[CLAUDE] Docs: Session 26 chốt cuối — 6 Plan AG series PE tree view + Plan AI RAG global MCP setup

Update:
- docs/STATUS.md: Last updated S26 cumulative wrap
- docs/HANDOFF.md: TL;DR S26 chốt cuối với 3 pattern reusable NEW
- docs/changelog/sessions/2026-05-21-s26-pe-tree-view-rag-setup.md: NEW session log đầy đủ
- docs/guides/multi-agent-setup-guide.md: NEW ~750 lines onboarding 4 dự án future
- .claude/agent-memory/*/MEMORY.md: 4 agent flush S26 entries
- .claude/rag.json: NEW project config cho RAG bootstrap

Plans done S26:
- Plan AG/AG2/AG3/AG4/AG5/AG6 — 6 commits 0bf6c7e..d99069a PE List tree view UI iteration
- Plan AI Phase 0-4 — RAG global MCP setup (Voyage-4-large + Qdrant Windows native binary v1.18.0 NO Docker + FastMCP 3.3.1 stdio + SQLite FTS5 BM25 + RRF k=60 + Anthropic Contextual Retrieval prepend)
- SOLUTION_ERP bootstrap: 126 files → 2,392 chunks indexed 60.9s (~484K Voyage tokens = 0.24% free tier 200M/month)

Multi-agent ROI S26: 5 spawn (Inv 2 audit 5Q + RAG distribution research 4 study cases + Imp 1 Case 2 + Rev 1 pre-commit + CICD 1 Run #222) ~123K + em main solo Plan AG2-AG6 polish + Plan AI Phase 0-4 ~280K = ~28% solo equivalent.

3 patterns reusable cross-project NEW S26:
1. Pattern 19 Implementer — HTML native <details>/<summary> + Tailwind named groups (group/proj+year+sup) + localStorage Set<string> cho hierarchical 3-level tree UI when no Accordion lib
2. RAG User-level Global MCP — 1 server localhost serve N project + per-project .claude/rag.json (Approach A — 1 dev solo scenario, không phải team VPS)
3. Qdrant Windows native binary deployment — no Docker overhead, qdrant-x86_64-pc-windows-msvc.zip 28.3MB chính thức GitHub release

Pending S27+:
- Memory CURATE 4 agent (cicd-monitor 74KB OVER 50KB hard threshold URGENT)
- Plan AI Phase 5 bootstrap 4 project còn lại (NamGroup/DH Y Dược/Ashico/Vipix)
- Plan AI Phase 6 file watcher + Windows Task Scheduler

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

.claude/agent-memory/investigator/MEMORY.md

@@ -153,3 +153,7 @@ Last curate: 2026-05-11 (initial seed)
 - **2026-05-19 (S25 t1 spawn audit 2 bug critical UAT — Changelog logging missing):** Em main report 2 bug UAT: (1) Budget Adjust 2×click → "Lịch sử thay đổi" KHÔNG show 2 entry. (2) Return Assignee mode F1 click → "Lịch sử thay đổi" KHÔNG show return action. Audit 5Q: Q1-DB state Prod SSH fail (auth), fallback code inspect. Q2-Budget handler: `PurchaseEvaluationFeatures.cs:379-387` **ĐÃ log changelog** EntityType=Header + Action=Update + summary diff → **code ĐÚNG, likelihood FE filter bug or UAT DB stale**. Q3-Return handler: `PurchaseEvaluationWorkflowService.cs:215-378 ApplyReturnModeAsync` **ZERO changelog log** — 4 mode branches (OneLevel/OneStep/Assignee/Drafter) mutate pointers only, return summary, caller `TransitionAsync:100 LogTransitionAsync` only logs phase transition KHÔNG log mode side-effect. Q4-Schema: `PurchaseEvaluationChangelog.cs` support EntityType=Workflow(5) enum but **ZERO code populate** — schema incomplete missing Kind/ChangeType subtype enum to disambiguate "budget adjust" vs "detail edit" vs "return mode" (all Header/Workflow entity cùng Action.Update). Q5-FE: Query handler `ListPurchaseEvaluationChangelogsQueryHandler:1050-1064` **KHÔNG filter logic** — returns ALL entities. FE query component unknown (fe-user source not easily searchable) but **likely filter EntityType == (Supplier|Detail|Quote)** omit Header+Workflow → Bug 1 Budget (Header type) hidden, Bug 2 Return (Workflow type unlogged vậy). Root cause B1: FE filter skip Header updates OR Schema gap (Header.Update collision: budget vs section 2 edit). Root cause B2: Handler **intentionally skip logging** (no companion audit table like PurchaseEvaluationApprovals for return history). Fix path B1: Option A (add Kind enum — 30 LOC schema+handlers) vs B (extend EntityType — 20 LOC) vs C (FE filter conditional show Budget — 10 LOC FE). Fix path B2: Add `db.PurchaseEvaluationChangelogs.Add()` per mode (15-25 LOC 1 file). Cross-ref pattern memory: S23 t3 lookup bug (Level.Order collision per-NV Mig 29) — schema Mig 29 OR-of-N refactor created similar subtype ambiguity. Surprise: `EntityType.Workflow=5` enum value design-only, unused 4+ mig history. Recommendation: B2 fix first (clear win, 1 file BE), then B1 design (schema + cross-module pattern audit Contracts). Token ~28k.
 - **2026-05-19 (S25 wrap — Plan AB Bug 1+2 audit + 6 follow-up plans em main solo):** Pre-Plan AB audit ~28K confirm root cause: Bug 1 Budget Adjust Handler ĐÃ log Changelog (Header+Update) nhưng FE HistoryTab filter strict TraLai-only loại. Bug 2 ApplyReturnModeAsync 4 mode KHÔNG add Changelog.Add() — chỉ caller LogTransitionAsync log phase transition. Recommended fix path: BE add log return mode + FE filter relax + Decision badge differentiation. **Subsequent em main solo follow-ups** (Plan AC..AF, KHÔNG re-spawn Investigator): AC capture pre-call Step/Level state + skipToFinal comment enrich + FE Decision badge; AC2 FE merge synthetic Reject rows từ Changelog (reversible recovery KHÔNG DB touch); AD drop misleading dual-phase badges + parse semantic next-target hint via regex `Chuyển phase X → Y` + comment keyword "Trả về"/"vượt cấp"; AE BE preventive batch UserName fix 9 Changelog.Add sites (Drafter+Detail+Quote+Adjust); AF FE userMap fallback từ embedded domain data PeDetailBundle (drafter+approvals+approvalFlow+levelOpinions+departmentOpinions) — no extra API fetch. **Patterns reusable Contract V2 audit recovery + Budget changelog UI:** (a) synthetic recovery FE merge từ existing audit table, (b) userMap fallback từ embedded domain data, (c) drop misleading badges + semantic hint via regex parsing, (d) preventive batch fix systemic gap. Cross-ref `feedback_service_hook_vs_endpoint` (audit log derived state) + new pattern user-level: FE merge synthetic + userMap fallback. **Tokens cumulative em main this session ~340K (mostly solo iterations, sub-agent leverage ~62K total).** **CICD result:** 7 commits `e23f51c..506cada` push remote, runs #215 FAIL (Plan M tests SQLite tie-break re-emerge) → #216 PASS (Chunk A2 SQLite tie-break fix) → #217-#221 PASS streak. Gotcha #48 SQLite tie-break pending docs.
 
+- **2026-05-21 (S26 spawn Plan AG 5Q audit — PE List tree view feedback Tra Sol):** Bro UAT 2026-05-21 screenshot phàn nàn UI Duyệt NCC flat list "đám rừng" + propose Outlook-style folder template. Em main spawn Investigator audit 5Q: Q1 prod data scale (SSH auth fail, fallback code inspect — ~50-200 projects × 5-15 PE typical). Q2 Project entity `Master/Project.cs:5-14` extends AuditableEntity với Code+Name+dates+Budget+PM+Note, **MISSING MaxLength + navigation tới PE** → ready for Phase 2 ProjectPackage table mở rộng. Q3 PE List structure `PurchaseEvaluationsListPage.tsx:133` 3-panel grid `lg:grid-cols-[340px_1fr_360px]` flat `<ul><li>` line 199-252 + pageSize 50 + filter type/phase/awId/search. Q4 shadcn fe-user component GAP **THIẾU Accordion/Collapsible/Tree/Card/Badge** (chỉ Button/Dialog/Input/Label/Select/Textarea) → Phase 1 fallback HTML `<details>/<summary>` native + inline badge `<div>` (verified S24 Plan AA pattern). Q5 Mig 32 prep naming `AddProjectPackageTable` follow cumulative pattern. **Recommend Approach C Hybrid: Phase 1 FE-only group view ~160 LOC × 2 app (1-2 ngày) + Phase 2 ProjectPackage schema (3-5 ngày defer post-UAT). Implementer Case 2 cookie-cutter mirror 2 app ACCEPT.** Token ~30k.
+
+- **2026-05-21 (S26 spawn Plan AI RAG distribution research — 4 study cases industry-validated):** Bro plan setup RAG cho 5 dự án cùng máy localhost share infrastructure. Investigator deep research 7Q: Anthropic patterns + community tools + multi-tenant architecture + distribution mechanism + auth + cost + sync. **Findings 4 study cases:** (1) **Cursor** — Merkle tree + Turbopuffer + 92% similarity reuse pattern, secret key derived Git commit hash, per-team index sharing cut indexing hours→seconds. (2) **Cline Memory Bank** — markdown + JIT retrieval + git-native sync. (3) **Continue.dev Hub** — slug-based YAML config sharing centralized. (4) **Sourcegraph Cody** — interesting anti-pattern, BỎ embeddings sang Search API + graph IR scale 100K+ repos enterprise (alternative if good search infra sẵn). **Multi-tenant 3 patterns:** Single index (pool, cheap), Per-tenant isolated (silo, expensive), Hybrid base+delta (Cursor pattern). **Distribution combo winner cho 5-dev team:** FastMCP HTTP server VPS Hetzner $15/tháng + git-sync embeddings snapshots committed to private repo. Auth JWT RS256 + role scopes + document-level ACL. Cost reality 5 dev ~$20/month vs cloud Qdrant $50/month. Sync git-based weekly cron (real-time chỉ cần khi team > 10 dev). **Surprise:** Voyage AI **200M tokens/month free tier cover 5 devs** comfortably — coi như chỉ tốn $15 VPS. **Anthropic standards 9/9 matched** + community best practices 6/7 matched (Sourcegraph alternative N/A cho memory-heavy). Plan AI architecture chốt với em main: User-level Global MCP (Pattern C — 1 server localhost serve 5 project) thay vì team VPS pattern (anh single dev 5 dự án). Token ~40k.
+