Repository navigation
v1.13.0
Token Efficiency: Agent-Isolated ddd-develop Iterations
Each ddd-develop cycle in ddd-auto now runs in an isolated Agent (subagent) context instead of the main session.
Before
- Each iteration accumulated 30-80K tokens in the main session
- 10 items = 300-800K tokens, risking context window limits and prompt cache misses
After
- Each iteration returns only a ~200 token structured summary to the main session
- 10 items = ~10-15K tokens in the main session
- Prompt cache stays warm across iterations
Changes
ddd-auto/SKILL.md: Step 6 now dispatches Agent tool instead of direct/ddd-developcallstop-hook.sh: Re-inject prompt updated to instruct Agent dispatch- Version bumped to 1.13.0