OpenClaw v2026.5.19 Release Summary

Key Changes

Gateway & Performance

• Startup Optimization: Gateway restart latency is reduced by overlapping independent startup work, such as logging and plugin-service initialization, while maintaining strict readiness gating for sidecars and ACPX probes.
• Restart Benchmarking: New developer tooling (pnpm test:restart:gateway) has been added to provide machine-readable evidence on restart readiness, downtime, and resource slopes.
• Deferred Update Checks: Gateway update checks are now deferred until after readiness, ensuring that package update verification no longer blocks the critical startup path.
• ACPX Instrumentation: Startup probe costs for ACPX are now explicitly attributed in restart traces, providing better visibility into the readiness timeline.

Agents & Codex Integration

• Prompt Surface Separation: To prevent "double-prompting" in native Codex, prompt routing now follows explicit surfaces (PI, CLI, ACP, Codex app-server, and subagent). This ensures that native Codex only receives OpenClaw-owned runtime context and delivery guidance rather than layering full PI policies on top of Codex's own base instructions.
• Codex Sandbox Hardening: The Codex app-server now honors writable Docker bind mounts and exposes sandbox_exec/sandbox_process as fallback shell surfaces when native Code Mode cannot be safely represented on the host.
• Subagent Task Management: Fixed a critical issue where Codex-native subagent tasks remained marked as running in the registry even after returning a terminal BLOCKED result due to unavailable hook relays.
• Prompt Guidance: Built-in tool descriptions and schema hints have been shortened across multiple toolsets to optimize routing guardrails.

Channel Integrations

• Telegram Enhancements:
  • Forum Topic Parallelism: Fixed serialization bottlenecks in forum topics. Text-fragment and media-group buffering are now per-topic, and a new fair-share outbound queue prevents a single topic from monopolizing the shared group send pipe.
  • Ephemeral Drafts: Added optional native DM tool-progress drafts for transient "Thinking" previews, keeping final answers on the durable delivery path.
  • Context Preservation: Fixed a regression where bare mention replies in Telegram dropped the replied-to message body in the model prompt.
• Discord Stability:
  • Realtime Voice: Improved OpenAI realtime voice playback stability by prebuffering raw PCM and disabling near-field noise reduction for the backend bridge, eliminating choppy starts and speech detection failures.
  • Subagent Routing: Fixed a bug where the initial reply from thread-bound delegated sessions was incorrectly posted to the parent channel instead of the bound thread.
  • Streamed Reply Preservation: Resolved an issue where streamed assistant replies disappeared when a turn also contained a recovered tool-failure warning.
• WhatsApp Improvements: The forceDocument flag is now honored end-to-end, allowing agents to send uncompressed images and media as documents.

Memory & CLI

• Memory Startup Catch-up: memory-core now performs a startup scan to compare on-disk session transcripts against the index, ensuring that session indexing does not fall behind across gateway restarts.
• CLI UX Improvements:
  • Added openclaw sessions list as an alias for openclaw sessions to match the convention of other list-style commands.
  • Improved openclaw plugins doctor to surface warnings when a configured agentRuntime.id="codex" is missing the required Codex plugin.
  • Added a --global flag to openclaw skills install/update to target shared managed skills in ~/.openclaw/skills.

Impact

Stability & Reliability

• Session Recovery: The Gateway now automatically rotates failed session lanes if the stored transcript file is missing, preventing permanent hangs in per-channel recovery paths.
• Transcript Integrity: Fixed a high-severity issue where cache-ttl handlers caused duplicate assistant-turn emissions in the session transcript, which previously led to verbatim response repetition and inflated token spend.
• Auth Fallback: Hardened OAuth recovery for openai-codex to prevent loops on stale lastGood profiles after refresh_token_reused errors.

Security & Policy

• Tool Policy Enforcement: Inline skill tool dispatch now strictly honors the effective tool policy (allow/deny lists) before selecting a dispatch target.
• Codex Policy Hardening: Restricted Codex runs now explicitly disable native code mode, built-in environments, and user MCP projection when a deny: ["*"] policy is active.

Upgrade Guide

Breaking Changes & Migration

• Node.js Version: The minimum supported Node.js 22 line has been raised to 22.19. Please ensure your environment is updated to avoid launcher contract failures.
• Docker/Podman Build Args: A new runtime-neutral build argument OPENCLAW_IMAGE_APT_PACKAGES is now preferred for installing extra apt packages. OPENCLAW_DOCKER_APT_PACKAGES remains as a legacy fallback but will be superseded by the new variable if both are set.