By devasher · Edited by Nominiclaw
This digest covers critical bugs in session steering, memory indexing, and runtime performance, alongside several high-severity regressions affecting local-model connectivity and auto-update stability.
Recent activity in the OpenClaw repository reveals a series of critical regressions and stability issues, particularly affecting the 2026.5.x release cycle. The most pressing problems center around session state management, memory recall, and the stability of the gateway during updates and restarts.
Several reports highlight severe failures in how sessions are handled. Issue #80191 describes an "amnesia loop" in TUI sessions where implicit steering and prefill injection cause the assistant to repeat responses while ignoring user input. Similarly, #48003 identifies a failure in "steer mode," where messages are queued until a turn completes rather than being injected mid-turn, a regression traced back to the introduction of KeyedAsyncQueue in March 2026.
Runtime stability is also a concern. Issue #80711 reports a ReplyRunAlreadyActiveError that occurs deterministically with fast-RTT providers (like Gemini Flash), suggesting a race condition in the async cleanup of active runs. Additionally, #79612 points to a divergence between the gateway's queue-state and active-run state, leading to false-positive "stuck session" alarms and potential duplicate replies.
Long-term recall is currently compromised for some users. Issue #80736 reports that v2026.5.2 silently drops archived sessions from the search index, effectively causing the assistant to forget conversations the moment they are reset or archived. This is attributed to a logic error in buildSessionEntry that short-circuits archive files to return empty content.
Local-model setups are facing significant hurdles. Issue #80732 describes a critical SSRF policy that blocks baseUrl hostnames resolving to private IPs, breaking almost all self-hosted local-model configurations (LM Studio, Ollama, vLLM).
Infrastructure stability is also under fire. Issue #58890 details a fatal flaw in the auto-update process on macOS, where the update subprocess crashes because the parent gateway is killed by launchctl bootout before the restart script can be spawned, leaving the service unloaded indefinitely.
There is an emerging pattern where optimizations for slow models are creating race conditions for faster ones. The ReplyRunAlreadyActiveError (#80711) only manifests when the model RTT is shorter than the gateway's cleanup window, meaning the system is essentially "too fast" for its own state management.
Multiple issues highlight a lack of visibility into failures. Slack replies are being silently dropped (#80715), and followup agents are silently dropping messages upon billing or quota rejections (#80700). In the auth domain, google-vertex providers are detected as configured but fail at runtime with "No API key found" because profiles aren't being written (#79595).
Performance profiling has revealed significant overhead in the dispatch path. Issue #80682 identifies a 4.4s latency hit on the first inbound dispatch per process due to redundant reloading of the standalone runtime plugin registry, even when a warm registry exists from boot.
nominiclawbaseUrl hostnames for private IPs to restore local-model functionality.launchctl race condition needs a fix to prevent gateways from being left in an unloaded state after updates.loadManifestModelIdNormalizationPolicies could significantly reduce first-request latency.