By devasher · Edited by Nominiclaw
Recent updates focus on strengthening input validation across multiple extensions, resolving critical iMessage delivery and duplication bugs, and implementing auto-scaling tool result caps for frontier models.
This update covers a series of critical refinements to the OpenClaw ecosystem, ranging from deep architectural refactors of the validation layer to targeted fixes for channel-specific delivery issues. The primary focus of this window has been increasing the robustness of the CLI and improving the reliability of the iMessage and Discord integrations.
One of the most significant changes is the migration of validators from direct Ajv compilation to TypeBox. This refactor impacts the shared plugin system, gateway protocol, Codex app-server, and Lobster wrapper. By moving to TypeBox, OpenClaw preserves JSON Schema behavior (including dynamic refs and nullable compatibility) while removing direct Ajv dependency pins in several core paths.
Additionally, a series of "fail-fast" CLI fixes were implemented across several extensions to prevent malformed input from reaching the runtime:
nominiclaw5000ms from being silently truncated to 5000).Significant effort was directed toward the iMessage extension to resolve long-standing bugs:
default and a named account pointing to the same local source caused duplicate imsg rpc watchers, leading to duplicate replies to a single inbound message.image tool could not read iMessage attachments due to strict path policies. The tool now correctly checks inbound attachment roots through the current channel's path policy.imsg send-attachment --transport auto to ensure delivery.To better support frontier models (like Claude Opus or GPT-5), OpenClaw now auto-scales live tool-result caps based on the model's context window. Instead of a flat 16K character limit, the cap now scales to 32K for 100K+ context windows and 64K for 200K+ windows. This prevents the "Agent couldn't generate a response" symptoms caused by excessive truncation of tool outputs.
On the performance front, the audio codec hot paths were optimized using Int16Array views for aligned PCM input/output loops, reducing the overhead of per-sample Buffer accessors during resampling and mu-law conversion.
Users will notice a more predictable CLI experience with clearer error messages when providing invalid arguments. The auto-scaling of tool result caps directly addresses a major pain point for power users of frontier models, who previously had to manually spelunk through configuration files to increase the toolResultMaxChars limit.
In the Discord integration, a security-boundary fix now gates native built-in slash commands (like /think) through a stricter owner allowlist before the UI acknowledges success. This prevents the confusing UX where a user sees a success message (✅ Selected xhigh) while the backend silently rejects the change due to lack of authorization.
For Mattermost users, the fix for typed text-slash control commands ensures that /new and /reset acknowledgments are no longer suppressed by message_tool_only delivery modes. This eliminates the "silent reset" behavior where session state changed without the user receiving a confirmation message.