OpenClaw Issue Digest: Event-Loop Starvation and Session Stability Regressions

Open Issues

The current reporting window reveals a significant cluster of stability issues centered around event-loop starvation and session state corruption. Most notably, users on Windows and macOS are reporting severe performance degradation and silent failures during complex agent runs.

Critical Stability & Performance

• Event-Loop Starvation: Multiple reports (#86599, #86031, #86951) indicate that local model calls and certain sidecar operations are blocking the Node.js event loop. On Windows, trivial prompts are taking minutes to complete, and the Gateway becomes unresponsive to health probes and WebSocket RPCs.
• Resource Leaks: A high-severity leak in the memory-core plugin (#86613) is causing gateways on macOS to accumulate thousands of file descriptors (FDs) during memory_search calls, eventually leading to EBADF errors and requiring process restarts.
• Codex App-Server Failures: The in-process Codex plugin is reportedly dropping turns silently (#86948), with event-loop utilization hitting 100% and P99 delays exceeding 5 seconds, which is categorized as a beta-blocker.
• Session Lock Races: A recurring issue with EmbeddedAttemptSessionTakeoverError (#86966, #86508) is causing agent runs to abort when session files change while an embedded prompt lock is released, leading to silent message loss in Discord and Telegram.

Functional Regressions

• Kubernetes Permission Errors: A regression in src/infra/exec-approvals.ts (#83619) is blocking all exec tool calls on Kubernetes deployments using fsGroup-mounted PVCs due to an unconditional chmodSync call that returns EPERM.
• Anthropic Session Poisoning: Sessions using the standard anthropic-messages transport are being bricked by empty thinking-block signatures (#86886), a bug previously fixed for Bedrock but missing from the direct API path.
• WebUI Performance: Users report that v2026.5.22 has significantly degraded WebUI responsiveness (#86951), with sessions.list latency jumping from <1s to 8s.

Feature Requests & Improvements

• Security Hardening: A proposal for an outbound DM allowlist (dmAllowTo) to prevent agents from messaging arbitrary users (#86983).
• UX Enhancements: Requests for auto-playing assistant audio attachments in the Control UI (#86894) and a configurable greeting model for the /new command to reduce token costs (#12931).
• Architectural Changes: A proposal to rename the cron subsystem to scheduler to avoid collisions with Unix system cron (#86237).

Key Themes

Event-Loop Saturation

There is a clear pattern of synchronous, CPU-heavy work leaking into the main event loop. Whether it is the sidecars.session-locks phase (#86509) or the Codex app-server's SSE stream handling (#86948), the result is a "frozen" gateway where internal timers (like the Telegram polling watchdog) fire late or not at all.

Session State Fragility

Session management is currently a primary source of silent failures. The combination of optimistic UI updates in WebChat (#86946) and server-side session lock races (#86966) means users often see messages disappear or agents fail to reply without any error being surfaced in the UI.

Environment-Specific Failures

Significant friction is appearing for non-standard deployments. Kubernetes users are blocked by strict permission checks (#83619), and macOS users are facing unique FD leak patterns (#86613) and launchd silent-parking after crashes (#86688).

Action Required

Immediate Attention (High Severity)

• #86613 (Memory FD Leak): Needs urgent investigation into why memory_search is not releasing file handles on macOS.
• #86948 (Codex Turn Drops): This is a beta-blocker causing 100% failure rates for multi-turn tool-heavy sessions.
• #83619 (K8s EPERM): A simple fix (whitelisting EPERM in chmodSync) is needed to unblock Kubernetes users.
• #86599 (Windows Event-Loop Block): Critical for Windows users where local model calls render the gateway unusable.

Blocked or High-Risk

• #86966 / #86508 (Session Takeover Errors): These are causing silent message loss and need a robust fix for the embedded prompt lock lifecycle.
• #86886 (Anthropic Signature Poisoning): This is permanently bricking sessions and requires the Bedrock fix to be backported to the standard transport.