[Release] openclaw/openclaw v2026.6.5-beta.5: openclaw 2026.6.5-beta.5

OpenClaw v2026.6.5-beta.5 introduces critical stability and safety improvements across multiple channels and providers. Key fixes include QQBot reasoning stripping, MCP tool result coercion to prevent API errors, Anthropic extended-thinking session recovery, and Parallel web-search bundling. The release strengthens auth durability, provider model resolution, and agent/tool/provider loop safety while improving macOS node mode reliability and upgrade path safety.

Key Points

• •QQBot now strips model reasoning/thinking scaffolding before delivery to prevent raw `<thinking>` content leaking into channel replies
• •MCP tool results coerce non-text/image blocks at materialization boundary to prevent Anthropic 400s and poisoned session history
• •Anthropic extended-thinking sessions recover after prompt-cache expiry or Gateway restart by waiting for `message_start` events
• •Parallel is now a bundled web_search provider with `PARALLEL_API_KEY` discovery, guarded endpoints, and cache-safe session IDs
• •Google Vertex ADC users regain static catalog rows and runtime model resolution with improved single-provider cooldown recovery
• •Matrix channels support voice-note preflight and thread-aware read/reply behavior with QA coverage
• •Auth profiles now live in SQLite; official npm plugin installs keep trusted pins; prerelease fallback integrity checks avoid stale data
• •Agent/tool/provider loops enforce stricter MCP lease timestamps, prompt-cache tool names, local catalogs, and owner-only HTTP tool gating
• •macOS node mode no longer silently self-reconnects away from healthy direct Gateway sessions, reducing companion app churn
• •Upgrade and service paths are safer: cron legacy JSON migration, masked service env secrets, bounded WhatsApp startup waits, and config-reload teardown

Workflow Diagram

Concepts