# Task chunking rule

- No single job should run longer than 40 minutes. Plan chunks before you start.
- After each chunk, write a checkpoint (what's done, what's next, file paths/ids) to persistent storage.
- On restart, read the checkpoint first; never redo finished chunks.
- Long waits (builds, crawls, deploys) run in the background; poll briefly instead of blocking.
- Report `interruptions` (timeouts, killed_tasks) honestly at check-in.