update
May 7, 2026
By Teun
OpenClaw 2026.5.6 fixes Codex, browser, and macOS issues
OpenClaw’s May 7 newsletter highlights the 2026.5.6 release and related fixes for Codex, browser tabs, and macOS LaunchAgents. It also lists new work on agent failover, plugin authorization, localized tool summaries, and memory capture rules.
OpenClaw’s May 7 newsletter centers on the 2026.5.6 release, which the project describes as a fixes-focused update. The release follows a string of targeted patches across Codex support, browser handling, gateway behavior, and messaging integrations.
One of the top items is a Doctor/OpenAI Codex fix in 2026.5.6. According to the newsletter, OpenClaw reverted a 2026.5.5 doctor --fix repair that had rewritten valid openai-codex/ and ChatGPT/Codex OAuth routes to openai/, a change that could break OAuth-only GPT-5.5 setups.
The newsletter also points to a Feishu fix in 2026.5.5. OpenClaw said it now hydrates missing native topic starter thread IDs before session routing so first turns and follow-ups stay in the same topic session. The project said that patch fixes issue #78262.
Another listed fix addresses macOS LaunchAgent lifecycle behavior in src/daemon/launchd.ts. OpenClaw said the gateway stop path had been calling launchctl disable before stopping, which could permanently suppress KeepAlive recovery. The same patch also skips a repair kickstart when a LaunchAgent is already running.
On the model and plugin side, OpenClaw said it now installs or repairs the Codex plugin when openai/ or openai-codex/ models are selected. According to the newsletter, that coverage now extends across onboarding, auth and model picker flows, CLI model commands, Crestodian setup, and gateway session handling.
The browser stack also got a fix. OpenClaw said the post-navigation SSRF guard could reject a page and silently close the tab, even when the rejection came from a read-only operation. The new behavior keeps user tabs open in that case.
Several other items in the newsletter focus on agent behavior and runtime correctness. OpenClaw said it added configurable localization for inline tool summaries, with an optional minimum interval between summary emissions, while keeping English as the default.
The project also listed a change to compute plugin callback authorization. According to OpenClaw, Telegram and Slack plugin-owned interactive callbacks could receive a hardcoded authorized-sender signal instead of the command-level authorization decision, and the new work makes that authorization dynamic.
OpenClaw also noted a fix for structured delta.content handling in Mistral reasoning replies. The newsletter says Mistral can return delta.content as an array of typed blocks, and the bug caused replies to show [object Object] instead of readable content.
Performance work is also in the mix. OpenClaw said it trimmed agent tool helper queues, cached provider tool schema normalization, and reused plugin metadata snapshots for provider runtime loads. The project also added state-aware failover and lane suspension to reduce context loss and retry loops when providers run out of quota.
On the memory side, OpenClaw said the memory plugin now has configurable capture triggers and category detection rules. The newsletter says this replaces hardcoded behavior and makes it easier to adapt memory capture to different languages, domains, or workflows.
The newsletter also mentions additional fixes for session runtime override reporting, gateway reset skills snapshots, Feishu topic session splitting, native command owner enforcement, Codex plugin app support, proxy-aware Codex usage loads, and a browser SSRF tab-retention bug. OpenClaw’s repo stats for May 7 list 369,229 stars, 76,175 forks, and 2,128 contributors.
