OpenClaw v2026.4.23 Adds gpt-image-2 Direct OAuth Support, Introduces Forked Context Mode for Sub-agents

Gate News message, April 25 — OpenClaw, an open-source AI agent framework, released v2026.4.23 on April 23, introducing updates across image generation, sub-agent mechanisms, and security hardening.

Image generation enhancements allow gpt-image-2 to be called directly via Codex OAuth without separate OPENAI_API_KEY configuration. OpenRouter image models are now integrated through the image_generate tool. Multi-reference image editing has been migrated from JSON data URLs to multipart uploads, resolving previous failures with complex edits. Agents can now specify quality, output format, and background transparency parameters as needed.

Sub-agents gain a new “forked context” mode, enabling child agents to inherit parent agent conversation context rather than starting from blank sessions; isolation mode remains the default. Image, video, music, and TTS generation tools now support per-call timeout settings independent of global limits. Security improvements include restricting MCP tool bridge access to prevent privilege escalation, enforcing whitelist-based gateway runtime configuration, isolating WhatsApp contact and location text to prevent prompt injection, and requiring correct appid in Teams Bot Framework tokens. The memory system decoupled dreaming (periodic memory consolidation) from heartbeat, and added memorySearch.local.contextSize configuration (default 4096) for context window tuning on low-spec devices.