[CLAUDE] Docs: chốt Session 21 cuối (turn 1-5) — gotcha #46 + 2 memory mới + 4 agent MEMORY flush
Session 21 5-turn timeline chốt cuối (2026-05-12 0030 → 2026-05-13 1530):
| Turn | Topic | Commits |
|---|---|---|
| t1 | Add cicd-monitor sub-agent (4th, Path A) | 2 |
| t2 | RAG Hybrid setup planning Cách A | 2 |
| t3 | Fix gotcha #45 PE button "Trả lại" mismatch | 3 |
| t4 | F1+F2+F3 PE Workflow Mig 28 workflow-level | 5 |
| t5 | Refactor Allow* sang per-NV Mig 29 | 4 |
Cumulative 12 commits pushed remote `3a34831..c0af9e0`. No pending push.
**Gotcha mới #46** (`docs/gotchas.md`):
- Gitea Actions API path `/actions/tasks` not `/actions/runs` (Gitea v1 vs
GitHub naming khác)
- Cache `updated_at` stale ~2 min → cross-check VPS file LastWriteTime
- Discovery từ CICD Monitor Run #186 (S21 t4) + #187 (S21 t5)
- Saved Bash command preset cho future CICD spawn
**2 Memory user-level mới** (`C:\Users\pqhuy\.claude\projects\D--Dropbox-CONG-VIEC-SOLUTION\memory\`):
1. `feedback_ef_migration_backfill_reorder.md` — Cross-project pattern:
- EF auto-generated drop-then-add WRONG cho data preservation
- Manual reorder ADD → BACKFILL SQL via migrationBuilder.Sql() → DROP
- Anti-patterns: trust EF order, backfill separate migration, C# foreach
- Down() rollback chấp nhận data loss
- Bài học S21 t5 SOLUTION_ERP Mig 29 (48/48 Levels + 0/13 Users backfill OK)
2. `feedback_per_nv_permission_scope.md` — Cross-project pattern:
- Multi-role workflow flag KHÔNG gắn parent table cho "tiện"
- Split scope theo role context: Approver → Level table, Drafter → User table
- Decision tree: role context → entity natural carry
- UX implication: per-Level inline checkbox + User Mgmt per-user toggle
- Bài học S21 t4 (Mig 28 SAI scope) → S21 t5 (Mig 29 ĐÚNG per-NV)
- Trigger: user feedback "cấu hình cho từng người chứ ko phải toàn bộ"
**4 agent MEMORY.md flush:**
- 🟦 Investigator: seeds-only S21 t3-t5 (em main solo cross-stack reasoning chain)
- 🟨 Implementer: REFUSE 3× per criteria #3+#4 (correct — Anthropic warning match)
- 🟥 Reviewer: seeds-only (em main self-review build+test + CICD post-deploy)
- 🟩 CICD Monitor: 2 runs PASS (#186 + #187, ~110-120K cost each, all 5-stage green)
**Plan G Trial Week 1 evidence:**
- CICD Monitor: 2/2 PASS green = 0 fail catch (deploy clean)
- Cost: ~110-120K per spawn, under 150K budget
- CI baseline: 3-3.5 min stable
- Bonus discoveries saved: Gitea API path + prod credential fallback
- Other 3 agents: seeds-only ROI track pending future spawn opportunity
**STATUS + HANDOFF updates:**
- STATUS: Last updated S21 chốt + count 45→46 gotcha + 17→19 memory
- HANDOFF: Insert section "Session 21 chốt cuối — 5 turn timeline" trên cùng:
- Turn-by-turn table với commits + CICD verify
- Major schema evolution Mig 28 → Mig 29 (workflow-level → per-NV)
- 2 pattern reusable saved memory
- Plan G Trial Week 1 evidence table
- Pending S22+ tree (Plan C test bundle / F2 UI / Plan B Contract V2 / etc)
- Audit cron 2026-06-01 unchanged (threshold KHÔNG đạt sớm)
**MEMORY index user-level +2 entry** (memory MEMORY.md).
State final S21:
- 29 mig · 59 tables · ~143 endpoints · 34 FE pages
- 84 test pass (58 Domain + 26 Infra)
- 46 gotcha (+2 từ baseline 44 sau S20: #45 + #46)
- 19 memory entries (+3 từ baseline 16 sau S20: RAG + EF backfill + per-NV scope)
- 6 skills unchanged
- 4 sub-agents (3 seeds-only + 1 cicd-monitor 2-run PASS)
Pending: bro UAT continue. Plan C test-after bundle defer sau UAT 2-3 lần ổn.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
3d725c42f7