aidp-engineer-overview | oracle-ai-data-platform-workbench-…

Stats

Actions

Tags

aidp-engineer-overview | oracle-ai-data-platform-workbench-…

`aidp-engineer-overview` — router & shared environment rules

Pick the right skill for an AIDP task, and own the cross-cutting setup/auth rules every other skill relies on.

When to use

The user mentions AIDP but the specific operation is unclear, or names several at once.
The user asks "how do I … in AIDP" / "what can this do".
Any aidp-* skill fails with an auth, workspace, or cluster error.

First run

If .aidp/catalog.md does not exist, route to aidp-engineer-bootstrap (one-time setup), then aidp-catalog-init (grounding), before answering data questions. The plugin is self-contained — it does not require a bootstrap of any MCP server to run.

Routing table

User intent	Skill
First-run setup / resolve env (region, DataLake OCID, workspace, cluster)	`aidp-engineer-bootstrap`
Create a DataLake instance or workspace (incl. private network)	`aidp-workspace-admin`
Build the catalog map / "discover my lakehouse"	`aidp-catalog-init`
Answer a business question / run read Spark SQL	`aidp-analyzing-data`
Write SQL — INSERT/UPDATE/DELETE/MERGE, CREATE/ALTER/DROP, OPTIMIZE/VACUUM/time-travel	`aidp-sql-ddl`
Create/alter/drop table·view·schema·catalog (control-plane); register an external catalog	`aidp-table-management`
Browse catalogs/schemas/tables/volumes, resolve a name→key	`aidp-catalog-explore`
Profile a table (nulls/distinct/min-max)	`aidp-profiling-tables`
Data-quality rule checks	`aidp-data-quality`
Recent activity / async operations	`aidp-observability`
Load a file into a table	`aidp-ingest-file-to-table`
Workspace files / notebooks CRUD	`aidp-workspace-files`
Volumes (PAR up/down, dirs)	`aidp-volumes`
Build/run/monitor a Job (pipeline, schedule)	`aidp-pipelines`
Author/run a notebook + kernel session	`aidp-notebooks`
Cluster lifecycle / libraries	`aidp-cluster-ops`
Debug a slow/failed Spark job	`aidp-spark-debugging`
Tune/optimize a slow Spark job (skew, spill, shuffle, small files, joins, AQE, Delta, JDBC reads)	`aidp-spark-optimization`
Define metrics / logical model	`aidp-semantic-model`
Register/validate reusable verified queries	`aidp-verified-queries`
Federate / join across multiple sources	`aidp-federate`
LLM inside SQL (`ai_generate`)	`aidp-ai-sql`
Secrets / credential store	`aidp-credentials`
Delta sharing (shares/recipients)	`aidp-data-sharing`
Git in the workspace	`aidp-git`
Author/deploy/run an agent flow (13 node types, guardrails)	`aidp-agent-flows`
Code an agent in Python (LangGraph / aidputils, high-code)	`aidp-agent-highcode`
Create a reusable standalone tool (SQL/Prompt/RAG/HTTP/Custom/MCP)	`aidp-tools`
Build a Knowledge Base / RAG / vector index	`aidp-knowledge-bases`
Deploy/purge a resource bundle	`aidp-bundle`
Roles / permissions / access (incl. per-resource grants, masking)	`aidp-roles-access`
MLflow experiments / models	`aidp-mlops`
Installed models catalog	`aidp-models-catalog`
Audit logs (enable/retention, search)	`aidp-audit`
User settings / preferences	`aidp-user-settings`
Databricks → AIDP migration	`aidp-migration`

Out of scope → point elsewhere

Connecting to an external source (Oracle ADB/ExaCS, Fusion, EPM, Essbase, Snowflake, S3, Kafka, …) — including a request like "a notebook that connects to Fusion" → the oracle-ai-data-platform-workbench-spark-connectors plugin: use its aidp-<source> skill for the connection recipe — do NOT hand-roll the connection. Check it's installed (claude plugin list); if not, tell the user to install it. Its oracle_ai_data_platform_connectors helper package is installed once via that plugin's aidp-connectors-bootstrap skill (pushes it to /Workspace/Shared via the AIDP MCP; if the MCP can't reach the instance, upload manually). A single external source needs only that connector skill (then author/run via aidp-notebooks); aidp-federate composes several for cross-source joins.
OCI networking (VCN/NAT/ACL) and OAC dashboards/registration — not handled here.

Shared environment & auth rules (every skill inherits these)

Engine precedence — self-contained, no MCP/private-repo dependency. For every CONTROL-PLANE op (catalogs, schemas, tables, clusters, jobs, workspaces, roles, volumes, files, credentials, sharing, bundle, mlops, audit, user-settings, …), use this order — see references/aidp-cli-map.md:
1. Preferred: the official aidp CLI — aidp <group> <command> --auth api_key --profile DEFAULT --instance-id <DATALAKE_OCID> (public, Oracle-supported: github.com/oracle-samples/aidataplatform-sdk; aidp command-groups / aidp search). Use it when installed.
2. Fallback: oci raw-request against the same REST API (https://aidp.<region>.oci.oraclecloud.com/20240831/dataLakes/<DATALAKE_OCID>/…, api_key --profile DEFAULT) — see references/oci-raw-request.md / references/no-mcp-rest-map.md. Same endpoint + auth as the CLI, so it is a drop-in when the CLI isn't installed or doesn't expose the op (e.g. full Git, agent-flow authoring). Do NOT invent endpoints — use the references.
- Interactive Spark-SQL / notebook cells run via the bundled helper: python "$PLUGIN_DIR/scripts/aidp_sql.py" --region <r> --datalake <ocid> --workspace <ws> --cluster <key> --code <python/spark code>. It mints a UPST from the api_key DEFAULT profile, auto-creates a scratch notebook, and returns JSON (status / outputs / spark_job_ids). No AIDP_SESSION required; --session-profile is optional. Use a SELECT 1 cell as the trivial smoke test.
- The aidp MCP is an OPTIONAL accelerator only. If an MCP happens to be configured you MAY use its tools, but it is NOT required and NOT assumed. Never frame the plugin as depending on it, and never route to bootstrap to install an MCP — there is nothing to install for the plugin to work.
Workspace-first. AIDP operations are workspace-scoped. Confirm/select the workspace before acting; pass the workspace key/name explicitly (--workspace to the helper, …/workspaces/<ws>/… in REST URLs) when the default isn't the target.
Cluster must be RUNNING for any data/SQL op. Check cluster status (GET /workspaces/<ws>/clusters/<key>); start it (POST …/actions/start with a {} body) if stopped.
Persist + confirm every mutation. Before any create/update/delete/run/deploy/grant, write the request body to .aidp/payloads/<verb>-<resource>.json, show it to the user, and confirm before running it (auditable + re-runnable — see references/payloads.md). Especially for deploy/purge/delete/grant/share.
Auth ladder: --profile DEFAULT (api_key) → on 401/403/"NotAuthenticated"/"Security Token": oci session refresh --profile AIDP_SESSION then retry with --auth security_token --profile AIDP_SESSION; if refresh fails, oci session authenticate --profile AIDP_SESSION --region <region>. The helper mints its own UPST from DEFAULT; pass --session-profile AIDP_SESSION only if a tenancy rejects IAD api keys (session-token only). Some tenancies do reject IAD API keys outright.
Never hardcode or print OCIDs/keys/tokens. Never trust a local .env for region/OCID/profile.

References