OpenClaw Signals: Concurrency Fixes, Plugin Sandboxing, and the Backoff Tuning Debate

Tools

PR #329: Fixing a Race Condition in Plugin Resolution Under Load

This patch closes a race window that showed up when multiple plugins tried to resolve dependencies concurrently under heavy load. If you've been running OpenClaw in production with a lot of plugins loaded at once, this is the fix you want before you scale further.

PR #328: Manifest Cache Invalidation on Hot Reload

Turns out the manifest cache wasn't invalidating properly on hot reload, which meant stale plugin configs could silently persist. Small fix, big deal if you're iterating fast on plugin configs without full restarts.

PR #327: Sandbox Filesystem Isolation Tightened for Untrusted Plugins

More sandbox hardening, this time locking down filesystem access for plugins that haven't been explicitly trusted. Good move if you're running third-party plugins you didn't write yourself, which is basically everyone at this point.

Analysis

Issue #1560: OOM Recovery Still Flaky Under Sustained Concurrency

The memory leak hunt continues and it's still not fully resolved, with maintainers now debating whether the fix needs to happen at the scheduler level instead of patching individual plugins. Worth watching closely if you're running anything long-lived or high-throughput, this is the one bug that keeps biting people in prod.

Discussion Thread: Should Retry Backoff Be Configurable Per-Plugin

A side debate spun out of the concurrency fix PR about whether retry backoff timing should be global or per-plugin configurable. The practical answer is per-plugin, since a flaky external API plugin shouldn't force conservative backoff on everything else in your pipeline.