aidp-fusion-bundle

aidp-fusion-bundle — Fusion ERP/HCM/SCM → AIDP, batteries included

Productizes the official Oracle blog Bring Fusion Data into Oracle AI Data Platform Workbench Using BICC plus the ateam companion How to Extract Fusion Data using Oracle AI Data Platform. The current path is: configure → connect OAC MCP → bootstrap → seed AIDP gold → advise the OAC dataset → use oac-dataset-setup for the governed manual OAC connection/dataset checkpoint → author the workbook via OAC MCP.

When a run, bootstrap, validation, dashboard, or workbook flow reports an AIDPF-* code, start with aidpf-error-triage. It is read-only and routes the failure to the recovery skill that owns it.

When to use

• User wants Fusion data in AIDP and asks "where do I start"
• User has BICC privileges and wants curated bronze/silver/gold layers without writing the pipeline
• User is preparing a CFO/analytics demo and needs OAC dashboards on Fusion data
• User wants to use OAC MCP (Preview) to chat with Fusion data via Claude/Cline/Copilot

When NOT to use

• For a single one-off PVO read → use aidp-fusion-bicc (sibling plugin, smaller scope).
• For Fusion REST queries with <50k rows → aidp-fusion-rest.
• For EPM Cloud Planning data slices → aidp-epm-cloud.
• For Essbase MDX → aidp-essbase.

Positioning

This bundle is additive to and complementary with Oracle's managed Fusion data offerings. It productizes Option 1 of pdf1's three-option architecture (BICC into AIDP for "Custom AI and ML, raw data access, data engineering"). Never positioned as a replacement for FDI, OAC, OTBI, BIP, or Data Transforms.

What you get

Mirrors pdf1 §"What Can You Do Once the Data is in Oracle AI Data Platform":

1. Custom ML/AI training on operational ERP/HCM/SCM data (PySpark + Python in AIDP notebooks)
2. Cross-source enrichment — join Fusion data with non-Fusion sources via the aidataplatform connector family
3. Medallion architecture — bronze (raw audit) → silver (typed + dim-joined) → gold (business marts) in Delta
4. GenAI agent grounding — ai_generate("which suppliers had >$1M Q1 spend?") against gold marts via OCI Generative AI
5. BI & reporting via JDBC — OAC, Tableau, Power BI consume the gold layer
6. Delta Sharing (v3 roadmap) — share curated datasets with other teams or external partners

Quickstart

Recommended route: install the Claude Code plugin, open Claude Code from a clean customer bundle directory such as Workspace/demo-fusion-cfo/, then invoke aidp-fusion-autopilot with the dashboard goal. Autopilot installs/uses the bundled CLI if needed, scaffolds customer files, and conducts the whole chain (configure → OAC MCP setup → bootstrap → seed → advise → dataset → workbook → optional MCP chat), pausing only for real decisions. Keep the customer directory separate from the plugin source. The manual quickstart below is the step-by-step path autopilot automates.

1. Install the CLI on your laptop:

   pip install -e /path/to/oracle-ai-data-platform-fusion-bundle
   

2. Scaffold a bundle in your repo:

   aidp-fusion-bundle init
   

   Edits bundle.yaml and aidp.config.yaml to match your environment (Fusion pod URL, AIDP workspace, OAC URL, OCI Vault refs for credentials).

3. Connect operator OAC MCP early:

   env -u OAC_URL -u OAC_MCP_USER -u OAC_MCP_PASSWORD -u OAC_ADMIN_USER -u OAC_ADMIN_PASSWORD \
   aidp-fusion-bundle dashboard mcp-setup \
     --connector-js <path to oac-mcp-connect.js>
   

   Run from the customer project directory. The env -u ... wrapper lets the local .env provide the OAC values instead of any global shell profile. Restart/reconnect Claude Code after this. Autopilot and workbook-authoring need OAC MCP for search_catalog, describe_data, and save_catalog_content. If setup happens mid-journey, autopilot writes .aidp/autopilot/resume.md; resume with: "Resume the Fusion dashboard workflow from .aidp/autopilot/resume.md."

4. Probe prerequisites and pin tenant variation:

   aidp-fusion-bundle bootstrap --check-iam
   

   Prefer to drive this conversationally? Use aidp-fusion-bootstrap. It confirms BICC role, BICC External Storage profile (set in BICC console), AIDP catalog, IAM policies, Vault access, and routes unresolved variation to medallion-author.

5. Run the orchestrator:

   aidp-fusion-bundle run --mode seed     # first-time full extract
   aidp-fusion-bundle run --mode incremental  # daily delta
   

   Prefer to drive this conversationally? The aidp-fusion-seed skill turns "seed", "seed supplier_spend", "seed just bronze", or "resume the seed" into the correct guarded run --mode seed invocation — it parses the scope, auto-satisfies preconditions (validate / /aidp-fusion-bootstrap / cluster), and fail-closed-confirms before overwriting populated silver/gold marts. If the run reports an AIDPF-* code, use aidpf-error-triage before choosing a recovery path.

6. Build dashboards (MCP-native — the current path). Ask oac-dataset-advisor what OAC dataset your goal needs (grounded in the live AIDP gold layer), use oac-dataset-setup to guide the manual AIDP connection/dataset step and verify it through MCP, then have workbook-authoring generate the visualization(s) and write them via the OAC MCP save_catalog_content tool. If the gold layer can't serve the goal, mart-author authors a new mart (then use-pack + seed). Legacy alternative: the .bar snapshot dashboard install flow (snapshot register + restore via OAC REST) still ships — see docs/oac_rest_api_setup.md — but the MCP-native family above supersedes it for authoring.

7. End users chat with the data via OAC MCP. Set up the connector for Claude Code (non-interactive basic auth, the path that actually works in a terminal client):

   env -u OAC_URL -u OAC_MCP_USER -u OAC_MCP_PASSWORD -u OAC_ADMIN_USER -u OAC_ADMIN_PASSWORD \
   aidp-fusion-bundle dashboard mcp-setup \
     --connector-js <path to oac-mcp-connect.js>
   

   Then ask "what's our AR aging?" and watch MCP call search_catalog → describe_data → execute_logical_sql against fusion_catalog.gold.*. Scope the OAC user to least privilege — the v1.4 connector exposes catalog write/delete/ACL tools governed by that user's grants.

Key gotchas

• BICC role required — Fusion user must hold BIA_ADMINISTRATOR_DUTY or ORA_ASM_APPLICATION_IMPLEMENTATION_ADMIN_ABSTRACT. Without it, /biacm/api/v[12]/* endpoints 302-redirect to IDCS. Bootstrap probes for this.
• BICC External Storage profile — must be configured once in the BICC console (admin task: BICC Console → Configure External Storage → OCI Object Storage Connection tab → bucket name + namespace + region + OCI username + auth token → Test Connection → Save). The fusion.external.storage Spark option references this BICC profile name. There is no parallel AIDP-side registration. Bundle does not provision the BICC profile; bootstrap verifies it exists.
• First extract is slow — BICC builds a full snapshot on first call; subsequent runs use fusion.initial.extract-date for incremental.
• 499 row/page hard cap on Fusion REST (per MOS Doc ID 2429019.1) — bundle's REST fallback enforces this; anything >5k rows must use BICC.
• OAC MCP (v1.4) is NOT read-only — it exposes catalog write tools too. The bundle authors workbooks via save_catalog_content (live-verified 2026-06-15: created gold_balance_2viz on a real OAC). It still cannot create datasets (no create-dataset tool — dataset modeling is an OAC UI step), and the write/delete/ACL tools run with the connecting user's grants → use a least-privilege MCP user. (Supersedes the earlier "MCP is read-only" note.)
• POST /catalog/connections REST validator does not bless AIDP idljdbc — Oracle's validator falls through to generic Oracle DB schemas requiring serviceName/password/connectionString. The realistic flow is therefore: customer creates the connection via OAC UI once (using the 6-key JSON written by --print-only). Legacy dashboard install can re-use that connection via the precheck on subsequent .bar snapshot deployments.
• Snapshot BAR URI shape is file:///<folder>/<name>.bar — NOT oci://..., NOT bare object name, NOT the OCI Object Storage HTTPS URL.
• OAC catalog browse needs search=* — GET /catalog?type=connections (no search) returns a single-element TypeInfo header ([{"type":"connections"}]), NOT the actual list. Bundle's list_connections defaults search="*" so the precheck works.
• Use ExtractPVOs for bulk, NOT OTBI reporting PVOs — pdf1 Pro Tip; bundle's catalog refuses OTBI PVOs with a clear warning.

References

• Plan: C:\Users\anuma\.claude\plans\oracle-ai-data-platform-fusion-bundle.md
• Sibling plugin (single-PVO connector): oracle-ai-data-platform-workbench-spark-connectors
• Official Oracle BICC blog: https://blogs.oracle.com/ai-data-platform/bring-fusion-data-into-oracle-ai-data-platform-workbench-using-bicc
• Ateam blog (saas-batch path): https://www.ateam-oracle.com/how-to-extract-fusion-data-using-oracle-ai-data-platform
• Official sample notebook: oracle-aidp-samples/data-engineering/ingestion/Read_Only_Ingestion_Connectors.ipynb
• OAC MCP Preview docs: https://docs.oracle.com/en/cloud/paas/analytics-cloud/acsdv/access-oracle-analytics-cloud-mcp-server-preview.html