OpenClaw 2026.4.27 adds desktop control checks and bug fixes

OpenClaw’s April 30 newsletter centers on the 2026.4.27 release, which adds new setup and safety features for Codex Computer Use, along with a wide set of fixes across the gateway, agents, memory, chat, and integrations.

According to the newsletter, Codex Computer Use setup now includes status and install commands, marketplace discovery, and fail-closed MCP checks for desktop control in Codex mode. MCP refers to the Model Context Protocol, a standard way for tools and agents to exchange structured context and capabilities.

The release notes also mention updates to OpenGrep’s pull request scanning. The CI change restricts scanning to relevant changed paths, skips draft PRs until they are ready for review, and reduces the PR diff runner from Blacksmith 16 vCPU to 4 vCPU.

Several storage and memory items landed in the release stream as well. One fix in memory-lancedb adds support for reindex and drop-index commands, after a bug where reindex could report success before LanceDB had finished creating the new vector index.

On the bug side, the newsletter says openclaw infer could hang indefinitely in 2026.4.27, with a child process spinning at 100% CPU and no network I/O. Another report says both the web app and CLI became increasingly laggy after 2026.4.23, even after upgrading again.

Gateway and config behavior also saw multiple fixes. The newsletter lists a bug where v2026.4.27 could fail during config load with EACCES on /var/lib/openclaw/plugin-runtime-deps, plus another issue where a config.patch to agents.defaults.model.primary triggered a gateway restart even though the docs say that setting should hot-reload.

Other gateway-related fixes include tolerating Paperclip metadata on agent invocation payloads, skipping web-fetch provider discovery when tools.web.fetch is explicitly disabled, and fixing a package export error for @mariozechner/pi-ai/oauth that could break gateway sessions after the upgrade.

The newsletter also flags a memory-wiki bridge issue that still returns 0 artifacts on 2026.4.27 because plugin-runtime-deps appears to be split from memory-state singleton behavior. In another memory-related change, the team notes that a publicArtifacts registration bug had already been corrected earlier, but the bridge problem remains.

In agent behavior, the release stream includes fixes for timeout and loop problems. One bug report says embedded Kimi sub-sessions could time out with no fallback even when fallbacks were configured. Another fix prevents a zombie Agent loop when an abort signal arrives before prompt() is called.

Messaging and media handling also received attention. The newsletter includes fixes to preserve the audioAsVoice flag through outbound Telegram send paths, both in the legacy adapter and the plugin sendPayload handler. That keeps text-to-speech replies from being sent as plain file attachments when they should be delivered as voice messages.

A separate Telegram issue says audioAsVoice was previously dropped in the legacy outbound media path. Discord auto-reply also appears in the bug list, where the bot could ignore messages if role-mention autocomplete was chosen instead of user-mention.

The newsletter closes with several additional fixes and regressions, including loop detection for web_fetch calls, reasoning stream support in live chat, hiding legacy runtime transcript wrappers, reaping orphaned Codex ACP processes, and a bug where the image tool failed after a global npm install on macOS because a sharp postinstall step was silently dropped.

Overall, the April 30 newsletter is a release-and-bug roundup rather than a single feature announcement. It shows OpenClaw continuing to harden its agent runtime, gateway startup, and integration paths while shipping controls meant to make desktop automation safer to set up.