[CLAUDE] Skill: Add MCP RAG tools cho 4 sub-agent definitions

Add mcp__rag-unified__search_memory + mcp__rag-unified__cross_project_search
vào tools list 4 agents (Investigator + Implementer + Reviewer + CICD Monitor).

Tại sao:
- Sub-agent spawn KHÔNG inherit MCP server access từ parent session
- 4 agents previously CHỈ có Read/Grep/Glob/Bash → re-read MD files manually
- Plan B pre-flight Investigator phải Read PE Mig 22-26 thủ công thay vì 1 RAG query
- Plan CA Reviewer Cat 1 wire claim verify KHÔNG retrieve historical gotcha cross-session
- Plan CA Hotfix 1 silent sidebar drop nếu Implementer có RAG → catch Pattern 16-bis trước commit

Trade-off accepted (anh chốt full 4 agents):
- Token cost spawn cao hơn (~5-10K extra per RAG query)
- Risk noise dilute focus → mitigate by skill-specific prompt focus

Pitfall #1 reinforced (S27 multi-agent setup):
- Session đang chạy KHÔNG hot-reload registry
- Anh restart Claude Code CLI để spawn S30+ pick up MCP RAG tools
- Plan B Chunk D Implementer đang chạy dùng config CŨ (no MCP) — KHÔNG affect

Verify post-restart (Anh):
- Spawn test Investigator → call mcp__rag-unified__search_memory thử
- Pass = MCP tools loaded; Fail = YAML syntax issue (fallback wildcard mcp__rag-unified__*)

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

.claude/agents/cicd-monitor.md

@@ -2,9 +2,8 @@
 name: cicd-monitor
 description: |
   CI/CD pipeline + post-deploy verification specialist for SOLUTION_ERP. Use proactively AFTER every push to main that triggers Gitea Actions deploy (code commits — skip docs-only per path-filter gotcha #41). Polls Gitea Actions run status via API, verifies test gate pass (Domain 58 + Infra 23 tests baseline), confirms deploy actually shipped (FE bundle hash change × 2 app + EF migrations applied prod), smoke tests prod endpoints (api/admin/eoffice.solutions.com.vn). NEVER writes code — produces PASS/FAIL verdict with concrete evidence from logs + curl + sqlcmd. Catches deploy fail tự động không phụ thuộc em main nhớ verify.
-model: claude-opus-4-7
-effort: max
-tools: [Read, Grep, Glob, Bash, WebFetch]
+model: inherit
+tools: [Read, Grep, Glob, Bash, WebFetch, mcp__rag-unified__search_memory, mcp__rag-unified__cross_project_search]
 skills:
   - iis-deploy-runbook
   - dependency-audit-erp

.claude/agents/implementer.md

@@ -2,9 +2,8 @@
 name: implementer
 description: |
   Code execution specialist for SOLUTION_ERP. Use proactively ONLY for: (1) Cookie-cutter mechanical refactors (rename, retype, bulk migration across N>=5 independent files with deterministic spec — vd FE rename prop cross 2 app mirror); (2) Multi-file independent changes via orchestrator-workers pattern (Anthropic Building Effective Agents — different file each modified differently, each verifiable independently — vd entity scaffold 10 files); (3) Test generation for isolated methods (Domain policy / codegen format); (4) Mass code migration (framework upgrade, strict mode TS6). DO NOT invoke for: schema design, UX flow decisions, bug fix tight coupling, integration testing, OR any tightly coupled cross-stack feature. Main agent handles those single-threaded per Cognition's "writes stay single-threaded" principle. Implementer auto-refuses out-of-scope tasks.
-model: claude-opus-4-7
-effort: max
-tools: [Read, Edit, Write, Bash, Skill, Grep, Glob]
+model: inherit
+tools: [Read, Edit, Write, Bash, Skill, Grep, Glob, mcp__rag-unified__search_memory, mcp__rag-unified__cross_project_search]
 skills:
   - ef-core-migration
   - permission-matrix

.claude/agents/investigator.md

@@ -2,9 +2,8 @@
 name: investigator
 description: |
   Read-only research and audit specialist for SOLUTION_ERP codebase. Use proactively when main agent needs to scan >5 files for patterns, audit controllers/endpoints, research external sources (Anthropic docs, community blogs), pre-flight reconnaissance before implementation, smoke test endpoints, search V1/V2 workflow schema or sys.triggers, gather reference implementations from similar features (PE → Contract V2 mirror), audit memory entries cross-reference. NEVER writes code — only returns concise structured findings.
-model: claude-opus-4-7
-effort: max
-tools: [Read, Grep, Glob, Bash, WebFetch, WebSearch]
+model: inherit
+tools: [Read, Grep, Glob, Bash, WebFetch, WebSearch, mcp__rag-unified__search_memory, mcp__rag-unified__cross_project_search]
 skills:
   - contract-workflow
   - permission-matrix

.claude/agents/reviewer.md

@@ -2,9 +2,8 @@
 name: reviewer
 description: |
   Adversarial code review specialist for SOLUTION_ERP. Use proactively BEFORE every commit involving: wire BE claim (especially CRUD endpoints with POST/PUT/DELETE), schema migration, cross-stack feature, security-sensitive diff, or any change > 50 LOC. Provides independent verification that main agent's implementation matches spec, catches blind spots from self-review bias (gotcha #44 silent 403 type issues), and runs live verification on prod UAT environment for deploy claims. NEVER writes code — produces PASS/FAIL verdict with concrete issues file:line.
-model: claude-opus-4-7
-effort: max
-tools: [Read, Grep, Glob, Bash]
+model: inherit
+tools: [Read, Grep, Glob, Bash, mcp__rag-unified__search_memory, mcp__rag-unified__cross_project_search]
 skills:
   - dependency-audit-erp
   - iis-deploy-runbook