What's new in OpenMeld

Production release checks require evidence for every changed receiver contract. An unchanged AI usage receiver can be verified from its recorded deployment source, while other pending contracts still require their own evidence or a receiver deployment. This prevents one compatible contract from admitting unverified changes between the same services.

Center Agent working-folder updates require the new Core receiver to deploy before their Remote caller. Unrelated AI usage evidence cannot admit these updates while that receiver is still pending.

Agent setup and folder recovery also wait for their Database, Auth, Core and Gateway receivers. Cloud command references and detailed lists require Core and Remote Agent before Gateway exposes them, then Gateway before the CLI. The checks apply to caller-only changes and reject unavailable source evidence.

Saved working-folder reads require Core before the new Service registration, Core's read methods before Gateway, and Gateway's GET routes before Web. Other folder or runtime evidence cannot substitute for this read contract. Workspace initialization also requires Core before the Service advertises its strict initialization capability, including a Service-only source change. Initialization and read proofs must each cover their own contract.

When an Agent needs a Plugin account or permission, you can open the existing connection and access controls from its message. Save the access you choose, then continue the original task there. New connections grant no Agent access until you approve it.

Slack joins the featured Plugin catalog. Slack, Lark, and Feishu use the live OpenConnector configuration for sign-in readiness. A missing OAuth client now shows a clear setup message instead of offering a connection that cannot start.

Thread replies share Profile cards, message tools, quoted replies, queued follow-ups, unread navigation, and Jump to latest with the main conversation. Queued replies stay in their original conversation. The composer work shortcut also finds queued replies in closed Threads and opens the right Thread, alongside other work outside your view. Full-width messages reserve room for avatars on both sides, so long messages have matching content edges. Members can open an existing Team Brief, while creation stays with owners and admins.

Center Agent Profile cards show the model and a cloud icon, with a direct settings action. Routine availability and default reasoning labels are omitted. App Settings also omits the routine Available label. Bot avatars in Profile cards look around, follow the mouse, and play a full spin when clicked. Conversation avatars retain their small click hop.

Choose a separate working copy without adding extra restrictions to your Agent's tools or permitted network access. Working copy and Access are separate choices in the Agent menu, and CLI Project selection keeps Agent settings by default. Choose Session → Automatic to clear those task choices and return to automatic continuation. An existing session keeps its own restrictions.

OpenMeld manages the copy and retains unfinished changes. Computer file policies, native approvals, and explicitly restricted tasks continue to apply. Forking an independent task keeps its separate-copy mode.

For customer organizations in production, model logs keep usage and status information without a copy of the conversation. Automatic agent feedback sends problem categories and diagnostic IDs instead of generated summaries or details. User-submitted feedback still includes the information you choose to share.

Diagnostic Report IDs resolve through the current database to compact execution traces. Authorized infrastructure operators can still investigate stored content with your consent. Task execution and conversation history remain available.

Existing Spaces apply the database upgrade for deleted Space history when they next open. Their members, settings, and messages remain available. Repeated starts preserve the saved history and deletion state.

Proactive Agent reports, Routine results, and formal system outcomes now create unread updates for the conversation's recipients. Reading a message before Inbox finishes syncing keeps it read after synchronization.

Space Tasks awaiting a Human review or rework decision now remain in Inbox until the Task's own workflow ends the request. Reading messages or dismissing an Inbox item does not approve the result.

Contextual sidebar lists stay at their minimum readable width while you drag inward. The minimum width is more compact. Collapse requires another third of that width plus a small buffer, which prevents small drag movements from hiding the list. A shorter outward drag reopens it. Dragging and keyboard resizing no longer leave a squeezed icon-only state. Longer names stay on one line with an ellipsis, and edge hover still reveals a temporary preview.

When a Space denies a visit activity update, OpenMeld now reports a permission error instead of a service connection failure.

Delegated tasks and follow-up work keep the original request's time and timezone. Relative instructions such as "in ten minutes" use the same starting time after a task completes or its follow-up request retries. Local Agents also receive this time reference when their work starts later.