Add --interactive flag for human-in-the-loop checkpoints between waves

[nickmaglowsch]

Problem

The Anthropic article uses "feature-at-a-time decomposition" with evaluation between features. Currently, once the orchestrator starts, it runs all waves autonomously. For high-risk tasks (migrations, breaking API changes), the user has no opportunity to review intermediate results.

Proposed Fix

Add an --interactive flag to /build:

1. build/SKILL.md — parse --interactive flag from arguments
2. parallel-task-orchestrator — when interactive mode is enabled:
   • After each wave completes, pause and report:

     Wave N complete: [task-01 ✅, task-03 ✅]
     Files changed: [list]
     Next wave: [task-02, task-04]
     Continue? (Yes / Abort / Review changes first)
     

   • Use AskUserQuestion for the checkpoint
   • Default behavior (no flag) remains fully autonomous

This is opt-in — zero overhead for normal builds.

Effort

Medium — flag parsing in skill + conditional checkpoint in orchestrator.