version: "1.0.0" name: shopify-customer-events description: >- Build a pipeline that turns the Shopify Partner API app-events feed into clean customer lifecycle events (subscribed, upgraded, downgraded, unsubscribed, uninstalled, trial_expired) for MRR and churn analytics. Use when the user wants to ingest Shopify partner/app events, derive subscription lifecycle events, compute MRR deltas (netChange), suppress plan-change "cancel" noise, derive trials, normalize uninstall reasons, or migrate off Mantle's customer-events pipeline into their own codebase.

Shopify Customer Events

What this skill does

Helps you build the pipeline that compiles Shopify's raw Partner API app-events feed into clean customer lifecycle events — the subscribed / upgraded / downgraded / unsubscribed / uninstalled / trial_expired stream every MRR and churn chart is built on.

The mechanism: keep the raw feed verbatim in one table, then derive clean events from it with a per-install state machine replayed in time order. The hard parts — and the reasons this isn't a simple type-to-type map — are the cancel trap (a lone cancel is usually a plan change, not churn, including the no-activation-webhook paid→free case), monthly normalization of netChange MRR deltas, trial derivation (nothing in the feed fires at trial end), and uninstall-reason normalization (preset map → LLM fallback, with store closures flagged separately).

The full, precise mechanics live in [`references/implementation-spec.md`](./references/implementation-spec.md). Read it before writing code — it has the exact state-machine branches, the suppression windows, the normalization math, and the edge cases. This file is the procedure; the spec is the reference.

When to use it

Trigger on intents like: "ingest Shopify partner/app events," "turn the app-events feed into lifecycle events," "build MRR/churn from Shopify subscription events," "my plan changes are showing up as churn," "compute netChange / MRR deltas," "detect trials from Shopify events," "categorize uninstall reasons," or "migrate off Mantle's customer-events pipeline."

How to drive the implementation

Work through these phases. Do not dump code blindly — discover the user's stack and existing event/subscription code first, then map the spec onto it.

Phase 1 — Orient (ask, then read)

1. Determine the stack: language, web framework, ORM/database, job/cron runner,

and how they already talk to the Shopify Partner API (do they have Partner API access and an appEvents feed query?).

1. Find existing code: any current partner/app-event ingestion, plan model,

subscription model, a customer-events/analytics table, cron/scheduled tasks, and any LLM/translation helpers (for uninstall reasons).

1. Ask which slices they need:

• raw ingestion + basic lifecycle events only?
• + the cancel trap (plan-change suppression)?
• + netChange/MRR deltas and an MRR rollup?
• + trial derivation? + uninstall-reason normalization?
• + a search/columnar fan-out, or is a SQL query over the clean table enough?

Phase 2 — Read the reference

Read `references/implementation-spec.md` end to end. The invariants in its §0 are non-negotiable; the §9 MUST/MUST NOT rules are the bugs you're being paid to avoid.

Phase 3 — Data model (spec §1)

Add/confirm: a verbatim raw_app_events table with the (install, type, occurredAt, chargeGid) dedup key; a clean app_events table with a unique deterministic `platformEventId`, netChange, previousSubscriptionId, and a soft-delete column; a uninstall_app_events child table; and enough of the Shopify charge/plan to normalize money. Write a migration. Settle the clean lifecycle-type enum (spec §1.5).

Phase 4 — Poll the feed (spec §2)

Implement cursor-paginated polling of the appEvents feed with an overlapping since-buffer, persisting raw events first and deduping on the unique key. Catch per-event errors. Wire it to the user's scheduler.

Phase 5 — The state machine (spec §3)

Implement the per-install, time-ordered fold carrying currentState, currentAmount, and mostRecentSubscription. Implement monthly normalization (spec §5.1) and the activation branch (subscribe / resubscribe / upgrade / downgrade). Log-and-skip unknown types. Set previousSubscriptionId (spec §3.5).

Phase 6 — The cancel trap (spec §4) — do this carefully

Implement the three suppression checks (activation-after 60s, activation-before-same-charge 5s, chained paid→free upgraded/downgraded ±60s), the soft-delete-on-suppress, and the state re-baseline for the paid→free case. Add the offline reconciler for stray unsubscribed events.

Phase 7 — netChange / MRR (spec §5)

Attach the signed monthly delta to each event, reflect discounts (effective price, not list, unless app-credit discounts), and prove MRR = sum of netChange over non-ignored events.

Phase 8 — Trials (spec §6, optional)

Derive trial status from local dates; add the nightly idempotent trial_expired sweep; handle trial-converts-via-plan-change so it isn't counted as churn.

Phase 9 — Uninstall reasons (spec §7, optional)

Build the enrichment job: translate → preset map → LLM fallback → cache, store reasonCodes[] + primary, and flag the store-closure subset separately.

Phase 10 — Verify

Walk the spec §10 acceptance scenarios as a checklist. At minimum prove: first subscribe + annual normalization; a mid-cycle plan change emits one event and zero churn; paid→free suppresses the cancel and re-baselines; real churn emits unsubscribed; full replay is idempotent.

Guardrails

• Preserve the five invariants (spec §0): two layers (raw + derived), a

per-install time-ordered state machine, monthly-normalized money, the guilty-until-proven cancel trap, and idempotency on a deterministic platformEventId. Breaking any of them produces phantom churn or MRR drift.

• The cancel trap is the highest-value rule. A lone cancel is presumed a plan

change; only suppress-or-emit per spec §4 — never blindly record churn.

• Handle paid→free explicitly — it emits no activation webhook; detect the

chained plan-change event and re-baseline the state machine.

• Normalize money to monthly before comparing (annual ÷ 12; flex from plan

nominal) — or annual plans break every chart.

• Don't delete non-counting events — mark them ignored at the read boundary.
• Never throw on an unknown raw type — log and skip.
• Treat search/columnar stores as derived projections — their failures never

block the authoritative clean-event write.

• Prefer editing the user's existing models/clients over inventing parallel ones.

Match their conventions.