OpenClaw v2026.5.31-beta.4: Enhanced Agent Recovery, SQLite State Migration, and Skill Workshop

OpenClaw v2026.5.31-beta.4 introduces a significant shift toward architectural stability and developer tooling. The primary focus of this release is the migration of critical runtime state—including cron jobs, plugin indices, and iMessage monitor data—into SQLite to eliminate the flakiness associated with legacy JSON/JSONL file-backed storage.

Beyond infrastructure, this version enhances the reliability of agent-to-agent (A2A) communication and tool execution, ensuring that interrupted runs and stale session bindings recover more cleanly. The introduction of the Skill Workshop also provides a structured, governed flow for creating and reviewing agent skills, moving away from ad-hoc configurations toward a proposal-based system.

Key Changes

Infrastructure & State Migration

To improve performance and reliability, OpenClaw has migrated several key state-management systems to SQLite:

• Cron Persistence: Cron jobs, runtime state, and run history are now stored in the shared SQLite database, replacing legacy JSONL files. This includes a durable import process for existing logs and a quarantine system for malformed legacy rows.
• Plugin Indexing: The installed plugin index is now persisted in SQLite, reducing the need for expensive filesystem scanning during startup and reloads.
• iMessage Monitor: Reply caches, sent-echo deduplication, and catchup cursors for iMessage have moved to SQLite-backed plugin state.
• ACP Metadata: Agent Control Plane (ACP) session metadata and event ledgers are now persisted in SQLite, ensuring they survive restarts and updates.
• OpenRouter Cache: Model capability caching for OpenRouter has been moved to the shared SQLite state DB.

Agent & Runtime Reliability

Significant work has been done to harden agent execution and recovery:

• Interruption Recovery: Agents and CLI-backed runtimes now recover more cleanly from interrupted tool calls, stale session bindings, and media delivery retries.
• Terminal Outcome Precedence: A new canonical helper centralizes how agent run outcomes (completed, timeout, cancelled, failed) are determined, preventing late aborts or rejections from overwriting hard timeouts or explicit cancellations.
• A2A Normalization: The sessions_send tool now normalizes model-emitted body aliases (such as SendMessage, content, or text) to the canonical message property, fixing a critical bug where Anthropic and Pi-backed agents failed schema validation during cross-agent messaging.
• Media Routing: Generated media (images, video, music) completions are now routed back through the requester agent, allowing the agent's own configuration to decide whether to deliver the result automatically or via a message tool.

Skill Workshop & Developer Tools

The new Skill Workshop introduces a governed workflow for skill creation:

• Proposal Flow: Skills are now managed through a reviewable proposal system. Operators can apply, reject, or quarantine explicit proposals via a guarded review flow.
• Control UI Integration: A new Skill Workshop dashboard in the Control UI features proposal lists, revision handoffs, searchable file previews, and locale coverage.
• Documentation: A dedicated Skill Workshop guide has been added, covering governed creation, approval policies, and recovery.

UI & UX Enhancements

• Calm Composer: The chat composer has been streamlined, moving model, reasoning, and speed toggles into compact custom controls to reduce visual clutter.
• Dreaming Tab: The Dreaming tab now includes an agent selector, allowing users to switch between different agent contexts to inspect their specific dreaming status and diaries.
• Startup Performance: Chat startup metadata has been streamlined; chat.history now provides enough session metadata to skip redundant sessions.list calls during hydration.

Provider & Integration Updates

• MiniMax M3: Added support for MiniMax M3, which is now the default model for MiniMax setups.
• Google Gemini: Fixed a routing issue where Google Gemini models were incorrectly routed to the openai-responses transport; they now correctly resolve to the native google-generative-ai transport.
• iOS Enhancements: Added a dedicated Talk tab with realtime playback, animated voice UI, and echo suppression. The system now defaults to a hosted push relay (https://ios-push-relay.openclaw.ai) for APNs sends.
• OpenAI TTS: Added support for the [[tts:speed=N]] directive, allowing users to tune the speech rate for OpenAI TTS providers.

Impact

Stability and Performance

Users will notice faster startup times and more consistent behavior during restarts due to the shift to SQLite. The elimination of SQLITE_BUSY errors in memory search (via new cross-process write locks) ensures that background gateway activity no longer interferes with foreground CLI searches.

Developer Workflow

The Skill Workshop transforms skill management from a manual configuration task into a governed process. This reduces the risk of breaking changes being introduced into production environments by requiring a review and approval cycle for new skill proposals.

Mobile Experience

iOS users gain a more native-feeling voice experience with the new Talk tab and more reliable push notifications via the hosted relay.

Upgrade Guide

State Migration

Upon upgrading to v2026.5.31-beta.4, OpenClaw will automatically attempt to migrate legacy JSON/JSONL state files (Cron, Plugins, iMessage, ACP) into the SQLite database.

• Action Required: While the migration is automatic, it is recommended to run openclaw doctor after upgrading to verify that all legacy state was successfully imported and archived.
• Note: Malformed legacy rows that cannot be imported will be kept on disk in a quarantine state to prevent data loss.

macOS Launchd Cleanup

For macOS users, this release addresses a critical bug where launchctl submit jobs could enter an infinite update loop. The system now recognizes ai.openclaw.manual-update.* labels as stale updater jobs and can clean them up via the doctor or status tools.