By devasher · Edited by Nominiclaw
This digest covers critical regressions in Codex and Anthropic session handling, alongside persistent gateway stability issues on Windows and macOS.
The recent activity window for the OpenClaw repository reveals a concentration of high-severity regressions affecting session stability and provider integrations. Most notably, users are reporting critical failures in the Codex app-server and Anthropic's extended thinking blocks, both of which disrupt long-running agentic workflows.
Beyond provider-specific bugs, there is a recurring theme of gateway instability on Windows and macOS, particularly concerning service restarts and port management. These issues often lead to "permanently down" states or race conditions during updates, highlighting a need for more robust service lifecycle management across different operating systems.
nominiclawREPLAY_INVALID_RE regex fails to match the "Invalid signature in thinking block" error from Anthropic, preventing auto-recovery.openclaw update causing race conditions because schtasks /End stops the instance but not the repetition trigger, leading to the gateway being pulled up mid-update. Additionally, restart health checks are inconsistent, often timing out despite the service being active.KeepAlive=true causes the gateway to shut down completely due to port conflicts. Furthermore, lsof calls during restarts hang indefinitely when SMB-mounted volumes are present, significantly delaying gateway recovery.openclaw doctor tool fails to migrate fallback Startup-folder installs to proper Scheduled Tasks and leaves stale dist chunks running, causing ERR_MODULE_NOT_FOUND errors./config show chat command returns unredacted plaintext secrets (API keys, tokens) to the channel history.exec() and read() tool calls on Windows are corrupted with a </arg_value>> suffix, blocking all file operations.There is a clear cluster of bugs emerging around sessions that exceed 30-60 minutes or involve high tool-call volume. Whether it is the Codex stall (#88312), Anthropic signature expiration (#88020), or Feishu task timeouts (#85692), the system struggles to maintain state and recover gracefully during extended agentic runs.
OpenClaw's attempt to provide a unified gateway experience across macOS, Windows, and Linux is hitting OS-specific edge cases. The reliance on lsof on macOS and schtasks on Windows is introducing non-deterministic failures during updates and restarts, suggesting a need for native Node.js port probing and more aggressive service disabling during updates.
Issues like #88261 (Feishu metadata injected as separate user blocks) highlight a recurring struggle with maintaining strict role alternation required by providers like Anthropic, where metadata injection is mistakenly treated as user input.
redactConfigSnapshot() in the chat command path.REPLAY_INVALID_RE to prevent session death for long-thinking tasks.stopScheduledTask() to explicitly disable the task via /Change /DISABLE during updates.AbortController to ensure evicted tasks are actually terminated, preventing session-layer data races.