To Tasks

Turn a plan, or what this session has worked out, into a task list a less capable agent can retrace step by step: verbatim tests, verbatim code, and a runnable check per task. You walk the code once so the executor doesn't have to. Write the list; don't run it.

Stop and ask. Follow the source as given, but when something is missing, contradictory, or looks like the wrong design, tell the user and wait. Don't fill the gap with a guess.

1. Pick the source. Use the file the user names, and the slice they give (slice-N) if it has slices. With no file, use this session's discussion. If neither says enough to build tasks, ask.

2. Read the code it touches. Ground every task in live code, not the plan's wording. Load the convention and testing skills for each language the tasks will contain. Done when you know every signature, path, and test pattern the tasks rely on.