sg-parallel-execute

Parallel execution of independent plan groups requires a parallel_groups.json file defining wave-structured tasks. The orchestrator dispatches up to three Task agents concurrently per wave while preserving dependency barriers, reading each group's PLAN.md directly without invoking superpowers:executing-plans.

Wave-based parallel task dispatch executes independent groups concurrently in batches of up to three Task agents per wave. It processes waves sequentially from lowest to highest wave number, ensuring dependency barriers between waves are preserved while maximizing concurrency within each wave.

Setting up parallel_groups.json for concurrent plan execution requires defining independent plan groups in a JSON array format, each with its own PLAN.md file. Groups must be organized into waves where dependencies dictate execution order, with each group executed via bare Task() calls.

When a Task agent fails during parallel wave execution, the orchestrator appends a parallel-failed row to .planning/HANDOFF.md and halts all further wave advancement. This prevents inconsistent progress across independent plan groups and stops subsequent waves from executing.

Parallel plan execution explicitly avoids invoking superpowers:executing-plans and instead uses bare Task() execution to run each group's PLAN.md directly. This non-invasive approach ensures Task agents do not touch GSD or Superpowers internal state files during concurrent dispatch.

Parallel orchestration limits concurrency to three Task dispatches per wave to balance throughput with resource constraints and dependency safety. This maximum prevents overwhelming the system while still reducing slow sequential execution time by running independent plan groups concurrently within each wave barrier.