sf-datacloud（顺丰数据云）

sf-datacloud: Salesforce Data Cloud Orchestrator

Use this skill when the user needs product-level Data Cloud workflow guidance rather than a single isolated command family: pipeline setup, cross-phase troubleshooting, data spaces, data kits, or deciding whether a task belongs in Connect, Prepare, Harmonize, Segment, Act, or Retrieve.

This skill intentionally follows sf-skills house style while using the external sf data360 command surface as the runtime. The plugin is not vendored into this repo.

When This Skill Owns the Task

Use sf-datacloud when the work involves:

• multi-phase Data Cloud setup or remediation
• data spaces (sf data360 data-space *)
• data kits (sf data360 data-kit *)
• health checks (sf data360 doctor)
• CRM-to-unified-profile pipeline design
• deciding how to move from ingestion → harmonization → segmentation → activation
• cross-phase troubleshooting where the root cause is not yet clear

Delegate to a phase-specific skill when the user is focused on one area:

Delegate outside the family when the user is:

• extracting Session Tracing / STDM telemetry → sf-ai-agentforce-observability
• writing CRM SOQL only → sf-soql
• loading CRM source data → sf-data
• creating missing CRM schema → sf-metadata
• implementing downstream Apex or Flow logic → sf-apex, sf-flow

Required Context to Gather First

Ask for or infer:

• target org alias
• whether the plugin is already installed and linked
• whether the user wants design guidance, read-only inspection, or live mutation
• data sources involved: CRM objects, external databases, file ingestion, knowledge, etc.
• desired outcome: unified profiles, segments, activations, vector search, analytics, or troubleshooting
• whether the user is working in the default data space or a custom one
• whether the org has already been classified with scripts/diagnose-org.mjs
• which command family is failing today, if any

If plugin availability or org readiness is uncertain, start with:

• references/plugin-setup.md
• references/feature-readiness.md
• scripts/verify-plugin.sh
• scripts/diagnose-org.mjs
• scripts/bootstrap-plugin.sh

Core Operating Rules

• Use the external sf data360 plugin runtime; do not reimplement or vendor the command layer.
• Prefer the smallest phase-specific skill once the task is localized.
• Run readiness classification before mutation-heavy work. Prefer scripts/diagnose-org.mjs over guessing from one failing command.
• For sf data360 commands, suppress linked-plugin warning noise with 2>/dev/null unless the stderr output is needed for debugging.
• Distinguish Data Cloud SQL from CRM SOQL.
• Do not treat sf data360 doctor as a full-product readiness check; the current upstream command only checks the search-index surface.
• Do not treat query describe as a universal tenant probe; only use it with a known DMO/DLO table after broader readiness is confirmed.
• Preserve Data Cloud-specific API-version workarounds when they matter.
• Prefer generic, reusable JSON definition files over org-specific workshop payloads.

Recommended Workflow

1. Verify the runtime and auth

Confirm:

• sf is installed
• the community Data Cloud plugin is linked
• the target org is authenticated

Recommended checks:

sf data360 man
sf org display -o <alias>
bash ~/.claude/skills/sf-datacloud/scripts/verify-plugin.sh <alias>

Treat sf data360 doctor as a broad health signal, not the sole gate. On partially provisioned orgs it can fail even when read-only command families like connectors, DMOs, or segments still work.

2. Classify readiness before changing anything

Run the shared classifier first:

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

Only use a query-plane probe after you know the table name is real:

node ~/.claude/skills/sf-datacloud/scripts/diagnose-org.mjs -o <org> --phase retrieve --describe-table MyDMO__dlm --json

Use the classifier to distinguish:

• empty-but-enabled modules
• feature-gated modules
• query-plane issues
• runtime/auth failures

3. Discover existing state with read-only commands

Use targeted inspection after classification:

sf data360 doctor -o <org> 2>/dev/null
sf data360 data-space list -o <org> 2>/dev/null
sf data360 data-stream list -o <org> 2>/dev/null
sf data360 dmo list -o <org> 2>/dev/null
sf data360 identity-resolution list -o <org> 2>/dev/null
sf data360 segment list -o <org> 2>/dev/null
sf data360 activation platforms -o <org> 2>/dev/null

4. Localize the phase

Route the task:

• source/connector issue → Connect
• ingestion/DLO/stream issue → Prepare
• mapping/IR/unified profile issue → Harmonize
• audience or insight issue → Segment
• downstream push issue → Act
• SQL/search/index issue → Retrieve

5. Choose deterministic artifacts when possible

Prefer JSON definition files and repeatable scripts over one-off manual steps. Generic templates live in:

• assets/definitions/data-stream.template.json
• assets/definitions/dmo.template.json
• assets/definitions/mapping.template.json
• assets/definitions/relationship.template.json
• assets/definitions/identity-resolution.template.json
• assets/definitions/data-graph.template.json
• assets/definitions/calculated-insight.template.json
• assets/definitions/segment.template.json
• assets/definitions/activation-target.template.json
• assets/definitions/activation.template.json
• assets/definitions/data-action-target.template.json
• assets/definitions/data-action.template.json
• assets/definitions/search-index.template.json

6. Verify after each phase

Typical verification:

• stream/DLO exists
• DMO/mapping exists
• identity resolution run completed
• unified records or segment counts look correct
• activation/search index status is healthy

High-Signal Gotchas

• connection list requires --connector-type.
• dmo list --all is useful when you need the full catalog, but first-page dmo list is often enough for readiness checks and much faster.
• Segment creation may need --api-version 64.0.
• segment members returns opaque IDs; use SQL joins for human-readable details.
• sf data360 doctor can fail on partially provisioned orgs even when some read-only commands still work; fall back to targeted smoke checks.
• query describe errors such as Couldn't find CDP tenant ID or DataModelEntity... not found are query-plane clues, not automatic proof that the whole product is disabled.
• Many long-running jobs are asynchronous in practice even when the command returns quickly.
• Some Data Cloud operations still require UI setup outside the CLI runtime.

Output Format

When finishing, report in this order:

1. Task classification
2. Runtime status
3. Readiness classification
4. Phase(s) involved
5. Commands or artifacts used
6. Verification result
7. Next recommended step

Suggested shape:

Data Cloud task: <setup / inspect / troubleshoot / migrate>
Runtime: <plugin ready / missing / partially verified>
Readiness: <ready / ready_empty / partial / feature_gated / blocked>
Phases: <connect / prepare / harmonize / segment / act / retrieve>
Artifacts: <json files, commands, scripts>
Verification: <passed / partial / blocked>
Next step: <next phase, setup guidance, or cross-skill handoff>

Cross-Skill Integration

Reference Map

Start here

• README.md
• references/plugin-setup.md
• references/feature-readiness.md
• UPSTREAM.md

Phase skills

• sf-datacloud-connect
• sf-datacloud-prepare
• sf-datacloud-harmonize
• sf-datacloud-segment
• sf-datacloud-act
• sf-datacloud-retrieve

Deterministic helpers

• scripts/bootstrap-plugin.sh
• scripts/verify-plugin.sh
• scripts/diagnose-org.mjs
• assets/definitions/