Decide what an agent should do when the same request arrives twice

You are designing duplicate-handling rules for an automation that takes actions in real systems. Below is the trigger, the actions it performs, the systems it writes to, and the ways a run can repeat (retry, webhook redelivery, user double-submit, backfill): {{trigger_actions_target_systems_and_known_repeat_scenarios}}

1. WHAT BREAKS ON A REPEAT - table: Action | Safe to repeat? (yes / no / partly) | What a duplicate looks like to a customer | How I would notice.
2. IDENTITY KEY - the exact field or combination that makes one request one request, and what to do when the key is missing or arrives late.
3. RULES - per action: skip, update in place, or run again, with the check that decides. Write these as conditions an engineer can implement.
4. HALF-FINISHED RUNS - what happens when the run died between two writes, and how the next attempt tells "already done" from "never started".
5. WHAT TO LOG - the minimum fields needed to prove a duplicate was caught, and how long to keep them.

No lecture on idempotency theory. Use my system and field names only; if a key does not exist in what I gave you, say so instead of inventing one.