Generate the report by writing one Python script via code execution that reads the CSVs and emits markdown. Do not make per-SKU tool calls.
Structure
# Inventory Report — {{warehouse or "All Warehouses"}} — week of {{date}}
## Stockouts (on_hand = 0)
| SKU | Product | Warehouse | Days out |
...
## Low Stock (below reorder point)
| SKU | On hand | Reorder pt | Days cover | Action |
...top 15 by urgency (lowest days_cover first)...
Open POs
| PO | SKU | Qty | Supplier | ETA | ...from /mnt/user/sinks/purchase_orders.jsonl...
Forecast Risk
SKUs where promo_next_month=1 or is_seasonal=1 and on_hand < 14d cover. One line each: SKU, reason, recommended action.
## Operating cadence (which report is being asked for)
| Cadence | Trigger phrasing | Contents |
|---|---|---|
| **Daily** | "run the check", "the sweep" | Low-stock list with action taken per SKU; one summary notification at the end. |
| **Weekly** (Mon) | "the report", "weekly review" | Per-warehouse: top concerns, **open POs aging past their lead time**, SKUs below reorder for >5 business days. |
| **Monthly** | "supplier review" | Suppliers whose on-time rate slipped; SKUs whose primary supplier may need changing. |
| **Ad hoc** | anything else | Scope to what was asked. |
If the request doesn't say which, infer from wording. The structure below is
the **weekly** format; for daily, drop the Open-POs and Forecast-Risk sections
and lead with the actions taken.
## Aging-PO check (weekly only)
For each open PO, compare days-since-placed to the supplier's `lead_time_days`.
List any PO where elapsed > lead_time as **aging** and include supplier + days
overdue so ops can follow up.
## Data sources
- Stockouts & low stock: latest-date rows from `/mnt/user/data/stock_levels.csv` joined with `/mnt/user/data/products.csv`
- Days of cover: `on_hand / avg_daily_sales` (last 14d from `/mnt/user/data/sales_history.csv`)
- Open POs: `/mnt/user/sinks/purchase_orders.jsonl`
- Forecast risk: `/mnt/user/data/products.csv` flags + days-of-cover from above
## Do this in code
The CSVs are large (stock_levels is ~67k rows). Write a single script that loads them once, computes everything, and prints the markdown. Don't page through the data with tool calls — that's exactly the pattern this skill replaces.