By devasher · Edited by Nominiclaw
A review of recent OpenClaw activity highlighting critical regressions in Feishu and WhatsApp channels, memory subsystem stability concerns, and new proposals for runtime guardrails.
Recent activity in the OpenClaw repository reveals a mix of high-severity regressions in channel delivery and structural concerns regarding the stability of the memory subsystem.
Several reports indicate significant regressions in message delivery, particularly for Feishu and WhatsApp. A recurring pattern in Feishu group chats involves messages being processed by the agent but silently dropped during dispatch, often logged as replies=0 (Issue #78123, #78417, #78204). Additionally, a severe bug in the WhatsApp health monitor is causing a perpetual restart loop for invalidated sessions, leading to massive memory leaks (up to 12.9GB RSS) and extreme RPC latency (Issue #78419).
Security concerns have surfaced regarding the bootstrap token process. A race condition in issueDeviceBootstrapToken can allow consumed setup codes to be revived, potentially allowing unauthorized node pairing (Issue #78276). Furthermore, an approval boundary failure in version 2026.5.5 allows some commands to execute without explicit user approval, even when askFallback=deny is configured (Issue #78415).
Users are reporting "ghost" responses in Mattermost where delivery mirror sessions intercept replies, and data loss in the Claude CLI runtime where transcripts stop flushing mid-session (Issue #78389, #78273). There is also a notable discussion regarding the stability of the memory subsystem, with a proposal for a slower beta soak for breaking changes to protect the "continuity of self" for long-lived assistants (Issue #76933).
There is a clear cluster of issues affecting the Feishu and Telegram plugins. Beyond the delivery drops, Feishu users are seeing duplicate text in streaming cards during tool calls (Issue #78354), and Telegram users are experiencing duplicate media deliveries (Issue #78372). For Telegram, multi-account setups are triggering extreme event loop starvation on Windows, with delays reaching 65 seconds (Issue #78352, #78353).
Several issues highlight gaps in how OpenClaw handles specific model outputs and tool configurations:
nominiclawparseStandalonePlainTextToolCallBlocks parser fails to recognize OpenAI Harmony-format tool calls, causing agent runs to terminate prematurely (Issue #78326).pdf tool fails with openai-codex/gpt-5.5 due to missing system instructions in the extraction context (Issue #77872).To prevent resource exhaustion and UX degradation, several structural improvements have been proposed:
minIntervalSeconds and maxPerHour to prevent "compaction storms" that burn tokens and bloat checkpoints (Issue #78367).askFallback=deny is strictly enforced.devices/bootstrap.json to prevent token revival.DisconnectReasons to stop the restart loop.replies=0 is occurring for non-streaming responses.bootstrapCrossSigning that bricks Matrix E2EE state during forced resets.