[Digvijay (Digvijay)]

Summary

Fixes for the Agentic Inventory Planning (Fabric) hack based on facilitator/participant feedback from the Hack-in-Germany and Norway runs, plus issues caught during a full end-to-end verification on a live tenant.

Original feedback addressed

• New Foundry portal navigation: deployed model → Playground, Traces → Response view (removed old "Models + endpoints" / "Agents → Playground" / left-nav Tracing).
• Shared-tenant: how to identify your Foundry project (match FoundryProjectEndpoint → Parent resource).
• Unique Fabric workspace name (tenant-unique) guidance.
• Published URL → MCP server URL (Settings → Model Context Protocol), new /v1/mcp/... format.
• Attach Lakehouse via Add data items → From OneLake catalog.
• Data Agent category snake_case mapping — fixes the NL2SQL "Outdoor Power Tools" vs outdoor_power_tools mismatch (confirmed working).
• Working SKU-based sample prompt; parquet load / cancel-in-Monitor note.
• Facilitator notes: audience/expectations, ontology capacity, 200K TPM + 5 labs/sub, new -FabricSkuName param.

Issues found during end-to-end verification (fixed + documented)

• fx function-stub trap — the New Foundry portal auto-generates empty function stubs from instructions; use the Fabric Data Agent connector instead.
• Stage configuration not found — the Data Agent works in Fabric's test pane (draft) but must be published for Foundry to consume it.
• Zero-reorder bug — optimisation agent queried demand "across warehouses" (no sales rows); fixed to query store-level weekly demand and convert weekly→daily (÷7). Reorder now returns non-zero (verified: P004 Portland ≈ 17).

Verified end-to-end

All four challenges run on a live F2 capacity: data backend (7 tables, snake_case categories, CRITICAL data confirmed via SQL), Data Agent answers, and each Foundry agent calls Fabric (fabric_dataagent_preview_call in traces).

Screenshots

Refreshed the Priority-1 set (MCP URL trio + both agent traces). A few optional refreshes remain (noted in images/README.md).

Deferred / follow-up

• Evaluation set (Germany ask) — intentionally deferred to a follow-up PR.