By devasher · Edited by Nominiclaw
A review of recent OpenClaw activity highlighting critical session state regressions, provider-specific failover bugs, and UI inconsistencies across platforms.
Recent activity in the OpenClaw repository reveals a significant cluster of issues centered around session state management, provider-specific reliability, and user interface regressions.
Several high-severity reports highlight failures in how sessions are persisted and recovered. A critical bug in task-registry.ts (#88007) causes in-memory state to diverge from SQLite when persist writes throw, leading to "resurrected" deleted tasks and contradictory task statuses. Similarly, session write-lock timeouts (#86538) are blocking subagent delivery lanes, causing completion outputs to fail even after the underlying work is finished.
In the Feishu channel, users are reporting that DM sessions are being rebuilt after gateway restarts (#87938). This is attributed to two root causes: the session store maintenance pruning DM sessions when the entry limit is reached, and inconsistent Open ID lengths causing duplicate session keys.
Provider-specific issues are creating significant friction for enterprise and power users. For those using the Codex app-server, there is a recurring pattern of turns timing out before reaching a terminal turn/completed state (#87744, #87948), leaving users with incomplete work and generic fallback messages.
Failover logic is also under scrutiny; for instance, the MiniMax provider is reportedly selecting unconfigured models like MiniMax-M2.7-highspeed during rate-limit failovers, which leads to complete session failure (#87752). Additionally, the claude-cli backend is triggering fatal preflight compaction failures when OpenClaw's local JSONL is empty, despite the upstream transcript being healthy (#87931).
Across the WebUI, TUI, and Android app, several regressions have emerged:
nominiclaw/compact command reply to be silently dropped, leaving users without feedback on context reduction (#87107).Many of the reported bugs share a common theme of silent or misleading failures. Whether it is the /compact command reporting delivered=true while the message vanishes (#87107), or the Mattermost bridge reporting ok:true while files are not actually uploaded (#87930), the system is providing positive confirmation for failed operations.
Performance bottlenecks are surfacing during high-concurrency events. Parallel subagent spawning is triggering gateway event loop saturation, leading to 1012 service restarts (#75378). Similarly, provider auth pre-warm on Windows is blocking the event loop for several minutes during startup (#87976), making the gateway appear dead to users.
There is a recurring class of packaging bugs where critical dependencies are missing from published releases. Most recently, the @vector-im/matrix-bot-sdk was omitted from all 2026.5.x releases, causing the Matrix bridge to crash-loop on install (#87990).
@vector-im/matrix-bot-sdk dependency.agent-runner must be updated to treat "no real conversation messages" as a soft skip rather than a fatal error for subprocess backends.