quest-exec

Subagent orchestration prevents context contamination by spawning a fresh subagent per task. This approach ensures isolated execution contexts and atomic commits, keeping multi-task phase execution deterministic and free of cross-task state bleeding.

Executing planned phases requires reading task definitions from PLAN.md and validating prerequisites first. The orchestrator then processes tasks sequentially or in parallel, emitting phase_executing events for tracking, and aggregates results into PHASES.md and STATE.md.

Atomic commit automation works by provisioning isolated subagents that complete individual tasks and commit changes independently. This guarantees that each task's modifications are tracked separately, preventing partial or overlapping commits across the phase.

Yes, you need a PLAN.md file to define task structures. The orchestrator reads this file to identify planned phases under phases/NN-name directories, validates prerequisites, and uses the defined tasks to provision fresh subagents for execution.

Phase execution state is tracked by emitting phase_executing events during processing. After subagents finish, the system aggregates task results and updates PHASES.md with phase status, while STATE.md logs changes for full traceability.

Context rot occurs when tasks share the same execution context. Spawning fresh subagents for each task eliminates this issue by ensuring isolated contexts, preventing accumulated state from degrading determinism in later phases.