OpenClaw v2026.4.26

## 2026.4.26 ### Changes - Control UI/Talk: add a generic browser realtime transport contract, Google Live browser Talk sessions with constrained ephemeral tokens, and a Gateway relay for backend-only realtime voice plugins. Thanks @VACInc. - CLI/models: route provider-filtered model listing through an explicit source plan so user config, installed manifest rows, Provider Index previews, and scoped runtime fallbacks keep a stable authority order without adding another catalog cache. Thanks @shakkernerd. - Providers: add Cerebras as a bundled plugin with onboarding, static model catalog, docs, and manifest-owned endpoint metadata. - Memory/OpenAI-compatible: add optional `memorySearch.inputType`, `queryInputType`, and `documentInputType` config for asymmetric embedding endpoints, including direct query embeddings and provider batch indexing. Carries forward #63313 and #60727. Thanks @HOYALIM and @prospect1314521. - Ollama/memory: add model-specific retrieval query prefixes for `nomic-embed-text`, `qwen3-embedding`, and `mxbai-embed-large` memory-search queries while leaving document batches unchanged. Carries forward #45013. Thanks @laolin5564. - Plugins/providers: move pre-runtime model-id normalization, provider endpoint host metadata, and OpenAI-compatible request-family hints into plugin manifests so core no longer carries bundled-provider routing tables. - Plugins/config: deprecate direct plugin config load/write helpers in favor of passed runtime snapshots plus transactional mutation helpers with explicit restart follow-up policy, scanner guardrails, runtime warnings, and revision-based cache invalidation. - Plugins/install: allow `OPENCLAW_PLUGIN_STAGE_DIR` to contain layered runtime-dependency roots, resolving read-only preinstalled deps before installing missing deps into the final writable root. Fixes #72396. Thanks @liorb-mountapps. - Control UI: add a raw config pending-changes diff panel that parses JSON5, redacts sensitive values until reveal, and avoids fake raw-edit callbacks when opening the panel. Refs #39831; supersedes #48621 and #46654. Thanks @JiajunBernoulli and @BunsDev. - Control UI: polish the quick settings dashboard grid so common cards align across desktop, tablet, and mobile layouts without wasting horizontal space. Thanks @BunsDev. - Matrix/E2EE: add `openclaw matrix encryption setup` to enable Matrix encryption, bootstrap recovery, and print verification status from one setup flow. Thanks @gumadeiras. - Agents/compaction: add an opt-in `agents.defaults.compaction.maxActiveTranscriptBytes` preflight trigger that runs normal local compaction when the active JSONL grows too large, requiring transcript rotation so successful compaction moves future turns onto a smaller successor file instead of raw byte-splitting history. Thanks @vincentkoc. - CLI/migration: add a bundled Claude importer that previews and applies Claude Code and Claude Desktop instructions, MCP servers, skills, command prompts, and safe archive/manual-review state. Thanks @vincentkoc. - CLI/migration: add `openclaw migrate` with plan, dry-run, JSON, pre-migration backup, onboarding detection, archive-only report copies, and a bundled Hermes importer for configuration, memory/plugin hints, model providers, MCP servers, skills, and supported credentials. Thanks @NousResearch. ### Fixes - Gateway/device tokens: stop echoing rotated bearer tokens from shared/admin `device.token.rotate` responses while preserving the same-device token handoff needed by token-only clients before reconnect. (#66773) Thanks @MoerAI. - Agents/sessions_spawn: resolve configured bare model aliases for spawn model overrides using the target agent runtime default provider, carrying forward the alias-specific #69029 review fixes from #59681 without the unrelated active-session pruning path. Fixes #59681. Thanks @HowdyDooToYou. - Control UI/Talk: keep Google Live browser sessions on the WebSocket transport instead of falling back to WebRTC, validate browser Google Live WebSocket endpoints, cap Gateway relay sessions per browser connection, and remove stale browser-native voice buttons that did not use the configured Talk/TTS provider. Thanks @BunsDev. - Gateway/startup: reuse config snapshot plugin manifests for startup auto-enable before plugin bootstrap plans plugin loading. Thanks @shakkernerd. - Agents/subagents: enforce `subagents.allowAgents` for explicit same-agent `sessions_spawn(agentId=...)` calls instead of auto-allowing requester self-targets. Fixes #72827. Thanks @oiGaDio. - ACP/sessions_spawn: let explicit `sessions_spawn(runtime="acp")` bootstrap turns run while `acp.dispatch.enabled=false` still blocks automatic ACP thread dispatch. Fixes #63591. Thanks @moeedahmed. - CLI/update: install npm global updates into a verified temporary prefix before swapping the package tree into place, preventing mixed old/new installs and stale packaged files from breaking `openclaw update` verification. Thanks @shakkernerd. - Gateway: skip CLI startup self-respawn for foreground gateway runs so low-memory Linux/Node 24 hosts start through the same path as direct `dist/index.js` without hanging before logs. Fixes #72720. Thanks @sign-2025. - Google Meet: grant Meet media permissions through browser control and pin local Chrome audio defaults to `BlackHole 2ch`, so joined agents no longer show `Permission needed` or use macOS default audio devices. Thanks @DougButdorf. - Gateway: treat uncaught broken-pipe stream errors like `EPIPE` as non-fatal so Discord delivery or closed pipes no longer crash the Gateway after a reply is ready. - Google Meet: route local Chrome joins through OpenClaw browser control instead of raw default Chrome, so agents use the configured OpenClaw browser profile when opening Meet. Thanks @oromeis. - Plugins/discovery: follow symlinked plugin directories in global and workspace plugin roots while keeping broken links ignored and existing package safety checks in place. Fixes #36754; carries forward #72695 and #63206. Thanks @Quackstro, @ming1523, and @xsfX20. - Plugins/install: skip test files and directories during install security scans while still force-scanning declared runtime entrypoints, so packaged test mocks no longer block plugin installs. Fixes #66840; carries forward #67050. Thanks @saurabhjain1592 and @Magicray1217. - Plugins/install: allow exact package-manager peer links back to the trusted OpenClaw host package during install security scans while continuing to block spoofed or nested escaping `node_modules` symlinks. Carries forward #70819. Thanks @fgabelmannjr. - Plugins/install: resolve plugin install destinations from the active profile state dir across CLI, ClawHub, marketplace, local path, and channel setup installs, so `openclaw --profile <name> plugins install ...` no longer writes into the default profile. Fixes #69960; carries forward #69971. Thanks @FrancisLyman and @Sanjays2402. - Plugins/registry: suppress duplicate-plugin startup warnings when a tracked npm-installed plugin intentionally overrides the bundled plugin with the same id. Carries forward #48673. Thanks @abdushsk. - Plugins/startup: reuse canonical realpath lookups throughout each plugin discovery pass, including package and manifest boundary checks, so Windows npm-global startups no longer repeat expensive path resolution for the same plugin roots. Fixes #65733. Thanks @welfo-beo. - Gateway/proxy: pass `ALL_PROXY` / `all_proxy` into the global Undici env-proxy dispatcher and provider proxy-fetch helper while keeping SSRF trusted-proxy auto-upgrade on `HTTP_PROXY` / `HTTPS_PROXY` only, so gateway/provider calls honor all-proxy setups without weakening guarded fetches. Fixes #43821; carries forward #43919. Thanks @RickyTong1. - Providers/LiteLLM: honor `--custom-base-url` during non-interactive API-key onboarding without adding proxy discovery side effects, so scripted remote LiteLLM setup keeps the requested endpoint instead of falling back to localhost. Carries forward #66160. Thanks @dongs0104. - Reply/link understanding: keep media and link preprocessing on stable runtime entrypoints and continue with raw message content if optional enrichment fails, so URL-bearing messages are no longer dropped after stale runtime chunk upgrades. Fixes #68466. Thanks @songshikang0111. - Discord: persist routed model-picker overrides when the hidden `/model` dispatch succeeds but the bound thread session store is still stale, including LM Studio suffixed model ids. Carries forward #61473. Thanks @Nanako0129. - Nodes/CLI: add `openclaw nodes remove --node <id|name|ip>` and `node.pair.remove` so stale gateway-owned node pairing records can be cleaned without hand-editing state files. - Gateway: include the connecting client and fresh presence version in the initial `hello-ok` snapshot, so clients no longer need a follow-up event before seeing themselves online. - Docker: install the CA certificate bundle in the slim runtime image so HTTPS calls from containerized gateways no longer fail TLS setup after the `bookworm-slim` base switch. Fixes #72787. Thanks @ryuhaneul. - Providers/OpenRouter: remove retired Hunter Alpha and Healer Alpha static catalog rows and disable proxy reasoning injection for stale Hunter Alpha configs, so replies are not hidden when OpenRouter returns answer text in reasoning fields. Fixes #43942. Thanks @EvanDataForge. - Providers/reasoning: let Groq and LM Studio declare provider-native reasoning effort values, so Qwen thinking models receive `none`/`default` or `off`/`on` instead of OpenAI-only `low`/`medium` values. Fixes #32638. Thanks @Aqu1bp, @mgoulart, @Norpps, and @BSTail. - Local models: default custom providers with only `baseUrl` to the Chat Completions adapter and trust loopback model requests automatically, so local OpenAI-compatible proxies receive `/v1/chat/completions` without timing out. Fixes #40024. Thanks @parachuteshe. - Channels/message tool: surface Discord, Slack, and Mattermost `user:`/`channel:` target syntax in the shared message target schema and Discord ambiguity errors, so DM sends by numeric id stop burning retries before finding `user:<id>`. Fixes #72401. Thanks @garyd9, @hclsys, and @praveen9354. - Agents/tools: scope tool-loop detection history to the active run when available, so scheduled heartbeat cycles no longer inherit stale repeated-call counts from previous runs. Fixes #40144. Thanks @mattbrown319. - Agents/subagents: preserve requester delivery for completion announces when a child agent is bound to a different channel account while keeping same-channel thread completions routed to the child thread. Thanks @sfuminya. - Agents/subagents: fail closed instead of selecting a single child thread binding when completion delivery lacks requester conversation signal. Thanks @suyua9. - Agents/status: persist the post-compaction token estimate from auto-compaction when providers omit usage metadata, so `/status` and session lists keep showing fresh context usage after compaction. Fixes #67667; carries forward #72822. Thanks @Jimmy-xuzimo and @skylight-9. - Control UI: show loading, reload, and retry states when a lazy dashboard panel cannot load after an upgrade, so the Logs tab no longer appears blank on stale browser bundles. Fixes #72450. Thanks @sobergou. - Gateway/plugins: start the Gateway in degraded mode when a single plugin entry has invalid schema config, and let `openclaw doctor --fix` quarantine that plugin config instead of crash-looping every channel. Fixes #62976 and #70371. Thanks @Doraemon-Claw and @pksidekyk. - Agents/plugins: skip malformed plugin tools with missing schema objects and report plugin diagnostics, so one broken tool no longer crashes Anthropic agent runs. Fixes #69423. Thanks @jmnickels. - Agents/reasoning: recover fully wrapped unclosed `<think>` replies that would otherwise sanitize to empty text while keeping strict stripping for closed reasoning blocks and unclosed tails after visible text. Fixes #37696; supersedes #51915. Thanks @druide67 and @okuyam2y. - Control UI/Gateway: bind WebChat handshakes to their active socket and reject post-close server registrations, so aborted connects no longer leave zombie clients or misleading duplicate WebSocket connection logs. Fixes #72753. Thanks @LumenFromTheFuture. - Agents/fallback: split ambiguous provider failures into `empty_response`, `no_error_details`, and `unclassified`, and add flat fallback-step fields to structured fallback logs so primary-model failures stay visible when later fallbacks also fail. Fixes #71922; refs #71744. Thanks @andyk-ms and @nikolaykazakovvs-ux. - Gateway/startup: reuse the plugin manifest registry inside config validation so restrictive plugin allowlists avoid a duplicate manifest pass during startup. Thanks @shakkernerd. - Gateway/startup: run plugin auto-enable from authored source config and skip disabled setup probes, avoiding runtime-default plugin allowlist writes and a second config snapshot read during startup. Thanks @shakkernerd. - Plugins/Windows: normalize Windows absolute paths before handing bundled plugin modules to Jiti, so Feishu/Lark message sending no longer fails with unsupported `c:` ESM loader URLs. Fixes #72783. Thanks @jackychen-png. - CLI/doctor: run bundled plugin runtime-dependency repairs through the async npm installer with spinner/line progress and heartbeat updates, so long `openclaw doctor --fix` installs no longer look hung in TTY or piped output. Fixes #72775. Thanks @dfpalhano. - Feishu/Windows: normalize bundled channel sidecar loads before Jiti evaluates them, so Feishu outbound sends no longer fail with raw `C:` ESM loader errors on Windows. Fixes #72783. Thanks @jackychen-png. - Agents/tools: ignore volatile `exec` runtime metadata when comparing tool-loop outcomes, so enabled loop detection can stop repeated identical shell-command results instead of resetting on duration, PID, session, or cwd changes. Fixes #34574; supersedes #41502. Thanks @gucasbrg and @Zcg2021. - Agents/fallback: classify internal live-session model switch conflicts as unknown fallback failures instead of provider overloads, preventing local vLLM endpoints from receiving misleading overloaded cooldowns. Refs #63229. Thanks @clawdia-lobster. - Discord: let thread sessions inherit the parent channel's session-level `/model` override as a model-only fallback without enabling parent transcript inheritance. Fixes #72755. Thanks @solavrc. - Gateway/plugins: skip stale configured channels whose matching plugin is no longer discoverable, point cleanup at `openclaw doctor --fix`, and keep unrelated channel typos fatal so one missing channel plugin no longer crash-loops the Gateway. Fixes #53311. Thanks @futhgar. - Control UI: keep session-specific assistant identity loads authoritative after WebSocket connect, so non-main agent chat sessions do not show the main agent name in the header after bootstrap refreshes. Fixes #72776. Thanks @rockytian-top. - Agents/Qwen: preserve exact custom `modelstudio` provider configs with foreign `api` owners so explicit OpenAI-compatible Model Studio endpoints no longer get normalized into the bundled Qwen plugin path. Fixes #64483. Thanks @FiredMosquito831. - MCP/bundle-mcp: normalize CLI-native `type: "http"` MCP server entries to OpenClaw `transport: "streamable-http"` on save, repair existing configs with doctor, and keep embedded Pi from falling back to legacy SSE GET-first startup for those servers. Fixes #72757. Thanks @Studioscale. - OpenCode: expose Anthropic Opus/Sonnet 4.x thinking levels for proxied Claude models, so `/think xhigh`, `/think adaptive`, and `/think max` validate consistently with the direct Anthropic provider. Fixes #72729. Thanks @haishmg and @aaajiao. - Media-understanding/audio: migrate deprecated `{input}` placeholders in legacy `audio.transcription.command` configs to `{{MediaPath}}`, so custom audio transcribers no longer receive the literal placeholder after doctor repair. Fixes #72760. Thanks @krisfanue3-hash. - Ollama/WSL2: warn when GPU-backed WSL2 installs combine CUDA visibility with an autostarting `ollama.service` using `Restart=always`, and document the systemd, `.wslconfig`, and keep-alive mitigation for crash loops. Carries forward #61022; fixes #61185. Thanks @yhyatt. - Ollama/onboarding: de-dupe suggested bare local models against installed `:latest` tags and skip redundant pulls, so setup shows the installed model once and no longer says it is downloading an already available model. Fixes #68952. Thanks @tleyden. - Memory-core/doctor: keep `doctor.memory.status` on the cached path by default and only run live embedding pings for explicit deep probes, preventing slow local embedding backends from blocking Gateway status checks. Fixes #71568. Thanks @apex-system. - Memory/QMD: group same-source collections into one QMD search invocation when the installed QMD supports multiple `-c` filters, while keeping older QMD builds on the per-collection fallback. Fixes #72484; supersedes #72485 and #69583. Thanks @BsnizND and @zeroaltitude. - Memory/QMD: accept QMD status vector-count variants such as `Vectors = 42`, `Vectors:42`, and `Vectors: 42 embedded`, so `memory status --deep` no longer reports embeddings unavailable for healthy QMD wrappers. Fixes #63652; carries forward #63678. Thanks @apoapostolov and @WarrenJones. - Memory/QMD: skip QMD vector status probes and embedding maintenance in lexical `searchMode: "search"`, so BM25-only QMD setups on ARM do not trigger llama.cpp/Vulkan builds during status checks or embed cycles. Fixes #59234 and #67113. Thanks @PrinceOfEgypt, @Vksh07, @Snipe76, @NomLom, @t4r3e2q1-commits, and @dmak. - Memory/QMD: report the live watcher dirty state in memory status, so changed QMD-backed memory files show as dirty until the queued sync finishes. Fixes #60244. Thanks @xinzf. - Compaction: skip oversized pre-compaction checkpoint snapshots and prune duplicate long user turns from compaction input and rotated successor transcripts, preventing retry storms from being preserved across checkpoint cycles. Fixes #72780. Thanks @SweetSophia. - Control UI/Cron: render cron job prompts and run summaries as sanitized markdown in the dashboard, with full-width block content, safer link clicks, and no duplicate error text when a failed run has no summary. Supersedes #48504. Thanks @garethdaine. - Control UI/Gateway: preserve WebChat client version labels across localhost, 127.0.0.1, and IPv6 loopback aliases on the same port, avoiding misleading `vcontrol-ui` connection logs while investigating duplicate-message reports. Refs #72753 and #72742. Thanks @LumenFromTheFuture and @allesgutefy. - Agents/reasoning: treat orphan closing reasoning tags with following answer text as a privacy boundary across delivery, history, streaming, and Control UI sanitizers so malformed local-model output cannot leak chain-of-thought text. Fixes #67092. Thanks @AnildoSilva. - Memory-core: run one-shot memory CLI commands through transient builtin and QMD managers so `memory index`, `memory status --index`, and `memory search` no longer start long-lived file watchers that can hit macOS `EMFILE` limits. Fixes #59101; carries forward #49851. Thanks @mbear469210-coder and @maoyuanxue. - Agents/ACP: ship the Claude ACP adapter with OpenClaw and require Claude result messages before idle can complete a prompt, preventing parent agents from waking early on long-running `sessions_spawn(runtime: "acp", agentId: "claude")` children. Fixes #72080. Thanks @siavash-saki and @iannwu. - CLI/tasks: route `tasks --json`, `tasks list --json`, and `tasks audit --json` through a lean JSON path so read-only task inspection no longer loads unrelated plugin/runtime command graphs. Fixes #66238. Thanks @ChuckChambers. - Memory-core: re-resolve the active runtime config whenever `memory_search` or `memory_get` executes, so provider changes made by `config.patch` stop leaving stale embedding backends behind in existing tool instances. Fixes #61098. Thanks @BradGroux and @Linux2010. - WebChat: keep bare `/new` and `/reset` startup instructions out of visible chat history while preserving `/reset <note>` as user-visible transcript text. Fixes #72369. Thanks @collynes and @haishmg. - Tasks/memory: checkpoint and truncate SQLite WAL sidecars on a timer and before close for task, Task Flow, proxy capture, and builtin memory databases, bounding long-running gateway `*.sqlite-wal` growth. Fixes #72774. Thanks @dfpalhano. - CLI/doctor: remove dangling channel config, heartbeat targets, and channel model overrides when stale plugin repair removes a missing channel plugin, preventing Gateway boot loops after failed plugin reinstalls. Fixes #65293. Thanks @yidecode. - Control UI/Gateway: cache, coalesce, stale-refresh, and invalidate effective tool inventory on channel registry changes while reusing the gateway-bound plugin registry and avoiding model/auth discovery, so chat runs no longer stall Control UI requests on repeated plugin/model setup. Fixes #72365; supersedes #72558. Thanks @Gabiii2398 and @1yihui. - Channels/setup: treat bundled channel plugins as already bundled during `channels add` and onboarding, enabling them without writing redundant `plugins.load.paths` entries or path install records. Fixes #72740. Thanks @iCodePoet. - WhatsApp: honor gateway `HTTPS_PROXY` / `HTTP_PROXY` env vars for QR-login WebSocket connections, while respecting `NO_PROXY`, so proxied networks no longer fall back to direct `mmg.whatsapp.net` connections that time out with 408. Fixes #72547; supersedes #72692. Thanks @mebusw and @SymbolStar. - Bonjour: default mDNS advertisements to the system hostname when it is DNS-safe, avoiding `openclaw.local` probing conflicts and Gateway restart loops on hosts such as `Lobster` or `ubuntu`. Fixes #72355 and #72689; supersedes #72694. Thanks @mscheuerlein-bot, @gcusms, @moyuwuhen601, @pavan987, @zml-0912, @hhq365, and @SymbolStar. - Agents/OpenAI-compatible: retry replay-safe empty `stop` turns once for `openai-completions` endpoints, so transient empty local backend responses no longer surface as “Agent couldn't generate a response” when a continuation succeeds, and restore `openclaw agent --model` for one-shot CLI runs. Fixes #72751. Thanks @moooV252. - Git hooks: skip ignored staged paths when formatting and restaging pre-commit files, so merge commits no longer abort when `.gitignore` newly ignores staged merged content. Fixes #72744. Thanks @100yenadmin. - Memory-core/dreaming: add a supported `dreaming.model` knob for Dream Diary narrative subagents, wired through phase config and the existing plugin subagent model-override trust gate. Refs #65963. Thanks @esqandil and @mjamiv. - Agents/Anthropic: remove trailing assistant prefill payloads when extended thinking is enabled, so Opus 4.7/Sonnet 4.6 requests do not fail Anthropic's user-final-turn validation. Fixes #72739. Thanks @superandylin. - Agents/vLLM/Qwen: add plugin-owned Qwen thinking controls for vLLM chat-template kwargs and DashScope-style top-level `enable_thinking` flags, including preserved thinking for agent loops. Fixes #72329. Thanks @stavrostzagadouris. - Memory-core/dreaming: treat request-scoped narrative fallback as expected, skip session cleanup when no subagent run was created, and remove duplicate phase-level cleanup so fallback no longer emits warning noise. Fixes #67152. Thanks @jsompis. - Agents/exec: apply configured `tools.exec.timeoutSec` to background, `yieldMs`, and node `system.run` commands when no per-call timeout is set, preventing auto-backgrounded and remote node commands from running indefinitely. Fixes #67600; supersedes #67603. Thanks @dlmpx and @kagura-agent. - Config/doctor: stop masking unknown-key validation diagnostics such as `agents.defaults.llm`, and have `openclaw doctor --fix` remove the retired `agents.defaults.llm` timeout block. Thanks @aidiffuser. - CLI/startup: keep the built pre-dispatch CLI graph free of package-level imports and extend packaged CLI smoke coverage to onboard and doctor help paths, preventing missing runtime dependencies such as tslog from killing onboarding before repair code can run. Fixes #63024. Thanks @hu19940121. - CLI/plugins: preserve unversioned ClawHub install specs so `plugins update` can follow newer ClawHub releases instead of pinning to the initially resolved version. Fixes #63010; supersedes #58426. Thanks @kangsen1234 and @robinspt. - Memory-core/subagents: tag plugin-created subagent sessions with their plugin owner so dreaming narrative cleanup can delete its own ephemeral sessions without granting broad admin session deletion. Fixes #72712. Thanks @BSG2000. - Gateway/models: move local-provider pricing opt-outs, OpenRouter/LiteLLM aliases, and proxy passthrough pricing lookup into plugin manifest metadata so core no longer carries extension-specific pricing tables. - CLI/update: honor `OPENCLAW_NO_AUTO_UPDATE=1` as a gateway startup kill-switch for configured background package auto-updates, so operators can hold a deliberate downgrade during incident recovery without editing config first. Fixes #72715. Thanks @Xivi08. - Agents/Claude CLI: force live-session launches to include `--output-format stream-json` whenever OpenClaw adds `--input-format stream-json`, so new Claude CLI sessions no longer fail immediately while reusable sessions keep working. Fixes #72206. Thanks @kwangwonkoh and @Xivi08. - CLI/plugins: accept ClawHub plugin API wildcard ranges such as `*` without rejecting compatible plugin installs, while still requiring a valid runtime API version. Fixes #56446; supersedes #56466. Thanks @darconada and @claygeo. - CLI/plugins: add an explicit `npm:<package>` install prefix that skips ClawHub lookup for known npm packages while keeping bare package specs ClawHub-first. Fixes #55805; supersedes #54377. Thanks @Zeoy2020 and @vagusX. - CLI/plugins: let config-gated bundled plugins install without persisting invalid placeholder config entries, so install/uninstall sweeps can cover plugins such as memory-lancedb before the user configures credentials. Thanks @vincentkoc. - CLI/plugins: reject malformed ClawHub plugin specs with trailing `@` before registry lookup, so empty-version typos report as invalid specs instead of package-not-found errors. Fixes #56579; supersedes #56582. Thanks @Kansodata. - Agents/sessions: acquire the session write lock only after cold bootstrap, plugin, and tool setup so fallback runs are not blocked by stalled pre-model startup work. - Browser/plugins: auto-start the bundled browser plugin when root `browser` config is present, including restrictive plugin allowlists, and ignore stale persisted plugin registries whose package paths no longer exist. - Browser: circuit-break repeated managed Chrome launch failures per profile so browser requests stop spawning Chromium indefinitely when CDP cannot start. Fixes #64271. Thanks @TheophilusChinomona. - Gateway/models: skip external OpenRouter and LiteLLM pricing refreshes for local/self-hosted model endpoints so startup does not wait on remote pricing catalogs for local-only Ollama, vLLM, and compatible providers. - CLI/plugins: stop security-blocked plugin installs from retrying as hook packs, so normal plugin packages report the scanner failure without a misleading "not a valid hook pack" follow-up. Fixes #61175; supersedes #64102. Thanks @KonsultDigital and @ziyincody. - Agents/Anthropic: strip stale trailing assistant prefill turns from outbound replay so context-engine short circuits cannot send unsupported assistant-prefill payloads to provider APIs. Fixes #72556. Thanks @Veda-openclaw. - Agents/Google: strip stale trailing assistant/model prefill turns from Gemini outbound replay so Google Generative AI requests end with a user turn or function response. Follow-up to #72556. Thanks @Veda-openclaw. - Control UI/Dreaming: require explicit confirmation before applying restart-impacting Dreaming mode changes, with restart warning copy and loading feedback. Fixes #63804. (#63807) Thanks @bbddbb1. - CLI/agent: mark Gateway-to-embedded fallback runs with `meta.transport: "embedded"` and `meta.fallbackFrom: "gateway"` in JSON output, and make the terminal diagnostic explicit so scripts and operators can distinguish fallback runs from Gateway runs. Fixes #71416. Thanks @amknight. - Agents/tools: normalize `null` or missing tool-call arguments to `{}` for parameterless object schemas before Pi validation, so empty-argument tools run instead of failing argument validation. Fixes #72587. Thanks @amknight. - Agents/subagents: clear active embedded-run state before terminal lifecycle events so post-completion cleanup no longer treats finished child runs as still active and skips archive or announcement bookkeeping. (#70187) Thanks @amknight. - CLI/update: keep the automatic post-update completion refresh on the core-command tree so it no longer stages bundled plugin runtime deps before the Gateway restart path, avoiding `.24` update hangs and 1006 disconnect cascades. Fixes #72665. Thanks @sakalaboator and @He-Pin. - Control UI: make explicit Reload Config actions discard stale local config edits while passive refreshes and failed-save recovery keep pending drafts intact. Fixes #40352; carries forward #40443. Thanks @realmikechong-dotcom. - Agents/Bedrock: stop heartbeat runs from persisting blank user transcript turns and repair existing blank user text messages before replay, preventing AWS Bedrock `ContentBlock` blank-text validation failures. Fixes #72640 and #72622. Thanks @goldzulu. - Agents/LM Studio: promote standalone bracketed local-model tool requests into registered tool calls and hide unsupported bracket blocks from visible replies, so MemPalace MCP lookups do not print raw `[tool]` JSON scaffolding in chat. Fixes #66178. Thanks @detroit357. - Local models: warn when an assistant reply looks like a tool call but the provider emitted plain text instead of a structured tool invocation, making fake/non-executed tool calls visible in logs. Fixes #51332. Thanks @emilclaw. - Local models: accept persisted non-secret local auth markers for private-LAN custom OpenAI-compatible providers, so LAN Ollama configs no longer fail with missing auth when `ollama-local` is saved as the key. Fixes #49736. Thanks @charles-zh. - TUI/local models: treat visible gateway client labels such as `openclaw-tui` as the current requester session for session-aware tools, so Ollama tool calls no longer fail by resolving the UI label as a session id. Fixes #66391. Thanks @kickingzebra. - Local models: route self-hosted OpenAI-compatible model discovery through the guarded fetch path pinned to the configured host, covering vLLM and SGLang setup without reopening local/LAN SSRF probes. Supersedes #46359. Thanks @cdxiaodong. - Local models: classify terminated, reset, closed, timeout, and aborted model-call failures and attach a process memory snapshot to the diagnostic event, making LM Studio/Ollama RAM-pressure failures easier to prove from stability bundles. Refs #65551. Thanks @BigWiLLi111. - Local models: pass configured provider request timeouts through OpenAI SDK transports and the model idle watchdog so long-running local or custom OpenAI-compatible streams use one timeout knob instead of hitting the SDK's 10-minute default or the 120s idle default. Fixes #63663. Thanks @aidiffuser. - LM Studio: trust configured LM Studio loopback, LAN, and tailnet endpoints for guarded model requests by default, preserving explicit private-network opt-outs. Refs #60994. Thanks @tnowakow. - Docker/setup: route Docker onboarding defaults for host-side LM Studio and Ollama through `host.docker.internal` and add the Linux host-gateway mapping to the bundled Compose file, so containerized gateways can reach local providers without using container loopback. Fixes #68684; supersedes #68702. Thanks @safrano9999 and @skolez. - Agents/LM Studio: strip prior-turn Gemma 4 reasoning from OpenAI-compatible replay while preserving active tool-call continuation reasoning. Fixes #68704. Thanks @chip-snomo and @Kailigithub. - LM Studio: allow interactive onboarding to leave the API key blank for unauthenticated local servers, using local synthetic auth while clearing stale LM Studio auth profiles. Fixes #66937. Thanks @olamedia. - Plugins/startup: use a `PluginLookUpTable` during Gateway startup so channel ownership, deferred channel loading, and startup plugin IDs reuse the same installed manifest registry instead of rebuilding manifest metadata on the boot path. Thanks @shakkernerd. - Plugins/startup: pass the Gateway `PluginLookUpTable` through plugin loading so auto-enable checks and startup-scope fallback reuse the same manifest registry instead of doing another manifest pass. Thanks @shakkernerd. - Plugins/startup: carry the Gateway `PluginLookUpTable` into deferred channel full-runtime reloads so post-listen startup does not rebuild manifest metadata after the provisional setup-runtime load. Thanks @shakkernerd. - Gateway/models: reuse Gateway plugin manifest metadata during the initial model-pricing refresh so pricing policies and configured plugin web-search models do not rebuild plugin lookups during startup. Thanks @shakkernerd. - Gateway/startup: extend `OPENCLAW_GATEWAY_STARTUP_TRACE=1` with per-phase event-loop delay plus plugin lookup-table timing and count metrics for installed-index, manifest, startup-plan, and owner-map work, and include the new timing fields in startup benchmark summaries. Thanks @shakkernerd. - Plugins/channels: resolve read-only channel command defaults from one plugin index plus manifest pass instead of reloading plugin metadata while checking candidate plugin enablement. Thanks @shakkernerd. - Plugins/capabilities: cache manifest-derived capability provider plugin IDs per config snapshot so repeated TTS, media, realtime, memory, image, video, and music provider resolution avoids redundant manifest scans. Thanks @shakkernerd. - Plugins/contracts: resolve runtime manifest-contract plugin owners from one plugin index plus manifest pass instead of rebuilding manifest metadata separately for all owners and enabled owners. Thanks @shakkernerd. - Plugins/extractors: reuse one manifest registry pass while resolving bundled document and web-content extractor plugins instead of rereading manifests for compatibility and enablement filtering. Thanks @shakkernerd. - Plugins/providers: reuse one plugin registry snapshot and manifest registry while resolving provider discovery entries instead of rebuilding manifest metadata after provider owner discovery. Thanks @shakkernerd. - Plugins/registry: resolve lookup-table owner maps for providers, CLI backends, setup providers, command aliases, model catalogs, channel configs, and manifest contracts while preserving setup-only CLI backend ownership. Thanks @shakkernerd. - Plugins/registry: cache repeated installed-index manifest registry fallback rebuilds behind a bounded invalidating cache so cold provider-discovery paths avoid rereading unchanged manifests. Thanks @mcaxtr. - Plugins/web: reuse manifest records already loaded for bundled web provider candidate discovery when falling back to public artifact provider loading. Thanks @shakkernerd. - Mattermost: keep direct-message replies top-level by suppressing reply roots for DM delivery while preserving channel and group thread roots, and derive inbound chat kind from the trusted channel lookup instead of the websocket event channel type. Carries forward #60115, #55186, #72305, and #72659; refs #59758, #59981, #59791, and #57565. Thanks @vincentkoc, @jwchmodx, and @hnykda. - Docker: pre-create `/home/node/.openclaw` with node ownership and private permissions so first-run Docker Compose named volumes no longer fail startup with EACCES. (#48072, #63959; fixes #61279) Thanks @timoxue and @jeanibarz. - CLI/Gateway: treat local restart probe policy closes for connect, exact `device required`, pairing, and auth failures as Gateway reachability proof without accepting empty, broad standalone token/password/scope/role, or pair-substring 1008 close reasons. Fixes #48771; carries forward #48801; related #63491. Thanks @MarsDoge and @genoooool. - Feishu: send outgoing interactive reply payloads as native cards with clickable buttons while preserving text, media, and document-comment fallbacks. Fixes #13175 and #58298; carries forward #47891. Thanks @Horacehxw. - Control UI/WebChat: skip redundant final-event history reloads when the assistant payload already rendered, and keep deferred `session.message` reloads attached to the active run so final reconciliation no longer splits, duplicates, or drops assistant bubbles. Fixes #66875 and #66274; follows #66997 and #67037. Thanks @BiznessFish, @scotthuang, and @hansolo949. - CLI/Agents: route new `openclaw agent --to` sessions through the configured default agent while migrating legacy `agent:main:<mainKey>` rows into the default-agent store, preserving the default-agent fix from #64108. Fixes #63992; related #56370, #56453, and #42009. Thanks @mushuiyu886 and @voocel. - Process/Windows: decode command stdout and stderr from raw bytes with console-codepage awareness, while preserving valid UTF-8 output and multibyte characters split across chunks. Fixes #50519. Thanks @iready, @kevinten10, @zhangyongjie1997, @knightplat-blip, @heiqishi666, and @slepybear. - Bonjour/Windows: hide the bundled mDNS advertiser's Windows ARP shell probe so Gateway startup no longer flashes command-prompt windows. Fixes #70238. Thanks @alexandre-leng, @PratikRai0101, @infinitypacific, and @tomerpeled. - Agents/bootstrap: dedupe hook-injected bootstrap context files by workspace-relative path and store normalized resolved paths so duplicate relative and absolute hook paths no longer depend on the process cwd. (#59344; fixes #59319; related #56721, #56725, and #57587) Thanks @koen666. - Agents/bootstrap: refresh cached workspace bootstrap snapshots on long-lived main-session turns when `AGENTS.md`, `SOUL.md`, `MEMORY.md`, or `TOOLS.md` change on disk, while preserving unchanged snapshot identity through the workspace file cache. (#64871; related #43901, #26497, #28594, #30896) Thanks @aimqwest and @mikejuyoon. - macOS Gateway: detect installed-but-unloaded LaunchAgent split-brain states during status, doctor, and restart, and re-bootstrap launchd supervision before falling back to unmanaged listener restarts. Fixes #67335, #53475, and #71060; refs #58890, #60885, and #70801. Thanks @ze1tgeist88, @dafacto, and @vishutdhar. - Plugins/install: treat mirrored core logger dependencies as staged bundled runtime deps so packaged Gateway starts do not crash when the external plugin-runtime-deps root is missing `tslog`. Fixes #72228; supersedes #72493. Thanks @deepujain. - Build/plugins: preserve active bundled runtime-dependency staging temp directories owned by live build processes so overlapping postbuild runs no longer delete each other's staged deps mid-prune. Supersedes #72220. Thanks @VACInc. - Plugins/install: hide bundled runtime-dependency npm child windows on Windows across Gateway startup, postinstall, and packaged staging paths so Telegram/Anthropic dependency repair no longer flashes shell windows. Fixes #72315. Thanks @athuljayaram and @joshfeng. - Agents/Windows: normalize lazy agent runtime imports before Node ESM loading so Windows drive-letter `subagent-registry` runtime paths no longer fail every agent task with `ERR_UNSUPPORTED_ESM_URL_SCHEME`. Fixes #72636; carries forward #72716. Thanks @Andyz-CData and @xialonglee. - Plugins/Windows: normalize lazy plugin service override imports before Node ESM loading so drive-letter browser-control module paths no longer fail with `ERR_UNSUPPORTED_ESM_URL_SCHEME`. Fixes #72573; supersedes #72599 and #72582. Thanks @llzzww316, @feineryonah-byte, and @WuKongAI-CMU. - Browser/plugins: load `playwright-core` through the browser runtime shim so packaged installs can run Playwright actions from staged plugin runtime deps after doctor/startup repair. Fixes #72168; supersedes #72238. Thanks @zdg1110 and @yetval. - Plugins/install: stage bundled plugin runtime dependencies before Gateway startup, drain update restarts, and materialize plugin-owned root chunks in external mirrors so staged deps resolve under native ESM. Fixes #72058; supersedes #72084. Thanks @amnesia106 and @drvoss. - TTS/SecretRef: resolve `messages.tts.providers.*.apiKey` from the active runtime snapshot so SecretRef-backed MiniMax and other TTS provider keys work in runtime reply/audio paths. Fixes #68690. Thanks @joshavant. - Gateway/install: surface systemd user-bus recovery hints during Linux service activation and retry via the target user scope when `systemctl --user` reports no-medium bus failures, without letting stale `SUDO_USER` override `sudo -u` installs. Fixes #39673; refs #44417 and #63561. Thanks @Arbor4, @myrsu, @mssteuer, and @boyuaner. - CLI/nodes: make unfiltered `openclaw nodes list` prefer the effective paired-node view used by `nodes status` while preserving pending rows, pairing-scope fallback, terminal-safe table rendering, and paired JSON metadata. Fixes #46871; carries forward #65772 through the ProjectClownfish #72619 repair. Thanks @skainguyen1412. - CLI/startup: read generated startup metadata from the bundled `dist` layout before falling back to live help rendering, so root/browser help and channel-option bootstrap stay on the fast path. Thanks @vincentkoc. - Feishu/Lark: stop treating broadcast-only `@all`/`@_all` messages as bot mentions while preserving direct bot mentions, including messages that also include `@all`. Fixes #37706. Thanks @JosepLee. - CLI/help: treat positional `help` invocations like `openclaw channels help` as help paths for startup gating, avoiding model/auth warmup while preserving positional arguments such as `openclaw docs help`. Thanks @gumadeiras. - Web search: route plugin-scoped web_search SecretRefs through the active runtime config snapshot so provider execution receives resolved credentials across app/runtime paths, including `plugins.entries.brave.config.webSearch.apiKey`. Fixes #68690. Thanks @VACInc. - Voice Call: allow SecretRef-backed Twilio auth tokens and call-specific OpenAI/ElevenLabs TTS API keys through the plugin config surface. Fixes #68690. Thanks @joshavant. - Google Meet: clean stale chrome-node realtime audio bridges by URL before rejoining, expose active node bridge inspection, and tolerate transient node input pull failures instead of dropping the Meet session. Fixes #72371. (#72372) Thanks @BsnizND. - Google Meet: use 24 kHz PCM16 for Chrome command-pair realtime audio by default, preserve legacy 8 kHz G.711 mu-law custom command pairs, and let realtime providers negotiate the selected bridge audio format. Fixes #72525. Thanks @BsnizND. - Google Meet: clear queued Gemini Live playback when realtime interruptions arrive, restart Chrome command-pair audio output after clears, and expose Google Live interruption/VAD config knobs for Meet and Voice Call realtime bridges. Fixes #72523. (#72524) Thanks @BsnizND. - Google Meet: add `realtime.agentId` so live meeting consults can target a named OpenClaw agent instead of always using `main`. (#72381) Thanks @BsnizND. - Google Meet: route stateful `google_meet` tool actions through the gateway-owned runtime so created or joined realtime sessions remain visible to status, speak, and leave after the agent turn ends. Fixes #72440. (#72441) Thanks @BsnizND. - Google Meet/Voice Call: send Gemini Live a non-blocking consult continuation before long OpenClaw agent consults finish, then deliver the final result when idle so calls and meetings do not sit silent during tool-backed answers. (#72189) Thanks @VACInc. - Google Meet: preserve Gemini Live function names when replying to realtime tool calls so Google SDK validation accepts the `FunctionResponse` payload. Fixes #72425. (#72426) Thanks @BsnizND. - Discord/media: keep incidental Markdown image badges in final replies as text unless a channel opts into Markdown-image media extraction, while preserving Telegram Markdown-image media replies and explicit `MEDIA:` attachments. Fixes #72642. Thanks @solavrc and @Bartok9. - Matrix/E2EE: stabilize recovery and broken-device QA flows while avoiding Matrix device-cleanup sync races that could leave shutdown-time crypto work running. Thanks @gumadeiras. - Cron: apply `cron.maxConcurrentRuns` to a dedicated `cron-nested` isolated agent-turn lane as well as cron dispatch, so parallel cron jobs no longer serialize on inner LLM execution while non-cron nested flows keep their existing lane behavior. Fixes #72707. Thanks @kagura-agent. - Cron: report isolated runs as successful when verified cron delivery already delivered the reply, while keeping unresolved Message/Canvas tool failures fatal. Fixes #72732 and #50170; follow-up to #54188. Thanks @zNatix, @pixeldyn, and @ChickenEggRoll. - Cron: treat isolated run-level agent failures as job errors even when no reply payload is produced, synthesizing a safe error payload so model/provider failures increment error counters and trigger failure notifications instead of clearing as successful. Fixes #43604; carries forward #43631. Thanks @SPFAdvisors. - Cron: preserve exact `NO_REPLY` tool results from isolated jobs with empty final assistant turns as quiet successes instead of surfacing incomplete-turn errors. Fixes #68452; carries forward #68453. Thanks @anyech. - Cron: resolve failure alerts and failure-destination announcements against `session:<id>` targets before falling back to the creator session, so jobs created from group chats can notify the targeted direct session without cross-account routing errors. Refs #62777; carries forward #68535. Thanks @slideshow-dingo and @likewen-tech. - Discord: preserve explicit `user:` and `channel:` delivery targets through plugin routing so cron announcements and failure alerts keep their intended recipient kind. Refs #62777; carries forward #62798. Thanks @neeravmakwana. - Cron: add `failureAlert.includeSkipped` and `openclaw cron edit --failure-alert-include-skipped` so persistently skipped jobs can alert without counting skips as execution errors or affecting retry backoff. Fixes #60846. Thanks @slideshow-dingo. - Cron: invalidate stale pending runtime slots after live or offline `jobs.json` schedule edits, while preserving due slots for formatting-only rewrites. Fixes #27996 and #71607; carries forward #71651. Thanks @xialonglee and @fagnersouza666. - Cron: keep legacy flat `jobs.json` rows loadable while comparing split-state schedule identities, so old cron stores do not crash before in-memory hydration can normalize them. - Cron: start isolated agent-turn execution timeouts after the runner enters its effective execution lane, so queued cron/manual runs no longer spend their whole timeout budget before useful work begins. Fixes #41783. Thanks @ayanesakura and @Hurray0. - Cron/Telegram: preserve direct-chat thread IDs and optional account IDs when inferring reminder delivery from Telegram direct-thread session keys. Fixes #44270; carries forward #44325, #44351, #44412, and #72657. Thanks @RunMintOn, @arkyu2077, @0xsline, and @vincentkoc. - Cron: omit synthetic `delivery.resolved` errors from `--no-deliver` run records while preserving explicit no-deliver target traces for agent-initiated messages. Fixes #72210; carries forward #72219. Thanks @hatemclawbot-collab and @xydigit-sj. - Cron: classify isolated runs as errors from structured embedded-run execution-denial metadata, with final-output marker fallback for `SYSTEM_RUN_DENIED`, `INVALID_REQUEST`, and approval-binding refusals, so blocked commands no longer appear green in cron history. Fixes #67172; carries forward #67186. Thanks @oc-gh-dr, @hclsys, and @1yihui. - Subagents: keep the delegated task only in the subagent system prompt and send a short initial kickoff message, avoiding duplicate task tokens while preserving multiline task formatting. Fixes #72019; carries forward #72053. Thanks @Wizongod and @ly85206559. - Onboarding/GitHub Copilot: add manifest-owned `--github-copilot-token` support for non-interactive setup, including env fallback, tokenRef storage in ref mode, saved-profile reuse, and current Copilot default-model wiring. Refs #50002 and supersedes #50003. Thanks @scottgl9. - Gateway/install: add a validated `--wrapper`/`OPENCLAW_WRAPPER` service install path that persists executable LaunchAgent/systemd wrappers across forced reinstalls, updates, and doctor repairs instead of falling back to raw node/bun `ProgramArguments`. Fixes #69400. (#72445) Thanks @willtmc. - Plugins: fail plugin registration when loader-owned acceptance gates reject missing hook names or memory-only capability registration from non-memory plugins, surfacing the issue through plugin status and doctor instead of silently dropping the registration. Fixes #72459. Thanks @amknight. - macOS Gateway: write launchd services with a state-dir `WorkingDirectory`, use a durable state-dir temp path instead of freezing macOS session `TMPDIR`, create that temp directory before bootstrap, and label abort-shaped launchd exits as `SIGABRT/abort` in status output. Fixes #53679 and #70223; refs #71848. Thanks @dlturock, @stammi922, and @palladius. - Control UI/update: make `Update now` require a real gateway process replacement, report skipped/error update outcomes with stable reasons, and verify the running gateway version after restart so global installs cannot silently keep old code in memory. Fixes #62492; addresses #64892 and #63562. Thanks @IAMSamuelRodda. - Exec approvals: accept runtime-owned `source: "allow-always"` and `commandText` allowlist metadata in gateway and node approval-set payloads so Control UI round-trips no longer fail with `unexpected property 'source'`. Fixes #60000; carries forward #60064. Thanks @sd1471123, @sharkqwy, and @luoyanglang. - Exec/node: skip approval-plan preparation for full-trust `host=node` runs so interpreter and script commands no longer fail with `SYSTEM_RUN_DENIED: approval cannot safely bind` when effective policy is `security=full` and `ask=off`. Fixes #48457 and duplicate #69251. Thanks @ajtran303, @jaserNo1, @Blakeshannon, @lesliefag, and @AvIsBeastMC. - Exec/node: synthesize a local approval plan when a paired node advertises `system.run` without `system.run.prepare`, unblocking approval-required `host=node` exec on current macOS companion nodes while preserving remote prepare for node hosts that support it. Fixes #37591 and duplicate #66839; carries forward #69725. Thanks @soloclz. - Memory/QMD: prefer QMD's `--mask` collection pattern flag so root memory indexing stays scoped to `MEMORY.md` instead of widening to every markdown file in the workspace. Fixes #65480; supersedes #65481 and #66259. Thanks @ccage-simp, @Bortlesboat, @seank-com, and @crazyscience. - Memory/doctor: treat the specific `gateway timeout after ...` gateway memory probe result as inconclusive instead of reporting embeddings not ready, while preserving warnings for explicit failures. Fixes #44426; carries forward #46576 with the Greptile review feedback applied. Thanks Cengiz (@ghost). - Gateway/memory: defer QMD startup for implicit non-default agents and scope memory runtime loading to the selected memory slot so Gateway boot and first memory recall avoid broad plugin runtime fanout. Thanks @vincentkoc. - Gateway/startup: keep core request handlers, setup wizard, and channel runtime helpers off the boot path until the first matching request, wizard run, or channel start, reducing no-plugin Gateway ready RSS and avoidable startup imports. Thanks @vincentkoc. - Gateway/startup: keep CLI outbound channel send dependencies as lazy request-time senders so Gateway boot no longer imports channel plugin registration just to construct default deps. Thanks @vincentkoc. - Gateway/startup: split lightweight HTTP auth helpers away from model-override helpers so Gateway bind no longer imports model catalog selection while wiring base HTTP routes. Thanks @vincentkoc. - Gateway/startup: lazy-load plugin HTTP route dispatch when active plugin routes exist so no-plugin Gateway boot skips plugin route runtime scope setup. Thanks @vincentkoc. - Gateway/startup: move chat run/subscriber registries onto a lightweight state module and defer chat/session event projection until the first event so Gateway boot skips session IO imports. Thanks @vincentkoc. - Gateway/startup: keep node session runtime on a lightweight JSON parser instead of importing gateway method validation helpers during boot. Thanks @vincentkoc. - Gateway/startup: read embedded-run activity from a lightweight shared state module so restart deferral no longer imports the embedded runner during Gateway boot. Thanks @vincentkoc. - Gateway/startup: defer MCP loopback server imports until Gateway shutdown so normal boot no longer loads the loopback HTTP/tool schema stack just to register close handlers. Thanks @vincentkoc. - Gateway/startup: resolve channel runtime helpers asynchronously only when an enabled/configured channel starts, so no-channel Gateway boot skips auto-reply, media, pairing, and outbound channel helper imports. Thanks @vincentkoc. - Gateway/startup: lazy-load HTTP auth, canvas auth, and plugin route scope helpers from their request paths so Gateway bind no longer pays those utility graphs during boot. Thanks @vincentkoc. - Gateway/startup: defer isolated cron runner imports until `/hooks/agent` dispatch so Gateway boot skips the agent-turn runtime on installs that only need normal HTTP bind. Thanks @vincentkoc. - Gateway/startup: split hook request parsing into a request-path module and load the Gateway hook dispatcher only when a request matches the hooks base path, keeping hook mapping and throttle helpers off plain HTTP bind. Thanks @vincentkoc. - CLI/Gateway: use a parse-only config snapshot for plain `gateway status` reads and reuse same-path service config context so status no longer spends tens of seconds in full config validation before printing. Thanks @vincentkoc. - Lobster/Gateway: memoize repeated Ajv schema compilation before loading the embedded Lobster runtime so scheduled workflows and `llm.invoke` loops stop growing gateway heap on content-identical schemas. Fixes #71148. Thanks @cmi525, @vsolaz, and @vincentkoc. - Codex harness: normalize cached input tokens before session/context accounting so prompt cache reads are not double-counted in `/status`, `session_status`, or persisted `sessionEntry.totalTokens`. Fixes #69298. Thanks @richardmqq. - Hooks/session-memory: use the host local timezone for memory filenames, fallback timestamp slugs, and markdown headers instead of UTC dates. Fixes #46703. (#46721) Thanks @Astro-Han. - Gateway health: preserve live runtime-backed channel/account state in `gateway.health` snapshots and cached refreshes while keeping raw probe payloads on sensitive/admin paths only. (#39921, #42586, #46527, #52770, #42543) Thanks @FAL1989, @rstar327, @0xble, and @ajayr. - Feishu: extract quoted/replied interactive-card text across schema 1.0, schema 2.0, i18n, template-variable, and post-format fallback shapes without carrying broad generated/config churn from related parser experiments. (#38776, #60383, #42218, #45936) Thanks @lishuaigit, @lskun, @just2gooo, and @Br1an67. - Telegram/agents: hide raw failed write/edit warning messages in Telegram when the assistant already explicitly acknowledges the failed action, while keeping warnings when the reply claims success or omits the failure; #39406 remains the broader configurable delivery-policy follow-up. Fixes #51065; covers #39631. Thanks @Bartok9 and @Bortlesboat. - Exec approvals: accept a symlinked `OPENCLAW_HOME` as the trusted approvals root while still rejecting symlinked `.openclaw` path components below it. (#64663) Thanks @FunJim. - Logging: add top-level `hostname`, flattened `message`, and available `agent_id`, `session_id`, and `channel` fields to file-log JSONL records for multi-agent filtering without removing existing structured log arguments. Fixes #51075. Thanks @stevengonsalvez. - ACP: route server logs to stderr before Gateway config/bootstrap work so ACP stdout remains JSON-RPC only for IDE integrations. Fixes #49060. Thanks @Hollychou924. - Logging: propagate internal request trace scopes through Gateway HTTP requests and WebSocket frames so file logs, diagnostic events, agent run traces, model-call traces, OTEL spans, and trusted provider `traceparent` headers share a correlatable `traceId` without logging raw request or model content. Fixes #40353. Thanks @liangruochong44-ui. - Diagnostics/OTEL: capture privacy-safe model-call request payload bytes, streamed response bytes, first-response latency, and total duration in diagnostic events, plugin hooks, stability snapshots, and OTEL model-call spans/metrics without logging raw model content. Fixes #33832. Thanks @wwh830. - Logging: write validated diagnostic trace context as top-level `traceId`, `spanId`, `parentSpanId`, and `traceFlags` fields in file-log JSONL records so traced requests and model calls are easier to correlate in log processors. Refs #40353. Thanks @liangruochong44-ui. - Logging/sessions: apply configured redaction patterns to persisted session transcript text and accept escaped character classes in safe custom redaction regexes, so transcript JSONL no longer keeps matching sensitive text in the clear. Fixes #42982. Thanks @panpan0000. - Agents/sessions: let `sessions_spawn runtime="subagent"` ignore ACP-only `streamTo` and `resumeSessionId` fields while keeping ACP passthrough and documenting `streamTo` as ACP-only. Fixes #43556 and #63120; covers #56326, #61724, #64714, and #67248; carries forward #68397, #65282, #58686, #56342, and #40102. Thanks @skernelx, @damselem, @Br1an67, @Mintalix, @IsaacAPerez, @vvitovec, @Sanjays2402, @shenkq97, and @1034378361. - Providers/Ollama: honor `/api/show` capabilities when registering local models so non-tool Ollama models no longer receive the agent tool surface, and keep native Ollama thinking opt-in instead of enabling it by default. Fixes #64710 and duplicate #65343. Thanks @yuan-b, @netherby, @xilopaint, and @Diyforfun2026. - Image tool/media: honor `tools.media.image.timeoutSeconds` and matching per-model image timeouts in explicit image analysis, including the MiniMax VLM fallback path, so slow local vision models are not capped by hardcoded 30s/60s aborts. Fixes #67889; supersedes #67929. Thanks @AllenT22 and @alchip. - Providers/Ollama: read larger custom Modelfile `PARAMETER num_ctx` values from `/api/show` so auto-discovered Ollama models with expanded context no longer stay pinned to the base model context. Fixes #68344. Thanks @neeravmakwana. - Providers/Ollama: honor configured model `params.num_ctx` in native and OpenAI-compatible Ollama requests so local models can cap runtime context without rebuilding Modelfiles. Fixes #44550 and #52206; supersedes #69464. Thanks @taitruong, @armi0024, and @LokiCode404. - Providers/Ollama: stop forcing native Ollama requests to use the full configured `contextWindow` as `options.num_ctx` unless `params.num_ctx` is explicit, so local models can keep Ollama's VRAM/env default instead of looking hung on first turns. Fixes #49684 and #68662. Thanks @zhouZcong and @dshenster-byte. - Providers/Ollama: forward whitelisted native Ollama model params such as `temperature`, `top_p`, and top-level `think` so users can disable API-level thinking or tune local models from config without proxy shims. Fixes #48010. Thanks @tangzhi, @pandego, @maweibin, @Adam-Researchh, and @EmpireCreator. - Providers/Ollama: expose native Ollama thinking effort levels so `/think max` is accepted for reasoning-capable Ollama models and maps to Ollama's highest supported `think` effort. Fixes #71584. Thanks @g0st1n. - Providers/Ollama: strip the active custom Ollama provider prefix before native chat and embedding requests, so custom provider ids like `ollama-spark/qwen3:32b` reach Ollama as the real model name. Fixes #72353. Thanks @maximus-dss and @hclsys. - Providers/Ollama: parse stringified native tool-call arguments before dispatch, preserving unsafe integer values so Ollama tool use receives structured parameters. Fixes #69735; supersedes #69910. Thanks @rongshuzhao and @yfge. - Providers/Ollama: skip ambient localhost discovery unless Ollama auth or meaningful config opts in, preventing unexpected probes to `127.0.0.1:11434` for users who are not using Ollama. Fixes #56939; supersedes #57116. Thanks @IanxDev and @tsukhani. - Providers/Ollama: skip implicit localhost discovery when a custom remote `api: "ollama"` provider is configured, while still treating `127/8` loopback hosts as local. Carries forward #43224. Thanks @issacthekaylon. - Providers/models: honor provider-level `contextWindow`, `contextTokens`, and `maxTokens` as defaults when resolving discovered models, so local Ollama and other self-hosted providers can cap all models without repeating per-model entries. Fixes #44786; carries forward #44955. Thanks @voltwake and @maweibin. - Providers/Ollama: move memory embeddings to Ollama's current `/api/embed` endpoint with batched `input` requests while preserving vector normalization and custom provider auth/header overrides. Fixes #39983. Thanks @sskkcc and @LiudengZhang. - Providers/Ollama: route local web search through Ollama's signed `/api/experimental/web_search` daemon proxy, use hosted `/api/web_search` directly for `ollama.com`, and keep `OLLAMA_API_KEY` scoped to cloud fallback auth. Fixes #69132. Thanks @yoon1012 and @hyspacex. - Providers/Ollama: accept OpenAI SDK-style `baseURL` as an alias for `baseUrl` across discovery, streaming, setup pulls, embeddings, and web search so remote Ollama hosts are not silently ignored. Fixes #62533; supersedes #62549. Thanks @Julien-BKK and @Linux2010. - Providers/Ollama: scope synthetic local auth and embedding bearer headers to declared Ollama host boundaries so cloud keys are not sent to local/self-hosted embedding endpoints and remote/cloud Ollama endpoints no longer receive the `ollama-local` marker as if it were a real token. Supersedes #69261 and #69857; refs #43945. Thanks @hyspacex, @maxramsay, and @Meli73. - Providers/Ollama: resolve custom-named local Ollama providers such as `ollama-remote` through the Ollama synthetic-auth hook so subagents no longer miss `ollama-local` auth and silently fall back to cloud models. Fixes #43945. Thanks @Meli73 and @maxramsay. - Providers/Ollama: add provider-scoped model request timeouts, thread them through guarded fetch connect/header/body/abort handling, and document `params.keep_alive` for cold local models so first-turn Ollama loads no longer require global agent timeout changes. Fixes #64541 and #68796; supersedes #65143 and #66511. Thanks @LittleJakub, @Juankcba, @uninhibite-scholar, and @yfge. - Providers/Ollama: preserve explicit configured model input modalities when merging discovered provider metadata so custom vision models keep image support instead of silently dropping attachments. Fixes #39690; carries forward #39785. Thanks @Skrblik and @Mriris. - Providers/Ollama: estimate native Ollama transcript usage when `/api/chat` omits prompt/eval counters while preserving exact zero counters, keeping local model runs visible in usage surfaces. Carries forward #39112. Thanks @TylonHH. - Agents/Ollama: retry native Ollama turns that finish without user-visible text, including unsigned thinking-only responses, so constrained reasoning turns can continue instead of surfacing an empty reply. Carries forward #66552 and #61223. Thanks @yfge and @L3G. - Docs/Ollama: expand setup recipes for local, LAN, cloud, multi-host, web search, embeddings, thinking control, and large-context troubleshooting. - Providers/PDF/Ollama: add bounded network timeouts for Ollama model pulls and native Anthropic/Gemini PDF analysis requests so unresponsive provider endpoints no longer hang sessions indefinitely. Fixes #54142; supersedes #54144 and #54145. Thanks @jinduwang1001-max and @arkyu2077. - LLM Task/Ollama: accept model overrides that already include the selected provider prefix, avoiding doubled ids such as `ollama/ollama/llama3.2:latest`, and live-verify local Ollama JSON tasks return parsed output. Fixes #50052. Thanks @ralphy-maplebots and @Hollychou924. - Memory/doctor: treat Ollama memory embeddings as key-optional so `openclaw doctor` no longer warns about a missing API key when the gateway reports embeddings are ready. Fixes #46584. Thanks @fengly78. - Agents/Ollama: apply provider-owned replay turn normalization to native Ollama chat so Cloud models no longer reject non-alternating replay history in agent/Gateway runs. Fixes #71697. Thanks @ismael-81. - Control UI/Ollama: show the resolved configured thinking default in chat and session thinking dropdowns so inherited `adaptive`/per-model thinking config no longer appears as `Default (off)` or a generic inherit value. Fixes #72407. Thanks @NotecAG. - Agents/Ollama: validate explicit `--thinking max` against catalog-discovered Ollama reasoning metadata so local agent runs accept the same native thinking levels shown in the model catalog. Fixes #71584. Thanks @g0st1n. - CLI/models: include explicitly configured provider models in `openclaw models list --provider <id>` without requiring the full catalog path, so configured Ollama models are visible. Fixes #65207. Thanks @drzeast-png. - Docker/QA: add observability coverage to the normal Docker aggregate so QA-lab OTEL and Prometheus diagnostics run inside Docker. Thanks @vincentkoc. - Auto-reply: poison inbound message dedupe after replay-unsafe provider/runtime failures so retries stay safe before visible progress but cannot duplicate messages after block output, tool side effects, or session progress. Fixes #69303; keeps #58549 and #64606 as duplicate validation. Thanks @martingarramon, @NikolaFC, and @zeroth-blip. - Agents/model fallback: keep auto-persisted fallback model overrides selected across turns until `/new` or reset clears them, avoiding repeated probes of a known-bad primary while `/status` shows the selected and active models. Thanks @kibedu. - Agents/model fallback: jump directly to a known later live-session model redirect instead of walking unrelated fallback candidates, while preserving the already-landed live-session/fallback loop guard. Fixes #57471; related loop family already closed via #58496. Thanks @yuxiaoyang2007-prog. - Gateway/Bonjour: keep @homebridge/ciao cancellation handlers registered across advertiser restarts so late probing cancellations cannot crash Linux and other mDNS-churned gateways. - 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"`. - 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. - 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. - 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. - WebChat/Control UI: support non-video file attachments in chat uploads while preserving the existing image attachment path and MIME-sniff fallback for generic image uploads. (#70947) Thanks @IAMSamuelRodda. - Skills/memory: restore Chokidar v5 hot reloads by watching concrete skill and memory roots with filters, including SKILL.md removals and deleted skill folders without broad workspace recursion. Fixes #27404, #33585, and #41606. Thanks @shelvenzhou, @08820048, and @rocke2020. - 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. - Gateway/session rows: report the same config-resolved thinking default that runtime sessions use, including global and per-agent defaults, so Control UI and TUI default labels stay aligned. (#71779, #70981, #71033, #70302) Thanks @chen-zhang-cs-code, @SymbolStar, and @cholaolu-boop. - Plugins: share package entrypoint resolution between install and discovery, reject mismatched `runtimeExtensions`, and cache bundled runtime-dependency manifest reads during scans. - WhatsApp/Web: keep quiet but healthy linked-device sessions connected by basing the watchdog on WhatsApp Web transport activity, while retaining a longer app-silence cap so frame activity cannot mask a stuck session forever. Fixes #70678; carries forward the focused #71466 approach and keeps #63939 as related configurable-timeout follow-up. Thanks @vincentkoc and @oromeis. - Discord/gateway: count failed health-monitor restart attempts toward cooldown and hourly caps, and evict stale account lifecycle state during channel reloads so repeated Discord gateway recovery cannot loop on old status. Fixes #38596. (#40413) Thanks @jellyAI-dev and @vashquez. - Cron/context engine: run isolated cron jobs under run-scoped context-engine session keys so prior runs of the same job are not inherited unless the job is explicitly session-bound. (#72292) Thanks @jalehman. - Control UI: localize command palette labels, categories, skill shortcuts, footer hints, and connect-command copy labels while preserving localized command palette search matching. (#61130, #61119) Thanks @rubensfox20. - Plugins/memory-lancedb: request float embedding responses from OpenAI-compatible servers so local providers that default SDK requests to base64 no longer return dimension-mismatched LanceDB vectors while preserving configured dimensions. Fixes #45982. (#59048, #46069, #45986) Thanks @deep-introspection, @xiaokhkh, @caicongyang, and @thiswind. - Plugins/memory-lancedb: advance auto-capture cursors per session only after messages are processed or intentionally skipped, retry failed messages, survive compacted histories, and clear cursor state on session end. Fixes #71349; carries forward #42083. Thanks @as775116191. - Plugins/memory-core: respect configured memory-search embedding concurrency during non-batch indexing so local Ollama embedding backends can serialize indexing instead of flooding the server. Fixes #66822. (#66931) Thanks @oliviareid-svg and @LyraInTheFlesh. - Docker/update smoke: keep the package-derived update-channel fixture on package-shipped files and make its UI build stub create the asset the updater verifies. Thanks @vincentkoc. - Gateway/models: repair legacy `models.providers.*.api = "openai"` config values to `openai-completions`, and skip providers with future stale API enum values during startup instead of bricking the gateway. Fixes #72477. (#72542) Thanks @JooyoungChoi14 and @obviyus. - Gateway/skills: redact `apiKey` and secret-named `env` values from the `skills.update` RPC response to prevent leaking credentials into WebSocket traffic, client logs, or session transcripts. Config is still written to disk in full; only the response payload is redacted. (#69998) Thanks @Ziy1-Tan. - 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. - 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. - Telegram: send a fresh final message for long-lived preview-streamed replies so the visible Telegram timestamp reflects completion time instead of the preview creation time. Thanks @rubencu.