OpenClaw 2026.4.22 Unifies Plugin Lifecycle Across Codex and Pi Harnesses, Reduces Plugin Load Time by Up to 90%

Gate News message, April 23 — OpenClaw, an open-source AI Agent platform, released version 2026.4.22 on April 22, with its biggest change being the alignment of Codex harness and Pi harness lifecycles. Previously, plugins behaved inconsistently across the two harness pathways, with some hooks missing in certain environments. The new version consolidates critical hooks including before_prompt_build, before_compaction/after_compaction, after_tool_call, before_message_write, and llm_input/llm_output/agent_end, eliminating the need for developers to maintain separate implementations for each pathway. The update also introduces async tool_result middleware support for Codex-side plugin extensions.

Performance improvements are significant: plugin loading now uses native Jiti, reducing startup time by 82% to 90%, while doctor --non-interactive runtime decreased by approximately 74%. The default thinking level for inference models has been upgraded from off/low to medium, enabling reasoning outputs by default for unmodified configurations. Additional fixes include Kimi K2.6 multi-turn agent calls no longer interrupting due to tool_call ID corruption, enhanced Linux subprocess memory management with automatic oom_score_adj adjustment, and a new last-known-good configuration recovery system to prevent Gateway crashes from accidental config overwrites.

New provider integrations include xAI's image generation (grok-imagine-image and grok-imagine-image-pro), TTS, and STT capabilities; Tencent Cloud as an official provider plugin with Hy3 preview models and pricing; and OpenAI's native web_search tool, replacing OpenClaw's managed search channel when web search is enabled.