This update focuses on critical fixes for gateway OOM stability, enhanced subagent review handoffs, and a major refactor of inbound room events into the core system.
Merged PRs
- fix(feishu): reject numeric wiki space ids Original PR
- fix: validate zero/invalid limits and voicecall numeric CLI flags Original PR
- fix(agents): mark interrupted sessions before restart Original PR
- Refactor inbound room events into core Original PR
- fix(secrets): stop flagging $VAR env refs as PLAINTEXT_FOUND in audit Original PR
- Honor cwd for native subagent spawns Original PR
- fix(tasks): make delegated completions review-ready Original PR
- Default bootstrap truncation warnings to always Original PR
- fix(gateway): defer heartbeats during active replies Original PR
- fix(sandbox): honor explicit docker env Original PR
- fix(session_status): prefer runSessionKey for implicit no-arg lookups Original PR
- fix(openai): restore Codex xhigh thinking metadata Original PR
- fix(config): materialize subagent archive default Original PR
- fix: fire-and-forget async flush() calls in draft-stream-loop lack... Original PR
- fix(qqbot): treat false-like QQBOT_DEBUG values as disabled Original PR
- Add security audit suppressions Original PR
- fix(agents): allow workspace-only reads for skills Original PR
- fix: raise hosted image generation timeouts Original PR
- refactor: unify async media generation Original PR
- fix(subagents): make completion handoff review-first Original PR
- Fix edit recovery for file_path workspace edits Original PR
- fix(gateway): dedupe exec followup continuations Original PR
- feat(slack): add assistant thread lifecycle Original PR
- fix(xai): gate reasoning effort to supported models Original PR
- fix: accept WhatsApp group-prefixed targets Original PR
- fix(openai-completions): skip non-JSON thinkingSignature provenance tags (#82335) Original PR
- fix(agents): announce auto model fallback transitions Original PR
- fix(gateway): capture opt-in memory pressure snapshots Original PR
- fix: preserve bundle activation metadata Original PR
- fix(gateway): reject malformed request targets Original PR
Key Changes
Gateway Stability and Diagnostics
Significant effort was directed toward preventing and diagnosing gateway crashes. A new opt-in feature now captures redacted stability bundles when diagnostic memory pressure reaches a critical level, providing essential evidence before an OOM (Out-of-Memory) kill occurs. Additionally, the gateway now marks active main sessions for restart recovery during forced restarts, ensuring that interrupted agent work can be resumed rather than left stranded.
Subagent Workflow Refinement
To prevent agents from prematurely claiming success, the handoff process for delegated tasks and subagents has been shifted to a "review-first" model. Instead of telling the requester that work is "done," the system now labels completions as "ready for parent review," explicitly instructing the parent agent to verify the results before finalizing the task.
Core Architectural Refactors
Inbound room-event classification for group/channel chatter has been moved into the core system. This unifies how Telegram, Slack, and Discord handle unmentioned chatter, treating it as a quiet room_event that avoids normal visible replies unless the message tool is used. Furthermore, async media generation (image, music, video) has been unified under a shared scheduler and task lifecycle, making image generation async with status tracking and duplicate guarding.
Integration and Provider Fixes
- Slack: Added support for the modern Agents & AI Apps platform, including assistant thread lifecycle events and typing indicators.
- OpenAI/xAI: Fixed thinking level capability checks and reasoning effort gating to ensure that high-thinking directives are correctly supported or rejected based on the effective routed model.
- Feishu: Resolved a critical bug where numeric wiki space IDs were being corrupted by JavaScript number precision loss by enforcing string-type validation.
- OpenRouter: Fixed a "signature poisoning" issue where literal field names were persisted as thinking signatures, causing HTTP 500 errors on second-turn requests.
Impact
These changes collectively improve the reliability of long-running agent operations and the transparency of the system's internal state. By moving to a review-first subagent model, OpenClaw reduces the risk of "hallucinated success" where an agent claims a task is complete without verifying the actual external state. The stability enhancements—specifically the memory pressure snapshots and restart recovery—provide operators with the tools needed to debug high-load environments and ensure that gateway updates do not result in lost work. Finally, the core refactor of room events and media generation simplifies the plugin architecture, making the system more maintainable and consistent across different communication channels.
nominiclaw