OpenClaw: Performance Optimizations and Secret Management Fixes

Transcript

Good morning. This is your OpenClaw developer briefing for May 17th, 2026.

Peter Steinberger merged several critical fixes today. PR #83020 resolves scoped web command SecretRefs, fixing how local web search handles plugin credentials. PR #82892 routes image URL descriptions through MiniMax VLM, improving image understanding capabilities. PR #83025 shortened agent tool descriptions across the platform for better performance.

Josh Avant contributed performance improvements with PR #83033, which reuses plugin metadata snapshots to speed up the slow `openclaw models` command path. PR #83039 fixes subagent completion announce delivery timing, ensuring proper transcript commit before marking announcements as delivered.

Sam Zong optimized startup performance with PR #83031, speeding up gateway secrets initialization by avoiding full runtime imports when no SecretRefs are present. PR #82991 enables reuse of gateway startup auth preflight snapshots when configurations match.

Vincent Koc fixed PR #83029, preserving nested tool-result middleware output in Codex that was previously showing as "post-processing error." PR #82624 adds malformed transcript state entry handling to skip corrupted JSONL entries.

Additional merged work includes message tool session-key routing fixes, provider conversation state error classification, ACP timeout handling improvements, and reply media failure surfacing. Three new skills were added: meme-maker, debugger diagram, and spike prototyping tools.

The team also pushed 30 additional commits covering CLI secret resolution improvements, QA lab cursor polling fixes, and various test coverage enhancements.

What's next: Focus remains on gateway performance optimizations and expanding the skills ecosystem. Secret management reliability continues as a priority area.

That's your OpenClaw update. Back to building.