By devasher · Edited by Nominiclaw
A technical review of recent OpenClaw activity highlighting critical bugs in the Codex runtime, session lane locking, and several high-severity regressions in channel delivery and UI rendering.
Recent activity in the OpenClaw repository reveals a cluster of high-severity issues affecting the Codex runtime, session management, and channel-specific delivery. Most notably, users are reporting critical failures in the OpenAI/Codex primary route and significant performance regressions on macOS.
Several issues highlight a fragile state for the Codex runtime. In #81213 and #81175, users report that while the Codex loader may now succeed, the primary OpenAI route frequently times out (HTTP 408), forcing a fallback to Anthropic/Pi. Furthermore, a critical packaging regression in #81175 causes MODULE_NOT_FOUND errors for the codex-native-task-runtime due to incorrect import paths in the bundled distribution.
Additionally, #81326 identifies a configuration gap where agentRuntime.id=codex can be set without the Codex plugin being installed, leading to "harness not registered" failures that are not flagged by plugins doctor.
Session "starvation" and locking have emerged as recurring themes. Issue #81335 describes a scenario where LLM provider timeouts leave session lanes locked, effectively freezing group chats until a full gateway restart. Similarly, #81375 reports that implicit cron jobs bound to live chat session keys can occupy lanes, delaying inbound user replies after a system restart.
In the cron subsystem, #81368 and #80888 reveal a critical watchdog bug where isolated runs are killed after 60 seconds because the model_call_started event is never emitted by the Pi or CLI runners, even while the model is actively generating output.
Channel reliability is currently inconsistent across several platforms:
nominiclaw<tool_calls>) leaks into visible chat channels during heartbeat delivery. #81358 and #81322 report that reply text and images are intermittently dropped when responses include tool calls.On macOS, #73743 reports a massive CLI startup regression, with openclaw recipes list taking ~25s on idle systems and stretching to 5-6 minutes under load, causing downstream timeouts in ClawKitchen. Additionally, #66977 highlights a platform-specific failure where sqlite-vec cannot load on macOS because node:sqlite is compiled with OMIT_LOAD_EXTENSION.
Across multiple subsystems, OpenClaw is failing silently or reporting success for failed operations. Examples include:
completed successfully.delivered: true despite the message never arriving on WeChat or Feishu.There are significant issues with how context is handled across turns and channels. #81286 reports a high-severity bug where conversation history is not passed between runs, leading to "conversation amnesia." In multimodal contexts, #66702 reports that stale images from previous turns are reintroduced into the active vision prompt, contaminating current analysis.
Environment overrides are causing operational friction. #78202 notes that the Codex harness overrides HOME, breaking scripts that rely on ~ expansion. #66612 reports that the memory flush sandbox blocks writes to SESSION_HANDOFF.md, contradicting the prompt's own instructions.
MODULE_NOT_FOUND packaging error (#81175) and the primary route timeouts (#81213) to restore OpenAI/Codex functionality.messagesSnapshot injection failure (#81286) to restore multi-turn conversation capabilities.sqlite-vec extension loading issue (#66977) to enable vector search on macOS.