OpenClaw v2026.4.25-beta.3 预发布

## 2026.4.25 ### Highlights - Voice replies get a full TTS upgrade: `/tts latest`, chat-scoped auto-TTS controls, personas, per-agent/per-account overrides, and new Azure Speech, Xiaomi, Local CLI, Inworld, Volcengine, and ElevenLabs v3 provider coverage. Thanks @leonchui, @zoujiejun, @solar2ain, @cshape, @xuruiray, @itsuzef, and @barronlroth. - Plugin startup and install paths move to the cold persisted registry, cutting broad manifest scans while making plugin update, repair, provider discovery, and install metadata more deterministic. Thanks @vincentkoc and @shakkernerd. - OpenTelemetry coverage expands across model calls, token usage, tool loops, harness runs, exec processes, outbound delivery, context assembly, and memory pressure with bounded low-cardinality attributes. Thanks @vincentkoc, @jlapenna, @Lidang-Jiang, and @oc-factus. - Browser automation gets safer tab URLs, iframe-aware role snapshots, CDP readiness tuning, headless one-shot launch, and deeper browser doctor probes for slow hosts. Thanks @beat843796 and @BenediktSchackenberg. - Control UI and setup flows add PWA/Web Push support, Crestodian first-run repair, TUI setup, context mode selection, and a shorter startup greeting. Thanks @eduardocruz, @SebTardif, and @kevinlin-openai. - Install/update hardening covers Windows, macOS, Linux, Docker, bundled plugin runtime deps, Node service restarts, LaunchAgent token rotation, and mixed-version gateway verification. Thanks @Kobevictor, @igormf, @abhinas90, @jsompis, @Solvely-Colin, and @gucasbrg. ### Changes - TTS/WhatsApp: add `/tts latest` read-aloud support with duplicate suppression and `/tts chat on|off|default` session-scoped auto-TTS overrides, completing the on-demand voice-note UX for current-chat replies. Fixes #66032. - TTS/channels: resolve channel and account TTS overrides generically, enabling Feishu and QQBot accounts to deep-merge `channels.<channel>.accounts.<id>.tts` over global and per-agent TTS config. Thanks @sahilsatralkar. - TTS/agents: allow `agents.list[].tts` to override global `messages.tts` for per-agent voices, and make `/tts audio`, `/tts status`, and the `tts` agent tool honor the active voice/provider override while keeping shared provider credentials and preferences in the existing TTS config surface. - Providers/Azure Speech: add Azure Speech as a bundled TTS provider with Speech-resource auth, voice listing, SSML escaping, native Ogg/Opus voice-note output, and telephony output. (#51776) Thanks @leonchui. - Google Meet: add calendar-backed attendance export workflows, export manifests, dry-run previews, and tool parity for meeting records. - Control UI: add PWA install support and Web Push notifications for Gateway chat. (#44590) Thanks @eduardocruz. - Browser automation: add safe tab URLs in agent responses plus a CDP-native role snapshot fallback with iframe-aware refs, cursor-clickable detection, target attach preparation, and `openclaw browser doctor --deep` live snapshot probing. - CLI/image generation: expose generic `--background` on `openclaw infer image generate` and `openclaw infer image edit`, keep `--openai-background` as an OpenAI alias, and let fal image generation honor `--output-format png|jpeg`. - Browser/config: allow local managed Chrome launch discovery and post-launch CDP readiness timeouts to be raised for slower hosts such as Raspberry Pi. Fixes #66803. Thanks @beat843796. - Discord: allow `channels.discord.voice.model` to override the LLM used for voice channel responses while keeping STT and TTS on their existing media settings. (#64368) Thanks @mrdavey. - Browser/CLI: add `openclaw browser start --headless` as a one-shot local managed browser launch override without rewriting persisted browser config. Thanks @BenediktSchackenberg. - CLI/Crestodian/TUI: add the first-run setup helper, local planner fallback, full-TUI interactive Crestodian, startup progress indicators, context mode selector, and a shorter startup greeting. (#71720, #71760) Thanks @SebTardif and @kevinlin-openai. - Plugins: migrate the local plugin registry automatically during package install/update, keeping install metadata in the plugin index while indexing existing plugin manifests for the new cold registry path. Thanks @vincentkoc and @shakkernerd. - Plugins/doctor: make `openclaw doctor --fix` refresh the plugin index and cold registry index when needed without treating plugin install records as authored config. Thanks @vincentkoc and @shakkernerd. - Plugins/hooks: add before-agent-finalize hooks, cron `jobId` hook context, bounded native permission fingerprints, and Codex MCP hook relay support. (#71765, #71758, #71707) Thanks @vincentkoc and @pashpashpash. - Plugins/tokenjuice: bump the bundled tokenjuice runtime to 0.6.3. Thanks @vincentkoc. - Diagnostics/OTEL: align model-call GenAI span attributes with OpenTelemetry stability opt-in semantics, keeping legacy `gen_ai.system` by default while emitting `gen_ai.provider.name` under `OTEL_SEMCONV_STABILITY_OPT_IN=gen_ai_latest_experimental`. Thanks @vincentkoc. - Diagnostics/OTEL: support signal-specific OTLP endpoint overrides for traces, metrics, and logs via config or standard OTEL environment variables. Thanks @vincentkoc. - Diagnostics/OTEL: emit bounded telemetry exporter health diagnostics for startup and log-export failures without exporting raw error text. Thanks @vincentkoc. - Diagnostics/OTEL: export agent harness lifecycle telemetry as bounded `openclaw.harness.run` spans and `openclaw.harness.duration_ms` metrics so QA-lab, Codex, and future harnesses share one trace shape. Thanks @vincentkoc. - Diagnostics/trace: propagate W3C `traceparent` headers from trusted model-call trace context to provider transports while replacing caller-supplied traceparent values. Thanks @vincentkoc. - Diagnostics/Prometheus: add a bundled `diagnostics-prometheus` plugin with a protected gateway scrape route for low-cardinality diagnostics metrics. Thanks @vincentkoc. - Plugins/CLI: add `openclaw plugins registry` for explicit persisted-registry inspection and `--refresh` repair without making normal startup rescan plugin locations. Thanks @vincentkoc. - Plugins/CLI: make `openclaw plugins list` read the cold persisted registry snapshot by default, leaving module-aware diagnostics to `plugins doctor` and `plugins inspect`. Thanks @vincentkoc. - Plugins/startup: move gateway startup plugin planning onto the versioned cold registry index, with postinstall repair for older registry files that predate startup metadata. Thanks @vincentkoc. - Plugins/startup: normalize startup and provider plugin enablement through registry aliases so boot paths do not need the legacy manifest alias scan. Thanks @vincentkoc. - Providers/plugins: resolve provider ownership, provider discovery scopes, and catalog-hook provider ids from the cold plugin registry instead of rescanning manifests on those paths. Thanks @vincentkoc. - Plugins/registry: keep installed plugin index records focused on install/state/load paths and resolve plugin capabilities from manifests scoped to indexed plugins. Thanks @shakkernerd. - Plugins/registry: route cold manifest and capability lookups through the installed plugin index so setup, channels, config, secrets, doctor, and provider metadata paths avoid broad plugin-root scans before runtime execution. Thanks @shakkernerd. - CLI/models: speed up `models list --all --provider <id>` for static manifest-backed providers by loading catalog rows through the installed plugin index instead of broad manifest scans or runtime suppression hooks. Thanks @shakkernerd. - CLI/models: use OpenClaw Provider Index preview rows as the final cold fallback for installable providers, while keeping user config, installed manifests, and refreshed cache rows above provider-index metadata. Thanks @vincentkoc. - Providers/plugins: keep onboarding and auth-choice setup lists on cold manifest/install metadata and add Provider Index install metadata for not-yet-installed provider plugins. Thanks @vincentkoc. - Providers/plugins: keep provider setup guidance and configure auth imports on cold manifest metadata, with a regression guard against static provider-runtime imports on setup/configure list paths. Thanks @vincentkoc. - CLI/capabilities: keep capability command registration from importing the models auth runtime until `model auth login` actually runs. Thanks @vincentkoc. - CLI/configure: keep web-search configure prompts on cold plugin registry metadata until the user chooses managed search setup. Thanks @vincentkoc. - Plugins/chat commands: refresh the persisted plugin registry after `/plugins enable` and `/plugins disable`, matching the CLI mutation path. Thanks @vincentkoc. - Plugins/compat: mark `OPENCLAW_DISABLE_PERSISTED_PLUGIN_REGISTRY` as a deprecated break-glass switch and point operators at registry repair instead. Thanks @vincentkoc. - Plugins/compat: expand the central compatibility registry with dated owners, replacements, and maximum three-month removal targets for legacy SDK, manifest, setup, registry-migration, and agent-runtime surfaces. Thanks @vincentkoc. - Plugins/registry: ignore stale persisted registry reads when plugin policy no longer matches current config, and stamp generated registry files with a do-not-edit warning. Thanks @vincentkoc. - Config/plugins: keep plugin command-alias validation on cold manifest metadata instead of importing the runtime alias resolver. Thanks @vincentkoc. - Security/plugins: keep web-search credential presence checks on cold config, env, and manifest metadata instead of importing web-search provider runtime. Thanks @vincentkoc. - Diagnostics/OTEL: surface provider request identifiers as bounded hashes on model-call diagnostics and span events, without exporting raw request IDs or metric labels. Thanks @Lidang-Jiang and @vincentkoc. - Plugins/diagnostics: add metadata-only `model_call_started` and `model_call_ended` hooks for provider/model call telemetry without exposing prompts, responses, headers, request bodies, or raw provider request IDs. Thanks @vincentkoc. - Diagnostics/OTEL: emit bounded context assembly diagnostics and export `openclaw.context.assembled` spans with prompt/history sizes but no prompt, history, response, or session-key content. Thanks @vincentkoc. - Diagnostics/OTEL: export existing tool-loop diagnostics as `openclaw.tool.loop` counters and spans without loop messages, session identifiers, params, or tool output. Thanks @vincentkoc. - Diagnostics/OTEL: export diagnostic memory samples and pressure as bounded memory histograms, counters, and pressure spans to help spot leak regressions without session or payload data. Thanks @vincentkoc. - Diagnostics/OTEL: add the GenAI `gen_ai.client.token.usage` histogram for input/output model usage while keeping session identifiers and aggregate cache counters out of the semantic metric. Thanks @vincentkoc. - Diagnostics/OTEL: add a bounded `openclaw.agent` label to OpenClaw token metrics so per-agent Grafana dashboards can group usage without exporting session identifiers. Thanks @oc-factus. - Plugins/install: consolidate managed plugin install metadata into the state-managed plugin index at `plugins/installs.json`, replacing the temporary `plugins/installed-index.json` path and removing `plugins.installs` as an authored config surface. Thanks @vincentkoc and @shakkernerd. - Diagnostics/OTEL: add the GenAI `gen_ai.client.operation.duration` histogram for model-call latency in seconds with bounded provider/model/API and error attributes. Thanks @vincentkoc. - Diagnostics/OTEL: add GenAI usage token attributes to model-usage spans, including cache read/write input token counts without session identifiers or prompt/response content. Thanks @vincentkoc. - Diagnostics/OTEL: include bounded GenAI operation, provider, and request-model attributes on model-usage spans so token usage remains self-describing without diagnostic identifiers. Thanks @vincentkoc. - Diagnostics/OTEL: keep model-usage span GenAI provider attributes aligned with the existing semantic-convention opt-in policy, using legacy `gen_ai.system` unless latest experimental GenAI conventions are enabled. Thanks @vincentkoc. - Diagnostics/OTEL: keep `gen_ai.request.model` present on GenAI token usage metrics with a bounded `unknown` fallback when model usage events do not include a model. Thanks @vincentkoc. - Docs/OTEL: document the GenAI token and model-call duration metrics, model-usage span attributes, and `OTEL_SEMCONV_STABILITY_OPT_IN=gen_ai_latest_experimental` provider-attribute behavior. Thanks @vincentkoc. - Docs: refresh the MCP, model provider, doctor, troubleshooting, BlueBubbles, media generation, TTS, subagents, skills, cron/tasks, exec approvals, and voice-call guides with structured Steps, Tabs, and Accordion content. - Diagnostics/trace: add an internal traceparent propagation helper that only formats trusted dispatcher metadata, keeping plugin-emitted diagnostic traces out of outbound propagation by default. Thanks @vincentkoc. - Diagnostics/OTEL: add bounded outbound message delivery lifecycle diagnostics and export them as low-cardinality delivery spans/metrics without message body, recipient, room, or media-path data. (#71471) Thanks @vincentkoc and @jlapenna. - Diagnostics/OTEL: emit bounded exec-process diagnostics and export them as `openclaw.exec` spans without exposing command text, working directories, or container identifiers. (#71451) Thanks @vincentkoc and @jlapenna. - Diagnostics/OTEL: support `OPENCLAW_OTEL_PRELOADED=1` so the plugin can reuse an already-registered OpenTelemetry SDK while keeping OpenClaw diagnostic listeners wired. (#71450) Thanks @vincentkoc and @jlapenna. - Providers/Xiaomi: add MiMo TTS as a bundled speech provider with MP3/WAV output and voice-note Opus transcoding. Fixes #52376. (#55614) Thanks @zoujiejun. - Providers/ElevenLabs: include `eleven_v3` in the bundled TTS model catalog so model selection surfaces can offer ElevenLabs v3. (#68321) Thanks @itsuzef. - Providers/Local CLI TTS: add a bundled local command speech provider with file/stdout input, voice-note Opus conversion, and telephony PCM output. (#56239) Thanks @solar2ain. - Providers/Inworld: add Inworld as a bundled speech provider with streaming TTS synthesis, voice listing, voice-note output, and PCM telephony output. (#55972) Thanks @cshape. - Providers/Volcengine: add Volcengine/BytePlus Seed Speech as a bundled TTS provider with API-key auth, native Ogg/Opus voice-note output, and MP3 audio-file output. (#55641) Thanks @xuruiray. - Android/Talk Mode: expose Talk Mode in the Voice tab with runtime-owned voice capture modes and microphone foreground-service escalation. Thanks @alex-latitude. - Providers/LiteLLM: register `litellm` as an image-generation provider so `image_generate model=litellm/...` calls and `agents.defaults.imageGenerationModel.fallbacks` entries resolve through the LiteLLM proxy. Thanks @zqchris. - Providers/fal: add Seedance 2.0 reference-to-video models with multi-image, video, and audio reference input mapping plus model-specific capability limits for `video_generate`. Thanks @shivanker. - Codex harness: require Codex app-server `0.125.0` or newer and cover native MCP `PreToolUse`, `PostToolUse`, and `PermissionRequest` payloads through the OpenClaw hook relay. - Agents/Codex: teach prompts and `agents_list` to surface native Codex app-server availability so agents prefer `/codex ...` over Codex ACP unless ACP/acpx is explicit. Thanks @vincentkoc. - ACPX/Droid: add Factory Droid to the live ACP bind Docker matrix, including `.factory` settings staging, `FACTORY_API_KEY` forwarding, and the single-agent `test:docker:live-acp-bind:droid` recipe. - TTS/personas: add provider-aware TTS personas with deterministic provider binding merges, `/tts persona` controls, gateway/CLI persona state, Google Gemini `audio-profile-v1` prompt wrapping, and OpenAI instruction mapping. (#70748) Thanks @barronlroth. - Voice Wake: add trigger-based routing so macOS voice wake phrases can select a configured agent or session target, with Gateway routing APIs and node update events. (#30354) Thanks @longbiaochen. ### Fixes - Plugins/CLI: let flag-driven `openclaw channels add` install the selected channel plugin from its default source without opening an interactive prompt, fixing published npm Telegram setup in stdin-closed automation. Thanks @codex. - Plugins/startup: load the default `memory-core` slot during Gateway startup when permitted so active-memory recall can call `memory_search` and `memory_get` without requiring an explicit `plugins.slots.memory` entry, while preserving `plugins.slots.memory: "none"`. Thanks @codex. - Plugins/CLI: prefer native require for compiled bundled plugin JavaScript before jiti so read-only config, status, device, and node commands avoid unnecessary transform overhead on slow hosts. Fixes #62842. Thanks @Effet. - Plugins/compat: inventory doctor-side deprecation migrations separately from runtime plugin compatibility so release sweeps preserve needed repairs while enforcing dated removal windows. Thanks @vincentkoc. - Plugins/compat: add missing dated compatibility records for legacy extension-api, memory registration, provider hook/type aliases, runtime aliases, channel SDK helpers, and approval/test utility shims. Thanks @vincentkoc. - Plugins/CLI: refresh the persisted registry after managed plugin files are removed so ClawHub uninstall cannot leave stale `plugins list` entries. Thanks @codex. - Plugins/CLI: make plugin install and uninstall config writes conflict-aware, clear stale denylist entries on explicit reinstall/removal, and delete managed plugin files only after config/index commit succeeds. Thanks @codex. - Plugins: fail `plugins update` when tracked plugin or hook updates error, keep bundled runtime-dependency repair behind restrictive allowlists, and reject package installs with unloadable extension entries. Thanks @codex. - Gateway/chat: keep duplicate attachment-backed `chat.send` retries with the same idempotency key on the documented in-flight path so aborts still target the real active run. Fixes #70139. Thanks @Feelw00. - Plugins: share package entrypoint resolution between install and discovery, reject mismatched `runtimeExtensions`, and cache bundled runtime-dependency manifest reads during scans. Thanks @codex. - Onboarding/setup: keep first-run config reads, plugin compatibility notices, and post-model sanity checks on cold metadata paths unless the user chooses to browse all models, avoiding full plugin/runtime catalog work between prompts. Thanks @shakkernerd. - Onboarding/auth: run manifest-owned provider auth choices through scoped setup providers so selecting OpenAI Codex browser/device auth no longer loads every provider runtime before OAuth starts. Thanks @shakkernerd. - Onboarding/auth: keep the post-auth default-model policy lookup on manifest/setup metadata so the next prompt appears without loading broad provider runtime. Thanks @shakkernerd. - Onboarding/models: keep skip-auth and provider-scoped model picker prompts off the full global model catalog path, and cache provider catalog hook resolution so setup no longer stalls after auth on large plugin registries. Thanks @shakkernerd. - Gateway/Bonjour: suppress known @homebridge/ciao cancellation and network assertion failures through scoped process handlers so malformed mDNS packets or restricted VPS networking disable/restart Bonjour instead of crashing the gateway. Fixes #67578. Thanks @zenassist26-create. - Discord: keep late clicks on already-resolved exec approval buttons quiet when elevated mode auto-resolved the request, while still surfacing real approval submission failures. Fixes #66906. Thanks @rlerikse. - Agents/subagents: deliver completed yielded-subagent results back to no-thread requester routes via direct fallback when the dormant parent announce turn produces no visible reply, and add QA-lab coverage for the regression. Thanks @vincentkoc. - Gateway/Tailscale: let Tailscale-authenticated Control UI operator sessions with browser device identity skip the device-pairing round trip while still rejecting device-less and node-role connections. Refs #71986. Thanks @jokedul. - Doctor: honor `OPENCLAW_SERVICE_REPAIR_POLICY=external` by reporting gateway service health while skipping service install/start/restart/bootstrap, supervisor rewrites, and legacy service cleanup for externally managed environments. Thanks @shakkernerd. - CLI/update: run package post-update doctor with `--fix` so package updates repair config migrations before restart. Thanks @shakkernerd. - CLI/update: retry failed npm global updates with `--omit=optional` and ignore the superseded first failure when the fallback succeeds. Thanks @shakkernerd. - Plugins/uninstall: migrate and reset `plugins.slots.contextEngine` alongside memory slots when plugin ids change or selected plugins are removed. Thanks @shakkernerd. - Agents/Discord: keep raw `Agent failed before reply` runner failures out of Discord group/channel chats and show detailed runner errors in direct chats only when `/verbose` is enabled. Thanks @codex. - UI/Windows: quote resolved pnpm `.cmd` launcher paths before spawning UI install/build/test commands so Node installs under `C:\Program Files` no longer fail as `C:\Program`. Fixes #45275. Thanks @Kobevictor, @stoppieboy, and @iubns. - Codex/agent: translate `--thinking minimal` to `low` for modern Codex models (gpt-5.5, gpt-5.4, gpt-5.4-mini, gpt-5.2) at request build time so the first turn is accepted instead of paying a wasted call + retry-with-low fallback. Older Codex models still receive `minimal` directly. Fixes #71946. Thanks @hclsys. - Plugins/uninstall: remove tracked plugin files from their recorded managed extensions root even when the current state directory points somewhere else, so `openclaw plugins uninstall --force` does not leave the plugin discoverable. Thanks @shakkernerd. - Agents/runtime: add `agentRuntime.id` as the canonical config key, migrate legacy runtime-policy configs with `openclaw doctor --fix`, route canonical Anthropic models through `claude-cli` without passing CLI backend aliases to embedded harness selection, and load CLI backend owner plugins before channel startup. Fixes #71957. Thanks @WolvenRA. - CLI/update: guard Windows scheduled-task stops by state and timeout so auto-update restart cannot hang indefinitely on `schtasks /End` before stale-listener cleanup. Fixes #69970. Thanks @yangswld and @sherlock-huang. - Windows install/Lobster: execute `pnpm.exe` directly when `npm_execpath` points at the native pnpm binary, add an installed-package fallback for the Lobster embedded runtime, and include the Lobster runner regression test in Windows CI. Fixes #69456. Thanks @igormf. - Gateway/install: refresh loaded gateway service installs when the current service embeds stale gateway auth instead of returning already-installed, avoiding LaunchAgent token-mismatch loops after token rotation. Fixes #70752. Thanks @hyspacex. - Update: ignore bundled plugin `.openclaw-install-stage` directories during global install verification and packaged dist pruning so leftover runtime-dep staging files do not turn successful updates into `unexpected packaged dist file` failures. Fixes #71752. Thanks @waynegault. - CLI/update: fail package updates when post-update plugin sync fails and refresh legacy npm plugin install records before trusting unchanged artifacts, preventing successful updates from restarting with stale or failed plugin state. Thanks @vincentkoc and @shakkernerd. - Release/update: reject pre-populated bundled plugin `.openclaw-install-stage` directories, including mixed-case path variants, before package inventory generation so release tarballs cannot ship poisoned runtime-dependency staging debris. Fixes #71752. Thanks @hclsys. - Node runtime: keep node-host retry timers alive across Gateway restarts and exit on terminal credential pauses so supervised nodes do not become silent zombies. Fixes #69800. Thanks @meroli28. - Gateway/plugins: stop persisted WhatsApp auth state from activating bundled channel runtime-dependency repair during startup when `channels.whatsapp` is absent, avoiding npm/git stalls on packaged Linux installs. Fixes #71994. Thanks @xiao398008. - Gateway/device tokens: enforce caller-scope containment inside token rotation and revocation so pairing-only sessions cannot mutate higher-scope operator tokens. Fixes #71990. Thanks @coygeek. - Plugins/channels: keep security checks, thread-binding placement, provider summaries, health formatting, and message action labels on read-only or already-loaded channel metadata instead of importing full channel runtime. Thanks @shakkernerd. - Plugins/status: keep config-only channel labels and status security summaries from importing plugin runtime modules just to render metadata. Thanks @shakkernerd. - Sessions/channels: stop group-session metadata from loading bundled channel runtime just to classify `#channel` subjects, using only already-loaded channel capabilities on that path. Thanks @shakkernerd. - Plugins/channels: keep native command and native skill `auto` defaults on static channel metadata so config, audit, and command-list checks do not load channel runtime just to read those defaults. Thanks @shakkernerd. - CLI/channels: keep channel remove selection and all-channel capabilities summaries on read-only plugin metadata, loading channel runtime only for the selected mutation path. Thanks @shakkernerd. - CLI/models: keep Provider Index preview rows out of `models list --all --provider <id>` when the owning provider plugin is disabled, preserving config authority for cold catalog fallbacks. Thanks @shakkernerd. - CLI/model runs: keep `openclaw infer model run` on explicit OpenRouter models from loading the full provider catalog or inheriting chat-agent silent-reply policy, restoring non-empty one-shot probe output. Fixes #68791. Thanks @limpredator. - Installer/macOS: rerun Homebrew install steps without the gum spinner when raw-mode ioctl failures occur, and avoid claiming `node@24` was installed when the Homebrew keg binary is missing. Fixes #70411. Thanks @1fanwang and @dad-io. - Installer: load nvm before Node.js detection so `curl | bash` installs respect nvm-managed Node instead of stale system Node. Fixes #49556. Thanks @heavenlxj. - Installer/Windows: route PowerShell install failures through a top-level handler so `iwr ... | iex` returns control to the current shell while direct script-file runs still exit non-zero. Fixes #38054. Thanks @PwrSrg. - CLI/Volta: respawn raw `openclaw` CLI runs through the named `node` shim when the current Node executable resolves to `volta-shim`, avoiding direct shim execution failures in non-interactive shells. Fixes #68672. Thanks @sanchezm86. - Installer: warn when multiple npm global roots contain OpenClaw installs, showing active Node/npm/openclaw plus each install path and version so stale version-manager installs are visible. Fixes #40839. Thanks @zhixianio. - Cron/tasks: recover completed cron task ledger records from durable run logs and job state before marking them `lost`, reducing false `backing session missing` audit errors for isolated cron runs and keeping offline CLI audit from treating its empty local cron active-job set as authoritative. Fixes #71963. - Docker: copy patched dependency files into runtime images so downstream `pnpm install` layers keep working. Fixes #69224. Thanks @gucasbrg. - Package: include patched dependency files in the published npm package so downstream installs can resolve `patchedDependencies`. (#69224) Thanks @gucasbrg and @vincentkoc. - Plugins/channels: treat malformed bundled channel plugin loaders that return `undefined` as unavailable instead of crashing config and help paths. Fixes #69044. Thanks @frankhli843 and @vincentkoc. - Scripts/watch: show corrupted dependency package-config recovery guidance when `gateway:watch` fails during watcher startup, without double-logging unrelated import failures. (#58780) Thanks @roytong9 and @vincentkoc. - Signal: read signal-cli RPC, health checks, and SSE events through Node's HTTP client so Node 24/25 fetch regressions do not break Signal sends or inbound events. Fixes #51716 and #53040. Thanks @Barukimang, @minupla, and @vincentkoc. - Skills/Docker: run npm-backed skill dependency installs with an OpenClaw-managed user prefix so non-root Docker images do not write to `/usr/local`. Fixes #59601. Thanks @chanjarster and @vincentkoc. - Agents/runtime: submit heartbeat, cron, and exec wakeups as transient runtime context instead of visible user prompts, keeping synthetic system work out of chat transcripts. Fixes #66496 and #66814. Thanks @jeades and @mandomaker. - Telegram: include native quote excerpts automatically for threaded replies and reply tags when the original Telegram text is available, without adding another config knob. Fixes #6975. Thanks @rex05ai. - Node/Linux: make `openclaw node install` enable and restart the `openclaw-node` systemd unit instead of the gateway unit on node-only VMs. Fixes #68287. Thanks @dlebee-agent. - Browser/CDP: retry transient raw-CDP WebSocket handshake failures before any browser command is sent, and reconnect stale persistent Playwright CDP sessions for safe tab-list reads without replaying mutating browser actions. Fixes #67728. - Gateway/Linux: retry `systemctl --user enable` after a second daemon reload when the freshly written gateway unit is not visible yet on migrated systemd installs. Fixes #65184. Thanks @liushuaiiu. - Telegram: preserve exact selected quote text when sending native quote replies, and retry with legacy replies if Telegram rejects quote parameters. (#71952) Thanks @rubencu. - Plugins/CLI: preserve manifest name, description, format, and source metadata in cold `openclaw plugins list` output without importing plugin runtime. Thanks @shakkernerd. - Security/audit: read channel exposure and plugin allowlist ownership from read-only plugin index metadata so cold audits do not depend on loaded channel runtime. Thanks @shakkernerd. - Plugins/chat: keep `/plugins list`, `/plugins enable`, and `/plugins disable` on the persisted plugin index path so chat plugin management does not load diagnostic/runtime plugin registries before execution. Thanks @shakkernerd. - Plugins/doctor: read workspace plugin status and legacy web-search ownership through installed-index manifest metadata instead of broad manifest registry scans. Thanks @shakkernerd. - CLI/agents: read channel provider status from read-only plugin index metadata for text `agents list` output instead of the loaded channel registry. Thanks @shakkernerd. - Logging: redact configured secret patterns at console and file-log sink exits so credentials that reach the logger are masked before terminal display or JSONL persistence. Fixes #67953. Thanks @Ziy1-Tan. - Gateway/services: refuse process and service mutations from an older OpenClaw binary when the config was last written by a newer version, preventing split-brain installs from stopping or rewriting newer gateway services. Fixes #57079. - Gateway: reserve `/healthz` and `/readyz` ahead of plugin, canvas, and Control UI HTTP stages so liveness/readiness probes still answer when a later route handler stalls. Fixes #69674. Thanks @Xike-Creek. - Logging: load `logging.file` and redaction settings directly from the active OpenClaw config path in bundled runtimes, so packaged gateways stop falling back to `/tmp/openclaw`. Fixes #59370, #67168, and #61295. Thanks @KeaneYan, @Pan9hu, and @zsjlovelike. - Logging: rotate file logs at `logging.maxFileBytes`, keep bounded numbered archives, and make long-lived rolling loggers follow the current-day file instead of suppressing diagnostics or writing stale dated files. Fixes #58583 and #62381. Thanks @jpeghead and @zhaoleink. - Agents/groups: treat clean empty assistant stops as silent `NO_REPLY` only for always-on groups where silent replies are allowed, while keeping direct and mention-gated sessions on the incomplete-turn retry path. Thanks @MagnaAI. - macOS/Node: keep native remote app nodes from advertising `browser.proxy`, start browser-capable CLI node services through the restored `openclaw node start` command, and show an actionable browser-control error when the local control service is missing. Fixes #66637. - Gateway/update: fail package updates when the restarted managed gateway reports the wrong version, including fallback restarts and JSON mode, avoiding false-success mixed-version restarts after macOS LaunchAgent updates. Fixes #71835. Thanks @abhinas90 and @jsompis. - Gateway/update: warn before package updates and bundled plugin runtime-dependency repairs when the target volume appears low on disk space, without blocking installs on best-effort filesystem checks. Fixes #71835. Thanks @abhinas90 and @jsompis. - Plugins/runtime deps: surface activated plugin load failures in health and fail package-update restart verification or doctor repair when bundled runtime deps still cannot load, avoiding false-success repairs. (#71883) Thanks @Solvely-Colin. - Gateway/Linux: include fnm `aliases/default/bin` in generated service PATHs and let doctor accept either modern fnm aliases or the legacy `current/bin` symlink, avoiding false PATH repair prompts. Fixes #68169. Thanks @richard-scott. - Installer/Linux: run apt installs with noninteractive dpkg and needrestart settings so fresh Ubuntu 24.04 `curl | bash` installs do not hang while installing Node.js, Git, or build tools. Fixes #41146. Thanks @iht76, @alexcarv318, @cs3gallery, @firofame, and @cgdusek. - Providers/Bedrock: defer the AWS SDK import until Bedrock discovery actually runs so plugin registration and setup stay lightweight on cold start. Fixes #71690. Thanks @jarvis-ai-gregmoser. - Installer/macOS: stop immediately when Homebrew `node@24` installation fails and avoid printing PATH advice for missing Homebrew Node installs. Fixes #70411. Thanks @1fanwang. - WhatsApp: remove ack reactions after a visible reply when `messages.removeAckAfterReply` is enabled, matching other reaction-capable channels. Fixes #26183. Thanks @MrUnforsaken. - Providers/Z.AI: map OpenClaw thinking controls to Z.AI's `thinking` payload and add opt-in preserved thinking replay via `params.preserveThinking`, so GLM 5.x can keep prior `reasoning_content` when requested. Fixes #58680. Thanks @xuanmingguo. - Channels/status: keep read-only channel lists on manifest and package metadata by default, loading setup runtime only for explicit fallback callers. Thanks @shakkernerd. - Plugins: scope setup and web-provider metadata manifest reads to explicit plugin ids when callers already know the owning plugin set. Thanks @vincentkoc. - Plugins/onboarding: defer onboarding install-record index writes until the guarded config commit so setup failures cannot leave the plugin index ahead of `openclaw.json`. Thanks @shakkernerd. - Plugins/registry: resolve web provider ownership from the installed plugin index instead of broad manifest scans on secret, tool, and pricing paths. Thanks @shakkernerd. - Config/providers: accept `video` and `audio` in configured model `input` values and preserve them in provider catalog entries. Fixes #20721. Thanks @alvinttang. - Models/auth: honor the parent `--agent` flag for auth write commands (`add`, `login`, `setup-token`, `paste-token`, and the GitHub Copilot shortcut) so OAuth/API-key/token results are written to the requested agent store instead of the default agent. Fixes #71864. (#71933) Thanks @balric-seo. - TTS: strip model-emitted TTS directives from streamed block text before channel delivery, including directives split across adjacent blocks, while preserving the accumulated raw reply for final-mode synthesis. Fixes #38937. - TTS: keep explicit `provider=...` directive keys scoped to that provider and warn on unsupported keys instead of letting another speech provider consume overlapping keys. Fixes #60131. - TTS/Feishu: normalize final-mode streamed TTS-only audio before delivery so generated voice-note files use the same safe media path and native voice routing as normal final replies. Fixes #71920. - Feishu: transcribe inbound voice-note audio with the shared media audio path before agent dispatch and keep raw Feishu `file_key` payloads out of message text. Fixes #67120 and #61876. - Tasks: terminalize async Gateway agent task records from the Gateway run result while preserving aborted, failed, and cancelled outcomes instead of leaving completed runs stuck as active or lost. (#71905) Thanks @likewen-tech. - WhatsApp: let authorized group voice-note transcripts satisfy mention gating before reply dispatch, while keeping unmentioned transcripts in pending group history. Fixes #44908. - Media understanding: carry channel voice-note preflight state into attachment selection so WhatsApp, Feishu, Telegram, and Discord do not transcribe the same inbound audio twice. Fixes #70580. - TTS/BlueBubbles: deliver compatible auto-TTS audio as iMessage voice memo bubbles instead of plain MP3/CAF file attachments. Fixes #16848. - TTS: resolve voice-note and voice-memo routing from channel plugin capabilities instead of speech-core-owned channel id lists. - ACP: send subagent and async-task completion wakes to external ACP harnesses as plain prompts instead of OpenClaw internal runtime-context envelopes, while keeping those envelopes out of ACP transcripts. - TTS/status: show configured TTS model, voice, and sanitized custom endpoint in `/status`, preserve OpenAI-compatible TTS instructions on custom endpoints, and retry empty Microsoft/Edge TTS output once. Addresses #46602, #47232, and #43936. Thanks @leekuangtao, @Huntterxx, and @rex993. - Agents/Gateway: steer agent-driven config edits and restarts through the owner-only `gateway` tool, document `config.schema.lookup` as the field-doc source, and warn against using `gateway stop && gateway start` as a restart substitute on macOS. Fixes #71929. Thanks @ygc3817922006-sketch. - Media understanding/audio: inject a deterministic transcript placeholder for too-small voice notes so agents do not hallucinate transcription or provider failures. Fixes #48944. Thanks @eulicesl. - Providers/vLLM: send Nemotron 3 chat-template kwargs when thinking is off and honor configured `params.chat_template_kwargs` for OpenAI-compatible completions, so vLLM/Nemotron replies stay visible instead of becoming thinking-only. Fixes #71891. Thanks @jmystaki-create and @dennis-lynch. - Channels/replies: strip copied inbound metadata blocks from user-facing assistant replies and model replay history, so Discord/vLLM sessions do not leak `Conversation info` / `UNTRUSTED ... message body` envelopes after a model echoes them. Fixes #71847. Thanks @jmystaki-create. - Subagents/memory: keep inter-session completion wakes out of memory and dreaming session exports, and strip internal runtime-context blocks from realtime Control UI chat events. - Agents/Claude: treat zero-token empty `stop` turns as failed provider output, retry once, repair replay, and allow configured model fallback instead of preserving them as successful silent replies. Fixes #71880. Thanks @MagnaAI. - Tasks: normalize task lifecycle timestamps at create, update, and restore time, and report retained lost tasks as audit warnings until their cleanup window expires. (#71871) Thanks @likewen-tech. - Diagnostics/OTEL: treat normal early model stream cleanup as a completed model call instead of exporting a misleading `StreamAbandoned` error span. Thanks @vincentkoc. - Gateway/pairing: stop corrupt or unreadable device/node pairing stores from being treated as empty state, preserving `paired.json` for repair instead of overwriting approved pairings. Fixes #71873. Thanks @iret77. - ACP: keep `/acp` management commands, plus local `/status` and `/unfocus`, on the Gateway path inside ACP-bound threads so they are not consumed as ACP prompt text. Fixes #66298. Thanks @kindomLee. - ACPX: stop probing ACP agents during normal Gateway startup; the embedded backend now registers without spawning Codex/ACP child processes unless `OPENCLAW_ACPX_RUNTIME_STARTUP_PROBE=1` is explicitly set. - CLI/image edit: accept `--size`, `--aspect-ratio`, and `--resolution` on `openclaw infer image edit` and report all supported edit flags from `capability inspect image.edit`. Thanks @Pinghuachiu. - ACP: wait for the configured runtime backend to become healthy before startup identity reconciliation, avoiding transient acpx warnings during Gateway boot. Fixes #40566. - Channels/ACP bindings: time out configured binding readiness checks instead of letting Discord preflight hang forever when an ACP target never settles. Fixes #68776. - Control UI: hide the chat loading skeleton during background history reloads when existing messages or active stream content are already visible, avoiding reload flashes on high-latency local gateways. Fixes #71844. Thanks @WolvenRA. - Control UI: keep locally optimistic chat messages visible when a history reload temporarily returns empty, avoiding lost first-turn messages on high-latency gateways. Fixes #71878. Thanks @WolvenRA. - Control UI: keep chat history limits based on visible messages after filtering heartbeat and control-only transcript rows, so recent hidden entries no longer make older visible replies disappear. Thanks @WolvenRA. - Agents/images: scrub old `[media attached: ...]`, `[Image: source: ...]`, and `media://inbound/...` markers from pruned model replay context so stale media refs are not rehydrated as fresh prompt images. Fixes #71868. Thanks @jmeadlock. - Docker/Bonjour: disable Bonjour/mDNS advertising by default for bundled Compose gateways on bridge networking, while keeping host/macvlan opt-in with `OPENCLAW_DISABLE_BONJOUR=0`. Fixes #71879. Thanks @gbballpack. - CLI/status: label the OpenClaw Serve/Funnel setting as `Tailscale exposure` and show daemon state separately when available, so `gateway.tailscale.mode: "off"` no longer reads like the Tailscale daemon is stopped. Fixes #71790. Thanks @pesvobodak. - Plugins/Bonjour: stop ciao mDNS watchdog failures from looping forever when the advertiser stays stuck in `probing` or `announcing`; Bonjour now disables itself for the current Gateway process after repeated failed restarts while the Gateway keeps running. Fixes #69011. Thanks @siddharthaagarwalofficial-ux, @FiredMosquito831, and @spikefcz. - Gateway/Fly.io: seed Control UI allowed origins from the actual runtime bind and port so CLI-driven non-loopback starts do not crash before config exists. Fixes #71823. - macOS/remote SSH: keep discovered gateway hosts in `gateway.remote.sshTarget` while pinning SSH transport URLs to the local loopback tunnel, so browser automation does not regress into blocked non-loopback `ws://` endpoints. Fixes #67336. - Gateway/proxy: bootstrap env proxy dispatching from direct Gateway startup so provider and plugin network requests honor `HTTPS_PROXY`/`HTTP_PROXY` before the first embedded agent attempt runs. (#71833) Thanks @mjamiv. - Plugins/runtime deps: verify clean npm installs actually place requested bundled runtime packages in the managed install root, reporting exact missing specs instead of a false successful repair. (#71883) Thanks @Solvely-Colin. - Plugins/discovery: ignore stale `plugins.load.paths` aliases that point back at packaged bundled plugin directories and have doctor remove them, keeping bundled plugins on the runtime-deps staging path. Thanks @codex. - Models/LM Studio: preserve `@iq*` quant suffixes in model refs and provider matching so `/model lmstudio/...@iq3_xxs` keeps the exact LM Studio variant. Fixes #71474. (#71486) Thanks @Bartok9, @XinwuC, and @Sanjays2402. - Matrix/cron: preserve the live Matrix delivery target when creating implicit announce reminder jobs so mixed-case room IDs are not reconstructed from lowercased session keys. Fixes #71798. - Feishu: accept Schema 2.0 card action callbacks that report `context.open_chat_id` instead of legacy `context.chat_id`, so button callbacks no longer drop as malformed. Fixes #71670. Thanks @eddy1068. - Feishu: keep synthetic card-action and bot-menu ids out of platform reply targets, using the real card callback message id when Feishu provides one and plain-sending otherwise. Fixes #71673. Thanks @eddy1068. - Plugins/QQ Bot: prefer an installed QQ Bot plugin that declares it replaces the bundled `qqbot` channel, preventing duplicate `qqbot_channel_api` and `qqbot_remind` tool registration noise. Fixes #63102. - Browser automation: keep stable tab ids and labels attached when Chromium replaces the raw target after form submissions or other action-triggered navigations, and return the replacement `targetId` from `/act` when the match is provable. Fixes #46137. - QQ Bot: make `qqbot_remind` schedule, list, and remove Gateway cron jobs directly for owner-authorized senders instead of returning `cronParams` and relying on a follow-up generic `cron` tool call. Fixes #70865. (#70937) Thanks @GaosCode. - Agents/ACP: hide `sessions_spawn` ACP runtime options unless an ACP backend is loaded, and make `/acp doctor` call out `plugins.allow` blocking bundled `acpx`. Thanks @vincentkoc. - Agents/Codex: keep ACP prompt/skill routing hidden unless an ACP runtime backend is available, and warn in doctor when enabled Codex plugin configs still route `openai-codex/*` models through PI. Thanks @vincentkoc. - Media delivery: avoid sending generated image attachments twice when the assistant reply already includes explicit `MEDIA:` lines for the same turn, and reject unsafe remote `MEDIA:` URLs before delivery. Thanks @pashpashpash. - Codex harness: ignore retryable app-server error notifications after Codex recovers, and preserve the real nested error message for terminal app-server failures instead of replacing it with a generic failure. Thanks @pashpashpash. - Agents/Codex: prepare native Codex sub-agent session metadata without a nested Gateway session patch and add a focused Docker smoke for the app-server sub-agent path. Thanks @vincentkoc. - Agents/subagents: keep queued subagent announces session-only when the requester has no external channel target, avoiding ambiguous multi-channel delivery failures. Fixes #59201. Thanks @larrylhollan. - Image understanding: preserve configured provider-prefixed vision model metadata when callers request the model without the provider prefix, so custom image models keep their `input: ["text", "image"]` capability. Fixes #33185. Thanks @Kobe9312 and @vincentkoc. - Plugins/install: restore the previous plugin index records if a concurrent config write conflict interrupts install, update, or uninstall metadata commits. Thanks @shakkernerd. - Plugins/install: reject native plugin archives that do not include a valid `openclaw.plugin.json`, preventing manifestless archives from writing install records that later show missing-manifest diagnostics. Thanks @shakkernerd. - Plugins/uninstall: remove tracked managed plugin install directories even when the persisted install path differs from the default id-derived target, while still refusing deletes outside the managed extensions root. Thanks @shakkernerd. - Plugins/update: restore previous plugin index records if core update or channel setup hits a concurrent config write conflict after plugin metadata changes. Thanks @shakkernerd. - Plugins/onboarding: defer channel/provider plugin install records until the owning config write commits, keeping setup failures from advancing the plugin index ahead of `openclaw.json`. Thanks @shakkernerd. - Plugins/config: route configure and agent setup writes with pending plugin install records through the plugin index commit helper so provider onboarding metadata is not stripped by plain config writes. Thanks @shakkernerd. - Plugins/channels: merge pending channel plugin install records with the existing plugin index before config writes, preserving unrelated tracked installs during channel setup, resolve, remove, and capability repair flows. Thanks @shakkernerd. - Plugins/config: defer shipped `plugins.installs` index migration during config writes until the guarded config commit window and roll it back if the config write fails before commit. Thanks @shakkernerd. - Sessions: keep embedded runtime context out of the visible user prompt by sending it as a hidden next-turn custom message, and teach doctor to repair affected 2026.4.24 transcripts with duplicated prompt-rewrite branches. Fixes #71761. - Gateway/subagents: keep direct-loopback backend RPCs authenticated with the shared gateway token/password off stale CLI paired-device scope baselines, so internal calls no longer hit `scope-upgrade` pairing prompts while remote, browser, node, device-token, and explicit-device paths still require normal pairing approval. Fixes #63548. - Providers/Azure OpenAI: give deployment-scoped image generation requests a longer 600s default timeout so slow `gpt-image-2` generations can complete without a per-call `timeoutMs`. Fixes #71705. Thanks @voytas75. - Gateway/plugins: link source-checkout bundled runtime dependency caches instead of recursively copying `node_modules` on the gateway main thread, preventing local status, node, and skill probes from timing out during startup cache restores. - Skills/remote nodes: only expose remote macOS skill bins for connected nodes, clear stale bin matches when node probes fail, and include probe command, timeout, bin count, and connection state in timeout logs. - Skills/remote nodes: recognize `system.which` object-map responses when probing connected macOS nodes, so Linux gateways can expose macOS-only skills such as Apple Notes when the required binaries are installed remotely. Fixes #71877. Thanks @miguelarios. - CLI/gateway: keep diagnostic probes from creating first-time read-only device pairings, while still reusing cached device tokens for detailed read probes. Fixes #71766. Thanks @SunboZ. - CLI/plugins: keep `message` startup, `channels logs`, `agents delete`, and `agents set-identity` off broad plugin preloading; message delivery still loads plugins when the action actually runs. - Image understanding: resolve configured image models such as local LM Studio vision entries before reporting `Unknown model` when the discovery registry has not registered that provider. Fixes #66486. Thanks @zhanggpcsu. - QQ Bot: ignore self-echoed bot messages using the outbound ref-index marker, preventing mirrored replies from re-entering the agent loop while still allowing users to quote bot replies. Fixes #71912. Thanks @wangyc6003. - Sessions: separate reset freshness from session-store `updatedAt`, so heartbeat, cron, exec, and gateway bookkeeping no longer prevent configured daily/idle resets from rolling long-running channel sessions. Fixes #68315, #63732, #63820, and #69083. Thanks @maxatv, @longhairedsi, @bradfreels, and @akessel56. - Sessions: clear queued system-event notices during `/new`, `/reset`, gateway `sessions.reset`, and daily/idle rollover so stale background updates cannot leak into the first prompt of the fresh session. Fixes #66864. Thanks @opeyio, @Magicray1217, and @cedillarack. - CLI/agents: keep `agents bind`, `agents unbind`, and `agents bindings` on setup-safe channel metadata paths so they do not preload bundled plugin runtimes or stage runtime dependencies. Fixes #71743. - Plugins/registry: preserve explicit disabled plugin records during registry migration without persisting every unused bundled plugin discovered on disk. Thanks @shakkernerd. - Windows/native: keep CLI startup and bundled provider plugin loading off Windows ESM raw-path failure paths, fixing native onboarding/install smoke on Node 24. - Plugins/doctor: read bundled channel doctor capabilities through the same packaged plugin directory resolver used by plugin loading, so published installs keep Matrix DM allowlist repairs on `channels.matrix.dm.*` instead of writing invalid top-level `dmPolicy` keys. Fixes #71757. - Plugins/Windows: keep bundled plugin Jiti loaders off the native import path on Windows so channel plugins such as Telegram no longer crash with `ERR_UNSUPPORTED_ESM_URL_SCHEME` on `C:\...` paths. Fixes #71749. Thanks @smeyer9. - Providers/Ollama: use Ollama's current `/api/web_search` endpoint and honor `https://ollama.com` model-provider base URLs for Ollama Web Search. Fixes #71741. Thanks @madhvidua. - Memory/Ollama: serialize Ollama memory embedding batches and add an inline batch timeout override, with longer defaults for local/self-hosted embedding providers. - Sessions/usage: exclude compaction checkpoint transcript snapshots from usage totals and session discovery, while keeping old checkpoint files removable. - CLI/agents: keep `openclaw agents list --json` on the config-only path by default, avoiding bundled plugin loading unless callers request `--bindings`. Fixes #71739. Thanks @kaloster. - Plugins/install: force plugin dependency installs to stay project-local even when inherited npm config requests global installs, so successful installs still materialize the plugin's staged `node_modules`. - Providers/Google: transcode Gemini TTS PCM to Opus for voice-note targets so WhatsApp and other native voice-note replies can play as voice messages. - TTS/WhatsApp: mark non-Opus provider output as voice-note intent so channel delivery transcodes MP3/WebM replies to Ogg/Opus PTT audio. - Plugins/runtime deps: reuse existing external bundled-plugin stage roots when mirrored plugin roots are inspected again, avoiding second-generation `openclaw-unknown-*` stages and repeated first-turn restaging. Fixes #71599. - iOS/macOS Talk Mode: allow `talk.speechLocale` to set the speech recognition locale for non-English voice conversations. Fixes #44688. - Plugins/providers: honor explicit plugin candidate lists instead of reading a persisted registry snapshot from local state, keeping candidate-scoped provider discovery hermetic. - Plugins/doctor: keep bundled plugin runtime-dependency repairs inside the managed OpenClaw stage even when user npm prefix/global config points npm at `$HOME/node_modules`. Fixes #71730. - ACP/sessions_spawn: reject normal OpenClaw config agent ids when callers explicitly request `runtime="acp"`, while allowing agents configured with `runtime.type="acp"` to resolve to their ACP harness id. Fixes #63914. - ACP/sessions_spawn: apply `runTimeoutSeconds` to ACP child turns and dispatch those turns on the background subagent lane, so quota-stalled ACP harnesses do not occupy the main agent lane indefinitely. Fixes #68823. - ACP/oneshot: reconcile runtime session identity before closing completed oneshot ACP runs, so finished `sessions.json` entries do not stay stuck with `acp.identity.state="pending"`. - ACPX: bundle `[email protected]` so unsupported generic model overrides fail clearly instead of silently falling back to the target adapter default. - ACP/models: document that non-Codex ACP model overrides require adapter support for ACP `models` plus `session/set_model`, so unsupported harnesses fail clearly instead of silently falling back to their defaults. - Plugins/Voice Call: treat missing provider credentials as setup-incomplete during Gateway startup and log the missing keys as a warning instead of a runtime startup error, while keeping explicit command/tool errors when used. - Android/Talk Mode: prevent duplicate TTS playback when fast or repeated final chat events arrive while Talk Mode is waiting for its own response. Fixes #46546. - Tooling/check:changed: pass parent heavy-check lock markers to lint lanes so `pnpm check:changed` no longer waits on its own `lint:extensions` child. - CLI/completion: dedupe provider auth flags before registering `openclaw onboard` options, so completion-cache refresh during update no longer fails when stale core fallback flags overlap plugin manifest flags. Fixes #71667. - Diagnostics/trace: report live context usage from the current prompt snapshot instead of provider turn totals, avoiding false near-full context spikes on cached or tool-heavy runs. - Providers/Google: honor `models.providers.google.request.allowPrivateNetwork` for Gemini TTS and telephony TTS, matching Google image generation and media understanding. (#71723) Thanks @ro-hansolo. - Providers/MiniMax: register `minimax-portal` for music and video generation, preserving OAuth auth and regional MiniMax base URLs across the shared `music_generate` and `video_generate` tools. (#63241) Thanks @tars90percent. - Providers/onboarding: keep Runway and Alibaba Model Studio out of the text-inference setup picker by scoping their video-generation auth choices to the media setup flow. (#65856) Thanks @Jah-yee. - Plugins/Bonjour: stop the gateway from crash-looping on `CIAO PROBING CANCELLED` when the mDNS watchdog cancels a stuck probe. Restores the rejection-handler wiring dropped during the bonjour plugin migration and shares unhandled-rejection state across module instances so plugin-staged copies of `openclaw/plugin-sdk/runtime` register into the same handler set the host consults. Especially affects Docker on macOS, where mDNS probing reliably hits the watchdog. Thanks @troyhitch. - Google Meet: report pinned Chrome nodes as offline or missing capabilities in setup/join diagnostics, keep inaccessible nodes out of auto-selection, and preflight local BlackHole/SoX requirements before agents try local Chrome. - Providers/MiniMax: route `image-01` requests to the dedicated image generation endpoint while preserving CN endpoint selection. Fixes #61149. Thanks @mushuiyu886. - Plugins/startup: remove ownerless bundled runtime-dependency install locks after a short grace window and include lock owner details when startup times out waiting for a plugin runtime-deps lock. - Plugins/install: anchor bundled runtime-dependency npm installs with an OpenClaw-owned package manifest so Linux updates cannot accidentally write to a parent `$HOME/node_modules` tree. Fixes #71730. - Plugins/install: pass onboarding plugin config into plugin index writes so local plugin installs outside default discovery roots keep their install records. Thanks @shakkernerd. - Plugins/install: migrate shipped `plugins.installs` config records into the plugin index while stripping them from runtime config and future writes. Thanks @shakkernerd. - Plugins/install: durably remove shipped `plugins.installs` from `openclaw.json` after its records are copied into the plugin index, while rolling back the index write if config cleanup fails. Thanks @shakkernerd. - Plugins/install: keep migrated plugin install records in the plugin index even when the plugin manifest is missing or invalid, so update, uninstall, inspect, and audit can still recover broken installs. Thanks @shakkernerd. - Plugins/security: keep plugin audit JSON check ids stable while reporting plugin index install-record findings with updated wording. Thanks @shakkernerd. - CLI/config: reject direct `plugins.installs` edits with guidance to use `openclaw plugins install`, `openclaw plugins update`, or `openclaw plugins uninstall` instead. Thanks @shakkernerd. - Live tests/voice: accept common STT variants for OpenClaw and ElevenLabs brand names so provider smoke tests fail on real regressions rather than equivalent transcripts. - Agents/replies: forward sanitized underlying agent failure details on external channels instead of replacing unknown failures with a generic retry message. - CLI/MCP: translate OpenClaw `mcp.servers.*.transport` entries into Claude/Gemini CLI `type` fields so streamable HTTP MCP servers load in CLI backend sessions. (#71724) Thanks @Blockchain-Oracle. - Browser/CDP: honor configured remote and `attachOnly` CDP HTTP/WebSocket timeouts when opening tabs through raw CDP or `/json/new` fallback. (#54238) Thanks @FuncWei. - WhatsApp/TTS: send visible text separately from PTT voice-note audio instead of relying on hidden voice-note captions. Fixes #51081. - Browser/client: avoid telling agents to restart OpenClaw for dispatcher timeouts on external browser profiles such as `attachOnly`, remote CDP, and existing-session. (#40815) Thanks @0xsline. - Agents/TTS: preserve `[[audio_as_voice]]` directives on trusted text tool-result `MEDIA:` payloads so generated audio still delivers as a voice note. (#46535) Thanks @azade-c. - Agents/TTS: keep queued tool media when an assistant ends with `NO_REPLY` on non-block delivery paths, so media-only generated audio replies still send. (#60025) Thanks @bradlind1. - Telegram/STT: frame inbound voice-note transcripts as machine-generated, untrusted text in agent context while preserving raw transcript mention detection. Closes #33360. Thanks @smartchainark. - Subagents/browser: show an actionable `/tools` notice when browser automation is configured but filtered out by the active tool profile, and document that coding-profile agents should use `tools.alsoAllow: ["browser"]` rather than subagent allowlists alone. - Control UI/Quick Settings: persist the assistant avatar override to browser local storage (mirroring the user avatar) so uploaded image data URLs no longer fail config validation with "Too big: expected string to have <=200 characters". Also lift the gateway-side `ui.assistant.avatar` length cap to match the user avatar size budget for non-UI clients writing the field directly. Thanks @BunsDev. - Plugin SDK: share diagnostic event subscriptions across duplicate source/dist module graphs so legacy root SDK imports still receive runtime diagnostic events. - Agents/Bedrock: prevent empty assistant stream-error turns from poisoning Converse replay by persisting, repairing, and replaying a non-empty fallback block. Fixes #71572. (#71627) Thanks @openperf. - Agents/Anthropic/Bedrock: strip thinking blocks with missing, empty, or blank replay signatures before provider conversion, falling back to non-empty omitted-reasoning text when needed so corrupted signed-thinking history no longer poisons subsequent turns. Fixes #45010. (#70054) Thanks @castaples. - Agents/Anthropic/Bedrock: preserve stripped thinking-only assistant replay turns with non-empty omitted-reasoning text so provider adapters keep strict user/assistant turn shape. Thanks @wujiaming88. - ACP/Codex: pass `sessions_spawn(runtime="acp")` model and thinking overrides into Codex ACP startup, normalize `openai-codex/*` refs and slash reasoning suffixes, and recognize managed Codex ACP wrapper commands without blocking current `gpt-5.5` sessions. Fixes #40393. (#71643) Thanks @91wan. - Browser/CDP: make readiness diagnostics use the same discovery-first fallback as reachability for bare `ws://` Browserless and Browserbase CDP URLs. Fixes #69532. - Browser/CDP: explain that loopback Browserless or other externally managed CDP services need `attachOnly: true` and matching Browserless `EXTERNAL` endpoint when reporting local port ownership conflicts, and fall back to the configured bare WebSocket root when a discovered Browserless endpoint rejects CDP. Fixes #49815. - Gateway/reload: preserve indefinite `gateway.reload.deferralTimeoutMs: 0` semantics for channel hot reload deferrals so active agent runs are not interrupted by a forced channel restart. (#71637) Thanks @Poo-Squirry. - Agents/tool results: cap persisted Pi tool-result details and strip hidden diagnostics before provider conversion, preventing large debug payloads from bloating session transcripts. (#71637) Thanks @Poo-Squirry. - ACP/OpenCode: update the bundled acpx runtime to 0.6.0 and cover the OpenCode ACP bind path in Docker live tests. - Providers/OpenCode Go: add DeepSeek V4 Pro and DeepSeek V4 Flash to the Go catalog while the bundled Pi registry catches up. Fixes #71587. - Providers/OpenCode Go: route DeepSeek V4 Pro/Flash through the OpenAI-compatible Go endpoint and suppress invalid `reasoning_effort: "off"` payloads, fixing tool-enabled requests for `opencode-go/deepseek-v4-flash`. Fixes #71683. - Plugins/model defaults: run Skill Workshop review, Active Memory recall, and session-memory slug generation on the configured agent default model instead of the hardcoded OpenAI SDK fallback when hook context lacks model metadata. Fixes #71659. - Providers/Venice: fill the required DeepSeek V4 `reasoning_content` placeholder for `venice/deepseek-v4-pro` and `venice/deepseek-v4-flash` replay turns without sending native DeepSeek `thinking` controls that Venice rejects. Fixes #71628. - Browser/existing-session: support per-profile Chrome MCP command/args, map `cdpUrl` to `--browserUrl` or `--wsEndpoint`, and avoid combining endpoint flags with `--userDataDir`. Fixes #47879, #48037, and #62706. Thanks @puneet1409, @zhehao, and @madkow1001. - Media/plugins: bound MIME sniffing and ZIP archive preflight before handing untrusted files to `file-type` or `jszip`, reducing parser CPU and memory exposure for attachments and ClawHub plugin archives. Thanks @vincentkoc. - Memory-host SDK: use trusted env-proxy mode for remote embedding and batch HTTP calls only when Undici will proxy that target, preserving SSRF DNS pinning for `ALL_PROXY`-only and `NO_PROXY` bypass cases. Fixes #52162. (#71506) Thanks @DhtIsCoding. - Gateway/dashboard: render Control UI and WebSocket links with `https://`/`wss://` when `gateway.tls.enabled=true`, including `openclaw gateway status`. Fixes #71494. (#71499) Thanks @deepkilo. - Agents/OpenAI-compatible: default proxy/local completions tool requests to `tool_choice: "auto"` when tools are present, so providers enter native tool-calling mode instead of replying with plain-text tool directives. (#71472) Thanks @Speed-maker. - OpenAI image generation: use `gpt-5.5` for the Codex OAuth responses transport instead of the retired `gpt-5.4` model, fixing 500s from ChatGPT Codex image generation. Fixes #71513. Thanks @baolongl. - OpenAI image generation: route transparent-background default-model requests to `gpt-image-1.5`, document the expected `image_generate` call shape, and keep Azure/custom OpenAI-compatible deployment names untouched. - Google video generation: download direct MLDev Veo `video.uri` results instead of passing them through the Files API path, fixing 404s after successful generation/polling. Fixes #71200. Thanks @panhaishan. - Google video generation: fall back to the REST `predictLongRunning` Veo endpoint for text-only SDK 404s while keeping reference image/video generation on the SDK path. Fixes #62309 and #63008. (#62343) Thanks @leoleedev. - MiniMax music generation: switch the bundled default model from the unsupported `music-2.5+` id to the current `music-2.6` API model. Fixes #64870 and addresses the music default from #62315. Thanks @noahclanman and @edwardzheng1. - Cron: record jobs interrupted by a gateway restart as failed at their original `runningAtMs`, skip unsafe startup replay, and disable interrupted one-shot jobs so they show a visible failure instead of silently disappearing or duplicating work. Fixes #59056, #61343, #63657, and #59301. Thanks @ponchoooPenguin, @daemic24, @myradon, and @hikiwibot. - Cron tool: recover flat top-level schedule shorthand such as `cron`, `tz`, and `staggerMs` before gateway validation, so model-generated cron add/update calls preserve cron jitter settings. Thanks @tyxben. - Cron: hydrate flat legacy job rows with top-level `cron`, `tz`, `session`, and `message` fields into canonical schedule, target, and payload objects before startup recomputes run times. Fixes #43351. - Agents/replies: let pending group chat history trigger bare mentioned turns without treating metadata-only inbound context as user input. Fixes #71489. (#71520) Thanks @SymbolStar. - Google media generation: strip a configured trailing `/v1beta` from Google music/video provider base URLs before calling the Google GenAI SDK, preventing doubled `/v1beta/v1beta` paths. Fixes #63240. (#63258) Thanks @Hybirdss. - Discord: restore direct-message voice-note preflight transcription and classify URL-only Ogg/Opus voice attachments as audio while skipping partial attachments without usable URLs. Fixes #61314 and #64803. - Plugins/build: copy bundled plugin skill trees into `dist-runtime`, broaden Windows symlink-copy fallbacks, and fingerprint runtime dependencies from `lstat` so symlink-like directory entries cannot crash staging. - Google Chat: preserve reply text when a typing indicator message is deleted or can no longer be updated, so media captions and first text chunks are resent instead of silently disappearing. (#71498) Thanks @colin-lgtm. - Cron: tolerate malformed legacy job rows in startup, main-session system-event payloads, and human-readable `cron list` output so missing `state`, `payload.text`, or display fields no longer crash the scheduler or CLI. Fixes #66016, #65916, #64137, #57872, #59968, #63813, #52804, and #43163. (#71509) Thanks @vincentkoc. - CLI/models: make `openclaw models scan` fall back to public OpenRouter free-model metadata when no `OPENROUTER_API_KEY` is configured, avoid config secret resolution for explicit `--no-probe` scans, and apply the scan timeout to the OpenRouter catalog request. - Feishu: keep streaming cards to one live card per turn, flush throttled card edits after meaningful text boundaries, and skip exact block/partial repeats so tool-heavy replies do not duplicate card output. Thanks @allan0509. - Feishu: finish the streaming-card duplicate closeout by stripping leaked reasoning tags, preserving cross-block partial snapshots, enabling topic-thread streaming cards, omitting the generic `main` card header, surfacing transient tool/compaction status, and cleaning streaming state after close failures. Thanks @sesame437, @Vicky-v7, @maoku-family, @Pengxiao-Wang, and @Maple778. - Telegram: recover incomplete partial-stream previews by falling back to a final send when an ambiguous final edit failure would otherwise retain a strict prefix of the answer. Fixes #71525. (#71554) Thanks @sahilsatralkar. - Control UI/chat: collapse assistant token/model context details behind an explicit Context disclosure and show full dates in message footers, making historical transcript timing clear without noisy default metadata. (#71337) Thanks @BunsDev. - OpenAI/Codex OAuth: explain `unsupported_country_region_territory` token-exchange failures with a proxy/region hint instead of surfacing a generic OAuth error. Fixes #51175. (#71501) Thanks @vincentkoc and @wulala-xjj. - Browser/Linux: fall back to headless mode for local managed profiles on hosts without a display server, while preserving explicit per-profile headed overrides and reporting the headless source. (#60953) Thanks @rrpsantos. - Telegram: remove the startup persisted-offset `getUpdates` preflight so polling restarts do not self-conflict before the runner starts. Fixes #69304. (#69779) Thanks @chinar-amrutkar. - Telegram: keep the polling stall watchdog active even when grammY reports the runner as not running while its task is still pending, so a rebuilt transport cannot leave `getUpdates` silent until a manual gateway restart. Fixes #69064. Thanks @LDLoeb. - Subagents: fall back to direct completion delivery when the parent announce turn finishes without a visible payload, so child results still reach channel-backed requester sessions. - Subagents: tell parent agents to use `sessions_yield` while waiting for child completion events, preventing GPT-5 fast runs from ending silently after spawning workers. - Browser/Playwright: ignore benign already-handled route races during guarded navigation so browser-page tasks no longer fail when Playwright tears down a route mid-flight. (#68708) Thanks @Steady-ai. - Browser/CLI: lazy-load browser command groups and plugin runtime services so `openclaw browser --help` can render without loading the full browser automation stack. Fixes #65400. (#65460, #66640) Thanks @pandego and @Tianworld. - Browser/CLI: serve precomputed `openclaw browser --help` text from CLI startup metadata, avoiding the full plugin/config startup path for the common help invocation. - Browser/downloads: seed managed Chrome profiles with OpenClaw download prefs and capture unmanaged click-triggered downloads under the guarded downloads directory, while explicit download waiters still own their target file. (#64558) Thanks @Pearcekieser. - Browser/Chrome: stop passing redundant `--disable-setuid-sandbox` when `browser.noSandbox` is enabled; `--no-sandbox` remains the effective sandbox opt-out. (#67939) Thanks @sebykrueger. - Browser/client: stop telling agents to permanently avoid the browser after transient timeout or cancellation failures; keep the no-retry hint for persistent unavailable/rate-limit cases. (#46505) Thanks @jriff. - Browser/aria snapshots: bind `format=aria` `axN` refs to live DOM nodes through backend DOM ids when Playwright is available, so follow-up browser actions can use those refs without timing out. (#62434) Thanks @MrKipler. - Telegram: prevent duplicate in-process long pollers for the same bot token and add clearer `getUpdates` conflict diagnostics for external duplicate pollers. Fixes #56230. Thanks @Co-Messi. - Browser/Linux: detect Chromium-based installs under `/opt/google`, `/opt/brave.com`, `/usr/lib/chromium`, and `/usr/lib/chromium-browser` before asking users to set `browser.executablePath`. (#48563) Thanks @lupuletic. - Sessions/browser: close tracked browser tabs when idle, daily, `/new`, or `/reset` session rollover archives the previous transcript, preventing tabs from leaking past the old session. Thanks @jakozloski. - Sessions/forking: fall back to transcript-estimated parent token counts when cached totals are stale or missing, so oversized thread forks start fresh instead of cloning the full parent transcript. Thanks @jalehman. - OpenAI/Codex: send Codex Responses system prompts through top-level `instructions` while preserving the existing native Codex payload controls. - MCP/CLI: retire bundled MCP runtimes at the end of one-shot `openclaw agent` and `openclaw infer model run` gateway/local executions, so repeated scripted runs do not accumulate stdio MCP child processes. Fixes #71457. Thanks @spartoviMD. - OpenAI/Codex image generation: canonicalize legacy `openai-codex.baseUrl` values such as `https://chatgpt.com/backend-api` to the Codex Responses backend before calling `gpt-image-2`, matching the chat transport. Fixes #71460. Thanks @GodsBoy. - Control UI: make `/usage` use the fresh context snapshot for context percentage, and include cache-write tokens in the Usage overview cache-hit denominator. Fixes #47885. Thanks @imwyvern and @Ante042. - GitHub Copilot: preserve encrypted Responses reasoning item IDs during replay so Copilot can validate encrypted reasoning payloads across requests. (#71448) Thanks @a410979729-sys. - GitHub Copilot: never rewrite connection-bound reasoning item IDs regardless of whether `encrypted_content` is present, fixing a 400 "Encrypted content item_id did not match" error with `gpt-5.3-codex` and future Codex models that fall through to the forward-compat catch-all with `reasoning: false`. Also recognize Codex-named models as reasoning-capable so they inherit the correct capability flags. Refs #68735. Thanks @InvalidPandaa. - Agents/replies: recover final-answer text when streamed assistant chunks contain only whitespace, preventing completed turns from surfacing as empty-payload errors. Fixes #71454. (#71467) Thanks @Sanjays2402. - Feishu/TTS: transcode voice-intent MP3 and other audio replies to Ogg/Opus before sending native Feishu audio bubbles, while keeping ordinary MP3 attachments as files. Fixes #61249 and #37868. Thanks @sg1416-zg and @ycjlb2023-peteryi. - WhatsApp/TTS: transcode MP3/WebM audio, including Microsoft Edge TTS output, to Ogg/Opus before sending PTT voice notes. - QQBot/TTS: honor plain `audioAsVoice` replies by synthesizing TTS to native QQ voice messages, and mark inbound voice-only messages as audio media without exposing raw voice paths to generic media context. - Providers/SenseAudio: add bundled SenseAudio batch audio transcription through `tools.media.audio` with `SENSEAUDIO_API_KEY` auth. (#66943) Thanks @Fl0rencess720. - Providers/MiniMax: let TTS use MiniMax portal OAuth and Token Plan credentials before falling back to `MINIMAX_API_KEY`, and include current TTS HD model ids. Fixes #55017. Thanks @zx15210404690-hash. - Telegram/webhook: acknowledge validated webhook updates before running bot middleware, keeping slow agent turns from tripping Telegram delivery retries while preserving per-chat processing lanes. Fixes #71392. Thanks @joelforsberg46-source. - MCP/config reload: hot-apply `mcp.*` changes by disposing cached session MCP runtimes, and dispose bundled MCP runtimes during gateway shutdown so removed `mcp.servers` entries reap child processes promptly. Fixes #60656. Thanks @xieyuanqing. - Active Memory: keep silent recall sub-agent billing/auth failures out of shared auth-profile cooldown state, so a Claude CLI extra-usage rejection cannot disable normal Claude-backed turns. Fixes #71284. (#71539) Thanks @vishutdhar and @obviyus. - Auth/Claude CLI: sync refreshed Claude CLI OAuth credentials into the managed auth profile so long-running Claude CLI runs stop falling back to stale OpenClaw snapshots. (#70902) Thanks @starvex. - Sessions: make `sessions_spawn(mode="session")` errors name usable alternatives when the current channel cannot bind subagent threads. Fixes #67400. (#67790) Thanks @stainlu. - Agents/Claude CLI: pass the OpenClaw system prompt through Claude's prompt-file flag so Windows runs avoid argv length failures without changing system prompt semantics. Fixes #69158. (#69211) Thanks @skylee-01, @cassioanorte, @Syu0, and @Stache73. - Agents/CLI sessions: bind `google-gemini-cli` session auth-epoch to the Google account identity in `~/.gemini/oauth_creds.json`, so Gemini-backed agents resume their conversation after gateway restart instead of minting a fresh session, and stale bindings are invalidated when the authenticated Google account changes. Fixes #70973. (#71076) Thanks @openperf. - Slack: stop treating user mentions in assistant-authored message edit blocks as sender attribution, preventing edited bot messages from spoofing a mentioned DM user. (#71700) Thanks @vincentkoc. - Codex: consume unauthorized bound conversation inbound claims before they can fall through to other claim handlers or enqueue Codex turns. (#71702) Thanks @vincentkoc. - Codex media understanding: require approval-checked app-server image turns while explicitly declining tool, file, permission, and elicitation approval requests for the bounded image worker. (#71703) Thanks @vincentkoc. - Agents/Claude CLI: allow large live `stream-json` JSONL lines up to the existing per-turn raw limit, preventing large Telegram, WebChat, MCP, and image turns from aborting on the old stdout buffer cap. Fixes #71793, #71080, and #70766. (#71897) Thanks @chacher86, @shivamgrover21, and @tpjordan. - Agents/Claude CLI: unwrap nested Claude result envelopes in CLI JSON output so delegated agent responses surface as final text instead of raw result JSON. (#66819) Thanks @mraleko. - Agents/Claude CLI: apply the configured 1M context window override to eligible Claude CLI Opus and Sonnet models when `context1m` is enabled. (#70863) Thanks @bidadh. - Models/status: report fresh Claude CLI native auth instead of stale stored `anthropic:claude-cli` profile expiry when local credentials are current. Fixes #71256. (#71332) Thanks @matthiasjanke and @neeravmakwana. - CLI backends: compact OpenClaw transcripts after over-budget CLI turns and reseed fresh CLI sessions from the compacted transcript instead of stale external resume state. Fixes #68329. (#71916) Thanks @obviyus. - Telegram: keep default tool progress messages visible when answer preview streaming is disabled. (#71825) Thanks @VACInc. - Configure/models: clear deselected model fallbacks when updating the model picker allowlist, including provider-scoped setup flows. (#71596) Thanks @rubencu. - Agents/streaming: strip namespaced `<antml:thinking>` reasoning tags from streamed assistant replies before user-visible text is emitted. (#69288) Thanks @xialonglee.