OpenClaw Release Highlights

2026.6.10 introduces /fast auto, letting short conversational turns start quickly while longer or fallback work automatically returns to normal mode — with the effective state still shown in status rather than collapsed to a simple on/off. Much of this release is the follow-through: keeping that automatic fast-mode timing, progress display, and reset behaviour consistent across model fallbacks, live model switches, connected-agent sessions, and scheduled cron runs, so the feature behaves the same whichever path a turn takes through it.

Provider routing gets several targeted fixes: Zhipu/GLM overload responses are now correctly treated as overloads so the right fallback engages, bundled Z.ai GLM-5 models no longer fall through to OpenAI with a misleading API-key error, and GLM-5.2's thinking levels expand beyond a binary on/off. On the channel side, a scheduled message's delivery outcome no longer gets lost on the next turn, and streamed progress updates stop dropping a repeated status that actually represented a separate step.

On the Gateway side, approval-sensitive tools stay protected when connected plugins or extensions change, and — the more consequential fix — authenticated package-source tokens are no longer sent to an allowed redirect on a different origin, while the legitimate redirected download still completes. Elsewhere, a durable session-transcript SDK contract lands for plugin authors, cross-channel direct-message sessions stop carrying over the wrong channel's identity after a switch, and Docker/Podman setup no longer runs unbounded on hosts where GNU timeout is installed as gtimeout.

The June 21 release pours its attention into Telegram, which becomes markedly richer to read. Replies now send proper HTML, preserve markdown formatting and sticker paths, render progress drafts and command output more faithfully, and normalise HTML tables safely — so a model's tabular answer arrives readable rather than mangled. Mentions and spooled message handlers stay on the correct delivery path. If Telegram is your primary channel, this is the most visible upgrade in months.

Agent reliability continues its steady climb. Retries, terminal outcomes, post-compaction usage accounting, session-history repair, and reply reconciliation all push interrupted or partial turns toward a visible final result instead of leaving them stranded. This is the kind of unglamorous resilience work you only notice by its absence — conversations that used to occasionally stall now finish.

The Codex integration gets stronger across the board: automatic plugin approvals, OAuth routing for GPT-5.3 Spark, remote-node exec exposed as a dynamic tool, and more dependable app-server teardown. Teams running delegated Codex workflows get a meaningfully smoother experience, with fewer cases where a helper failure leaves the runtime in a bad state.

Structurally, official provider packages are now first-class npm releases — externally installed channel plugins load at Gateway startup, and StepFun ships from both npm and ClawHub. The clients grow too: the Control UI adds a session workspace rail and extension-health reporting, iOS gains Apple Watch controls, and Android surfaces chat context. Codex Hosted Search arrives, key-free search providers stay deliberate opt-ins rather than surprise fallbacks, and ClawHub skill installs now retain verified source provenance so you can trust where a skill came from.

A broad reliability release with two clear themes: better-formatted messages and sturdier internal state. Telegram learns to render structured text properly — tables, lists, expandable blockquotes, and intentional line breaks all survive the trip — and WhatsApp now honours configured ACP bindings. New model support lands too, with GLM-5.2 and Claude Haiku 4.5 added to the catalog behind normalised provider IDs and safer OpenAI/Anthropic tool-schema recovery.

Agent runs get more dependable in the specific ways that bite in production: account-scoped direct messages, generated-media completions, restart shutdown aborts, and session-identity prompts all stay on the right recovery path. The /usage command gains a proper footer renderer with credential-aware limits and clear warnings, rather than silently producing bad output when a template breaks.

Underneath, memory and state grow more resilient — oversized embedding batches split before they trip request-size limits, search stays available in transient mode, SQLite avoids WAL on NFS volumes (a classic source of corruption), and full reindexes preserve rollback and cache recovery. Web search also gets more predictable: key-free providers like DuckDuckGo and Ollama remain explicit opt-ins instead of surprising automatic fallbacks.

This release is, above all, a security release. The project has drawn firmer lines around several places where content crosses trust boundaries: conversation transcripts, sandbox file binds, MCP servers speaking over stdio, and a number of individual integrations. The most consequential change for anyone running exec approvals is that they now fail closed — if an approval times out, the command does not run. The previous behaviour could let a timed-out approval fall through, which is exactly the kind of edge case you want resolved in the strict direction on a system that can execute commands.

Messaging reliability gets sustained attention. Telegram gains account-scoped topic routing and keeps streamed text intact through tool calls, so long replies no longer risk losing their thread when the assistant pauses mid-message to use a tool. iMessage — historically one of the trickier channels to keep healthy — picks up an always-on restart for inbound monitoring, durable echo markers so the assistant can reliably tell its own messages from yours, and a hardened outbound transport.

Browser automation can now attach to an existing session over CDP rather than always launching fresh, WebSocket connections are validated more strictly, and the boundary around browser output is safer. On performance, the Control UI starts faster and produces its first reply sooner thanks to cached model metadata and lazily loaded slash commands. Provider support keeps pace too: OpenRouter gets a proper OAuth onboarding flow, and Claude Fable 5 is supported including its adaptive thinking behaviour.

This release fixes a class of bug that corrupts conversations slowly: MCP tool results containing unexpected block types are now coerced properly, which stops Anthropic API 400 errors and — worse — poisoned session history that degrades every later reply. In the same spirit, extended-thinking sessions recover cleanly after the prompt cache expires, and QQBot strips the model's reasoning scaffolding before delivery so raw thinking tags stop leaking into channel replies.

Authentication gets more durable: auth profiles move into SQLite, and npm plugin installs keep their trusted pins across updates. Validation tightens around MCP lease timestamps, prompt-cache tool names, and provider catalogs. macOS users in node mode get a fix for an irritating behaviour where a healthy direct Gateway session would be silently abandoned in favour of a reconnect.

Elsewhere: Parallel ships as a bundled web-search provider with key discovery and onboarding, Google Vertex users on application-default credentials get their model catalog back, and Matrix gains voice-note preflight checks plus thread-aware read and reply behaviour. Upgrades are gentler too — legacy cron JSON stores migrate automatically during doctor preflight, and WhatsApp startup waits are bounded so a slow channel cannot stall the whole boot.

The theme of 2026.6.1 is graceful recovery. Agents and CLI-backed runtimes handle interrupted tool calls, stale session bindings, and compaction handoffs without losing their footing, and media deliveries retry sensibly. Channel reliability improves across the board — Telegram, WhatsApp, iMessage, Slack, Discord, Microsoft Teams, Google Chat, Google Meet, and iOS realtime Talk all get steadier delivery. A long list of timers, retries, OAuth lifetimes, and polling paths are now bounded, closing off ways a run could hang indefinitely.

The chat experience tightens noticeably: sends stay alive while history loads, streams render incrementally without redundant markdown work, drafts stay local while you type, and first-output latency is traced so slowness can be diagnosed rather than guessed at. Skill Workshop matures into a fuller Control UI flow with proposal lists, review states, and searchable file previews.

New surfaces arrive too — Workboard for orchestration, SecretRef plugin manifests, a hosted iOS push relay — alongside provider additions including MiniMax M3 and SQLite-backed model caching for OpenRouter. iMessage monitor state and plugin install ledgers move to SQLite, so restarts recover without rescanning the filesystem.