kl3inIT · kl3inIT · May 31, 2026 · May 26, 2026 · May 26, 2026 · May 26, 2026
diff --git a/.ci-backend-gradle.log b/.ci-backend-gradle.log
diff --git a/.ci-frontend-web.log b/.ci-frontend-web.log
diff --git a/.ci-playwright.log b/.ci-playwright.log
diff --git a/.claude/skills/react-doctor/SKILL.md b/.claude/skills/react-doctor/SKILL.md
@@ -0,0 +1,46 @@
+---
+name: react-doctor
+description: Use when finishing a feature, fixing a bug, before committing React code, or when the user types `/doctor`, asks to scan, triage, or clean up React diagnostics. Covers lint, accessibility, bundle size, architecture. Includes a regression check and a full local-triage workflow that fetches the canonical playbook.
+version: "1.1.0"
+---
+
+# React Doctor
+
+Scans React codebases for security, performance, correctness, and architecture issues. Outputs a 0–100 health score.
+
+## After making React code changes:
+
+Run `npx react-doctor@latest --verbose --diff` and check the score did not regress.
+
+If the score dropped, fix the regressions before committing.
+
+## For general cleanup or code improvement:
+
+Run `npx react-doctor@latest --verbose` (without `--diff`) to scan the full codebase. Fix issues by severity — errors first, then warnings.
+
+## /doctor — full local triage workflow
+
+When the user types `/doctor`, says "run react doctor", or asks for a full triage / cleanup pass (not just a regression check), fetch the canonical local-triage playbook and follow every step in it:
+
+```bash
+curl --fail --silent --show-error \
+  --header 'Cache-Control: no-cache' \
+  https://www.react.doctor/prompts/react-doctor-agent.md
+```
+
+The playbook is the single source of truth — a scan → filter → triage → fix → validate loop that edits the working tree directly (never commits, never opens PRs). Updating the prompt at its source updates every agent on its next fetch — no skill reinstall needed.
+
+Pair it with the matching per-rule prompts at `https://www.react.doctor/prompts/rules/<plugin>/<rule>.md` (fetched on demand inside the playbook) so each fix uses the canonical, reviewer-tested recipe.
+
+## Command
+
+```bash
+npx react-doctor@latest --verbose --diff
+```
+
+| Flag        | Purpose                                       |
+| ----------- | --------------------------------------------- |
+| `.`         | Scan current directory                        |
+| `--verbose` | Show affected files and line numbers per rule |
+| `--diff`    | Only scan changed files vs base branch        |
+| `--score`   | Output only the numeric score                 |
diff --git a/.github/workflows/react-doctor.yml b/.github/workflows/react-doctor.yml
@@ -0,0 +1,21 @@
+name: React Doctor
+
+on:
+  pull_request:
+    types: [opened, synchronize, reopened, ready_for_review]
+
+permissions:
+  contents: read
+  pull-requests: write
+  issues: write
+
+concurrency:
+  group: react-doctor-${{ github.event.pull_request.number || github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  react-doctor:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v5
+      - uses: millionco/react-doctor@b612664043a9be414166e3c6a69b355e39a8dcf4 # main as of 2026-05-31
diff --git a/.planning/REQUIREMENTS.md b/.planning/REQUIREMENTS.md
@@ -70,35 +70,35 @@
 
 ### Settings Page — Personalization (carried from v1.1)
 
-- [ ] **SET-VOICE-01**: User can edit free-text writing style description (200–500 words) that influences AI draft tone
-- [ ] **SET-VOICE-02**: User can edit free-text personal instructions ("About me") that gets injected into the system prompt for chat/triage/draft (XML-fenced + sanitized for prompt-injection sentinels + length cap 2000 chars)
-- [ ] **SET-VOICE-03**: User can edit free-text email signature appended to AI drafts
-- [ ] **SET-VOICE-04**: User can manage a list of titled knowledge-base snippets the AI consults when drafting
-- [ ] **SET-VOICE-05**: User can pick a tone preset (professional / friendly / casual / formal / custom) as a quick baseline
-- [ ] **SET-VOICE-06**: User can pick AI output language (VI / EN, default VI) — separate from UI language
-- [ ] **SET-VOICE-07**: User can trigger a "Generate from recent sent emails" action inside the writing-style edit dialog. The action fetches the most recent N sent emails (default N=20, max 50), feeds them transiently to the LLM along with a style-extraction prompt, and returns a concise style guide (≤500 words) that pre-populates the writing-style textarea for the user to review and edit before saving. Privacy invariant: raw email bodies and the LLM prompt/completion exchange MUST be in-memory-only (no DB, no log file, no audit row); only the user-reviewed-and-saved style guide is persisted (into the existing `assistant_settings.writing_style` column). Pulled into v1.2 Phase 9 scope from `SET-VOICE-FUT-03` on 2026-05-26 during discuss-phase.
+- [x] **SET-VOICE-01**: User can edit free-text writing style description (200–500 words) that influences AI draft tone
+- [x] **SET-VOICE-02**: User can edit free-text personal instructions ("About me") that gets injected into the system prompt for chat/triage/draft (XML-fenced + sanitized for prompt-injection sentinels + length cap 2000 chars)
+- [x] **SET-VOICE-03**: User can edit free-text email signature appended to AI drafts
+- [x] **SET-VOICE-04**: User can manage a list of titled knowledge-base snippets the AI consults when drafting
+- [x] **SET-VOICE-05**: User can pick a tone preset (professional / friendly / casual / formal / custom) as a quick baseline
+- [x] **SET-VOICE-06**: User can pick AI output language (VI / EN, default VI) — separate from UI language
+- [x] **SET-VOICE-07**: User can trigger a "Generate from recent sent emails" action inside the writing-style edit dialog. The action fetches the most recent N sent emails (default N=20, max 50), feeds them transiently to the LLM along with a style-extraction prompt, and returns a concise style guide (≤500 words) that pre-populates the writing-style textarea for the user to review and edit before saving. Privacy invariant: raw email bodies and the LLM prompt/completion exchange MUST be in-memory-only (no DB, no log file, no audit row); only the user-reviewed-and-saved style guide is persisted (into the existing `assistant_settings.writing_style` column). Pulled into v1.2 Phase 9 scope from `SET-VOICE-FUT-03` on 2026-05-26 during discuss-phase.
 
 ### Settings Page — Behavior Toggles (carried from v1.1)
 
-- [ ] **SET-BEHV-01**: User can toggle auto-draft replies (master switch for v1.0 DRFT-01..04 background drafts)
-- [ ] **SET-BEHV-02**: User can set a draft confidence threshold (0.0–1.0); AI only saves drafts ≥ threshold
+- [x] **SET-BEHV-01**: User can toggle auto-draft replies (master switch for v1.0 DRFT-01..04 background drafts)
+- [x] **SET-BEHV-02**: User can set a draft confidence threshold (0.0–1.0); AI only saves drafts ≥ threshold
 - [ ] **SET-BEHV-03**: User can toggle daily digest (reuses v1.0 ANL-03 config)
-- [ ] **SET-BEHV-04**: User can toggle sensitive-data protection (controls v1.0 LLM-05 PII redaction behavior; default ON)
+- [x] **SET-BEHV-04**: User can toggle sensitive-data protection (controls v1.0 LLM-05 PII redaction behavior; default ON)
 - [ ] **SET-BEHV-05**: User can surface the shadow-mode toggle (reuses v1.0 TRG-07) from the assistant Settings page
 
 ### Settings Page — Sender Safety Net (carried from v1.1)
 
-- [ ] **SET-SAFE-01**: User can view, add, and remove sender safety net entries (email or domain pattern) via the Settings page (exposes existing v1.0 TRG-07..08 tables to end users for the first time)
+- [x] **SET-SAFE-01**: User can view, add, and remove sender safety net entries (email or domain pattern) via the Settings page (exposes existing v1.0 TRG-07..08 tables to end users for the first time)
 - [ ] **SET-SAFE-02**: User can paste-import multiple entries at once with a parsed preview before save
 - [ ] **SET-SAFE-03**: User can pick per-entry mode (`protect` = never auto-act, `escalate` = notify but don't act)
-- [ ] **SET-SAFE-04**: User sees a visual indicator in the audit log when a rule was blocked by the safety net ("Was going to archive, blocked by VIP rule for ceo@acme.com")
+- [x] **SET-SAFE-04**: User sees a visual indicator in the audit log when a rule was blocked by the safety net ("Was going to archive, blocked by VIP rule for ceo@acme.com")
 
 ### Settings Page — AI Provider/Model (carried from v1.1, rewired onto curated catalog)
 
-- [ ] **SET-AI-01**: User has ONE BYOK card with an `Active` switch as the only on/off control. When the row is `active=true` AND has a tested model, every AI feature (chat / triage / draft / voice-generate) runs through that key+URL+model. When `active=false` (or no row), every feature falls back to the admin-curated catalog default. **Updated 2026-05-26 round 2** — no per-feature picker, no separate `Platform default ↔ Use my key` mode card; the `active` flag on the BYOK row replaces both
-- [ ] **SET-AI-02**: BYOK row holds provider (OpenAI / Anthropic / Google / DeepSeek only — NEVER OpenRouter, NEVER 9Router), base URL (auto-filled per provider, user-editable to support OpenAI-compatible / Anthropic-compatible endpoints; validated as `https://` except `http://localhost*` for dev), API key (AES-GCM encrypted via v1.0 LLM-04 / `RefreshTokenCipher`, never logged, never returned to the frontend after save — only `lastFourChars`), and a user-picked model from the Test-connection response. Saving any field clears `last_test_result` and `last_tested_at` and forces `active=false`. Switching providers/URLs/keys replaces the single tenant row
-- [ ] **SET-AI-03**: User sees a single tenant-wide last-7d AI cost figure below the BYOK card (e.g. `Chi phí AI 7 ngày qua: $2.43`). **Updated 2026-05-26** — per-feature cost rows removed; aggregation is a single tenant-scoped sum from existing `llm_call_audit` rows, no `call_site=CHAT` schema change required
-- [ ] **SET-AI-04**: User can test the BYOK connection (either against the stored row OR an inline-payload pre-save) using the same `/v1/models` probe and enum-only response shape (`OK / INVALID_KEY / RATE_LIMITED / NETWORK_ERROR / TIMEOUT`) as admin MKEY-03. On `OK` the response additionally carries `models[]` (provider's chat-completion-capable model IDs, capped at 100) so the user can pick a model from the result. Both admin and user paths delegate to a shared `ProviderConnectionTester` (D-14). Rate-limited to 10 tests/hour per tenant. Activating the BYOK row requires the last test to be `OK` AND a model to be picked, otherwise `PUT /api/byok/active` returns HTTP 400 `code=ai.byok.no_model_picked`
+- [x] **SET-AI-01**: User has ONE BYOK card with an `Active` switch as the only on/off control. When the row is `active=true` AND has a tested model, every AI feature (chat / triage / draft / voice-generate) runs through that key+URL+model. When `active=false` (or no row), every feature falls back to the admin-curated catalog default. **Updated 2026-05-26 round 2** — no per-feature picker, no separate `Platform default ↔ Use my key` mode card; the `active` flag on the BYOK row replaces both
+- [x] **SET-AI-02**: BYOK row holds provider (OpenAI / Anthropic / Google / DeepSeek only — NEVER OpenRouter, NEVER 9Router), base URL (auto-filled per provider, user-editable to support OpenAI-compatible / Anthropic-compatible endpoints; validated as `https://` except `http://localhost*` for dev), API key (AES-GCM encrypted via v1.0 LLM-04 / `RefreshTokenCipher`, never logged, never returned to the frontend after save — only `lastFourChars`), and a user-picked model from the Test-connection response. Saving any field clears `last_test_result` and `last_tested_at` and forces `active=false`. Switching providers/URLs/keys replaces the single tenant row
+- [x] **SET-AI-03**: User sees a single tenant-wide last-7d AI cost figure below the BYOK card (e.g. `Chi phí AI 7 ngày qua: $2.43`). **Updated 2026-05-26** — per-feature cost rows removed; aggregation is a single tenant-scoped sum from existing `llm_call_audit` rows, no `call_site=CHAT` schema change required
+- [x] **SET-AI-04**: User can test the BYOK connection (either against the stored row OR an inline-payload pre-save) using the same `/v1/models` probe and enum-only response shape (`OK / INVALID_KEY / RATE_LIMITED / NETWORK_ERROR / TIMEOUT`) as admin MKEY-03. On `OK` the response additionally carries `models[]` (provider's chat-completion-capable model IDs, capped at 100) so the user can pick a model from the result. Both admin and user paths delegate to a shared `ProviderConnectionTester` (D-14). Rate-limited to 10 tests/hour per tenant. Activating the BYOK row requires the last test to be `OK` AND a model to be picked, otherwise `PUT /api/byok/active` returns HTTP 400 `code=ai.byok.no_model_picked`
 
 ### Rule Actions and Examples Catalog (NEW — Phase 08.1)
 
@@ -260,26 +260,26 @@ Phase-to-requirement mapping (populated by gsd-roadmapper 2026-05-19).
 | RACT-10 | Phase 08.1 | Complete |
 | RACT-11 | Phase 08.1 | Pending |
 | RACT-12 | Phase 08.1 | Complete |
-| SET-VOICE-01 | Phase 9 | Pending |
-| SET-VOICE-02 | Phase 9 | Pending |
-| SET-VOICE-03 | Phase 9 | Pending |
-| SET-VOICE-04 | Phase 9 | Pending |
-| SET-VOICE-05 | Phase 9 | Pending |
-| SET-VOICE-06 | Phase 9 | Pending |
-| SET-VOICE-07 | Phase 9 | Pending |
-| SET-BEHV-01 | Phase 9 | Pending |
-| SET-BEHV-02 | Phase 9 | Pending |
+| SET-VOICE-01 | Phase 9 | Complete |
+| SET-VOICE-02 | Phase 9 | Complete |
+| SET-VOICE-03 | Phase 9 | Complete |
+| SET-VOICE-04 | Phase 9 | Complete |
+| SET-VOICE-05 | Phase 9 | Complete |
+| SET-VOICE-06 | Phase 9 | Complete |
+| SET-VOICE-07 | Phase 9 | Complete |
+| SET-BEHV-01 | Phase 9 | Complete |
+| SET-BEHV-02 | Phase 9 | Complete |
 | SET-BEHV-03 | Phase 9 | Pending |
-| SET-BEHV-04 | Phase 9 | Pending |
+| SET-BEHV-04 | Phase 9 | Complete |
 | SET-BEHV-05 | Phase 9 | Pending |
-| SET-SAFE-01 | Phase 9 | Pending |
+| SET-SAFE-01 | Phase 9 | Complete |
 | SET-SAFE-02 | Phase 9 | Pending |
 | SET-SAFE-03 | Phase 9 | Pending |
-| SET-SAFE-04 | Phase 9 | Pending |
-| SET-AI-01 | Phase 9 | Pending |
-| SET-AI-02 | Phase 9 | Pending |
-| SET-AI-03 | Phase 9 | Pending |
-| SET-AI-04 | Phase 9 | Pending |
+| SET-SAFE-04 | Phase 9 | Complete |
+| SET-AI-01 | Phase 9 | Complete |
+| SET-AI-02 | Phase 9 | Complete |
+| SET-AI-03 | Phase 9 | Complete |
+| SET-AI-04 | Phase 9 | Complete |
 | ARCH-08 | Phase 8 | Complete |
 | ARCH-09 | Phase 8 | Complete |
 | ARCH-10 | Phase 8 | Complete |

diff --git a/.planning/ROADMAP.md b/.planning/ROADMAP.md
@@ -32,7 +32,7 @@ Full details: [milestones/v1.1-ROADMAP.md](milestones/v1.1-ROADMAP.md)
 
 - [x] **Phase 8: Admin Console & Operator Tooling** — Operators can log in to a hardened `/admin/*` console with RBAC + append-only audit, configure 6 LLM providers with master keys, curate the per-feature model catalog, inspect tenant health / worker queue / platform LLM spend, and deploy v1.2 infrastructure via the re-platformed reverse proxy (completed 2026-05-20)
 - [ ] **Phase 08.1: Inbox Zero-style Rule Actions & Admin-managed Examples Catalog** — Users can build rules from Inbox Zero-style examples/personas and enable expanded actions including send replies, forward, and send email behind one default-ON global auto-send setting, runtime safety gates, fallback-to-draft behavior, and auditable send boundaries
-- [ ] **Phase 9: User Settings UI on Curated Catalog** — Users can configure voice, behavior, safety net, and AI provider/model across four tabs backed by the admin-curated catalog
+- [x] **Phase 9: User Settings UI on Curated Catalog** — Users can configure voice, behavior, safety net, and AI provider/BYOK on the single `/ai` settings surface backed by the admin-curated catalog (completed 2026-05-29)
 
 ## Phase Details
 
@@ -160,13 +160,13 @@ Plans:
 5. In `AI Provider`, user fills a single BYOK card with: provider `<Select>` (OpenAI / Anthropic / Google / DeepSeek — never OpenRouter or 9Router), base URL `<Input>` (auto-filled per provider, user-editable for OpenAI-compatible / Anthropic-compatible endpoints), API key (AES-GCM encrypted, no plaintext echo, masked display on re-render), model `<Select>` (populated from the provider's `/v1/models` response returned by Test connection), an Active `<Switch>` (default OFF; disabled until a model is picked AND the last Test result is `OK`), `Kiểm tra kết nối`, and `Lưu`. When the row is `active=true` AND has a tested model, every AI feature (chat, triage, draft, voice-generate) runs through that BYOK row; otherwise the admin-curated catalog default applies. Test connection uses the SAME enum-only response (`OK / INVALID_KEY / RATE_LIMITED / NETWORK_ERROR / TIMEOUT`) as admin MKEY-03 via the shared `ProviderConnectionTester`, plus a `models[]` list on `OK` so the user can pick a model. A single tenant-wide last-7d cost figure renders below the card. BYOK lives on `/ai` because Zero Mail is single-tenant-per-user. **Updated 2026-05-26 during plan-phase round 2** — per-feature picker, per-feature `Platform default ↔ Use my key` toggle, AND the tenant-wide mode card all removed; replaced by a single BYOK card with an Active switch.
 
 **Plans** (7 plans across 4 waves):
-- [ ] 09-01-PLAN.md — Wave 0: Liquibase changesets 094..097 + JPA entity scaffolding + 33 Wave-0 test stubs
-- [ ] 09-02-PLAN.md — Wave 1: Voice + Behavior + Knowledge backend (services/controllers/DTOs) + DraftReplyWorker + SensitiveDataRedactor wiring
-- [ ] 09-03-PLAN.md — Wave 1: Safety Net DELETE + DOMAIN pattern + triage audit blocked_by_safety_net_pattern badge
-- [ ] 09-04-PLAN.md — Wave 1: ProviderConnectionTester extraction + UserByokService + ByokProviderResolver + UserByokController + AiCostQueryService (D-17)
-- [ ] 09-05-PLAN.md — Wave 1: SET-VOICE-07 generate-from-sent (in-memory privacy invariant + Spring AI observation hardening)
-- [ ] 09-06-PLAN.md — Wave 2: OpenAPI regen + FE sections + Knowledge feature + AiProviderSection + ByokForm removal from /settings
-- [ ] 09-07-PLAN.md — Wave 3: Playwright e2e ai-settings.spec.ts + Phase9ArchitectureTest aggregate + manual UX checkpoint
+- [x] 09-01-PLAN.md — Wave 0: Liquibase changesets 094..097 + JPA entity scaffolding + 33 Wave-0 test stubs
+- [x] 09-02-PLAN.md — Wave 1: Voice + Behavior + Knowledge backend (services/controllers/DTOs) + DraftReplyWorker + SensitiveDataRedactor wiring
+- [x] 09-03-PLAN.md — Wave 1: Safety Net DELETE + DOMAIN pattern + triage audit blocked_by_safety_net_pattern badge
+- [x] 09-04-PLAN.md — Wave 1: ProviderConnectionTester extraction + UserByokService + ByokProviderResolver + UserByokController + AiCostQueryService (D-17)
+- [x] 09-05-PLAN.md — Wave 1: SET-VOICE-07 generate-from-sent (in-memory privacy invariant + Spring AI observation hardening)
+- [x] 09-06-PLAN.md — Wave 2: OpenAPI regen + FE sections + Knowledge feature + AiProviderSection + ByokForm removal from /settings
+- [x] 09-07-PLAN.md — Wave 3: Playwright e2e ai-settings.spec.ts + Phase9ArchitectureTest aggregate + manual UX checkpoint
 
 **UI hint**: yes
 
@@ -180,7 +180,7 @@ Plans:
 | 7. Chat Email Assistant | v1.1 | 6/6 | Complete | 2026-05-18 |
 | 8. Admin Console & Operator Tooling | v1.2 | 6/6 | Complete   | 2026-05-20 |
 | 08.1. Inbox Zero-style Rule Actions & Admin-managed Examples Catalog | v1.2 | 5/6 | In Progress|  |
-| 9. User Settings UI on Curated Catalog | v1.2 | 0/7 | In Progress | — |
+| 9. User Settings UI on Curated Catalog | v1.2 | 7/7 | Complete | 2026-05-29 |
 
 ---