OpenClaw Update: Enhancing Session Routing, Memory Performance, and Plugin Reliability

Merged PRs

• fix(doctor): include channel model provider repairs #83328
• fix(export): preserve explicit trajectory session keys #83308
• fix(update): guide EACCES manual recovery #83757
• refactor(channels): unify session route projection #83644
• fix: bypass npm freshness for managed installs #83761
• fix(anthropic): preserve Claude image capability #83756
• fix(memory-core): yield event loop during fallback vector search #83758
• fix(github): preserve clawsweeper proof labels #83781
• fix(cli): fix flaky config set help text test #83423
• fix: keep inter-session provenance out of transcripts #83755
• fix(outbound): resolve send-capable channel registry #83733
• fix(auto-reply): honor webchat textChunkLimit/chunkMode config overrides #83742
• fix(ui): render session-scoped tool events #83734
• fix(ci): preserve Barnacle proof labels #83735
• Doctor: convert read-only health checks #83198
• fix(doctor): anchor WhatsApp TUI process matching #83313
• fix(models): label picker auth via effective provider order #83726
• test(qa-lab): add personal share-safe diagnostics scenario #83717
• fix(nextcloud-talk): dispatch react action so agents can send reactions #72348
• fix(subagents): collect unresolved announce batches #83701

Key Changes

Core Architecture & Session Routing

One of the most significant changes is the unification of session route projection. Previously, channel delivery route metadata was handled via ad hoc plugin hints, which led to routing failures in complex scenarios like Discord subagent threads. The new ChannelRouteRef provides a canonical route metadata structure owned by the core, allowing the system to derive compatibility delivery fields (like lastChannel and lastThreadId) from a single source of truth. This ensures that follow-up deliveries are routed correctly regardless of whether they originate from a main session or a subagent thread.

Performance & Stability

Critical improvements were made to the memory-core plugin to prevent event-loop starvation. In cases where a native sqlite-vec index was missing, the fallback vector search performed a synchronous cosine-similarity loop over the entire memory corpus. For large datasets, this pinned the Node.js main thread for tens of seconds, causing gateway heartbeats to fail and agents to hang.

"The fallback vector path scans every row of the chunks table with a synchronous JS-side cosine-similarity loop... that loop pins the Node.js main thread for tens of seconds."

To resolve this, the search now operates in bounded batches (LIMIT 256) and yields to the event loop using setImmediate between batches, maintaining gateway responsiveness even during heavy memory scans.

Plugin & Provider Fixes

• Anthropic Vision: Fixed a bug where native images were dropped during sanitization for Claude 4.x models due to stale configuration rows marking the models as text-only. The system now normalizes these models back to text+image capability.
• Nextcloud Talk: Wired the sendReactionNextcloudTalk HTTP sender to the agent-facing action surface, enabling agents to finally send emoji reactions.
• Outbound Delivery: Resolved an issue where pinned setup-only channel registries could hide full runtime plugins, causing intermittent "Outbound not configured" errors for channels like Discord.
• Subagent Announces: Fixed a bug in collect mode where a single unkeyed item would poison the entire batch, forcing individual delivery and leading to silent message loss when the orchestrator session was busy.

Tooling & DX

• Doctor Command: Converted read-only health checks into structured detect() findings, allowing doctor --lint to report issues without triggering repair mode. Additionally, the WhatsApp TUI process matching was anchored to prevent the accidental termination of unrelated processes.
• Update Process: Improved recovery guidance for EACCES permission failures during global npm installs, specifically instructing managed-Gateway operators to stop the Gateway before performing manual sudo recovery to avoid transient half-swapped package states.
• WebChat: Removed hardcoded truncation limits for live responses in WebChat, allowing users to configure textChunkLimit and chunkMode as they do for other channels.

Impact

These updates collectively harden the OpenClaw gateway against production-scale failures. The memory-core fix eliminates a primary cause of agent "freezes" during large-scale RAG operations, while the route projection refactor provides a more robust foundation for multi-channel subagent orchestration.

For users, the most immediate impacts are the restoration of vision capabilities for Anthropic models, the ability to use reactions in Nextcloud Talk, and a more reliable experience when using the WebChat interface for long-form responses. The improved doctor and update guidance reduces the risk of operational errors during system maintenance on Linux environments.