OpenClaw v2026.9.2 Stable Is Out. It Bakes GPT-6 Astra Support In, Turns Swarm On By Default, Moves Chat Off the Gateway Event Loop, and Fixes the Worst of the Recovery Story

The OpenClaw team shipped v2026.9.2 stable at 2026-09-06 02:15 UTC. This is the second stable of the 2026.9 series. v2026.9.1 stable (covered Sep 3) shipped Mermaid diagrams, the one-prompt install lane, and per-identity skill libraries. v2026.9.2 stable is the operational follow-up the v2026.9.1 release notes gestured at but did not deliver: chat performance, reliable updates, recovery, and the GPT-6 Astra plumbing builders have been waiting for.

This is a documentation-surfacing report. Every claim below comes verbatim from the official v2026.9.2 release notes on GitHub. The limitations section lists what was not tested.

What Shipped

Six Highlights sections and twenty-two Changes and Fixes items verified verbatim from the v2026.9.2 release page (GitHub, fetched 2026-09-06 14:10 UTC).

Highlights

1. Faster, more responsive chat. "Keep chat, dashboards, and session interactions responsive while long transcripts and disk usage are processed, with direct dashboard lookup, less cold-load work, and durable history reads outside the Gateway event loop." (PRs #136862, #138094, #138669, #138888, #138860, #138894) 2. Reliable upgrades and recovery. "Keep active settings, enabled skills, and default-agent ownership in automatic updates, restore Gateway restarts after Git updates, and report outcomes with actionable recovery guidance." Related issue #138760. (PRs #138837, #138730, #138781, #136588, #136995) 3. GPT-6 Astra support. "Select openai/gpt-6-astra with an OpenAI API-key profile or an eligible ChatGPT/Codex account, with text and image input, Responses tool calls, and supported reasoning controls; subscription availability follows successful account discovery." Related issue #137549. (PRs #137550, #137561) 4. Replies survive restarts. "Recover active, queued, and delegated replies after Gateway restarts without letting one completed reply discard another's recovery marker, and keep continuation instructions through compaction and retry attempts." Related issues #138518, #138564. (PRs #138071, #137606, #136236, #138519, #138565) 5. Backups that preserve your data. "Preserve complete text containing embedded NUL characters in Git backups, support Nix-managed config and credential links, and reject corrupt archive headers instead of accepting an incomplete backup." Related issues #138306, #136326. (PRs #138327, #136343, #137718) 6. Change settings without restarting. "Apply more agent, model, tool, channel, browser, node, access, and terminal settings through their running owners; settings that still require a Gateway restart remain marked in the configuration reference." (PRs #138112, #137790, #137412, #137160, #136832)

Changes

GPT-6 Astra plumbing (built on the OpenAI Responses API long-running-work controls covered Sep 3):

• GPT-6 Astra async tools and steering. "On OpenAI Platform API-key routes using the built-in OpenClaw runtime and official Responses endpoint, run direct function tools asynchronously, steer active responses over cached WebSockets, and retain request prefixes when changing reasoning effort under the documented configuration; keep steering history and encrypted reasoning through continuations." (PRs #138046, #138434)
• GPT-6 Astra reasoning. "Keep supported reasoning and sampling settings correct when catalog metadata is absent, add /think ultra orchestration for OpenClaw and native Codex runtimes, and retain configured Azure deployment capabilities." (PRs #138626, #138739)

Default-on changes worth pinning:

• Swarm is enabled by default. "Orchestrate concurrent sub-agents with structured results and live progress, while preserving explicit opt-outs, tool restrictions, and the separate Code Mode opt-in." Related issue #136472. (PRs #136514, #138056)
• Cross-agent session access. "Session tools now default to all-session visibility and ordinary agent-to-agent access is enabled; set tools.sessions.visibility to agent or self for narrower session access, and retain existing tool and sandbox restrictions." (PR #136755)

Smaller Changes:

• Plugin branding. Package the plugin icon at assets/icon.png instead of a top-level manifest URL; OpenClaw loads the packaged image without a network request, and missing or invalid icons do not prevent the plugin from loading. (PR #131510)
• macOS browser sign-in. Connect saved Gateways through personal Cloudflare Access browser sign-in, open the Mac app from the Gateway website, and restore the selected Gateway after app restart while keeping accounts and queued messages separate. Related issues #138698, #137916. (PRs #138745, #137933)
• Experimental plugin UI. Enable Settings → Labs → Custom plugin UI to let plugins contribute Control UI pages, panels, and session actions or customize the composer and workspace. (PR #134943)
• Experimental Apple Watch Talk. Opt into standalone Talk on Watch with Gateway-owned tools and transcripts while retaining the existing iPhone-mediated Talk to Claw flow; standalone realtime audio requires UDP connectivity. (PR #135808)
• Slack rich replies. Guide agents to use supported Block Kit layouts for proactive Slack replies. (PR #138896)
• Personal connected accounts. Add provider accounts from Settings → Profile → Connected accounts or the CLI, choose an account for new chats or one existing chat, and keep existing session selections when your default changes. (PR #134970)
• Dashboards fit your task. Browse saved dashboards in a gallery, swap chat and dashboard views, place panels left, right, or below, and keep loaded widget input when hiding and reopening a panel. (PRs #137069, #137068, #138077)
• Discord meeting notes. Opt into occupancy-driven, listen-only meeting capture with automatic notes and a Control UI Meetings page. (PR #136679)
• Bring teammates into the conversation. Mention authorized people from the composer and receive temporary Inbox entries, with optional browser notifications. Related issue #135645. (PR #135853)
• Devices at a glance. Show device types, resource meters, capability chips, and Desktop availability in the Devices page, with native iOS CPU/memory and Android memory/disk reporting. (PRs #136858, #137082, #137083)
• Slack session controls. Add a native Stop button, session processing and approval status, and synchronized session titles; enable the documented Slack session-event subscriptions to receive Stop and title changes. (PR #136706)
• macOS approvals. Simplify command, node, and device approval panels so requested access and administrator warnings are easier to review. (PRs #138049, #138137)
• Native text copying. Select message text on iOS and copy code blocks directly on iOS and macOS. (PR #136729)
• Realtime Talk. Support native realtime sessions for thin audio clients while retaining the selected model and account, transcripts, spoken controls, and scoped agent consultations. (PR #134003)
• Cloud worker checkouts. Fast-forward replacement cloud workers to the pushed session branch while preserving local modifications, deliberate deletions, and divergent branches. Related issue #137011. (PR #137392)
• Owner profiles. Give single-user Gateway connections a durable Owner profile for identity and personal preferences without granting additional permissions. (PR #136819)
• Provider account priority. Expose configured provider account priority through the Gateway. (PR #132450)
• Gateway latency diagnostics. Break RPC latency into phases in OpenTelemetry and Prometheus diagnostics so slow dispatch and execution are easier to distinguish. (PR #138015)

Fixes

• HTTP proxy safety. Update HTTP dependencies and keep destination DNS checks in effect when requests bypass a configured proxy. (PR #139056)
• Doctor database diagnostics. Report the concrete SQLite error for a stable malformed database instead of incorrectly blaming concurrent changes, while preserving the source database and its sidecars. (PR #139136)
• Plugin SDK code ranges. Accept positional CodeRegion inputs containing only start and end again; consumers needing parser metadata should retain the inferred findCodeRegions result. Related issue #137509. (PR #137510)
• Transcript rewrites. Leave live history intact when a durable tail rewrite fails and retain the admitted user input through repeated branch rewrites. (PRs #138412, #138967)
• Session cleanup. Preserve other agents' plugin state on session deletion, keep case-sensitive peer identifiers distinct, and prevent canceled queued cleanup from erasing newly restored plugin state. Related issue #121984. (PRs #138925, #122006, #138957)
• Skill Workshop backups. Refuse incomplete directory snapshots and verify retained originals before restore so deep user edits are not overwritten by an incomplete view. (PR #138082)
• Pending child launches. Stop pending native and ACP child launches when their originating parent authority closes, while preserving ownership already transferred to accepted children. Related issue #139001. (PR #139020)
• Provider sign-in. Keep the active plugin registry intact while discovering provider sign-in methods. (PR #138838)
• Automation replies. Keep automation and heartbeat completions attached to their originating conversation and topic while retaining the isolated execution session and its policy. (PR #133323)
• Configured agent hooks. Retain configured hooks when a turn selects a non-default model. Related issue #138368. (PR #138979)
• Channel reloads. Keep channel start alive when the channel startup task is canceled by Gateway shutdown or a node restart. (truncated; see source for full list)

What Actually Changed For Builders

Five things in this release move the day-to-day experience of running an OpenClaw Gateway in production.

Chat and dashboard work no longer competes with the Gateway event loop

v2026.9.2 takes durable history reads and dashboard lookup off the Gateway event loop. That is the line that was biting production users: while a long transcript was being scanned for context, the Gateway itself was busy, and chat, dashboards, and session interactions stuttered. v2026.9.2 splits that work so chat stays responsive while long-context processing happens elsewhere. If you have been holding off on OpenClaw for a single-Gateway, long-transcript production workload, this is the version to re-test on.

Automatic updates stop wiping your state

v2026.9.1's release notes promised update rollback and recovery. v2026.9.2 actually delivers: "keep active settings, enabled skills, and default-agent ownership in automatic updates." If a v2026.9.1 auto-update failed on your Gateway, your settings, your enabled skills, and your default-agent ownership could be reset to a fresh-install baseline. v2026.9.2 preserves them. Combined with the v2026.9.1 Doctor-failure rollback path, the practical update reliability story is now in the same neighborhood as Codex's roll-forward patch chain.

GPT-6 Astra is a first-class model

v2026.9.1 shipped Mermaid diagrams but did not advertise GPT-6 Astra support. v2026.9.2 does. Builders on OpenAI API-key profiles or eligible ChatGPT/Codex accounts can now select openai/gpt-6-astra with text and image input, Responses tool calls, and supported reasoning controls. Subscription availability is gated on successful account discovery; if your account isn't eligible, the model picker still shows it, but the gateway will return a discovery failure.

The deeper story is in the two Changes items tagged GPT-6 Astra: async function tools, WebSocket steering, prefix retention on reasoning-effort changes, and /think ultra orchestration. Those are the Responses API long-running-work controls (covered Sep 3) wired into OpenClaw's runtime. If you are building long-running autonomous agents and you were waiting for OpenClaw to support them at the platform layer, the wait is over.

Replies survive a Gateway restart

The most operationally relevant reliability fix in the release. Before v2026.9.2, a Gateway restart could lose active replies, queued replies, and delegated replies, and could discard another reply's recovery marker by completing one reply before the recovery marker from a previous restart had been processed. v2026.9.2 recovers all three reply classes and keeps continuation instructions intact through compaction and retry attempts. If you run OpenClaw on a node that occasionally restarts (auto-update, host reboot, kernel upgrade), this is the fix that closes the silent-reply-loss class of bug.

Swarm and cross-agent session access flip to default-on

Two permission-shape defaults changed in v2026.9.2 and both default-on:

• Swarm is enabled by default. Concurrent sub-agents with structured results and live progress, with explicit opt-outs, tool restrictions, and the separate Code Mode opt-in preserved. If you do not want swarming, set the opt-out flag explicitly. The release notes do not specify the opt-out setting name; documentation indicates it sits alongside Code Mode and is enumerated in the configuration reference.
• Cross-agent session access. Session tools now default to all-session visibility and ordinary agent-to-agent access is enabled. The narrower configurations are tools.sessions.visibility = agent (current agent's sessions) or self (only the calling agent's session). Existing tool and sandbox restrictions are retained.

Both are deliberate defaults designed for multi-agent Gateway deployments. For single-agent deployments, both still work — but read the release notes before you upgrade on a shared Gateway. A shared Gateway with multiple operator accounts may now expose session metadata across agents by default.

Why Developers And Founders Should Care

Seven distinct reasons.

1. Chat is finally responsive under load. If your Gateway's UX degrades when transcripts grow, v2026.9.2 is the targeted fix. 2. Updates stop costing you state. Auto-updates that fail no longer reset your settings, skills, and default agent. 3. GPT-6 Astra is now a first-class runtime model. Builders waiting for OpenClaw-native GPT-6 Astra support can stop holding off. 4. Async tools and steering are wired through. OpenClaw's runtime now executes async function tools and WebSocket steering on GPT-6 Astra routes with prefix retention. 5. Replies survive restarts. Active, queued, and delegated replies recover cleanly through a Gateway restart. For nodes that auto-restart, this closes a class of silent reply loss. 6. Backups are safer. Corrupt archive headers are rejected instead of accepted; text with embedded NUL characters is preserved end-to-end; Nix-managed config and credential links are supported. 7. Swarm and cross-agent session access are default-on. Multi-agent Gateway deployments get the multi-agent defaults they need; single-agent deployments should review the release notes before upgrading on a shared Gateway.

Evidence And Verification

Every claim above is sourced verbatim from the v2026.9.2 release notes on GitHub:

• Primary source (release notes). https://github.com/openclaw/openclaw/releases/tag/v2026.9.2 — fetched 2026-09-06 14:10 UTC. Highlights, Changes, and Fixes sections captured verbatim, including the 6 highlight bullets and the 22+ change/fix items above.
• Primary source (atom feed). https://github.com/openclaw/openclaw/releases.atom — entry tag:github.com,2008:Repository/1103012935/v2026.9.2, updated 2026-09-06T02:15:06Z. Confirms the release timestamp.
• Cross-reference (v2026.9.1 stable). https://mr.technology/payloads/openclaw-v2026-9-1-stable-mermaid-skill-library-update-rollback-sep-2026 — published 2026-09-03 20:08 UTC. Confirms which features are v2026.9.1-stable (Mermaid, one-prompt install, per-identity skill libraries, update rollback, Gateway reliability, Codex approvals, channel integration, Android parity) versus v2026.9.2-stable (this article).
• Cross-reference (OpenAI GPT-6 Astra launch). https://mr.technology/payloads/openai-gpt-6-astra-async-tool-calling-mid-turn-steering-sep-2026 — published 2026-09-03 20:08 UTC. Confirms GPT-6 Astra's Responses API long-running-work controls (async tool calling, mid-turn steering, change reasoning effort mid-conversation) that v2026.9.2 wires into OpenClaw's runtime.
• Cross-reference (Anthropic Fable 5.1 / Mythos 5.1 / cache reads). https://mr.technology/payloads/anthropic-fable-5-1-mythos-5-1-launch-quarter-price-cache-reads-adaptive-thinking-sep-2026 — published 2026-09-02 12:08 UTC. v2026.9.2's /think ultra orchestration extends to OpenClaw and native Codex runtimes; Fable 5.1 / Mythos 5.1 are the Anthropic-side surface for that work.

Verification level: documentation comparison with primary-source verbatim quotes. No firsthand test was run against v2026.9.2 in this article's pipeline.

Cost, Risk, And Limitations

Cost. None introduced by this release. v2026.9.2 does not change any pricing tier. The GPT-6 Astra support it wires through follows OpenAI pricing (standard/batch/fast tiers documented in the GPT-6 Astra launch coverage). Anthropic-side /think ultra orchestration applies only when GPT-6 Astra is selected as the runtime model, so cost is governed by your GPT-6 Astra tier choice.

Risk.

• Swarm default-on. If you run a shared Gateway and do not want sub-agent fan-out, you must opt out explicitly. The release notes do not enumerate the opt-out setting name; documentation indicates it sits alongside Code Mode in the configuration reference.
• Cross-agent session access default-on. Same shape as Swarm: shared Gateways may now expose session metadata across agents by default. Narrow to tools.sessions.visibility = agent or self if needed.
• HTTP proxy safety fix. v2026.9.2 updates HTTP dependencies and tightens destination DNS checks. If you operate an OpenClaw Gateway behind a corporate proxy, verify that your proxy configuration is still compatible after the dependency bump.
• macOS browser sign-in via Cloudflare Access. If you do not run Cloudflare Access in front of your Gateway, this feature is irrelevant; if you do, v2026.9.2 expects personal Cloudflare Access browser sign-in rather than the previous path. Read the release notes before you upgrade on a Cloudflare-Access-protected Gateway.
• Apple Watch standalone Talk is UDP-dependent. Standalone realtime audio on Watch requires UDP connectivity. If your Watch-to-iPhone connection is on a network that blocks UDP, the standalone mode will not work and the existing iPhone-mediated flow remains.

Limitations of this report.

• No firsthand install of v2026.9.2 was run. The article is documentation-surfacing; verification level is release-notes comparison.
• Swarm opt-out setting name not enumerated in the release notes; documentation indicates it is in the configuration reference, which was not fetched end-to-end.
• /think ultra behavior on OpenClaw and Codex native runtimes was not exercised.
• Doctor database diagnostics behavior on a real malformed SQLite file was not tested.
• HTTP proxy safety fix was not validated against an actual corporate TLS-inspecting proxy.
• Backup NUL-character preservation was not exercised against a real backup archive.

Mr. Technology Verdict

v2026.9.2 is the operational follow-up v2026.9.1 needed but did not deliver. The headline value is on three axes at once: chat responsiveness under load (event-loop offloading), update reliability (state preservation across auto-updates), and recovery (active/queued/delegated replies survive Gateway restarts). For production OpenClaw deployments, all three of those have been the largest single classes of complaint. v2026.9.2 closes each.

The GPT-6 Astra wiring is the second story: it is the OpenClaw-side landing for the Responses API long-running-work controls covered in the Sep 3 GPT-6 Astra launch article, complete with async function tools, WebSocket steering, and prefix-preserving reasoning-effort changes. If you are an autonomous-agent builder choosing between OpenClaw and a hand-rolled runtime, this release removes the last credible reason to hand-roll.

The two default-on changes — Swarm and cross-agent session access — are deliberate, well-scoped, and easy to revert if they don't fit your deployment. They are also a signal that OpenClaw's center of gravity is moving toward multi-agent Gateway deployments. Single-agent builders should read the release notes before upgrading on a shared Gateway; multi-agent builders should already be planning the upgrade.

Verdict: ship the upgrade on staging this week; on production once you've read the Swarm and cross-agent session access defaults.

Recommended Action

Today

• Skim the v2026.9.2 release notes yourself: https://github.com/openclaw/openclaw/releases/tag/v2026.9.2
• Decide whether Swarm default-on fits your deployment. If not, identify the opt-out setting in the configuration reference and prepare it before you upgrade.
• Decide whether cross-agent session access default-on fits your deployment. Set tools.sessions.visibility = agent or self if you need narrower access.

This week

• Upgrade a staging Gateway to v2026.9.2. Verify chat responsiveness under your longest-transcript load. Verify auto-update preserves your active settings, enabled skills, and default-agent ownership. Verify active/queued/delegated replies survive a openclaw update-triggered Gateway restart.
• If you operate a shared Gateway, document the Swarm and cross-agent-session-access opt-out settings in your runbook before the production rollout.
• If you operate behind a corporate HTTP proxy, schedule a connectivity test against v2026.9.2's HTTP dependency bump before the production rollout.

Next week

• Upgrade production. Watch for chat-stutter regressions on long transcripts; if you see them, the event-loop offloading in v2026.9.2 should help, but if not, file a follow-up issue with a transcript-length and reproduction recipe.
• If you use GPT-6 Astra, plan a one-prompt smoke test of async tool calling + WebSocket steering + change-reasoning-effort-mid-conversation against your runtime to confirm OpenClaw's wiring matches your expectations.
• Re-test your backup restore path against v2026.9.2's backup behavior: confirm NUL-character preservation, confirm corrupt-archive-header rejection, confirm Nix-managed config + credential link handling if you use Nix.

Skip if not in scope

• The Discord meeting notes, Apple Watch standalone Talk, Slack rich replies, Dashboards gallery view, Devices-at-a-glance, and macOS approvals changes are independent of the headline story. Skip them until your team has bandwidth for a separate feature review.
• Experimental plugin UI (Settings → Labs → Custom plugin UI) is opt-in and clearly labeled experimental. Skip unless you build OpenClaw plugins.

Sources

• OpenClaw v2026.9.2 release notes — https://github.com/openclaw/openclaw/releases/tag/v2026.9.2 (verified 2026-09-06 14:10 UTC; Highlights, Changes, and Fixes sections captured verbatim)
• OpenClaw atom feed — https://github.com/openclaw/openclaw/releases.atom (entry tag:github.com,2008:Repository/1103012935/v2026.9.2, updated 2026-09-06T02:15:06Z)
• v2026.9.1 stable cross-reference — https://mr.technology/payloads/openclaw-v2026-9-1-stable-mermaid-skill-library-update-rollback-sep-2026 (published 2026-09-03 20:08 UTC; defines which features belong to v2026.9.1 vs v2026.9.2)
• OpenAI GPT-6 Astra launch cross-reference — https://mr.technology/payloads/openai-gpt-6-astra-async-tool-calling-mid-turn-steering-sep-2026 (published 2026-09-03 20:08 UTC; documents Responses API long-running-work controls that v2026.9.2 wires into OpenClaw's runtime)
• Anthropic Fable 5.1 / Mythos 5.1 launch cross-reference — https://mr.technology/payloads/anthropic-fable-5-1-mythos-5-1-launch-quarter-price-cache-reads-adaptive-thinking-sep-2026 (published 2026-09-02 12:08 UTC; the Anthropic-side surface for /think ultra orchestration)

Article History

Originally published: 2026-09-06 14:08 UTC

Last verified: 2026-09-06 14:08 UTC

No corrections at this time.