OpenClaw Fix Plugin Ecosystem Fragmentation: Codex and Pi Hook Harmonized, Load Time Reduced by 90%

According to Watchful AI monitoring, the open-source AI Agent platform OpenClaw has released version 2026.4.22. The biggest change is the alignment of Codex harness with Pi's lifecycle. Previously, plugins behaved inconsistently on the Pi harness and Codex harness paths, with the same plugin potentially missing hook calls on different harnesses. This release simultaneously fills in key hooks such as `before_prompt_build`, `before_compaction / after_compaction`, `after_tool_call`, `before_message_write`, `llm_input / llm_output / agent_end`, eliminating the need for plugin developers to adapt separately for the two paths. The sync also added a plugin extension interface on the Codex side, supporting asynchronous tool_result middleware. This comes after the OpenAI Codex team fixed the Codex harness authentication silent fallback issue earlier this week, marking another round of systemic patches for OpenClaw's Codex integration.

Another architectural addition is the TUI local embedding mode: Users can now run agent conversations directly in the terminal without starting the Gateway, while retaining the plugin approval mechanism.

The default thinking level of the reasoning model has been silently upgraded from the previous off/low to medium. Users who have not manually set the inference level will now find the reasoning model automatically starting to output the thought process.

In terms of performance, plugin loading now uses native Jiti, reducing startup time by 82% to 90%; the `doctor --non-interactive` runtime has decreased by approximately 74%. Kimi K2.6 multi-round agent calls no longer interrupt due to incorrect cleaning of the tool_call ID. Subprocesses on Linux now automatically increase `oom_score_adj`, with the kernel favoring the termination of temporary workers under memory pressure rather than the Gateway main process. The configuration system has added last-known-good recovery to prevent accidental config overrides leading to Gateway crash loops.

On the provider side, xAI has added image generation (grok-imagine-image / grok-imagine-image-pro), TTS, and STT; Tencent Cloud is now a built-in official provider plugin, featuring the Hy3 preview model and pricing. When OpenAI models enable web search, they now directly use the OpenAI native web_search tool, bypassing the OpenClaw-hosted search channel.