CS Engine setup

Build the stack profile: the mapping from the engine's tool categories to the connectors this user actually has. Every other skill in this plugin reads the profile first.

Step 1: Inventory what is connected

List the tools available in this session. Look at the MCP/connector tool names and descriptions (mcp__<server>__<tool> names; a ListConnectors tool when present). Do not ask the user what they have connected before looking: the point of this skill is to see for them.

Classify every server into one category from ${CLAUDE_PLUGIN_ROOT}/CONNECTORS.md:

A server can fill more than one category (Zoom = conversation intelligence + calendar; HubSpot = CRM + survey + support when those hubs are in use). If two servers compete for one category, prefer the one with richer read tools and note the alternate.

Step 2: Confirm the mapping

Show the user a short table: category → connector found → what it unlocks → what is missing. Use AskUserQuestion only for genuine ambiguity (two CRMs; a warehouse and product analytics, in which case ask which is the source of truth for usage). Otherwise present the mapping and ask for a one-word confirmation.

For each missing category, say in one line what the engine loses without it, taken from the "Minimum viable setup" section of CONNECTORS.md. Do not block setup on missing optional categories, and name the path from CONNECTORS.md ("When your tool isn't here") when the user's tool has no connector. Block only if ~~crm is missing; then explain that the CRM is the account spine and suggest connecting one before continuing.

Step 2b: Check each mapped tool

For every connected tool that has a map in ${CLAUDE_PLUGIN_ROOT}/skills/customer-360/references/tools/, run the read described under that map's "Setup check" and record the result in the stack profile's status column using the wording the map gives (for example connected (mapped; confirmed) or connected (mapped; CSAT not on this plan)). These maps were written from vendor documentation; the check is what confirms them in this workspace. For chat, email, calendar and documents tools, one list call proving the connector answers is enough. A failed check is recorded, not fatal: the block falls back to the generic rule in field-sources.md and its flags.

Step 3: Capture the operator context

Ask, in one AskUserQuestion call, only what cannot be inferred:

1. The user's name as it appears in the CRM as an owner (so "my book" resolves); skip if the CRM connector exposes the current user.
2. Segment names used in the CRM and how they map to enterprise / mid_market / smb / digital; skip if the CRM uses those words.
3. Internal email domains (so transcript speakers can be split internal vs customer).
4. Whether the user wants the engine to propose CRM writes (default) or never touch the CRM at all (read-only mode).

Step 4: Save the stack profile

Write the profile to cs-engine/stack-profile.md in the working folder (create the folder). Use exactly this shape so other skills can parse it:

# CS Engine stack profile
updated: <today>
operator: <name> · crm_owner_id: <id or unknown>
mode: propose-writes | read-only
internal_domains: [<domain>, ...]
segment_map: { "<crm value>": "enterprise", ... }

| category | connector | server | status |
|---|---|---|---|
| crm | HubSpot | mcp__HubSpot | connected |
| support desk | Intercom | mcp__Intercom | connected (mapped; confirmed) |
| conversation intelligence | none | none | missing (sentiment pillar reduced) |
...

If memory tools are available, also save a two-line summary under /areas/cs-claude-engine-stack.md (category → connector only; no credentials, no customer data) so a new conversation can skip re-discovery.

Step 5: Prove the install

Run the bundled sample through the deterministic scorer and show the result:

python3 ${CLAUDE_PLUGIN_ROOT}/src/models/health_score.py ${CLAUDE_PLUGIN_ROOT}/examples/sample_customer_360.json

If Python is unavailable in this session, read ${CLAUDE_PLUGIN_ROOT}/examples/sample_customer_360.json and explain the health score from its health.components instead. Either way, present it the way every reply reads (${CLAUDE_PLUGIN_ROOT}/skills/customer-360/references/writing-standard.md): a headline, the account in two sentences, and what is driving the score in plain words. Then end with:

• the three skills to try first: customer-360 on a real account, book-review for the whole book, and whichever playbook matches their biggest worry (red account → red-account-turnaround; renewal coming → renewal-defense; expansion → expansion-whitespace; EBR → ebr-generator);
• one sentence on the approval model: the engine drafts and proposes; nothing reaches a customer or changes a CRM field without the user approving it.

Rules

• Never ask the user for API keys or tokens. The connectors are already authorized in Claude.
• Never store customer data in the stack profile: it describes tools, not accounts.
• If a connector is present but its tools fail (auth expired, scope missing), record it as connected (failing: <reason>) and continue; do not abort setup.
• Keep the whole interaction under five exchanges. Setup is a door, not a destination.