Public tool that builds a file inventory, analyzes parseable code, and returns the analysis artifact for grouping.

Prepare mode returns a compact packet for semantic grouping; it is not the final grouping. The LLM should treat folder names as hints, not truth, and group by responsibility, runtime role, data flow, and dependencies. Prefer glob patterns like folder/** instead of enumerating files, and use exact file paths only for entry points, exceptions, or cross-cutting files. Apply mode consumes a small LLM-authored GroupingPlan and deterministically assigns files.

Apply LLM group decisions after a Blueprint refresh. Use this tool only for assigning unassigned files to existing groups, creating a new group for unassigned files, or deleting groups that are already empty. Do not use this tool for updated files already assigned to a real group, deleted file cleanup, editing markdown memory, or rewriting blueprint-output.json manually. Validation rules: assignments[].fileId must refer to a file whose groupId is "unassigned"; assignments[].groupId must be an existing group id; newGroups[].id must not already exist; newGroups[].fileIds must all be unassigned file ids; deleteGroups[] may only contain groups with no fileIds. The tool writes .blueprint/blueprint-output.json and creates group markdown templates for new groups.

Deterministically refresh .blueprint/blueprint-output.json from the current filesystem snapshot. The tool compares .blueprint/refresh-scan.json with a fresh full scan, writes refreshed Blueprint JSON and scan state, and returns the maintenance prompt the assistant should follow. It does not send raw git diffs to the assistant. Use blueprint.group.update afterwards only for unassigned files or empty group decisions, then update affected group Markdown docs.

Compose the final frontend-ready Blueprint JSON from a grouping artifact. If the response contains assistantNextSteps with required=true and executionPolicy=must_execute_before_final_response, the assistant must execute those steps before giving the user a final answer. For hydrate-group-docs, spawn one sub-agent per target group doc when sub-agents are available; otherwise edit the target docs yourself. Do not ask the user unless the required step is impossible.

Return a compact deterministic context slice for a natural-language task from a Blueprint output artifact.