Operating procedure for the ApexCloud Retention Operations API (http://127.0.0.1:8074). Use this skill WHENEVER a task asks you to build, score, or summarize anything in the ApexCloud retention / customer-success / revenue-ops domain: renewal risk queues, retention action boards, QBR metric packets, receivables + pipeline operations reviews, churn-model validation, outreach rankings, or board/leadership summaries. Trigger it even when the request only mentions pieces of this world (current ARR, "clean" ticket counts, latest NPS, overdue/A-R aging buckets, risk_score / risk_level, primary_action, reason_codes, expansion pipeline, win rate, policy_codes like RS-6 / REV-4 / SUP-8 / ACT-5 / RCP-7 / CM-5 / MOD-7, or account ids like acct_*). These tasks share fixed business conventions that are easy to get wrong from intuition; this skill encodes the exact, verified rules.
Operating procedure for the ApexCloud Retention Operations API (http://127.0.0.1:8074). Use this skill WHENEVER a task asks you to build, score, or summarize anything in the ApexCloud retention / customer-success / revenue-ops domain: renewal risk queues, retention action boards, QBR metric packets, receivables + pipeline operations reviews, churn-model validation, outreach rankings, or board/leadership summaries. Trigger it even when the request only mentions pieces of this world (current ARR, "clean" ticket counts, latest NPS, overdue/A-R aging buckets, risk_score / risk_level, primary_action, reason_codes, expansion pipeline, win rate, policy_codes like RS-6 / REV-4 / SUP-8 / ACT-5 / RCP-7 / CM-5 / MOD-7, or account ids like acct_*). These tasks share fixed business conventions that are easy to get wrong from intuition; this skill encodes the exact, verified rules.
ApexCloud Retention Operations SOP
This skill captures the business conventions for the ApexCloud Retention Operations environment.
The data is realistic but full of traps (decoy CRM lookalikes, spam/cancelled tickets, retracted NPS,
multiple revenue sources, several aging buckets). The grader checks exact numbers, controlled enum
labels, sort order, and fixed policy_codes. Reproduce the rules, never guess.
All data comes from the read-only HTTP API. Do not read local files. Health/base:
GET http://127.0.0.1:8074/api/health # row counts + seed, good sanity check
How to approach any task
Identify the task family (decides output shape, enums, and which policy_codes to emit):
Renewal Risk Queue / Retention Action Board -> the risk model (RS/REV/SUP/ACT, + board extras)
Churn Model Validation & Outreach Ranking -> CSV exports + logistic model (MOD/PRB/DEP/OUT)
Pull the canonical fields using the verified definitions below (current ARR, clean tickets,
latest NPS, overdue balance). These four are reused everywhere and are the most common source of error.
Apply the family-specific procedure (see references/playbook.md for full step-by-step recipes,
formulas, and the per-family policy_codes).
Match output exactly: only the keys in the answer template, controlled enums verbatim, currency
2 decimals, percentages 1 decimal, counts and risk scores as integers. Respect the requested sort.
references/playbook.md is the detailed companion — read it for the family you are working on. It also
contains the worked rule-derivations and edge cases. The four canonical definitions below are the heart
of the system, so they live here.
The four canonical field definitions (memorize these)
1. Current ARR -> latest POSTED billing snapshot (NOT the account record, NOT CRM)
GET /api/billing/snapshots?account_id=<id> returns quarterly snapshots. current_arr =
billing_arr of the snapshot with the greatest as_of date that is posted == true AND
as_of <= assessment/as-of date.
Do NOT use account.billing_arr_current (a rounded headline number) and never use account.crm_arr.
Example: assessment 2026-06-30 -> use the 2026-06-30 (Q2) snapshot, e.g. 1416439.47, not the
account's 1425000.0 and not the later Q3/Q4 snapshots.
The QBR task is the one exception for monthly revenue: it uses recognized_revenue from the
metrics endpoint, but still labels the source enum crm_closed_won (see playbook).
2. Clean support ticket count -> count tickets minus hygiene exclusions
GET /api/accounts/<id>/tickets?start=YYYY-MM-DD&end=YYYY-MM-DD. clean_ticket_count = tickets where
NONE of these hold: is_spam == true, is_duplicate == true, status == "cancelled".
Do NOT use metrics.support_ticket_count — it is a different (unfiltered) number.
Per-month QBR ticket counts also use this cleaned set, bucketed by created_date[:7].
3. Latest NPS -> latest valid monthly reading, with detractor/recovery logic for nps_drop
The latest NPS value to report is the last valid monthly reading. Use metrics[].nps_score in month
order, treating null AND the sentinel -1 AND any retracted response as missing; take the last
non-missing month. (The /nps endpoint corroborates this; exclude retracted == true and score < 0.)
nps_drop reason code fires when: latest_valid_monthly_nps < 50AND
(every valid monthly reading was < 50OR the latest month is not a recovery, i.e.
latest <= previous valid reading). A score that dipped below 50 but rose at the end (e.g. 51->44->46)
does NOT count as a drop; a series stuck in the detractor band (e.g. 17->39, or 28->29) does.
If no valid reading exists, report null and do not fire nps_drop.
4. Overdue balance -> ONLY the older aging buckets
GET /api/finance/ar-aging?as_of=YYYY-MM-DD. Each row has current, 1_30, 31_60, 61_90, 90_plus.
overdue_balance = 61_90 + 90_plus (the "older aging buckets"). Do NOT include current,
1_30, or 31_60. A client is "overdue" only when 61_90 + 90_plus > 0.
A/R rows have NO account_id; they key on customer_name. Link to CRM by exact legal_name
match to an account — never fuzzy-match aliases (lookalikes like "Globex North Subsidiary LLC" or
"North Star Finance Services" are deliberate decoys and stay unlinked).
Retention risk model (one model, two output families)
Used by the Renewal Risk Queue and the Retention Action Board. Compute boolean risk factors, sum
weighted points (cap 100), bucket into a level, map to an action, and emit the reason codes that fired.
This reproduces every train score exactly; full point table and edge cases are in the playbook.
Risk factors (and the reason code each one emits):
renewal_window: renewal_date is in the future within 90 days, i.e. 0 <= (renewal_date - as_of) <= 90.
sla_degradation: ANY clean ticket has first_response_sla_met == false OR resolution_sla_met == false.
usage_decline: latest-month product_usage < first-month AND latest-month product_usage < ~62
(a low-usage floor; steep declines that stay above the floor do NOT count).
expansion_offset: account has open expansion pipeline whose opp close_date is inside the window
(state == "open" and close_date in [window_start, window_end]); emits as an offsetting reason.
clean_billings: overdue_balance == 0. (Informational; included as a reason ONLY in the Renewal Risk
Queue family, never in the Action Board family.)
risk_score = segment base + factor points, capped at 100:
renewal_window +25, overdue_receivable +15, nps_drop +15, sla_degradation +15, usage_decline +10, low_tenure_high_churn +10; segment base: Strategic +20, Enterprise +5, Mid-Market / SMB +0.
(expansion_offset and clean_billings add 0 points; they surface as reasons only.)
risk_level thresholds:critical >= 80, high 50-79, medium 30-49, low < 30.
primary_action priority ladder:
In the Action Board family ONLY: if risk_level == low -> no_action.
Else if overdue_balance > 0 -> collections_followup.
Else if a significant SLA problem exists (any first-response miss, OR any SLA miss on a
P1/P2/P3 ticket) -> technical_recovery.
Else if in renewal window (only minor SLA issues, e.g. a lone P4/resolution-only miss) -> renewal_save.
Else -> nurture_monitor.
(executive_qbr exists in the enum for high-touch strategic escalations but did not fire in training.)
Sort / tie-break: order by risk_score descending, then current_arr descending. The Renewal Risk
Queue returns the top 5; the Action Board returns all requested accounts in this same order.
Using account.billing_arr_current / crm_arr instead of the latest posted billing snapshot.
Counting all tickets (or metrics.support_ticket_count) instead of cleaned tickets.
Summing ALL overdue buckets instead of just 61_90 + 90_plus.
Fuzzy-linking A/R lookalikes to CRM accounts (only exact legal_name links).
Treating NPS -1 / null / retracted as real, or using a flat NPS threshold without the recovery rule.
Computing SLA compliance from metrics.sla_compliance (QBR SLA% is derived from clean tickets'
first-response SLA — see playbook).
Emitting clean_billings in the Action Board family, or no_action in the Renewal Risk Queue family.
Forgetting the current_arr descending tie-break, or returning more/fewer rows than requested.
Inventing policy_codes instead of using the fixed values above.
Always read references/playbook.md for the exact recipe of the family you are solving, including the
QBR, receivables/pipeline, and churn procedures and their precision rules.
