Release Overview
OpenClaw 2026.4.23 focuses on refining media capabilities and agent orchestration. Key additions include image generation via Codex and OpenRouter OAuth, a new "forked context" mode for subagents, and improved memory configuration for local embeddings. This release also squashes critical bugs in Telegram media replies, WhatsApp onboarding, and Slack group DM classification.
Headline Features
Codex & OpenRouter Image Generation
Agents can now generate and edit images using openai/gpt-image-2 through active Codex OAuth or OpenRouter profiles. This removes the requirement for a direct OpenAI API key for image tasks when these providers are configured. Theimage_generate tool now also supports provider-specific hints for quality, background, and moderation.
Subagent Forked Context
The sessions_spawn tool gains an optional forked context mode. When enabled, the child agent inherits the current requester transcript, allowing it to maintain continuity without the parent agent needing to manually summarize the conversation. This is ideal for complex follow-up tasks where full context is required.
Configurable Local Embeddings
Local embedding context size now defaults to 4096 and can be adjusted viamemorySearch.local.contextSize. This provides better performance out of the box while allowing power users to tune memory usage for constrained environments.
Stability & Fixes
Telegram & WhatsApp Improvements
Telegram group chats now correctly parse remote markdown image syntax into outbound media payloads, preventing the fallback to plain-text URLs. WhatsApp onboarding is now decoupled from the Baileys runtime path, allowing the QuickStart setup to display correctly before all dependencies are staged.
Codex Harness Hardening
The Codex harness receives multiple fixes: request_user_input prompts are now correctly routed back to the originating chat, and npm-installed codex.cmdshims on Windows are resolved automatically. Structured debug logging has also been added for harness selection decisions.
Slack Group DMs
Multi-person direct messages (MPIMs) are now correctly classified as group chat context. This automatically suppresses verbose tool/plan progress traces in these surfaces, keeping shared conversations clean.
Other Notable Changes
- Media Understanding: Explicit image-model configuration is now honored before native-vision skips, improving support for text-only primary models.
- Memory/CLI: The built-in local embedding provider is now declared in the core manifest, fixing resolution issues for standalone CLI commands.
- Plugins/Google Meet: Improved cleanup for Twilio calls and audio bridges.
- Control UI: Assistant-generated images are now persisted as authenticated media, ensuring they remain visible after history reloads.
- Dependencies: Bundled Pi packages updated to 0.70.0 with GPT-5.5 catalog metadata.
How to Update
Update to 2026.4.23 with one command:
openclaw updateFAQ
Should I upgrade to 2026.4.23 immediately?
Yes, this release fixes critical media handling issues in Telegram and WhatsApp, improves stability for the Codex harness, and adds powerful context inheritance for subagents.
How does the Codex image generation work?
OpenClaw now routes image generation through Codex OAuth (openai/gpt-image-2) if active, meaning you can generate and edit images without needing a separate OpenAI API key.
What is forked context for subagents?
It's an optional mode for sessions_spawn that lets a subagent inherit the current requester's transcript. This allows the child agent to 'know' what was just discussed without manual summary passing.
Why was the local embedding context size changed?
It now defaults to 4096 and is configurable via memorySearch.local.contextSize, allowing you to tune performance for resource-constrained hosts like Raspberry Pi.
Need help from people who already use this stuff?
Join the OpenClaw Community
Get help with updates, share your use cases, and connect with other OpenClaw builders.