ldy26098
/
BMAD-METHOD
mirror of https://github.com/bmad-code-org/BMAD-METHOD.git
1
0
Fork 0
Code Packages Projects Releases Wiki Activity
1,187 Commits 62 Branches 141 Tags 74 MiB
Commit Graph

4 Commits

Author SHA1 Message Date
Jonah Schulte 2c84b29cb6 fix(batch-super-dev): Fix conditional questions + add 'Mend the Gap' context 
PROBLEM 1: Confusing Conditional Question Logic
- Step 3.5 asked BOTH questions (Sequential/Parallel AND agent count) simultaneously
- User selects Sequential but still had to answer 'how many agents?'
- LLM misinterpreted agent count answer and launched parallel agents incorrectly

FIX 1: Split into Proper Conditional Flow
- First <ask>: Sequential or Parallel? (required)
- Second <ask>: ONLY if Parallel selected, ask agent count (conditional)
- Sequential mode now skips agent count question entirely
- Matches proper workflow.xml semantics (one <ask> = one response)

PROBLEM 2: Missing Use Case Context
- Workflow didn't explain it's primarily for gap analysis/reconciliation
- LLM didn't understand 'Sequential' means process in THIS session (not background)
- Missing mental model: verify existing → build gaps → check boxes → done

FIX 2: Add 'Mend the Gap' Header with Use Case Guide
- Added description: gap analysis + reconciliation workflow
- Explained 3 modes: Reconciliation (most common), Greenfield, Brownfield
- Clarified Sequential = in-session processing (not background agents)
- Added simple mental model for LLM understanding

IMPROVEMENT 3: Safety Limit on Parallel Agents
- Added 10-agent maximum (prevents resource exhaustion)
- 'all' option caps at min(story_count, 10)
- Added warning when count capped

RESULT:
- Sequential mode works as intended (no confusion)
- Parallel mode asks for agent count (conditional)
- LLMs understand primary use case (gap analysis)
- Safety limits prevent runaway agent spawning
- Proper workflow.xml conditional flow

Resolves: Execution mode confusion that caused incorrect parallel agent launches
 2026-01-07 19:54:26 -05:00
Jonah Schulte 9bdf489438 fix: resolve critical complexity routing issues from multi-agent review 
Fixes 6 critical issues discovered in multi-agent code review:

1. **Parameter propagation** - CRITICAL FIX
   - Added complexity_level parameter to super-dev-pipeline invocations
   - Fixed both sequential and parallel execution paths
   - Without this, complexity routing was completely non-functional

2. **Keyword matching rules** - CRITICAL FIX
   - Defined explicit matching algorithm in workflow.yaml
   - Case insensitive, word boundary matching, exact strategy
   - Added keyword variants (auth -> authentication, authorize, etc.)
   - Scan locations: story_title, task_descriptions, subtask_descriptions

3. **Threshold decision tree** - CRITICAL FIX
   - Rewrote overlapping logic to be mutually exclusive
   - Priority order: COMPLEX → MICRO → STANDARD
   - Prevents stories from matching multiple categories

4. **Task counting method** - CRITICAL FIX
   - Defined method: "top_level_only" (ignore subtasks)
   - Added documentation with examples
   - Eliminates ambiguity in complexity scoring

5. **max_files implementation** - FIX
   - Added file_count ≤ 5 check to MICRO classification
   - Previously extracted but never used (dead code)

6. **Version synchronization** - FIX
   - Updated super-dev-pipeline to v1.3.0 (was 1.2.0)
   - Matches batch-super-dev version for consistency

Impact: These fixes make complexity routing actually functional. The original
implementation computed complexity but never passed it to the pipeline,
rendering the entire feature non-operational.
 2026-01-07 16:34:44 -05:00
Jonah Schulte e5ede9ec3f feat: add complexity-based routing and pipeline optimizations (v1.3.0) 
Phase 1 & 2 improvements to reduce token waste and improve robustness:

1. Complexity scoring in batch-super-dev (step 2.6)
   - Score stories as micro/standard/complex based on task count + risk keywords
   - Micro (≤3 tasks, low risk): lightweight path
   - Standard: full pipeline
   - Complex (≥16 tasks or high-risk): enhanced validation

2. Lightweight path for micro stories
   - Skip pre-gap analysis (step 2) for micro complexity
   - Skip code review (step 5) for micro complexity
   - Estimated 50-70% token savings on simple stories

3. Smart batching clarification
   - Fixed contradictory instructions in step-03-implement.md
   - Default: one task at a time
   - Exception: batch low-risk patterns (package installs, imports)

4. Gap analysis optimization
   - dev-story skips step 1.5 if recent gap analysis exists (<24h)
   - Added skip_gap_analysis flag for pre-validated stories

5. Early bailout pre-flight check (step 4.5)
   - Bail early if story already complete (all tasks checked)
   - Bail if no tasks found (malformed story)
   - Bail if missing required sections
   - Saves tokens on stories that don't need processing

6. Multi-agent review integration (step 5)
   - For complex stories: recommend /multi-agent-review
   - Provides architecture, security, and performance review
   - Dynamic agent selection based on changed files
 2026-01-07 16:20:34 -05:00
Jonah Schulte b517a8b1d0 Add batch-super-dev workflow with auto-installation and quality gates 
Features:
- Interactive batch selector for processing multiple ready-for-dev stories
- 5-layer quality gate system (pre-gap, post-validation, reconciliation, code review, sprint-status sync)
- Smart story reconciliation (auto-updates checkboxes based on Dev Agent Record)
- Anti-vibe-coding enforcement (verifies files exist, tests pass, implementations aren't stubs)
- Story validation & auto-creation (validates 12 BMAD sections, creates missing stories)
- Auto-generates /batch-super-dev and /super-dev-pipeline slash commands on installation

Platform Support:
- Claude Code: Auto-creates .claude-commands/batch-super-dev.md
- Windsurf: Auto-creates slash commands for cross-compatibility

Workflow Location:
- Source: src/modules/bmm/workflows/4-implementation/batch-super-dev/
- Installed: _bmad/bmm/workflows/4-implementation/batch-super-dev/

Quality Gates:
1. Pre-gap analysis validates story completeness
2. Post-implementation validation verifies every task (anti-vibe-coding)
3. Smart reconciliation updates checkboxes based on evidence
4. Multi-agent code review finds issues before commit
5. Python script validates sprint-status.yaml accuracy
 2026-01-07 12:12:13 -05:00