OpenClaw v2026.5.16-beta.3 Release Summary

Key Changes

Codex App-Server & Runtime Enhancements

This release brings substantial improvements to the Codex app-server integration, focusing on stability, policy enforcement, and context management:

• Native Tool Policy Enforcement: OpenClaw now injects a trusted native hook relay into the Codex app-server. This ensures that Codex-native shell and file operations are intercepted by the before_tool_call policy, preventing security bypasses.
• Context-Engine Binding: Codex app-server threads are now bound to context-engine projection epochs. This allows the system to reuse bootstrap bindings and automatically start fresh backend threads when context-engine compaction occurs, keeping the model and context engine in sync.
• Progress-Aware Turns: Codex turns are now more responsive to activity. The system tracks progress from user input, dynamic tools, and notifications, preventing turns from timing out while active work is still being performed.
• Watchdog Improvements: New idle watchdogs have been implemented to fail fast when Codex emits raw tool-output or assistant completions without a terminal turn/completed event, reducing unnecessary wait times.

Gateway & Infrastructure

• Restart Observability: A new restart trace instrumentation (OPENCLAW_GATEWAY_RESTART_TRACE) provides parseable logs for the drain, shutdown, and ready phases of a gateway restart, enabling deeper performance diagnostics.
• Startup Reliability: The gateway run-loop now queues restart and shutdown signals received during the startup phase, ensuring that update churn does not drop critical lifecycle requests.
• Lifecycle Hook Budgets: Default timeout budgets for gateway lifecycle hooks have been increased (5s for shutdown, 10s for pre-restart) to allow notification handlers more time to complete.

Channel & Provider Updates

• Telegram:
  • Added "room event" handling for ambient group chatter, allowing always-on context to run quietly and speak only via the message tool.
  • Fixed a critical bug where expired approval callbacks caused infinite retry loops, blocking all incoming DMs.
  • Improved restart replay to prevent context loss in group-topic sessions.
• Slack:
  • Improved user mention guidance, instructing agents to use stable <@USER_ID> tokens.
  • Fixed a bug where tool warnings could overwrite finalized assistant drafts during streaming.
  • Routed DM thread replies to the main session instead of creating invisible thread-scoped sessions.
• LINE: Fixed a silent push failure where lowercased recipient IDs (generated during session-key fallback) were rejected by the LINE API.
• Ollama: Added logic to skip the think parameter for models marked as non-reasoning, avoiding 400 errors from Ollama servers.
• Xiaomi MiMo: Fixed an issue where reasoning-only completions appeared blank by promoting reasoning_content to visible text for terminal outputs.

CLI & DX

• Onboarding i18n: The CLI setup wizard and bundled channel setup flows are now localized for English, Simplified Chinese, and Traditional Chinese.
• Cron Enhancements: Added a --wait mode for openclaw cron run, allowing automation to block until a manual run reaches a terminal status.
• Doctor Tool: openclaw doctor --fix now repairs source-only official plugin installs (e.g., Brave, Slack) and ensures the Codex plugin is enabled when required by configured OpenAI routes.

Impact

Performance & Stability

• Memory Leak Fix: A significant file descriptor leak in the memory search watcher was resolved by removing awaitWriteFinish polling, preventing spawn EBADF errors in environments with large Markdown trees.
• Session Repair: Fixed a bug where successful session repairs left behind unbounded .bak files, which previously caused gigabytes of debris for stuck sessions.
• Session Usage: Local/OpenAI-compatible sessions now use a conservative transcript-derived estimate when provider usage telemetry is missing, ensuring the context meter remains accurate.

Security

• System Event Authority: Authority for queued system events is now carried as structured metadata (forceSenderIsOwnerFalse) rather than prompt-visible text, preventing trust markers from being spoofed or transformed by the model.
• Auth Revocation: Removing provider auth through the Gateway control plane now immediately aborts active matching runs with a stopReason: "auth-revoked" signal.

User Experience

• Android App: URLs in chat messages are now tappable by preserving Compose URL annotations in rendered markdown.
• TUI: The status bar now updates the displayed model in real-time during provider fallback, providing better visibility into which model is currently being attempted.