Plans can now be refined through conversation. When you open a plan thread, the AI is scoped to that specific task and its subtasks — it reads the current state before responding, so you can ask it to add, remove, or restructure tasks and it will update them directly. The conversation stays in place and each message operates on the live plan.
Images you upload to briefs and blueprints are now automatically described by AI in the background. The description is attached to the image and flows into every downstream AI consumer — when agents read your brief or generate a plan, they see the image description as part of the document content. The description appears as a tooltip when you hover over an image in the editor.
Thread spawn cards for plan generation and task creation have been updated with consistent status labels, type-specific icons, and CTA buttons. The brief title now shows as the shimmer text while a plan is being generated. The "View Plan" / "View Task" button is visible for both in-progress and completed threads so you can open the thread before it finishes.
The plan agent now requires verified search results before populating file paths, infrastructure references, and codebase patterns in tasks. It runs code search against your connected repositories before referencing any technical details, preventing fabricated paths and service names from appearing in generated tasks.