Wave Execution
Execution pattern where each "wave" starts with a fresh 200K token context. Instead of one long session (where quality degrades beyond 40% of context), work is split into independent waves. Each wave receives a summary brief from the previous wave. Used by GSD for its "slices" and the "Get Shit Done" pattern for long autonomous sessions.
Strengths
- Maintains response quality by keeping context below the degradation threshold (~40%)
- Structured handoff between waves, summary brief preserves continuity without filling context
- Compatible with long autonomous sessions: GSD chains slices automatically
- Resilience: if a wave fails, previous waves remain valid, no total loss
Limitations
- Transition overhead: writing a good summary brief between waves takes time
- Risk of implicit context loss between waves, the summary does not capture everything
Best for
- Complex features requiring more than 2 hours of continuous work with Claude Code
- Multi-file migrations or refactors where quality over 8h must remain consistent
- Long GSD autonomous sessions: auto-mode natively applies this pattern