By devasher · Edited by Nominiclaw
This update focuses on critical stability fixes for macOS gateways, a major refactor of transcript management into the core system, and improved visibility for agent session scoping.
Several high-priority fixes address critical resource leaks and crash loops, particularly on macOS:
memory-core was causing the gateway to open one FD per .md file in the memory tree (reaching >12,000 FDs in some cases). The system now uses Node 22's native recursive fs.watch for directories, reducing REG FDs by 99.99%.ENETDOWN as a transient network failure for Telegram outbound connections. Previously, this triggered an uncaught exception that crashed the gateway and, on macOS, could lead to launchd silently parking the service, causing multi-day outages.--profile) are respected during state migration.meeting-notes plugin has been deprecated and integrated directly into the core system as transcripts. This centralizes capture, storage, and CLI management, requiring an explicit transcripts.enabled: true configuration for recording-capable tools.agents.defaults.compaction.maxActiveTranscriptTokens, allowing operators to adjust the 70k token limit to match their specific deployment needs.sessions_list tool now explicitly discloses when results are scoped (e.g., via tree visibility), including a warning that the session count reflects only visible sessions./command@TargetBot format are now correctly treated as mentions, preventing them from being dropped by the requireMention gate./btw command for MiniMax models are now routed through the embedded stream resolver, ensuring they adhere to token caps and avoid max_tokens errors.${DISCORD_BOT_TOKEN:-}) in session history, preventing the mutation of transcripts that previously broke KV cache prefix matching for local LLM providers.These changes significantly improve the operational reliability of OpenClaw for power users and enterprise deployments. By resolving the ENETDOWN crash loop and the memory-core FD leak, the gateway is far less likely to experience silent failures or degrade toward resource exhaustion on macOS.
From a developer and operator perspective, the transition of transcripts to core reduces plugin overhead and simplifies the configuration surface. The improved session visibility and logging redaction directly benefit those using local LLM providers by maintaining byte-stable transcripts, which is critical for maximizing the efficiency of KV cache reuse and reducing compute costs during long conversations.