OpenClaw v2026.4.15

### Changes - Anthropic/models: default Anthropic selections, `opus` aliases, Claude CLI defaults, and bundled image understanding to Claude Opus 4.7. - Google/TTS: add Gemini text-to-speech support to the bundled `google` plugin, including provider registration, voice selection, WAV reply output, PCM telephony output, and setup/docs guidance. (#67515) Thanks @barronlroth. - Control UI/Overview: add a Model Auth status card showing OAuth token health and provider rate-limit pressure at a glance, with attention callouts when OAuth tokens are expiring or expired. Backed by a new `models.authStatus` gateway method that strips credentials and caches for 60s. (#66211) Thanks @omarshahine. - Memory/LanceDB: add cloud storage support to `memory-lancedb` so durable memory indexes can run on remote object storage instead of local disk only. (#63502) Thanks @rugvedS07. - GitHub Copilot/memory search: add a GitHub Copilot embedding provider for memory search, and expose a dedicated Copilot embedding host helper so plugins can reuse the transport while honoring remote overrides, token refresh, and safer payload validation. (#61718) Thanks @feiskyer and @vincentkoc. - Agents/local models: add experimental `agents.defaults.experimental.localModelLean: true` to drop heavyweight default tools like `browser`, `cron`, and `message`, reducing prompt size for weaker local-model setups without changing the normal path. (#66495) Thanks @ImLukeF. - Packaging/plugins: localize bundled plugin runtime deps to their owning extensions, trim the published docs payload, and tighten install/package-manager guardrails so published builds stay leaner and core stops carrying extension-owned runtime baggage. (#67099) Thanks @vincentkoc. - QA/Matrix: split Matrix live QA into a source-linked `qa-matrix` runner and keep repo-private `qa-*` surfaces out of packaged and published builds. (#66723) Thanks @gumadeiras. - Docs/showcase: add a scannable hero, complete section jump links, and a responsive video grid for community examples. (#48493) Thanks @jchopard69. ### Fixes - Gateway/tools: anchor trusted local `MEDIA:` tool-result passthrough on the exact raw name of this run's registered built-in tools, and reject client tool definitions whose names normalize-collide with a built-in or with another client tool in the same request (`400 invalid_request_error` on both JSON and SSE paths), so a client-supplied tool named like a built-in can no longer inherit its local-media trust. (#67303) - Agents/replay recovery: classify the provider wording `401 input item ID does not belong to this connection` as replay-invalid, so users get the existing `/new` session reset guidance instead of a raw 401-style failure. (#66475) Thanks @dallylee. - Gateway/webchat: enforce localRoots containment on webchat audio embedding path [AI-assisted]. (#67298) Thanks @pgondhi987. - Matrix/pairing: block DM pairing-store entries from authorizing room control commands [AI-assisted]. (#67294) Thanks @pgondhi987. - Docker/build: verify `@matrix-org/matrix-sdk-crypto-nodejs` native bindings with `find` under `node_modules` instead of a hardcoded `.pnpm/...` path so pnpm v10+ virtual-store layouts no longer fail the image build. (#67143) thanks @ly85206559. - Matrix/E2EE: keep startup bootstrap conservative for passwordless token-auth bots, still attempt the guarded repair pass without requiring `channels.matrix.password`, and document the remaining password-UIA limitation. (#66228) Thanks @SARAMALI15792. - Cron/announce delivery: suppress mixed-content isolated cron announce replies that end with `NO_REPLY` so trailing silent sentinels no longer leak summary text to the target channel. (#65004) thanks @neo1027144-creator. - Plugins/bundled channels: partition bundled channel lazy caches by active bundled root so `OPENCLAW_BUNDLED_PLUGINS_DIR` flips stop reusing stale plugin, setup, secrets, and runtime state. (#67200) Thanks @gumadeiras. - Packaging/plugins: prune common test/spec cargo from bundled plugin runtime dependencies and fail npm release validation if packaged test cargo reappears, keeping published tarballs leaner without plugin-specific special cases. (#67275) thanks @gumadeiras. - Agents/context + Memory: trim default startup/skills prompt budgets, cap `memory_get` excerpts by default with explicit continuation metadata, and keep QMD reads aligned with the same bounded excerpt contract so long sessions pull less context by default without losing deterministic follow-up reads. - Matrix/commands: skip DM pairing-store reads on room traffic now that room control-command authorization ignores pairing-store entries, keeping the room path narrower without changing room auth behavior. (#67325) Thanks @gumadeiras. - Memory-core/dreaming: skip dreaming narrative transcripts from session-store metadata before bootstrap records land so dream diary prompt/prose lines do not pollute session ingestion. (#67315) thanks @jalehman. - Agents/local models: clarify low-context preflight hints for self-hosted models, point config-backed caps at the relevant OpenClaw setting, and stop suggesting larger models when `agents.defaults.contextTokens` is the real limit. (#66236) Thanks @ImLukeF. - Dreaming/memory-core: change the default `dreaming.storage.mode` from `inline` to `separate` so Dreaming phase blocks (`## Light Sleep`, `## REM Sleep`) land in `memory/dreaming/{phase}/YYYY-MM-DD.md` instead of being injected into `memory/YYYY-MM-DD.md`. Daily memory files no longer get dominated by structured candidate output, and the daily-ingestion scanner that already strips dream marker blocks no longer has to compete with hundreds of phase-block lines on every run. Operators who want the previous behavior can opt in by setting `plugins.entries.memory-core.config.dreaming.storage.mode: "inline"`. (#66412) Thanks @mjamiv. - Control UI/Overview: fix false-positive "missing" alerts on the Model Auth status card for aliased providers, env-backed OAuth with auth.profiles, and unresolvable env SecretRefs. (#67253) Thanks @omarshahine. - Dashboard: constrain exec approval modal overflow on desktop so long command content no longer pushes action buttons out of view. (#67082) Thanks @Ziy1-Tan. - Agents/CLI transcripts: persist successful CLI-backed turns into the OpenClaw session transcript so google-gemini-cli replies appear in session history and the Control UI again. (#67490) Thanks @obviyus. - Discord/tool-call text: strip standalone Gemma-style `<function>...</function>` tool-call payloads from visible assistant text without truncating prose examples or trailing replies. (#67318) Thanks @joelnishanth. - WhatsApp/web-session: drain the pending per-auth creds save queue before reopening sockets so reconnect-time auth bootstrap no longer races in-flight `creds.json` writes and falsely restores from backup. (#67464) Thanks @neeravmakwana. - BlueBubbles/catchup: add a per-message retry ceiling (`catchup.maxFailureRetries`, default 10) so a persistently-failing message with a malformed payload no longer wedges the catchup cursor forever. After N consecutive `processMessage` failures against the same GUID, catchup logs a WARN, skips that message on subsequent sweeps, and lets the cursor advance past it. Transient failures still retry from the same point as before. Also fixes a lost-update race in the persistent dedupe file lock that silently dropped inbound GUIDs on concurrent writes, a dedupe file naming migration gap on version upgrade, and a balloon-event bypass that let catchup replay debouncer-coalesced events as standalone messages. (#67426, #66870) Thanks @omarshahine. - Ollama/chat: strip the `ollama/` provider prefix from Ollama chat request model ids so configured refs like `ollama/qwen3:14b-q8_0` stop 404ing against the Ollama API. (#67457) Thanks @suboss87. - Agents/tools: resolve non-workspace host tilde paths against the OS home directory and keep edit recovery aligned with that same path target, so `~/...` host edit/write operations stop failing or reading back the wrong file when `OPENCLAW_HOME` differs. (#62804) Thanks @stainlu. - Speech/TTS: auto-enable the bundled Microsoft and ElevenLabs speech providers, and route generic TTS directive tokens through the explicit or active provider first so overrides like `[[tts:speed=1.2]]` stop silently landing on the wrong provider. (#62846) Thanks @stainlu. - OpenAI Codex/models: normalize stale native transport metadata in both runtime resolution and discovery/listing so legacy `openai-codex` rows with missing `api` or `https://chatgpt.com/backend-api/v1` self-heal to the canonical Codex transport instead of routing requests through broken HTML/Cloudflare paths, combining the original fixes proposed in #66969 (saamuelng601-pixel) and #67159 (hclsys). (#67635) - Agents/failover: treat HTML provider error pages as upstream transport failures for CDN-style 5xx responses without misclassifying embedded body text as API rate limits, while still preserving auth remediation for HTML 401/403 pages and proxy remediation for HTML 407 pages. (#67642) Thanks @stainlu. - Gateway/skills: bump the cached skills-snapshot version whenever a config write touches `skills.*` (for example `skills.allowBundled`, `skills.entries.<id>.enabled`, or `skills.profile`). Existing agent sessions persist a `skillsSnapshot` in `sessions.json` that reuses the skill list frozen at session creation; without this invalidation, removing a bundled skill from the allowlist left the old snapshot live and the model kept calling the disabled tool, producing `Tool <name> not found` loops that ran until the embedded-run timeout. (#67401) Thanks @xantorres. - Agents/tool-loop: enable the unknown-tool stream guard by default. Previously `resolveUnknownToolGuardThreshold` returned `undefined` unless `tools.loopDetection.enabled` was explicitly set to `true`, which left the protection off in the default configuration. A hallucinated or removed tool (for example `himalaya` after it was dropped from `skills.allowBundled`) would then loop "Tool X not found" attempts until the full embedded-run timeout. The guard has no false-positive surface because it only triggers on tools that are objectively not registered in the run, so it now stays on regardless of `tools.loopDetection.enabled` and still accepts `tools.loopDetection.unknownToolThreshold` as a per-run override (default 10). (#67401) Thanks @xantorres. - TUI/streaming: add a client-side streaming watchdog to `tui-event-handlers` so the `streaming · Xm Ys` activity indicator resets to `idle` after 30s of delta silence on the active run. Guards against lost or late `state: "final"` chat events (WS reconnects, gateway restarts, etc.) leaving the TUI stuck on `streaming` indefinitely; a new system log line surfaces the reset so users know to send a new message to resync. The window is configurable via the new `streamingWatchdogMs` context option (set to `0` to disable), and the handler now exposes a `dispose()` that clears the pending timer on shutdown. (#67401) Thanks @xantorres. - Extensions/lmstudio: add exponential backoff to the inference-preload wrapper so an LM Studio model-load failure (for example the built-in memory guardrail rejecting a load because the swap is saturated) no longer produces a WARN line every ~2s for every chat request. The wrapper now records consecutive preload failures per `(baseUrl, modelKey, contextLength)` tuple with a 5s → 10s → 20s → … → 5min cooldown and skips the preload step entirely while a cooldown is active, letting chat requests proceed directly to the stream (the model is often already loaded via the LM Studio UI). The combined `preload failed` log line now reports consecutive-failure count and remaining cooldown so operators can act on the real issue instead of drowning in repeated warnings. (#67401) Thanks @xantorres. - Agents/replay: re-run tool/result pairing after strict replay tool-call ID sanitization on outbound requests so Anthropic-compatible providers like MiniMax no longer receive malformed orphan tool-result IDs such as `...toolresult1` during compaction and retry flows. (#67620) Thanks @stainlu. - Gateway/startup: fix spurious SIGUSR1 restart loop on Linux/systemd when plugin auto-enable is the only startup config write; the config hash guard was not captured for that write path, causing chokidar to treat each boot write as an external change and trigger a reload → restart cycle that corrupts manifest.db after repeated cycles. Fixes #67436. (#67557) thanks @openperf - Codex/harness: auto-enable the Codex plugin when `codex` is selected as an embedded agent harness runtime, including forced default, per-agent, and `OPENCLAW_AGENT_RUNTIME` paths. (#67474) Thanks @duqaXxX. - OpenAI Codex/CLI: keep resumed `codex exec resume` runs on the safe non-interactive path without reintroducing the removed dangerous bypass flag by passing the supported `--skip-git-repo-check` resume arg plus Codex's native `sandbox_mode="workspace-write"` config override. (#67666) Thanks @plgonzalezrx8. - Codex/app-server: parse Desktop-originated app-server user agents such as `Codex Desktop/0.118.0`, keeping the version gate working when the Codex CLI inherits a multi-word originator. (#64666) Thanks @cyrusaf. - Cron/announce delivery: keep isolated announce `NO_REPLY` stripping case-insensitive across direct and text delivery, preserve structured media-only sends when a caption strips silent, and derive main-session awareness from the cleaned payloads so silent captions no longer leak stale `NO_REPLY` text. (#65016) Thanks @BKF-Gitty. - Sessions/Codex: skip redundant `delivery-mirror` transcript appends only when the latest assistant message has the same visible text, preventing duplicate visible replies on Codex-backed turns without suppressing repeated answers across turns. (#67185) Thanks @andyylin. - Auto-reply/prompt-cache: keep volatile inbound chat IDs out of the stable system prompt so task-scoped adapters can reuse prompt caches across runs, while preserving conversation metadata for the user turn and media-only messages. (#65071) Thanks @MonkeyLeeT. - BlueBubbles/inbound: restore inbound image attachment downloads on Node 22+ by stripping incompatible bundled-undici dispatchers from the non-SSRF fetch path, accept `updated-message` webhooks carrying attachments, use event-type-aware dedup keys so attachment follow-ups are not rejected as duplicates, and retry attachment fetch from the BB API when the initial webhook arrives with an empty array. (#64105, #61861, #65430, #67510) Thanks @omarshahine. - Agents/skills: sort prompt-facing `available_skills` entries by skill name after merging sources so `skills.load.extraDirs` order no longer changes prompt-cache prefixes. (#64198) Thanks @Bartok9. - Agents/OpenAI Responses: add `models.providers.*.models.*.compat.supportsPromptCacheKey` so OpenAI-compatible proxies that forward `prompt_cache_key` can keep prompt caching enabled while incompatible endpoints can still force stripping. (#67427) Thanks @damselem. - Agents/context engines: keep loop-hook and final `afterTurn` prompt-cache touch metadata aligned with the current assistant turn so cache-aware context engines retain accurate cache TTL state during tool loops. (#67767) thanks @jalehman. - Memory/dreaming: strip AI-facing inbound metadata envelopes from session-corpus user turns before normalization so REM topic extraction sees the user's actual message text, including array-shaped split envelopes. (#66548) Thanks @zqchris. - Agents/errors: detect standalone Cloudflare/CDN HTML challenge pages before transport DNS classification so provider block pages no longer appear as local DNS lookup failures. (#67704) Thanks @chris-yyau. - Security/approvals: redact secrets in exec approval prompts so inline approval review can no longer leak credential material in rendered prompt content. (#61077, #64790) - CLI/configure: re-read the persisted config hash after writes so config updates stop failing with stale-hash races. (#64188, #66528) - CLI/update: prune stale packaged `dist` chunks after npm upgrades and keep downgrade/verify inventory checks compat-safe so global upgrades stop failing on stale chunk imports. (#66959) Thanks @obviyus. - Onboarding/CLI: fix channel-selection crashes on globally installed CLI setups during onboarding. (#66736) - Video generation/live tests: bound provider polling for live video smoke, default to the fast non-FAL text-to-video path, and use a one-second lobster prompt so release validation no longer waits indefinitely on slow provider queues. - Memory-core/QMD `memory_get`: reject reads of arbitrary workspace markdown paths and only allow canonical memory files (`MEMORY.md`, `memory.md`, `DREAMS.md`, `dreams.md`, `memory/**`) plus exact paths of active indexed QMD workspace documents, so the QMD memory backend can no longer be used as a generic workspace-file read shim that bypasses `read` tool-policy denials. (#66026) Thanks @eleqtrizit. - Cron/agents: forward embedded-run tool policy and internal event params into the attempt layer so `--tools` allowlists, cron-owned message-tool suppression, explicit message targeting, and command-path internal events all take effect at runtime again. (#62675) Thanks @hexsprite. - Setup/providers: guard preferred-provider lookup during setup so malformed plugin metadata with a missing provider id no longer crashes the wizard with `Cannot read properties of undefined (reading 'trim')`. (#66649) Thanks @Tianworld. - Matrix/security: normalize sandboxed profile avatar params, preserve `mxc://` avatar URLs, and surface gmail watcher stop failures during reload. (#64701) Thanks @slepybear. - Telegram/documents: drop leaked binary caption bytes from inbound Telegram text handling so document uploads like `.mobi` or `.epub` no longer explode prompt token counts. (#66663) Thanks @joelnishanth. - Gateway/auth: resolve the active gateway bearer per-request on the HTTP server and the HTTP upgrade handler via `getResolvedAuth()`, mirroring the WebSocket path, so a secret rotated through `secrets.reload` or config hot-reload stops authenticating on `/v1/*`, `/tools/invoke`, plugin HTTP routes, and the canvas upgrade path immediately instead of remaining valid on HTTP until gateway restart. (#66651) Thanks @mmaps. - Agents/compaction: cap the compaction reserve-token floor to the model context window so small-context local models (e.g. Ollama with 16K tokens) no longer trigger context-overflow errors or infinite compaction loops on every prompt. (#65671) Thanks @openperf. - Agents/OpenAI Responses: classify the exact `Unknown error (no error details in response)` transport failure as failover reason `unknown` so assistant/model fallback still runs for that no-details failure path. (#65254) Thanks @OpenCodeEngineer. - Models/probe: surface invalid-model probe failures as `format` instead of `unknown` in `models list --probe`, and lock the invalid-model fallback path in with regression coverage. (#50028) Thanks @xiwuqi. - Agents/failover: classify OpenAI-compatible `finish_reason: network_error` stream failures as timeout so model fallback retries continue instead of stopping with an unknown failover reason. (#61784) thanks @lawrence3699. - Onboarding/channels: normalize channel setup metadata before discovery and validation so malformed or mixed-shape channel plugin metadata no longer breaks setup and onboarding channel lists. (#66706) Thanks @darkamenosa. - Slack/native commands: fix option menus for slash commands such as `/verbose` when Slack renders native buttons by giving each button a unique action ID while still routing them through the shared `openclaw_cmdarg*` listener. Thanks @Wangmerlyn. - Feishu/webhook: harden the webhook transport and card-action replay guards to fail closed on missing `encryptKey` and blank callback tokens — refuse to start the webhook transport without an `encryptKey`, reject unsigned requests when no key is present instead of accepting them, and drop blank card-action tokens before the dedupe claim and dispatcher. Defense-in-depth over the already-closed monitor-account layer. (#66707) Thanks @eleqtrizit. - Agents/workspace files: route `agents.files.get`, `agents.files.set`, and workspace listing through the shared `fs-safe` helpers (`openFileWithinRoot`/`readFileWithinRoot`/`writeFileWithinRoot`), reject symlink aliases for allowlisted agent files, and have `fs-safe` resolve opened-file real paths from the file descriptor before falling back to path-based `realpath` so a symlink swap between `open` and `realpath` can no longer redirect the validated path off the intended inode. (#66636) Thanks @eleqtrizit. - Gateway/MCP loopback: switch the `/mcp` bearer comparison from plain `!==` to constant-time `safeEqualSecret` (matching the convention every other auth surface in the codebase uses), and reject non-loopback browser-origin requests via `checkBrowserOrigin` before the auth gate runs. Loopback origins (`127.0.0.1:*`, `localhost:*`, same-origin) still go through, including the `localhost`↔`127.0.0.1` host mismatch that browsers flag as `Sec-Fetch-Site: cross-site`. (#66665) Thanks @eleqtrizit. - Auto-reply/billing: classify pure billing cooldown fallback summaries from structured fallback reasons so users see billing guidance instead of the generic failure reply. (#66363) Thanks @Rohan5commit. - Agents/fallback: preserve the original prompt body on model fallback retries with session history so the retrying model keeps the active task instead of only seeing a generic continue message. (#66029) Thanks @WuKongAI-CMU. - Reply/secrets: resolve active reply channel/account SecretRefs before reply-run message-action discovery so channel token SecretRefs (for example Discord) do not degrade into discovery-time unresolved-secret failures. (#66796) Thanks @joshavant. - Agents/Anthropic: ignore non-positive Anthropic Messages token overrides and fail locally when no positive token budget remains, so invalid `max_tokens` values no longer reach the provider API. (#66664) thanks @jalehman - Agents/context engines: preserve prompt-only token counts, not full request totals, when deferred maintenance reuses after-turn runtime context so background compaction bookkeeping matches the active prompt window. (#66820) thanks @jalehman. - BlueBubbles/inbound: add a persistent file-backed GUID dedupe so MessagePoller webhook replays after BB Server restart or reconnect no longer cause the agent to re-reply to already-handled messages. (#19176, #12053, #66816) Thanks @omarshahine. - Secrets/plugins/status: align SecretRef inspect-vs-strict handling across plugin preload, read-only status/agents surfaces, and runtime auth paths so unresolved refs no longer crash read-only CLI flows while runtime-required non-env refs stay strict. (#66818) Thanks @joshavant. - Memory/dreaming: stop ordinary transcripts that merely quote the dream-diary prompt from being classified as internal dreaming runs and silently dropped from session recall ingestion. (#66852) Thanks @gumadeiras. - Telegram/documents: sanitize binary reply context and ZIP-like archive extraction so `.epub` and `.mobi` uploads can no longer leak raw binary into prompt context through reply metadata or archive-to-`text/plain` coercion. (#66877) Thanks @martinfrancois. - Telegram/native commands: restore plugin-registry-backed auto defaults for native commands and native skills so Telegram slash commands keep registering when `commands.native` and `commands.nativeSkills` stay on `auto`. (#66843) Thanks @kashevk0. - OpenRouter/Qwen3: parse `reasoning_details` stream deltas as thinking content without skipping same-chunk tool calls, so Qwen3 replies no longer fail empty on OpenRouter and mixed reasoning/tool-call chunks still execute normally. (#66905) Thanks @bladin. - BlueBubbles/catchup: replay missed webhook messages after gateway restart via a persistent per-account cursor and `/api/v1/message/query?after=<ts>` pass, so messages delivered while the gateway was down no longer disappear. Uses the existing `processMessage` path and is deduped by #66816's inbound GUID cache. (#66857, #66721) Thanks @omarshahine. - Telegram/native commands: keep Telegram command-sync cache process-local so gateway restarts re-register the menu instead of trusting stale on-disk sync state after Telegram cleared commands out-of-band. (#66730) Thanks @nightq. - Audio/self-hosted STT: restore `models.providers.*.request.allowPrivateNetwork` for audio transcription so private or LAN speech-to-text endpoints stop tripping SSRF blocks after the v2026.4.14 regression. (#66692) Thanks @jhsmith409. - Auto-reply/media: allow workspace-rooted absolute media paths in auto-reply send flows so valid local media references no longer fail path validation. (#66689) - WhatsApp/Baileys media upload: harden encrypted upload handling so large outbound media sends avoid buffer spikes and reliability regressions. (#65966) Thanks @frankekn. - QQBot/cron: guard against undefined `event.content` in `parseFaceTags` and `filterInternalMarkers` so cron-triggered agent turns with no content payload no longer crash with `TypeError: Cannot read properties of undefined (reading 'startsWith')`. (#66302) Thanks @xinmotlanthua. - CLI/plugins: stop `--dangerously-force-unsafe-install` plugin installs from falling back to hook-pack installs after security scan failures, while still preserving non-security fallback behavior for real hook packs. (#58909) Thanks @hxy91819. - Claude CLI/sessions: classify `No conversation found with session ID` as `session_expired` so expired CLI-backed conversations clear the stale binding and recover on the next turn. (#65028) thanks @Ivan-Fn. - Context Engine: gracefully fall back to the legacy engine when a third-party context engine plugin fails at resolution time (unregistered id, factory throw, or contract violation), preventing a full gateway outage on every channel. (#66930) Thanks @openperf. - Control UI/chat: keep optimistic user message cards visible during active sends by deferring same-session history reloads until the active run ends, including aborted and errored runs. (#66997) Thanks @scotthuang and @vincentkoc. - Media/Slack: allow host-local CSV and Markdown uploads only when the fallback buffer actually decodes as text, so real plain-text files work without letting opaque non-text blobs renamed to `.csv` or `.md` slip past the host-read guard. (#67047) Thanks @Unayung. - Ollama/onboarding: split setup into `Cloud + Local`, `Cloud only`, and `Local only`, support direct `OLLAMA_API_KEY` cloud setup without a local daemon, and keep Ollama web search on the local-host path. (#67005) Thanks @obviyus. - Webchat/security: reject remote-host `file://` URLs in the media embedding path. (#67293) Thanks @pgondhi987. - Dreaming/memory-core: use the ingestion day, not the source file day, for daily recall dedupe so repeat sweeps of the same daily note can increment `dailyCount` across days instead of stalling at `1`. (#67091) Thanks @Bartok9. - Node-host/tools.exec: let approval binding distinguish known native binaries from mutable shell payload files, while still fail-closing unknown or racy file probes so absolute-path node-host commands like `/usr/bin/whoami` no longer get rejected as unsafe interpreter/runtime commands. (#66731) Thanks @tmimmanuel.