yellow-bank-soal/handoff.md

## HANDOFF CONTEXT GOAL

Continue implementation of IRT-Powered Adaptive Question Bank System after user configures GLM-5 model mapping for specific subagent categories. WORK COMPLETED

• Created comprehensive PRD (v1.1) from project-brief.md
• Resolved 10 critical clarification questions with client:
  1. Excel Import: Standardized across ALL tryouts
  2. AI Generation: 1 request = 1 question, admin playground for testing, no approval workflow
  3. Normalization: Optional manual/automatic control (system handles auto when sufficient data)
  4. Rollback: Preserve IRT history, apply CTT to new sessions only
  5. Admin Permissions: Not needed (WordPress handles per-site admins)
  6. Dashboards: FastAPI Admin only
  7. Rate Limiting: User-level reuse check + AI generation toggle
  8. Student UX: Admin sees internal metrics, students only see primary score
  9. Data Retention: Keep all data
  10. Reporting: All 4 report types required
• Created detailed technical implementation plan with 10 parallel subagents:
  • Deep Agent 1: Core API + CTT Scoring
  • Deep Agent 2: IRT Calibration Engine (recommended for GLM-5)
  • Deep Agent 3: CAT Selection Logic (recommended for GLM-5)
  • Deep Agent 4: AI Generation + OpenRouter (recommended for GLM-5)
  • Deep Agent 5: WordPress Integration
  • Deep Agent 6: Reporting System (recommended for GLM-5)
  • Unspecified-High Agents: Database Schema, Excel Import/Export, Admin Panel, Normalization CURRENT STATE

---

• PRD.md file created (746 lines, v1.1)
• project-brief.md exists (reference document)
• No code implementation started yet
• No git repository initialized
• Working directory: /Users/dwindown/Applications/tryout-system
• Session ID: ses_2f1bf9e3cffes96exBxyheOiYT PENDING TASKS

---

1. User configures GLM-5 model mapping for deep category (GLM-5 for algorithmic complexity)
2. User configures GLM-4.7 model mapping for unspecified-high category (general implementation)
3. Initialize git repository
4. Create project structure (app/, models/, routers/, services/, tests/)
5. Launch Unspecified-High Agent 1: Database Schema + ORM (BLOCKS all other agents)
6. After schema complete: Launch Deep Agents 1-3 in parallel (Core API, IRT Calibration, CAT Selection)
7. Launch Deep Agents 4-6 + Unspecified-High Agents 2-4 in parallel (AI Generation, WordPress, Reporting, Excel, Admin, Normalization)
8. Integration testing and validation KEY FILES

---

• PRD.md - Complete product requirements document (v1.1, 746 lines)
• project-brief.md - Original technical specification reference IMPORTANT DECISIONS

---

• 1 request = 1 question for AI generation (no batch)
• Admin playground for AI testing (no approval workflow for student tests)
• Normalization: Admin chooses manual/automatic; system handles auto when data sufficient
• Rollback: Keep IRT historical scores, apply CTT only to new sessions
• No admin permissions system (WordPress handles per-site admin access)
• FastAPI Admin only (no custom dashboards)
• Global AI generation toggle for cost control
• User-level question reuse check (prevent duplicate difficulty exposure)
• Admin sees internal metrics, students only see primary score
• Keep all data indefinitely
• All 4 report types required (Student, Item, Calibration, Tryout comparison) EXPLICIT CONSTRAINTS

---

• Excel format is standardized across ALL tryouts (strict parser)
• CTT formulas must match client Excel 100% (p = Σ Benar / Total Peserta)
• IRT 1PL Rasch model only (b parameter, no a/c initially)
• θ and b ∈ [-3, +3], NM and NN ∈ [0, 1000]
• Normalization target: Mean=500±5, SD=100±5
• Tech stack: FastAPI, PostgreSQL, SQLAlchemy, FastAPI Admin, OpenRouter (Qwen3 Coder 480B / Llama 3.3 70B)
• Deployment: aaPanel VPS with Python Manager
• No type error suppression (no as any, @ts-ignore)
• Zero disruption to existing operations (non-destructive, additive) GLM-5 MODEL ALLOCATION RECOMMENDATION

---

Use GLM-5 for:

• Deep Agent 2: IRT Calibration Engine (mathematical algorithms, sparse data handling)
• Deep Agent 3: CAT Selection Logic (adaptive algorithms, termination conditions)
• Deep Agent 4: AI Generation + OpenRouter (prompt engineering, robust parsing)
• Deep Agent 6: Reporting System (complex aggregation, multi-dimensional analysis) Use GLM-4.7 for:
• Deep Agent 1: Core API + CTT Scoring (straightforward formulas)
• Deep Agent 5: WordPress Integration (standard REST API)
• Unspecified-High Agents: Database Schema, Excel Import/Export, Admin Panel, Normalization (well-defined tasks) NOTE: Model mapping is controlled by category configuration in system, not by direct model specification in task() function. CONTEXT FOR CONTINUATION

---

• User is currently configuring GLM-5 model mapping for specific categories
• After model mapping is configured, implementation should start with Database Schema (Unspecified-High Agent 1) as it blocks all other work
• Parallel execution strategy: Never run sequential when parallel is possible - all independent work units run simultaneously
• Use task(category="...", load_skills=[], run_in_background=true) pattern for parallel delegation
• All delegated work must include: TASK, EXPECTED OUTCOME, REQUIRED TOOLS, MUST DO, MUST NOT DO, CONTEXT (6-section prompt structure)
• Verify results after delegation: DOES IT WORK? DOES IT FOLLOW PATTERNS? EXPECTED RESULT ACHIEVED?
• Run lsp_diagnostics on changed files before marking tasks complete
• This is NOT a git repository yet - will need to initialize before any version control operations

---