OpenClaw Update: Enhancing Codex Runtime, iMessage Reactions, and Auth Diagnostics

Recent updates to the OpenClaw repository have introduced significant improvements to the Codex runtime, expanded communication channel capabilities, and hardened the authentication pipeline for native operator clients. These changes collectively improve the reliability of large-context model runs and the user experience across multiple integrations.

Merged PRs

• fix(codex): default plugin destructive actions on Original PR
• test: stabilize core fast tests Original PR
• Make OpenAI auth login use ChatGPT by default Original PR
• fix(imessage): route inbound tapbacks as reactions Original PR
• fix(cli/completion): re-apply shell-profile source-line guard Original PR
• docs: reorganize tools automation nav Original PR
• fix(codex): scale context engine projection Original PR
• fix: stabilize code-mode follow-up tool display and replay Original PR
• fix: Fix the build: annotate provider-http test-helper exports for portable dts Original PR
• Telegram: keep verbose tool results separate from final answers Original PR
• docs: add refactor docs skill Original PR
• feat(discord): remove harness runtime dropdown from /models picker Original PR
• fix(codex): match connector app approvals by id Original PR
• Fix iOS operator bootstrap scope diagnostics Original PR
• Mirror Codex native subagents into task registry Original PR
• fix(browser): skip internal CDP targets Original PR
• docs(subagents): document announce timeout Original PR

Key Changes

Codex Runtime and Context Scaling

One of the most critical fixes addresses a context-delivery regression where the Codex app-server projection was silently capping rendered context at 24,000 characters. This effectively hid full-fit context from the model even when the context engine had assembled a larger frontier. The projection cap is now budget-aware, deriving a larger cap from the active contextTokenBudget while honoring user-configured reserve tokens.

Additionally, native Codex subagent lifecycle events are now mirrored into the Task Registry as silent subagent tasks. This provides better visibility into native Codex threads without treating them as OpenClaw-managed child sessions.

Communication Channel Enhancements

• iMessage: Inbound tapbacks (reactions like ❤️ or 👍) are now correctly classified as reaction system events rather than regular text messages. This prevents the agent from treating a "Loved..." notification as a new prompt, which previously triggered unnecessary reply loops.
• Telegram: Verbose tool progress and results are now kept separate from final assistant answers. This prevents the final response from editing or reusing the last tool-result message, ensuring a cleaner conversation history.
• Discord: The /models picker has been simplified by removing the harness runtime dropdown, as runtime resolution is handled via provider routing and should not be a user-facing knob.

Authentication and Security

A significant fix was implemented for iOS and macOS operator bootstrap diagnostics. Previously, a scope mismatch between the client's request and the bootstrap token's allowlist was collapsed into a generic "device token mismatch" error. The system now propagates AUTH_SCOPE_MISMATCH specifically, allowing clients to distinguish between an invalid token and an insufficient scope approval.

Browser and CLI Tooling

• Browser: The CDP (Chrome DevTools Protocol) implementation now filters out internal Chrome targets (e.g., chrome://version/), ensuring that navigation occurs in actual web tabs rather than internal UI components.
• CLI: The openclaw completion --install command's shell-profile source-line guard was re-applied to prevent missing-file errors in login shells after OpenClaw is uninstalled.

Impact

These updates resolve several high-friction pain points for power users and maintainers. The scaling of the Codex context projection ensures that frontier models can actually utilize the large contexts assembled by the LCM engine, removing a silent bottleneck in model performance. The iMessage tapback fix directly addresses a reported bug where reactions were causing agent "hallucinations" or loops by being treated as text.

From a DX perspective, the improved authentication diagnostics for iOS/macOS reduce troubleshooting time for operator pairing, and the stabilization of core fast tests ensures a more reliable CI pipeline. The reorganization of the documentation under a "Capabilities" tab further streamlines the onboarding process for tools, skills, and automation.