[Release] openclaw/openclaw v2026.5.12-beta.2: openclaw 2026.5.12-beta.2

OpenClaw v2026.5.12-beta.2 releases critical fixes and enhancements across authentication, provider streams, memory management, and UI/UX. Key improvements include auth-profile-backed media tools availability, subagent session visualization, error handling in auto-reply, and provider stream reliability. The release also upgrades to pnpm 11, adds per-agent tool policies, expands Slack/Gemini/Fal provider support, and strengthens TypeScript/linting standards.

Key Points

• •Auth-profile-backed media tools (e.g., image_generate) now available when OpenAI auth lives in agent's auth-profile store instead of environment variables
• •Subagent sessions now nested under parent sessions in session picker dropdown with visual `└─` prefix for clarity
• •Auto-reply surfaces visible errors when model backend fails and fallback produces no reply, while preserving intentional silent turns
• •Provider streams improved: OpenAI-compatible SSE/JSON fallback streams drain across split chunks; Azure Response streams fail with bounded diagnostics instead of stalling
• •Per-sender tool policies added with channel-scoped sender keys to restrict dangerous tools by requester identity across all tool surfaces
• •Workspace upgraded to pnpm 11 with aligned Docker, install, update, and release workflows
• •Gemini 3 Pro Preview IDs normalized to google/gemini-3.1-pro-preview across SDK OAuth, CLI auth, and provider catalog
• •Fal provider enhanced: GPT Image 2 and Nano Banana 2 edit requests routed to `/edit` with image_urls array; input-image caps lifted to 10 and 14 respectively
• •Per-agent tool policy overrides added for `tools.message.crossContext` and `tools.message.actions.allow` to restrict message sends and enforce send-only tools
• •Slack enhancements: `unfurlLinks`/`unfurlMedia` config added for bot replies; explicit `replyBroadcast` support for thread replies

Workflow Diagram

Concepts