subagent-driven-development

You can orchestrate multi-task plans by dispatching a fresh subagent per task. This sequential execution model isolates context between tasks, preventing pollution and maintaining spec compliance across the entire session.

Two-stage code review validates each task sequentially by first checking spec compliance, then evaluating code quality. This gating mechanism ensures changes meet requirements before assessing implementation structure.

Track changes by recording BASE_SHA before task execution and HEAD_SHA after completion. This diff-based validation pinpoints exact modifications, reporting changed files, test results, and commands for each task.

Yes, executing multiple tasks in a single session is possible by applying a dedicated, fresh subagent per task. This approach explicitly avoids context leakage and isolates execution environments for each step.

Diff-based validation is needed when executing sequential, plan-driven tasks where quality gating is required. It uses BASE_SHA and HEAD_SHA tracking to verify changes and enforce spec compliance before proceeding to the next task.

The best way to isolate execution is using a fresh subagent per task with two-stage reviews. This approach isolates context, gates quality through diff validation, and tracks changes using SHA references.