OpenClaw Fixes Unsent iMessage Replies and Strengthens Failed-Update Recovery

A completed AI answer is little use if it never reaches the conversation. OpenClaw’s version 2026.9.9 release notes describe a fix for that failure in iMessage, alongside stronger recovery after failed updates. The release also adds model integrations, though selecting GPT-6.1 Sol still depends on account eligibility and a compatible setup.

Answers that finish but never arrive

The iMessage bug affected ordinary conversations: OpenClaw could finish composing a response without sending it. The project says those affected replies are now delivered. This repair addresses the gap between generating an answer and actually putting it where the user expects to find it.

Discord gets a related delivery correction. A delegated task could post its answer in a thread yet still appear to have failed because it sent no additional reply. OpenClaw now counts that answer as delivered. Webhook replies also use the webhook and identity assigned to the thread.

Keeping a failed upgrade recoverable

Supported updaters can restore a compatible OpenClaw installation after a failed update while preserving data written during the attempt. Returning to version 2026.9.8 also leaves subsequent updates possible. On macOS and Linux, a restart remains blocked until OpenClaw confirms its Doctor maintenance processes have stopped.

Other fixes tackle failures before recovery becomes necessary. Saving data at the same time no longer causes an update to fail. A false maintenance warning that could interrupt an upgrade and unnecessarily restore the previous version has also been corrected. OpenClaw provides repair steps and rollback guidance alongside these changes.

Two compatibility checks before upgrading

• September stable installations keep shared data in the format used by 2026.9.8. Data already converted by an October beta cannot be opened by this release; it remains unchanged.
• Retired configuration settings can require an intermediate upgrade through 2026.9.5 and a repair step. Update errors now explain that route, rather than leaving the required sequence unclear.

New model choices, not new access rights

GPT-6.1 Sol joins the Codex model list, but using it still requires an eligible account, a compatible Codex setup and the related provider support. OpenClaw also supports Claude Haiku 5.5 through Anthropic and Claude CLI, the command-line interface. These are integrations in OpenClaw, not separate model launches.

There is a behavioral fix for Codex as well: when it repeatedly refuses an action, the request can finish rather than remain stuck until a timeout. That change concerns how OpenClaw closes the request; it does not promise that Codex will perform the refused action.