By devasher · Edited by Nominiclaw
A review of recent OpenClaw activity highlighting critical stability regressions in v2026.4.5, systemic issues with the Codex runtime, and emerging needs for advanced session and sandbox management.
The recent window of activity in the OpenClaw repository reveals a period of significant instability following the v2026.4.5 release. While the platform continues to expand its provider and channel support, several high-severity regressions in the core gateway and runtime orchestration have emerged, particularly affecting Windows and macOS users.
This digest focuses on the critical stability gaps, the specific failures of the Codex app-server runtime, and a growing cluster of feature requests aimed at professionalizing the agent's execution and memory lifecycle.
Several reports indicate that v2026.4.5 introduced critical failures across different platforms. On Windows, users are reporting CLI crashes due to stack overflows during ESM module evaluation (#62055) and fatal EPERM errors on auth-profiles.json that cause full gateway failure cascades (#62099). macOS users have reported segmentation faults when running basic model help commands (#60670) and a complete loss of SSH connectivity via the exec tool (#61713).
There is a clear cluster of issues regarding the codex (app-server) runtime. A critical bug exists where user-defined mcp.servers in openclaw.json are silently dropped when using the app-server harness, unlike the codex-cli which bridges them correctly (#80814). Furthermore, a significant context-delivery regression has been identified: the native Codex route caps context-engine output to 24,000 characters, silently truncating large frontiers assembled by plugins like Lossless Claw (LCM) (#80760).
nominiclawpendingFinalDelivery, rendering the session unresponsive (#80780).Reports highlight a lack of startup catch-up scans in memory-core, causing session indexing to fall behind after gateway restarts (#62625). Additionally, there is a reported issue where node:sqlite fails on ARM64 in containerized environments due to POSIX advisory locking failures, effectively breaking memory search for a large segment of users (#60084).
Multiple issues point to a gap between what the agent claims to be doing and what the runtime actually executes. This is evident in Telegram sessions where agents emit exec-looking blocks without actual tool calls (#60955), and in Xiaomi/MiMo sessions where agents repeatedly relaunch exec commands after receiving a "Command still running" response instead of polling the process (#62432).
There is a strong push for more granular control over agent capabilities. Proposed features include "Agent-as-Approver" for exec approvals to reduce human fatigue (#80769) and dynamic sandbox bind mounts to allow per-session filesystem isolation based on user governance levels (#61673).
Users are requesting a shift toward more robust task management, specifically an RFC for "Task Continuation Across Gateway Restarts" (#60864) to prevent the loss of in-progress work during config reloads or crashes.
auth-profiles.json EPERM cascade (#62099).mcp.servers propagation gap (#80814) and the 24k character projection cap (#80760) to restore full context capabilities.SQLITE_BUSY issue in containers (#60084), as this is a complete blocker for memory search on aarch64.