Scale-adaptive intelligence: automatic scope→workflow-depth mapping

[azalio]

Summary

map-framework has map-efficient vs map-fast as a user choice based on token cost, not on project complexity. BMAD has scale-adaptive intelligence: the system automatically adjusts planning depth based on project scope — Quick Flow track skips phases 1-3 for small work, full 4-phase pipeline for enterprise-scale projects.

What BMAD does

Two-track system

1. Quick Flow (bmad-quick-dev):

   • Skip phases 1-3 (analysis, planning, solutioning)
   • Unified: clarify intent → plan → implement → review → present
   • For small, well-understood work (bug fixes, minor features, refactors)

2. Full Pipeline (4 phases):

   • Phase 1: Analysis (brainstorming, market/domain/tech research, product brief/PRFAQ)
   • Phase 2: Planning (PRD, UX design)
   • Phase 3: Solutioning (architecture, epics & stories, implementation readiness check)
   • Phase 4: Implementation (sprint planning, create story, dev story, code review, retrospective)

Spec scope standards

• bmad-spec targets a single user-facing goal within 900-1600 tokens
• Scope auto-detection: the agent decides if the input warrants full pipeline or Quick Flow

Automatic adaptation

• User says "fix typo in README" → Quick Flow, no spec, no plan, one-step implement
• User says "add OAuth2 to authentication module" → Full pipeline, architecture review, multiple stories

Current map-framework state

• map-efficient (full MAP loop) vs map-fast (skip Predictor/Reflector) — based on user choice
• No automatic scope detection triggering the right workflow depth
• No "quick path" that bypasses planning entirely for trivial changes
• map-plan always decomposes, even for 1-subtask tasks

Proposed design

Scope auto-detection at entry

1. Classify intent before routing to skill:

   TRIVIAL: < 3 files, < 50 lines, no new modules → map-fast (skip Predictor, Reflector)
   SMALL:   1 component, well-understood → map-plan-light (spec only, no research)
   MEDIUM:  2-5 components, moderate risk → map-efficient (full MAP loop)
   LARGE:   6+ components, high risk → map-efficient + map-tdd + adversarial review
   

2. Configurable thresholds: .map/config.yaml keys for scope classification
3. Override: --force-full / --force-fast to bypass classification

Light plan mode (map-plan --light)

• Generate spec from intent (no research phase)
• Minimal subtasks (2-5, bite-sized)
• Skip Predictor, use Actor+Monitor only
• No Evaluator/Reflector unless Monitor raises issues

Deep plan mode (map-plan --deep)

• Full research phase (SOFA, codebase exploration, dependency analysis)
• Architecture review before task decomposition
• Cross-artifact consistency check
• Mandatory review gate after each phase

Integration

• .map/config.yaml:

  scale:
    auto: true
    thresholds:
      trivial: { max_files: 3, max_lines: 50 }
      small:   { max_files: 10, max_lines: 200 }
      medium:  { max_files: 30, max_lines: 1000 }

References

• BMAD: src/bmm-skills/4-implementation/bmad-quick-dev/SKILL.md, src/core-skills/bmad-spec/SKILL.md
• BMAD module-help.csv: workflow dependency graph with Quick Flow as parallel track

Acceptance criteria

• Scope auto-detection before routing to map-* skill
• Configurable thresholds for trivial/small/medium/large classification
• map-plan --light mode (spec only, no research)
• map-plan --deep mode (full research + architecture review)
• Override flags --force-full / --force-fast
• .map/config.yaml scale section

[azalio]

Отложено пользователем — вернёмся позже.