OpenClaw v2026.2.21

### Changes - Models/Google: add Gemini 3.1 support (`google/gemini-3.1-pro-preview`). - Providers/Onboarding: add Volcano Engine (Doubao) and BytePlus providers/models (including coding variants), wire onboarding auth choices for interactive + non-interactive flows, and align docs to `volcengine-api-key`. (#7967) Thanks @funmore123. - Channels/CLI: add per-account/channel `defaultTo` outbound routing fallback so `openclaw agent --deliver` can send without explicit `--reply-to` when a default target is configured. (#16985) Thanks @KirillShchetinin. - Channels: allow per-channel model overrides via `channels.modelByChannel` and note them in /status. Thanks @thewilloftheshadow. - Telegram/Streaming: simplify preview streaming config to `channels.telegram.streaming` (boolean), auto-map legacy `streamMode` values, and remove block-vs-partial preview branching. (#22012) thanks @obviyus. - Discord/Streaming: add stream preview mode for live draft replies with partial/block options and configurable chunking. Thanks @thewilloftheshadow. Inspiration @neoagentic-ship-it. - Discord/Telegram: add configurable lifecycle status reactions for queued/thinking/tool/done/error phases with a shared controller and emoji/timing overrides. Thanks @wolly-tundracube and @thewilloftheshadow. - Discord/Voice: add voice channel join/leave/status via `/vc`, plus auto-join configuration for realtime voice conversations. Thanks @thewilloftheshadow. - Discord: add configurable ephemeral defaults for slash-command responses. (#16563) Thanks @wei. - Discord: support updating forum `available_tags` via channel edit actions for forum tag management. (#12070) Thanks @xiaoyaner0201. - Discord: include channel topics in trusted inbound metadata on new sessions. Thanks @thewilloftheshadow. - Discord/Subagents: add thread-bound subagent sessions on Discord with per-thread focus/list controls and thread-bound continuation routing for spawned helper agents. (#21805) Thanks @onutc. - iOS/Chat: clean chat UI noise by stripping inbound untrusted metadata/timestamp prefixes, formatting tool outputs into concise summaries/errors, compacting the composer while typing, and supporting tap-to-dismiss keyboard in chat view. (#22122) thanks @mbelinky. - iOS/Watch: bridge mirrored watch prompt notification actions into iOS quick-reply handling, including queued action handoff until app model initialization. (#22123) thanks @mbelinky. - iOS/Gateway: stabilize background wake and reconnect behavior with background reconnect suppression/lease windows, BGAppRefresh wake fallback, location wake hook throttling, and APNs wake retry+nudge instrumentation. (#21226) thanks @mbelinky. - Auto-reply/UI: add model fallback lifecycle visibility in verbose logs, /status active-model context with fallback reason, and cohesive WebUI fallback indicators. (#20704) Thanks @joshavant. - MSTeams: dedupe sent-message cache storage by removing duplicate per-message Set storage and using timestamps Map keys as the single membership source. (#22514) Thanks @TaKO8Ki. - Agents/Subagents: default subagent spawn depth now uses shared `maxSpawnDepth=2`, enabling depth-1 orchestrator spawning by default while keeping depth policy checks consistent across spawn and prompt paths. (#22223) Thanks @tyler6204. - Security/Agents: make owner-ID obfuscation use a dedicated HMAC secret from configuration (`ownerDisplaySecret`) and update hashing behavior so obfuscation is decoupled from gateway token handling for improved control. (#7343) Thanks @vincentkoc. - Security/Infra: switch gateway lock and tool-call synthetic IDs from SHA-1 to SHA-256 with unchanged truncation length to strengthen hash basis while keeping deterministic behavior and lock key format. (#7343) Thanks @vincentkoc. - Dependencies/Tooling: add non-blocking dead-code scans in CI via Knip/ts-prune/ts-unused-exports to surface unused dependencies and exports earlier. (#22468) Thanks @vincentkoc. - Dependencies/Unused Dependencies: remove or scope unused root and extension deps (`@larksuiteoapi/node-sdk`, `signal-utils`, `ollama`, `lit`, `@lit/context`, `@lit-labs/signals`, `@microsoft/agents-hosting-express`, `@microsoft/agents-hosting-extensions-teams`, and plugin-local `openclaw` devDeps in `extensions/open-prose`, `extensions/lobster`, and `extensions/llm-task`). (#22471, #22495) Thanks @vincentkoc. - Dependencies/A2UI: harden dependency resolution after root cleanup (resolve `lit`, `@lit/context`, `@lit-labs/signals`, and `signal-utils` from workspace/root) and simplify bundling fallback behavior, including `pnpm dlx rolldown` compatibility. (#22481, #22507) Thanks @vincentkoc. ### Fixes - Security/Agents: cap embedded Pi runner outer retry loop with a higher profile-aware dynamic limit (32-160 attempts) and return an explicit `retry_limit` error payload when retries never converge, preventing unbounded internal retry cycles (`GHSA-76m6-pj3w-v7mf`). - Telegram: detect duplicate bot-token ownership across Telegram accounts at startup/status time, mark secondary accounts as not configured with an explicit fix message, and block duplicate account startup before polling to avoid endless `getUpdates` conflict loops. - Agents/Tool images: include source filenames in `agents/tool-images` resize logs so compression events can be traced back to specific files. - Providers/OAuth: harden Qwen and Chutes refresh handling by validating refresh response expiry values and preserving prior refresh tokens when providers return empty refresh token fields, with regression coverage for empty-token responses. - Models/Kimi-Coding: add missing implicit provider template for `kimi-coding` with correct `anthropic-messages` API type and base URL, fixing 403 errors when using Kimi for Coding. (#22409) - Auto-reply/Tools: forward `senderIsOwner` through embedded queued/followup runner params so owner-only tools remain available for authorized senders. (#22296) thanks @hcoj. - Discord: restore model picker back navigation when a provider is missing and document the Discord picker flow. (#21458) Thanks @pejmanjohn and @thewilloftheshadow. - Memory/QMD: respect per-agent `memorySearch.enabled=false` during gateway QMD startup initialization, split multi-collection QMD searches into per-collection queries (`search`/`vsearch`/`query`) to avoid sparse-term drops, prefer collection-hinted doc resolution to avoid stale-hash collisions, retry boot updates on transient lock/timeout failures, skip `qmd embed` in BM25-only `search` mode (including `memory index --force`), and serialize embed runs globally with failure backoff to prevent CPU storms on multi-agent hosts. (#20581, #21590, #20513, #20001, #21266, #21583, #20346, #19493) Thanks @danielrevivo, @zanderkrause, @sunyan034-cmd, @tilleulenspiegel, @dae-oss, @adamlongcreativellc, @jonathanadams96, and @kiliansitel. - Memory/Builtin: prevent automatic sync races with manager shutdown by skipping post-close sync starts and waiting for in-flight sync before closing SQLite, so `onSearch`/`onSessionStart` no longer fail with `database is not open` in ephemeral CLI flows. (#20556, #7464) Thanks @FuzzyTG and @henrybottter. - Providers/Copilot: drop persisted assistant `thinking` blocks for Claude models (while preserving turn structure/tool blocks) so follow-up requests no longer fail on invalid `thinkingSignature` payloads. (#19459) Thanks @jackheuberger. - Providers/Copilot: add `claude-sonnet-4.6` and `claude-sonnet-4.5` to the default GitHub Copilot model catalog and add coverage for model-list/definition helpers. (#20270, fixes #20091) Thanks @Clawborn. - Auto-reply/WebChat: avoid defaulting inbound runtime channel labels to unrelated providers (for example `whatsapp`) for webchat sessions so channel-specific formatting guidance stays accurate. (#21534) Thanks @lbo728. - Status: include persisted `cacheRead`/`cacheWrite` in session summaries so compact `/status` output consistently shows cache hit percentages from real session data. - Heartbeat/Cron: restore interval heartbeat behavior so missing `HEARTBEAT.md` no longer suppresses runs (only effectively empty files skip), preserving prompt-driven and tagged-cron execution paths. - WhatsApp/Cron/Heartbeat: enforce allowlisted routing for implicit scheduled/system delivery by merging pairing-store + configured `allowFrom` recipients, selecting authorized recipients when last-route context points to a non-allowlisted chat, and preventing heartbeat fan-out to recent unauthorized chats. - Heartbeat/Active hours: constrain active-hours `24` sentinel parsing to `24:00` in time validation so invalid values like `24:30` are rejected early. (#21410) thanks @adhitShet. - Heartbeat: treat `activeHours` windows with identical `start`/`end` times as zero-width (always outside the window) instead of always-active. (#21408) thanks @adhitShet. - CLI/Pairing: default `pairing list` and `pairing approve` to the sole available pairing channel when omitted, so TUI-only setups can recover from `pairing required` without guessing channel arguments. (#21527) Thanks @losts1. - TUI/Pairing: show explicit pairing-required recovery guidance after gateway disconnects that return `pairing required`, including approval steps to unblock quickstart TUI hatching on fresh installs. (#21841) Thanks @nicolinux. - TUI/Input: suppress duplicate backspace events arriving in the same input burst window so SSH sessions no longer delete two characters per backspace press in the composer. (#19318) Thanks @eheimer. - TUI/Heartbeat: suppress heartbeat ACK/prompt noise in chat streaming when `showOk` is disabled, while still preserving non-ACK heartbeat alerts in final output. (#20228) Thanks @bhalliburton. - TUI/History: cap chat-log component growth and prune stale render nodes/references so large default history loads no longer overflow render recursion with `RangeError: Maximum call stack size exceeded`. (#18068) Thanks @JaniJegoroff. - Memory/QMD: diversify mixed-source search ranking when both session and memory collections are present so session transcript hits no longer crowd out durable memory-file matches in top results. (#19913) Thanks @alextempr. - Memory/Tools: return explicit `unavailable` warnings/actions from `memory_search` when embedding/provider failures occur (including quota exhaustion), so disabled memory does not look like an empty recall result. (#21894) Thanks @XBS9. - Session/Startup: require the `/new` and `/reset` greeting path to run Session Startup file-reading instructions before responding, so daily memory startup context is not skipped on fresh-session greetings. (#22338) Thanks @armstrong-pv. - Auth/Onboarding: align OAuth profile-id config mapping with stored credential IDs for OpenAI Codex and Chutes flows, preventing `provider:default` mismatches when OAuth returns email-scoped credentials. (#12692) thanks @mudrii. - Provider/HTTP: treat HTTP 503 as failover-eligible for LLM provider errors. (#21086) Thanks @Protocol-zero-0. - Slack: pass `recipient_team_id` / `recipient_user_id` through Slack native streaming calls so `chat.startStream`/`appendStream`/`stopStream` work reliably across DMs and Slack Connect setups, and disable block streaming when native streaming is active. (#20988) Thanks @Dithilli. Earlier recipient-ID groundwork was contributed in #20377 by @AsserAl1012. - CLI/Config: add canonical `--strict-json` parsing for `config set` and keep `--json` as a legacy alias to reduce help/behavior drift. (#21332) thanks @adhitShet. - CLI: keep `openclaw -v` as a root-only version alias so subcommand `-v, --verbose` flags (for example ACP/hooks/skills) are no longer intercepted globally. (#21303) thanks @adhitShet. - Memory: return empty snippets when `memory_get`/QMD read files that have not been created yet, and harden memory indexing/session helpers against ENOENT races so missing Markdown no longer crashes tools. (#20680) Thanks @pahdo. - Telegram/Streaming: always clean up draft previews even when dispatch throws before fallback handling, preventing orphaned preview messages during failed runs. (#19041) thanks @mudrii. - Telegram/Streaming: split reasoning and answer draft preview lanes to prevent cross-lane overwrites, and ignore literal `<think>` tags inside inline/fenced code snippets so sample markup is not misrouted as reasoning. (#20774) Thanks @obviyus. - Telegram/Streaming: restore 30-char first-preview debounce and scope `NO_REPLY` prefix suppression to partial sentinel fragments so normal `No...` text is not filtered. (#22613) thanks @obviyus. - Telegram/Status reactions: refresh stall timers on repeated phase updates and honor ack-reaction scope when lifecycle reactions are enabled, preventing false stall emojis and unwanted group reactions. Thanks @wolly-tundracube and @thewilloftheshadow. - Telegram/Status reactions: keep lifecycle reactions active when available-reactions lookup fails by falling back to unrestricted variant selection instead of suppressing reaction updates. (#22380) thanks @obviyus. - Discord/Streaming: apply `replyToMode: first` only to the first Discord chunk so block-streamed replies do not spam mention pings. (#20726) Thanks @thewilloftheshadow for the report. - Discord/Components: map DM channel targets back to user-scoped component sessions so button/select interactions stay in the main DM session. Thanks @thewilloftheshadow. - Discord/Allowlist: lazy-load guild lists when resolving Discord user allowlists so ID-only entries resolve even if guild fetch fails. (#20208) Thanks @zhangjunmengyang. - Discord/Gateway: handle close code 4014 (missing privileged gateway intents) without crashing the gateway. Thanks @thewilloftheshadow. - Discord: ingest inbound stickers as media so sticker-only messages and forwarded stickers are visible to agents. Thanks @thewilloftheshadow. - Auto-reply/Runner: emit `onAgentRunStart` only after agent lifecycle or tool activity begins (and only once per run), so fallback preflight errors no longer mark runs as started. (#21165) Thanks @shakkernerd. - Auto-reply/Tool results: serialize tool-result delivery and keep the delivery chain progressing after individual failures so concurrent tool outputs preserve user-visible ordering. (#21231) thanks @ahdernasr. - Auto-reply/Prompt caching: restore prefix-cache stability by keeping inbound system metadata session-stable and moving per-message IDs (`message_id`, `message_id_full`, `reply_to_id`, `sender_id`) into untrusted conversation context. (#20597) Thanks @anisoptera. - iOS/Watch: add actionable watch approval/reject controls and quick-reply actions so watch-originated approvals and responses can be sent directly from notification flows. (#21996) Thanks @mbelinky. - iOS/Watch: refresh iOS and watch app icon assets with the lobster icon set to keep phone/watch branding aligned. (#21997) Thanks @mbelinky. - CLI/Onboarding: fix Anthropic-compatible custom provider verification by normalizing base URLs to avoid duplicate `/v1` paths during setup checks. (#21336) Thanks @17jmumford. - iOS/Gateway/Tools: prefer uniquely connected node matches when duplicate display names exist, surface actionable `nodes invoke` pairing-required guidance with request IDs, and refresh active iOS gateway registration after location-capability setting changes so capability updates apply immediately. (#22120) thanks @mbelinky. - Gateway/Auth: require `gateway.trustedProxies` to include a loopback proxy address when `auth.mode="trusted-proxy"` and `bind="loopback"`, preventing same-host proxy misconfiguration from silently blocking auth. (#22082, follow-up to #20097) thanks @mbelinky. - Gateway/Auth: allow trusted-proxy mode with loopback bind for same-host reverse-proxy deployments, while still requiring configured `gateway.trustedProxies`. (#20097) thanks @xinhuagu. - Gateway/Auth: allow authenticated clients across roles/scopes to call `health` while preserving role and scope enforcement for non-health methods. (#19699) thanks @Nachx639. - Gateway/Hooks: include transform export name in hook-transform cache keys so distinct exports from the same module do not reuse the wrong cached transform function. (#13855) thanks @mcaxtr. - Gateway/Control UI: return 404 for missing static-asset paths instead of serving SPA fallback HTML, while preserving client-route fallback behavior for extensionless and non-asset dotted paths. (#12060) thanks @mcaxtr. - Gateway/Pairing: prevent device-token rotate scope escalation by enforcing an approved-scope baseline, preserving approved scopes across metadata updates, and rejecting rotate requests that exceed approved role scope implications. (#20703) thanks @coygeek. - Gateway/Pairing: clear persisted paired-device state when the gateway client closes with `device token mismatch` (`1008`) so reconnect flows can cleanly re-enter pairing. (#22071) Thanks @mbelinky. - Gateway/Config: allow `gateway.customBindHost` in strict config validation when `gateway.bind="custom"` so valid custom bind-host configurations no longer fail startup. (#20318, fixes #20289) Thanks @MisterGuy420. - Gateway/Pairing: tolerate legacy paired devices missing `roles`/`scopes` metadata in websocket upgrade checks and backfill metadata on reconnect. (#21447, fixes #21236) Thanks @joshavant. - Gateway/Pairing/CLI: align read-scope compatibility in pairing/device-token checks and add local `openclaw devices` fallback recovery for loopback `pairing required` deadlocks, with explicit fallback notice to unblock approval bootstrap flows. (#21616) Thanks @shakkernerd. - Cron: honor `cron.maxConcurrentRuns` in the timer loop so due jobs can execute up to the configured parallelism instead of always running serially. (#11595) Thanks @Takhoffman. - Agents/Compaction: restore embedded compaction safeguard/context-pruning extension loading in production by wiring bundled extension factories into the resource loader instead of runtime file-path resolution. (#22349) Thanks @Glucksberg. - Agents/Subagents: restore announce-chain delivery to agent injection, defer nested announce output until descendant follow-up content is ready, and prevent descendant deferrals from consuming announce retry budget so deep chains do not drop final completions. (#22223) Thanks @tyler6204. - Agents/System Prompt: label allowlisted senders as authorized senders to avoid implying ownership. Thanks @thewilloftheshadow. - Agents/Tool display: fix exec cwd suffix inference so `pushd ... && popd ... && <command>` does not keep stale `(in <dir>)` context in summaries. (#21925) Thanks @Lukavyi. - Tools/web_search: handle xAI Responses API payloads that emit top-level `output_text` blocks (without a `message` wrapper) so Grok web_search no longer returns `No response` for those results. (#20508) Thanks @echoVic. - Agents/Failover: treat non-default override runs as direct fallback-to-configured-primary (skip configured fallback chain), normalize default-model detection for provider casing/whitespace, and add regression coverage for override/auth error paths. (#18820) Thanks @Glucksberg. - Docker/Build: include `ownerDisplay` in `CommandsSchema` object-level defaults so Docker `pnpm build` no longer fails with `TS2769` during plugin SDK d.ts generation. (#22558) Thanks @obviyus. - Docker/Browser: install Playwright Chromium into `/home/node/.cache/ms-playwright` and set `node:node` ownership so browser binaries are available to the runtime user in browser-enabled images. (#22585) thanks @obviyus. - Hooks/Session memory: trigger bundled `session-memory` persistence on both `/new` and `/reset` so reset flows no longer skip markdown transcript capture before archival. (#21382) Thanks @mofesolapaul. - Dependencies/Agents: bump embedded Pi SDK packages (`@mariozechner/pi-agent-core`, `@mariozechner/pi-ai`, `@mariozechner/pi-coding-agent`, `@mariozechner/pi-tui`) to `0.54.0`. (#21578) Thanks @Takhoffman. - Config/Agents: expose Pi compaction tuning values `agents.defaults.compaction.reserveTokens` and `agents.defaults.compaction.keepRecentTokens` in config schema/types and apply them in embedded Pi runner settings overrides with floor enforcement via `reserveTokensFloor`. (#21568) Thanks @Takhoffman. - Docker: pin base images to SHA256 digests in Docker builds to prevent mutable tag drift. (#7734) Thanks @coygeek. - Docker: run build steps as the `node` user and use `COPY --chown` to avoid recursive ownership changes, trimming image size and layer churn. Thanks @huntharo. - Config/Memory: restore schema help/label metadata for hybrid `mmr` and `temporalDecay` settings so configuration surfaces show correct names and guidance. (#18786) Thanks @rodrigouroz. - Skills/SonosCLI: add troubleshooting guidance for `sonos discover` failures on macOS direct mode (`sendto: no route to host`) and sandbox network restrictions (`bind: operation not permitted`). (#21316) Thanks @huntharo. - macOS/Build: default release packaging to `BUNDLE_ID=ai.openclaw.mac` in `scripts/package-mac-dist.sh`, so Sparkle feed URL is retained and auto-update no longer fails with an empty appcast feed. (#19750) thanks @loganprit. - Signal/Outbound: preserve case for Base64 group IDs during outbound target normalization so cross-context routing and policy checks no longer break when group IDs include uppercase characters. (#5578) Thanks @heyhudson. - Anthropic/Agents: preserve required pi-ai default OAuth beta headers when `context1m` injects `anthropic-beta`, preventing 401 auth failures for `sk-ant-oat-*` tokens. (#19789, fixes #19769) Thanks @minupla. - Security/Exec: block unquoted heredoc body expansion tokens in shell allowlist analysis, reject unterminated heredocs, and require explicit approval for allowlisted heredoc execution on gateway hosts to prevent heredoc substitution allowlist bypass. Thanks @torturado for reporting. - macOS/Security: evaluate `system.run` allowlists per shell segment in macOS node runtime and companion exec host (including chained shell operators), fail closed on shell/process substitution parsing, and require explicit approval on unsafe parse cases to prevent allowlist bypass via `rawCommand` chaining. Thanks @tdjackey for reporting. - WhatsApp/Security: enforce allowlist JID authorization for reaction actions so authenticated callers cannot target non-allowlisted chats by forging `chatJid` + valid `messageId` pairs. Thanks @aether-ai-agent for reporting. - ACP/Security: escape control and delimiter characters in ACP `resource_link` title/URI metadata before prompt interpolation to prevent metadata-driven prompt injection through resource links. Thanks @aether-ai-agent for reporting. - TTS/Security: make model-driven provider switching opt-in by default (`messages.tts.modelOverrides.allowProvider=false` unless explicitly enabled), while keeping voice/style overrides available, to reduce prompt-injection-driven provider hops and unexpected TTS cost escalation. Thanks @aether-ai-agent for reporting. - Security/Agents: keep overflow compaction retry budgeting global across tool-result truncation recovery so successful truncation cannot reset the overflow retry counter and amplify retry/cost cycles. Thanks @aether-ai-agent for reporting. - BlueBubbles/Security: require webhook token authentication for all BlueBubbles webhook requests (including loopback/proxied setups), removing passwordless webhook fallback behavior. Thanks @zpbrent. - iOS/Security: force `https://` for non-loopback manual gateway hosts during iOS onboarding to block insecure remote transport URLs. (#21969) Thanks @mbelinky. - Gateway/Security: remove shared-IP fallback for canvas endpoints and require token or session capability for canvas access. Thanks @thewilloftheshadow. - Gateway/Security: require secure context and paired-device checks for Control UI auth even when `gateway.controlUi.allowInsecureAuth` is set, and align audit messaging with the hardened behavior. (#20684) Thanks @coygeek and @Vasco0x4 for reporting. - Gateway/Security: scope tokenless Tailscale forwarded-header auth to Control UI websocket auth only, so HTTP gateway routes still require token/password even on trusted hosts. Thanks @zpbrent for reporting. - Docker/Security: run E2E and install-sh test images as non-root by adding appuser directives. Thanks @thewilloftheshadow. - Skills/Security: sanitize skill env overrides to block unsafe runtime injection variables and only allow sensitive keys when declared in skill metadata, with warnings for suspicious values. Thanks @thewilloftheshadow. - Security/Commands: block prototype-key injection in runtime `/debug` overrides and require own-property checks for gated command flags (`bash`, `config`, `debug`) so inherited prototype values cannot enable privileged commands. Thanks @tdjackey for reporting. - Security/Browser: block non-network browser navigation protocols (including `file:`, `data:`, and `javascript:`) while preserving `about:blank`, preventing local file reads via browser tool navigation. Thanks @q1uf3ng for reporting. - Security/Exec: block shell startup-file env injection (`BASH_ENV`, `ENV`, `BASH_FUNC_*`, `LD_*`, `DYLD_*`) across config env ingestion, node-host inherited environment sanitization, and macOS exec host runtime to prevent pre-command execution from attacker-controlled environment variables. Thanks @tdjackey. - Security/Exec (Windows): canonicalize `cmd.exe /c` command text across validation, approval binding, and audit/event rendering to prevent trailing-argument approval mismatches in `system.run`. Thanks @tdjackey for reporting. - Security/Gateway/Hooks: block `__proto__`, `constructor`, and `prototype` traversal in webhook template path resolution to prevent prototype-chain payload data leakage in `messageTemplate` rendering. (#22213) Thanks @SleuthCo. - Security/OpenClawKit/UI: prevent injected inbound user context metadata blocks from leaking into chat history in TUI, webchat, and macOS surfaces by stripping all untrusted metadata prefixes at display boundaries. (#22142) Thanks @Mellowambience, @vincentkoc. - Security/OpenClawKit/UI: strip inbound metadata blocks from user messages in TUI rendering while preserving user-authored content. (#22345) Thanks @kansodata, @vincentkoc. - Security/OpenClawKit/UI: prevent inbound metadata leaks and reply-tag streaming artifacts in TUI rendering by stripping untrusted metadata prefixes at display boundaries. (#22346) Thanks @akramcodez, @vincentkoc. - Security/Agents: restrict local MEDIA tool attachments to core tools and the OpenClaw temp root to prevent untrusted MCP tool file exfiltration. Thanks @NucleiAv and @thewilloftheshadow. - Security/Net: strip sensitive headers (`Authorization`, `Proxy-Authorization`, `Cookie`, `Cookie2`) on cross-origin redirects in `fetchWithSsrFGuard` to prevent credential forwarding across origin boundaries. (#20313) Thanks @afurm. - Security/Systemd: reject CR/LF in systemd unit environment values and fix argument escaping so generated units cannot be injected with extra directives. Thanks @thewilloftheshadow. - Security/Tools: add per-wrapper random IDs to untrusted-content markers from `wrapExternalContent`/`wrapWebContent`, preventing marker spoofing from escaping content boundaries. (#19009) Thanks @Whoaa512. - Shared/Security: reject insecure deep links that use `ws://` non-loopback gateway URLs to prevent plaintext remote websocket configuration. (#21970) Thanks @mbelinky. - macOS/Security: reject non-loopback `ws://` remote gateway URLs in macOS remote config to block insecure plaintext websocket endpoints. (#21971) Thanks @mbelinky. - Browser/Security: block upload path symlink escapes so browser upload sources cannot traverse outside the allowed workspace via symlinked paths. (#21972) Thanks @mbelinky. - Security/Dependencies: bump transitive `hono` usage to `4.11.10` to incorporate timing-safe authentication comparison hardening for `basicAuth`/`bearerAuth` (`GHSA-gq3j-xvxp-8hrf`). Thanks @vincentkoc. - Security/Gateway: parse `X-Forwarded-For` with trust-preserving semantics when requests come from configured trusted proxies, preventing proxy-chain spoofing from influencing client IP classification and rate-limit identity. Thanks @AnthonyDiSanti and @vincentkoc. - Security/Sandbox: remove default `--no-sandbox` for the browser container entrypoint, add explicit opt-in via `OPENCLAW_BROWSER_NO_SANDBOX` / `CLAWDBOT_BROWSER_NO_SANDBOX`, and add security-audit checks for stale/missing sandbox browser Docker hash labels. Thanks @TerminalsandCoffee and @vincentkoc. - Security/Sandbox Browser: require VNC password auth for noVNC observer sessions in the sandbox browser entrypoint, plumb per-container noVNC passwords from runtime, and emit short-lived noVNC observer token URLs while keeping loopback-only host port publishing. Thanks @TerminalsandCoffee for reporting. - Security/Sandbox Browser: default browser sandbox containers to a dedicated Docker network (`openclaw-sandbox-browser`), add optional CDP ingress source-range restrictions, auto-create missing dedicated networks, and warn in `openclaw security --audit` when browser sandboxing runs on bridge without source-range limits. Thanks @TerminalsandCoffee for reporting. > Note: npm shipped a follow-up update fix as `[email protected]`.