Thomas Jetzinger
|
d2d6328be2
|
feat(bmm): add story-pipeline workflow for 60-70% token savings
Single-session step-file architecture replacing 6 separate claude calls:
Structure:
- 8 step files (~200 lines each) with just-in-time loading
- Role switching (SM → TEA → DEV) within session
- State YAML for checkpoint/resume capability
- Checklists for validation, ATDD, implementation, code review
Steps:
1. Init - Load context, detect mode
2. Create Story - SM generates story from epic
3. Validate Story - Adversarial validation (find issues)
4. ATDD - TEA generates failing tests (RED phase)
5. Implement - DEV implements to pass tests (GREEN)
6. Code Review - Adversarial review (must find 3-10 issues)
7. Complete - Git commit, status update
8. Summary - Audit trail generation
Token efficiency:
- Legacy (6 calls): ~71K tokens/story
- Interactive mode: ~25K tokens (~65% savings)
- Batch mode: ~30K tokens (~58% savings)
Tested with real story (1-4 User Invitation): 17 files, +2800 lines,
context exhaustion mid-session recovered via state file.
|
2025-12-26 13:12:40 +01:00 |