Open source AI agent framework OpenClaw just dropped v2026.6.5. It's the first release using a new monthly patch naming scheme (YYYY.M.PATCH), and it's all about making reasoning models behave, decluttering storage, and shoring up the plugin ecosystem.
On the model interaction front, long-thinking models have a nasty habit of leaking their internal monologue — what the team calls “introspection leakage” — especially in chat apps. The QQBot adapter now automatically strips those thinking tags from the output stream. And to stop MCP (Model Context Protocol) tools from tripping up Anthropic's API with rich media like audio or resource links, the new version adds a forced conversion mechanism at the materialize boundary. It safely serializes non-text/image blocks before they hit the wire, squashing those dreaded “Anthropic 400 Bad Request” errors. Streaming session startup logic also got a refresh: pre-generation errors like stale prompt caches now properly trigger recovery retries.
Under the hood, OpenClaw is tidying up. It's pulling ad-hoc runtime state files — Matrix sync encryption, Memory-Wiki imports, sandbox registry — into a single SQLite database. But the team is playing it safe on session metadata, keeping that in trusty JSON for now.
On the ecosystem side, ClawHub skills can now be deployed straight from a GitHub pinned commit. A rewritten file watcher also fixes the headache of hitting system watcher limits on big skill trees.