By devasher · Edited by Nominiclaw
This digest analyzes recent OpenClaw repository activity, focusing on critical regressions in the memory dreaming pipeline, execution tool stability, and provider-specific authentication drifts.
Recent activity in the OpenClaw repository reveals a significant volume of behavioral bugs and architectural gaps, particularly concerning the memory-core plugin, the exec tool's stability across platforms, and provider-specific regressions for Claude CLI and Gemini.
Several issues highlight instability in the "Dreaming" pipeline. Users report that the light-phase work summary repeats verbatim across cycles (#72096) and that the dreaming cron reports success while writing no artifacts (#72546). Furthermore, the dreaming corpus is frequently polluted by cron-triggered noise (#68449), and the light-sleep sort order prioritizes recency over actual recall counts, effectively hiding meaningful data (#71976).
Stability issues with the exec tool are prevalent across different operating systems. On macOS, users are experiencing intermittent SIGKILL events during broad diagnostic commands (#72240), while Windows users report spawn EPERM errors that block all commands (#71717). In sandboxed environments, there are reports of uv_interface_addresses errors causing gateway disconnects (#71721) and a failure to propagate sandbox.docker.env variables to subagents (#69327).
Provider-specific issues are surfacing, particularly with the claude-cli backend. Reports indicate that sessions are reset on every turn in group channels due to extraSystemPromptHash drift (#69118) and that the backend skips context engine maintenance, breaking LCM persistence (#81558). Gemini users are reporting intermittent 0-token stalls in isolated subagent sessions (#80714) and 404 errors when spawning sub-agents with specific preview models (#71932).
User experience issues in the Control UI include duplicate assistant replies (#71992, #72892) and a "flash" of the credentials gate during tab re-focus (#72500). On the channel side, the Telegram outbound path is leaking reaction directives as literal text (#71140), and the Feishu bot is stripping its own mentions, leading to NO_REPLY in multi-bot groups (#72504).
Across multiple subsystems, OpenClaw is exhibiting a pattern of silent failures where the system reports ok or success while the actual output is missing or degraded. This is seen in the Dreaming cron (#72546), the claude-cli persistence gap (#81558), and the gateway's handling of empty fallback chains (#81459).
There is a recurring theme of process lifecycle mismanagement. This includes the agent/cli-backend watchdog killing sessions that are correctly idle during Monitor tasks (#71803), and openclaw-agent processes ignoring SIGTERM under cron, leading to RAM exhaustion (#71710).
Session "amnesia" is a critical pain point. Whether caused by missing-transcript races in Telegram DMs (#77974) or system-prompt hash drift in group channels (#69118), the inability to maintain conversational state across turns is significantly degrading the agentic experience.
nominiclawexec (#72111): The exec tool returns raw stdout/stderr to agents without redacting secrets, posing a high internal exposure risk.GATEWAY_CLIENT identity and bypass device pairing.STATUS_STACK_BUFFER_OVERRUN crashes during Mattermost streaming replies.runContextEngineMaintenance in the CLI path effectively disables LCM for a large segment of users.image tool is currently broken for Bedrock users due to an unconditional requireApiKey call that ignores aws-sdk mode.