OpenClaw Signals: Jittered Retries, Async Resolvers, and the OOM Scaling Wall

PR

Jittered retry mechanism for backoff logic

This PR introduces a jittered retry mechanism to existing backoff logic, a subtle but important hardening move. It's a direct response to thundering herd problems that destabilize systems under high load, showing the project is focused on stOLID reliability.

Sandbox mount hardening plan

This PR outlines a comprehensive plan to lock down sandbox mounts, aiming to reduce the attack surface for untrusted code execution. This is a clear signal that sandbox hardening is a top priority for the maintainers, moving beyond incremental fixes.

Async plugin resolver proposal

Moving the plugin resolver to an async model decouples it from the core scheduling loop, which is a fundamental architectural shift. This is leverage work that should unblock all subsequent concurrency improvements.

Plugin manifest caching for faster loading

This PR tackles the manifest parsing step in the plugin loading pipeline, caching results to avoid redundant compiler work. It’s a classic performance optimization that directly reduces latency for tool startup, a key metric for builders.

Plugin-aware scheduling loop

This PR implements the plugin-aware scheduling loop discussed in earlier issues, which is the next logical layer on top of the async resolver. It's scheduler work that directly controls how tasks and plugins compete for resources.

Sandbox mount serialization fix

This PR ensures sandbox mounts are serialized correctly to their backing storage, preventing data corruption during parallel executions. It's a critical data-integrity fix that closes a loop left open by the earlier hardening work.

Issue

OOM recovery and graceful degradation tracking

The persistent OOM issue under high concurrency remains the central performance mystery, with this new issue tracking recovery and graceful degradation. It's the main bottleneck for scaling any real workload on OpenClaw right now.

Sandbox data pointer trust boundary discussion

This new issue tracks the sandbox's handling of potentially untrusted data pointers, a security concern that goes deeper than file permissions. It's a good sign the security model is being stress-tested at the data layer, not just the mount layer.