maestro-flow 0.4.10 → 0.4.12

package/workflows/brainstorm.md

@@ -1,496 +1,603 @@
-# Workflow: Brainstorm
-
-Unified brainstorming workflow with dual-mode operation: auto pipeline (full multi-role analysis) and single role analysis mode.
-
-## Architecture
-
-```
-┌─────────────────────────────────────────────────────────┐
-│                  /maestro-brainstorm                    │
-│           Unified Entry Point + Interactive Routing      │
-└───────────────────────┬─────────────────────────────────┘
-                        │
-              ┌─────────┴─────────┐
-              ↓                   ↓
-    ┌─────────────────┐  ┌──────────────────┐
-    │   Auto Mode     │  │ Single Role Mode │
-    └────────┬────────┘  └────────┬─────────┘
-             │                    │
-    ┌────────┼────────┐          │
-    ↓        ↓        ↓          ↓
- Phase 2  Phase 3  Phase 4    Phase 3
-Artifacts  N×Role  Synthesis  1×Role
-           并行               Analysis
-```
-
-## Dual-Mode Routing
-
-### Auto Mode (full pipeline)
-Triggered by `--yes`/`-y` flag or user selection.
-
-```
-Phase 1: Mode Detection → Parse args, detect mode
-Phase 1.5: Terminology & Boundary → Extract terms, collect Non-Goals
-Phase 2: Interactive Framework → 7 sub-phases (context → topic → roles → questions → conflicts → check → spec)
-Phase 3: Parallel Role Analysis → N concurrent role analyses via conceptual-planning-agent
-Phase 4: Synthesis Integration → Cross-role analysis → user clarification → spec generation
-```
-
-### Single Role Mode
-Triggered when first arg is a valid role name.
-
-```
-Phase 1: Mode Detection → Parse args, detect mode
-Phase 3: Single Role Analysis → Detection → Context → Agent → Validation
-```
-
-## Input
-
-- `$ARGUMENTS`: topic text (auto mode) or role name (single role mode)
-- All output goes to `.workflow/scratch/brainstorm-{slug}-{date}/`
-- Registers artifact (type=brainstorm) in state.json on completion
-
-### Parameters
-
-| Parameter | Description | Default |
-|-----------|-------------|---------|
-| `--yes`, `-y` | Auto mode, skip all questions | - |
-| `--count N` | Number of roles to select | 3 |
-| `--session ID` | Use existing session | - |
-| `--update` | Update existing analysis | - |
-| `--include-questions` | Interactive context gathering | - |
-| `--skip-questions` | Use default answers | - |
-| `--style-skill PKG` | Style package for ui-designer | - |
-
-### Available Roles
-
-| Role ID | Title | Focus Area |
-|---------|-------|------------|
-| `data-architect` | 数据架构师 | Data models, storage strategies, data flow |
-| `product-manager` | 产品经理 | Product strategy, roadmap, prioritization |
-| `product-owner` | 产品负责人 | Backlog management, user stories, acceptance criteria |
-| `scrum-master` | 敏捷教练 | Process facilitation, impediment removal |
-| `subject-matter-expert` | 领域专家 | Domain knowledge, business rules, compliance |
-| `system-architect` | 系统架构师 | Technical architecture, scalability, integration |
-| `test-strategist` | 测试策略师 | Test strategy, quality assurance |
-| `ui-designer` | UI设计师 | Visual design, mockups, design systems |
-| `ux-expert` | UX专家 | User research, information architecture, journey |
-
-## Output
-
-### Directory Structure
-
-All brainstorm output goes to scratch:
-```
-.workflow/scratch/brainstorm-{slug}-{date}/
-├── guidance-specification.md         # Phase 2 output
-├── feature-index.json               # Phase 4 output
-├── synthesis-changelog.md            # Phase 4 audit trail
-├── feature-specs/                   # Phase 4 feature specs
-│   ├── F-001-{slug}.md
-│   └── F-00N-{slug}.md
-├── {role}/                          # Phase 3 role analyses (immutable)
-│   ├── {role}-context.md
-│   ├── analysis.md
-│   ├── analysis-cross-cutting.md
-│   └── analysis-F-{id}-{slug}.md
-└── synthesis-specification.md       # Non-feature mode fallback
-```
-
----
-
-## Process
-
-### Step 1: Parse & Route (Mode Detection)
-
-Parse $ARGUMENTS to determine execution mode:
-
-**Mode Detection (ordered by priority)**:
-1. `--yes` or `-y` flag → **Auto Mode** (no question asked)
-2. First non-flag arg matches valid role name → **Single Role Mode**
-3. First non-flag arg is a number → **Phase Mode** (resolve phase dir, then auto)
-4. Text provided without flags → Ask user via AskUserQuestion:
-   - "自动模式 (推荐)" — 完整流程：框架生成 → 多角色并行分析 → 跨角色综合
-   - "单角色分析" — 为单个角色生成分析文档
-
-**Parameter Parsing**:
-- `--count N`: cap at 9, default 3
-- `--session ID`: target specific session
-- `--style-skill PKG`: validate `.claude/skills/style-{PKG}/SKILL.md` exists
-- Missing/empty args without flags = error E001
-
-**Session Detection**:
-- Check `.workflow/scratch/brainstorm-*/` for existing sessions
-- Multiple → AskUserQuestion to select | Single → use it
-- None + auto mode → will create new session
-- None + single role mode → error E002
-
-**Output Directory Resolution**:
-- Phase mode (number): resolve `state.json.artifacts[phase == phaseNum].path` → `.workflow/{path}/.brainstorming/` (ERROR if phase not found)
-- All output: `.workflow/scratch/brainstorm-{slug}-{date}/`
-- Existing session: use existing session directory
-
----
-
-### Step 1.5: Load Project Specs
-
-```
-specs_content = maestro spec load --category arch
-```
-
-Pass to conceptual-planning-agent in Step 4 for architecture-aware role analysis.
-
----
-
-### Auto Mode Steps (Phase 1.5 → Phase 1.7 → Phase 2 → Phase 3 → Phase 4)
-
-### Step 1.7: External Research — Design Routes (Auto Mode, Optional)
-
-Spawn `workflow-external-researcher` agent to discover design alternatives, architecture patterns, and competitive approaches for the brainstorm topic. This enriches the framework generation and role analyses with external knowledge.
-
-**Trigger**: Always in auto mode. Skip if `--skip-questions` and no tech keywords detected.
-
-**Auto-suggest when**: Topic contains technology keywords, architecture terms, or "design" / "pattern" / "alternative" in the description.
-
-```
-// Step 1.7.1: Spawn external researcher for design routes
-Agent(
-  subagent_type="workflow-external-researcher",
-  prompt="""
-<objective>
-Research design alternatives and architecture patterns for: {topic}
-Mode: Design Research
-</objective>
-
-<context>
-Project specs: {specs_content or "none"}
-Topic keywords: {extracted_keywords}
-</context>
-
-<task>
-Search for:
-1. Reference projects — how 2-3 similar projects/products solve this problem (architecture, key decisions, what worked)
-2. Extractable patterns — reusable design patterns distilled from those projects, with applicability notes
-3. Architecture approaches (at least 2-3 alternatives with trade-offs)
-4. UX/UI patterns if applicable (interaction models, layout strategies)
-5. Common design pitfalls and anti-patterns to avoid
-
-IMPORTANT: Output MUST include "Reference Projects / Implementations" and "Extractable Patterns" sections.
-Focus on design ROUTES — alternative approaches the brainstorm roles can evaluate.
-Be prescriptive where evidence is strong, present alternatives where trade-offs exist.
-Return structured markdown only — do NOT write files.
-</task>
-  """,
-  run_in_background=false
-)
-
-// Step 1.7.2: Store as designResearchContext (in-memory)
-designResearchContext = agent_output
-```
-
-`designResearchContext` is passed into:
-- Step 2 (Terminology): enriches domain term extraction
-- Step 3 Phase 1 (Topic Analysis): provides external design alternatives
-- Step 4 (Parallel Role Analysis): each role agent receives design research as additional context
-
-If research fails (W005): `designResearchContext = null`, continue without external context.
-
----
-
-### Step 1.8: Load Project Context (if `.workflow/` exists)
-
-Load existing project history to ground brainstorming in what's already been built:
-
-- From `.workflow/project.md`: `### Validated` → already_shipped, `### Active` → current_scope, `## Context` → project_history
-- From `.workflow/state.json.accumulated_context`: `deferred[]` → deferred_items, `key_decisions[]` → existing_constraints
-
-Pass `project_context` into Step 2 (terminology) and Step 3 (framework generation):
-- `already_shipped` informs what exists — brainstorm should explore extensions, not re-invent
-- `deferred_items` are high-value brainstorming seeds
-- `lessons_learned` surface pitfalls to avoid
-
----
-
-### Step 2: Terminology & Boundary Definition (Auto Mode)
-
-Extract core terminology and define scope boundaries before framework generation.
-
-1. Analyze topic description and any project context (project.md, roadmap.md, project_context from Step 1.8)
-2. Extract 5-10 core domain terms:
-   - term (canonical), definition, aliases, category (core|technical|business)
-3. AskUserQuestion for Non-Goals (multiSelect=true):
-   - Generate 4-5 context-aware exclusion candidates based on topic
-   - Include "其他（请补充）" option for custom exclusions
-   - If user selects "其他", follow up with free-text question
-4. Store terminology table and non_goals to session state
-
-**Skip if**: `--yes` flag (use auto-generated terms, empty non-goals)
-
-### Step 3: Interactive Framework Generation (Auto Mode)
-
-Seven sub-phases producing guidance-specification.md:
-
-**Phase 0: Context Collection**
-- Read init outputs directly: `.workflow/project.md` (tech stack, requirements, decisions), `.workflow/state.json` (project state), `.workflow/specs/` (conventions)
-- If `.workflow/` does not exist: continue without project context
-
-**Phase 1: Topic Analysis**
-- Load Phase 0 context (tech_stack, modules, conflict_risk)
-- Deep topic analysis (entities, challenges, constraints, metrics)
-- Generate 2-4 context-aware probing questions via AskUserQuestion
-- Questions MUST reference topic keywords (no generic questions)
-- Store to `session.intent_context`
-
-**Phase 2: Role Selection**
-- Analyze Phase 1 keywords → recommend count+2 roles with rationale
-- AskUserQuestion (multiSelect=true) for user to select `count` roles
-- If `--yes`: auto-select recommended roles
-- Store to `session.selected_roles`
-
-**Phase 3: Role-Specific Questions**
-- FOR each selected role, generate 3-4 deep questions mapping role expertise to Phase 1 challenges
-- AskUserQuestion per role (sequential, one role at a time)
-- Questions must include: implementation depth, trade-offs, edge cases
-- Store to `session.role_decisions[role]`
-- If `--yes`: skip all role questions
-
-**Phase 4: Conflict Resolution**
-- Analyze Phase 3 answers for contradictions, missing integrations, implicit dependencies
-- Generate clarification questions referencing SPECIFIC Phase 3 choices
-- AskUserQuestion (max 4 per round)
-- If NO conflicts detected: skip with notification
-- Store to `session.cross_role_decisions`
-
-**Phase 4.5: Final Clarification + Feature Decomposition**
-- Ask: "是否有前面未澄清的重点需要补充？" (无需补充 / 需要补充)
-- If "需要补充": progressive questions until resolved
-- Extract candidate features from all Phase 1-4 decisions (max 8)
-- Each feature: F-{3-digit} ID, kebab-case slug, description, related roles, priority
-- Validate: independence, completeness, granularity balance, boundary clarity
-- AskUserQuestion for user to confirm or adjust feature list
-- Store to `session.feature_list`
-
-**Phase 5: Generate Specification**
-- Load all decisions + terminology + non_goals + feature_list
-- Transform Q&A to declarative statements (CONFIRMED/SELECTED)
-- Apply RFC 2119 keywords (MUST, SHOULD, MAY, MUST NOT, SHOULD NOT)
-- Write `guidance-specification.md` with sections:
-  1. Project Positioning & Goals
-  2. Concepts & Terminology (table)
-  3. Non-Goals (Out of Scope)
-  4-N. [Role] Decisions (with RFC 2119)
-  Cross-Role Integration
-  Risks & Constraints
-  Feature Decomposition (table)
-  Appendix: Decision Tracking
-- Validate: no interrogative sentences, all decisions traceable, RFC keywords applied
-
-**Output**: `{output_dir}/guidance-specification.md`, session metadata (workflow-session.json)
-
-### Step 3.5: Visual Style Foundation (Auto Mode, conditional)
-
-When `ui-designer` is among the selected roles, establish the project's visual direction before role analysis begins. This ensures all downstream UX analysis works within a confirmed design system.
-
-**Condition**: Skip if `.workflow/impeccable/DESIGN.md` already exists (visual direction already locked).
-
-**Execution** (sequential):
-
-1. **Product context** — if `.workflow/impeccable/PRODUCT.md` missing:
-   ```
-   Skill({ skill: "maestro-impeccable", args: "teach" })
-   ```
-   This runs the teach interview to establish brand, users, personality, anti-references.
-
-2. **Visual exploration** — multi-style comparison and selection:
-   ```
-   Skill({ skill: "maestro-impeccable", args: "explore" })
-   ```
-   This generates multiple design system variants as HTML prototypes, launches visual comparison, and lets the user select/mix. Produces DESIGN.md on completion.
-
-3. Record in session metadata: `design_system_established: true`, `design_md_path: ".workflow/impeccable/DESIGN.md"`
-
-**`--yes` mode**: `explore` auto-selects variant 1 without visual comparison. `teach` still requires minimal input if PRODUCT.md is missing.
-
-**Skip mode**: If user explicitly passes `--skip-design` to brainstorm, skip this step entirely. ui-designer role will generate its own independent theme in Phase 2.
-
-### Step 4: Parallel Role Analysis (Auto Mode)
-
-For EACH selected role, spawn a conceptual-planning-agent in parallel:
-
-```
-Agent({
-  subagent_type: "conceptual-planning-agent",
-  prompt: "[role analysis prompt with framework + role template]",
-  run_in_background: false
-})
-```
-
-Each agent receives:
-- `guidance-specification.md` for framework context
-- Role-specific template from `~/.maestro/templates/planning-roles/{role}.md`
-- Feature list for feature-point organization
-- `--skip-questions` flag (context already gathered in Phase 2)
-- For ui-designer: `--style-skill {package}` if provided
-- If `designResearchContext` is set: include as "External Design Research" section in agent prompt (design alternatives, patterns, competitive analysis for the role to evaluate and reference)
-
-**Feature-Point Organization** (when feature list available):
-- `analysis.md` — Role overview INDEX only (< 1500 words)
-- `analysis-cross-cutting.md` — Cross-feature decisions (< 2000 words)
-- `analysis-F-{id}-{slug}.md` — Per-feature analysis (< 2000 words each)
-
-**Fallback Organization** (no feature list):
-- `analysis.md` — Main analysis (< 3000 words)
-- Optional `analysis-{slug}.md` sub-documents (max 5)
-
-**system-architect specific requirements**:
-- MUST include: Data Model (3-5 entities), State Machine (ASCII + transition table), Error Handling, Observability (5+ metrics), Configuration Model, Boundary Scenarios
-- All constraints MUST use RFC 2119 keywords
-
-**Quality Validation** (after each role completes, orchestrator self-check):
-- Verify `analysis.md` exists and is non-empty
-- Feature mode: verify `analysis-cross-cutting.md` + `analysis-F-*.md` files match feature list
-- Grep for RFC 2119 keywords (MUST/SHOULD/MAY) — warn if < 5 occurrences
-- Check word count (wc -w) against limits — warn if exceeded
-- system-architect: verify Data Model and State Machine sections exist
-
-**Parallel Safety**: Each role operates on its own directory. No cross-agent dependencies.
-
-### Step 5: Synthesis Integration (Auto Mode)
-
-Seven sub-phases producing feature specs from cross-role analysis:
-
-**Sub-phase 1: Discovery**
-- Detect session, validate analysis files exist
-- Load user intent from session metadata
-- Detect feature mode (feature decomposition table + analysis-F-*.md files)
-
-**Sub-phase 2: File Discovery**
-- Glob `{output_dir}/{role}/analysis*.md`
-- Extract role_analysis_paths and participating_roles
-- Feature mode optimization: read only analysis.md index files (~4.5K total) not sub-documents
-
-**Sub-phase 3A: Cross-Role Analysis Agent**
-- Spawn conceptual-planning-agent for cross-role analysis
-- Input: analysis index files (feature mode) or all analysis files (fallback)
-- Output: `enhancement_recommendations` (EP-001, EP-002, ...) + `feature_conflict_map` (per-feature consensus/conflicts/cross_refs)
-- Conflict resolution quality: actionable, justified ("because...tradeoff:..."), scoped, confidence-tagged ([RESOLVED]|[SUGGESTED]|[UNRESOLVED])
-
-**Sub-phase 3B: Brainstorm Confidence Scoring**
-
-Compute after cross-role analysis, before user interaction. Dimensions (5): role_coverage, cross_role_consistency, feature_completeness, spec_quality, design_feasibility. Factors (weights): analysis_depth(.30), evidence_strength(.25), coverage_breadth(.20), user_validation(.15), consistency(.10).
-
-Scoring points: per-role completion → `role_coverage`; after cross-role → `cross_role_consistency`; after user interaction → `user_validation`.
-
-**Quality mechanisms**: Pressure Pass (before spec gen) — verify claims backed by ≥2 role analyses. Devil's Advocate — challenge when dimension > 0.7. Conflict Detection (replaces stall) — >3 `[UNRESOLVED]` conflicts → challenge.
-
-**Readiness Gate** (before Sub-phase 4): Block if role_coverage < 100% | cross_role_consistency < 40% | feature_completeness below intent count | no pressure pass. If blocked → AskUserQuestion: fill gaps or proceed with residual risks. Append confidence summary to synthesis-changelog.md.
-
-**Sub-phase 4: User Interaction**
-- Enhancement selection: AskUserQuestion (multiSelect=true, batched by 4)
-- Clarification questions: 9-category taxonomy scan, AskUserQuestion (single-select, multi-round)
-- Build spec_context: selected_enhancements + clarification_answers + original_user_intent
-- If `--yes`: auto-select all enhancements, skip clarifications
-
-**Sub-phase 5: Spec Generation + Conditional Review**
-- Single conceptual-planning-agent generates all specs sequentially:
-  - Feature mode: one `feature-specs/F-{id}-{slug}.md` per feature (7 sections, 1500-2500 words)
-  - Fallback mode: single `synthesis-specification.md`
-  - `feature-index.json` (feature mode only)
-  - `synthesis-changelog.md` (enhancements applied, clarifications resolved, conflicts resolved)
-- Self-evaluate complexity_score (0-8 scale: feature count, unresolved conflicts, roles, cross-deps)
-- If complexity_score >= 4: trigger review agent for cross-feature consistency check
-  - Minor fixes applied directly, major issues flagged with [REVIEW-FLAG]
-
-**Feature Spec Template (7 Sections)**:
-1. Requirements Summary (RFC 2119 keywords)
-2. Design Decisions [CORE — 40%+ word count]
-3. Interface Contract
-4. Constraints & Risks
-5. Acceptance Criteria
-6. Detailed Analysis References (@-links)
-7. Cross-Feature Dependencies
-
-**Four-Layer Aggregation Rules**:
-- Layer 1: Direct Reference (consensus → quote roles)
-- Layer 2: Structured Extraction (complementary → merge, de-duplicate)
-- Layer 3: Conflict Distillation ([RESOLVED] → decision, [SUGGESTED] → recommended, [UNRESOLVED] → [DECISION NEEDED])
-- Layer 4: Cross-Feature Annotation (dependency notes, integration points)
-
-**Sub-phase 6: Finalization**
-- Update context-package.json with spec paths
-- Update session metadata (enhancements_applied, questions_asked, complexity_score, review results)
-- Completion report with next step suggestion
-
----
-
-### Single Role Mode Steps
-
-### Step 6: Single Role Analysis
-
-Execute analysis for ONE specified role with optional interactive context gathering.
-
-**Step 6.1: Detection & Validation**
-- Validate role_name against VALID_ROLES list
-- Detect session (--session or find existing)
-- Check for guidance-specification.md → framework_mode
-- Extract feature list → feature_mode
-- Check existing analysis → update_mode (ask: update/regenerate/cancel)
-
-**Step 6.2: Interactive Context Gathering**
-- Skip if `--skip-questions`
-- Force if `--include-questions`
-- Generate 3-5 role-specific questions (Chinese, with business context)
-- AskUserQuestion per batch (max 4 per round)
-- Save context to `{role}/{role}-context.md`
-
-**Step 6.3: Agent Execution**
-- Spawn conceptual-planning-agent with:
-  - Role name, framework (if exists), feature list (if exists)
-  - User context (if gathered), session metadata
-  - Role-specific template
-  - For ui-designer: style-skill package
-- Agent generates analysis files in `{output_dir}/{role}/`
-
-**Step 6.4: Validation**
-- Verify `analysis.md` exists
-- Check framework reference if framework_mode
-- Update session metadata with completion status
-- Report results with next step suggestions
-
----
-
-### Step 7: Final Report
-
-**Auto mode report:**
-- Session ID and output directory
-- Roles analyzed (N)
-- Features specified (N)
-- Enhancements applied (EP-IDs)
-- Complexity score and review status
-- Next:
-  Skill({ skill: "maestro-roadmap", args: "--mode full --from-brainstorm {sessionId}" })  — Generate full spec package
-  Skill({ skill: "maestro-analyze", args: "{topic}" })   — Evaluate feasibility + lock decisions
-  Skill({ skill: "maestro-analyze", args: "{phase} -q" })   — Quick decision extraction only
-  Skill({ skill: "maestro-plan", args: "{phase}" })       — Plan directly (if scope is clear)
-
-**Single role mode report:**
-- Role analyzed
-- Framework alignment status
-- Context questions answered
-- Output file location
-- Next: run more roles or Skill({ skill: "maestro-brainstorm", args: "--session {sessionId}" }) for synthesis
-
----
-
-## Quality Criteria
-
-- If `designResearchContext` is set: guidance-specification.md references external design findings
-- guidance-specification.md uses RFC 2119 keywords (MUST/SHOULD/MAY)
-- Concepts & Terminology section with 5-10 core terms
-- Non-Goals section with rationale
-- Feature Decomposition table (max 8 features, independently implementable)
-- Role analyses follow role-specific templates
-- system-architect includes: Data Model, State Machine, Error Handling, Observability
-- Feature specs have 7 sections, Section 2 (Design Decisions) is 40%+ of content
-- All conflicts resolved or marked [DECISION NEEDED]
-- synthesis-changelog.md records all synthesis decisions as audit trail
+# Workflow: Brainstorm
+
+Unified brainstorming workflow with dual-mode operation: auto pipeline (full multi-role analysis) and single role analysis mode.
+
+## Architecture
+
+```
+┌─────────────────────────────────────────────────────────┐
+│                  /maestro-brainstorm                    │
+│           Unified Entry Point + Interactive Routing      │
+└───────────────────────┬─────────────────────────────────┘
+                        │
+              ┌─────────┴─────────┐
+              ↓                   ↓
+    ┌─────────────────┐  ┌──────────────────┐
+    │   Auto Mode     │  │ Single Role Mode │
+    └────────┬────────┘  └────────┬─────────┘
+             │                    │
+    ┌────────┬──────────┬───────┐    │
+    ↓        ↓          ↓       ↓    ↓
+ Phase 2  Phase 3    Phase 4 Phase 5 Phase 3
+Artifacts N×Role     Cross-  Apply  1×Role
+          Analysis   Role    Resol. Analysis
+          (multi-    Review
+           file)     (digest)
+```
+
+## Dual-Mode Routing
+
+### Auto Mode (full pipeline)
+Triggered by `--yes`/`-y` flag or user selection.
+
+```
+Phase 1: Mode Detection → Parse args, detect mode
+Phase 1.5: Terminology & Boundary → Extract terms, collect Non-Goals
+Phase 2: Interactive Framework → 7 sub-phases (context → topic → roles → questions → conflicts → check → spec)
+Phase 3: Parallel Role Analysis → N concurrent {role}/ multi-file via role-design-author
+Phase 4: Cross-Role Review → cross-role-reviewer compares Decision Digests (read-only)
+Phase 5: Apply Resolutions → AskUserQuestion per finding → patch role files + guidance §12
+```
+
+### Single Role Mode
+Triggered when first arg is a valid role name.
+
+```
+Phase 1: Mode Detection → Parse args, detect mode
+Phase 3: Single Role Analysis → Detection → Context → Agent → Validation
+```
+
+## Input
+
+- `$ARGUMENTS`: topic text (auto mode) or role name (single role mode)
+- All output goes to `.workflow/scratch/brainstorm-{slug}-{date}/`
+- Registers artifact (type=brainstorm) in state.json on completion
+
+### Parameters
+
+| Parameter | Description | Default |
+|-----------|-------------|---------|
+| `--yes`, `-y` | Auto mode, skip all questions | - |
+| `--count N` | Number of roles to select | 3 |
+| `--session ID` | Use existing session | - |
+| `--update` | Update existing analysis | - |
+| `--include-questions` | Interactive context gathering | - |
+| `--skip-questions` | Use default answers | - |
+| `--style-skill PKG` | Style package for ui-designer | - |
+| `--review-only` | Skip Phase 3 (role analysis); run only Step 4.5 + Step 5 on existing role analyses | - |
+
+### Available Roles
+
+| Role ID | Title | Focus Area |
+|---------|-------|------------|
+| `data-architect` | 数据架构师 | Data models, storage strategies, data flow |
+| `product-manager` | 产品经理 | Product strategy, roadmap, prioritization |
+| `product-owner` | 产品负责人 | Backlog management, user stories, acceptance criteria |
+| `scrum-master` | 敏捷教练 | Process facilitation, impediment removal |
+| `subject-matter-expert` | 领域专家 | Domain knowledge, business rules, compliance |
+| `system-architect` | 系统架构师 | Technical architecture, scalability, integration |
+| `test-strategist` | 测试策略师 | Test strategy, quality assurance |
+| `ui-designer` | UI设计师 | Visual design, mockups, design systems |
+| `ux-expert` | UX专家 | User research, information architecture, journey |
+
+## Output
+
+### Directory Structure
+
+All brainstorm output goes to scratch:
+```
+.workflow/scratch/brainstorm-{slug}-{date}/
+├── guidance-specification.md          # Phase 2 output — machine contract (downstream consumes this)
+├── design-research.md                 # Optional Step 1.7 output
+├── system-architect/                  # Phase 3 per-role analysis (one folder per selected role)
+│   ├── analysis.md                    # INDEX — digest + cross-cutting foundations + file index
+│   ├── analysis-F-001-agent-loop.md   # per-feature analysis
+│   ├── analysis-F-002-skill-engine.md
+│   └── findings-perf-risk.md          # additional discoveries
+├── ux-expert/
+│   ├── analysis.md
+│   ├── analysis-F-001-agent-loop.md
+│   └── analysis-F-003-tui-frontend.md
+└── {role}/
+    └── ...
+```
+
+Clarifications and audit trail live in `guidance-specification.md` §11 Appendix. Cross-Role Resolutions in §12.
+
+---
+
+## Process
+
+### Step 1: Parse & Route (Mode Detection)
+
+Parse $ARGUMENTS to determine execution mode:
+
+**Mode Detection (ordered by priority)**:
+1. `--review-only` flag → **Review-Only Mode** (requires `--session ID`; runs Step 4.5 + Step 5 only)
+2. `--yes` or `-y` flag → **Auto Mode** (no question asked)
+3. First non-flag arg matches valid role name → **Single Role Mode**
+4. First non-flag arg is a number → **Phase Mode** (resolve phase dir, then auto)
+5. Text provided without flags → Ask user via AskUserQuestion:
+   - "自动模式 (推荐)" — 完整流程：框架生成 → 多角色并行分析 → 跨角色复审 → 决议回流
+   - "单角色分析" — 为单个角色生成 {role}/ 分析文件
+   - "跨角色复审" — 已有多个角色分析，仅运行复审与决议回流
+
+**Parameter Parsing**:
+- `--count N`: cap at 9, default 3
+- `--session ID`: target specific session
+- `--style-skill PKG`: validate `.claude/skills/style-{PKG}/SKILL.md` exists
+- Missing/empty args without flags = error E001
+
+**Session Detection**:
+- Check `.workflow/scratch/brainstorm-*/` for existing sessions
+- Multiple → AskUserQuestion to select | Single → use it
+- None + auto mode → will create new session
+- None + single role mode → error E002
+
+**Output Directory Resolution**:
+- Phase mode (number): resolve `state.json.artifacts[phase == phaseNum].path` → `.workflow/{path}/.brainstorming/` (ERROR if phase not found)
+- All output: `.workflow/scratch/brainstorm-{slug}-{date}/`
+- Existing session: use existing session directory
+
+---
+
+### Step 1.5: Load Project Specs
+
+```
+specs_content = maestro spec load --category arch
+```
+
+Pass to role-design-author in Step 4 for architecture-aware role analysis.
+
+---
+
+### Auto Mode Steps (Phase 1.5 → Phase 1.7 → Phase 2 → Phase 3 → Phase 4)
+
+### Step 1.7: External Research — Design Routes (Auto Mode, Optional)
+
+Spawn `workflow-external-researcher` agent to discover design alternatives, architecture patterns, and competitive approaches for the brainstorm topic. This enriches the framework generation and role analyses with external knowledge.
+
+**Trigger**: Always in auto mode. Skip if `--skip-questions` and no tech keywords detected.
+
+**Auto-suggest when**: Topic contains technology keywords, architecture terms, or "design" / "pattern" / "alternative" in the description.
+
+```
+// Step 1.7.1: Spawn external researcher for design routes
+Agent(
+  subagent_type="workflow-external-researcher",
+  prompt="""
+<objective>
+Research design alternatives and architecture patterns for: {topic}
+Mode: Design Research
+</objective>
+
+<context>
+Project specs: {specs_content or "none"}
+Topic keywords: {extracted_keywords}
+</context>
+
+<task>
+Search for:
+1. Reference projects — how 2-3 similar projects/products solve this problem (architecture, key decisions, what worked)
+2. Extractable patterns — reusable design patterns distilled from those projects, with applicability notes
+3. Architecture approaches (at least 2-3 alternatives with trade-offs)
+4. UX/UI patterns if applicable (interaction models, layout strategies)
+5. Common design pitfalls and anti-patterns to avoid
+
+IMPORTANT: Output MUST include "Reference Projects / Implementations" and "Extractable Patterns" sections.
+Focus on design ROUTES — alternative approaches the brainstorm roles can evaluate.
+Be prescriptive where evidence is strong, present alternatives where trade-offs exist.
+Return structured markdown only — do NOT write files.
+</task>
+  """,
+  run_in_background=false
+)
+
+// Step 1.7.2: Store as designResearchContext (in-memory)
+designResearchContext = agent_output
+```
+
+`designResearchContext` is passed into:
+- Step 2 (Terminology): enriches domain term extraction
+- Step 3 Phase 1 (Topic Analysis): provides external design alternatives
+- Step 4 (Parallel Role Analysis): each role-design-author receives design research as additional context
+
+Also persisted to `{output_dir}/design-research.md` for future reference.
+
+If research fails (W005): `designResearchContext = null`, continue without external context.
+
+---
+
+### Step 1.8: Load Project Context (if `.workflow/` exists)
+
+Load existing project history to ground brainstorming in what's already been built:
+
+- From `.workflow/project.md`: `### Validated` → already_shipped, `### Active` → current_scope, `## Context` → project_history
+- From `.workflow/state.json.accumulated_context`: `deferred[]` → deferred_items, `key_decisions[]` → existing_constraints
+
+Pass `project_context` into Step 2 (terminology) and Step 3 (framework generation):
+- `already_shipped` informs what exists — brainstorm should explore extensions, not re-invent
+- `deferred_items` are high-value brainstorming seeds
+- `lessons_learned` surface pitfalls to avoid
+
+---
+
+### Step 2: Terminology & Boundary Definition (Auto Mode)
+
+Extract core terminology and define scope boundaries before framework generation.
+
+1. Analyze topic description and any project context (project.md, roadmap.md, project_context from Step 1.8)
+2. Extract 5-10 core domain terms:
+   - term (canonical), definition, aliases, category (core|technical|business)
+3. AskUserQuestion for Non-Goals (multiSelect=true):
+   - Generate 4-5 context-aware exclusion candidates based on topic
+   - Include "其他（请补充）" option for custom exclusions
+   - If user selects "其他", follow up with free-text question
+4. Store terminology table and non_goals to session state
+
+**Skip if**: `--yes` flag (use auto-generated terms, empty non-goals)
+
+### Step 3: Interactive Framework Generation (Auto Mode)
+
+Seven sub-phases producing guidance-specification.md:
+
+**Phase 0: Context Collection**
+- Read init outputs directly: `.workflow/project.md` (tech stack, requirements, decisions), `.workflow/state.json` (project state), `.workflow/specs/` (conventions)
+- If `.workflow/` does not exist: continue without project context
+
+**Phase 1: Topic Analysis**
+- Load Phase 0 context (tech_stack, modules, conflict_risk)
+- Deep topic analysis (entities, challenges, constraints, metrics)
+- Generate 2-4 context-aware probing questions via AskUserQuestion
+- Questions MUST reference topic keywords (no generic questions)
+- Store to `session.intent_context`
+
+**Phase 2: Role Selection**
+- Analyze Phase 1 keywords → recommend count+2 roles with rationale
+- AskUserQuestion (multiSelect=true) for user to select `count` roles
+- If `--yes`: auto-select recommended roles
+- Store to `session.selected_roles`
+
+**Phase 3: Role-Specific Questions**
+- FOR each selected role, generate 3-4 deep questions mapping role expertise to Phase 1 challenges
+- AskUserQuestion per role (sequential, one role at a time)
+- Questions must include: implementation depth, trade-offs, edge cases
+- Store to `session.role_decisions[role]`
+- If `--yes`: skip all role questions
+
+**Phase 4: Conflict Resolution**
+- Analyze Phase 3 answers for contradictions, missing integrations, implicit dependencies
+- Generate clarification questions referencing SPECIFIC Phase 3 choices
+- AskUserQuestion (max 4 per round)
+- If NO conflicts detected: skip with notification
+- Store to `session.cross_role_decisions`
+
+**Phase 4.5: Final Clarification + Feature Decomposition**
+- Ask: "是否有前面未澄清的重点需要补充？" (无需补充 / 需要补充)
+- If "需要补充": progressive questions until resolved
+- Extract candidate features from all Phase 1-4 decisions (max 8)
+- Each feature: F-{3-digit} ID, kebab-case slug, description, related roles, priority
+- Validate: independence, completeness, granularity balance, boundary clarity
+- AskUserQuestion for user to confirm or adjust feature list
+- Store to `session.feature_list`
+
+**Phase 5: Generate Specification**
+- Load all decisions + terminology + non_goals + feature_list
+- Transform Q&A to declarative statements (CONFIRMED/SELECTED)
+- Apply RFC 2119 keywords (MUST, SHOULD, MAY, MUST NOT, SHOULD NOT)
+- Write `guidance-specification.md` with sections:
+  1. Project Positioning & Goals
+  2. Concepts & Terminology (table)
+  3. Non-Goals (Out of Scope)
+  4-N. [Role] Decisions (with RFC 2119)
+  Cross-Role Integration
+  Risks & Constraints
+  §10. Feature Decomposition (table)
+  §11. Appendix: Decision Tracking
+  §12. Cross-Role Resolutions (initially empty — populated by Step 5)
+- Validate: no interrogative sentences, all decisions traceable, RFC keywords applied
+
+**Output**: `{output_dir}/guidance-specification.md`, session metadata (workflow-session.json)
+
+### Step 3.5: Visual Style Foundation (Auto Mode, conditional)
+
+When `ui-designer` is among the selected roles, establish the project's visual direction before role analysis begins. This ensures all downstream UX analysis works within a confirmed design system.
+
+**Condition**: Skip if `.workflow/impeccable/DESIGN.md` already exists (visual direction already locked).
+
+**Execution** (sequential):
+
+1. **Product context** — if `.workflow/impeccable/PRODUCT.md` missing:
+   ```
+   Skill({ skill: "maestro-impeccable", args: "teach" })
+   ```
+   This runs the teach interview to establish brand, users, personality, anti-references.
+
+2. **Visual exploration** — multi-style comparison and selection:
+   ```
+   Skill({ skill: "maestro-impeccable", args: "explore" })
+   ```
+   This generates multiple design system variants as HTML prototypes, launches visual comparison, and lets the user select/mix. Produces DESIGN.md on completion.
+
+3. Record in session metadata: `design_system_established: true`, `design_md_path: ".workflow/impeccable/DESIGN.md"`
+
+**`--yes` mode**: `explore` auto-selects variant 1 without visual comparison. `teach` still requires minimal input if PRODUCT.md is missing.
+
+**Skip mode**: If user explicitly passes `--skip-design` to brainstorm, skip this step entirely. ui-designer role will generate its own independent theme in Phase 2.
+
+### Step 4: Parallel Role Analysis (Auto Mode)
+
+For EACH selected role, spawn a `role-design-author` agent in parallel. Each agent produces a multi-file analysis under `{output_dir}/{role}/`.
+
+```
+Agent({
+  subagent_type: "role-design-author",
+  prompt: """
+    role_name: {role}
+    role_template_path: ~/.maestro/templates/planning-roles/{role}.md
+    guidance_path: {output_dir}/guidance-specification.md
+    output_dir: {output_dir}/{role}/
+    feature_list: <F-id + slug + title rows from guidance §10>
+    design_research: {output_dir}/design-research.md if exists, else null
+    project_specs: {specs_content or null}
+    user_context: {session.role_decisions[role] if any}
+    style_skill: {style_skill_path if role == ui-designer and provided}
+
+    Write the analysis files per the role template's "Brainstorming Analysis Structure".
+    Reference guidance decisions by ID (e.g., SA-03) — do NOT copy decision text.
+    All behavioural statements MUST use RFC 2119 keywords.
+  """,
+  run_in_background: false
+})
+```
+
+**Output Contract**: each agent writes files under `{output_dir}/{role}/`:
+- `analysis.md` — INDEX document containing §1 Role Mandate, §2 Decision Digest (4 tables), §3 Cross-Cutting Foundations, §4 File Index, §5 Outstanding TODOs
+- `analysis-F-{id}-{slug}.md` — one per feature (< 2000 words each)
+- `findings-{slug}.md` — additional discoveries (0 or more, < 1000 words each)
+
+See `.claude/agents/role-design-author.md` for the full output contract.
+
+**Digest-Based Validation** (orchestrator reads ONLY `analysis.md` per role, not sub-files):
+- Read `{role}/analysis.md` §2 Decision Digest:
+  - Decisions table: verify ≥ 1 row per feature in feature_list → feature coverage check
+  - RFC 2119 keywords in Constraints column → keyword count check
+- Read `{role}/analysis.md` §4 File Index:
+  - Verify listed files actually exist on disk (glob check)
+  - Verify feature coverage matches feature_list
+- system-architect specifically: verify §3 contains "Data Model" and "State Machine" headings
+- Warn if any analysis-F-*.md > 2000 words (from agent's return protocol `total_lines`)
+
+**Parallel Safety**: Each agent writes only to its own `{role}/` directory. No shared output, no cross-agent dependencies.
+
+### Step 4.5: Cross-Role Review (Auto Mode)
+
+After all role analyses are produced, spawn ONE `cross-role-reviewer` agent. The agent compares Decision Digests from each role's `analysis.md` — it does NOT read per-feature sub-files unless it needs deeper context.
+
+```
+Agent({
+  subagent_type: "cross-role-reviewer",
+  prompt: """
+    analysis_indexes:
+      - {output_dir}/{role_1}/analysis.md
+      - {output_dir}/{role_2}/analysis.md
+      - ...
+    guidance_path: {output_dir}/guidance-specification.md
+    feature_list: <F-id + slug + title rows from guidance §10>
+
+    Compare Decision Digests (§2) across these role analysis index files.
+    Identify conflicts, gaps, and synergy opportunities.
+    Use §4 File Index for accurate patch target file paths and headings.
+    Return the structured report — do NOT write files.
+  """,
+  run_in_background: false
+})
+```
+
+**Agent output** (parsed by orchestrator):
+- `conflicts[]` — C-IDs with role-A vs role-B contradictions from §2 Decisions/Positions comparison
+- `gaps[]` — G-IDs from §2 Interfaces consumer references without matching definitions
+- `synergies[]` — S-IDs from §2 Findings Summary with concrete patch suggestions
+- `need_deeper_context[]` — (optional) specific sub-files the reviewer needs to read for ambiguous cases
+
+**Deeper context handling**: If the reviewer returns `need_deeper_context` entries, the orchestrator reads those specific files and re-invokes the reviewer via SendMessage with the injected content. Expected to happen in < 10% of reviews.
+
+Store as `review_findings` in session memory. Skip Step 5 entirely if all three arrays are empty.
+
+If `--yes` flag set: auto-apply each finding's suggested resolution without AskUserQuestion (still proceed to Step 5 writeback).
+
+### Step 5: Apply Cross-Role Resolutions (Auto Mode)
+
+Consume `review_findings` from Step 4.5 and apply user-confirmed resolutions to the relevant role analysis files and `guidance-specification.md`.
+
+**Sub-phase 5.1: Interactive Confirmation (skip if `--yes`)**
+- For each finding (conflicts → gaps → synergies, in that order), call AskUserQuestion (max 4 per round):
+  - **Conflict question** options: "Accept suggested resolution" (recommended) / "Pick role A's stance" / "Pick role B's stance" / "Defer to TODO"
+  - **Gap question** options: "Accept suggested addition" (recommended) / "Skip this gap" / "Defer to TODO"
+  - **Synergy question** options: "Apply patch" (recommended) / "Skip"
+- Question text MUST cite the exact role folder + `target_heading` from the agent output (heading text, not section number).
+- Store user choices as `resolutions[]`.
+
+**Sub-phase 5.2: Apply Patches to Role Files**
+
+The reviewer's output includes a `patch_targets[]` block per finding with closed-vocabulary `edit_type` (`annotate_after_heading` / `annotate_and_strikeout` / `append_to_section`). The orchestrator consumes these directly — do NOT re-parse the prose.
+
+For each accepted finding, iterate over its `patch_targets[]`:
+
+**Conflict patches** (edit_type usually `annotate_and_strikeout`):
+- Locate `target_heading` in `target_file` (e.g., `system-architect/analysis-F-002-skill-engine.md`) via Edit string match (heading text MUST match verbatim).
+- Insert `edit_content` (a `> blockquote` Cross-Role Resolution line) immediately after the matched heading.
+- Wrap the original paragraph below the heading in `<!-- superseded by C-XXX -->` … `<!-- /superseded -->` HTML comments so the original text stays readable but downstream tooling treats it as non-authoritative.
+- If the user chose "Pick role A's stance" → strikeout role B's site only; vice versa. If "Accept suggested resolution" → strikeout BOTH role-A and role-B sites (both supersede to the new resolution).
+
+**Gap patches** (edit_type pair: `annotate_after_heading` at reference site + `append_to_section` at owner site):
+- At reference site: insert the breadcrumb annotation pointing to the owner site.
+- At owner site: append the actual definition content (1-3 lines from `edit_content`).
+- Both edits MUST succeed; if owner-site append fails, roll back reference-site annotation.
+
+**Synergy patches** (edit_type `annotate_after_heading` × 2):
+- Insert cross-reference annotations in BOTH role files. Synergy is symmetric — alignment must be visible from either entry point.
+- Original content untouched.
+
+**Edit failure handling**:
+- If `target_heading` does not match verbatim, log the finding ID and target, skip the patch, and surface to the user in Step 7 final report ("3 patches skipped — heading drift").
+- Never invent the heading; refusal to patch is safer than wrong patch.
+
+**Sub-phase 5.3: Append to guidance §12**
+Append a new subsection to `guidance-specification.md` §12 Cross-Role Resolutions capturing the audit trail:
+```markdown
+### Cross-Role Resolutions (added {date})
+| ID | Type | Source(s) | Resolution | Applied to |
+|---|---|---|---|---|
+| C-001 | conflict | system-architect/analysis-F-002-skill-engine.md "## Architecture" / sme/analysis-F-002-skill-engine.md "## Constraints" | {answer} | both files patched |
+| G-001 | gap | ux-expert/analysis.md "### Interfaces" | {answer} | data-architect/analysis.md "### Decisions" |
+| S-001 | synergy | ux-expert/analysis-F-005-tui.md "## Interface Contract" / sme/findings-ndjson.md "## Recommendation" | applied | both annotated |
+```
+
+**Sub-phase 5.4: Finalization**
+- Update session metadata (`review_findings_count`, `resolutions_applied`, `completion_status`).
+- Emit completion report (see Step 7).
+
+---
+
+### Single Role Mode Steps
+
+### Step 6: Single Role Analysis
+
+Execute analysis for ONE specified role with optional interactive context gathering.
+
+**Step 6.1: Detection & Validation**
+- Validate role_name against VALID_ROLES list
+- Detect session (--session or find existing)
+- Check for guidance-specification.md → framework_mode
+- Extract feature list from guidance §10 → feature_mode
+- Check existing `{role}/analysis.md` → update_mode (ask: update/regenerate/cancel)
+
+**Step 6.2: Interactive Context Gathering**
+- Skip if `--skip-questions`
+- Force if `--include-questions`
+- Generate 3-5 role-specific questions (Chinese, with business context)
+- AskUserQuestion per batch (max 4 per round)
+- Pass collected answers as `user_context` to the agent (in-memory; not persisted to a separate file)
+
+**Step 6.3: Agent Execution**
+- Spawn role-design-author with full Inputs block (see Step 4 schema)
+- Agent writes files under `{output_dir}/{role}/`
+
+**Step 6.4: Validation**
+- Read `{role}/analysis.md` — verify exists and non-empty
+- Check §2 Decision Digest tables present
+- Check §4 File Index matches files on disk
+- Check framework reference if framework_mode
+- Update session metadata with completion status
+- Report results with next step suggestions
+
+---
+
+### Review-Only Mode Steps
+
+### Step 6.5: Review-Only Mode
+
+Triggered by `--review-only --session ID`. Skips Phase 3 entirely; runs only Step 4.5 + Step 5 against the existing role analysis files.
+
+**Step 6.5.1: Validation**
+- Require `--session ID`. Error E002 if missing.
+- Glob `{output_dir}/*/analysis.md`. Error E006 if zero files found (no role analyses to review).
+- Require `guidance-specification.md` to exist (for decision-ID context). Error E007 if missing.
+
+**Step 6.5.2: Execute Step 4.5 + Step 5**
+- Reuse the exact same logic as auto mode Step 4.5 (spawn cross-role-reviewer with analysis indexes) and Step 5 (interactive confirmation + patch application + guidance §12 audit).
+- Skip Step 5.1 interactive confirmation if `--yes` also passed.
+
+**Step 6.5.3: Report**
+- Same as review-only mode report in Step 7.
+
+---
+
+### Step 7: Final Report
+
+**Auto mode report:**
+- Session ID and output directory
+- Roles analyzed (N) — list each `{role}/analysis.md` path
+- Files written per role (from agent return protocol)
+- Features in scope (N, from guidance §10)
+- Cross-role review: conflicts / gaps / synergies counts
+- Resolutions applied: count + a breakdown by type (C/G/S)
+- Next:
+  Skill({ skill: "maestro-roadmap", args: "--mode full --from brainstorm:{artifactId}" })  — Generate full spec package
+  Skill({ skill: "maestro-analyze", args: "{topic}" })   — Evaluate feasibility + lock decisions
+  Skill({ skill: "maestro-analyze", args: "{phase} -q" })   — Quick decision extraction only
+  Skill({ skill: "maestro-plan", args: "{phase}" })       — Plan directly (if scope is clear)
+
+**Single role mode report:**
+- Role analyzed
+- Framework alignment status (does analysis reference guidance decisions?)
+- Context questions answered (count)
+- Files written: list from §4 File Index
+- Next:
+  - Run more roles: `Skill({ skill: "maestro-brainstorm", args: "{another-role} --session {sessionId}" })`
+  - When 2+ roles are done, trigger review: `Skill({ skill: "maestro-brainstorm", args: "--review-only --session {sessionId}" })`
+
+**Review-only mode report:**
+- Session ID and existing roles found (N)
+- Cross-role review: conflicts / gaps / synergies counts
+- Resolutions applied: count + breakdown by type (C/G/S)
+- Patches skipped due to heading drift: count (if any)
+- Same Next-step routing as auto mode
+
+---
+
+### Step 7.5: Generate Context Package
+
+Write `{output_dir}/context-package.json` by extracting from session artifacts:
+
+```jsonc
+{
+  "$schema": "context-package/1.0",
+  "source": {
+    "type": "brainstorm",
+    "artifact_id": "{artifact_id}",
+    "session_path": "{output_dir relative to .workflow/}",
+    "generated_at": "{ISO-8601}"
+  },
+  "requirements": [],      // From guidance-specification.md §10 Feature Decomposition
+  "constraints": [],        // From guidance-specification.md §4-N MUST/MUST NOT + {role}/analysis.md §2 Decisions[locked]
+  "domain": {},             // From guidance-specification.md §1-3 (problem, terminology, audience)
+  "non_goals": [],          // From guidance-specification.md Non-Goals section
+  "insights": [],           // From {role}/analysis.md §3 Cross-Cutting subsections
+  "open_questions": [],     // From guidance-specification.md §4-N SHOULD/MAY items
+  "references": []          // List all key files: guidance-specification.md + {role}/analysis.md per role
+}
+```
+
+**Extraction mapping**:
+- `requirements[]`: each row from §10 Feature Decomposition → `{ id: "F-{id}", title, description, priority: "must|should|may" (from RFC 2119 keywords), acceptance, ref: "guidance-specification.md#§10" }`
+- `constraints[]`: each MUST/MUST NOT from §4-N role decisions → `{ id: "C-{NNN}", area, constraint, rationale, status: "locked", ref: "{source}#§{N}" }`; append locked decisions from each `{role}/analysis.md` §2 Decisions table where status=locked
+- `domain.problem_statement`: from §1-3
+- `domain.terminology[]`: from Concepts & Terminology section → `{ term, definition, ref: "guidance-specification.md#§5" }`
+- `domain.audience`, `domain.industry`: from §1-3
+- `non_goals[]`: each Non-Goal → `{ title, rationale, ref: "guidance-specification.md#§6" }`
+- `insights[]`: from each `{role}/analysis.md` §3 Cross-Cutting subsections → `{ role, area, summary, ref: "{role}/analysis.md#§3-{heading}" }`
+- `open_questions[]`: from §4-N SHOULD/MAY items → `{ area, question, options[], ref }`
+- `references[]`: `{ type: "guidance", path: "guidance-specification.md" }` + `{ type: "role-analysis", path: "{role}/analysis.md" }` per role
+
+Register artifact in state.json with additional field:
+  `context_package: "{output_dir}/context-package.json"`   (relative to .workflow/)
+
+---
+
+## Quality Criteria
+
+- If `designResearchContext` is set: guidance-specification.md references external design findings
+- guidance-specification.md uses RFC 2119 keywords (MUST/SHOULD/MAY)
+- Concepts & Terminology section with 5-10 core terms
+- Non-Goals section with rationale
+- Feature Decomposition table (max 8 features, independently implementable)
+- Each `{role}/analysis.md` contains §2 Decision Digest with all four tables (Decisions, Interfaces, Cross-Cutting Positions, Findings Summary)
+- Each `{role}/analysis.md` §4 File Index accurately lists all written files with headings
+- Each `{role}/analysis-F-*.md` < 2000 words, each `findings-*.md` < 1000 words
+- system-architect's `analysis.md` §3 includes: Data Model, State Machine, Error Handling, Observability
+- Cross-role review runs (Step 4.5) and produces structured findings from §2 Digest comparison with `patch_targets[]` blocks
+- **If review yielded findings**: each accepted resolution is annotated in the affected role analysis file AND logged in guidance §12 "Cross-Role Resolutions"
+- **If review yielded zero findings**: guidance §12 unchanged; final report explicitly notes "No cross-role issues detected"
+- Heading-drift patch failures (if any) are surfaced in the final report, not silently dropped