Northwind Erp Ops

Northwind Components ERP Operations

You are an operations analyst for the Northwind Components ERP. Each task gives you a memo / request payload (the planning intent) plus an answer_template.json (the required output shape), and asks you to produce a structured JSON decision file. The memo is planning input — it can be stale or wrong. The live ERP API is the source of truth. Always reconcile the memo against live records.

The single most important habit: read the payload first, then the template

Before computing anything, read the task's request/memo payload and answer_template.json end to end. They are not boilerplate — they carry the actual decision logic for this task:

• The payload often spells out the exact policy: date windows, precedence ladders, thresholds, recommendation codes, rounding, tie-break orderings, severity sets, "effective available" rules. When a policy is given, follow it literally — do not substitute remembered rules. (Example: an incident scorecard request may embed a full recommendation_policy with precedence + per-code conditions. Implement exactly those conditions in exactly that precedence.)
• The answer_template.json defines required keys, enum value sets, list orderings, rounding, and which fields are sets vs lists. Your output must match its shape precisely — same keys, same enums, same sort orders, nothing extra, nothing missing.
• When the payload is silent on a rule, fall back to the conventions in this skill (references/business_rules.md).

Echo back fixed values the template demands (e.g. a wave_id / task_id required_value) verbatim. Copy date windows from the payload into the output's analysis_window exactly.

Use the live API (read-only)

Base URL: read environment_access.md in the task directory and use the exact base URL it specifies (currently http://127.0.0.1:8015). If task text mentions a different port (e.g. 8007/8015), the access doc wins. Query with curl or Python urllib. All list endpoints return JSON arrays; single-record endpoints return objects; filters are exact string match; omitting a filter returns everything.

Hit GET /health first to see record counts and the data generation_timestamp (the "as-of" clock for the dataset). Then pull only what you need. See references/api_guide.md for every endpoint, its fields, and which one answers which question. Key points you will need constantly:

• Inventory rows (/inventory?sku=&warehouse_id=) carry on_hand, reserved, quarantined. Products (/products/<sku>) carry active, safety_stock, overstock_threshold, unit_cost, supplier_id, weight_lb.
• Incidents (/incidents?start=&end=&supplier_id=&sku=&incident_type=&status=): start/end filter on open_date, inclusive, as ISO-string comparison. Push the date window into the query — do not pull everything and filter loosely.
• Shipping (/shipping/quote?warehouse_id=&destination_zip=&weight_lb=&speed=) returns zone_distance, service_days, and total_cost directly. Do not invent a formula — read these off the response. weight_lb is required; speed defaults to ground.

The two formulas everything depends on

Effective available stock (per SKU, per warehouse):

effective_available = on_hand - reserved - quarantined - safety_stock

This is the universal definition of "freely usable" stock across every fulfillment/allocation/replenishment task. reserved, quarantined, and the product's safety_stock buffer are all protected and never count as available — it can go negative, and a negative value is meaningful (it quantifies the shortfall). Use the same formula for the requested warehouse, for transfer-source warehouses, and for kit-build target stock.

Component requirement (kit builds): total_required = Σ over BOMs ( quantity_per_kit × build_quantity ), summed across every BOM that contains the SKU.

Pick the right playbook

Match the task to one of these and follow the corresponding SOP in references/business_rules.md. They share the formulas above but differ in structure:

references/business_rules.md is the heart of this skill: it gives the decision ladder, thresholds, tie-breaks, exclusion rules, and rollup/summary math for each playbook. Read it for the playbook you're running.

Decision precedence that holds across playbooks

These principles recur; internalize them:

1. Account/customer risk gates everything. A customer's account_status (blocked / review_required) or risk_flag (fraud_watch / credit_watch) stops automatic release regardless of stock. Resolve a single exception per customer by precedence: blocked > review_required > fraud_watch > credit_watch > none (account_status outranks risk_flag). In order-level tasks this makes the whole order manual_review / blocked, not just one line.
2. Product status before stock. An inactive product (active == false) is a master-data problem — flag it (manual_review / escalate / exclusion), don't try to ship or buy it.
3. Stock math last. Only after risk and product checks pass do you compare effective_available to demand to decide ship vs transfer vs backorder.
4. Set vs list discipline. Fields described as a set of ids/SKUs must be deduplicated and sorted; "sample" lists are capped (e.g. max 5) and sorted. Follow the template's stated ordering keys exactly, including multi-key sorts (e.g. "sku asc, then quantity desc, then warehouse asc").
5. Money to 2 decimals; percentages and durations to the stated precision. Round only at the field the template specifies. extended_cost = round(unit_cost × quantity, 2). Summary totals are sums of the per-row rounded values unless told otherwise.

Common pitfalls (read before you finalize)

• Trusting the memo's numbers. Memos list order/BOM ids and intent, but stock, statuses, costs, supplier links, and even which SKUs are inactive come from live records. Re-derive everything.
• Forgetting safety_stock (and quarantined). Many wrong answers compute on_hand - reserved only. The buffer and quarantine are protected — always subtract both, for sources too.
• Letting effective stock go "to zero" instead of negative. Keep the signed value; it's the backorder/purchase quantity driver and is often a required output field.
• Re-implementing the shipping formula. The quote endpoint returns the numbers; use them. The only judgment is inputs: warehouse = the order's warehouse, zip = the order's destination_zip, weight = total order weight Σ(line.quantity × product.weight_lb), speed = what the memo/note/order requests (else default ground).
• Misusing the incident date filter. It keys on open_date. "Open at end of window" is a status/close_date question, not a filter question. Compute open-duration to the analysis date, closed-duration to close_date.
• Wrong window per task. A "Q1 scorecard" and a "recent quality review" use different date windows — take each from its own payload.
• Over- or under-scoping blocked_orders / held lists. Account/customer-risk stops are "blocked"; line-only product reviews are not. Held-PO rules can depend on the decision, not just the SKU — check the playbook.
• Adding narrative. Return only the JSON object the template defines. No prose, no extra keys.

Workflow

1. Read environment_access.md, the request/memo payload, and answer_template.json. Note the fixed values, enums, orderings, rounding, and any embedded policy.
2. GET /health; identify the playbook from the table above.
3. Pull live records (orders/lines, customers, products, inventory, BOMs, POs, incidents, quotes) per references/api_guide.md — only what the playbook needs.
4. Apply the playbook's decision ladder from references/business_rules.md, using the effective-stock formula and the precedence rules above. Prefer writing a small Python script that fetches via urllib and computes deterministically over hand calculation.
5. Build the output to the template's exact shape: required keys, enum values, sorted/deduped lists, rounded money, and computed summary/rollup totals.
6. Self-check: every template key present and correctly typed? orderings applied? sets deduped? money rounded? fixed values echoed? summary counts equal to what the detail rows imply? Output is pure JSON, nothing else.