Amazon Inventory Orchestrator

Amazon Inventory Orchestrator

Use this Skill as a thin entrypoint. Keep inventory-ai-runner as the only calculation, snapshot, and state center. Do not reproduce or modify the clearance, replenishment, or repricing formulas in this Skill.

Resolve the run

1. Resolve exactly one sellerId + marketplace from the user request or config. If neither supplies an unambiguous scope, ask the user to select one store and site. Never traverse all stores implicitly.
2. Choose one run type:
   • WEEKLY_FULL: complete store/site weekly shadow run.
   • SKU: incremental recalculation; require one Seller SKU.
   • PILOT: representative-SKU validation; require one Seller SKU.
3. Collect the inventory/age report, SKU mapping, sales windows, and config. Add costs, exact fees, current effective prices, competitor events, and image cache when available. Competitor events must be assessed once through $amazon-competitor-monitor; do not copy its seller-level FBA-transition scoring into clearance, replenishment, or repricing.
4. Reject a formal conclusion when the inventory-age snapshot is more than seven days old or SKU/ASIN/snapshot identities conflict. A confirmed A-grade competitor observation older than four days may remain in the review queue but cannot produce a formal price suggestion.

Before invoking any calculator, remove pencil-case records from the normal SKU set. They may appear only in a separate 历史遗留清退 output and must not affect sales, profit, success rate, inventory, selection, or replenishment totals.

Run the canonical workflow

From ${INVENTORY_AI_RUNNER}, call:

node run_inventory_decision.mjs \
  --mode <WEEKLY_FULL|SKU|PILOT> \
  --report <inventory-age-report.csv> \
  --mapping <sku-mapping.json> \
  --windows <sales-windows.json> \
  --config <inventory-ai.config.json> \
  --as-of-date <YYYY-MM-DD> \
  --output-dir <shadow-output-directory>

Pass --pricing, --competitors, and --image-cache when supplied. For SKU and PILOT, also pass --sku. Pass --pilot-baseline only in PILOT and only when its snapshot date matches the current report.

The runner must call the canonical exported calculators: calculateClearance, calculateReplenishmentScenarios, and repricing calculate. Never use a copied formula or a silent fallback.

Interpret decisions

• Treat clearance_decision.main_status as the only inventory master status.
• For 紧急清退, 主动清货, or 停补观察, require replenishment quantities and FBM reminders to be zero.
• For 待补货, explain both the 7-day fast and 14-day formal scenarios.
• Carry competitor_fba_transition_risk, reasons, stress-test requirement, and manual-quantity-review flag into the replenishment decision. A high-risk result may recommend a smaller trial, but must preserve the calculator's original quantities until the user confirms a conservative quantity.
• Route repricing only for urgent/active clearance, low-stock replenishment risk with recent sales, or a confirmed A-grade competitor price/delivery anomaly.
• A confirmed A-grade FBM→FBA event routes repricing review and replenishment recalculation. FBA-transition risk alone never changes the clearance master status and never opens the clearance route.
• Treat TEMP_A results as provisional. Treat B/C competitors as manual-review evidence only. Allow a healthy SKU to return HOLD.
• If cost, exact fees, current effective price, fresh confirmed-A evidence, or identity checks are missing, keep formal price fields empty and report the blocking codes.

Preserve safety and state

• Generate read-only previews only. Never purchase, create an FBA shipment, change a price, or activate/modify an FBM offer.
• Keep project-level price_write as FORBIDDEN even if a calculation is otherwise ready.
• Never recommend, create, or execute an own-store Coupon.
• Use inventory-actions.json schema 1.1.0; do not create a parallel state file.
• Preserve user decisions and historical snapshots idempotently. Do not overwrite them during a model refresh.
• Reuse the existing ai heartbeat. Do not create another automation.
• Keep full-store output in a shadow directory until representative-SKU review and safety checks pass.
• The Feishu operations dashboard and HTML dashboard are retired. Never build, validate, sync, publish, link, or refresh them. dashboard-payload.json may remain an internal compatibility payload, but it is not a user entrypoint.
• Generate reports with build_inventory_skill_workbooks.mjs in SEPARATE_WORKBOOKS mode. Never create or send a combined workbook.

Report the result

State the run type, seller/site, as-of date, source snapshot, status counts, replenishment count, repricing queue/ready counts, blockers, and confirmation that all write policies remained forbidden. Link inventory-actions.json, then show three independent attachment sections in this order: repricing, replenishment, clearance. Each section must contain only its own summary and XLSX. On Thursday, send a changed-domain workbook only when that domain changed; otherwise send the domain-specific text 本周期无关键变化. Feishu delivery is three independent bot file messages to the configured Vicky open_id and must use domain-level idempotency. Do not link dashboard-payload.json to the user.