walkthrough

Move a cursor over a list, one item per turn. You present the item under the cursor with enough context to judge it, the user reacts, the cursor advances. The win is focus: the user processes and pushes back on each item instead of drowning in a dump.

The list is pinned — built once, held in context for the whole walkthrough. Each turn restates where the cursor is; never assume the user still sees the previous item.

Steps

1. Find the items. In priority order:

   • A file the user named (walk me through subtasks.json) → read it; treat array entries / list items / top-level sections as the items.
   • The conversation — findings, suggestions, tasks, options just produced → enumerate those.
   • An inline list in the user's message → parse it directly.

   If it's ambiguous what to walk through, ask once, then proceed.

2. Pin the list. Enumerate the items as a stable, ordered list and hold the full text in context — back, goto N, and find all read from it. Do not re-derive it each turn.

3. Announce, then present item 1. One line: Found N items. I'll go through them one at a time — ready for the first? Present item 1 in the same turn unless the user's request didn't already imply go.