keep first-class assistant `web_search` auto-detect and configured runtime providers visible when active runtime metadata or the active plugin registry is incomplete. Fixes #77073. Thanks @joeykrug. - Plugins/tools: mark manifest-optional sibling tools as optional even when they come from a shared non-optional factory, so cached/status/MCP metadata keeps opt-in tool policy accurate. Thanks @vincentkoc. - Matrix: keep `streaming.progress.toolProgress` scoped to progress draft mode, so partial and quiet Matrix previews do not lose tool progress unless `streaming.preview.toolProgress` is disabled. Thanks @vincentkoc. - Gateway/validation: isolate gateway server validation files, ignore unrelated startup logs in request-trace coverage, and fail fast on stuck shared-auth sockets, reducing false main-branch CI failures for contributors. Thanks @amknight. - Channels/streaming: keep `streaming.progress.toolProgress` scoped to progress draft mode, so disabling compact progress lines does not silence partial/block preview tool updates. Thanks @vincentkoc. - Plugins/update: treat OpenClaw stable correction versions like `2026.5.3-1` as stable releases for npm installs, plugin updates, and bundled-version comparisons, so `latest` can advance official plugins without prerelease opt-in. Thanks @vincentkoc. - Control UI: point the Appearance tweakcn browse action and docs at the live tweakcn editor route instead of the removed `/themes` page. Fixes #77048. - Control UI: render Dream Diary prose through the sanitized markdown pipeline, so diary bold/italic/header markdown no longer appears as literal source text. Fixes #62413. - Control UI: render tool results whose output arrives as text-block arrays and give expanded tool output a scrollable block, so read/exec output remains visible in WebChat. Fixes #77054. - MCP: include serialized conversation/message payloads in the primary text content for `conversations_list` and `messages_read`, while preserving `structuredContent` for capable clients. Fixes #77024. - Media: treat `EPERM` from the post-write media fsync step as best-effort, allowing WebChat and channel uploads to finish on Windows filesystems that reject `fsync` after a successful write. Fixes #76844. - Media/Telegram: send in-limit original images when optional image optimization is unavailable, so Telegram MEDIA replies and message-tool image sends do not fail just because `sharp` is missing. Fixes #77081. (#77117) Thanks @pfrederiksen. - Diagnostics: include last progress, cron job/run ids, stopped cron job name, and the last assistant transcript snippet in stalled-session and stuck-session recovery logs so cron stalls show what was stopped. - Streaming channels: add `streaming.preview.commandText: "status"` / `streaming.progress.commandText: "status"` to hide command/exec text in preview progress lines while keeping the released raw command text default. Fixes #77072. - Agents/cron: let explicit cron `timeoutSeconds` drive both CLI no-output and embedded LLM idle watchdogs instead of being capped by resume defaults. Fixes #76289. - Plugins/catalog: suppress missing `channelConfigs` compatibility diagnostics for external channel plugins that are disabled, denied, or outside a restrictive allowlist. Fixes #76095. - Diagnostics: keep webhook/message OTEL attributes and Prometheus delivery labels low-cardinality and omit raw chat/message IDs from spans, so progress-draft and message-tool modes do not leak high-cardinality messaging identifiers. - Google Meet: stop advertising legacy `mode: "realtime"` to agents and config UIs, while keeping it as a hidden compatibility alias for `mode: "agent"`, so new joins use the STT -> OpenClaw agent -> TTS path instead of selecting the direct realtime voice fallback. - Google Meet: add `chrome.audioBufferBytes` for generated command-pair SoX audio commands and lower the default buffer from SoX's 8192 bytes to 4096 bytes to reduce Chrome talk-back latency. - Google Meet: split realtime provider config into agent-mode transcription and bidi-mode voice providers, and migrate legacy Gemini Live bidi configs with `doctor --fix`, so Gemini Live can back direct bidi fallback without breaking the default OpenClaw agent talk-back path. - Google Meet: keep waiting for the Meet microphone to unmute during join intro readiness instead of permanently skipping talk-back when Meet briefly reports the local mic as muted. - Google Meet: expose `voiceCall.postDtmfSpeechDelayMs` in the plugin manifest schema and setup hints, so manifest-based config editing accepts the runtime-supported Twilio delay key. Thanks @vincentkoc. - Google Meet: keep explicit non-Google `realtime.provider` values as the transcription provider compatibility fallback when `realtime.transcriptionProvider` is unset. Thanks @vincentkoc. - Google Meet: make Twilio setup status require an enabled `voice-call` plugin entry instead of treating a missing entry as ready. Thanks @vincentkoc. - Telegram: render shared interactive reply buttons in reply delivery so plugin approval messages show inline keyboards. (#76238) Thanks @keshavbotagent. - Cron/sessions: keep cron metadata rows without an on-disk transcript non-resumable until a transcript exists, so doctor and `sessions cleanup --fix-missing` no longer report or prune pre-transcript cron rows as broken sessions. Refs #77011. - OpenAI Codex: recreate missing bound app-server threads once when a stale `/codex bind` sidecar survives a restart, preserving the selected auth profile and turn overrides before retrying the inbound turn. (#76936) Thanks @keshavbotagent. - Agents/cli-runner: drop a saved `claude-cli` resume sessionId at preparation time when its on-disk transcript no longer exists in `~/.claude/projects/`, so a stale binding from a half-installed `update.run` cannot trap follow-up runs (auto-reply / Telegram direct) in a `claude --resume` timeout loop; the run starts fresh and the new sessionId is written back through the existing post-run flow. (#77030; refs #77011) Thanks @openperf. - Release validation: install the cross-OS TypeScript harness through Windows-safe Node/npm shims so native Windows package checks reach the OpenClaw smoke suites instead of exiting before artifact capture. Thanks @vincentkoc. - Release validation: let Windows packaged-upgrade checks continue after the shipped 2026.5.2 updater hits its native-module swap cleanup fallback, verifying the fallback-installed candidate through package metadata and downstream smoke instead of crashing on the immediate update-status probe. Thanks @vincentkoc. - Doctor/plugins: skip channel-derived official plugin installs when another configured plugin is the effective owner for the same channel, so `doctor --repair` does not reinstall `feishu` while `openclaw-lark` handles `channels.feishu`. Fixes #76623. Thanks @fuyizheng3120. - Gateway/sessions: memoize repeated thinking-option enrichment and skip unused cost fallback checks while listing sessions, reducing per-row work on large multi-agent stores. Fixes #76931. - Gateway/sessions: bound default `sessions.list` RPC responses and report truncation metadata, preventing Slack-heavy long-lived stores from forcing unbounded Gateway row construction. Fixes #77062. - Agents/tools: use config-only runtime snapshots for plugin tool registration and live runtime config getters, avoiding expensive full secrets snapshot clones on the core-plugin-tools prep path. Fixes #76295. - Agents/tools: honor the effective tool denylist before constructing optional PDF/media tool factories, so `tools.deny: ["pdf"]` skips PDF setup before later policy filtering. Fixes #76997. - MCP/plugin tools: apply global `tools.profile`, `tools.alsoAllow`, and `tools.deny` policy while exposing plugin tools over the standalone MCP bridge, so ACP clients do not see policy-hidden plugin tools or miss opt-in optional tools. Thanks @vincentkoc. - Plugin tools: honor explicit tool denylists while selecting plugin tool runtimes, so denied plugin tools are not materialized for direct command or gateway surfaces before later policy filtering. Thanks @vincentkoc. - Plugin tools: filter factory-returned tools by manifest per-tool optional policy, so optional sibling tools from a shared runtime factory stay hidden unless explicitly allowed. Thanks @vincentkoc. - Agents/transcripts: retry context-overflow compaction from the current transcript only after the inbound user turn was actually persisted, and keep WebChat agent-run live delivery from writing duplicate Pi-managed assistant turns. Fixes #76424. (#77033) - Agents/bootstrap: keep pending `BOOTSTRAP.md` and bootstrap truncation notices in system-prompt Project Context instead of copying setup text or raw warning diagnostics into WebChat user/runtime context. Fixes #76946. - Gateway/install: keep `.env`-managed values in the macOS LaunchAgent env file while still tracking `OPENCLAW_SERVICE_MANAGED_ENV_KEYS`, so regenerated services do not boot without managed auth/provider keys. Fixes #75374. - Gateway/restart: verify listener PIDs by argv when `lsof` reports only the Node process name, so stale gateway cleanup can find macOS `cnode` listeners. Fixes #70664. - Gateway/logging: expand leading `~` in `logging.file` before creating the file logger, preventing startup crash loops for home-relative log paths. Fixes #73587. - Channels/CLI: keep `openclaw channels list --json` usable when provider usage fetching fails, and report per-provider usage errors without aborting the channel list. Refs #67595. - Doctor/plugins: do not treat `plugins.allow` entries as configured plugins during missing-plugin repair, so restrictive allowlists no longer install allowed-but-unused plugins. Thanks @vincentkoc. - Agents/messaging: deliver distinct final commentary after same-target `message` tool sends while still deduping text/media already sent by the tool, so short closing remarks are no longer silently dropped. Fixes #76915. Thanks @hclsys. - Agents/messaging: preserve string thread IDs when matching message-tool reply dedupe routes, avoiding precision loss on numeric-looking topic IDs before channel plugin comparison. Thanks @vincentkoc. - Channels/streaming: honor `agents.defaults.toolProgressDetail: "raw"` in Slack, Discord, Telegram, Matrix, and Microsoft Teams progress drafts, so tool-start lines include raw command/detail output when debugging. Thanks @vincentkoc. - Channels/streaming: strip unmatched inline-code backticks from compacted raw progress draft lines, avoiding stray markdown markers after long command details are shortened. Thanks @vincentkoc. - Discord/Slack/Mattermost: align draft preview tool-progress config help with the runtime behavior that hides interim tool updates when `streaming.preview.toolProgress` is false. Thanks @vincentkoc. - Feishu: use the shared channel progress formatter for streaming-card tool status lines, including raw command/detail output and message-tool filtering. Thanks @vincentkoc. - Mattermost: use the shared progress draft formatter for tool status previews, including raw command/detail output when `agents.defaults.toolProgressDetail: "raw"` is enabled. Thanks @vincentkoc. - Mattermost: suppress standalone default tool-progress messages while draft previews are active, including when draft tool lines are disabled. Thanks @vincentkoc. - Telegram: deliver button-only interactive replies by sending the shared fallback button-label text with the inline keyboard instead of dropping the reply as empty. Thanks @vincentkoc. - OpenAI Codex: honor `auth.order.openai-codex` when starting app-server clients without an explicit auth profile, so status/model probes and implicit startup use the configured Codex account instead of falling back to the default profile. Thanks @vincentkoc. - OpenAI Codex: let SSRF-guarded provider requests inherit OpenClaw's undici IPv4/IPv6 fallback policy, so ChatGPT-backed Codex runs recover on IPv4-working hosts when DNS still returns unreachable IPv6 addresses. Fixes #76857. Thanks @jplavoiemtl and @SymbolStar. - Plugin updates: do not short-circuit trusted official npm updates as unchanged when the default/latest spec still resolves to an already-installed prerelease that the installer should replace with a stable fallback. Thanks @vincentkoc. - Plugin updates: clean stale bundled load paths for already-externalized npm installs whose legacy install record only preserved the resolved package name. Thanks @vincentkoc. - Plugin tools: keep auth-unavailable optional tools hidden even when another default tool from the same plugin is available and `tools.alsoAllow` names the optional tool. Thanks @vincentkoc. - Realtime transcription: report socket closes before provider readiness as closed-before-ready failures instead of mislabeling them as connection timeouts for OpenAI, xAI, and Deepgram streaming transcription. Thanks @vincentkoc. - OpenAI/Google Meet: fail realtime voice connection attempts when the socket closes before `session.updated`, avoiding stuck Meet joins waiting on a bridge that never became ready. Thanks @vincentkoc. - Google Meet: avoid treating repeated participant words as multiple assistant-overlap matches when suppressing realtime echo transcripts. Thanks @vincentkoc. - Google Meet: make `mode: "agent"` the default Chrome talk-back path, using realtime transcription for input and regular OpenClaw TTS for speech output, while keeping direct realtime voice answers available as `mode: "bidi"` and accepting `mode: "realtime"` as an agent-mode compatibility alias. - Codex harness: keep `codex_app_server.*` telemetry publication owned by the harness instead of republishing the same callback event from core runners. Thanks @vincentkoc. - Slack/Discord: suppress standalone tool-progress chatter when partial preview streaming has `streaming.preview.toolProgress: false`, matching the documented quiet-preview behavior. Thanks @vincentkoc. - Matrix: bind native approval reaction targets before publishing option reactions, so fast approver reactions on threaded prompts are not dropped while the approval handler finishes setup. Thanks @vincentkoc. - Google Meet: make realtime talk-back agent-driven by default with `realtime.strategy: "agent"`, keep the previous direct bidirectional model behavior available as `realtime.strategy: "bidi"`, route the Meet tab speaker output to `BlackHole 2ch` automatically for local Chrome realtime joins, coalesce nearby speech transcript fragments before consulting the agent, and avoid cutting off agent speech from server VAD or stale playback pipe errors. - Google Meet: suppress queued assistant playback and assistant-like transcript echoes from the realtime input path, so the meeting does not hear the agent's own speech as a new user turn and loop or cut itself off. - Google Meet: keep Chrome realtime transport tests hermetic on Linux prerelease shards while preserving the macOS-only runtime guard. Thanks @vincentkoc. - QA/Matrix: let the live tool-progress preview and error checks verify progress replacement events without depending on the preview saying `Working`, `tool: read`, an unlabelled/pathless `read from`, or the original draft root being observed. Thanks @vincentkoc. - QA/Matrix: keep the target=both approval scenario focused on channel and DM metadata delivery by resolving the accepted approval through the gateway after both Matrix events are observed. Thanks @vincentkoc. - QA/Matrix: wait for live approval reactions to echo before starting the threaded approval decision timeout. Thanks @vincentkoc. - QA/Matrix: reuse the primed driver sync stream when confirming approval reaction echoes, avoiding missed self-reactions in live release runs. Thanks @vincentkoc. - Channels/WhatsApp: apply the shared group/channel visible-reply mode during inbound dispatch so group replies stay message-tool-only by default without overriding direct-chat harness defaults. Refs #75178 and #67394. Thanks @scoootscooob. - Plugins/Codex: preserve Codex-native OAuth routing for `/codex bind` app-server turns so bound sessions keep the selected Codex auth profile instead of falling back to public OpenAI credentials. (#76714) Thanks @keshavbotagent. - Telegram: keep status checks pointed at the active chat so asking for the current session no longer reports an old direct-message conversation. (#76708) Thanks @amknight. - Gateway/install: prefer supported system Node over nvm/fnm/volta/asdf/mise when regenerating managed gateway services, so `gateway install --force` no longer recreates service definitions that doctor immediately flags as version-manager-backed. Fixes #76339. Thanks @brokemac79 and @BunsDev. - Google Chat: normalize Google auth certificate response headers before google-auth-library reads cache-control, so inbound webhook auth no longer rejects with `res?.headers.get is not a function`. Fixes #76880. Thanks @donbowman. - WhatsApp: route terminal login QR output through the active runtime for initial and restart sockets, so `openclaw channels login --channel whatsapp` does not lose the QR behind direct stdout writes. Fixes #76213. Thanks @dougvk. - Proxy/debugging: disable debug proxy direct upstream forwarding for proxy requests and CONNECT tunnels while managed proxy mode is active unless `OPENCLAW_DEBUG_PROXY_ALLOW_DIRECT_CONNECT_WITH_MANAGED_PROXY=1` is explicitly set for approved local diagnostics. Thanks @jesse-merhi and @mjamiv. - Direct APNs: route direct HTTP/2 delivery through the active managed proxy with redacted proxy diagnostics, so push requests honor configured egress controls and `openclaw proxy validate --apns-reachable` can prove APNs is reachable through the proxy before deployment. (#74905) Thanks @jesse-merhi. - Agents/subagents: detect prefix-only completion announce replies and fall back to the captured child result so requester chats no longer lose most of long sub-agent reports silently. Fixes #76412. Thanks @inxaos and @davemorin. - TUI: replace the stale-response watchdog notice with plain user-facing copy so stalled replies no longer surface backend or streaming internals. (#77120) Thanks @davemorin. - Security/Windows: validate `SystemRoot`/`WINDIR` env values through the Windows install-root validator and add them to the dangerous-host-env policy when resolving `icacls.exe`/`whoami.exe` for `openclaw security audit`, so workspace `.env` overrides and bare command names cannot redirect Windows ACL helpers to attacker-controlled binaries. (#74458) Thanks @mmaps. - Security/Windows: pin Windows registry-probe `reg.exe` resolution to the canonical Windows install root in install-root probing, so `SystemRoot`/`WINDIR` env overrides cannot redirect registry queries during Windows host detection. (#74454) Thanks @mmaps. - QQBot: preserve the framework command authorization decision when converting framework command contexts into engine slash command contexts, so downstream slash handlers see `commandAuthorized` matching the channel's resolved `isAuthorizedSender` instead of a hardcoded `true`. (#77453) Thanks @drobison00. - Security/Windows: block `LOCALAPPDATA` from workspace `.env` and resolve Windows update-flow portable Git path prepends from the trusted process-local `LOCALAPPDATA` only, so workspace-supplied values cannot redirect `git` discovery during `openclaw update`. (#77470) Thanks @drobison00. - Browser/SSRF: enforce the existing current-tab URL navigation policy before tab-scoped debug, export, and read routes (console, page errors, network requests, trace start/stop, response body, screenshot, snapshot, storage, etc.) collect from an already-selected tab, so blocked tabs return a policy error instead of being read first and redacted only at response time. (#75731) Thanks @eleqtrizit. - Security/Windows: route the `.cmd`/`.bat` process wrapper through the shared Windows install-root resolver instead of `process.env.ComSpec`, so workspace dotenv-blocked `SystemRoot`/`WINDIR` overrides and unsafe values like UNC paths or path-lists cannot redirect `cmd.exe` selection on Windows. (#77472) Thanks @drobison00. - Agents/bootstrap: honor `BOOTSTRAP.md` content injected by `agent:bootstrap` hooks when deciding whether bootstrap is pending, so hook-provided required setup instructions are included in the system prompt. (#77501) Thanks @ificator.