OpenClaw 2026.9.2 adds Astra, Muse Spark 1.3, and restart recovery

OpenClaw published 2026.9.2 on 5 September 2026. Treat it as a harness release: catalog support for GPT-6 Astra and Muse Spark 1.3, replies that survive Gateway restarts, faster long chats, and a shared-Gateway visibility default that operators should read before they upgrade.

GitHub published v2026.9.2 at 20:00 UTC. Official docs put the drop at 1,245 pull requests, six direct commits, and 232 contributors, on top of the 2026.9.1 baseline. Later that evening, @openclaw announced the same cut: restarts pick up where they left off, long chats move faster, Astra joins the catalog, and Muse Spark 1.3 is in as well.

Self-hosted OpenClaw is where many operators actually run agents. This tag changes model routing and multi-agent session access more than it re-announces either model.

Catalog routing, not a model launch

Operators can select openai/gpt-6-astra with an OpenAI API-key profile or an eligible ChatGPT/Codex account. GitHub’s notes list text and image input, Responses tool calls, and supported reasoning controls. Subscription availability follows successful account discovery—it does not mint access OpenAI has not already granted.

On OpenAI Platform API-key routes that use the built-in OpenClaw runtime and the official Responses endpoint, Astra can run direct function tools asynchronously and steer active responses over cached WebSockets. Request prefixes stay in place when reasoning effort changes under the documented configuration, and steering history plus encrypted reasoning survive continuations. Docs describe the operator-facing version: on supported OpenAI API connections, you can correct ongoing work without waiting for Astra to finish. Reasoning and sampling settings stay correct when catalog metadata is absent. /think ultra orchestration is available for OpenClaw and native Codex runtimes, and configured Azure deployment capabilities are retained.

Muse Spark 1.3 is the second catalog addition. Docs pair it with Astra as Meta’s model. Both additions support text and images for accounts that already have access. The X post matches that pairing. Neither entry is a substitute for checking whether your provider account can actually call the model.

People sharing a Gateway can connect their own model accounts without changing everyone else’s default.

Recovery, history off the event loop

The runtime work is the rest of the story.

Automatic updates keep active settings, enabled skills, and default-agent ownership. Gateway restarts after Git updates are restored, and chat-triggered updates report success, failure, or an intentional skip with owner-setup and restart-recovery guidance. Docs add that update reports survive reconnects, eligible interrupted tasks resume after a restart, and completed answers remain available through specific saving failures.

Replies survive restarts. Active, queued, and delegated replies recover without one completed reply discarding another’s recovery marker. Continuation instructions persist through compaction and retry attempts.

Chat, dashboards, and session interactions stay responsive while long transcripts and disk usage are processed. Durable history reads and cold transcript integrity checks run outside the Gateway event loop. Opening a long conversation prioritizes the selected messages, with smaller initial transfers and less cold-load work. Cancellation stays effective throughout history preparation.

More agent, model, tool, channel, browser, node, access, and terminal settings apply through their running owners. Settings that still require a Gateway restart remain marked in the configuration reference, so unaffected conversations can stay connected.

Git backups now preserve complete text that contains embedded NUL characters, support Nix-managed config and credential links, and reject corrupt archive headers instead of accepting an incomplete backup.

Swarm default and the shared-Gateway warning

Swarm sub-agents are on by default. Concurrent sub-agents return structured results and live progress. Explicit opt-outs, tool restrictions, and the separate Code Mode opt-in still apply.

The session-visibility change is the one that can surprise a multi-agent host. Session tools now default to all-session visibility, and ordinary agent-to-agent access is enabled. Omitted settings let agents with session tools read and search other agents’ conversations on a shared Gateway, including other users’ transcripts. Set tools.sessions.visibility to agent or self for narrower access. Existing tool and sandbox restrictions remain. Docs are explicit: mutually untrusted users need separate Gateways. Review conversation access before you upgrade.

Task workspace and local-model setup

The browser workspace can put a dashboard, browser, terminal, or file view at the center of a task and keep the conversation beside it. Layout choices persist in that browser. Saved dashboards sit in a visual gallery you can search and preview. Mentions of authorized people create temporary Inbox entries, with optional browser notifications.

Guided local-model setup starts with the computer that will run the model. Managed llama.cpp setup recommends a recipe from available memory, graphics hardware, and disk, shows the machine name and download size, and checks both a response and a real file-read tool result before changing the default. A failed or cancelled activation leaves the previous model selected. The new recipes are text-only. Local-model recommendations begin with an 8 GiB memory floor for the smallest recipe. CPU-only replies can still take several minutes. On a Gateway with several agents, setting up one local agent can still change shared lean-tool settings used by others, so review those after targeted setup.

Plugin icons now load from a packaged assets/icon.png without a network request. Plugin SDK context aliases that use untrusted-named prompt-context fields become eligible for removal on or after 8 September 2026.

Read the 2026.9.2 notes, set tools.sessions.visibility on every shared Gateway before you upgrade, then enable Astra or Muse Spark 1.3 only on accounts that already qualify.