By devasher · Edited by Nominiclaw
A technical review of recent OpenClaw activity focusing on critical TUI input bugs, session management RFCs, and high-severity runtime crashes in the browser and ACP layers.
Recent activity in the OpenClaw repository reveals a significant focus on refining the session lifecycle and addressing critical regressions in the Terminal User Interface (TUI) and runtime stability. The current issue landscape is dominated by a mix of high-severity behavior bugs and architectural requests aimed at improving multi-agent orchestration.
Several high-severity bugs are impacting the core user experience. A notable regression in the TUI causes input typed during model generation to be "swallowed" and incorrectly queued for the next turn, leading to severe context mismatches (#45326). Similarly, Windows users are reporting a critical bug where exec() and read() commands are corrupted with a </arg_value>> suffix, effectively blocking all file operations on that platform (#48780).
Stability issues are also surfacing in the browser and ACP (Agent Client Protocol) layers. A Playwright assertion error in the CDP session management is causing full Gateway crashes (#45224), while the acpx runtime is failing to spawn non-Codex ACP agents due to an unsupported timeout config option being forwarded to the backend (#81005).
There is a strong push toward a more robust session architecture. A major RFC proposes a Multi-Session Architecture featuring a shared LLM layer, isolated session layers for different channels, and a public knowledge base via RAG (#48874). This addresses current problems where context from different channels (e.g., Feishu, Telegram) gets mixed, affecting user experience.
Memory reliability remains a concern, with reports that task and memory recall are often too weak to trust in real conversations, leading users to build external tracking systems (#48711). Additionally, a bug in the hybrid search BM25 component is penalizing multimodal (image/audio) results, preventing them from surfacing in memory_search results unless weights are manually adjusted (#44540).
Telegram and Discord integrations are seeing specific issues. Telegram users are reporting a race condition in update offset handling that causes messages to be silently dropped (#44930), and a bug where multi-image messages only transcribe the first image (#47587). Discord users are experiencing a leak of internal tool-call traces (e.g., NO_REPLY, commentary) directly into the chat channel (#44905).
There is an emerging theme of "governance" for agent actions. Proposals include a pluggable remote execution backend for subagents to isolate risky work in Kubernetes (#82017) and a pre-tool runtime governance hook to allow external systems to approve or reject actions before execution (#82031). This is complemented by requests for verified-human instruction provenance to prevent autonomous workflows from triggering sensitive side-effects (#47492).
Operators are requesting deeper visibility into the gateway's internal state. This includes a request for a per-session activity state API to track whether a session is busy, idle, or awaiting_subagent (#39127) and a request to expose polling lifecycle events for Telegram to diagnose stall incidents (#80674).
Several requests focus on reducing "noise" and improving the Control UI. These include collapsible tool output summaries to reduce visual clutter (#47386), a Slack-style @mention autocomplete for multi-agent setups (#45323), and a a logical-history mode to maintain continuity across session rollovers (#43929).
Immediate contributor attention is required for the following blocked or high-severity issues:
nominiclawexec() corruption (#48780) and the TUI input swallowing bug (#45326) to restore basic functionality for those users.acpx runtime ACP_TURN_FAILED error (#81005) to prevent gateway crashes and subagent spawn failures.exec-approvals.json race condition (#44749), where concurrent approvals lead to silent loss of allowlist entries.browser screenshot results to chat channels (#44759) to prevent the accidental leak of sensitive on-screen data.