---
name: feedback-7agent-split-upgrade
description: SOLUTION_ERP S39 multi-agent 4→7 upgrade — BVAAU pattern adapted (backend/frontend split + test-specialist) + budget +50% + 5 RAG MCP + anti-truncation rules
metadata: 
  node_type: memory
  type: project
  tags: 
    - decision
    - multi-agent
    - governance
    - infra
  phase: phase-10-to-11
  originSessionId: 2a3ce6e1-edc5-464d-a436-38c8ed876dde
---

# Multi-agent 4→7 Upgrade (Session 39 — 2026-05-29)

**Trigger:** Anh main "tham khảo cấu hình 7 sub-agent BVAAU → tăng tốc + chất lượng → budget +50% + thêm 3 agent". Chốt qua AskUserQuestion 2 câu.

## Decision chốt (AskUserQuestion S39)

1. **Implementer split = Backend / Frontend** (KHÔNG domain/infra như BVAAU)
   - Lý do: SOLUTION_ERP có 2 React FE app riêng (admin + user) + FE mirror SHA256 là pain point chính (Pattern 16-bis 9×). Tách FE specialist → giảm gotcha #53 truncation khi BE+FE cùng spawn (đã gặp 5×). Mỗi spawn nhẹ hơn.
2. **test-specialist dedicated** = full test layer (2 implementer KHÔNG viết test)

## 7-agent final (color-coded)

| Agent | Color | Scope |
|---|---|---|
| investigator-codebase | cyan | INTERNAL SQL/EF/grep/reference mirror |
| investigator-api | blue | EXTERNAL docs/CVE/lib eval/cross-project |
| implementer-backend | yellow | .NET Domain+App+Infra+Api |
| implementer-frontend | orange | FE 2 app cookie-cutter SHA256 mirror |
| test-specialist | purple | tests/ dedicated xUnit |
| reviewer | red | adversarial pre-commit + live curl |
| cicd-monitor | green | post-deploy Gitea + bundle + smoke |

## Adapt KHÁC BVAAU (anh cho phép "không bắt buộc giống")

- **backend/frontend** thay domain/infra (2-FE-app fit)
- **Giữ 6 skill proven** (BVAAU Phase 0 = 0 skill, codebase rỗng). Investigator extract verdict: BVAAU config = **aspirational template chưa battle-test** (VIPIX guide reference không tồn tại) → SOLUTION_ERP port trục split CÓ CHỌN LỌC, giữ lợi thế 38-session battle-test.

## Port TỪ BVAAU

- Trục split research(2) + implement(2) + quality(3)
- **+3 RAG MCP** all 7 agent: `search_code` (BM25, prefer over Read full file — tiết kiệm token) + `store_memory` + `list_projects` (2→5 tools). MCP server verified support.

## Token budget +50%

- Prior: ~700K opt / ~1.35M heavy → **Now: ~1.05M opt / ~2M heavy ceiling**
- Justify: parallel BE+FE+test giảm wall-clock + split nhẹ giảm truncate + test-specialist clear coverage gap backlog
- Max 20× plan absorbs ~5× comfortable

## Anti-truncation rules (gotcha #53/#54/#55 codified README)

- Brief WRITE agent ≤ 8K (heavy spec ~10K → truncate risk #55)
- Curate MEMORY pre-spawn nếu > 25KB (#53)
- Agent entry ≤ 1.5K chars (frontmatter rule mỗi agent)
- 529 Overload (#54) → em main solo fallback, KHÔNG retry loop
- Pre-supply reference snippet trong brief nếu cần > 4 reference file

## How to apply (cross-session)

- **CLI restart BẮT BUỘC** sau đổi agent definition (per S27 model:inherit lesson) → registry hot-reload. Verify smoke spawn 3 agent mới.
- Memory folder rename: investigator→investigator-codebase + implementer→implementer-backend (giữ history). Khi SendMessage flush agent CŨ trước restart → ghi nhầm folder rename → conflict. Wrap session đổi config: em main proxy synthesize, KHÔNG SendMessage flush.
- Decision tree split-routing: Read internal→codebase, Read external→api, WRITE .NET→backend, WRITE FE→frontend, test→test-specialist.

**Reusable pattern cross-project:** Khi scale multi-agent, split theo TRỤC WORKLOAD THỰC TẾ của project (SOLUTION_ERP = 2 FE app → backend/frontend) thay vì copy nguyên pattern reference project. Reference project có thể aspirational chưa battle-test → verify trước khi port.

Related: [[feedback-implementer-truncation-mitigation]] [[feedback-multi-agent-setup]]