PBI Bootstrap

Session setup and recovery for SemanticOps MCP. Use the bundled references as the working documentation.

Quick Start

1. Check the connection: manage_model_connection { "operation": "get_current" }.
2. If no model is connected — Desktop: { "operation": "list" }, present the choices, then { "operation": "select", "model_id": "<model_id>" } with the user's pick. Service: { "operation": "authenticate", "source": "service" } first when needed, then { "operation": "list_workspaces" } for the workspace choice, then { "operation": "list", "source": "service", "workspace": "<workspace>" } to list its datasets, and finally select with the returned model_id (workspaces themselves are not selectable).
3. Load saved preferences with manage_preferences { "action": "list", "resource": "rendered" } and apply them within their stated trust boundary.
4. Read tool-invocation-conventions before composing any non-trivial write or bulk request.
5. If a tool returns empty or stale results, switch to troubleshooting-guide before retrying.

Route to the right tool family

• list_model (list, search, analyze, info, report) for discovery, search, previews, and metadata inspection.
• run_query (execute, analyze, vertipaq, test_access) for DAX execution, performance analysis, storage inspection, and RLS checks.
• Write tools (manage_schema, manage_semantic, manage_security) only after the target object and operation are explicit.
• Once connected and healthy, hand off to the task skill: query authoring → mcp-engine-query; slow queries → mcp-engine-dax-performance; wrong values → mcp-engine-dax-debugging; tables, columns, relationships, partitions → mcp-engine-schema-authoring; measures and calc groups → mcp-engine-semantic-authoring; multi-object refactors → mcp-engine-refactoring; tests, checkpoints, rollback → mcp-engine-testing-changes; RLS, policy, masking, audit → mcp-engine-security-governance. If a named skill is not installed, continue with the tool's inputSchema or ask the user to add it.

Guard bulk and write requests

• Keep transaction, dry_run, include_items, and include_details as top-level request controls.
• Put per-item identifiers and spec values inside each item; items do not inherit from the request.
• Confirm destructive intent before deletes, renames, broad refreshes, or model-wide rewrites.

Recover from common failures

• Wrong or stale connection state: re-check with { "operation": "get_current" }, then { "operation": "reload" } after external model changes.
• Object seems missing: broaden discovery with list_model { "operation": "search", "spec": { "query": "<object name>", "mode": "name" } } (spec.query is required) before concluding it does not exist.
• Argument validation errors: fix key names and payload shape per tool-invocation-conventions before changing business logic.
• Full recovery flows: troubleshooting-guide.