OpenClaw published v2026.6.10-beta.1, a large beta release that bundles 109 merged pull requests across agent turns, channel delivery, Codex approvals, mobile clients, plugin installs, and operator diagnostics.
The release landed after the previous stable 2026.6.9 coverage, and the theme is clear: OpenClaw is hardening the parts of the system that decide whether an agent turn feels dependable in the real world. The changelog highlights "more reliable agent turns and session state," then backs that up with fixes for pending subagent completion announcements, non-empty chat history transcripts, media index alignment, dormant follow-up drains, and compaction model aliases.
What Changed
The biggest bucket is runtime reliability. OpenClaw now preserves more completion and transcript state across partial turns, retries, and recovery paths. That matters because many OpenClaw deployments are not single-command CLI sessions anymore. They are long-running assistants receiving channel messages, spawning subagents, compacting context, streaming progress, and recovering after mobile or gateway hiccups.
The release also strengthens Codex and approval flows. The notes call out app-server SecretRefs, thread context, bounded turn text, routed approval context, and typed SDK approval/session helpers. That combination should make approval-driven workflows less brittle when a request moves between a source channel, an embedded agent runtime, and the app server.
Channel delivery gets a broad pass too:
- Telegram gains richer progress previews and structured send-error handling.
- Discord and Slack preserve more reasoning, thread, and stream output.
- Slack shortcut support is wired through.
- Canonical sent threads are recorded more reliably.
That is not a single flashy feature, but it is the kind of cross-channel polish that prevents messages from being lost, duplicated, or shown in the wrong shape.
Safer Boundaries
The release includes several safety and network-boundary fixes. The release notes say SSH tunnel preflight is now loopback-scoped, device-backed node pairings have been removed, volatile SQLite state is surfaced by doctor, and legacy Codex routes are repaired instead of silently persisting stale state.
Those changes are important because they reduce ambiguity. Operators get clearer diagnostics when state is volatile, old routes are stale, or a network probe could otherwise classify too broadly. The release also rejects unsafe chat, tool, package, and response lengths, and keeps isolated cron delivery from proceeding without explicit target proof.
Operator-Facing Improvements
The beta also includes useful quality-of-life changes for everyday operators. Session renaming from chat, explicit session compaction, session duration display, command progress details, and dry-run previews for message sends and polls all appear in the release notes.
The mobile and desktop clients are part of the same polish pass. Android settings are grouped by intent, iOS notification state is cleaned up, the Watch app moves to an Xcode 27-compatible target layout, and macOS file inputs open through the native panel.
Release Verification
OpenClaw continues to publish verification artifacts with its releases. This beta includes dependency evidence, a release manifest, postpublish evidence, SHA-256 sidecars, npm package verification, registry tarball integrity, and links to the release validation and publish workflows.
That is useful for anyone running OpenClaw in production-like environments. The release is not just a tag and a changelog; it ships with machine-checkable evidence that the npm package, plugin publish steps, and validation workflow completed.
The practical read: v2026.6.10-beta.1 is a broad reliability beta. It is less about one new marquee feature and more about making long-running, channel-connected OpenClaw agents behave predictably across session state, approvals, diagnostics, and delivery.