What's new in OpenMeld

The Center Agent now runs new turns on GLM 5.3 Flash instead of Grok 4.6. Organizations that did not pick a specific model follow the default on their next turn; organizations that explicitly selected a model keep that exact model, and Grok 4.6 stays selectable.

Center Agent settings now also describes the image path accurately: images use the dedicated vision model before the selected primary model receives the description.

OpenCode Agent Profiles also gain the current curated OpenCode Go, free-tier, and Zen model lineup, including GLM 5.3 Flash on OpenCode Go, with each model's official reasoning variants.

Center Agent replies no longer show an unrelated Computer setup, Space, or Project button merely because the agent checked one of those resources while working. An action appears only when opening that destination is the useful next step for the current request.

Center Agent requests now recover when the private Computer reconnects while a turn is being prepared. OpenMeld resumes the same request from durable state before any model or tool work starts, without repeating completed external actions. If recovery cannot finish, the request reaches one clear failure instead of remaining indefinitely active.

The Center Agent can page through Organization members without loading unrelated Agent Profile readiness, and can read the requester's canonical OpenMeld Tasks separately from local coding-agent task handles.

When Center Agent work ends in failure, OpenMeld now reports that same final outcome in the conversation, progress checks, diagnostics, and the CLI Wake trace summary. Agents no longer keep waiting, treat a failed turn as an ordinary reply, or suggest that a failed response is still being published.

Browser notification setup now disappears completely after permission is granted, “Not now” is selected, or the new close button is used, so members can immediately continue setup or use the page underneath it.

Center Agent diagnostics now keep the time a task first completed when a final confirmation is safely replayed. Repeated confirmations no longer make a completed task look slower than it was.

Agents can now choose and record a Space reply, final status, or silent outcome through a dispatch-scoped OpenMeld tool without asking the member to approve an OpenMeld shell command. This does not relax the Agent Profile's normal Codex permissions. Multi-paragraph replies also preserve real line breaks, literal \n text, quotes, backslashes, emoji, and shell-like characters without rewriting the message. The dispatch-scoped CLI remains available as a fallback and for replies with attachments.

When a selected Local Agent installation clearly cannot run a model, OpenMeld now disables that model while an Agent Profile is created or edited. The dialog explains why and links directly to the Computer where the member can update or change the installation. Missing compatibility information does not block the member or make OpenMeld guess. Existing Agents that hit the same problem during a Wake now get the same direct Computer recovery path.

Center Agent can send formatted Slack messages again and retry the same message when OpenMeld fails while preparing it, before anything reaches Slack. Genuinely uncertain Slack delivery outcomes remain safely blocked so a retry cannot create a duplicate.