sf-datacloud-prepare（顺丰数据云准备）

sf-datacloud-prepare: Data Cloud Prepare Phase

Use this skill when the user needs ingestion and lake preparation work: data streams, Data Lake Objects (DLOs), transforms, Document AI, unstructured ingestion, or the handoff from connector setup into a live stream.

When This Skill Owns the Task

Use sf-datacloud-prepare when the work involves:

• sf data360 data-stream *
• sf data360 dlo *
• sf data360 transform *
• sf data360 docai *
• choosing how data should enter Data Cloud
• rerunning or rescanning ingestion after a source update
• preparing Ingestion API-backed streams after connector setup is complete

Delegate elsewhere when the user is:

• still creating/testing source connections → sf-datacloud-connect
• mapping to DMOs or designing IR/data graphs → sf-datacloud-harmonize
• querying ingested data → sf-datacloud-retrieve

Required Context to Gather First

Ask for or infer:

• target org alias
• source connection name
• source object / dataset / document source
• desired stream type
• DLO naming expectations
• whether the user is creating, updating, running, or deleting a stream
• whether the source is CRM, a database connector, an unstructured file source, or an Ingestion API feed

Core Operating Rules

• Verify the external plugin runtime before running Data Cloud commands.
• Run the shared readiness classifier before mutating ingestion assets: node ~/.claude/skills/sf-datacloud/scripts/diagnose-org.mjs -o <org> --phase prepare --json.
• Prefer inspecting existing streams and DLOs before creating new ingestion assets.
• Suppress linked-plugin warning noise with 2>/dev/null for normal usage.
• Treat DLO naming and field naming as Data Cloud-specific, not CRM-native.
• Confirm whether each dataset should be treated as Profile, Engagement, or Other before creating the stream.
• Distinguish stream-level refresh from connection-level reruns when working with unstructured sources.
• Use UI setup intentionally when initial stream or unstructured asset creation is platform-gated.
• Hand off to Harmonize only after ingestion assets are clearly healthy.

Recommended Workflow

1. Classify readiness for prepare work

node ~/.claude/skills/sf-datacloud/scripts/diagnose-org.mjs -o <org> --phase prepare --json

2. Inspect existing ingestion assets

sf data360 data-stream list -o <org> 2>/dev/null
sf data360 dlo list -o <org> 2>/dev/null

3. Confirm the stream category before creation

Use these rules when suggesting categories:

When the source is ambiguous, ask the user explicitly whether the dataset should be treated as Profile, Engagement, or Other.

4. Create or inspect streams intentionally

sf data360 data-stream get -o <org> --name <stream> 2>/dev/null
sf data360 data-stream create-from-object -o <org> --object Contact --connection SalesforceDotCom_Home 2>/dev/null
sf data360 data-stream create -o <org> -f stream.json 2>/dev/null
sf data360 data-stream run -o <org> --name <stream> 2>/dev/null

5. Check DLO shape

sf data360 dlo get -o <org> --name Contact_Home__dll 2>/dev/null

6. Choose the right refresh mechanism

Use the smaller refresh scope that matches the user goal:

sf data360 data-stream run -o <org> --name <stream> 2>/dev/null
sf data360 connection run-existing -o <org> --name <connection-id> 2>/dev/null

• data-stream run is the closest match to a stream-level refresh or re-scan.
• connection run-existing runs at the connection level and can be useful for some connector workflows, but it is not a reliable replacement for stream refresh on unstructured sources.
• For unstructured document connectors, prefer data-stream run when the goal is to re-scan newly added or changed files.

7. Handle unstructured sources deliberately

For SharePoint-style document ingestion, a minimal unstructured DLO payload can look like:

{
  "name": "my_udlo",
  "label": "My UDLO",
  "category": "Directory_Table",
  "dataSource": {
    "sourceType": "SF_DRIVE",
    "directoryAndFilesDetails": [
      {
        "dirName": "SPUnstructuredDocument/<CONNECTION_ID>/<SITE_ID>",
        "fileName": "*"
      }
    ],
    "sourceConfig": {
      "reservedPrefix": "$dcf_content$"
    }
  }
}

Use the UI for the first-time unstructured setup when the user needs the richer end-to-end pipeline. The UI path can seed additional document metadata fields and downstream assets that a bare CLI DLO create flow may not provision automatically.

8. Use the local Ingestion API example for send-data workflows

For external systems pushing records into Data Cloud:

1. create the connector in sf-datacloud-connect
2. upload the schema with sf data360 connection schema-upsert
3. create the stream in the UI when required
4. send records with the local example in examples/ingestion-api/

cd examples/ingestion-api
cp .env.example .env
python3 send-data.py

Key details:

• auth is a staged flow: JWT → Salesforce token → Data Cloud token
• the ingestion endpoint uses the tenant URL, not the Salesforce instance URL
• 202 means the payload was accepted for processing, not that records are queryable immediately
• validation failures often surface in the Problem Records DLO family

9. Only then move into harmonization

Once the stream and DLO are healthy, hand off to sf-datacloud-harmonize.

High-Signal Gotchas

• CRM-backed stream behavior is not the same as fully custom connector-framework ingestion.
• sf data360 data-stream run and sf data360 connection run-existing are not interchangeable; prefer stream-level refresh for unstructured rescans.
• SFDC streams sync on a platform-managed schedule; data-stream run is not the general control path for CRM connector refresh.
• Some external database connectors can be created via API while stream creation still requires UI flow or org-specific browser automation. Do not promise a pure CLI stream-creation path for every connector type.
• Initial SharePoint-style unstructured setup can be richer in the UI than in a minimal CLI DLO create flow.
• Stream deletion can also delete the associated DLO unless the delete mode says otherwise.
• DLO field naming differs from CRM field naming, including __c → _c transformations.
• Query DLO record counts with Data Cloud SQL instead of assuming list output is sufficient.
• CdpDataStreams means the stream module is gated for the current org/user; guide the user to provisioning/permissions review instead of retrying blindly.

Output Format

Prepare task: <stream / dlo / transform / docai>
Source: <connection + object>
Target org: <alias>
Artifacts: <stream names / dlo names / json definitions>
Verification: <passed / partial / blocked>
Next step: <harmonize or retrieve>

References

• README.md
• examples/ingestion-api/README.md
• ../sf-datacloud/assets/definitions/data-stream.template.json
• ../sf-datacloud/references/plugin-setup.md
• ../sf-datacloud/references/feature-readiness.md