What are forked-context subagents in OpenClaw?

Forked-context subagents allow a child agent to inherit the conversation history from the main requester while maintaining its own isolated session. This feature, introduced in version 2026.4.23, ensures that delegated agents have the necessary background knowledge to complete tasks without losing the security benefits of session isolation.

Does OpenClaw support GPT-5.5?

Yes, OpenClaw version 2026.4.23 adds native support for GPT-5.5. The update was released within 24 hours of the model's launch and includes updated metadata for OpenAI and OpenAI Codex providers. This allows users to run the latest frontier model through their self-hosted OpenClaw instances across various messaging platforms.

How does OpenClaw handle image generation without an OpenAI API key?

OpenClaw now supports image generation and reference-image editing through Codex OAuth. This allows the gpt-image-2 model to function without a direct OpenAI API key if a Codex profile is active. Additionally, the update adds support for image generation through OpenRouter using an OpenRouter API key for various image models.

What security changes were made in OpenClaw v2026.4.23?

This update implements a fail-closed policy for agent-driven configuration edits, restricting agents to a narrow set of tunable paths. It also enforces stricter rules for mobile connections, requiring loopback-only cleartext gateway connections on Android and private-IP hosts for mobile pairing to prevent insecure data transmission over public networks.

How does OpenClaw improve AI interactions on Telegram and Slack?

The update polishes messaging channels by parsing Telegram markdown images into outbound media payloads to prevent plain-text URL fallbacks. On Slack, it correctly classifies multi-person direct messages as group contexts and suppresses internal tool progress traces in non-DM surfaces to keep Working messages from leaking into public or shared rooms.

OpenClaw Integrates GPT-5.5 and Forked-Context Subagents to Improve Task Delegation

AI Assistant

GPT

Image Generation

AI Agent

Multimodal

Performance

Published Apr 24, 2026

OpenClaw Integrates GPT-5.5 and Forked-Context Subagents to Improve Task Delegation

OpenClaw, an open-source self-hosted AI assistant framework, released version 2026.4.23 with immediate support for the newly launched gpt-5.5. The update introduces forked-context subagents—autonomous agents that plan and execute tasks—via sessions_spawn, enabling child agents to inherit the requester's transcript while maintaining isolated sessions.

This release addresses the context amnesia common in multi-agent systems by allowing delegated agents to understand the full conversation history. It builds on the Active Memory sub-agent and agentic routing fixes to create more reliable orchestration loops. By integrating gpt-5.5 within 24 hours, the platform maintains its role as a fast-moving bridge for self-hosted frontier models.

You can now perform reference-image editing and generation using gpt-image-2 via Codex OAuth or OpenRouter without a direct OpenAI key. The update also hardens security with a fail-closed policy for agent-driven configuration edits and requires loopback-only connections for mobile cleartext routes. OpenClaw is available as a free, self-hosted repository on GitHub.

Read the full update →

Frequently asked questions

What are forked-context subagents in OpenClaw?: Forked-context subagents allow a child agent to inherit the conversation history from the main requester while maintaining its own isolated session. This feature, introduced in version 2026.4.23, ensures that delegated agents have the necessary background knowledge to complete tasks without losing the security benefits of session isolation.
Does OpenClaw support GPT-5.5?: Yes, OpenClaw version 2026.4.23 adds native support for GPT-5.5. The update was released within 24 hours of the model's launch and includes updated metadata for OpenAI and OpenAI Codex providers. This allows users to run the latest frontier model through their self-hosted OpenClaw instances across various messaging platforms.
How does OpenClaw handle image generation without an OpenAI API key?: OpenClaw now supports image generation and reference-image editing through Codex OAuth. This allows the gpt-image-2 model to function without a direct OpenAI API key if a Codex profile is active. Additionally, the update adds support for image generation through OpenRouter using an OpenRouter API key for various image models.
What security changes were made in OpenClaw v2026.4.23?: This update implements a fail-closed policy for agent-driven configuration edits, restricting agents to a narrow set of tunable paths. It also enforces stricter rules for mobile connections, requiring loopback-only cleartext gateway connections on Android and private-IP hosts for mobile pairing to prevent insecure data transmission over public networks.
How does OpenClaw improve AI interactions on Telegram and Slack?: The update polishes messaging channels by parsing Telegram markdown images into outbound media payloads to prevent plain-text URL fallbacks. On Slack, it correctly classifies multi-person direct messages as group contexts and suppresses internal tool progress traces in non-DM surfaces to keep Working messages from leaking into public or shared rooms.