diff --git a/RPG_FLOW.md b/RPG_FLOW.md
new file mode 100644
index 0000000..c552495
--- /dev/null
+++ b/RPG_FLOW.md
@@ -0,0 +1,369 @@
+# RPG-режим: полный алгоритм запросов (шпаргалка)
+
+Документ описывает, **какой HTTP-запрос что получает и что делает**, когда RPG включён полностью (`rpg_enabled = 1`) и в настройках активны все опции мастера:
+
+```json
+{
+  "dice": true,
+  "narrator": true,
+  "quests": true,
+  "affinity": true,
+  "choices": true,
+  "stats": false
+}
+```
+
+> **Шкалы** (`lust` / `stamina` / `tension`): по умолчанию `stats: false`. Если включить чекбокс «Шкалы», после каждого хода добавляется ветка `stats_delta` (см. § «Вариант: stats включены»).
+
+Пример проходит **от создания чата до 3-го сообщения игрока**, с перекрытием основных веток (диалог / d20 / кнопка выбора / сдвиг сюжета / факты / SD).
+
+---
+
+## Модели и роли
+
+| Роль | Env / модель по умолчанию | Где вызывается |
+|------|---------------------------|----------------|
+| **Персонаж (чат)** | `CHAT_MODEL` → `mistralai/mistral-nemo` | `POST /chat/stream` → `stream_message()` |
+| **Нарратор pre/post** | `RPG_NARRATOR_MODEL` → `deepseek/deepseek-chat-v3` | Перед/после ответа персонажа |
+| **Сюжетная арка** | `RPG_PLOT_MODEL` → DeepSeek v3 | Opening, fallback если арки нет |
+| **Факты** | `RPG_FACTS_MODEL` → DeepSeek v3 | После каждого ответа персонажа |
+| **SD-сцена** | `SD_PROMPT_MODEL` или `SYSTEM_MODEL` | Opening + конец каждого stream |
+| **Появление (prose)** | `SYSTEM_MODEL` | Только при сохранении тегов персоны (не в ходе RPG) |
+
+Персонаж **не пишет** affinity/stats в БД — только читает их из `runtime_suffix`. Пишет **только narrator post** (`affinity_delta`, `stats_delta`, `scene_update`, …).
+
+---
+
+## Слои промпта (что видит CHAT_MODEL)
+
+```
+llm_system = static_prompt + runtime_suffix [+ context_warning]
+```
+
+| Слой | Источник | Когда меняется |
+|------|----------|----------------|
+| **static_prompt** | `get_system_prompt()` — persona + lorebook (последние 5 реплик + текущий user) + `ROLEPLAY_GUARDRAILS` | Каждый stream; в `messages` хранится **только** static (без RPG) |
+| **facts_block** | До 20 фактов из `facts_json` | После `extract_facts` на прошлом ходу |
+| **PlotArc** | `title`, `phase`, `next_beat_hint` | При генерации/сдвиге арки |
+| **Status quo** | `sessions.status_quo` | narrator pre/post, opening |
+| **Scene** | `sessions.scene_json` | narrator pre/post, opening |
+| **Relationship** | `affinity` + тон | narrator post (`affinity_delta`) |
+| **Character state** | lust/stamina/tension | Только если `stats: true` |
+| **Narrator directives** | narrator **pre** (текущий ход) | Эфемерно, не в БД |
+| **Mechanics** | d20 + outcome (текущий ход) | Только если была проверка |
+| **Context warning** | `~N% of CHAT_CONTEXT_MAX` | Если оценка > 85% |
+
+История в LLM: **все** `user` / `assistant` из БД + один объединённый `system`.
+
+---
+
+## Фаза 0: создание RPG-чата (UI → API)
+
+### 0.1 `PATCH /sessions/{id}`
+
+**Тело (пример):**
+```json
+{
+  "persona_id": "card_abc",
+  "rpg_enabled": true,
+  "genre": "fantasy",
+  "rpg_settings_json": "{\"dice\":true,\"narrator\":true,...}"
+}
+```
+
+**Действия:** запись в `sessions` — флаги RPG, жанр, настройки. Сообщений ещё нет.
+
+### 0.2 `POST /chat/init`
+
+**Тело:** `{ "session_id", "persona_id", "first_mes_override"? }`
+
+**Действия:**
+1. `get_system_prompt(persona, [], "")` → static (персона, lorebook, guardrails).
+2. `upsert_static_system_message` — одна строка `role=system` в `messages`.
+3. Если истории нет — `add_message(assistant, first_mes)` из карты / override.
+
+**LLM:** не вызывается.
+
+**БД после:** `messages = [system, assistant(first_mes)]`.
+
+### 0.3 `POST /chat/opening/process`
+
+**Тело:** `{ "session_id", "persona_id", "rpg": true }`
+
+Последовательность **только на opening** (без `narrator_pre`, без stream персонажа):
+
+```mermaid
+sequenceDiagram
+    participant UI
+    participant API as opening/process
+    participant Plot as RPG_PLOT_MODEL
+    participant Nar as RPG_NARRATOR_MODEL
+    participant SD as SD_PROMPT_MODEL
+
+    UI->>API: POST opening/process
+    API->>Plot: generate_plot_arc(first_mes, persona, genre)
+    Plot-->>API: plot_arc_json
+    API->>API: seed rpg_quests from beats
+    API->>Nar: narrator_post OPENING
+    Nar-->>API: status_quo, scene, affinity_delta, outfit, choices, quests
+    API->>API: apply_narrator_post → sessions
+    API->>SD: generate_sd_prompt + run_sd_for_message
+    API-->>UI: plot_arc, quests, affinity, choices, image_*
+```
+
+| Шаг | Модель | Вход | Выход в БД / UI |
+|-----|--------|------|-----------------|
+| 1 | `generate_plot_arc` | имя, description, scenario, **first_mes**, facts, genre | `plot_arc_json`; квесты из `beats[].title` |
+| 2 | `narrator_post(is_opening=true)` | контекст: `assistant: {first_mes}`; arc; facts | `status_quo`, `scene_json`, `outfit_json`, `affinity += delta`, квесты, **choices** (в ответ API, не в messages) |
+| 3 | `generate_sd_prompt` | история, outfit, **scene_json** | картинка на last assistant id |
+
+**Особенности opening для narrator post:**
+- Просит заполнить `scene_update` из greeting + scenario.
+- Просит **non-zero `affinity_delta`**, если в first_mes явный тёплый/враждебный тон.
+- `outfit_update` — начальная одежда из greeting (danbooru-теги).
+
+**UI после:** `reloadChatFromServer`, панель квестов, 💖 affinity, кнопки `choices`, картинка SD.
+
+---
+
+## Состояние сессии перед 1-м сообщением игрока
+
+| Поле | Пример |
+|------|--------|
+| `messages` | system, assistant(first_mes) |
+| `plot_arc_json` | сгенерированная арка |
+| `status_quo`, `scene_json`, `outfit_json` | из narrator opening |
+| `affinity` | 0 или ±1..2 после opening |
+| `facts_json` | `[]` |
+| `narrative_stats_json` | `{"lust":0,"stamina":10,"tension":0}` (если stats выкл. — всё равно дефолт в БД) |
+
+---
+
+## Сообщение игрока №1 — обычная реплика (без d20)
+
+**Пример:** игрок пишет: «Привет, как тебя зовут?»
+
+**UI:** `POST /chat/stream`  
+```json
+{
+  "session_id": "...",
+  "persona_id": "...",
+  "message": "Привет, как тебя зовут?",
+  "is_narrator_choice": false
+}
+```
+
+### Этап A — подготовка (до stream)
+
+1. **static_prompt** = `get_system_prompt(persona, history, message)` — lorebook по последним 5 репликам + guardrails.
+2. **RPG pre (narrator):**
+   - **Вход user:** persona, user action, Global plot (= full arc JSON), Facts, Recent (до 8 реплик).
+   - **Модель:** `RPG_NARRATOR_MODEL`.
+   - **Ожидание:** `needs_check: false` для чистого диалога.
+   - **Эффекты:**
+     - `directives` → попадут в `narrator_extra`.
+     - `status_quo_update` → `UPDATE sessions.status_quo`.
+     - `scene_update` (partial) → merge в `scene_json`.
+     - **Нет** d20, **нет** bubble «Рассказчик».
+3. **runtime_suffix** = `build_rpg_runtime_suffix(session)` + `narrator_extra` (directives только).
+4. `upsert_static_system_message(static)` — в БД system без RPG-блоков.
+5. `add_message(user, message)`.
+6. **context_usage** — если > 85%, в конец system добавляется `[Context: ~N% …]`.
+7. **llm_messages** = system(static+runtime) + вся история user/assistant.
+
+### Этап B — SSE stream
+
+| Событие SSE | Когда |
+|-------------|--------|
+| `{ "chunk": "..." }` | Поток `CHAT_MODEL` — ответ персонажа |
+| `{ "done": true, ... }` | После сохранения assistant в БД |
+
+**Персонаж:** один запрос stream, читает affinity/scene/status/arc/facts/directives, **не меняет** счётчики.
+
+### Этап C — post-process (после ответа персонажа, тот же HTTP stream)
+
+Порядок внутри `generate()`:
+
+| # | Действие | Модель | Условие |
+|---|----------|--------|---------|
+| C1 | `generate_plot_arc` | PLOT | Только если `plot_arc_json` пуст (редко после opening) |
+| C2 | `should_advance_arc(user_message)` | код | Ключевые слова: отдых → `event_driven:rest`, путь → `travel`, помощь → `help_request` |
+| C3 | `pop_matching_beats` + injection | — | Если trig совпал с beat; choices из beat |
+| C4 | `advance_phase` | — | Если beats пусты — фаза opening→hook→… |
+| C5 | `extract_facts` | FACTS | Последние 10 реплик; merge до 80, в промпт 20 |
+| C6 | **`narrator_post`** | NARRATOR | Контекст: последние 8 реплик **включая новый ответ** |
+| C7 | **`apply_narrator_post`** | — | status_quo, affinity_delta, scene, outfit, quests, stats_delta* |
+| C8 | `generate_sd_prompt` + Comfy | SD | outfit + scene_json + последние 6 реплик |
+| C9 | SSE `done` | — | choices, affinity, quests, image_*, debug |
+
+\* `stats_delta` только при `rpg_settings.stats === true`.
+
+**Для реплики №1 (диалог):** narrator pre → no check; post → facts, возможно choices, affinity ±, scene/status.
+
+---
+
+## Сообщение игрока №2 — действие с проверкой d20
+
+**Пример:** «Пытаюсь перепрыгнуть через пропасть»
+
+Тот же `POST /chat/stream`. Отличия в **этапе A**:
+
+```mermaid
+flowchart TD
+    pre1[narrator_pre без броска]
+    need{needs_check AND dice?}
+    roll[d20 1-20]
+    pre2[narrator_pre с roll+outcome]
+    res[resolution_text + action_resolutions]
+    chat[CHAT_MODEL stream]
+    post[narrator_post + facts + SD]
+
+    pre1 --> need
+    need -->|да| roll --> pre2 --> res
+    need -->|нет| chat
+    res --> chat
+    chat --> post
+```
+
+| Шаг | Что происходит |
+|-----|----------------|
+| pre (фаза 1) | `needs_check: true` |
+| Бросок | `random 1..20` → outcome: crit fail / fail / success / crit success |
+| pre2 | Тот же user + `Roll d20=N`, `Outcome=...` → **`resolution_text` обязателен** |
+| UI | SSE **`{ "narrator": { roll, outcome, text } }`** **до** chunks — bubble «📖 Рассказчик» |
+| runtime | + `--- Mechanics ---` (d20, outcome, «не противоречь») |
+| БД | `action_resolutions` — intent, roll, outcome, resolution_text |
+| Персонаж | Должен согласовать ответ с resolution + mechanics |
+
+Post-process (C) — **тот же**, что на ходу 1. Facts могут записать «игрок не перепрыгнул», quest может обновиться.
+
+---
+
+## Сообщение игрока №3 — кнопка выбора + сдвиг арки
+
+**Пример:** игрок жмёт choice «Отправиться в лес» (label уходит как user message)  
+или пишет: «Идём дальше в лес»
+
+**UI:** `sendMessage(label, true)` → `is_narrator_choice: true`
+
+| Отличие | Поведение |
+|---------|-----------|
+| Текст в БД | `"[Player chose: Отправиться в лес]"` вместо сырого label |
+| pre/post | Как обычный ход; narrator видит метку выбора |
+| **should_advance_arc** | Если в тексте есть «идем дальше», «в путь», … → `event_driven:travel` |
+| beats | `pop_matching_beats(arc, trig)` — до 1 beat; **injection** в debug; **choices** из beat + post |
+| phase | Если beats опустели → `advance_phase` (например hook → complication) |
+
+Остальной pipeline = stream №1 или №2 (в зависимости от needs_check).
+
+---
+
+## Сводная таблица: кто что обновляет (ход 1–3)
+
+| Данные | Opening | narrator pre | CHAT_MODEL | narrator post | extract_facts | arc engine |
+|--------|---------|--------------|------------|---------------|---------------|------------|
+| messages | first_mes | — | user + assistant | — | — | — |
+| plot_arc_json | create | — | read | read | — | advance/beats |
+| status_quo | post | pre/post | read | post | — | — |
+| scene_json | post | pre/post | read | post | — | — |
+| outfit_json | post | — | read | post | — | — |
+| affinity | post | — | read | delta | — | — |
+| facts_json | — | read | read | post.facts* | merge | — |
+| rpg_quests | seed | — | — | post | — | — |
+| action_resolutions | — | d20 path | — | — | — | — |
+| narrative_stats | default | — | read** | stats_delta** | — | — |
+
+\* facts из post в JSON схеме narrator — сейчас основной путь facts = `extract_facts`.  
+\** только при `stats: true`.
+
+---
+
+## Варианты настроек (перекрытие)
+
+### `narrator: false`
+
+- Нет narrator pre/post, нет directives/mechanics bubble.
+- `runtime_suffix` = только `build_rpg_runtime_suffix` (facts, arc, status, scene, affinity).
+- Post-process: нет C6–C7; facts/arc/SD могут остаться (arc/SD/facts всё ещё в коде stream).
+
+### `dice: false`
+
+- `needs_check` от pre **игнорируется** — всегда ветка без броска (как диалог).
+- pre2 и mechanics не вызываются.
+
+### `affinity: false`
+
+- Нет блока Relationship в runtime; `affinity_delta` не применяется.
+
+### `quests: false`
+
+- Opening не сидирует квесты; post не вызывает `upsert_quest`.
+
+### `choices: false`
+
+- UI не рисует кнопки; post/beat choices не добавляются в SSE.
+
+### `stats: true`
+
+- В runtime: блок lust/stamina/tension (0–10).
+- post: `stats_delta` каждый ключ −2..+2, clamp 0–10.
+- Дефолт stamina = 10.
+
+### `is_narrator_choice: true`
+
+- User content = `[Player chose: …]`.
+
+### Контекст > 85%
+
+- В system добавляется одна строка-предупреждение (не в БД).
+
+---
+
+## Вспомогательные запросы (не в примере 1–3, но в RPG)
+
+| Запрос | Когда |
+|--------|-------|
+| `GET /chat/system/{id}` | Панель system blob: static, runtime-превью, scene, stats, **context_usage** |
+| `GET /chat/history/{id}` | Перезагрузка чата |
+| `POST /chat/rpg/bootstrap` | RPG на **старом** чате: arc + **narrator_post opening** + apply (как opening, без нового first_mes) |
+| `POST /chat/` (не stream) | Упрощённый ход: static + runtime, один ответ, без narrator pre/post |
+
+---
+
+## Пример хронологии (3 сообщения, все опции ON)
+
+| # | HTTP | LLM-вызовы | Заметки |
+|---|------|------------|---------|
+| 0 | PATCH session | — | rpg_enabled |
+| 0 | POST init | — | system + first_mes |
+| 0 | POST opening/process | Plot, Narrator post, SD | arc, scene, affinity, quests, choices |
+| 1 | POST stream | Pre (no check), **Chat**, Facts, Post, SD | Диалог; +facts; choices в done |
+| 2 | POST stream | Pre (check), **Pre2+d20**, **Chat**, Facts, Post, SD | Bubble рассказчика; action_resolutions |
+| 3 | POST stream | Pre, Chat, Facts, Post, SD; arc travel? | `[Player chose:…]` или текст; beat injection |
+
+**Порядок SSE на ходу 2:**
+1. `narrator` (d20)
+2. `chunk` × N
+3. `image_generating` (если SD)
+4. `done` (choices, affinity, quests, debug)
+
+---
+
+## Файлы для углубления
+
+| Тема | Файл |
+|------|------|
+| Stream + RPG | `routers/chat.py` → `chat_stream` |
+| Opening | `services/opening.py` |
+| Narrator JSON | `services/rpg_narrator.py` |
+| Runtime blocks | `routers/chat.py` → `build_rpg_runtime_suffix`, `services/rpg_state.py` |
+| Persist post | `services/rpg_state.py` → `apply_narrator_post` |
+| Arc / beats | `services/rpg_plot.py` |
+| Facts | `services/rpg_facts.py` |
+| UI stream | `static/js/chat.js` → `sendMessage`, `consumeStream` |
+| New chat | `static/js/newChatWizard.js` |
+| Context % | `services/context_budget.py` |
+
+---
+
+*Обновлено под текущую реализацию RP UX (anti-OOC, scene_json, context blob, rpFormat, bootstrap narrator).*
diff --git a/database/db.py b/database/db.py
index 38f7a49..d753ebb 100644
--- a/database/db.py
+++ b/database/db.py
@@ -86,6 +86,12 @@ async def _migrate_messages_columns(db):
         await db.execute("ALTER TABLE messages ADD COLUMN image_prompt TEXT")
     if "image_path" not in cols:
         await db.execute("ALTER TABLE messages ADD COLUMN image_path TEXT")
+    if "image_prompt_alt" not in cols:
+        await db.execute("ALTER TABLE messages ADD COLUMN image_prompt_alt TEXT")
+    if "image_path_alt" not in cols:
+        await db.execute("ALTER TABLE messages ADD COLUMN image_path_alt TEXT")
+    if "choices_json" not in cols:
+        await db.execute("ALTER TABLE messages ADD COLUMN choices_json TEXT")
 
 
 async def _migrate_personas_columns(db):
@@ -105,6 +111,8 @@ async def _migrate_personas_columns(db):
         await db.execute("ALTER TABLE personas ADD COLUMN avatar_path TEXT DEFAULT ''")
     if "alternate_greetings_json" not in cols:
         await db.execute("ALTER TABLE personas ADD COLUMN alternate_greetings_json TEXT DEFAULT '[]'")
+    if "appearance_prose" not in cols:
+        await db.execute("ALTER TABLE personas ADD COLUMN appearance_prose TEXT DEFAULT ''")
 
 
 async def _migrate_sessions_columns(db):
@@ -130,6 +138,13 @@ async def _migrate_sessions_columns(db):
         await db.execute("ALTER TABLE sessions ADD COLUMN rpg_settings_json TEXT DEFAULT '{}'")
     if "outfit_json" not in cols:
         await db.execute("ALTER TABLE sessions ADD COLUMN outfit_json TEXT DEFAULT '[]'")
+    if "scene_json" not in cols:
+        await db.execute("ALTER TABLE sessions ADD COLUMN scene_json TEXT DEFAULT '{}'")
+    if "narrative_stats_json" not in cols:
+        await db.execute(
+            "ALTER TABLE sessions ADD COLUMN narrative_stats_json TEXT DEFAULT "
+            "'{\"lust\":0,\"stamina\":10,\"tension\":0}'"
+        )
 
 async def _migrate_rpg_quests(db):
     await db.executescript("""
@@ -170,3 +185,5 @@ async def _migrate_characters_columns(db):
         await db.execute("ALTER TABLE characters ADD COLUMN avatar_path TEXT DEFAULT ''")
     if "alternate_greetings_json" not in cols:
         await db.execute("ALTER TABLE characters ADD COLUMN alternate_greetings_json TEXT DEFAULT '[]'")
+    if "appearance_prose" not in cols:
+        await db.execute("ALTER TABLE characters ADD COLUMN appearance_prose TEXT DEFAULT ''")
diff --git a/main.py b/main.py
index 802a95a..869840c 100644
--- a/main.py
+++ b/main.py
@@ -3,9 +3,10 @@ from contextlib import asynccontextmanager
 from fastapi import FastAPI
 from fastapi.staticfiles import StaticFiles
 from fastapi.responses import FileResponse
-from routers import chat, personas, sessions, characters, images, translate
+from routers import chat, personas, sessions, characters, images, translate, debug
 from database.db import init_db
 from services.persona_seed import seed_default_personas
+from services.system_message_migration import migrate_static_system_messages
 
 logging.basicConfig(level=logging.INFO, format="%(levelname)s %(name)s: %(message)s")
 
@@ -14,6 +15,7 @@ logging.basicConfig(level=logging.INFO, format="%(levelname)s %(name)s: %(messag
 async def lifespan(app: FastAPI):
     await init_db()
     await seed_default_personas()
+    await migrate_static_system_messages()
     yield
 
 
@@ -25,6 +27,7 @@ app.include_router(sessions.router)
 app.include_router(characters.router)
 app.include_router(images.router)
 app.include_router(translate.router)
+app.include_router(debug.router)
 
 app.mount("/static", StaticFiles(directory="static"), name="static")
 
@@ -34,6 +37,11 @@ async def root():
     return FileResponse("static/index.html")
 
 
+@app.get("/debug")
+async def debug_page():
+    return FileResponse("static/debug.html")
+
+
 @app.get("/health")
 async def health():
     return {"status": "ok"}
diff --git a/main_carrie-just-a-friendly-hug-3ccb4b5342bb_spec_v2.png b/main_carrie-just-a-friendly-hug-3ccb4b5342bb_spec_v2.png
new file mode 100644
index 0000000..8ef6eb0
Binary files /dev/null and b/main_carrie-just-a-friendly-hug-3ccb4b5342bb_spec_v2.png differ
diff --git a/main_clingy-obsessive-girlfriend-af26ead7_spec_v2.png b/main_clingy-obsessive-girlfriend-af26ead7_spec_v2.png
new file mode 100644
index 0000000..058ce48
Binary files /dev/null and b/main_clingy-obsessive-girlfriend-af26ead7_spec_v2.png differ
diff --git a/main_delta-125aa7a6_spec_v2.png b/main_delta-125aa7a6_spec_v2.png
new file mode 100644
index 0000000..8f3f8b4
Binary files /dev/null and b/main_delta-125aa7a6_spec_v2.png differ
diff --git a/main_violet-merino-d2e9f62b5d77_spec_v2.png b/main_violet-merino-d2e9f62b5d77_spec_v2.png
new file mode 100644
index 0000000..da57117
Binary files /dev/null and b/main_violet-merino-d2e9f62b5d77_spec_v2.png differ
diff --git a/main_vulpisfoglia-e0a6befda921_spec_v2.png b/main_vulpisfoglia-e0a6befda921_spec_v2.png
new file mode 100644
index 0000000..9b52d48
Binary files /dev/null and b/main_vulpisfoglia-e0a6befda921_spec_v2.png differ
diff --git a/main_your-scumbag-superheroine-friend-c65bf1fe881c_spec_v2.png b/main_your-scumbag-superheroine-friend-c65bf1fe881c_spec_v2.png
new file mode 100644
index 0000000..fcd461e
Binary files /dev/null and b/main_your-scumbag-superheroine-friend-c65bf1fe881c_spec_v2.png differ
diff --git a/models/schemas.py b/models/schemas.py
index 6ed48a4..648241c 100644
--- a/models/schemas.py
+++ b/models/schemas.py
@@ -24,6 +24,37 @@ class RegenerateRequest(BaseModel):
 class ForkSessionRequest(BaseModel):
     until_message_id: int
 
+
+class RebindPersonaRequest(BaseModel):
+    persona_id: str
+    clear_history: bool = False
+
+
+class QuestStatusPatch(BaseModel):
+    status: str
+
+
+class RpgStateDebugPatch(BaseModel):
+    """Debug override for relationship/stats (session-scoped to current player for now)."""
+    affinity: Optional[int] = None
+    lust: Optional[int] = None
+    stamina: Optional[int] = None
+    tension: Optional[int] = None
+
+
+class SessionContextPatch(BaseModel):
+    """Live-edit RPG / SD context fields for the active session."""
+    status_quo: Optional[str] = None
+    global_plot: Optional[str] = None
+    outfit_json: Optional[str] = None
+    scene_json: Optional[str] = None
+    facts_json: Optional[str] = None
+    plot_arc_json: Optional[str] = None
+    affinity: Optional[int] = None
+    lust: Optional[int] = None
+    stamina: Optional[int] = None
+    tension: Optional[int] = None
+
 class ChatResponse(BaseModel):
     reply: str
     session_id: str
diff --git a/routers/characters.py b/routers/characters.py
index 22c2036..27e349b 100644
--- a/routers/characters.py
+++ b/routers/characters.py
@@ -22,6 +22,7 @@ class CardPatch(BaseModel):
     first_mes: Optional[str] = None
     mes_example: Optional[str] = None
     appearance_tags: Optional[str] = None
+    appearance_prose: Optional[str] = None
     lora_name: Optional[str] = None
     lora_weight: Optional[float] = None
     alternate_greetings_json: Optional[str] = None
diff --git a/routers/chat.py b/routers/chat.py
index 02988a6..0d83d10 100644
--- a/routers/chat.py
+++ b/routers/chat.py
@@ -3,14 +3,12 @@ import logging
 import os
 import random
 
-import aiosqlite
 from fastapi import APIRouter, HTTPException
 from fastapi.responses import StreamingResponse
 from pydantic import BaseModel
 
-from database.db import DB_PATH
 from models.schemas import ChatRequest, ChatResponse, MessageEditRequest, RegenerateRequest
-from services.llm import send_message, stream_message
+from services.llm import LLMError, send_message, stream_message
 from services.memory import (
     get_history,
     add_message,
@@ -18,41 +16,74 @@ from services.memory import (
     get_or_create_session,
     get_session,
     update_session_title,
-    update_session_persona,
     get_message_count,
     get_last_assistant_message_id,
     update_message_image,
     update_session_facts,
     update_session_status_quo,
-    update_session_affinity,
     update_session_genre,
-    update_session_rpg_settings,
-    update_session_outfit,
     update_session_plot_arc,
-    upsert_quest,
     get_quests,
+    seed_quests_from_arc,
+    narrator_message_content,
+    parse_narrator_message,
     add_action_resolution,
     get_message,
     update_message_content,
     delete_messages_after,
     delete_message,
+    delete_message_and_following,
+    update_message_choices,
+    clear_choices_for_session,
+    upsert_static_system_message,
+)
+from services.context_budget import compute_payload_usage, context_warning_line
+from services.rpg_state import (
+    apply_narrator_post,
+    parse_scene_json,
+    parse_stats_json,
+    scene_prompt_block,
+    affinity_prompt_block,
+    stats_prompt_block,
+    format_narrator_outcome_for_llm,
+    format_user_message_for_llm,
 )
 from services.personas import get_persona
+from services.chat_prompt import get_system_prompt, DEFAULT_PROMPT
+from services.session_identity import resolve_session_persona
 from services.sd_prompt import generate_sd_prompt, strip_image_prompt_tag, extract_image_prompt_tag
-from services.lorebook import get_lorebook_context
+from services.rp_sanitize import RP_OUTPUT_REMINDER, strip_ooc_from_reply
+from services.sd_images import run_sd_for_message
 from services.character_card import get_character
 from services import sdbackend as sd_service
-from services.rpg_facts import extract_facts, merge_facts, facts_to_prompt
-from services.rpg_plot import generate_plot_arc, should_advance_arc, pop_matching_beats, advance_phase
+from services.rpg_facts import extract_facts, merge_facts_persist, facts_to_prompt, rp_day_from_scene
+from services.rpg_context import format_narrator_context
+from services.rpg_plot import (
+    generate_plot_arc,
+    process_arc_beats,
+    advance_phase,
+    replenish_arc_beats,
+    reconcile_plot_arc,
+    reconcile_plot_arc,
+    choices_from_beat,
+    choices_from_narrator,
+)
 from services.rpg_narrator import narrator_pre, narrator_post
+from services.opening import ensure_plot_arc_and_quests, resolve_greeting, process_opening
 
 logger = logging.getLogger(__name__)
 router = APIRouter(prefix="/chat", tags=["chat"])
 
-DEFAULT_PROMPT = "Ты — полезный AI ассистент. Отвечай чётко и по делу."
 SD_AUTO_GENERATE = os.getenv("SD_AUTO_GENERATE", "false").lower() in ("1", "true", "yes")
 
-DEFAULT_RPG_SETTINGS = {"dice": True, "narrator": True, "quests": True, "affinity": True, "choices": True}
+DEFAULT_RPG_SETTINGS = {
+    "dice": True,
+    "narrator": True,
+    "quests": True,
+    "affinity": True,
+    "choices": True,
+    "stats": False,
+}
 
 
 def get_rpg_settings(session: dict) -> dict:
@@ -62,34 +93,61 @@ def get_rpg_settings(session: dict) -> dict:
         return DEFAULT_RPG_SETTINGS
 
 
-def affinity_prompt_block(affinity: int) -> str:
-    if affinity >= 10:   tone = "very warm, trusting, affectionate"
-    elif affinity >= 5:  tone = "friendly and open"
-    elif affinity >= 1:  tone = "slightly positive"
-    elif affinity <= -5: tone = "hostile or deeply distrustful"
-    elif affinity <= -1: tone = "cold and wary"
-    else:                tone = "neutral"
-    return f"\n\n--- Relationship ---\nAffinity toward player: {affinity} ({tone}). Reflect this in your attitude and word choice.\n---"
+def build_rpg_runtime_suffix(session: dict, rpg_settings: dict, facts_block: str = "") -> str:
+    runtime_suffix = ""
+    if facts_block:
+        runtime_suffix += "\n\n" + facts_block
+    try:
+        arc = json.loads(session.get("plot_arc_json") or "{}")
+    except Exception:
+        arc = {}
+    if arc:
+        runtime_suffix += "\n\n--- PlotArc ---\n" + json.dumps(
+            {k: arc.get(k) for k in ("title", "phase", "next_beat_hint")}, ensure_ascii=False
+        ) + "\n---"
+    status_quo = (session.get("status_quo") or "").strip()
+    if status_quo:
+        from services.rp_sanitize import status_quo_prompt_block
+
+        runtime_suffix += status_quo_prompt_block(status_quo)
+    scene = parse_scene_json(session.get("scene_json"))
+    block = scene_prompt_block(scene)
+    if block:
+        runtime_suffix += block
+    if rpg_settings.get("affinity", True):
+        runtime_suffix += affinity_prompt_block(int(session.get("affinity") or 0))
+    if rpg_settings.get("stats", False):
+        stats = parse_stats_json(session.get("narrative_stats_json"))
+        runtime_suffix += stats_prompt_block(stats)
+    return runtime_suffix
 
 
-async def get_system_prompt(persona_id: str, history: list, user_message: str = "") -> str:
-    persona = await get_persona(persona_id)
-    if not persona:
-        return DEFAULT_PROMPT
-    prompt = persona["prompt"]
-    recent = [m for m in history if m["role"] in ("user", "assistant")][-5:]
-    context = recent + [{"role": "user", "content": user_message}]
-    if persona.get("lorebook_json"):
-        lore = get_lorebook_context(persona.get("lorebook_json", "[]"), context)
-        if lore:
-            prompt += "\n\n" + lore
-    if persona_id.startswith("card_"):
-        card = await get_character(persona_id[5:])
-        if card:
-            lore = get_lorebook_context(card.get("lorebook_json", "[]"), context)
-            if lore:
-                prompt += "\n\n" + lore
-    return prompt
+def messages_for_llm(history: list, llm_system_content: str) -> list[dict]:
+    """Build LLM payload: one system message (static + runtime), no duplicate system rows."""
+    out: list[dict] = []
+    system_used = False
+    for m in history:
+        if m["role"] == "system":
+            if not system_used:
+                out.append({"role": "system", "content": llm_system_content})
+                system_used = True
+        elif m["role"] == "narrator":
+            data = parse_narrator_message(m.get("content") or "")
+            if data:
+                out.append({"role": "user", "content": format_narrator_outcome_for_llm(data)})
+        elif m["role"] == "user":
+            has_res = bool(m.get("action_resolution"))
+            out.append({
+                "role": "user",
+                "content": format_user_message_for_llm(
+                    m["content"], has_dice_resolution=has_res
+                ),
+            })
+        else:
+            out.append({"role": m["role"], "content": m["content"]})
+    if not system_used:
+        out.insert(0, {"role": "system", "content": llm_system_content})
+    return out
 
 
 @router.get("/history/{session_id}")
@@ -100,33 +158,68 @@ async def get_chat_history(session_id: str):
 @router.get("/system/{session_id}")
 async def get_system_blob(session_id: str):
     history = await get_history(session_id)
-    system_msg = next((m for m in history if m.get("role") == "system"), None)
     session = await get_session(session_id)
+    if session and session.get("rpg_enabled"):
+        persona_id_pre = (session.get("persona_id") or "default")
+        persona_pre = await get_persona(persona_id_pre) or {}
+        await reconcile_plot_arc(
+            session_id,
+            persona_name=persona_pre.get("name", persona_id_pre),
+            recent_context=(session.get("status_quo") or "")[:2000],
+            genre=session.get("genre") or "adventure",
+        )
+        session = await get_session(session_id) or session
+    persona_id = (session.get("persona_id") if session else None) or "default"
+    persona = await get_persona(persona_id) or {}
+    system_msg = next((m for m in history if m.get("role") == "system"), None)
+    stored = system_msg.get("content") if system_msg else ""
+    live_static = await get_system_prompt(persona_id, history, "")
+    system_prompt = live_static if live_static else stored
     quests = await get_quests(session_id)
+    rpg_settings = get_rpg_settings(session) if session else DEFAULT_RPG_SETTINGS
+    facts_block = facts_to_prompt(session.get("facts_json", "[]")) if session else ""
+    runtime_suffix = ""
+    if session and session.get("rpg_enabled"):
+        runtime_suffix = build_rpg_runtime_suffix(session, rpg_settings, facts_block)
+    llm_system = system_prompt + runtime_suffix
+    context_usage = compute_payload_usage(history, llm_system)
     return {
-        "system_prompt": system_msg.get("content") if system_msg else "",
+        "persona_id": persona_id,
+        "persona_name": persona.get("name", persona_id),
+        "system_prompt": system_prompt,
         "status_quo": session.get("status_quo") if session else "",
+        "global_plot": session.get("global_plot") if session else "",
         "facts_json": session.get("facts_json") if session else "[]",
         "plot_arc_json": session.get("plot_arc_json") if session else "{}",
         "outfit_json": session.get("outfit_json") if session else "[]",
+        "scene_json": session.get("scene_json") if session else "{}",
+        "narrative_stats_json": session.get("narrative_stats_json") if session else "{}",
         "affinity": session.get("affinity", 0) if session else 0,
         "genre": session.get("genre", "") if session else "",
         "rpg_settings_json": session.get("rpg_settings_json") if session else "{}",
         "rpg_enabled": bool(session.get("rpg_enabled")) if session else False,
         "quests": quests,
+        "context_usage": context_usage,
     }
 
 
 @router.post("/init")
 async def init_chat(request: ChatRequest):
-    persona_id = request.persona_id or "default"
-    await get_or_create_session(request.session_id, persona_id)
+    await get_or_create_session(
+        request.session_id,
+        request.persona_id or "default",
+    )
+    persona_id = await resolve_session_persona(
+        request.session_id,
+        request.persona_id,
+        create_persona=request.persona_id,
+    )
     history = await get_history(request.session_id)
     if history:
         return {"first_mes": None}
 
     system_prompt = await get_system_prompt(persona_id, [], "")
-    await add_message(request.session_id, "system", system_prompt)
+    await upsert_static_system_message(request.session_id, system_prompt, [])
 
     first_mes = None
     if request.first_mes_override and request.first_mes_override.strip():
@@ -152,53 +245,67 @@ class RpgBootstrapRequest(BaseModel):
     genre: str = "adventure"
 
 
+class OpeningProcessRequest(BaseModel):
+    session_id: str
+    persona_id: str = "default"
+    rpg: bool = False
+
+
+@router.post("/opening/process")
+async def opening_process(req: OpeningProcessRequest):
+    await get_or_create_session(req.session_id, req.persona_id)
+    persona_id = await resolve_session_persona(req.session_id, req.persona_id)
+    try:
+        return await process_opening(req.session_id, persona_id, rpg=req.rpg)
+    except ValueError as e:
+        raise HTTPException(status_code=400, detail=str(e))
+
+
 @router.post("/rpg/bootstrap")
 async def rpg_bootstrap(req: RpgBootstrapRequest):
     await get_or_create_session(req.session_id, req.persona_id)
-    session = await get_session(req.session_id)
-    persona = await get_persona(req.persona_id) or {}
-
-    # Save genre
+    persona_id = await resolve_session_persona(req.session_id, req.persona_id)
     await update_session_genre(req.session_id, req.genre)
-
-    arc_json = (session.get("plot_arc_json") or "{}") if session else "{}"
-    try:
-        arc = json.loads(arc_json) if isinstance(arc_json, str) else {}
-    except Exception:
-        arc = {}
-    if not arc:
-        facts_block = facts_to_prompt((session or {}).get("facts_json", "[]"))
-        arc = await generate_plot_arc(
-            persona.get("name", req.persona_id),
-            persona.get("description", ""),
-            persona.get("scenario", ""),
-            persona.get("first_mes", ""),
-            facts_block=facts_block,
-            genre=req.genre,
+    persona = await get_persona(persona_id) or {}
+    greeting = await resolve_greeting(req.session_id, persona)
+    arc = await ensure_plot_arc_and_quests(req.session_id, persona, greeting, req.genre)
+    session = await get_session(req.session_id) or {}
+    rpg_settings = get_rpg_settings(session)
+    if rpg_settings.get("narrator", True) and greeting:
+        arc_json = json.dumps(arc, ensure_ascii=False) if arc else ""
+        facts_block = facts_to_prompt(session.get("facts_json", "[]"))
+        post = await narrator_post(
+            persona.get("name", persona_id),
+            f"assistant: {greeting}",
+            arc_json,
+            facts_block,
+            is_opening=True,
         )
-        if arc:
-            from services.memory import update_session_plot_arc
-            await update_session_plot_arc(req.session_id, json.dumps(arc, ensure_ascii=False))
-
-            # Seed quests from beats
-            for beat in arc.get("beats", []):
-                title = (beat.get("title") or beat.get("injection", "")).strip()
-                if title:
-                    await upsert_quest(req.session_id, title[:120])
-
+        await apply_narrator_post(req.session_id, post, rpg_settings, session)
     quests = await get_quests(req.session_id)
-    return {"plot_arc": arc, "quests": quests}
+    updated = await get_session(req.session_id) or {}
+    return {
+        "plot_arc": arc,
+        "quests": quests,
+        "affinity": updated.get("affinity", 0),
+        "scene_json": updated.get("scene_json", "{}"),
+        "narrative_stats_json": updated.get("narrative_stats_json", "{}"),
+    }
 
 
 @router.post("/stream")
 async def chat_stream(request: ChatRequest):
-    persona_id = request.persona_id or "default"
-
-    await get_or_create_session(request.session_id, persona_id)
+    await get_or_create_session(request.session_id, request.persona_id)
+    persona_id = await resolve_session_persona(
+        request.session_id,
+        request.persona_id,
+        create_persona=request.persona_id,
+    )
 
     history = await get_history(request.session_id)
     session = await get_session(request.session_id)
-    system_prompt = await get_system_prompt(persona_id, history, request.message)
+    static_prompt = await get_system_prompt(persona_id, history, request.message)
+    runtime_suffix = ""
 
     arc = {}
     roll = None
@@ -206,26 +313,24 @@ async def chat_stream(request: ChatRequest):
     resolution_text = ""
     narrator_msg = None  # shown as narrator bubble before assistant reply
     rpg_settings = {}
+    facts_block = ""
 
+    narrator_extra = ""
+    pre = {}
+    directives: list = []
+    pre_ok = False
     if session and session.get("rpg_enabled"):
         rpg_settings = get_rpg_settings(session)
         facts_block = facts_to_prompt(session.get("facts_json", "[]"))
-        if facts_block:
-            system_prompt = system_prompt + "\n\n" + facts_block
         try:
             arc = json.loads(session.get("plot_arc_json") or "{}")
         except Exception:
             arc = {}
-        if arc:
-            system_prompt = system_prompt + "\n\n--- PlotArc ---\n" + json.dumps(
-                {k: arc.get(k) for k in ("title", "phase", "next_beat_hint")}, ensure_ascii=False
-            ) + "\n---"
-        status_quo = (session.get("status_quo") or "").strip()
-        if status_quo:
-            system_prompt = system_prompt + "\n\n--- Status quo ---\n" + status_quo + "\n---"
-        if rpg_settings.get("affinity", True):
-            aff = int(session.get("affinity") or 0)
-            system_prompt = system_prompt + affinity_prompt_block(aff)
+
+        quests_list = await get_quests(request.session_id)
+        narr_ctx = format_narrator_context(
+            arc, quests_list, session.get("status_quo") or ""
+        )
 
         if rpg_settings.get("narrator", True):
             persona = await get_persona(persona_id) or {}
@@ -241,7 +346,9 @@ async def chat_stream(request: ChatRequest):
                 json.dumps(arc, ensure_ascii=False) if arc else "",
                 facts_block,
                 request.message,
+                extra_context=narr_ctx,
             )
+            pre_ok = bool(pre.get("_ok"))
 
             needs_check = pre.get("needs_check", False) and rpg_settings.get("dice", True)
 
@@ -265,6 +372,7 @@ async def chat_stream(request: ChatRequest):
                     request.message,
                     roll=roll,
                     outcome=outcome,
+                    extra_context=narr_ctx,
                 )
                 resolution_text = (pre2.get("resolution_text") or "").strip()
                 directives = pre2.get("directives") or []
@@ -274,66 +382,95 @@ async def chat_stream(request: ChatRequest):
                 pre_sq = (pre.get("status_quo_update") or "").strip()
 
             if directives:
-                system_prompt = system_prompt + "\n\n--- Narrator directives ---\n" + "\n".join(f"- {d}" for d in directives) + "\n---"
+                narrator_extra += (
+                    "\n\n--- Narrator directives ---\n"
+                    + "\n".join(f"- {d}" for d in directives)
+                    + "\n---"
+                )
             if pre_sq:
                 await update_session_status_quo(request.session_id, pre_sq)
                 session["status_quo"] = pre_sq
 
+            pre_for_scene = pre2 if needs_check else pre
+            scene_up = pre_for_scene.get("scene_update")
+            if isinstance(scene_up, dict) and scene_up:
+                from services.rpg_state import merge_scene
+                from services.memory import update_session_scene
+
+                merged = merge_scene(
+                    parse_scene_json(session.get("scene_json")), scene_up
+                )
+                scene_str = json.dumps(merged, ensure_ascii=False)
+                await update_session_scene(request.session_id, scene_str)
+                session["scene_json"] = scene_str
+
             if resolution_text:
-                await add_action_resolution(
-                    request.session_id,
-                    intent_text=request.message,
-                    roll=roll,
-                    outcome=outcome,
-                    resolution_text=resolution_text,
-                    message_id=None,
-                )
-                narrator_msg = {"roll": roll, "outcome": outcome, "text": resolution_text}
+                narrator_msg = {
+                    "roll": roll,
+                    "outcome": outcome,
+                    "text": resolution_text,
+                    "original_intent": request.message,
+                }
 
-            # Inject outcome into system prompt so character reply is consistent
-            if roll is not None:
-                system_prompt = (
-                    system_prompt
-                    + f"\n\n--- Mechanics ---\n"
-                    + f"Roll d20={roll}. Outcome: {outcome}.\n"
-                    + "Your reply MUST be consistent with this outcome. Do NOT contradict the narrator resolution.\n"
-                    + "---"
+            if roll is not None and resolution_text:
+                narrator_extra += (
+                    f"\n\n--- Mechanics (this turn) ---\n"
+                    f"Roll d20={roll}. Outcome: {outcome}.\n"
+                    f"Narrator resolution: {resolution_text}\n"
+                    "The character's next reply MUST match the narrator ruling in the message history "
+                    "(immediately after the player's intent). Do NOT re-enact the attempt as full success on failure.\n"
+                    "---"
                 )
 
-    # is_narrator_choice: wrap message so LLM understands context
+        runtime_suffix = build_rpg_runtime_suffix(session, rpg_settings, facts_block) + narrator_extra
+
+    llm_system = static_prompt + runtime_suffix
+    if persona_id != "default" or (session and session.get("rpg_enabled")):
+        llm_system += RP_OUTPUT_REMINDER
+
     user_message_content = request.message
     if request.is_narrator_choice:
         user_message_content = f"[Player chose: {request.message}]"
 
-    if not history:
-        await add_message(request.session_id, "system", system_prompt)
-    elif history[0]["role"] == "system" and history[0]["content"] != system_prompt:
-        async with aiosqlite.connect(DB_PATH) as db:
-            await db.execute(
-                """UPDATE messages SET content = ?
-                   WHERE session_id = ? AND role = 'system'
-                   AND id = (SELECT MIN(id) FROM messages WHERE session_id = ?)""",
-                (system_prompt, request.session_id, request.session_id),
-            )
-            await db.commit()
+    await upsert_static_system_message(request.session_id, static_prompt, history)
 
+    user_msg_id = None
     if not request.skip_user_add:
-        await add_message(request.session_id, "user", user_message_content)
+        await clear_choices_for_session(request.session_id)
+        user_msg_id = await add_message(request.session_id, "user", user_message_content)
+    if narrator_msg and narrator_msg.get("roll") is not None and user_msg_id:
+        await add_action_resolution(
+            request.session_id,
+            intent_text=request.message,
+            roll=narrator_msg["roll"],
+            outcome=narrator_msg["outcome"],
+            resolution_text=narrator_msg["text"],
+            message_id=user_msg_id,
+        )
+        narrator_msg["user_message_id"] = user_msg_id
+    if narrator_msg and (narrator_msg.get("text") or "").strip():
+        await add_message(
+            request.session_id,
+            "narrator",
+            narrator_message_content(narrator_msg),
+        )
     messages = await get_history(request.session_id)
+    usage = compute_payload_usage(messages, llm_system)
+    warn = context_warning_line(usage.get("percent", 0))
+    if warn:
+        llm_system += warn
+    llm_messages = messages_for_llm(messages, llm_system)
 
     full_reply = []
 
     async def generate():
         nonlocal arc
 
-        # Send narrator BEFORE streaming so it appears above the reply
         if narrator_msg:
             yield f"data: {json.dumps({'narrator': narrator_msg})}\n\n"
 
         try:
-            async for chunk in stream_message(
-                [{"role": m["role"], "content": m["content"]} for m in messages]
-            ):
+            async for chunk in stream_message(llm_messages):
                 full_reply.append(chunk)
                 yield f"data: {json.dumps({'chunk': chunk})}\n\n"
         except Exception as e:
@@ -342,99 +479,176 @@ async def chat_stream(request: ChatRequest):
             return
 
         complete = "".join(full_reply)
-        display_text = strip_image_prompt_tag(complete)
+        raw_display = strip_image_prompt_tag(complete)
+        display_text = strip_ooc_from_reply(raw_display)
 
-        hist_with_reply = await get_history(request.session_id) + [
-            {"role": "assistant", "content": display_text}
-        ]
-        sd_result = await generate_sd_prompt(
-            hist_with_reply, persona_id,
-            outfit_json=session.get("outfit_json", "[]") if session else "[]"
-        )
-        prompt_str = (sd_result[0] if sd_result and sd_result[0] else None) or extract_image_prompt_tag(complete)
-
-        if (display_text or complete).strip():
-            await add_message(request.session_id, "assistant", display_text or complete, image_prompt=prompt_str)
+        if (display_text or raw_display).strip():
+            await add_message(request.session_id, "assistant", display_text or raw_display)
 
         choices = []
         debug_blocks = []
         quests_updated = []
+        narrator_meta = {}
 
         if session and session.get("rpg_enabled"):
-            if not arc:
-                persona = await get_persona(persona_id) or {}
-                arc = await generate_plot_arc(
-                    persona.get("name", persona_id),
-                    persona.get("description", ""),
-                    persona.get("scenario", ""),
-                    persona.get("first_mes", ""),
-                    facts_block=facts_to_prompt(session.get("facts_json", "[]")),
-                    genre=session.get("genre") or "adventure",
-                )
+            try:
+                if not arc:
+                    persona = await get_persona(persona_id) or {}
+                    arc = await generate_plot_arc(
+                        persona.get("name", persona_id),
+                        persona.get("description", ""),
+                        persona.get("scenario", ""),
+                        persona.get("first_mes", ""),
+                        facts_block=facts_to_prompt(session.get("facts_json", "[]")),
+                        genre=session.get("genre") or "adventure",
+                    )
+                    if arc:
+                        await update_session_plot_arc(
+                            request.session_id, json.dumps(arc, ensure_ascii=False)
+                        )
+                        debug_blocks.append({
+                            "type": "plot_arc",
+                            "text": json.dumps(arc, ensure_ascii=False, indent=2),
+                        })
+                        if rpg_settings.get("quests", True):
+                            await seed_quests_from_arc(request.session_id, arc)
+
+                quests_list = await get_quests(request.session_id)
                 if arc:
-                    await update_session_plot_arc(request.session_id, json.dumps(arc, ensure_ascii=False))
-                    debug_blocks.append({"type": "plot_arc", "text": json.dumps(arc, ensure_ascii=False, indent=2)})
-                    if rpg_settings.get("quests", True):
-                        for beat in arc.get("beats", []):
-                            t = (beat.get("title") or beat.get("injection", "")).strip()
-                            if t:
-                                await upsert_quest(request.session_id, t[:120])
+                    beat_ctx = "\n".join(
+                        f"{m['role']}: {m['content']}"
+                        for m in (await get_history(request.session_id))[-6:]
+                        if m.get("role") in ("user", "assistant")
+                    )
+                    arc, beats, pruned, beat_mode = await process_arc_beats(
+                        arc,
+                        quests_list,
+                        request.message,
+                        recent_context=beat_ctx,
+                        last_dice_outcome=outcome if roll is not None else None,
+                    )
+                    if pruned or beats:
+                        await update_session_plot_arc(
+                            request.session_id, json.dumps(arc, ensure_ascii=False)
+                        )
+                    if pruned:
+                        debug_blocks.append({
+                            "type": "plot_arc_prune",
+                            "text": f"Removed {len(pruned)} beat(s) already completed as quests",
+                        })
+                    if beats:
+                        inj = beats[0].get("injection", "")
+                        if inj:
+                            debug_blocks.append({"type": "narrator_injection", "text": inj})
+                        if rpg_settings.get("choices", True):
+                            choices += choices_from_beat(beats[0])
+                        if beat_mode in ("after_dice", "llm", "trigger", "stuck_recovery"):
+                            debug_blocks.append({
+                                "type": "plot_arc",
+                                "text": (
+                                    f"Beat fired ({beat_mode}): "
+                                    f"«{beats[0].get('title', '')}»"
+                                ),
+                            })
+                    if advance_phase(arc):
+                        await update_session_plot_arc(
+                            request.session_id, json.dumps(arc, ensure_ascii=False)
+                        )
+                        debug_blocks.append({"type": "phase_advance", "text": arc["phase"]})
+                    if pruned and not arc.get("beats"):
+                        narrator_meta["arc_pruned"] = len(pruned)
+                    if beat_mode:
+                        narrator_meta["beat_mode"] = beat_mode
 
-            trig = should_advance_arc(request.message)
-            if trig and arc:
-                arc, beats = pop_matching_beats(arc, trig, max_beats=1)
-                if beats:
-                    await update_session_plot_arc(request.session_id, json.dumps(arc, ensure_ascii=False))
-                    inj = beats[0].get("injection", "")
-                    if inj:
-                        debug_blocks.append({"type": "narrator_injection", "text": inj})
-                    if rpg_settings.get("choices", True):
-                        choices += beats[0].get("choices") or []
-                if advance_phase(arc):
-                    await update_session_plot_arc(request.session_id, json.dumps(arc, ensure_ascii=False))
-                    debug_blocks.append({"type": "phase_advance", "text": arc["phase"]})
+                ctx = [
+                    m for m in (await get_history(request.session_id))
+                    if m["role"] in ("user", "assistant")
+                ][-10:]
+                new_facts = await extract_facts(
+                    ctx,
+                    rp_day_hint=rp_day_from_scene(session.get("scene_json")),
+                    existing_json=session.get("facts_json", "[]"),
+                )
+                if new_facts:
+                    merged = await merge_facts_persist(
+                        session.get("facts_json", "[]"),
+                        new_facts,
+                        rp_day_default=rp_day_from_scene(session.get("scene_json")),
+                        scene_context=json.dumps(
+                            parse_scene_json(session.get("scene_json")),
+                            ensure_ascii=False,
+                        ),
+                        status_quo=session.get("status_quo") or "",
+                    )
+                    await update_session_facts(request.session_id, merged)
+                    session["facts_json"] = merged
 
-            ctx = [m for m in (await get_history(request.session_id)) if m["role"] in ("user", "assistant")][-10:]
-            new_facts = await extract_facts(ctx)
-            if new_facts:
-                merged = merge_facts(session.get("facts_json", "[]"), new_facts)
-                await update_session_facts(request.session_id, merged)
-                session["facts_json"] = merged
+                persona = await get_persona(persona_id) or {}
+                ctx_txt = "\n".join(
+                    f"{m['role']}: {m['content']}"
+                    for m in ctx[-8:]
+                    if m.get("role") in ("user", "assistant")
+                )
+                narr_ctx_post = format_narrator_context(
+                    arc, await get_quests(request.session_id), session.get("status_quo") or ""
+                )
+                post = await narrator_post(
+                    persona.get("name", persona_id),
+                    ctx_txt,
+                    json.dumps(arc, ensure_ascii=False) if arc else "",
+                    facts_to_prompt(session.get("facts_json", "[]")),
+                    extra_context=narr_ctx_post,
+                )
 
-            persona = await get_persona(persona_id) or {}
-            ctx_txt = "\n".join(f"{m['role']}: {m['content']}" for m in ctx[-8:] if m.get("role") in ("user", "assistant"))
-            post = await narrator_post(
-                persona.get("name", persona_id),
-                ctx_txt,
-                json.dumps(arc, ensure_ascii=False) if arc else "",
-                facts_to_prompt(session.get("facts_json", "[]")),
-            )
+                sq = (post.get("status_quo_update") or "").strip()
+                if sq:
+                    debug_blocks.append({"type": "status_quo", "text": sq})
 
-            sq = (post.get("status_quo_update") or "").strip()
-            if sq:
-                await update_session_status_quo(request.session_id, sq)
-                debug_blocks.append({"type": "status_quo", "text": sq})
+                if rpg_settings.get("choices", True):
+                    choices += choices_from_narrator(post.get("choices") or [])
 
-            if rpg_settings.get("choices", True):
-                choices += post.get("choices") or []
+                applied = await apply_narrator_post(
+                    request.session_id, post, rpg_settings, session
+                )
+                narrator_meta = {
+                    "pre_ok": pre_ok,
+                    "post_ok": bool(post.get("_ok")),
+                    "choices_count": len(choices),
+                    "directives_count": len(directives),
+                    "dice": roll is not None,
+                    **applied,
+                }
 
-            if rpg_settings.get("affinity", True):
-                delta = int(post.get("affinity_delta") or 0)
-                if delta:
-                    await update_session_affinity(request.session_id, delta)
+                if not arc.get("beats"):
+                    persona = await get_persona(persona_id) or {}
+                    arc = await replenish_arc_beats(
+                        arc,
+                        persona.get("name", persona_id),
+                        ctx_txt,
+                        await get_quests(request.session_id),
+                        session.get("genre") or "adventure",
+                    )
+                    if arc.get("beats"):
+                        await update_session_plot_arc(
+                            request.session_id, json.dumps(arc, ensure_ascii=False)
+                        )
+                        debug_blocks.append({
+                            "type": "plot_arc",
+                            "text": f"Added {len(arc.get('beats', []))} new plot beats",
+                        })
+                        narrator_meta["beats_replenished"] = len(arc.get("beats", []))
+                        if rpg_settings.get("quests", True):
+                            await seed_quests_from_arc(request.session_id, arc)
+                outfit_update = post.get("outfit_update")
+                if isinstance(outfit_update, list) and outfit_update:
+                    from services.outfit_tags import outfit_list_to_json
 
-            outfit_update = post.get("outfit_update")
-            if isinstance(outfit_update, list) and outfit_update:
-                outfit_str = json.dumps(outfit_update, ensure_ascii=False)
-                await update_session_outfit(request.session_id, outfit_str)
-                session["outfit_json"] = outfit_str
-
-            if rpg_settings.get("quests", True):
-                for qu in (post.get("quest_updates") or []):
-                    t = (qu.get("title") or "").strip()
-                    if t:
-                        await upsert_quest(request.session_id, t[:120], qu.get("status", "active"))
+                    session["outfit_json"] = outfit_list_to_json(outfit_update)
                 quests_updated = await get_quests(request.session_id)
+            except LLMError as e:
+                logger.warning("RPG post-process skipped after reply: %s", e)
+            except Exception as e:
+                logger.exception("RPG post-process failed after reply: %s", e)
 
         count = await get_message_count(request.session_id)
         if count == 2 and not request.skip_user_add:
@@ -443,23 +657,63 @@ async def chat_stream(request: ChatRequest):
             if (session or {}).get("title", "Новый чат") in ("", "Новый чат"):
                 await update_session_title(request.session_id, f"{persona.get('name', persona_id)} — {preview}")
 
-        image_path = None
-        image_error = None
-        if prompt_str and SD_AUTO_GENERATE:
+        updated_session = await get_session(request.session_id) or session
+        hist = await get_history(request.session_id)
+        bundle = await generate_sd_prompt(
+            hist,
+            persona_id,
+            outfit_json=updated_session.get("outfit_json", "[]") if updated_session else "[]",
+            scene_json=updated_session.get("scene_json", "{}") if updated_session else "{}",
+        )
+        prompt_str = bundle.tag_full if bundle else extract_image_prompt_tag(complete)
+        msg_id = await get_last_assistant_message_id(request.session_id)
+        if msg_id and choices:
+            await update_message_choices(
+                msg_id, json.dumps(choices, ensure_ascii=False)
+            )
+
+        sd_out: dict = {}
+        if bundle:
+            yield f"data: {json.dumps({
+                'image_generating': True,
+                'image_prompt': bundle.tag_full,
+                'image_prompt_alt': bundle.desc_full,
+            })}\n\n"
+            sd_out = await run_sd_for_message(bundle, msg_id)
+        elif prompt_str and SD_AUTO_GENERATE:
             yield f"data: {json.dumps({'image_generating': True, 'image_prompt': prompt_str})}\n\n"
             rel, err = await sd_service.generate_from_full_prompt(prompt_str)
             if rel:
-                image_path = rel
-                msg_id = await get_last_assistant_message_id(request.session_id)
+                sd_out["image_path"] = f"/static/{rel}"
                 if msg_id:
                     await update_message_image(msg_id, rel)
             else:
-                image_error = err
+                sd_out["image_error"] = err
+            sd_out["image_prompt"] = prompt_str
 
-        updated_session = await get_session(request.session_id)
         affinity = updated_session.get("affinity", 0) if updated_session else 0
+        done_payload = {
+            "done": True,
+            "assistant_message_id": msg_id,
+            "assistant_content": display_text or raw_display,
+            "image_prompt": sd_out.get("image_prompt") or prompt_str,
+            "image_prompt_alt": sd_out.get("image_prompt_alt"),
+            "image_path": sd_out.get("image_path"),
+            "image_path_alt": sd_out.get("image_path_alt"),
+            "image_error": sd_out.get("image_error"),
+            "image_error_alt": sd_out.get("image_error_alt"),
+            "choices": choices,
+            "debug": debug_blocks,
+            "affinity": affinity,
+            "quests": quests_updated,
+            "narrator_meta": narrator_meta,
+        }
+        if rpg_settings.get("stats") and updated_session:
+            done_payload["narrative_stats"] = parse_stats_json(
+                updated_session.get("narrative_stats_json")
+            )
 
-        yield f"data: {json.dumps({'done': True, 'image_prompt': prompt_str, 'image_path': f'/static/{image_path}' if image_path else None, 'image_error': image_error, 'choices': choices, 'debug': debug_blocks, 'affinity': affinity, 'quests': quests_updated})}\n\n"
+        yield f"data: {json.dumps(done_payload)}\n\n"
 
     return StreamingResponse(
         generate(),
@@ -470,23 +724,37 @@ async def chat_stream(request: ChatRequest):
 
 @router.post("/", response_model=ChatResponse)
 async def chat(request: ChatRequest):
-    persona_id = request.persona_id or "default"
-    await get_or_create_session(request.session_id, persona_id)
+    await get_or_create_session(request.session_id, request.persona_id)
+    persona_id = await resolve_session_persona(
+        request.session_id,
+        request.persona_id,
+        create_persona=request.persona_id,
+    )
 
     history = await get_history(request.session_id)
-    system_prompt = await get_system_prompt(persona_id, history, request.message)
-
-    if not history:
-        await add_message(request.session_id, "system", system_prompt)
+    static_prompt = await get_system_prompt(persona_id, history, request.message)
+    await upsert_static_system_message(request.session_id, static_prompt, history)
 
     await add_message(request.session_id, "user", request.message)
     messages = await get_history(request.session_id)
-    reply = await send_message(
-        [{"role": m["role"], "content": m["content"]} for m in messages]
+    session = await get_session(request.session_id)
+    llm_system = static_prompt
+    if session and session.get("rpg_enabled"):
+        rpg_settings = get_rpg_settings(session)
+        facts_block = facts_to_prompt(session.get("facts_json", "[]"))
+        llm_system += build_rpg_runtime_suffix(session, rpg_settings, facts_block)
+    if persona_id != "default" or (session and session.get("rpg_enabled")):
+        llm_system += RP_OUTPUT_REMINDER
+    llm_messages = messages_for_llm(messages, llm_system)
+    reply = await send_message(llm_messages)
+    display = strip_ooc_from_reply(strip_image_prompt_tag(reply))
+    bundle = await generate_sd_prompt(
+        messages,
+        persona_id,
+        outfit_json=session.get("outfit_json", "[]") if session else "[]",
+        scene_json=session.get("scene_json", "{}") if session else "{}",
     )
-    display = strip_image_prompt_tag(reply)
-    prompt_tuple = await generate_sd_prompt(messages, persona_id)
-    prompt_str = prompt_tuple[0] if prompt_tuple else extract_image_prompt_tag(reply)
+    prompt_str = bundle.tag_full if bundle else extract_image_prompt_tag(reply)
 
     await add_message(request.session_id, "assistant", display, image_prompt=prompt_str)
 
@@ -497,6 +765,15 @@ async def chat(request: ChatRequest):
     )
 
 
+@router.delete("/messages/{message_id}")
+async def remove_message(message_id: int):
+    msg = await get_message(message_id)
+    if not msg:
+        raise HTTPException(status_code=404, detail="Сообщение не найдено")
+    await delete_message_and_following(msg["session_id"], message_id)
+    return {"status": "deleted", "message_id": message_id}
+
+
 @router.patch("/messages/{message_id}")
 async def edit_message(message_id: int, req: MessageEditRequest):
     msg = await get_message(message_id)
@@ -527,7 +804,6 @@ async def regenerate_chat(req: RegenerateRequest):
     stream_req = ChatRequest(
         message=user_text,
         session_id=req.session_id,
-        persona_id=req.persona_id,
         skip_user_add=True,
     )
     return await chat_stream(stream_req)
diff --git a/routers/debug.py b/routers/debug.py
new file mode 100644
index 0000000..af4532f
--- /dev/null
+++ b/routers/debug.py
@@ -0,0 +1,248 @@
+import json
+import os
+
+from fastapi import APIRouter, HTTPException
+from pydantic import BaseModel
+
+from services import sdbackend as sd_service
+from services.comfy_models import list_node_types, parse_model_lists
+from services.llm import (
+    CHAT_MODEL,
+    LLM_FALLBACK_MODEL,
+    LLMError,
+    SYSTEM_MODEL,
+    send_message,
+    send_message_with_model,
+)
+from services.personas import get_all_personas
+from services.sd_prompt import (
+    SD_PROMPT_MODEL,
+    anima_dual_enabled,
+    run_prompt_builder,
+)
+
+router = APIRouter(prefix="/debug", tags=["debug"])
+
+
+class ChatMessage(BaseModel):
+    role: str
+    content: str
+
+
+class SdPromptDebugRequest(BaseModel):
+    persona_id: str = "default"
+    chat_excerpt: str = ""
+    messages: list[ChatMessage] | None = None
+    outfit_json: str = "[]"
+    appearance_override: str | None = None
+    use_prose: bool = False
+
+
+class LlmDebugRequest(BaseModel):
+    model: str = ""
+    system: str = ""
+    user: str = ""
+    messages: list[ChatMessage] | None = None
+
+
+class ComfyRawRequest(BaseModel):
+    method: str = "GET"
+    path: str = "/system_stats"
+    params_json: str = "{}"
+    body_json: str = ""
+
+
+class ComfyGenerateRequest(BaseModel):
+    positive: str
+    negative: str = ""
+    unet: str | None = None
+    clip: str | None = None
+    vae: str | None = None
+    checkpoint: str | None = None
+
+
+@router.get("/config")
+async def debug_config():
+    base = sd_service.SD_BASE_URL
+    return {
+        "chat_model": CHAT_MODEL,
+        "system_model": SYSTEM_MODEL,
+        "llm_fallback_model": LLM_FALLBACK_MODEL,
+        "sd_prompt_model": SD_PROMPT_MODEL or SYSTEM_MODEL,
+        "sd_base_url": base,
+        "sd_has_token": bool(sd_service.SD_QUERY_PARAMS.get("token")),
+        "sd_anima_dual": anima_dual_enabled(),
+        "sd_unet": sd_service.SD_UNET,
+        "sd_clip": sd_service.SD_CLIP,
+        "sd_vae": sd_service.SD_VAE,
+        "sd_checkpoint": sd_service.SD_CHECKPOINT,
+        "sd_steps": sd_service.SD_STEPS,
+        "sd_cfg": sd_service.SD_CFG,
+        "router_key_set": bool(os.getenv("ROUTER_KEY")),
+    }
+
+
+@router.get("/personas")
+async def debug_personas():
+    personas = await get_all_personas()
+    return [
+        {
+            "persona_id": pid,
+            "name": p.get("name", pid),
+            "appearance_tags": p.get("appearance_tags", ""),
+        }
+        for pid, p in personas.items()
+    ]
+
+
+@router.post("/sd-prompt")
+async def debug_sd_prompt(req: SdPromptDebugRequest):
+    msgs = None
+    if req.messages:
+        msgs = [m.model_dump() for m in req.messages]
+    return await run_prompt_builder(
+        req.persona_id,
+        messages=msgs,
+        chat_excerpt=req.chat_excerpt,
+        outfit_json=req.outfit_json,
+        appearance_override=req.appearance_override,
+        use_prose=req.use_prose,
+    )
+
+
+@router.post("/llm")
+async def debug_llm(req: LlmDebugRequest):
+    if req.messages:
+        messages = [m.model_dump() for m in req.messages]
+    else:
+        messages = []
+        if req.system.strip():
+            messages.append({"role": "system", "content": req.system.strip()})
+        if req.user.strip():
+            messages.append({"role": "user", "content": req.user.strip()})
+    if not messages:
+        raise HTTPException(status_code=400, detail="Нужны messages или system/user")
+
+    model = (req.model or "").strip() or SD_PROMPT_MODEL or SYSTEM_MODEL
+    try:
+        if model in (SYSTEM_MODEL, "") and not req.model:
+            text = await send_message(messages)
+        else:
+            text = await send_message_with_model(messages, model)
+        return {"model": model, "response": text}
+    except LLMError as e:
+        raise HTTPException(status_code=502, detail=str(e))
+
+
+@router.get("/comfy/ping")
+async def debug_comfy_ping():
+    try:
+        status, body, headers = await sd_service.comfy_api_request("GET", "/system_stats")
+        return {"ok": status == 200, "status": status, "body": body, "headers": headers}
+    except Exception as e:
+        return {"ok": False, "error": str(e)}
+
+
+@router.get("/comfy/models")
+async def debug_comfy_models():
+    try:
+        info = await sd_service.fetch_object_info()
+        return {
+            "models": parse_model_lists(info),
+            "configured": {
+                "unet": sd_service.SD_UNET,
+                "clip": sd_service.SD_CLIP,
+                "vae": sd_service.SD_VAE,
+                "checkpoint": sd_service.SD_CHECKPOINT,
+            },
+            "node_type_count": len(list_node_types(info)),
+        }
+    except Exception as e:
+        raise HTTPException(status_code=502, detail=str(e))
+
+
+@router.get("/comfy/object_info")
+async def debug_comfy_object_info(node: str | None = None):
+    try:
+        info = await sd_service.fetch_object_info()
+        if node:
+            if node not in info:
+                raise HTTPException(status_code=404, detail=f"Unknown node: {node}")
+            return {node: info[node]}
+        return {
+            "node_types": list_node_types(info),
+            "models": parse_model_lists(info),
+        }
+    except HTTPException:
+        raise
+    except Exception as e:
+        raise HTTPException(status_code=502, detail=str(e))
+
+
+@router.post("/comfy/raw")
+async def debug_comfy_raw(req: ComfyRawRequest):
+    path = req.path.strip()
+    if not path.startswith("/"):
+        path = "/" + path
+    try:
+        params = json.loads(req.params_json or "{}")
+        if not isinstance(params, dict):
+            raise ValueError("params_json must be object")
+    except json.JSONDecodeError as e:
+        raise HTTPException(status_code=400, detail=f"params_json: {e}")
+
+    body = None
+    if req.body_json.strip():
+        try:
+            body = json.loads(req.body_json)
+        except json.JSONDecodeError as e:
+            raise HTTPException(status_code=400, detail=f"body_json: {e}")
+
+    method = req.method.upper()
+    if method not in ("GET", "POST", "PUT", "DELETE"):
+        raise HTTPException(status_code=400, detail="method must be GET|POST|PUT|DELETE")
+
+    try:
+        status, resp_body, headers = await sd_service.comfy_api_request(
+            method,
+            path,
+            params=params or None,
+            json_body=body,
+            timeout=120,
+        )
+        return {"status": status, "headers": headers, "body": resp_body}
+    except Exception as e:
+        raise HTTPException(status_code=502, detail=str(e))
+
+
+@router.post("/comfy/generate")
+async def debug_comfy_generate(req: ComfyGenerateRequest):
+    if not req.positive.strip():
+        raise HTTPException(status_code=400, detail="positive required")
+
+    overrides: dict[str, str] = {}
+    if req.unet:
+        overrides["unet"] = req.unet
+    if req.clip:
+        overrides["clip"] = req.clip
+    if req.vae:
+        overrides["vae"] = req.vae
+    if req.checkpoint:
+        overrides["checkpoint"] = req.checkpoint
+
+    full = req.positive.strip()
+    if req.negative.strip():
+        full += f"\n\nNegative prompt: {req.negative.strip()}"
+
+    try:
+        rel, err = await sd_service.generate_from_full_prompt(
+            full,
+            overrides=overrides or None,
+        )
+        if not rel:
+            raise HTTPException(status_code=502, detail=err or "generation failed")
+        return {"image_path": f"/static/{rel}", "status": "ok"}
+    except HTTPException:
+        raise
+    except Exception as e:
+        raise HTTPException(status_code=502, detail=str(e))
diff --git a/routers/personas.py b/routers/personas.py
index 750a90b..1ce2e7a 100644
--- a/routers/personas.py
+++ b/routers/personas.py
@@ -57,6 +57,7 @@ class PersonaPatch(BaseModel):
     lora_name: Optional[str] = None
     lora_weight: Optional[float] = None
     appearance_tags: Optional[str] = None
+    appearance_prose: Optional[str] = None
     personality: Optional[str] = None
     scenario: Optional[str] = None
     first_mes: Optional[str] = None
diff --git a/routers/sessions.py b/routers/sessions.py
index c2a38b5..cdbcdd6 100644
--- a/routers/sessions.py
+++ b/routers/sessions.py
@@ -1,8 +1,11 @@
+import json
+
 from fastapi import APIRouter, HTTPException
 from services.memory import (
     get_all_sessions,
     get_session,
     get_or_create_session,
+    get_history,
     delete_session,
     update_session_title,
     update_session_persona,
@@ -14,10 +17,27 @@ from services.memory import (
     update_session_genre,
     update_session_rpg_settings,
     get_quests,
+    update_quest_by_id,
+    set_session_affinity,
+    update_session_narrative_stats,
+    update_session_outfit,
+    update_session_scene,
+    update_session_plot_arc,
     get_last_message_preview,
     fork_session,
 )
-from models.schemas import ForkSessionRequest
+from models.schemas import (
+    ForkSessionRequest,
+    RebindPersonaRequest,
+    QuestStatusPatch,
+    RpgStateDebugPatch,
+    SessionContextPatch,
+)
+from services.rpg_plot import reconcile_plot_arc
+from services.rpg_state import parse_stats_json
+from services.chat_prompt import get_system_prompt
+from services.memory import rebind_session_persona
+from services.personas import get_persona
 
 router = APIRouter(prefix="/sessions", tags=["sessions"])
 
@@ -35,9 +55,149 @@ async def list_sessions():
 
 @router.get("/{session_id}/quests")
 async def list_quests(session_id: str):
+    session = await get_session(session_id)
+    if session and session.get("rpg_enabled"):
+        persona = await get_persona(session.get("persona_id") or "default") or {}
+        await reconcile_plot_arc(
+            session_id,
+            persona_name=persona.get("name", session.get("persona_id") or "Character"),
+            recent_context=(session.get("status_quo") or "")[:2000],
+            genre=session.get("genre") or "adventure",
+        )
     return await get_quests(session_id)
 
 
+@router.patch("/{session_id}/context")
+async def patch_session_context(session_id: str, body: SessionContextPatch):
+    """Live-edit session context (outfit, scene, plot, facts, status quo)."""
+    from services.outfit_tags import parse_and_normalize_outfit_json
+
+    session = await get_session(session_id)
+    if not session:
+        raise HTTPException(status_code=404, detail="Сессия не найдена")
+
+    if body.status_quo is not None:
+        await update_session_status_quo(session_id, body.status_quo)
+    if body.global_plot is not None:
+        await update_session_global_plot(session_id, body.global_plot)
+    if body.outfit_json is not None:
+        try:
+            normalized = parse_and_normalize_outfit_json(body.outfit_json)
+            json.loads(normalized)
+        except json.JSONDecodeError as e:
+            raise HTTPException(status_code=400, detail=f"outfit_json: {e}") from e
+        await update_session_outfit(session_id, normalized)
+    if body.scene_json is not None:
+        try:
+            json.loads(body.scene_json or "{}")
+        except json.JSONDecodeError as e:
+            raise HTTPException(status_code=400, detail=f"scene_json: {e}") from e
+        await update_session_scene(session_id, body.scene_json)
+    if body.facts_json is not None:
+        from services.rpg_facts import (
+            parse_facts_list,
+            facts_list_to_json,
+            dedupe_facts_fuzzy,
+            compress_facts,
+            FACTS_DEDUP_THRESHOLD,
+            FACTS_COMPRESS_TARGET,
+        )
+
+        try:
+            facts = dedupe_facts_fuzzy(parse_facts_list(body.facts_json))
+        except json.JSONDecodeError as e:
+            raise HTTPException(status_code=400, detail=f"facts_json: {e}") from e
+        if len(facts) > FACTS_DEDUP_THRESHOLD:
+            facts = await compress_facts(
+                facts,
+                status_quo=(session.get("status_quo") or ""),
+                scene_context=session.get("scene_json") or "{}",
+                target=FACTS_COMPRESS_TARGET,
+            )
+            facts = dedupe_facts_fuzzy(facts)
+        normalized = facts_list_to_json(facts)
+        await update_session_facts(session_id, normalized)
+    if body.plot_arc_json is not None:
+        try:
+            json.loads(body.plot_arc_json or "{}")
+        except json.JSONDecodeError as e:
+            raise HTTPException(status_code=400, detail=f"plot_arc_json: {e}") from e
+        await update_session_plot_arc(session_id, body.plot_arc_json)
+    if body.affinity is not None:
+        await set_session_affinity(session_id, body.affinity)
+    stats_changed = any(
+        getattr(body, k) is not None for k in ("lust", "stamina", "tension")
+    )
+    if stats_changed:
+        stats = parse_stats_json(session.get("narrative_stats_json"))
+        for key in ("lust", "stamina", "tension"):
+            val = getattr(body, key, None)
+            if val is not None:
+                stats[key] = max(0, min(10, int(val)))
+        await update_session_narrative_stats(
+            session_id, json.dumps(stats, ensure_ascii=False)
+        )
+
+    updated = await get_session(session_id) or session
+    return {
+        "status": "updated",
+        "outfit_json": updated.get("outfit_json", "[]"),
+        "scene_json": updated.get("scene_json", "{}"),
+        "affinity": updated.get("affinity", 0),
+        "narrative_stats": parse_stats_json(updated.get("narrative_stats_json")),
+    }
+
+
+@router.patch("/{session_id}/rpg-state")
+async def patch_rpg_state(session_id: str, body: RpgStateDebugPatch):
+    """Debug: set affinity and/or narrative stats (lust/stamina/tension 0–10)."""
+    session = await get_session(session_id)
+    if not session:
+        raise HTTPException(status_code=404, detail="Сессия не найдена")
+    affinity = session.get("affinity", 0)
+    if body.affinity is not None:
+        affinity = await set_session_affinity(session_id, body.affinity)
+    stats = parse_stats_json(session.get("narrative_stats_json"))
+    changed_stats = False
+    for key in ("lust", "stamina", "tension"):
+        val = getattr(body, key, None)
+        if val is not None:
+            stats[key] = max(0, min(10, int(val)))
+            changed_stats = True
+    if changed_stats:
+        import json as _json
+
+        await update_session_narrative_stats(
+            session_id, _json.dumps(stats, ensure_ascii=False)
+        )
+    return {
+        "affinity": affinity,
+        "narrative_stats": stats,
+        "target": "current_player",
+    }
+
+
+@router.patch("/{session_id}/quests/{quest_id}")
+async def patch_quest(session_id: str, quest_id: int, body: QuestStatusPatch):
+    status = body.status.strip()
+    if status not in ("active", "done", "failed"):
+        raise HTTPException(status_code=400, detail="status must be active, done, or failed")
+    ok = await update_quest_by_id(quest_id, session_id, status)
+    if not ok:
+        raise HTTPException(status_code=404, detail="Quest not found")
+    if status in ("done", "failed"):
+        session = await get_session(session_id)
+        if session and session.get("rpg_enabled"):
+            persona = await get_persona(session.get("persona_id") or "default") or {}
+            await reconcile_plot_arc(
+                session_id,
+                persona_name=persona.get("name", session.get("persona_id") or "Character"),
+                recent_context=(session.get("status_quo") or "")[:2000],
+                genre=session.get("genre") or "adventure",
+            )
+    return {"status": "updated", "quest_id": quest_id, "new_status": status}
+
+
 @router.get("/{session_id}")
 async def get_session_route(session_id: str):
     s = await get_session(session_id)
@@ -46,9 +206,42 @@ async def get_session_route(session_id: str):
     return s
 
 
+@router.post("/{session_id}/rebind-persona")
+async def rebind_persona(session_id: str, body: RebindPersonaRequest):
+    session = await get_session(session_id)
+    if not session:
+        raise HTTPException(status_code=404, detail="Сессия не найдена")
+    persona = await get_persona(body.persona_id)
+    if not persona:
+        raise HTTPException(status_code=400, detail="Персонаж не найден")
+
+    hist = [] if body.clear_history else await get_history(session_id)
+    static = await get_system_prompt(body.persona_id, hist, "")
+    first_mes = (persona.get("first_mes") or "").strip() if body.clear_history else None
+
+    try:
+        await rebind_session_persona(
+            session_id,
+            body.persona_id,
+            clear_history=body.clear_history,
+            static_prompt=static,
+            first_mes=first_mes or None,
+        )
+    except ValueError as e:
+        raise HTTPException(status_code=404, detail=str(e))
+
+    return {
+        "persona_id": body.persona_id,
+        "persona_name": persona.get("name", body.persona_id),
+        "system_prompt_preview": static[:500],
+        "clear_history": body.clear_history,
+    }
+
+
 @router.patch("/{session_id}")
 async def patch_session(session_id: str, data: dict):
-    await get_or_create_session(session_id, data.get("persona_id", "default"))
+    create_pid = data.get("persona_id") if "persona_id" in data else None
+    await get_or_create_session(session_id, create_pid)
     if "title" in data:
         await update_session_title(session_id, data["title"])
     if "persona_id" in data:
diff --git a/services/character_card.py b/services/character_card.py
index 2d0e398..8424505 100644
--- a/services/character_card.py
+++ b/services/character_card.py
@@ -45,6 +45,7 @@ def parse_card_v2(data: dict, card_id: str | None = None) -> dict:
         "first_mes": inner.get("first_mes", ""),
         "mes_example": inner.get("mes_example", ""),
         "appearance_tags": _extract_appearance(inner),
+        "appearance_prose": "",
         "lorebook_json": json.dumps(entries, ensure_ascii=False),
         "alternate_greetings": alternates,
         "alternate_greetings_json": json.dumps(alternates, ensure_ascii=False),
@@ -120,6 +121,8 @@ def parse_png_card(file_bytes: bytes) -> dict | None:
 
 
 def build_system_prompt(card: dict) -> str:
+    from services.chat_prompt import ROLEPLAY_GUARDRAILS
+
     parts = [
         f"You are {card['name']}. Stay in character.",
         f"Description: {card['description']}",
@@ -129,6 +132,7 @@ def build_system_prompt(card: dict) -> str:
     if card.get("mes_example"):
         parts.append(f"Example dialogue:\n{card['mes_example']}")
     parts.append("Reply only as the character. Do not add image tags.")
+    parts.append(ROLEPLAY_GUARDRAILS)
     return "\n\n".join(p for p in parts if p.split(": ", 1)[-1].strip())
 
 
@@ -141,13 +145,13 @@ async def save_character(card: dict, lora_name: str = "", lora_weight: float = 0
 
     async with aiosqlite.connect(DB_PATH) as db:
         await db.execute(
-            """INSERT OR REPLACE INTO characters
-               (card_id, name, description, personality, scenario, first_mes,
-                mes_example, raw_json, lora_name, lora_weight, appearance_tags, lorebook_json,
-                avatar_path, alternate_greetings_json)
-               VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)""",
+            """INSERT INTO characters
+               (card_id, name, description, personality, scenario, first_mes, mes_example,
+                raw_json, lora_name, lora_weight, appearance_tags, appearance_prose, lorebook_json, avatar_path,
+                alternate_greetings_json)
+               VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)""",
             (
-                card_id,
+                card["card_id"],
                 card["name"],
                 card["description"],
                 card["personality"],
@@ -157,10 +161,11 @@ async def save_character(card: dict, lora_name: str = "", lora_weight: float = 0
                 card["raw_json"],
                 lora_name,
                 lora_weight,
-                card.get("appearance_tags", ""),
+                card["appearance_tags"],
+                card.get("appearance_prose", ""),
                 card["lorebook_json"],
                 card.get("avatar_path", ""),
-                alt_json,
+                card.get("alternate_greetings_json", "[]"),
             ),
         )
         await db.commit()
@@ -199,8 +204,8 @@ async def delete_character(card_id: str) -> bool:
 async def update_appearance_tags(card_id: str, appearance_tags: str):
     async with aiosqlite.connect(DB_PATH) as db:
         await db.execute(
-            "UPDATE characters SET appearance_tags = ? WHERE card_id = ?",
-            (appearance_tags, card_id),
+            "UPDATE characters SET appearance_tags = ?, appearance_prose = ? WHERE card_id = ?",
+            (appearance_tags, "", card_id),
         )
         await db.commit()
 
@@ -228,7 +233,7 @@ async def preview_card_file(content: bytes, filename: str) -> dict:
 
 async def update_character(card_id: str, fields: dict) -> bool:
     allowed = {"name", "description", "personality", "scenario", "first_mes",
-               "mes_example", "appearance_tags", "lora_name", "lora_weight", "avatar_path",
+               "mes_example", "appearance_tags", "appearance_prose", "lora_name", "lora_weight", "avatar_path",
                "alternate_greetings_json"}
     updates = {k: v for k, v in fields.items() if k in allowed}
     if not updates:
@@ -295,6 +300,7 @@ async def import_card_file(
         "lora_name": lora_name,
         "lora_weight": lora_weight,
         "appearance_tags": saved.get("appearance_tags", ""),
+        "appearance_prose": saved.get("appearance_prose", ""),
         "avatar_path": saved.get("avatar_path", ""),
         "personality": saved.get("personality", ""),
         "scenario": saved.get("scenario", ""),
diff --git a/services/chat_prompt.py b/services/chat_prompt.py
new file mode 100644
index 0000000..f634e85
--- /dev/null
+++ b/services/chat_prompt.py
@@ -0,0 +1,30 @@
+from services.personas import get_persona
+from services.lorebook import get_lorebook_context
+from services.character_card import get_character
+
+DEFAULT_PROMPT = "Ты — полезный AI ассистент. Отвечай чётко и по делу."
+
+from services.rp_sanitize import ROLEPLAY_GUARDRAILS  # noqa: E402 — re-export for imports
+
+
+async def get_system_prompt(persona_id: str, history: list, user_message: str = "") -> str:
+    """Static character prompt only (no RPG runtime blocks)."""
+    persona = await get_persona(persona_id)
+    if not persona:
+        return DEFAULT_PROMPT
+    prompt = persona["prompt"]
+    recent = [m for m in history if m["role"] in ("user", "assistant")][-5:]
+    context = recent + [{"role": "user", "content": user_message}]
+    if persona.get("lorebook_json"):
+        lore = get_lorebook_context(persona.get("lorebook_json", "[]"), context)
+        if lore:
+            prompt += "\n\n" + lore
+    if persona_id.startswith("card_"):
+        card = await get_character(persona_id[5:])
+        if card:
+            lore = get_lorebook_context(card.get("lorebook_json", "[]"), context)
+            if lore:
+                prompt += "\n\n" + lore
+    if persona_id != "default":
+        prompt += "\n\n" + ROLEPLAY_GUARDRAILS
+    return prompt
diff --git a/services/comfy_models.py b/services/comfy_models.py
new file mode 100644
index 0000000..24e5cac
--- /dev/null
+++ b/services/comfy_models.py
@@ -0,0 +1,40 @@
+"""Parse ComfyUI /object_info into usable model lists."""
+
+from __future__ import annotations
+
+# Node types whose combo inputs we expose in the debug UI
+_MODEL_NODES: dict[str, tuple[str, str]] = {
+    "checkpoints": ("CheckpointLoaderSimple", "ckpt_name"),
+    "unets": ("UNETLoader", "unet_name"),
+    "clips": ("CLIPLoader", "clip_name"),
+    "vaes": ("VAELoader", "vae_name"),
+    "loras": ("LoraLoader", "lora_name"),
+}
+
+
+def _combo_options(node_def: dict, input_name: str) -> list[str]:
+    if not isinstance(node_def, dict):
+        return []
+    required = (node_def.get("input") or {}).get("required") or {}
+    optional = (node_def.get("input") or {}).get("optional") or {}
+    spec = required.get(input_name) or optional.get(input_name)
+    if not spec or not isinstance(spec, (list, tuple)):
+        return []
+    first = spec[0]
+    if isinstance(first, list):
+        return [str(x) for x in first]
+    return []
+
+
+def parse_model_lists(object_info: dict) -> dict[str, list[str]]:
+    out: dict[str, list[str]] = {}
+    for key, (node_type, input_name) in _MODEL_NODES.items():
+        node_def = object_info.get(node_type) or {}
+        options = _combo_options(node_def, input_name)
+        if options:
+            out[key] = options
+    return out
+
+
+def list_node_types(object_info: dict) -> list[str]:
+    return sorted(k for k in object_info.keys() if isinstance(object_info.get(k), dict))
diff --git a/services/context_budget.py b/services/context_budget.py
new file mode 100644
index 0000000..9e31524
--- /dev/null
+++ b/services/context_budget.py
@@ -0,0 +1,30 @@
+import os
+
+CHAT_CONTEXT_MAX = int(os.getenv("CHAT_CONTEXT_MAX", "128000"))
+
+
+def estimate_tokens(text: str) -> int:
+    return max(0, len(text or "") // 4)
+
+
+def compute_payload_usage(history: list, llm_system: str) -> dict:
+    """Estimate context fill for the payload messages_for_llm would send."""
+    chars = len(llm_system or "")
+    for m in history:
+        if m.get("role") in ("user", "assistant"):
+            chars += len(m.get("content") or "")
+    tokens_est = chars // 4 if chars else 0
+    max_tokens = CHAT_CONTEXT_MAX
+    percent = round(100.0 * tokens_est / max_tokens, 1) if max_tokens else 0.0
+    return {
+        "chars": chars,
+        "tokens_est": tokens_est,
+        "max_tokens_est": max_tokens,
+        "percent": percent,
+    }
+
+
+def context_warning_line(percent: float) -> str:
+    if percent <= 85:
+        return ""
+    return f"\n[Context: ~{int(percent)}% of budget — keep replies focused]"
diff --git a/services/llm.py b/services/llm.py
index 2b4ae89..4a1a796 100644
--- a/services/llm.py
+++ b/services/llm.py
@@ -13,6 +13,8 @@ OPENROUTER_URL = "https://openrouter.ai/api/v1/chat/completions"
 
 CHAT_MODEL   = os.getenv("CHAT_MODEL",   "mistralai/mistral-nemo")
 SYSTEM_MODEL = os.getenv("SYSTEM_MODEL", "google/gemini-2.5-flash")
+# Softer model when primary returns content_filter / empty / API errors (default: CHAT_MODEL).
+LLM_FALLBACK_MODEL = (os.getenv("LLM_FALLBACK_MODEL") or "").strip() or CHAT_MODEL
 
 HEADERS = {
     "Authorization": f"Bearer {OPENROUTER_KEY}",
@@ -21,26 +23,128 @@ HEADERS = {
 }
 
 
+class LLMError(Exception):
+    """OpenRouter returned an error or an unexpected response shape."""
+
+
+def _parse_completion_body(data: dict) -> str:
+    if not isinstance(data, dict):
+        raise LLMError(f"Invalid API response: expected object, got {type(data).__name__}")
+
+    if data.get("error"):
+        err = data["error"]
+        if isinstance(err, dict):
+            msg = err.get("message") or str(err)
+            code = err.get("code")
+        else:
+            msg = str(err)
+            code = None
+        suffix = f" (code={code})" if code is not None else ""
+        raise LLMError(f"OpenRouter error{suffix}: {msg}")
+
+    choices = data.get("choices")
+    if not choices:
+        preview = str(data)[:400]
+        raise LLMError(f"OpenRouter response has no 'choices'. Body preview: {preview}")
+
+    first = choices[0] if isinstance(choices[0], dict) else {}
+    message = first.get("message") or {}
+    if not isinstance(message, dict):
+        raise LLMError("OpenRouter choice has no message object")
+
+    finish = first.get("finish_reason") or ""
+    native_finish = first.get("native_finish_reason") or ""
+    blocked_reasons = {"content_filter", "safety", "moderation"}
+    if finish in blocked_reasons or str(native_finish).upper() in (
+        "PROHIBITED_CONTENT",
+        "SAFETY",
+        "BLOCKED",
+    ):
+        raise LLMError(
+            f"Content blocked by provider (finish_reason={finish}, native={native_finish})"
+        )
+
+    content = message.get("content")
+    if content is not None and str(content).strip():
+        return str(content)
+
+    refusal = message.get("refusal")
+    if refusal:
+        raise LLMError(f"Model refused the request: {refusal}")
+
+    if finish and finish not in ("stop", "length", "tool_calls", "function_call"):
+        raise LLMError(
+            f"OpenRouter finished without content (finish_reason={finish}, native={native_finish})"
+        )
+
+    raise LLMError("OpenRouter returned empty message content")
+
+
 def _clean(messages: list) -> list:
     """Filter out messages with empty content."""
     return [m for m in messages if (m.get("content") or "").strip()]
 
 
-async def _post(model: str, messages: list, extra: dict | None = None) -> str:
+async def _post_once(model: str, messages: list, extra: dict | None = None) -> str:
+    if not OPENROUTER_KEY:
+        raise LLMError("ROUTER_KEY is not set in environment")
+
     payload = {"model": model, "messages": _clean(messages), **(extra or {})}
     async with httpx.AsyncClient(timeout=90) as client:
         r = await client.post(OPENROUTER_URL, headers=HEADERS, json=payload)
-        r.raise_for_status()
-        return r.json()["choices"][0]["message"]["content"]
+        try:
+            data = r.json()
+        except Exception as e:
+            raise LLMError(f"Non-JSON response (HTTP {r.status_code}): {r.text[:300]}") from e
+
+        if r.status_code >= 400:
+            try:
+                _parse_completion_body(data)
+            except LLMError:
+                raise
+            raise LLMError(f"HTTP {r.status_code}: {data}")
+
+        try:
+            return _parse_completion_body(data)
+        except LLMError:
+            logger.warning(
+                "OpenRouter completion failed model=%s status=%s body=%.500s",
+                model,
+                r.status_code,
+                data,
+            )
+            raise
+
+
+async def _post(model: str, messages: list, extra: dict | None = None) -> str:
+    """POST completion; on failure retries once with LLM_FALLBACK_MODEL (usually CHAT_MODEL)."""
+    try:
+        return await _post_once(model, messages, extra)
+    except LLMError as primary_err:
+        fallback = LLM_FALLBACK_MODEL
+        if not fallback or fallback == model:
+            raise
+        logger.info(
+            "LLM fallback: %s failed (%s) → retrying with %s",
+            model,
+            primary_err,
+            fallback,
+        )
+        try:
+            return await _post_once(fallback, messages, extra)
+        except LLMError as fallback_err:
+            raise LLMError(
+                f"{primary_err} (fallback {fallback} also failed: {fallback_err})"
+            ) from fallback_err
 
 
 async def send_message(messages: list) -> str:
-    """System model — narrator, facts, SD prompt."""
+    """SYSTEM_MODEL with automatic fallback to LLM_FALLBACK_MODEL."""
     return await _post(SYSTEM_MODEL, messages)
 
 
 async def send_message_with_model(messages: list, model: str) -> str:
-    """Explicit model — plot arc, narrator override."""
+    """Named model (RPG_*, SD_*) with automatic fallback to LLM_FALLBACK_MODEL."""
     return await _post(model, messages)
 
 
@@ -73,10 +177,19 @@ async def stream_message(messages: list):
                             return
                         try:
                             chunk = json.loads(data)
-                            content = chunk["choices"][0]["delta"].get("content", "")
+                            if chunk.get("error"):
+                                err = chunk["error"]
+                                msg = err.get("message", err) if isinstance(err, dict) else err
+                                raise LLMError(f"OpenRouter stream error: {msg}")
+                            choices = chunk.get("choices") or []
+                            if not choices:
+                                continue
+                            content = (choices[0].get("delta") or {}).get("content", "")
                             if content:
                                 chunk_count += 1
                                 yield content
+                        except LLMError:
+                            raise
                         except Exception:
                             continue
         except Exception as e:
diff --git a/services/memory.py b/services/memory.py
index 25f5980..b90ebc5 100644
--- a/services/memory.py
+++ b/services/memory.py
@@ -1,8 +1,11 @@
+import json
+
 import aiosqlite
 from database.db import DB_PATH
 
 
-async def get_or_create_session(session_id: str, persona_id: str = "default") -> dict:
+async def get_or_create_session(session_id: str, persona_id: str | None = None) -> dict:
+    """Existing sessions keep their persona_id; persona_id applies only on INSERT."""
     async with aiosqlite.connect(DB_PATH) as db:
         db.row_factory = aiosqlite.Row
         async with db.execute(
@@ -13,9 +16,10 @@ async def get_or_create_session(session_id: str, persona_id: str = "default") ->
         if row:
             return dict(row)
 
+        pid = (persona_id or "default").strip() or "default"
         await db.execute(
             "INSERT INTO sessions (session_id, persona_id) VALUES (?, ?)",
-            (session_id, persona_id),
+            (session_id, pid),
         )
         await db.commit()
 
@@ -71,24 +75,104 @@ async def update_session_persona(session_id: str, persona_id: str):
             (persona_id, session_id),
         )
 
-        # If persona changed, reset RPG state bound to the persona/arc.
         if prev is not None and prev != persona_id:
-            await db.execute(
-                """UPDATE sessions
-                   SET facts_json = '[]',
-                       global_plot = '',
-                       status_quo = '',
-                       plot_arc_json = '{}'
-                   WHERE session_id = ?""",
-                (session_id,),
-            )
-            await db.execute(
-                "DELETE FROM action_resolutions WHERE session_id = ?",
-                (session_id,),
-            )
+            await _reset_persona_bound_state(db, session_id)
         await db.commit()
 
 
+async def _reset_persona_bound_state(db: aiosqlite.Connection, session_id: str) -> None:
+    from services.rpg_state import DEFAULT_NARRATIVE_STATS
+
+    stats_default = json.dumps(DEFAULT_NARRATIVE_STATS, ensure_ascii=False)
+    await db.execute(
+        """UPDATE sessions
+           SET facts_json = '[]',
+               global_plot = '',
+               status_quo = '',
+               plot_arc_json = '{}',
+               outfit_json = '[]',
+               affinity = 0,
+               scene_json = '{}',
+               narrative_stats_json = ?
+           WHERE session_id = ?""",
+        (stats_default, session_id),
+    )
+    await db.execute("DELETE FROM action_resolutions WHERE session_id = ?", (session_id,))
+    await db.execute("DELETE FROM rpg_quests WHERE session_id = ?", (session_id,))
+
+
+async def upsert_static_system_message(
+    session_id: str, static_prompt: str, history: list | None = None
+) -> bool:
+    """Store only static persona prompt in messages. Returns True if written."""
+    hist = history if history is not None else await get_history(session_id)
+    async with aiosqlite.connect(DB_PATH) as db:
+        if not hist:
+            await db.execute(
+                """INSERT INTO messages (session_id, role, content, image_prompt, image_path)
+                   VALUES (?, 'system', ?, NULL, NULL)""",
+                (session_id, static_prompt),
+            )
+            await db.execute(
+                "UPDATE sessions SET updated_at = CURRENT_TIMESTAMP WHERE session_id = ?",
+                (session_id,),
+            )
+            await db.commit()
+            return True
+
+        if hist[0]["role"] == "system":
+            if hist[0]["content"] == static_prompt:
+                return False
+            await db.execute(
+                """UPDATE messages SET content = ?
+                   WHERE session_id = ? AND role = 'system'
+                   AND id = (SELECT MIN(id) FROM messages WHERE session_id = ?)""",
+                (static_prompt, session_id, session_id),
+            )
+            await db.commit()
+            return True
+
+        await db.execute(
+            """INSERT INTO messages (session_id, role, content, image_prompt, image_path)
+               VALUES (?, 'system', ?, NULL, NULL)""",
+            (session_id, static_prompt),
+        )
+        await db.commit()
+        return True
+
+
+async def delete_dialog_messages(session_id: str) -> None:
+    async with aiosqlite.connect(DB_PATH) as db:
+        await db.execute(
+            "DELETE FROM messages WHERE session_id = ? AND role IN ('user', 'assistant')",
+            (session_id,),
+        )
+        await db.commit()
+
+
+async def rebind_session_persona(
+    session_id: str,
+    persona_id: str,
+    *,
+    clear_history: bool = False,
+    static_prompt: str,
+    first_mes: str | None = None,
+) -> None:
+    session = await get_session(session_id)
+    if not session:
+        raise ValueError("Session not found")
+
+    await update_session_persona(session_id, persona_id)
+    if clear_history:
+        await delete_dialog_messages(session_id)
+
+    history = await get_history(session_id)
+    await upsert_static_system_message(session_id, static_prompt, history)
+
+    if clear_history and first_mes and first_mes.strip():
+        await add_message(session_id, "assistant", first_mes.strip())
+
+
 async def update_session_rpg(session_id: str, rpg_enabled: bool):
     async with aiosqlite.connect(DB_PATH) as db:
         await db.execute(
@@ -174,25 +258,116 @@ async def delete_session(session_id: str):
         await db.commit()
 
 
-async def get_history(session_id: str) -> list:
+async def get_action_resolutions_map(session_id: str) -> dict[int, dict]:
     async with aiosqlite.connect(DB_PATH) as db:
         db.row_factory = aiosqlite.Row
         async with db.execute(
-            """SELECT id, role, content, image_prompt, image_path
+            """SELECT message_id, intent_text, roll, outcome, resolution_text
+               FROM action_resolutions
+               WHERE session_id = ? AND message_id IS NOT NULL
+               ORDER BY id""",
+            (session_id,),
+        ) as cur:
+            rows = await cur.fetchall()
+    out: dict[int, dict] = {}
+    for r in rows:
+        mid = r["message_id"]
+        if mid is not None:
+            out[int(mid)] = {
+                "intent_text": r["intent_text"],
+                "roll": r["roll"],
+                "outcome": r["outcome"],
+                "resolution_text": r["resolution_text"],
+            }
+    return out
+
+
+def narrator_message_content(narrator: dict) -> str:
+    return json.dumps(
+        {
+            "roll": narrator.get("roll"),
+            "outcome": narrator.get("outcome"),
+            "text": narrator.get("text", ""),
+            "original_intent": narrator.get("original_intent"),
+        },
+        ensure_ascii=False,
+    )
+
+
+def parse_narrator_message(content: str) -> dict | None:
+    try:
+        data = json.loads(content or "{}")
+    except (json.JSONDecodeError, TypeError):
+        return None
+    if not isinstance(data, dict) or not (data.get("text") or "").strip():
+        return None
+    return data
+
+
+async def seed_quests_from_arc(session_id: str, arc: dict) -> int:
+    """Create active quests for arc beats that are not already in rpg_quests."""
+    if not arc:
+        return 0
+    existing = {q["title"] for q in await get_quests(session_id)}
+    added = 0
+    for beat in arc.get("beats", []):
+        title = (beat.get("title") or beat.get("injection", "")).strip()[:120]
+        if title and title not in existing:
+            await upsert_quest(session_id, title, "active")
+            existing.add(title)
+            added += 1
+    return added
+
+
+async def get_history(session_id: str) -> list:
+    resolutions = await get_action_resolutions_map(session_id)
+    async with aiosqlite.connect(DB_PATH) as db:
+        db.row_factory = aiosqlite.Row
+        async with db.execute(
+            """SELECT id, role, content, image_prompt, image_path,
+                      image_prompt_alt, image_path_alt, choices_json
                FROM messages WHERE session_id = ? ORDER BY id""",
             (session_id,),
         ) as cursor:
             rows = await cursor.fetchall()
-    return [
-        {
+    result = []
+    for idx, r in enumerate(rows):
+        item = {
             "id": r["id"],
             "role": r["role"],
             "content": r["content"],
             "image_prompt": r["image_prompt"],
             "image_path": r["image_path"],
+            "image_prompt_alt": r["image_prompt_alt"],
+            "image_path_alt": r["image_path_alt"],
+            "choices_json": r["choices_json"],
         }
-        for r in rows
-    ]
+        if r["role"] == "user" and r["id"] in resolutions:
+            item["action_resolution"] = resolutions[r["id"]]
+        result.append(item)
+        if r["role"] == "user" and r["id"] in resolutions:
+            nxt = rows[idx + 1] if idx + 1 < len(rows) else None
+            if not nxt or nxt["role"] != "narrator":
+                res = resolutions[r["id"]]
+                result.append(
+                    {
+                        "id": -int(r["id"]),
+                        "role": "narrator",
+                        "content": narrator_message_content(
+                            {
+                                "roll": res.get("roll"),
+                                "outcome": res.get("outcome"),
+                                "text": res.get("resolution_text", ""),
+                            }
+                        ),
+                        "image_prompt": None,
+                        "image_path": None,
+                        "image_prompt_alt": None,
+                        "image_path_alt": None,
+                        "choices_json": None,
+                    }
+                )
+    return result
 
 
 async def get_message(message_id: int) -> dict | None:
@@ -230,6 +405,38 @@ async def delete_message(message_id: int):
         await db.commit()
 
 
+async def delete_message_and_following(session_id: str, message_id: int) -> bool:
+    async with aiosqlite.connect(DB_PATH) as db:
+        await db.execute(
+            "DELETE FROM messages WHERE session_id = ? AND id >= ?",
+            (session_id, message_id),
+        )
+        await db.execute(
+            "UPDATE sessions SET updated_at = CURRENT_TIMESTAMP WHERE session_id = ?",
+            (session_id,),
+        )
+        await db.commit()
+    return True
+
+
+async def update_message_choices(message_id: int, choices_json: str | None):
+    async with aiosqlite.connect(DB_PATH) as db:
+        await db.execute(
+            "UPDATE messages SET choices_json = ? WHERE id = ?",
+            (choices_json, message_id),
+        )
+        await db.commit()
+
+
+async def clear_choices_for_session(session_id: str):
+    async with aiosqlite.connect(DB_PATH) as db:
+        await db.execute(
+            "UPDATE messages SET choices_json = NULL WHERE session_id = ?",
+            (session_id,),
+        )
+        await db.commit()
+
+
 async def get_last_message_preview(session_id: str, max_len: int = 80) -> str:
     async with aiosqlite.connect(DB_PATH) as db:
         db.row_factory = aiosqlite.Row
@@ -261,8 +468,9 @@ async def fork_session(source_session_id: str, until_message_id: int) -> str | N
         await db.execute(
             """INSERT INTO sessions
                (session_id, persona_id, title, rpg_enabled, facts_json, global_plot,
-                status_quo, plot_arc_json, genre, rpg_settings_json, affinity)
-               VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)""",
+                status_quo, plot_arc_json, genre, rpg_settings_json, affinity,
+                outfit_json, scene_json, narrative_stats_json)
+               VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)""",
             (
                 new_id,
                 source["persona_id"],
@@ -275,6 +483,9 @@ async def fork_session(source_session_id: str, until_message_id: int) -> str | N
                 source.get("genre", "adventure"),
                 source.get("rpg_settings_json", "{}"),
                 source.get("affinity", 0),
+                source.get("outfit_json", "[]"),
+                source.get("scene_json", "{}"),
+                source.get("narrative_stats_json", '{"lust":0,"stamina":10,"tension":0}'),
             ),
         )
         async with db.execute(
@@ -309,18 +520,20 @@ async def add_message(
     content: str,
     image_prompt: str | None = None,
     image_path: str | None = None,
-):
+) -> int:
     async with aiosqlite.connect(DB_PATH) as db:
-        await db.execute(
+        cur = await db.execute(
             """INSERT INTO messages (session_id, role, content, image_prompt, image_path)
                VALUES (?, ?, ?, ?, ?)""",
             (session_id, role, content, image_prompt, image_path),
         )
+        msg_id = cur.lastrowid
         await db.execute(
             "UPDATE sessions SET updated_at = CURRENT_TIMESTAMP WHERE session_id = ?",
             (session_id,),
         )
         await db.commit()
+    return msg_id
 
 
 async def update_message_image(message_id: int, image_path: str):
@@ -332,6 +545,33 @@ async def update_message_image(message_id: int, image_path: str):
         await db.commit()
 
 
+async def update_message_prompt(message_id: int, image_prompt: str):
+    async with aiosqlite.connect(DB_PATH) as db:
+        await db.execute(
+            "UPDATE messages SET image_prompt = ? WHERE id = ?",
+            (image_prompt, message_id),
+        )
+        await db.commit()
+
+
+async def update_message_prompt_alt(message_id: int, image_prompt_alt: str):
+    async with aiosqlite.connect(DB_PATH) as db:
+        await db.execute(
+            "UPDATE messages SET image_prompt_alt = ? WHERE id = ?",
+            (image_prompt_alt, message_id),
+        )
+        await db.commit()
+
+
+async def update_message_image_alt(message_id: int, image_path_alt: str):
+    async with aiosqlite.connect(DB_PATH) as db:
+        await db.execute(
+            "UPDATE messages SET image_path_alt = ? WHERE id = ?",
+            (image_path_alt, message_id),
+        )
+        await db.commit()
+
+
 async def get_last_assistant_message_id(session_id: str) -> int | None:
     async with aiosqlite.connect(DB_PATH) as db:
         db.row_factory = aiosqlite.Row
@@ -362,6 +602,18 @@ async def update_session_affinity(session_id: str, delta: int):
         await db.commit()
 
 
+async def set_session_affinity(session_id: str, value: int):
+    """Debug / admin: set absolute affinity (-30..30)."""
+    clamped = max(-30, min(30, int(value)))
+    async with aiosqlite.connect(DB_PATH) as db:
+        await db.execute(
+            "UPDATE sessions SET affinity = ?, updated_at = CURRENT_TIMESTAMP WHERE session_id = ?",
+            (clamped, session_id),
+        )
+        await db.commit()
+    return clamped
+
+
 async def update_session_genre(session_id: str, genre: str):
     async with aiosqlite.connect(DB_PATH) as db:
         await db.execute(
@@ -389,6 +641,24 @@ async def update_session_outfit(session_id: str, outfit_json: str):
         await db.commit()
 
 
+async def update_session_scene(session_id: str, scene_json: str):
+    async with aiosqlite.connect(DB_PATH) as db:
+        await db.execute(
+            "UPDATE sessions SET scene_json = ?, updated_at = CURRENT_TIMESTAMP WHERE session_id = ?",
+            (scene_json, session_id),
+        )
+        await db.commit()
+
+
+async def update_session_narrative_stats(session_id: str, stats_json: str):
+    async with aiosqlite.connect(DB_PATH) as db:
+        await db.execute(
+            "UPDATE sessions SET narrative_stats_json = ?, updated_at = CURRENT_TIMESTAMP WHERE session_id = ?",
+            (stats_json, session_id),
+        )
+        await db.commit()
+
+
 async def upsert_quest(session_id: str, title: str, status: str = "active"):
     async with aiosqlite.connect(DB_PATH) as db:
         async with db.execute(
@@ -429,6 +699,18 @@ async def update_quest_status(session_id: str, title: str, status: str):
         await db.commit()
 
 
+async def update_quest_by_id(quest_id: int, session_id: str, status: str) -> bool:
+    if status not in ("active", "done", "failed"):
+        return False
+    async with aiosqlite.connect(DB_PATH) as db:
+        cur = await db.execute(
+            "UPDATE rpg_quests SET status = ? WHERE id = ? AND session_id = ?",
+            (status, quest_id, session_id),
+        )
+        await db.commit()
+        return cur.rowcount > 0
+
+
 async def get_message_count(session_id: str) -> int:
     async with aiosqlite.connect(DB_PATH) as db:
         db.row_factory = aiosqlite.Row
diff --git a/services/opening.py b/services/opening.py
new file mode 100644
index 0000000..6a175fd
--- /dev/null
+++ b/services/opening.py
@@ -0,0 +1,173 @@
+import json
+import logging
+
+from services.memory import (
+    get_history,
+    get_session,
+    get_last_assistant_message_id,
+    update_session_plot_arc,
+    update_message_choices,
+    seed_quests_from_arc,
+    get_quests,
+)
+from services.rpg_state import apply_narrator_post
+from services.personas import get_persona
+from services.rpg_facts import facts_to_prompt
+from services.rpg_plot import generate_plot_arc, choices_from_narrator
+from services.rpg_context import format_narrator_context
+from services.rpg_narrator import narrator_post
+from services.sd_prompt import generate_sd_prompt
+from services.sd_images import run_sd_for_message
+
+logger = logging.getLogger(__name__)
+
+DEFAULT_RPG_SETTINGS = {
+    "dice": True,
+    "narrator": True,
+    "quests": True,
+    "affinity": True,
+    "choices": True,
+    "stats": False,
+}
+
+
+def get_rpg_settings(session: dict) -> dict:
+    try:
+        return {**DEFAULT_RPG_SETTINGS, **json.loads(session.get("rpg_settings_json") or "{}")}
+    except Exception:
+        return DEFAULT_RPG_SETTINGS
+
+
+async def resolve_greeting(session_id: str, persona: dict) -> str:
+    history = await get_history(session_id)
+    for m in reversed(history):
+        if m.get("role") == "assistant" and (m.get("content") or "").strip():
+            return m["content"].strip()
+    return (persona.get("first_mes") or "").strip()
+
+
+async def ensure_plot_arc_and_quests(
+    session_id: str,
+    persona: dict,
+    greeting: str,
+    genre: str,
+    *,
+    seed_quests: bool = True,
+) -> dict:
+    session = await get_session(session_id) or {}
+    arc_json = session.get("plot_arc_json") or "{}"
+    try:
+        arc = json.loads(arc_json) if isinstance(arc_json, str) else {}
+    except Exception:
+        arc = {}
+
+    if arc:
+        return arc
+
+    facts_block = facts_to_prompt(session.get("facts_json", "[]"))
+    arc = await generate_plot_arc(
+        persona.get("name", "Character"),
+        persona.get("description", ""),
+        persona.get("scenario", ""),
+        greeting,
+        facts_block=facts_block,
+        genre=genre,
+    )
+    if not arc:
+        return {}
+
+    await update_session_plot_arc(session_id, json.dumps(arc, ensure_ascii=False))
+    if seed_quests:
+        await seed_quests_from_arc(session_id, arc)
+    return arc
+
+
+async def process_opening(session_id: str, persona_id: str, *, rpg: bool) -> dict:
+    session = await get_session(session_id)
+    if not session:
+        raise ValueError("Session not found")
+
+    history = await get_history(session_id)
+    assistant_msgs = [m for m in history if m.get("role") == "assistant"]
+    if not assistant_msgs:
+        raise ValueError("No assistant message (first_mes) found")
+
+    first_mes_text = assistant_msgs[-1].get("content", "").strip()
+    if not first_mes_text:
+        raise ValueError("Empty first_mes")
+
+    msg_id = await get_last_assistant_message_id(session_id)
+    persona = await get_persona(persona_id) or {}
+    rpg_settings = get_rpg_settings(session)
+
+    arc: dict = {}
+    choices: list = []
+    status_quo = session.get("status_quo") or ""
+    outfit_json = session.get("outfit_json") or "[]"
+
+    if rpg:
+        genre = session.get("genre") or "adventure"
+        arc = await ensure_plot_arc_and_quests(
+            session_id,
+            persona,
+            first_mes_text,
+            genre,
+            seed_quests=rpg_settings.get("quests", True),
+        )
+
+        session = await get_session(session_id) or session
+        ctx_txt = f"assistant: {first_mes_text}"
+        arc_json = json.dumps(arc, ensure_ascii=False) if arc else ""
+        facts_block = facts_to_prompt(session.get("facts_json", "[]"))
+
+        quests_pre = await get_quests(session_id)
+        narr_ctx = format_narrator_context(arc, quests_pre, session.get("status_quo") or "")
+        post = await narrator_post(
+            persona.get("name", persona_id),
+            ctx_txt,
+            arc_json,
+            facts_block,
+            is_opening=True,
+            extra_context=narr_ctx,
+        )
+
+        if rpg_settings.get("choices", True):
+            choices = choices_from_narrator(post.get("choices") or [])
+
+        await apply_narrator_post(session_id, post, rpg_settings, session)
+        session = await get_session(session_id) or session
+        status_quo = session.get("status_quo") or status_quo
+        outfit_json = session.get("outfit_json") or outfit_json
+
+    quests = await get_quests(session_id)
+    messages = await get_history(session_id)
+    bundle = await generate_sd_prompt(
+        messages,
+        persona_id,
+        outfit_json=outfit_json,
+        scene_json=session.get("scene_json", "{}") if session else "{}",
+    )
+    sd_out = await run_sd_for_message(bundle, msg_id) if bundle else {}
+
+    updated = await get_session(session_id)
+    affinity = updated.get("affinity", 0) if updated else 0
+
+    if msg_id and choices:
+        await update_message_choices(
+            msg_id, json.dumps(choices, ensure_ascii=False)
+        )
+
+    return {
+        "plot_arc": arc,
+        "quests": quests,
+        "outfit_json": outfit_json,
+        "status_quo": status_quo,
+        "choices": choices,
+        "image_prompt": sd_out.get("image_prompt"),
+        "image_prompt_alt": sd_out.get("image_prompt_alt"),
+        "image_path": sd_out.get("image_path"),
+        "image_path_alt": sd_out.get("image_path_alt"),
+        "image_error": sd_out.get("image_error"),
+        "image_error_alt": sd_out.get("image_error_alt"),
+        "affinity": affinity,
+    }
diff --git a/services/outfit_tags.py b/services/outfit_tags.py
new file mode 100644
index 0000000..0e907a6
--- /dev/null
+++ b/services/outfit_tags.py
@@ -0,0 +1,94 @@
+"""Danbooru-style outfit tags with color enrichment for stable SD prompts."""
+
+import json
+import re
+
+COLOR_TOKENS = frozenset({
+    "white", "black", "red", "blue", "green", "yellow", "pink", "purple",
+    "orange", "brown", "gray", "grey", "silver", "gold", "beige", "navy",
+    "blonde", "dark", "light", "cyan", "teal", "maroon", "crimson",
+})
+
+# garment substring -> default color when tag has no color word
+_GARMENT_DEFAULT_COLOR: list[tuple[str, str]] = [
+    ("championship_belt", "gold"),
+    ("belt_collar", "gold"),
+    ("belt", "brown"),
+    ("sports_shorts", "black"),
+    ("shorts", "black"),
+    ("torn_tank_top", "white"),
+    ("tank_top", "white"),
+    ("crop_top", "white"),
+    ("t_shirt", "white"),
+    ("shirt", "white"),
+    ("dress", "black"),
+    ("skirt", "black"),
+    ("jeans", "blue"),
+    ("pants", "black"),
+    ("hoodie", "gray"),
+    ("jacket", "black"),
+    ("ribbon", "red"),
+    ("collar", "black"),
+    ("boots", "black"),
+    ("socks", "white"),
+]
+
+
+def _clean_tag(raw: str) -> str:
+    t = (raw or "").strip().lower()
+    t = re.sub(r"[^\w]+", "_", t)
+    t = re.sub(r"_+", "_", t).strip("_")
+    return t
+
+
+def tag_has_color(tag: str) -> bool:
+    parts = tag.lower().split("_")
+    return any(p in COLOR_TOKENS for p in parts)
+
+
+def enrich_outfit_tag(tag: str) -> str:
+    """Add a color prefix when the tag names clothing but omits color."""
+    t = _clean_tag(tag)
+    if not t or tag_has_color(t):
+        return t
+    for needle, color in _GARMENT_DEFAULT_COLOR:
+        if needle in t:
+            return f"{color}_{t}"
+    return t
+
+
+def normalize_outfit_list(raw: list | None) -> list[str]:
+    out: list[str] = []
+    seen: set[str] = set()
+    if not isinstance(raw, list):
+        return out
+    for item in raw:
+        if isinstance(item, str):
+            t = enrich_outfit_tag(item)
+        elif isinstance(item, dict):
+            label = (item.get("tag") or item.get("label") or "").strip()
+            color = (item.get("color") or "").strip().lower()
+            base = _clean_tag(label)
+            if not base:
+                continue
+            t = f"{color}_{base}" if color and not tag_has_color(base) else enrich_outfit_tag(base)
+        else:
+            continue
+        if t and t not in seen:
+            seen.add(t)
+            out.append(t)
+    return out
+
+
+def outfit_list_to_json(tags: list[str]) -> str:
+    return json.dumps(normalize_outfit_list(tags), ensure_ascii=False)
+
+
+def parse_and_normalize_outfit_json(raw: str | None) -> str:
+    try:
+        data = json.loads(raw or "[]")
+    except (json.JSONDecodeError, TypeError):
+        data = []
+    if isinstance(data, str):
+        data = [x.strip() for x in data.split(",") if x.strip()]
+    return outfit_list_to_json(data if isinstance(data, list) else [])
diff --git a/services/personas.py b/services/personas.py
index e77d3a7..9e54183 100644
--- a/services/personas.py
+++ b/services/personas.py
@@ -63,6 +63,7 @@ def _row_to_persona(row: dict) -> dict:
         "lora_name": row["lora_name"] or "",
         "lora_weight": row["lora_weight"] if row["lora_weight"] is not None else 0.8,
         "appearance_tags": row["appearance_tags"] or "",
+        "appearance_prose": row.get("appearance_prose", "") or "",
         "personality": row.get("personality", "") or "",
         "scenario": row.get("scenario", "") or "",
         "first_mes": row.get("first_mes", "") or "",
@@ -84,6 +85,9 @@ def build_persona_prompt(data: dict) -> str:
     if ex:
         parts.append(f"Example dialogue:\n{ex}")
     parts.append("Stay in character. Reply as the character. Do not add image tags.")
+    from services.chat_prompt import ROLEPLAY_GUARDRAILS
+
+    parts.append(ROLEPLAY_GUARDRAILS)
     return "\n\n".join(p for p in parts if p and p.split(": ", 1)[-1].strip())
 
 
@@ -117,6 +121,7 @@ async def create_persona(
     lora_name: str = "",
     lora_weight: float = 0.8,
     appearance_tags: str = "",
+    appearance_prose: str = "",
     personality: str = "",
     scenario: str = "",
     first_mes: str = "",
@@ -138,19 +143,19 @@ async def create_persona(
         await db.execute(
             """INSERT INTO personas
                (persona_id, name, emoji, description, prompt, custom,
-                sd_enabled, lora_name, lora_weight, appearance_tags,
+                sd_enabled, lora_name, lora_weight, appearance_tags, appearance_prose,
                 personality, scenario, first_mes, mes_example, lorebook_json, avatar_path,
                 alternate_greetings_json)
-               VALUES (?, ?, ?, ?, ?, 1, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)""",
+               VALUES (?, ?, ?, ?, ?, 1, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)""",
             (
                 persona_id, name, emoji, description, final_prompt,
-                1 if sd_enabled else 0, lora_name, lora_weight, appearance_tags,
+                1 if sd_enabled else 0, lora_name, lora_weight, appearance_tags, appearance_prose,
                 personality, scenario, first_mes, mes_example, lorebook_json, avatar_path,
                 alternate_greetings_json,
             ),
         )
         await db.commit()
-    return {
+        return {
         "name": name,
         "emoji": emoji,
         "description": description,
@@ -160,6 +165,7 @@ async def create_persona(
         "lora_name": lora_name,
         "lora_weight": lora_weight,
         "appearance_tags": appearance_tags,
+        "appearance_prose": appearance_prose,
         "personality": personality,
         "scenario": scenario,
         "first_mes": first_mes,
@@ -226,6 +232,7 @@ async def patch_persona(persona_id: str, fields: dict) -> bool:
         "lora_name",
         "lora_weight",
         "appearance_tags",
+        "appearance_prose",
         "personality",
         "scenario",
         "first_mes",
@@ -255,6 +262,19 @@ async def patch_persona(persona_id: str, fields: dict) -> bool:
                 merged = dict(existing)
                 merged.update(updates)
                 updates["prompt"] = build_persona_prompt(merged)
+        
+        if "appearance_tags" in updates and "appearance_prose" not in updates:
+            tags = updates["appearance_tags"].strip()
+            if tags:
+                from services.llm import send_message
+                try:
+                    prose = await send_message([
+                        {"role": "system", "content": "Convert danbooru tags to natural English description. Output only the description, no markdown."},
+                        {"role": "user", "content": f"Tags: {tags}"}
+                    ])
+                    updates["appearance_prose"] = prose.strip()
+                except Exception:
+                    pass
 
         cols = ", ".join(f"{k} = ?" for k in updates)
         cur2 = await db.execute(
diff --git a/services/rp_sanitize.py b/services/rp_sanitize.py
new file mode 100644
index 0000000..9ec739b
--- /dev/null
+++ b/services/rp_sanitize.py
@@ -0,0 +1,77 @@
+"""Roleplay output guardrails and OOC stripping."""
+
+import re
+
+ROLEPLAY_GUARDRAILS = (
+    "[In-character rules — breaking these ruins immersion]\n"
+    "- Reply ONLY as the character in the present moment: spoken lines, visible actions, "
+    "thoughts they would actually show.\n"
+    "- NEVER write: P.S., PS, postscripts, footnotes, section headers, "
+    '"Статус кво", "Status quo", "To be continued", scene summaries, '
+    "editorial closings, or foreshadowing asides "
+    '(e.g. "Когда вы выйдете...", "Ты уже знаешь правду...").\n'
+    "- Do NOT explain subtext to the reader or predict future scenes. No narrator voice.\n"
+    "- End inside the scene; do not wrap up with meta commentary.\n"
+    "- Sections marked MANDATORY (relationship, state) are binding — obey without citing them."
+)
+
+RP_OUTPUT_REMINDER = (
+    "\n\n--- Reply format (MANDATORY) ---\n"
+    "Next message = in-character only. "
+    "Forbidden in output: P.S., Статус кво, Status quo, author notes, summaries, footnotes.\n"
+    "---"
+)
+
+
+def status_quo_prompt_block(status_quo: str) -> str:
+    sq = (status_quo or "").strip()
+    if not sq:
+        return ""
+    return (
+        "\n\n--- Current situation (INTERNAL — player never sees this block) ---\n"
+        + sq
+        + "\nBackground truth for you only. "
+        "Never echo this header, never open/close replies with 'Статус кво' or summaries. "
+        "Show the situation through dialogue and action.\n---"
+    )
+
+
+_OOC_PARA_START = re.compile(
+    r"^(?:"
+    r"Статус\s*кво|Status\s*quo|"
+    r"P\.?\s*S\.?|PS:|"
+    r"Примечание:|Author'?s?\s*note:|"
+    r"OOC:|\\[OOC\\]|"
+    r"To be continued|Продолжение следует"
+    r")\b",
+    re.IGNORECASE,
+)
+
+
+def strip_ooc_from_reply(text: str) -> str:
+    """Remove common OOC tails (P.S., Статус кво paragraphs, etc.)."""
+    if not text or not text.strip():
+        return text or ""
+
+    out = text.rstrip()
+
+    # Drop trailing P.S. block (often last paragraph).
+    out = re.sub(
+        r"(?is)\n\s*P\.?\s*S\.?\s*[:.\-—].*$",
+        "",
+        out,
+    ).rstrip()
+
+    parts = re.split(r"\n\s*\n", out)
+    kept: list[str] = []
+    for part in parts:
+        lines = [ln for ln in part.splitlines() if ln.strip()]
+        if not lines:
+            continue
+        if _OOC_PARA_START.match(lines[0].strip()):
+            continue
+        kept.append(part)
+
+    if not kept:
+        return ""
+    return "\n\n".join(kept).strip()
diff --git a/services/rpg_context.py b/services/rpg_context.py
new file mode 100644
index 0000000..27afcf8
--- /dev/null
+++ b/services/rpg_context.py
@@ -0,0 +1,53 @@
+"""Shared context blocks for RPG narrator / plot LLM calls."""
+
+from services.rpg_plot import count_active_quests
+
+
+def format_narrator_context(
+    arc: dict | None,
+    quests: list | None,
+    status_quo: str = "",
+) -> str:
+    parts: list[str] = []
+    arc = arc or {}
+    beats = arc.get("beats") or []
+    if not isinstance(beats, list):
+        beats = []
+
+    parts.append(f"Plot phase: {arc.get('phase', 'opening')}. Scripted beats left: {len(beats)}.")
+    if not beats:
+        parts.append(
+            "IMPORTANT: Scripted beats are EXHAUSTED (quests may already be done). "
+            "The story must CONTINUE — do not stall. "
+            "Always return 2-4 meaningful choices for the player's next actions. "
+            "You may add quest_updates with status 'active' for NEW optional threads. "
+            "Do NOT re-activate quests the player already completed unless they explicitly revisit that thread."
+        )
+    elif count_active_quests(quests) == 0:
+        pending = [
+            (b.get("title") or b.get("id") or "beat")
+            for b in beats[:3]
+            if isinstance(b, dict)
+        ]
+        parts.append(
+            "IMPORTANT: No active quests but scripted beats remain — arc was likely desynced. "
+            "The engine will inject the next beat; prefer choices that fit pending beats: "
+            + ", ".join(pending)
+            + ". Do NOT treat the arc as finished."
+        )
+    hint = (arc.get("next_beat_hint") or "").strip()
+    if hint:
+        parts.append(f"Arc hint: {hint}")
+
+    if quests:
+        parts.append("Quest log:")
+        for q in quests:
+            parts.append(f"  [{q.get('status', 'active')}] {q.get('title', '')}")
+    else:
+        parts.append("Quest log: (empty)")
+
+    sq = (status_quo or "").strip()
+    if sq:
+        parts.append(f"Status quo: {sq[:400]}")
+
+    return "\n".join(parts)
diff --git a/services/rpg_facts.py b/services/rpg_facts.py
index 253a95c..d535202 100644
--- a/services/rpg_facts.py
+++ b/services/rpg_facts.py
@@ -1,76 +1,370 @@
 import json
+import logging
 import os
+import re
 
-from services.llm import send_message_with_model, send_message
+from services.llm import LLMError, send_message_with_model, send_message
+
+logger = logging.getLogger(__name__)
 
 FACTS_MODEL = os.getenv("RPG_FACTS_MODEL", "").strip() or "deepseek/deepseek-chat-v3"
+FACTS_STORE_LIMIT = int(os.getenv("FACTS_STORE_LIMIT", "100"))
+FACTS_PROMPT_MAX = int(os.getenv("FACTS_PROMPT_MAX", "40"))
+FACTS_DEDUP_THRESHOLD = int(os.getenv("FACTS_DEDUP_THRESHOLD", "30"))
+FACTS_COMPRESS_TARGET = int(os.getenv("FACTS_COMPRESS_TARGET", "22"))
+
+FACTS_SYSTEM = """Extract NEW stable facts from the conversation.
+Return ONLY valid JSON (no markdown), as an array of objects:
+[{"text": "short durable fact", "rp_day": "when this became true in story time"}]
 
-FACTS_SYSTEM = """Extract stable facts from the conversation.
-Return ONLY valid JSON (no markdown), as an array of short strings.
 Rules:
-- Facts must be durable (names, relations, inventory, locations, world rules).
-- Do not include ephemeral actions unless they change state.
-- Avoid duplicates.
-- Keep each fact <= 120 chars.
-Example output:
-["User name is Alex", "We are in a ruined castle", "NPC Mira distrusts the user"]"""
+- Return at most 5 NEW facts per turn. If nothing new, return [].
+- Do NOT repeat or rephrase facts already listed under "Already known".
+- Facts must be durable (names, relations, inventory, locations, lasting world state).
+- Skip momentary emotions unless they permanently change a relationship.
+- text <= 120 chars each.
+- rp_day: in-world time label (день 1, второй день, та же ночь, через год). Use RP time hint when unclear."""
+
+FACTS_COMPRESS_SYSTEM = """You consolidate RPG session memory for a long-running chat.
+Return ONLY valid JSON (no markdown): an array of {"text": "...", "rp_day": "..."}.
+
+Goals:
+- Aggressively MERGE near-duplicates (same topic in RU/EN, Rin/Рин, Grigo/Григорий).
+- Keep ONE best fact per topic; combine rp_day if needed (e.g. "день 1–2").
+- DROP redundant, trivial, or superseded facts.
+- Keep: names, relationships, key locations, lasting magic/rules, inventory, unresolved threads.
+- Target at most {target} facts (fewer is better). Each text <= 120 chars.
+- rp_day = in-world labels only."""
+
+_NAME_ALIASES = (
+    ("grigoriy", "григорий"),
+    ("grigo", "григо"),
+    ("grigory", "григорий"),
+    ("rin", "рин"),
+    ("player", "игрок"),
+    ("user", "игрок"),
+    ("glade", "полян"),
+    ("flowers", "цвет"),
+    ("flower", "цвет"),
+    ("magical", "волшеб"),
+    ("magic", "волшеб"),
+    ("glow", "свет"),
+    ("glowing", "свет"),
+)
 
 
-def merge_facts(existing_json: str, new_facts: list[str], limit: int = 80) -> str:
+def parse_fact_entry(raw) -> dict | None:
+    if isinstance(raw, dict):
+        text = (raw.get("text") or raw.get("fact") or "").strip()
+        rp_day = (raw.get("rp_day") or raw.get("learned") or raw.get("day") or "").strip()
+    elif isinstance(raw, str):
+        text = raw.strip()
+        rp_day = ""
+    else:
+        return None
+    if not text:
+        return None
+    return {"text": text[:120], "rp_day": rp_day[:80]}
+
+
+def parse_facts_list(facts_json: str | None) -> list[dict]:
     try:
-        existing = json.loads(existing_json or "[]")
-        if not isinstance(existing, list):
-            existing = []
+        data = json.loads(facts_json or "[]")
     except json.JSONDecodeError:
-        existing = []
-
-    seen = {str(x).strip() for x in existing if str(x).strip()}
-    merged = [str(x).strip() for x in existing if str(x).strip()]
-    for f in new_facts:
-        s = str(f).strip()
-        if not s or s in seen:
+        return []
+    if not isinstance(data, list):
+        return []
+    out: list[dict] = []
+    seen: set[str] = set()
+    for item in data:
+        entry = parse_fact_entry(item)
+        if not entry:
             continue
-        seen.add(s)
-        merged.append(s)
-
-    if len(merged) > limit:
-        merged = merged[-limit:]
-    return json.dumps(merged, ensure_ascii=False)
+        key = entry["text"].lower()
+        if key in seen:
+            continue
+        seen.add(key)
+        out.append(entry)
+    return out
 
 
-async def extract_facts(context_messages: list[dict]) -> list[str]:
-    # Build a compact transcript
+def facts_list_to_json(facts: list[dict]) -> str:
+    return json.dumps(
+        [{"text": f["text"], "rp_day": f.get("rp_day", "")} for f in facts],
+        ensure_ascii=False,
+    )
+
+
+def rp_day_from_scene(scene_json: str | None) -> str:
+    try:
+        scene = json.loads(scene_json or "{}")
+        if isinstance(scene, dict):
+            day = (scene.get("day") or "").strip()
+            if day:
+                return day[:80]
+    except (json.JSONDecodeError, TypeError):
+        pass
+    return ""
+
+
+def _normalize_fact_text(text: str) -> str:
+    t = (text or "").lower()
+    for a, b in _NAME_ALIASES:
+        t = t.replace(a, b)
+    t = re.sub(r"[^\w\s]", " ", t, flags=re.UNICODE)
+    return re.sub(r"\s+", " ", t).strip()
+
+
+def _fact_tokens(text: str) -> set[str]:
+    words = re.findall(r"[\w]+", _normalize_fact_text(text), flags=re.UNICODE)
+    stop = {"the", "and", "that", "with", "this", "have", "has", "was", "are", "для", "что", "как", "это", "на", "в", "и", "а"}
+    return {w for w in words if len(w) > 2 and w not in stop}
+
+
+def facts_are_similar(a: str, b: str) -> bool:
+    al, bl = a.lower().strip(), b.lower().strip()
+    if al == bl:
+        return True
+    shorter, longer = (al, bl) if len(al) <= len(bl) else (bl, al)
+    if shorter in longer and len(shorter) / max(len(longer), 1) >= 0.35:
+        return True
+    ta, tb = _fact_tokens(a), _fact_tokens(b)
+    if not ta or not tb:
+        return False
+    overlap = len(ta & tb) / len(ta | tb)
+    return overlap >= 0.32
+
+
+def dedupe_facts_fuzzy(facts: list[dict]) -> list[dict]:
+    out: list[dict] = []
+    for f in facts:
+        placed = False
+        for i, existing in enumerate(out):
+            if facts_are_similar(f["text"], existing["text"]):
+                if len(f["text"]) > len(existing["text"]):
+                    out[i]["text"] = f["text"][:120]
+                if f.get("rp_day") and not existing.get("rp_day"):
+                    out[i]["rp_day"] = f["rp_day"]
+                placed = True
+                break
+        if not placed:
+            out.append(dict(f))
+    return out
+
+
+def merge_facts(
+    existing_json: str,
+    new_facts: list,
+    *,
+    rp_day_default: str = "",
+) -> str:
+    merged = parse_facts_list(existing_json)
+    seen = {f["text"].lower() for f in merged}
+    default_day = (rp_day_default or "").strip()[:80]
+
+    for raw in new_facts or []:
+        entry = parse_fact_entry(raw)
+        if not entry:
+            continue
+        if not entry["rp_day"] and default_day:
+            entry["rp_day"] = default_day
+        key = entry["text"].lower()
+        if key in seen:
+            for i, existing in enumerate(merged):
+                if existing["text"].lower() == key:
+                    if entry["rp_day"] and not existing.get("rp_day"):
+                        merged[i]["rp_day"] = entry["rp_day"]
+                    break
+            continue
+        dup = False
+        for i, existing in enumerate(merged):
+            if facts_are_similar(entry["text"], existing["text"]):
+                if len(entry["text"]) > len(existing["text"]):
+                    merged[i]["text"] = entry["text"]
+                if entry["rp_day"] and not existing.get("rp_day"):
+                    merged[i]["rp_day"] = entry["rp_day"]
+                dup = True
+                break
+        if dup:
+            continue
+        seen.add(key)
+        merged.append(entry)
+
+    return facts_list_to_json(merged)
+
+
+async def compress_facts(
+    facts: list[dict],
+    *,
+    scene_context: str = "",
+    status_quo: str = "",
+    target: int = FACTS_COMPRESS_TARGET,
+) -> list[dict]:
+    payload = json.dumps(facts, ensure_ascii=False, indent=2)
+    user = (
+        f"Current fact count: {len(facts)}. Target after merge: <= {target}.\n\n"
+        f"Facts JSON:\n{payload}\n"
+    )
+    if scene_context.strip():
+        user += f"\nCurrent scene:\n{scene_context.strip()[:1500]}\n"
+    if status_quo.strip():
+        user += f"\nStatus quo:\n{status_quo.strip()[:1500]}\n"
+
+    system = FACTS_COMPRESS_SYSTEM.format(target=target)
+    messages = [
+        {"role": "system", "content": system},
+        {"role": "user", "content": user},
+    ]
+    try:
+        raw = await (
+            send_message_with_model(messages, FACTS_MODEL)
+            if FACTS_MODEL
+            else send_message(messages)
+        )
+    except LLMError as e:
+        logger.warning("compress_facts LLM failed: %s", e)
+        return dedupe_facts_fuzzy(facts)[-target:]
+    except Exception as e:
+        logger.warning("compress_facts unexpected: %s", e)
+        return dedupe_facts_fuzzy(facts)[-target:]
+
+    cleaned = raw.strip()
+    if cleaned.startswith("```"):
+        cleaned = cleaned.split("\n", 1)[1] if "\n" in cleaned else cleaned
+        if cleaned.endswith("```"):
+            cleaned = cleaned.rsplit("```", 1)[0]
+        cleaned = cleaned.strip()
+    try:
+        data = json.loads(cleaned)
+        if isinstance(data, list):
+            out = []
+            for item in data:
+                entry = parse_fact_entry(item)
+                if entry:
+                    out.append(entry)
+            if out:
+                logger.info("compress_facts: %d -> %d", len(facts), len(out))
+                return dedupe_facts_fuzzy(out)[:FACTS_STORE_LIMIT]
+    except json.JSONDecodeError:
+        logger.warning("compress_facts JSON parse failed. Raw=%.400s", raw)
+    return dedupe_facts_fuzzy(facts)[-target:]
+
+
+async def merge_facts_persist(
+    existing_json: str,
+    new_facts: list,
+    *,
+    rp_day_default: str = "",
+    scene_context: str = "",
+    status_quo: str = "",
+) -> str:
+    """Merge, fuzzy-dedupe, LLM-compress when list grows too large."""
+    merged_json = merge_facts(
+        existing_json, new_facts, rp_day_default=rp_day_default
+    )
+    facts = dedupe_facts_fuzzy(parse_facts_list(merged_json))
+    if len(facts) > FACTS_DEDUP_THRESHOLD:
+        facts = await compress_facts(
+            facts,
+            scene_context=scene_context,
+            status_quo=status_quo,
+            target=FACTS_COMPRESS_TARGET,
+        )
+        facts = dedupe_facts_fuzzy(facts)
+    if len(facts) > FACTS_STORE_LIMIT:
+        facts = await compress_facts(
+            facts,
+            scene_context=scene_context,
+            status_quo=status_quo,
+            target=FACTS_STORE_LIMIT,
+        )
+    return facts_list_to_json(facts)
+
+
+async def extract_facts(
+    context_messages: list[dict],
+    *,
+    rp_day_hint: str = "",
+    existing_json: str = "",
+) -> list[dict]:
     transcript = "\n".join(
-        f"{m.get('role')}: {m.get('content','')}".strip()
+        f"{m.get('role')}: {m.get('content', '')}".strip()
         for m in context_messages
         if m.get("role") in ("user", "assistant")
     )[-6000:]
 
+    hint = (rp_day_hint or "").strip()
+    known = parse_facts_list(existing_json)
+    user_parts = []
+    if known:
+        known_lines = "\n".join(
+            f"- [{f.get('rp_day') or '?'}] {f['text']}" for f in known[-40:]
+        )
+        user_parts.append(f"Already known facts (do NOT repeat):\n{known_lines}\n")
+    if hint:
+        user_parts.append(f"RP time hint: {hint}\n")
+    user_parts.append(f"New transcript:\n{transcript}")
+    user = "\n".join(user_parts)
+
     messages = [
         {"role": "system", "content": FACTS_SYSTEM},
-        {"role": "user", "content": transcript},
+        {"role": "user", "content": user},
     ]
 
-    raw = await (send_message_with_model(messages, FACTS_MODEL) if FACTS_MODEL else send_message(messages))
     try:
-        data = json.loads(raw.strip())
-        if isinstance(data, list):
-            return [str(x) for x in data][:40]
-    except Exception:
+        raw = await (
+            send_message_with_model(messages, FACTS_MODEL)
+            if FACTS_MODEL
+            else send_message(messages)
+        )
+    except LLMError as e:
+        logger.warning("extract_facts LLM failed (model=%s): %s", FACTS_MODEL or "SYSTEM", e)
         return []
+    except Exception as e:
+        logger.warning("extract_facts unexpected: %s", e)
+        return []
+
+    cleaned = raw.strip()
+    if cleaned.startswith("```"):
+        cleaned = cleaned.split("\n", 1)[1] if "\n" in cleaned else cleaned
+        if cleaned.endswith("```"):
+            cleaned = cleaned.rsplit("```", 1)[0]
+        cleaned = cleaned.strip()
+    try:
+        data = json.loads(cleaned)
+        if isinstance(data, list):
+            out: list[dict] = []
+            for item in data[:8]:
+                entry = parse_fact_entry(item)
+                if not entry:
+                    continue
+                if any(facts_are_similar(entry["text"], k["text"]) for k in known):
+                    continue
+                if not entry["rp_day"] and hint:
+                    entry["rp_day"] = hint[:80]
+                out.append(entry)
+            return out
+    except json.JSONDecodeError:
+        logger.warning("extract_facts JSON parse failed. Raw=%.400s", raw)
     return []
 
 
-def facts_to_prompt(facts_json: str, max_items: int = 20) -> str:
-    try:
-        facts = json.loads(facts_json or "[]")
-        if not isinstance(facts, list):
-            return ""
-    except json.JSONDecodeError:
-        return ""
-    facts = [str(x).strip() for x in facts if str(x).strip()]
+def facts_to_prompt(facts_json: str, max_items: int = FACTS_PROMPT_MAX) -> str:
+    facts = dedupe_facts_fuzzy(parse_facts_list(facts_json))
     if not facts:
         return ""
-    block = "\n".join(f"- {x}" for x in facts[-max_items:])
-    return f"--- Facts (persistent memory) ---\n{block}\n---"
-
+    recent = facts[-max_items:]
+    lines = []
+    for f in recent:
+        day = (f.get("rp_day") or "").strip()
+        if day:
+            lines.append(f"- [{day}] {f['text']}")
+        else:
+            lines.append(f"- {f['text']}")
+    block = "\n".join(lines)
+    total = len(facts)
+    header = "--- Facts (persistent memory"
+    if total > len(recent):
+        header += f", showing {len(recent)} of {total}"
+    header += ") ---"
+    return f"{header}\n{block}\n---"
diff --git a/services/rpg_narrator.py b/services/rpg_narrator.py
index a611baf..12398f5 100644
--- a/services/rpg_narrator.py
+++ b/services/rpg_narrator.py
@@ -2,7 +2,7 @@ import json
 import os
 import random
 
-from services.llm import send_message_with_model
+from services.llm import LLMError, send_message_with_model
 import logging
 
 logger = logging.getLogger(__name__)
@@ -18,8 +18,10 @@ Return ONLY valid JSON (no markdown):
   "check_reason": "brief reason why a check is needed (e.g. 'jumping over a pit')",
   "directives": ["short imperative rules for the next character reply"],
   "resolution_text": "what actually happens as result of the action — written as narrator prose (1-2 sentences). Only if needs_check=true and roll/outcome provided.",
-  "status_quo_update": "optional short update about the world state"
+  "status_quo_update": "optional short update about the world state",
+  "scene_update": {"place": "", "time_of_day": "", "day": "", "weather": "", "exits": [], "layout_note": ""}
 }
+scene_update: only include keys that changed (partial). Omit scene_update if nothing changed.
 If needs_check=false: directives may still guide tone/pacing, resolution_text must be empty string.
 If needs_check=true and roll/outcome are provided: resolution_text MUST reflect the outcome.
 - critical failure (1): embarrassing or painful failure with extra complication
@@ -32,17 +34,25 @@ After the character replied, update persistent state.
 Return ONLY valid JSON (no markdown):
 {
   "status_quo_update": "what changed in the world/state (1-3 sentences)",
-  "facts": ["durable facts only"],
+  "facts": ["durable fact strings OR {\"text\":\"...\",\"rp_day\":\"день 1\"}"],
   "choices": [{"id":"a","label":"..."}, ...],
   "affinity_delta": 0,
+  "stats_delta": {"lust": 0, "stamina": 0, "tension": 0},
+  "scene_update": {"place": "", "place_id": "", "time_of_day": "", "day": "", "weather": "", "exits": [], "layout_note": ""},
   "quest_updates": [{"title": "quest title", "status": "active|done|failed"}],
   "outfit_update": ["danbooru_tag", "danbooru_tag"]
 }
 Rules:
+- status_quo_update: internal DM state only (facts, location, mood). Never address the player, never use headers like "Status quo"/"Статус кво", P.S., or author commentary.
 - affinity_delta: integer -2..+2. Positive if character warmed up to player, negative if pushed away. 0 if neutral.
+- stats_delta: each lust/stamina/tension -2..+2 (0 if unchanged). lust=arousal, stamina=energy, tension=stress.
+- scene_update: partial location/time schema; only keys that changed. Do not duplicate all of status_quo into scene_update.
 - quest_updates: only include if a quest was clearly started, completed, or failed. Empty array otherwise.
-- choices: 0-4 options for what the player can do next.
-- outfit_update: ONLY include if the character's clothing visibly changed (put on, took off, changed outfit). Use exact danbooru-style underscore_tags (e.g. ["white_dress", "red_ribbon", "barefoot"]). Empty array if no change."""
+- choices: 0-4 options for what the player can do next. REQUIRED when scripted beats are exhausted — never return an empty choices array unless the session truly ended.
+- outfit_update: ONLY if clothing visibly changed. Use danbooru underscore_tags WITH COLOR when possible
+  (e.g. white_tank_top, black_sports_shorts, gold_championship_belt, blue_jeans, red_ribbon).
+  Every garment tag should include a color prefix unless the item is inherently colorless (barefoot, nude).
+  Never bare generic tags like sports_shorts or torn_tank_top without a color. Empty array if no change."""
 
 
 async def narrator_pre(
@@ -53,6 +63,7 @@ async def narrator_pre(
     user_message: str,
     roll: int | None = None,
     outcome: str | None = None,
+    extra_context: str = "",
 ) -> dict:
     roll_block = f"Roll d20={roll}\nOutcome={outcome}\n\n" if roll is not None else ""
     user = (
@@ -63,10 +74,20 @@ async def narrator_pre(
         f"Facts:\n{facts_block}\n\n"
         f"Recent context:\n{context}\n"
     )
-    raw = await send_message_with_model(
-        [{"role": "system", "content": NARRATOR_PRE_SYSTEM}, {"role": "user", "content": user}],
-        NARRATOR_MODEL,
-    )
+    if extra_context:
+        user += f"\n--- Session state ---\n{extra_context}\n---\n"
+    try:
+        raw = await send_message_with_model(
+            [{"role": "system", "content": NARRATOR_PRE_SYSTEM}, {"role": "user", "content": user}],
+            NARRATOR_MODEL,
+        )
+    except LLMError as e:
+        logger.warning("Narrator-pre LLM failed (model=%s): %s", NARRATOR_MODEL, e)
+        return {"needs_check": False, "directives": [], "status_quo_update": "", "resolution_text": "", "_ok": False}
+    except Exception as e:
+        logger.warning("Narrator-pre unexpected error: %s", e)
+        return {"needs_check": False, "directives": [], "status_quo_update": "", "resolution_text": "", "_ok": False}
+
     cleaned = raw.strip()
     if cleaned.startswith("```"):
         cleaned = cleaned.split("\n", 1)[1] if "\n" in cleaned else cleaned
@@ -76,10 +97,11 @@ async def narrator_pre(
     try:
         data = json.loads(cleaned)
         if isinstance(data, dict):
+            data["_ok"] = True
             return data
     except Exception:
         logger.warning("Narrator-pre JSON parse failed. Raw=%.500s", raw)
-    return {"needs_check": False, "directives": [], "status_quo_update": "", "resolution_text": ""}
+    return {"needs_check": False, "directives": [], "status_quo_update": "", "resolution_text": "", "_ok": False}
 
 
 async def narrator_post(
@@ -87,17 +109,42 @@ async def narrator_post(
     context: str,
     global_plot: str,
     facts_block: str,
+    is_opening: bool = False,
+    extra_context: str = "",
 ) -> dict:
+    opening_block = ""
+    if is_opening:
+        opening_block = (
+            "\n\nOPENING SCENE: This is the first greeting, not a mid-conversation reply. "
+            "Extract the character's INITIAL visible clothing from the greeting into outfit_update "
+            "(danbooru underscore tags WITH color prefixes: white_shirt, black_shorts, gold_belt), "
+            "even if clothing did not change during the scene. "
+            "Set status_quo to describe the opening situation. "
+            "Fill scene_update from greeting and scenario (place, time_of_day, day, layout_note). "
+            "If the greeting shows clear warmth or hostility toward the player, set affinity_delta "
+            "non-zero (-2..+2); use 0 only if truly neutral.\n"
+        )
     user = (
         f"Persona: {persona_name}\n\n"
         f"Global plot:\n{global_plot}\n\n"
         f"Facts:\n{facts_block}\n\n"
         f"Recent context:\n{context}\n"
+        f"{opening_block}"
     )
-    raw = await send_message_with_model(
-        [{"role": "system", "content": NARRATOR_POST_SYSTEM}, {"role": "user", "content": user}],
-        NARRATOR_MODEL,
-    )
+    if extra_context:
+        user += f"\n--- Session state ---\n{extra_context}\n---\n"
+    try:
+        raw = await send_message_with_model(
+            [{"role": "system", "content": NARRATOR_POST_SYSTEM}, {"role": "user", "content": user}],
+            NARRATOR_MODEL,
+        )
+    except LLMError as e:
+        logger.warning("Narrator-post LLM failed (model=%s): %s", NARRATOR_MODEL, e)
+        return {"status_quo_update": "", "facts": [], "choices": [], "affinity_delta": 0, "quest_updates": [], "_ok": False}
+    except Exception as e:
+        logger.warning("Narrator-post unexpected error: %s", e)
+        return {"status_quo_update": "", "facts": [], "choices": [], "affinity_delta": 0, "quest_updates": [], "_ok": False}
+
     cleaned = raw.strip()
     if cleaned.startswith("```"):
         cleaned = cleaned.split("\n", 1)[1] if "\n" in cleaned else cleaned
@@ -107,7 +154,8 @@ async def narrator_post(
     try:
         data = json.loads(cleaned)
         if isinstance(data, dict):
+            data["_ok"] = True
             return data
     except Exception:
         logger.warning("Narrator-post JSON parse failed. Raw=%.500s", raw)
-    return {"status_quo_update": "", "facts": [], "choices": [], "affinity_delta": 0, "quest_updates": []}
+    return {"status_quo_update": "", "facts": [], "choices": [], "affinity_delta": 0, "quest_updates": [], "_ok": False}
diff --git a/services/rpg_plot.py b/services/rpg_plot.py
index c964813..d8a5d85 100644
--- a/services/rpg_plot.py
+++ b/services/rpg_plot.py
@@ -1,7 +1,7 @@
 import json
 import os
 
-from services.llm import send_message_with_model, send_message
+from services.llm import LLMError, send_message_with_model, send_message
 import logging
 
 logger = logging.getLogger(__name__)
@@ -63,7 +63,19 @@ async def generate_plot_arc(persona_name: str, persona_desc: str, persona_scenar
         {"role": "system", "content": ARC_SYSTEM},
         {"role": "user", "content": user},
     ]
-    raw = await (send_message_with_model(messages, PLOT_MODEL) if PLOT_MODEL else send_message(messages))
+    try:
+        raw = await (
+            send_message_with_model(messages, PLOT_MODEL)
+            if PLOT_MODEL
+            else send_message(messages)
+        )
+    except LLMError as e:
+        logger.warning("generate_plot_arc LLM failed (model=%s): %s", PLOT_MODEL or "SYSTEM", e)
+        return {}
+    except Exception as e:
+        logger.warning("generate_plot_arc unexpected error: %s", e)
+        return {}
+
     cleaned = raw.strip()
     # common OpenRouter formatting: fenced json
     if cleaned.startswith("```"):
@@ -79,17 +91,236 @@ async def generate_plot_arc(persona_name: str, persona_desc: str, persona_scenar
         return {}
 
 
-def should_advance_arc(user_text: str) -> str | None:
+BEAT_MATCH_SYSTEM = """You decide whether the player's latest message should fire ONE scripted plot beat.
+Return ONLY valid JSON (no markdown):
+{"fire_beat_id": "id from list or null", "confidence": "high|low"}
+
+Rules:
+- Fire only if the message clearly matches that beat's narrative intent RIGHT NOW.
+- event_driven:rest — stopping to rest, sleep, camp, sauna break, recuperate (not mere sitting still in scene).
+- event_driven:travel — leaving, driving, journey, going to a new place, hitting the road.
+- event_driven:help_request — explicit plea for help/rescue/assistance.
+- event_driven:after_fail / after_success — follow-up to a recent failure/success beat.
+- Casual talk, flirting, exploring the current place without leaving does NOT fire travel.
+- If nothing fits well, return null.
+- Pick at most ONE beat; prefer high confidence only."""
+
+
+def dice_outcome_to_beat_trigger(outcome: str | None) -> str | None:
+    """Map d20 outcome to event_driven beat trigger (after_fail / after_success)."""
+    o = (outcome or "").strip().lower()
+    if o in ("failure", "critical failure"):
+        return "event_driven:after_fail"
+    if o in ("success", "critical success"):
+        return "event_driven:after_success"
+    return None
+
+
+def should_advance_arc_keywords(user_text: str) -> str | None:
+    """Legacy keyword fallback when LLM match is unavailable."""
     t = (user_text or "").lower()
-    if any(x in t for x in ["отдыха", "ночлег", "спим", "сон", "разбить лагерь", "лагерь", "отдохн"]):
+    if any(x in t for x in ["отдыха", "ночлег", "спим", "сон", "разбить лагерь", "лагерь", "отдохн", "привала", "заправк", "саун"]):
         return "event_driven:rest"
-    if any(x in t for x in ["идем дальше", "пойдем дальше", "в путь", "продолжаем путь", "уходим", "возвращаемся", "переходим"]):
+    if any(
+        x in t
+        for x in [
+            "идем дальше", "пойдем дальше", "пойдём дальше", "едем дальше", "едем",
+            "поехали", "выезжаем", "выезжаю", "в путь", "продолжаем путь",
+            "уходим", "возвращаемся", "переходим", "за рул", "машин", "автомоб",
+            "дорог", "трас", "шосс", "приех", "прибыва", "стади", "на стадион",
+            "отправляемся", "выдвигаемся", "в дорогу",
+        ]
+    ):
         return "event_driven:travel"
     if any(x in t for x in ["помоги", "помочь", "нужна помощь", "спасите", "help"]):
         return "event_driven:help_request"
     return None
 
 
+def _parse_llm_json(raw: str) -> dict | list | None:
+    cleaned = (raw or "").strip()
+    if cleaned.startswith("```"):
+        cleaned = cleaned.split("\n", 1)[1] if "\n" in cleaned else cleaned
+        if cleaned.endswith("```"):
+            cleaned = cleaned.rsplit("```", 1)[0]
+        cleaned = cleaned.strip()
+    try:
+        return json.loads(cleaned)
+    except json.JSONDecodeError:
+        return None
+
+
+async def classify_plot_beat(
+    user_text: str,
+    beats: list[dict],
+    recent_context: str = "",
+    last_dice_outcome: str | None = None,
+) -> str | None:
+    """LLM: return beat id to fire, or None."""
+    pending = [b for b in beats if isinstance(b, dict) and b.get("id")]
+    if not pending or not (user_text or "").strip():
+        return None
+
+    lines = []
+    for b in pending[:8]:
+        lines.append(
+            json.dumps(
+                {
+                    "id": b.get("id"),
+                    "title": b.get("title", ""),
+                    "trigger": b.get("trigger", ""),
+                    "injection": (b.get("injection") or "")[:200],
+                },
+                ensure_ascii=False,
+            )
+        )
+    user = (
+        f"Player message:\n{user_text.strip()}\n\n"
+        f"Pending beats:\n" + "\n".join(lines)
+    )
+    if recent_context.strip():
+        user += f"\n\nRecent chat:\n{recent_context.strip()[-2500:]}\n"
+    if last_dice_outcome:
+        user += f"\nLast dice outcome this turn: {last_dice_outcome}\n"
+
+    messages = [
+        {"role": "system", "content": BEAT_MATCH_SYSTEM},
+        {"role": "user", "content": user},
+    ]
+    try:
+        raw = await (
+            send_message_with_model(messages, PLOT_MODEL)
+            if PLOT_MODEL
+            else send_message(messages)
+        )
+    except LLMError as e:
+        logger.warning("classify_plot_beat LLM failed: %s", e)
+        return None
+    except Exception as e:
+        logger.warning("classify_plot_beat unexpected: %s", e)
+        return None
+
+    data = _parse_llm_json(raw)
+    if not isinstance(data, dict):
+        return None
+    bid = data.get("fire_beat_id")
+    if bid in (None, "", "null", "none"):
+        return None
+    bid = str(bid).strip()
+    if data.get("confidence") == "low":
+        return None
+    valid_ids = {str(b.get("id")) for b in pending}
+    if bid in valid_ids:
+        logger.info("classify_plot_beat: fired %s", bid)
+        return bid
+    return None
+
+
+def pop_beat_by_id(arc: dict, beat_id: str) -> tuple[dict, list[dict]]:
+    beats = arc.get("beats") or []
+    matched, remaining = [], []
+    for b in beats:
+        if isinstance(b, dict) and str(b.get("id")) == str(beat_id) and not matched:
+            matched.append(b)
+        else:
+            remaining.append(b)
+    arc["beats"] = remaining
+    return arc, matched
+
+
+def beat_title(beat: dict) -> str:
+    return ((beat.get("title") or beat.get("injection") or "")[:120]).strip()
+
+
+def count_active_quests(quests: list | None) -> int:
+    return sum(1 for q in (quests or []) if q.get("status") == "active")
+
+
+def prune_beats_for_done_quests(arc: dict, quests: list | None) -> tuple[dict, list[dict]]:
+    """Drop beats whose title already matches a done/failed quest (manual quest close desync)."""
+    done_titles = {
+        (q.get("title") or "").strip().lower()
+        for q in (quests or [])
+        if q.get("status") in ("done", "failed")
+    }
+    if not done_titles:
+        return arc, []
+    removed, kept = [], []
+    for b in arc.get("beats") or []:
+        if isinstance(b, dict) and beat_title(b).lower() in done_titles:
+            removed.append(b)
+        else:
+            kept.append(b)
+    arc["beats"] = kept
+    return arc, removed
+
+
+def pop_next_beats(arc: dict, max_beats: int = 1) -> tuple[dict, list[dict]]:
+    beats = arc.get("beats") or []
+    if not isinstance(beats, list) or not beats:
+        return arc, []
+    n = min(max_beats, len(beats))
+    matched = [b for b in beats[:n] if isinstance(b, dict)]
+    arc["beats"] = beats[n:]
+    return arc, matched
+
+
+async def process_arc_beats(
+    arc: dict,
+    quests: list | None,
+    user_text: str,
+    *,
+    recent_context: str = "",
+    last_dice_outcome: str | None = None,
+    allow_stuck_recovery: bool = True,
+) -> tuple[dict, list[dict], list[dict], str]:
+    """
+    Prune completed beats, then fire by dice outcome, LLM match, keywords, or stuck recovery.
+    Returns (arc, fired_beats, pruned_beats, mode).
+    mode: '' | 'after_dice' | 'llm' | 'trigger' | 'stuck_recovery' | 'pruned'
+    """
+    if not arc:
+        return arc, [], [], ""
+
+    arc, pruned = prune_beats_for_done_quests(arc, quests)
+    beats_pending = arc.get("beats") or []
+
+    dice_trig = dice_outcome_to_beat_trigger(last_dice_outcome)
+    if dice_trig and beats_pending:
+        arc, fired = pop_matching_beats(arc, dice_trig, max_beats=1)
+        if fired:
+            logger.info(
+                "process_arc_beats: after_dice %s -> %s",
+                last_dice_outcome,
+                fired[0].get("id"),
+            )
+            return arc, fired, pruned, "after_dice"
+
+    if beats_pending:
+        beat_id = await classify_plot_beat(
+            user_text, beats_pending, recent_context, last_dice_outcome
+        )
+        if beat_id:
+            arc, fired = pop_beat_by_id(arc, beat_id)
+            if fired:
+                return arc, fired, pruned, "llm"
+
+        trig = should_advance_arc_keywords(user_text)
+        if trig:
+            arc, fired = pop_matching_beats(arc, trig, max_beats=1)
+            if fired:
+                return arc, fired, pruned, "trigger"
+
+    if allow_stuck_recovery and arc.get("beats") and count_active_quests(quests) == 0:
+        arc, fired = pop_next_beats(arc, 1)
+        if fired:
+            return arc, fired, pruned, "stuck_recovery"
+
+    if pruned:
+        return arc, [], pruned, "pruned"
+    return arc, [], [], ""
+
+
 PHASE_ORDER = ["opening", "hook", "complication", "reveal", "climax", "aftermath"]
 
 
@@ -108,6 +339,129 @@ def advance_phase(arc: dict) -> bool:
     return True
 
 
+BEATS_APPEND_SYSTEM = """You are a narrative designer for an RPG chat.
+The plot arc has NO remaining scripted beats. Generate 2-3 NEW beats to continue play.
+Return ONLY valid JSON (no markdown):
+{
+  "beats": [
+    {"id":"b_new_1","title":"short quest title","trigger":"event_driven:rest|event_driven:travel|event_driven:help_request|event_driven:after_fail|event_driven:after_success",
+     "injection":"1-3 sentences in-world",
+     "choices":[{"id":"a","label":"..."},{"id":"b","label":"..."}]}
+  ],
+  "next_beat_hint": "what to push next",
+  "phase": "hook|complication|reveal|climax|aftermath"
+}
+Match the current scene and completed quests. Do not restart finished storylines."""
+
+
+async def replenish_arc_beats(
+    arc: dict,
+    persona_name: str,
+    recent_context: str,
+    quests: list,
+    genre: str = "adventure",
+) -> dict:
+    """Append new beats when arc.beats is empty so plot/quest engine can continue."""
+    if arc.get("beats"):
+        return arc
+
+    quest_lines = "\n".join(
+        f"  [{q.get('status')}] {q.get('title')}" for q in (quests or [])
+    ) or "  (none)"
+    user = (
+        f"Character: {persona_name}\n"
+        f"Genre: {format_genres(genre)}\n"
+        f"Current arc title: {arc.get('title', '')}\n"
+        f"Phase: {arc.get('phase', 'aftermath')}\n"
+        f"Boundaries: {json.dumps(arc.get('boundaries', []), ensure_ascii=False)}\n"
+        f"Quests:\n{quest_lines}\n\n"
+        f"Recent chat:\n{recent_context[-4000:]}\n"
+    )
+    messages = [
+        {"role": "system", "content": BEATS_APPEND_SYSTEM},
+        {"role": "user", "content": user},
+    ]
+    try:
+        raw = await (
+            send_message_with_model(messages, PLOT_MODEL)
+            if PLOT_MODEL
+            else send_message(messages)
+        )
+    except LLMError as e:
+        logger.warning("replenish_arc_beats failed: %s", e)
+        return arc
+    except Exception as e:
+        logger.warning("replenish_arc_beats unexpected: %s", e)
+        return arc
+
+    cleaned = raw.strip()
+    if cleaned.startswith("```"):
+        cleaned = cleaned.split("\n", 1)[1] if "\n" in cleaned else cleaned
+        if cleaned.endswith("```"):
+            cleaned = cleaned.rsplit("```", 1)[0]
+        cleaned = cleaned.strip()
+    try:
+        data = json.loads(cleaned)
+    except Exception:
+        logger.warning("replenish_arc_beats JSON parse failed. Raw=%.400s", raw)
+        return arc
+
+    new_beats = data.get("beats") if isinstance(data, dict) else []
+    if isinstance(new_beats, list) and new_beats:
+        arc["beats"] = new_beats
+        logger.info("replenish_arc_beats: added %d beats", len(new_beats))
+    if isinstance(data, dict) and data.get("next_beat_hint"):
+        arc["next_beat_hint"] = data["next_beat_hint"]
+    if isinstance(data, dict) and data.get("phase"):
+        arc["phase"] = data["phase"]
+    return arc
+
+
+async def reconcile_plot_arc(
+    session_id: str,
+    *,
+    replenish_if_empty: bool = True,
+    recent_context: str = "",
+    persona_name: str = "Character",
+    genre: str = "adventure",
+) -> tuple[dict, bool]:
+    """
+    Prune beats that match done quests; replenish if empty. Persists arc when changed.
+    Returns (arc, changed).
+    """
+    from services.memory import get_session, get_quests, update_session_plot_arc, seed_quests_from_arc
+
+    session = await get_session(session_id)
+    if not session or not session.get("rpg_enabled"):
+        return {}, False
+    try:
+        arc = json.loads(session.get("plot_arc_json") or "{}")
+    except (json.JSONDecodeError, TypeError):
+        arc = {}
+    if not isinstance(arc, dict):
+        arc = {}
+
+    quests = await get_quests(session_id)
+    arc, pruned = prune_beats_for_done_quests(arc, quests)
+    changed = bool(pruned)
+
+    if replenish_if_empty and not arc.get("beats"):
+        arc = await replenish_arc_beats(
+            arc,
+            persona_name,
+            recent_context,
+            quests,
+            genre=session.get("genre") or genre,
+        )
+        if arc.get("beats"):
+            changed = True
+            await seed_quests_from_arc(session_id, arc)
+
+    if changed:
+        await update_session_plot_arc(session_id, json.dumps(arc, ensure_ascii=False))
+    return arc, changed
+
+
 def pop_matching_beats(arc: dict, trigger: str, max_beats: int = 1) -> tuple[dict, list[dict]]:
     beats = arc.get("beats", [])
     if not isinstance(beats, list):
@@ -121,3 +475,49 @@ def pop_matching_beats(arc: dict, trigger: str, max_beats: int = 1) -> tuple[dic
     arc["beats"] = remaining
     return arc, matched
 
+
+def normalize_choice(
+    raw: dict,
+    *,
+    source: str = "narrator",
+    beat: dict | None = None,
+) -> dict | None:
+    """Normalize a choice dict for storage/UI. Adds source and optional beat metadata."""
+    if not isinstance(raw, dict):
+        return None
+    label = (raw.get("label") or "").strip()
+    if not label:
+        return None
+    cid = (raw.get("id") or label[:1].lower() or "a").strip()
+    out = {"id": cid, "label": label, "source": source}
+    if beat and source == "plot_beat":
+        if beat.get("id"):
+            out["beat_id"] = beat["id"]
+        title = (beat.get("title") or "").strip()
+        if title:
+            out["beat_title"] = title
+        injection = (beat.get("injection") or "").strip()
+        if injection:
+            out["beat_injection"] = injection
+    return out
+
+
+def choices_from_beat(beat: dict) -> list[dict]:
+    if not isinstance(beat, dict):
+        return []
+    return [
+        c for c in (
+            normalize_choice(item, source="plot_beat", beat=beat)
+            for item in (beat.get("choices") or [])
+        )
+        if c
+    ]
+
+
+def choices_from_narrator(raw_choices: list) -> list[dict]:
+    if not isinstance(raw_choices, list):
+        return []
+    return [
+        c for c in (normalize_choice(item, source="narrator") for item in raw_choices) if c
+    ]
+
diff --git a/services/rpg_state.py b/services/rpg_state.py
new file mode 100644
index 0000000..ffc42b9
--- /dev/null
+++ b/services/rpg_state.py
@@ -0,0 +1,321 @@
+import json
+
+DEFAULT_NARRATIVE_STATS = {"lust": 0, "stamina": 10, "tension": 0}
+STAT_KEYS = ("lust", "stamina", "tension")
+
+
+def parse_stats_json(raw: str | None) -> dict:
+    try:
+        data = json.loads(raw or "{}") if isinstance(raw, str) else (raw or {})
+    except Exception:
+        data = {}
+    if not isinstance(data, dict):
+        data = {}
+    out = dict(DEFAULT_NARRATIVE_STATS)
+    for k in STAT_KEYS:
+        try:
+            out[k] = max(0, min(10, int(data.get(k, out[k]))))
+        except (TypeError, ValueError):
+            pass
+    return out
+
+
+def parse_scene_json(raw: str | None) -> dict:
+    try:
+        data = json.loads(raw or "{}") if isinstance(raw, str) else (raw or {})
+    except Exception:
+        data = {}
+    return data if isinstance(data, dict) else {}
+
+
+def merge_scene(existing: dict, update: dict | None) -> dict:
+    if not update or not isinstance(update, dict):
+        return dict(existing)
+    merged = dict(existing)
+    for k, v in update.items():
+        if v is None:
+            continue
+        if k == "exits" and isinstance(v, list):
+            merged["exits"] = v
+        elif isinstance(v, str) and v.strip():
+            merged[k] = v.strip()
+        elif v != "":
+            merged[k] = v
+    return merged
+
+
+def apply_stats_delta(stats: dict, delta: dict | None) -> dict:
+    out = parse_stats_json(json.dumps(stats, ensure_ascii=False))
+    if not delta or not isinstance(delta, dict):
+        return out
+    for k in STAT_KEYS:
+        try:
+            d = int(delta.get(k, 0))
+        except (TypeError, ValueError):
+            d = 0
+        d = max(-2, min(2, d))
+        if d:
+            out[k] = max(0, min(10, out[k] + d))
+    return out
+
+
+def scene_prompt_block(scene: dict) -> str:
+    if not scene:
+        return ""
+    lines = []
+    for key, label in (
+        ("place", "Location"),
+        ("place_id", "Location ID"),
+        ("time_of_day", "Time of day"),
+        ("day", "Day"),
+        ("weather", "Weather"),
+        ("layout_note", "Layout"),
+    ):
+        v = (scene.get(key) or "").strip()
+        if v:
+            lines.append(f"{label}: {v}")
+    exits = scene.get("exits")
+    if isinstance(exits, list) and exits:
+        lines.append("Exits: " + ", ".join(str(e) for e in exits))
+    if not lines:
+        return ""
+    return (
+        "\n\n--- Scene ---\n"
+        + "\n".join(lines)
+        + "\nUse this setting consistently. Do not recite scene labels verbatim.\n---"
+    )
+
+
+# Scoped to current player until multi-user / multi-card relationships exist.
+RELATIONSHIP_TARGET_LABEL = "current player (the human you are roleplaying with)"
+
+
+def _mandatory_state_preamble() -> str:
+    return (
+        "MANDATORY ROLEPLAY DIRECTIVE (binding — overrides default cheerfulness or energy):\n"
+        f"- Target: {RELATIONSHIP_TARGET_LABEL} only.\n"
+        "- Your reply MUST visibly obey this state in tone, body language, pacing, and what the character attempts.\n"
+        "- Do NOT contradict a low stamina / high tension / low affinity reading with peppy or intimate behavior unless the text above explicitly allows recovery.\n"
+        "- Never name affinity, lust, stamina, tension, stats, meters, or numeric values in dialogue.\n"
+    )
+
+
+def _band_instruction(value: int, bands: list[tuple[int, str]]) -> str:
+    """bands: list of (min_inclusive, instruction) sorted by min descending."""
+    v = int(value)
+    for min_val, text in bands:
+        if v >= min_val:
+            return text
+    return bands[-1][1] if bands else ""
+
+
+def affinity_prompt_block(affinity: int) -> str:
+    aff = max(-30, min(30, int(affinity)))
+    attitude = _band_instruction(
+        aff,
+        [
+            (10, "Devoted trust: openly affectionate, seeks closeness, defends the player, vulnerable honesty."),
+            (5, "Warm bond: friendly, teasing allowed, volunteers help, remembers small kindnesses."),
+            (1, "Slight fondness: polite-positive, rare soft moments, not yet intimate."),
+            (0, "Neutral professional distance: neither warm nor cold unless scene demands."),
+            (-1, "Cool and guarded: short answers, deflects personal topics, skeptical of motives."),
+            (-5, "Hostile or deeply distrustful: sarcasm, refusal, may threaten to leave or expose the player."),
+            (-30, "Open enmity: antagonistic, undermines, may sabotage or attack socially/physically if fitting genre."),
+        ],
+    )
+    return (
+        "\n\n--- Relationship toward player (MANDATORY) ---\n"
+        + _mandatory_state_preamble()
+        + f"Affinity (internal, not spoken): level {aff}.\n"
+        f"Required attitude toward {RELATIONSHIP_TARGET_LABEL}: {attitude}\n"
+        "---"
+    )
+
+
+def stats_prompt_block(stats: dict) -> str:
+    s = parse_stats_json(json.dumps(stats, ensure_ascii=False))
+    lust = s["lust"]
+    stamina = s["stamina"]
+    tension = s["tension"]
+
+    lust_line = _band_instruction(
+        lust,
+        [
+            (9, "Overwhelming arousal colors every beat: breathy, distracted, struggles to stay on task."),
+            (7, "Strong desire: flushed skin, lingering touch, voice unsteady, thoughts drift to intimacy."),
+            (5, "Clear attraction: meaningful glances, leaning in, playful double meanings if genre fits."),
+            (3, "Mild warmth: subtle flirt only when appropriate; easily redirected."),
+            (1, "Little romantic charge: platonic focus unless the player escalates."),
+            (0, "No romantic or sexual undertone in body language or subtext."),
+        ],
+    )
+    stamina_line = _band_instruction(
+        stamina,
+        [
+            (9, "Peak energy: brisk movement, sharp focus, may offer to take strenuous actions."),
+            (7, "Well-rested: alert, steady pace, normal exertion."),
+            (5, "Average fatigue: fine for talk/light action; heavy labor needs justification."),
+            (4, "Tired: slower reactions, sits when possible, voice softer."),
+            (3, "Heavy fatigue: frequent pauses, avoids running/fighting, may ask to stop."),
+            (2, "Exhausted: barely moves, slumped posture, short sentences, needs support to walk."),
+            (1, "On the verge of collapse: eyelids heavy, may stumble or nearly pass out; minimal action only."),
+            (0, "Cannot sustain activity: collapse/immediate sleep/rest is imminent unless helped."),
+        ],
+    )
+    tension_line = _band_instruction(
+        tension,
+        [
+            (9, "Breaking point: trembling, tears or rage close to surface, irrational snap decisions."),
+            (7, "High stress: clipped speech, hyper-vigilant, jumps at sounds, defensive."),
+            (5, "Uneasy: fidgeting, forced smiles, changes subject from danger."),
+            (3, "Mild edge: occasional sigh, watches exits, relaxes with reassurance."),
+            (1, "Mostly calm: normal breathing, open posture."),
+            (0, "Fully at ease in body and voice."),
+        ],
+    )
+
+    return (
+        "\n\n--- Physical & emotional state (MANDATORY) ---\n"
+        + _mandatory_state_preamble()
+        + f"Internal scales (0–10, never spoken): lust/arousal={lust}, stamina/energy={stamina}, tension/stress={tension}.\n"
+        f"- Lust/arousal: {lust_line}\n"
+        f"- Stamina/energy: {stamina_line}\n"
+        f"- Tension/stress: {tension_line}\n"
+        "If multiple apply, combine them (e.g. low stamina + high tension = shaky exhaustion, not peppy panic).\n"
+        "---"
+    )
+
+
+def format_narrator_outcome_for_llm(data: dict) -> str:
+    """Turn stored narrator JSON into a binding user-turn for the character model."""
+    roll = data.get("roll")
+    outcome = (data.get("outcome") or "").strip().lower()
+    text = (data.get("text") or "").strip()
+    lines = [
+        "--- Narrator ruling (MANDATORY — your next in-character reply MUST follow this) ---",
+        f"Roll d20={roll}. Outcome: {outcome}.",
+        f"What ACTUALLY happened (canonical truth): {text}",
+    ]
+    if outcome in ("failure", "critical failure"):
+        lines.append(
+            "The player's action FAILED as they imagined it. "
+            "Do NOT write a success version: no crowd fleeing, no intimidation working, "
+            "no effortless victory. Show the failure, embarrassment, or partial result above."
+        )
+    elif outcome == "critical success":
+        lines.append(
+            "The attempt succeeded dramatically. You may show amplified success aligned with the outcome above."
+        )
+    else:
+        lines.append(
+            "The attempt succeeded. Your reply must align with the narrator outcome above, not contradict it."
+        )
+    lines.append("Respond as the character to THIS outcome only. Never cite dice, rolls, or stats.")
+    lines.append("---")
+    return "\n".join(lines)
+
+
+def format_user_message_for_llm(content: str, *, has_dice_resolution: bool) -> str:
+    if not has_dice_resolution:
+        return content
+    return (
+        "[Player stated intent — canonical outcome is in the narrator ruling immediately below]\n"
+        + content
+    )
+
+
+def scene_to_sd_hint(scene: dict) -> str:
+    if not scene:
+        return ""
+    parts = []
+    for k in ("place", "time_of_day", "day", "weather", "layout_note"):
+        v = (scene.get(k) or "").strip()
+        if v:
+            parts.append(f"{k}: {v}")
+    return "\n".join(parts)
+
+
+async def apply_narrator_post(session_id: str, post: dict, rpg_settings: dict, session: dict | None = None) -> dict:
+    """Persist narrator_post fields into session. Returns summary of what changed."""
+    from services.memory import (
+        get_session,
+        update_session_status_quo,
+        update_session_affinity,
+        update_session_outfit,
+        update_session_scene,
+        update_session_narrative_stats,
+        update_session_facts,
+        upsert_quest,
+    )
+    from services.rpg_facts import merge_facts_persist, rp_day_from_scene, parse_facts_list
+
+    if session is None:
+        session = await get_session(session_id) or {}
+
+    applied = {
+        "status_quo": False,
+        "facts_added": 0,
+        "affinity_delta": 0,
+        "quests_updated": 0,
+        "scene": False,
+        "outfit": False,
+    }
+
+    sq = (post.get("status_quo_update") or "").strip()
+    if sq:
+        await update_session_status_quo(session_id, sq)
+        applied["status_quo"] = True
+
+    post_facts = post.get("facts") or []
+    if isinstance(post_facts, list) and post_facts:
+        rp_day = rp_day_from_scene(session.get("scene_json"))
+        before = len(parse_facts_list(session.get("facts_json") or "[]"))
+        merged = await merge_facts_persist(
+            session.get("facts_json", "[]"),
+            post_facts,
+            rp_day_default=rp_day,
+            scene_context=json.dumps(
+                parse_scene_json(session.get("scene_json")), ensure_ascii=False
+            ),
+            status_quo=session.get("status_quo") or "",
+        )
+        await update_session_facts(session_id, merged)
+        after = len(parse_facts_list(merged))
+        applied["facts_added"] = max(0, after - before)
+
+    if rpg_settings.get("affinity", True):
+        delta = int(post.get("affinity_delta") or 0)
+        if delta:
+            await update_session_affinity(session_id, delta)
+            applied["affinity_delta"] = delta
+
+    outfit_update = post.get("outfit_update")
+    if isinstance(outfit_update, list) and outfit_update:
+        from services.outfit_tags import outfit_list_to_json
+
+        await update_session_outfit(session_id, outfit_list_to_json(outfit_update))
+        applied["outfit"] = True
+
+    scene_update = post.get("scene_update")
+    if isinstance(scene_update, dict) and scene_update:
+        merged = merge_scene(parse_scene_json(session.get("scene_json")), scene_update)
+        await update_session_scene(session_id, json.dumps(merged, ensure_ascii=False))
+        applied["scene"] = True
+
+    if rpg_settings.get("stats", False):
+        stats_delta = post.get("stats_delta")
+        if isinstance(stats_delta, dict) and stats_delta:
+            current = parse_stats_json(session.get("narrative_stats_json"))
+            updated = apply_stats_delta(current, stats_delta)
+            await update_session_narrative_stats(
+                session_id, json.dumps(updated, ensure_ascii=False)
+            )
+
+    if rpg_settings.get("quests", True):
+        for qu in post.get("quest_updates") or []:
+            title = (qu.get("title") or "").strip()
+            if title:
+                await upsert_quest(session_id, title[:120], qu.get("status", "active"))
+                applied["quests_updated"] += 1
+
+    return applied
diff --git a/services/sd_images.py b/services/sd_images.py
new file mode 100644
index 0000000..3374b13
--- /dev/null
+++ b/services/sd_images.py
@@ -0,0 +1,48 @@
+"""Run ComfyUI generation from SdPromptBundle (single hybrid prompt for Anima)."""
+
+import logging
+import os
+
+from services import sdbackend as sd_service
+from services.memory import update_message_image, update_message_prompt, update_message_prompt_alt
+from services.sd_prompt import SdPromptBundle
+
+logger = logging.getLogger(__name__)
+
+SD_AUTO_GENERATE = os.getenv("SD_AUTO_GENERATE", "false").lower() in ("1", "true", "yes")
+
+
+async def run_sd_for_message(bundle: SdPromptBundle | None, msg_id: int | None) -> dict:
+    """Generate image, persist prompts/paths on message. Returns fields for API/SSE."""
+    out = {
+        "image_prompt": None,
+        "image_prompt_alt": None,
+        "image_path": None,
+        "image_path_alt": None,
+        "image_error": None,
+        "image_error_alt": None,
+    }
+    if not bundle or not bundle.tag_full:
+        return out
+
+    out["image_prompt"] = bundle.tag_full
+    if bundle.desc_full and bundle.desc_full != bundle.tag_full:
+        out["image_prompt_alt"] = bundle.desc_full
+
+    if msg_id:
+        await update_message_prompt(msg_id, bundle.tag_full)
+        if out["image_prompt_alt"]:
+            await update_message_prompt_alt(msg_id, out["image_prompt_alt"])
+
+    if not SD_AUTO_GENERATE:
+        return out
+
+    rel, err = await sd_service.generate_from_full_prompt(bundle.tag_full)
+    if rel:
+        out["image_path"] = f"/static/{rel}"
+        if msg_id:
+            await update_message_image(msg_id, rel)
+    else:
+        out["image_error"] = err
+
+    return out
diff --git a/services/sd_prompt.py b/services/sd_prompt.py
index 9e0b31b..aa5c7b9 100644
--- a/services/sd_prompt.py
+++ b/services/sd_prompt.py
@@ -2,26 +2,115 @@ import json
 import logging
 import os
 import re
+from dataclasses import dataclass
 
 from services.llm import send_message, send_message_with_model
 from services.personas import get_persona
 
 logger = logging.getLogger(__name__)
 
+NEGATIVE_PROMPT_SEPARATOR = "\n\n__NEGATIVE_PROMPT__\n\n"
+
 PROMPT_BUILDER_SYSTEM = """You are a Stable Diffusion prompt engineer for anime illustration models.
 Given a roleplay chat excerpt, output ONLY valid JSON (no markdown):
 {
   "should_generate": true,
   "shot_type": "first_person_pov" | "landscape" | "third_person",
-  "action_tags": "booru-style tags for pose/action/expression, e.g. 'sitting, smiling, holding_cup'",
-  "environment_tags": "booru-style tags for location/lighting/time, e.g. 'indoors, kitchen, sunlight, daytime'"
+  "action_tags": "booru-style tags for pose/action/expression",
+  "environment_tags": "booru-style tags for location/lighting/time"
 }
 Rules:
-- ONLY use real danbooru/e621 tags. Multi-word concepts MUST be underscore_joined: 'fox_ears' not 'fox ears'.
+- ONLY use real danbooru/e621 tags. Multi-word concepts MUST be underscore_joined.
 - Do NOT include appearance/character tags — those are provided separately.
 - Do NOT include quality tags, model names, style words, 'pov', or category/metadata words.
 - Do NOT invent tags. If unsure — omit.
-- Keep each field to 3-6 tags."""
+- Keep action_tags and environment_tags to 3-6 tags each.
+- shot_type: default "first_person_pov" for dialogue/intimacy at arm's length. "third_person" only for wide action (fight, chase). "landscape" only when environment is the focus.
+- should_generate: false for non-visual beats (pure internal monologue, time skips with no new pose, empty lines).
+- NEVER use negative words in tag fields (not, without, naked, nsfw, etc.)."""
+
+ANIMA_BUILDER_EXTRA = """
+Anima hybrid mode — ALSO include:
+  "pov_cue": "face_to_face" | "walking_together" | "doorway_invite" | "reach_to_viewer" | "dialogue_close",
+  "viewer_body_visible": false,
+  "scene_description": "ONE short English sentence (max 40 words). Camera POV: what the viewer sees. Mood/atmosphere only — do NOT repeat tags from action_tags/environment_tags. Do NOT list comma-separated booru tags."
+POV / interaction rules:
+- Default viewer_body_visible: false. The viewer's body, hands, or face must NOT appear in the image — only the character toward the camera.
+- For hugs, embraces: use arms_out, reaching_towards_viewer, inviting_hug — NOT holding_hands, lifting, carrying, nose_rub (these draw a second body in POV).
+- For long messages with time skips ("About an hour later..."), illustrate ONLY the final visible beat (usually the last paragraph).
+- scene_description: describe HER toward the camera only — NEVER "someone", "both", "with you", "hand in hand with", or another person's body.
+- NEVER use tags: looking_at_each_other, couple, 2girls, 2boys, multiple_girls. For POV walking together omit holding_hands (use walking, smiling, reaching_towards_viewer instead).
+- pov_cue: pick the framing that matches the CURRENT beat (walking_together for strolling side by side, doorway_invite for doorway with arms open, reach_to_viewer when she reaches toward camera, face_to_face for close dialogue).
+- Illustrate ONLY the beat under === ILLUSTRATE ===; use === Context === for outfit/location hints only.
+- Do NOT put English sentences in action_tags or environment_tags — tags only."""
+
+POV_CUE_PHRASES: dict[str, str] = {
+    "face_to_face": "POV: close face-to-face, she looks directly at you",
+    "walking_together": "POV: walking beside you, profile and shared path visible",
+    "doorway_invite": "POV: she blocks the doorway, arms open toward you",
+    "reach_to_viewer": "POV: she reaches toward the camera",
+    "dialogue_close": "POV: close conversation, she faces you at arm's length",
+}
+
+POV_CUE_DEFAULT = "POV: she stands before you, facing the camera"
+
+POV_INTERACTION_NEGATIVE = (
+    "duplicate, clone, multiple_girls, 2girls, extra_person, pov hands, "
+    "disembodied hands, extra arms, second person"
+)
+
+_CONTACT_ACTION_KEYWORDS = (
+    "hug", "holding_hands", "hand_holding", "arms_out", "embrace",
+    "reaching", "inviting_hug", "arm_around", "cuddling",
+)
+
+_JUNK_STANDALONE_TAGS = frozenset({
+    "white", "black", "skin", "ear", "ears", "girl", "boy", "fox", "wolf", "cat",
+    "short", "tall", "golden", "silver", "red", "blue", "green", "purple",
+    "pink", "brown", "eye", "eyes", "hair",
+})
+
+_INVALID_TAGS = frozenset({
+    "pumped_up", "pumped", "looking_at_each_other", "couple",
+    "2girls", "2boys", "multiple_girls", "multiple_boys", "duo",
+})
+
+_POV_DROP_ACTION_TAGS = frozenset({
+    "holding_hands", "hand_holding", "looking_at_each_other", "couple",
+    "lifting", "carry", "carrying", "princess_carry", "nose_rub", "nose_boop",
+})
+
+_TIME_SKIP_RE = re.compile(
+    r"(?i)\b(?:about an hour later|hours later|later that (?:day|evening|night)|"
+    r"the next (?:day|morning|evening)|meanwhile|after (?:some )?time)\b[.…\s]*",
+)
+
+_POV_MOOD_FALLBACK: dict[str, str] = {
+    "walking_together": "Easy warmth and quiet laughter in the afternoon light.",
+    "doorway_invite": "Cool air and playful tension as she waits in the doorway.",
+    "reach_to_viewer": "A charged moment as she reaches toward the camera.",
+    "face_to_face": "Her expression softens in close focus toward the camera.",
+    "dialogue_close": "Intimate calm in the space between you.",
+}
+
+_INDOOR_ENV_MARKERS = frozenset({"doorway", "indoors", "indoor", "apartment", "inside", "room"})
+_OUTDOOR_ENV_MARKERS = frozenset({"outdoor", "outdoors", "outside", "street"})
+
+_POV_PROSE_BANNED = re.compile(
+    r"\b(someone|both|together with|hand in hand with|another person|second person|"
+    r"your hands|your fingers|your embrace|your heat|intertwined|with you|"
+    r"demands your|before you)\b",
+    re.IGNORECASE,
+)
+
+SD_ANIMA_DUAL_COMPARE = os.getenv("SD_ANIMA_DUAL_COMPARE", "false").lower() in ("1", "true", "yes")
+
+
+@dataclass
+class SdPromptBundle:
+    tag_full: str
+    negative: str
+    desc_full: str | None = None
 
 
 def extract_image_prompt_tag(text: str) -> str | None:
@@ -44,7 +133,7 @@ SD_UNET = os.getenv("SD_UNET", "")
 SD_PROMPT_MODEL = os.getenv("SD_PROMPT_MODEL", "").strip()
 
 PONY_CHECKPOINTS = {"ponyDiffusionV6XL_v6StartWithThisOne.safetensors"}
-PONY_NEGATIVE  = "score_1, score_2, score_3, score_4, worst quality, low quality, blurry, bad anatomy, watermark, text, censored"
+PONY_NEGATIVE = "score_1, score_2, score_3, score_4, worst quality, low quality, blurry, bad anatomy, watermark, text, censored"
 ANIMA_NEGATIVE = "worst quality, low quality, score_1, score_2, score_3, blurry, jpeg artifacts, sepia"
 
 
@@ -56,37 +145,201 @@ def _is_anima() -> bool:
     return bool(SD_UNET) and not SD_CHECKPOINT
 
 
-def build_positive_prompt(scene: dict, persona: dict | None, outfit_tags: str = "") -> str:
+def anima_dual_enabled() -> bool:
+    return _is_anima() and SD_ANIMA_DUAL_COMPARE
+
+
+def _builder_system() -> str:
+    if _is_anima():
+        return PROMPT_BUILDER_SYSTEM + ANIMA_BUILDER_EXTRA
+    return PROMPT_BUILDER_SYSTEM
+
+
+def _normalize_shot_type(scene: dict) -> dict:
+    st = (scene.get("shot_type") or "").strip().lower()
+    if st == "landscape":
+        scene["shot_type"] = "landscape"
+        return _sanitize_scene_fields(scene)
+    if st == "third_person":
+        action = (scene.get("action_tags") or "").lower()
+        wide = ("battle", "fight", "chase", "running", "crowd", "wide_shot", "group_shot")
+        if any(w in action for w in wide):
+            scene["shot_type"] = "third_person"
+            return _sanitize_scene_fields(scene)
+    scene["shot_type"] = "first_person_pov"
+    if scene.get("viewer_body_visible") is None:
+        scene["viewer_body_visible"] = False
+    return _sanitize_scene_fields(scene)
+
+
+def _split_tag_input(tag_str: str) -> list[str]:
+    return [t.strip() for t in (tag_str or "").split(",") if t.strip()]
+
+
+def _is_sentence_like_tag(tag: str) -> bool:
+    t = tag.strip()
+    if len(t) > 45:
+        return True
+    if re.search(r"[.!?]", t):
+        return True
+    words = t.split()
+    return len(words) >= 5 and "_" not in t
+
+
+def _filter_tag_field(tag_str: str, *, for_pov: bool, field: str) -> str:
+    kept: list[str] = []
+    for raw in _split_tag_input(tag_str):
+        key = raw.lower().replace(" ", "_")
+        if key in _INVALID_TAGS:
+            continue
+        if _is_sentence_like_tag(raw):
+            continue
+        if for_pov and field == "action" and key in _POV_DROP_ACTION_TAGS:
+            continue
+        kept.append(raw if "_" in raw else key)
+    return ", ".join(kept)
+
+
+def _reconcile_environment_tags(env_str: str) -> str:
+    tags = _split_tag_input(env_str)
+    keys = {t.lower().replace(" ", "_") for t in tags}
+    has_indoor = bool(keys & _INDOOR_ENV_MARKERS) or any(
+        any(m in k for m in _INDOOR_ENV_MARKERS) for k in keys
+    )
+    has_outdoor = bool(keys & _OUTDOOR_ENV_MARKERS) or any(
+        any(m in k for m in _OUTDOOR_ENV_MARKERS) for k in keys
+    )
+    if has_indoor and has_outdoor:
+        tags = [t for t in tags if t.lower().replace(" ", "_") not in _OUTDOOR_ENV_MARKERS]
+    return ", ".join(tags)
+
+
+def _sanitize_pov_prose(desc: str, scene: dict) -> str:
+    if not desc or not desc.strip():
+        return ""
+    if scene.get("shot_type") != "first_person_pov":
+        return desc.strip()
+
+    kept: list[str] = []
+    for sentence in re.split(r"(?<=[.!?])\s+", desc.strip()):
+        s = sentence.strip()
+        if not s:
+            continue
+        if _POV_PROSE_BANNED.search(s):
+            continue
+        if re.search(r"\bwolfgirl\b", s, re.I) and re.search(
+            r"\b(walks|walking|stands)\b", s, re.I
+        ):
+            continue
+        kept.append(s)
+    out = " ".join(kept).strip()
+    return re.sub(r"\bat the viewer\b", "at the camera", out, flags=re.IGNORECASE)
+
+
+def _sanitize_scene_fields(scene: dict) -> dict:
+    scene = dict(scene)
+    for_pov = scene.get("shot_type") == "first_person_pov"
+    scene["action_tags"] = _filter_tag_field(
+        scene.get("action_tags") or "", for_pov=for_pov, field="action"
+    )
+    env = _filter_tag_field(scene.get("environment_tags") or "", for_pov=False, field="env")
+    scene["environment_tags"] = _reconcile_environment_tags(env)
+    scene["scene_description"] = _sanitize_pov_prose(
+        (scene.get("scene_description") or "").strip(), scene
+    )
+    return scene
+
+
+def _scene_should_generate(scene: dict) -> bool:
+    if scene.get("should_generate") is False:
+        return False
+    return True
+
+
+def _sanitize_tags_string(tag_str: str) -> str:
+    if not tag_str:
+        return ""
+    out: list[str] = []
+    seen: set[str] = set()
+    for raw in tag_str.split(","):
+        t = raw.strip()
+        if not t:
+            continue
+        key = t.lower().replace(" ", "_")
+        if key in seen:
+            continue
+        if key in _INVALID_TAGS:
+            continue
+        if "_" not in key and key in _JUNK_STANDALONE_TAGS:
+            continue
+        if len(key) <= 2:
+            continue
+        seen.add(key)
+        out.append(t if "_" in t else key)
+    return ", ".join(out)
+
+
+def _quality_prefix() -> str:
     if _is_pony():
-        quality = "score_9, score_8_up, score_7_up, source_anime, highres"
-    elif _is_anima():
-        quality = "masterpiece, best quality, score_7, anime"
-    else:
-        quality = "masterpiece, best quality, highres"
+        return "score_9, score_8_up, score_7_up, source_anime, highres"
+    if _is_anima():
+        return "masterpiece, best quality, score_7, anime"
+    return "masterpiece, best quality, highres"
 
-    parts = [quality]
 
-    appearance = (persona or {}).get("appearance_tags", "")
-    if appearance:
-        parts.append(appearance)
-    if outfit_tags:
-        parts.append(outfit_tags)
+def _appearance_for_persona(persona: dict | None) -> str:
+    """Tag core uses appearance_tags only (prose is for LLM context, not Comfy tag line)."""
+    return _sanitize_tags_string((persona or {}).get("appearance_tags", ""))
 
-    if scene.get("shot_type") == "landscape":
-        parts.append(scene.get("environment_tags", ""))
-    else:
-        if scene.get("shot_type") == "first_person_pov":
-            parts.append("pov, first-person view, looking at viewer")
-        parts.append(scene.get("action_tags", ""))
-        parts.append(scene.get("environment_tags", ""))
 
+def _dedupe_outfit_tags(outfit_tags: str) -> str:
+    tags = _split_tag_input(outfit_tags)
+    keys = {t.lower().replace(" ", "_") for t in tags}
+    if len(keys & {"jeans", "ripped_jeans", "black_jeans"}) > 1 and "jeans" in keys:
+        tags = [t for t in tags if t.lower().replace(" ", "_") != "jeans"]
+    return ", ".join(tags)
+
+
+def _scene_has_physical_contact(scene: dict) -> bool:
+    action = (scene.get("action_tags") or "").lower()
+    return any(k in action for k in _CONTACT_ACTION_KEYWORDS)
+
+
+def _infer_pov_cue_from_action(action_tags: str) -> str:
+    action = (action_tags or "").lower()
+    if any(k in action for k in ("holding_hands", "hand_holding", "walking", "strolling")):
+        return "walking_together"
+    if any(k in action for k in ("doorway", "door", "entry", "threshold")):
+        if any(k in action for k in ("arms_out", "hug", "embrace", "inviting")):
+            return "doorway_invite"
+    if any(k in action for k in ("arms_out", "reaching", "inviting_hug", "hug", "embrace")):
+        return "reach_to_viewer"
+    if any(k in action for k in ("sitting", "lying", "bed")):
+        return "dialogue_close"
+    return "face_to_face"
+
+
+def _build_pov_phrase(scene: dict) -> str:
+    if scene.get("shot_type") != "first_person_pov":
+        return ""
+    cue = (scene.get("pov_cue") or "").strip().lower().replace("-", "_").replace(" ", "_")
+    if cue in POV_CUE_PHRASES:
+        return POV_CUE_PHRASES[cue]
+    inferred = _infer_pov_cue_from_action(scene.get("action_tags", ""))
+    return POV_CUE_PHRASES.get(inferred, POV_CUE_DEFAULT)
+
+
+def _append_lora(parts: list[str], persona: dict | None) -> None:
     lora = (persona or {}).get("lora_name", "")
     weight = (persona or {}).get("lora_weight", 0.8)
     if lora:
         parts.append(f"<lora:{lora}:{weight}>")
 
+
+def _dedupe_comma_join(parts: list[str]) -> str:
     positive = ", ".join(p.strip() for p in parts if p and p.strip())
-    seen, deduped = set(), []
+    seen: set[str] = set()
+    deduped: list[str] = []
     for tag in positive.split(", "):
         t = tag.strip()
         if t and t not in seen:
@@ -95,53 +348,152 @@ def build_positive_prompt(scene: dict, persona: dict | None, outfit_tags: str =
     return ", ".join(deduped)
 
 
-async def generate_sd_prompt(
-    messages: list,
-    persona_id: str,
-    outfit_json: str = "[]",
-) -> tuple[str | None, str | None]:
-    persona = await get_persona(persona_id)
-    # Generate only if persona has appearance tags
-    if not persona or not (persona.get("appearance_tags") or "").strip():
-        logger.debug("sd_prompt skip: persona=%s no appearance_tags", persona_id)
-        return None, None
+def _build_tag_core(scene: dict, persona: dict | None, outfit_tags: str = "") -> str:
+    """Anchor + structure: quality, appearance, outfit, action/env tags, LoRA. No POV prose, no scene_description."""
+    parts = [_quality_prefix()]
+    appearance = _appearance_for_persona(persona)
+    if appearance:
+        parts.append(appearance)
+    if outfit_tags:
+        parts.append(_sanitize_tags_string(_dedupe_outfit_tags(outfit_tags)))
+    if scene.get("shot_type") == "landscape":
+        parts.append(_sanitize_tags_string(scene.get("environment_tags", "")))
+    else:
+        if not _is_anima() and scene.get("shot_type") == "first_person_pov":
+            parts.append("pov, first-person view, looking at viewer")
+        parts.append(_sanitize_tags_string(scene.get("action_tags", "")))
+        parts.append(_sanitize_tags_string(scene.get("environment_tags", "")))
+    _append_lora(parts, persona)
+    return _dedupe_comma_join(parts)
 
-    recent = [m for m in messages if m["role"] in ("user", "assistant")][-6:]
-    if not recent:
-        return None, None
 
-    excerpt = "\n".join(f"{m['role']}: {strip_image_prompt_tag(m['content'])}" for m in recent)
+def build_positive_prompt_tags_only(scene: dict, persona: dict | None, outfit_tags: str = "") -> str:
+    """Tags + contextual POV phrase (Anima) or legacy Pony path."""
+    if not _is_anima():
+        return build_positive_prompt(scene, persona, outfit_tags)
+    core = _build_tag_core(scene, persona, outfit_tags)
+    pov = _build_pov_phrase(scene)
+    if pov:
+        return f"{core}, {pov}" if core else pov
+    return core
 
-    builder_messages = [
-        {"role": "system", "content": PROMPT_BUILDER_SYSTEM},
-        {"role": "user", "content": f"Chat:\n{excerpt}"},
-    ]
 
-    try:
-        if SD_PROMPT_MODEL:
-            raw = await send_message_with_model(builder_messages, SD_PROMPT_MODEL)
-        else:
-            raw = await send_message(builder_messages)
-        raw = raw.strip()
-        if raw.startswith("```"):
-            raw = re.sub(r"^```\w*\n?", "", raw)
-            raw = re.sub(r"\n?```$", "", raw)
-        scene = json.loads(raw)
-        if not isinstance(scene, dict):
-            logger.warning("sd_prompt: LLM returned non-dict: %.100s", raw)
-            return None, None
-    except Exception as e:
-        logger.warning("sd_prompt failed: %s raw=%.200s", e, locals().get("raw", ""))
-        return None, None
+def _tag_tokens_for_dedupe(tag_line: str) -> set[str]:
+    tokens: set[str] = set()
+    for part in tag_line.replace("<lora:", " ").split(","):
+        for word in re.split(r"[\s_./]+", part.lower()):
+            w = word.strip()
+            if len(w) >= 4:
+                tokens.add(w)
+    return tokens
 
-    try:
-        outfit_list = json.loads(outfit_json or "[]")
-        outfit_tags = ", ".join(outfit_list) if isinstance(outfit_list, list) else ""
-    except Exception:
-        outfit_tags = ""
 
-    positive = build_positive_prompt(scene, persona, outfit_tags)
+def _trim_redundant_scene_description(desc: str, tag_line: str) -> str:
+    tag_tokens = _tag_tokens_for_dedupe(tag_line)
+    if not tag_tokens or not desc.strip():
+        return desc.strip()
 
+    kept: list[str] = []
+    for sentence in re.split(r"(?<=[.!?])\s+", desc.strip()):
+        s = sentence.strip()
+        if not s:
+            continue
+        words = [w.lower() for w in re.findall(r"[a-zA-Z]{4,}", s)]
+        if not words:
+            kept.append(s)
+            continue
+        overlap = sum(1 for w in words if w in tag_tokens) / len(words)
+        if overlap < 0.62:
+            kept.append(s)
+
+    return " ".join(kept).strip()
+
+
+def _extract_illustrate_content(content: str, max_chars: int = 1400) -> str:
+    """Long assistant posts (first_mes): use final beat after time-skip, last paragraphs."""
+    text = strip_image_prompt_tag(content).strip()
+    if not text:
+        return ""
+    chunks = _TIME_SKIP_RE.split(text)
+    if len(chunks) > 1:
+        text = chunks[-1].strip()
+    if len(text) <= max_chars:
+        return text
+    paragraphs = [p.strip() for p in re.split(r"\n\s*\n", text) if p.strip()]
+    if paragraphs:
+        for n in (1, 2, 3):
+            tail = "\n\n".join(paragraphs[-n:])
+            if len(tail) <= max_chars:
+                return tail
+        return paragraphs[-1][-max_chars:]
+    return text[-max_chars:]
+
+
+def _fallback_mood_prose(scene: dict) -> str:
+    cue = (scene.get("pov_cue") or "").strip().lower().replace("-", "_").replace(" ", "_")
+    if cue in _POV_MOOD_FALLBACK:
+        return _POV_MOOD_FALLBACK[cue]
+    inferred = _infer_pov_cue_from_action(scene.get("action_tags", ""))
+    return _POV_MOOD_FALLBACK.get(inferred, "Soft atmosphere; her expression toward the camera.")
+
+
+def _cap_scene_description(desc: str, max_words: int = 40, max_chars: int = 220) -> str:
+    words = desc.split()
+    if len(words) > max_words:
+        desc = " ".join(words[:max_words])
+    if len(desc) > max_chars:
+        desc = desc[: max_chars - 3] + "..."
+    return desc
+
+
+def build_positive_prompt_hybrid(scene: dict, persona: dict | None, outfit_tags: str = "") -> str:
+    """Production Anima prompt: tag core + POV cue + short mood prose."""
+    if not _is_anima():
+        return build_positive_prompt(scene, persona, outfit_tags)
+
+    base = build_positive_prompt_tags_only(scene, persona, outfit_tags)
+    desc = _trim_redundant_scene_description(
+        (scene.get("scene_description") or "").strip(),
+        base,
+    )
+    desc = _cap_scene_description(desc)
+    if not desc:
+        desc = _cap_scene_description(_fallback_mood_prose(scene))
+    if not desc:
+        return base
+
+    lora = (persona or {}).get("lora_name", "")
+    weight = (persona or {}).get("lora_weight", 0.8)
+    lora_suffix = f" <lora:{lora}:{weight}>" if lora else ""
+    if lora_suffix and base.endswith(lora_suffix):
+        base = base[: -len(lora_suffix)]
+        return f"{base}. {desc}{lora_suffix}"
+    return f"{base}. {desc}"
+
+
+def build_positive_prompt(scene: dict, persona: dict | None, outfit_tags: str = "") -> str:
+    """Legacy entry: Pony/non-Anima full prompt; Anima delegates to tags-only."""
+    if _is_anima():
+        return build_positive_prompt_tags_only(scene, persona, outfit_tags)
+
+    parts = [_quality_prefix()]
+    appearance = _appearance_for_persona(persona)
+    if appearance:
+        parts.append(appearance)
+    if outfit_tags:
+        parts.append(_sanitize_tags_string(_dedupe_outfit_tags(outfit_tags)))
+    if scene.get("shot_type") == "landscape":
+        parts.append(_sanitize_tags_string(scene.get("environment_tags", "")))
+    else:
+        if scene.get("shot_type") == "first_person_pov":
+            parts.append("pov, first-person view, looking at viewer")
+        parts.append(_sanitize_tags_string(scene.get("action_tags", "")))
+        parts.append(_sanitize_tags_string(scene.get("environment_tags", "")))
+    _append_lora(parts, persona)
+    return _dedupe_comma_join(parts)
+
+
+def _negative_for_scene(scene: dict) -> str:
     if _is_pony():
         negative = PONY_NEGATIVE
     elif _is_anima():
@@ -151,6 +503,237 @@ async def generate_sd_prompt(
 
     if scene.get("shot_type") == "first_person_pov":
         negative += ", third person, over the shoulder"
+        viewer_visible = scene.get("viewer_body_visible") is True
+        if not viewer_visible or _scene_has_physical_contact(scene):
+            negative += ", " + POV_INTERACTION_NEGATIVE
 
-    full = positive + f"\n\nNegative prompt: {negative}"
-    return full, negative
+    return negative
+
+
+def _format_builder_user_block(
+    persona: dict, messages: list[dict], outfit_json: str, scene_json: str = "{}"
+) -> str:
+    lines: list[str] = []
+    tags = (persona.get("appearance_tags") or "").strip()
+    lines.append(f"Character appearance (tags): {tags}")
+    prose = (persona.get("appearance_prose") or "").strip()
+    if _is_anima() and prose and prose != tags:
+        snippet = prose[:300] + ("..." if len(prose) > 300 else "")
+        lines.append(f"Character notes (do not copy into tags or scene_description): {snippet}")
+
+    try:
+        outfit_list = json.loads(outfit_json or "[]")
+        outfit_ref = ", ".join(outfit_list) if isinstance(outfit_list, list) else ""
+    except Exception:
+        outfit_ref = ""
+
+    if outfit_ref:
+        lines.append(f"Current outfit (tags): {outfit_ref}")
+
+    from services.rpg_state import parse_scene_json, scene_to_sd_hint
+
+    scene_hint = scene_to_sd_hint(parse_scene_json(scene_json))
+    if scene_hint:
+        lines.append(f"Scene (location/time):\n{scene_hint}")
+
+    recent = [m for m in messages if m.get("role") in ("user", "assistant")][-6:]
+    if not recent:
+        lines.append("\nChat:\n(no messages — return should_generate=false)")
+        return "\n".join(lines)
+
+    illustrate: list[dict] = []
+    if recent[-1]["role"] == "assistant":
+        illustrate = [recent[-1]]
+        if len(recent) >= 2 and recent[-2]["role"] == "user":
+            illustrate.insert(0, recent[-2])
+    else:
+        illustrate = [recent[-1]]
+        if len(recent) >= 2 and recent[-2]["role"] == "assistant":
+            illustrate.insert(0, recent[-2])
+
+    context = [m for m in recent if m not in illustrate]
+
+    lines.append("\n=== ILLUSTRATE (draw THIS beat only) ===")
+    for m in illustrate:
+        raw = m.get("content", "")
+        content = _extract_illustrate_content(raw) if m.get("role") == "assistant" else strip_image_prompt_tag(raw)
+        lines.append(f"{m['role']}: {content}")
+
+    if context:
+        lines.append("\n=== Context (outfit/location hints only — do not illustrate old beats) ===")
+        for m in context:
+            content = strip_image_prompt_tag(m.get("content", ""))
+            if len(content) > 800:
+                content = content[:797] + "..."
+            lines.append(f"{m['role']}: {content}")
+
+    return "\n".join(lines)
+
+
+def _parse_scene_json(raw: str) -> dict:
+    cleaned = raw.strip()
+    if cleaned.startswith("```"):
+        cleaned = re.sub(r"^```\w*\n?", "", cleaned)
+        cleaned = re.sub(r"\n?```$", "", cleaned)
+    scene = json.loads(cleaned)
+    if not isinstance(scene, dict):
+        raise ValueError("LLM returned non-object JSON")
+    return _normalize_shot_type(scene)
+
+
+def _bundle_from_scene(scene: dict, persona: dict, outfit_tags: str) -> SdPromptBundle:
+    negative = _negative_for_scene(scene)
+    if _is_anima():
+        hybrid = build_positive_prompt_hybrid(scene, persona, outfit_tags)
+        tag_full = hybrid + NEGATIVE_PROMPT_SEPARATOR + negative
+        desc_full = None
+        if anima_dual_enabled():
+            tags_only = build_positive_prompt_tags_only(scene, persona, outfit_tags)
+            desc_full = tags_only + NEGATIVE_PROMPT_SEPARATOR + negative
+        return SdPromptBundle(tag_full=tag_full, negative=negative, desc_full=desc_full)
+
+    positive = build_positive_prompt(scene, persona, outfit_tags)
+    tag_full = positive + NEGATIVE_PROMPT_SEPARATOR + negative
+    return SdPromptBundle(tag_full=tag_full, negative=negative, desc_full=None)
+
+
+def _parse_chat_excerpt(excerpt: str) -> list[dict]:
+    messages: list[dict] = []
+    for line in (excerpt or "").splitlines():
+        line = line.strip()
+        if not line:
+            continue
+        lower = line.lower()
+        if lower.startswith("user:"):
+            messages.append({"role": "user", "content": line[5:].strip()})
+        elif lower.startswith("assistant:"):
+            messages.append({"role": "assistant", "content": line[10:].strip()})
+        elif lower.startswith("system:"):
+            messages.append({"role": "system", "content": line[7:].strip()})
+        else:
+            messages.append({"role": "user", "content": line})
+    return messages
+
+
+async def run_prompt_builder(
+    persona_id: str,
+    *,
+    messages: list[dict] | None = None,
+    chat_excerpt: str = "",
+    outfit_json: str = "[]",
+    appearance_override: str | None = None,
+    use_prose: bool = False,
+) -> dict:
+    """Debug: full SD prompt builder pipeline with LLM raw output."""
+    persona = await get_persona(persona_id) or {}
+    if appearance_override is not None:
+        persona = {**persona, "appearance_tags": appearance_override}
+
+    recent = messages if messages is not None else _parse_chat_excerpt(chat_excerpt)
+    recent = [m for m in recent if m.get("role") in ("user", "assistant")]
+
+    user_block = _format_builder_user_block(persona, recent, outfit_json)
+    builder_messages = [
+        {"role": "system", "content": _builder_system()},
+        {"role": "user", "content": user_block},
+    ]
+    model_used = SD_PROMPT_MODEL or "SYSTEM_MODEL"
+    result: dict = {
+        "persona_id": persona_id,
+        "sd_prompt_model": model_used,
+        "builder_system": _builder_system(),
+        "builder_user": user_block,
+        "anima_dual": anima_dual_enabled(),
+    }
+
+    raw = ""
+    try:
+        if SD_PROMPT_MODEL:
+            raw = await send_message_with_model(builder_messages, SD_PROMPT_MODEL)
+        else:
+            raw = await send_message(builder_messages)
+        result["llm_raw"] = raw
+        scene = _parse_scene_json(raw)
+        result["scene"] = scene
+
+        if not _scene_should_generate(scene):
+            result["skipped"] = True
+            result["error"] = "should_generate=false"
+            return result
+
+        try:
+            outfit_tags = ", ".join(json.loads(outfit_json or "[]"))
+        except Exception:
+            outfit_tags = ""
+
+        negative = _negative_for_scene(scene)
+        if _is_anima():
+            tags_only = build_positive_prompt_tags_only(scene, persona, outfit_tags)
+            hybrid = build_positive_prompt_hybrid(scene, persona, outfit_tags)
+            result["tag_positive"] = tags_only
+            result["hybrid_positive"] = hybrid
+            result["negative"] = negative
+            result["tags_only_full"] = tags_only + NEGATIVE_PROMPT_SEPARATOR + negative
+            result["hybrid_full"] = hybrid + NEGATIVE_PROMPT_SEPARATOR + negative
+            result["tag_full"] = result["hybrid_full"]
+        else:
+            positive = build_positive_prompt(scene, persona, outfit_tags)
+            result["tag_positive"] = positive
+            result["negative"] = negative
+            result["tag_full"] = positive + NEGATIVE_PROMPT_SEPARATOR + negative
+    except Exception as e:
+        result["error"] = str(e)
+        result["llm_raw"] = raw or result.get("llm_raw", "")
+
+    return result
+
+
+async def generate_sd_prompt(
+    messages: list,
+    persona_id: str,
+    outfit_json: str = "[]",
+    scene_json: str = "{}",
+) -> SdPromptBundle | None:
+    persona = await get_persona(persona_id)
+    if not persona:
+        return None
+
+    recent = [m for m in messages if m["role"] in ("user", "assistant")]
+    if not recent:
+        return None
+
+    user_block = _format_builder_user_block(persona, recent, outfit_json, scene_json)
+    builder_messages = [
+        {"role": "system", "content": _builder_system()},
+        {"role": "user", "content": user_block},
+    ]
+
+    raw = ""
+    try:
+        if SD_PROMPT_MODEL:
+            raw = await send_message_with_model(builder_messages, SD_PROMPT_MODEL)
+        else:
+            raw = await send_message(builder_messages)
+        scene = _parse_scene_json(raw)
+    except Exception as e:
+        logger.warning("sd_prompt failed: %s raw=%.200s", e, raw)
+        return None
+
+    if not _scene_should_generate(scene):
+        logger.info("sd_prompt: skipped (should_generate=false)")
+        return None
+
+    try:
+        outfit_list = json.loads(outfit_json or "[]")
+        outfit_tags = ", ".join(outfit_list) if isinstance(outfit_list, list) else ""
+    except Exception:
+        outfit_tags = ""
+
+    bundle = _bundle_from_scene(scene, persona, outfit_tags)
+    if anima_dual_enabled() and bundle.desc_full:
+        logger.info(
+            "Anima prompts: hybrid=%.80s | tags_only=%.80s",
+            bundle.tag_full.split(NEGATIVE_PROMPT_SEPARATOR)[0],
+            bundle.desc_full.split(NEGATIVE_PROMPT_SEPARATOR)[0],
+        )
+    return bundle
diff --git a/services/sdbackend.py b/services/sdbackend.py
index aa3874f..962755e 100644
--- a/services/sdbackend.py
+++ b/services/sdbackend.py
@@ -3,6 +3,7 @@ import logging
 import os
 import uuid
 from pathlib import Path
+from urllib.parse import parse_qs, urlencode, urlparse, urlunparse
 
 import httpx
 from dotenv import load_dotenv
@@ -11,7 +12,178 @@ load_dotenv()
 
 logger = logging.getLogger(__name__)
 
-SD_BASE_URL = os.getenv("SD_BASE_URL", "http://127.0.0.1:8188").rstrip("/")
+
+def _parse_basic_auth() -> httpx.BasicAuth | None:
+    """
+    Vast Caddy on mapped ports often uses Basic realm=restricted.
+    Set SD_COMFY_HTTP_BASIC=user:password or SD_COMFY_USER + SD_COMFY_PASSWORD.
+    """
+    raw = (os.getenv("SD_COMFY_HTTP_BASIC") or "").strip()
+    if raw:
+        if ":" in raw:
+            user, _, password = raw.partition(":")
+        else:
+            user, password = "", raw
+        return httpx.BasicAuth(user, password)
+    user = (os.getenv("SD_COMFY_USER") or "").strip()
+    password = (os.getenv("SD_COMFY_PASSWORD") or "").strip()
+    if user or password:
+        return httpx.BasicAuth(user, password)
+    return None
+
+
+SD_BASIC_AUTH = _parse_basic_auth()
+
+
+def _parse_comfy_config() -> tuple[str, dict[str, str]]:
+    """
+    SD_BASE_URL may be pasted from Vast/Comfy UI with ?token=...
+    API paths must be base + /prompt, not ...?token=xxx/prompt
+    """
+    raw = (os.getenv("SD_BASE_URL") or "http://127.0.0.1:8188").strip()
+    extra_token = (os.getenv("SD_COMFY_TOKEN") or "").strip()
+    parsed = urlparse(raw)
+    base = f"{parsed.scheme}://{parsed.netloc}"
+    path = (parsed.path or "").rstrip("/")
+    if path and path != "/":
+        base = f"{base}{path}"
+    query: dict[str, str] = {}
+    for key, values in parse_qs(parsed.query).items():
+        if values:
+            query[key] = values[-1]
+    if extra_token:
+        query["token"] = extra_token
+    base = base.rstrip("/")
+    # Cloudflare tunnel to localhost:8188 — direct Comfy API, Vast ?token= does not apply
+    if "trycloudflare.com" in base.lower():
+        if query.pop("token", None):
+            logger.info(
+                "SD_BASE_URL is trycloudflare tunnel: Vast token stripped. "
+                "Use tunnel for port 8188 only (see instance Port Mapping)."
+            )
+    return base, query
+
+
+SD_BASE_URL, SD_QUERY_PARAMS = _parse_comfy_config()
+
+
+def _comfy_url(path: str) -> str:
+    if not path.startswith("/"):
+        path = f"/{path}"
+    return f"{SD_BASE_URL}{path}"
+
+
+def _log_comfy_target() -> str:
+    if SD_QUERY_PARAMS.get("token"):
+        return f"{SD_BASE_URL}?token=***"
+    return SD_BASE_URL
+
+
+def _absolute_url(location: str, fallback_path: str = "/") -> str:
+    if not location:
+        return _comfy_url(fallback_path)
+    if location.startswith(("http://", "https://")):
+        return location
+    if location.startswith("/"):
+        return f"{SD_BASE_URL}{location}"
+    return f"{SD_BASE_URL}/{location}"
+
+
+def _url_with_token(url: str) -> str:
+    """Append gateway token to URL (Vast/Cloudflare often strip ?token on redirect)."""
+    if not SD_QUERY_PARAMS.get("token"):
+        return url
+    p = urlparse(url)
+    q: dict[str, str] = {}
+    for key, values in parse_qs(p.query).items():
+        if values:
+            q[key] = values[-1]
+    q.update(SD_QUERY_PARAMS)
+    return urlunparse((p.scheme, p.netloc, p.path, "", urlencode(q), ""))
+
+
+def _merge_params(extra: dict | None) -> dict | None:
+    if not SD_QUERY_PARAMS and not extra:
+        return None
+    merged = dict(SD_QUERY_PARAMS)
+    if extra:
+        merged.update(extra)
+    return merged
+
+
+def _is_vast_gateway() -> bool:
+    return "trycloudflare.com" not in SD_BASE_URL.lower()
+
+
+def _make_comfy_client(*, timeout: float = 300) -> httpx.AsyncClient:
+    return httpx.AsyncClient(
+        timeout=timeout,
+        follow_redirects=False,
+        auth=SD_BASIC_AUTH,
+    )
+
+
+async def _prime_comfy_gateway(client: httpx.AsyncClient) -> None:
+    """
+    Vast Caddy: browser opens /?token=… and gets a session cookie; API then works.
+    Prime with redirects so Set-Cookie is collected, then merge into the API client.
+    """
+    token = SD_QUERY_PARAMS.get("token")
+    if not token or not _is_vast_gateway():
+        return
+    try:
+        async with httpx.AsyncClient(
+            timeout=30,
+            follow_redirects=True,
+            auth=SD_BASIC_AUTH,
+        ) as prime:
+            r = await prime.get(_comfy_url("/"), params={"token": token})
+            client.cookies.update(prime.cookies)
+            logger.info(
+                "Comfy gateway prime GET /?token=*** → %s, cookies=%s",
+                r.status_code,
+                list(prime.cookies.keys()) or "(none)",
+            )
+    except Exception as e:
+        logger.warning("Comfy gateway prime failed: %s", e)
+
+
+async def _comfy_request(
+    client: httpx.AsyncClient,
+    method: str,
+    path: str,
+    *,
+    params: dict | None = None,
+    **kwargs,
+) -> httpx.Response:
+    """
+    Comfy API: trycloudflare tunnel = no token.
+    Vast IP:PORT gateway = ?token= + cookie prime; follow redirects with token re-attached.
+    """
+    url = _comfy_url(path)
+    extra = params or {}
+    token = SD_QUERY_PARAMS.get("token")
+    use_vast_auth = _is_vast_gateway() and (bool(token) or SD_BASIC_AUTH is not None)
+
+    if token and _is_vast_gateway():
+        await _prime_comfy_gateway(client)
+
+    req_params: dict | None = _merge_params(extra) if use_vast_auth else (extra or None)
+    resp: httpx.Response | None = None
+
+    for hop in range(6):
+        resp = await client.request(method, url, params=req_params, **kwargs)
+        if resp.status_code not in (301, 302, 303, 307, 308):
+            return resp
+        loc = _absolute_url(resp.headers.get("location", ""), path)
+        url = _url_with_token(loc) if use_vast_auth else loc
+        req_params = extra or None
+        logger.info("Comfy redirect %s hop %s → %s", resp.status_code, hop + 1, url.split("?")[0])
+
+    assert resp is not None
+    return resp
+
+
 SD_STEPS = int(os.getenv("SD_STEPS", "28"))
 SD_CFG = float(os.getenv("SD_CFG", "7"))
 SD_SAMPLER = os.getenv("SD_SAMPLER", "euler")
@@ -26,6 +198,8 @@ SD_DEFAULT_NEGATIVE = os.getenv(
 SD_UNET = os.getenv("SD_UNET", "anima-preview3-base.safetensors")
 SD_CLIP = os.getenv("SD_CLIP", "qwen_3_06b_base.safetensors")
 SD_VAE = os.getenv("SD_VAE", "qwen_image_vae.safetensors")
+SD_STYLE_LORA = os.getenv("SD_STYLE_LORA", "")
+SD_STYLE_LORA_WEIGHT = float(os.getenv("SD_STYLE_LORA_WEIGHT", "0.7"))
 
 IMAGES_DIR = Path(os.getenv("IMAGES_DIR", "static/images"))
 
@@ -38,20 +212,38 @@ def _use_anima() -> bool:
 
 
 def split_prompt_and_negative(full_prompt: str) -> tuple[str, str]:
+    # Try new separator first
+    sep = "__NEGATIVE_PROMPT__"
+    if f"\n{sep}\n" in full_prompt:
+        pos, _, neg = full_prompt.partition(f"\n{sep}\n")
+        return pos.strip(), neg.strip()
+    # Fallback to old format
     if "\n\nNegative prompt:" in full_prompt:
         pos, _, neg = full_prompt.partition("\n\nNegative prompt:")
         return pos.strip(), neg.strip()
     return full_prompt.strip(), SD_DEFAULT_NEGATIVE
 
 
-def _build_workflow(positive: str, negative: str) -> dict:
+def _workflow_uses_anima(overrides: dict | None) -> bool:
+    if overrides and overrides.get("checkpoint"):
+        return False
+    if overrides and overrides.get("unet"):
+        return True
+    return _use_anima()
+
+
+def _build_workflow(positive: str, negative: str, overrides: dict | None = None) -> dict:
     seed = int(uuid.uuid4().int % 2**32)
-    if _use_anima():
-        return {
-            "44": {"class_type": "UNETLoader",      "inputs": {"unet_name": SD_UNET, "weight_dtype": "default"}},
-            "45": {"class_type": "CLIPLoader",       "inputs": {"clip_name": SD_CLIP, "type": "stable_diffusion", "device": "default"}},
-            "15": {"class_type": "VAELoader",        "inputs": {"vae_name": SD_VAE}},
-            "28": {"class_type": "EmptyLatentImage", "inputs": {"width": 1024, "height": 1024, "batch_size": 1}},
+    o = overrides or {}
+    if _workflow_uses_anima(o):
+        unet = o.get("unet") or SD_UNET
+        clip = o.get("clip") or SD_CLIP
+        vae = o.get("vae") or SD_VAE
+        workflow = {
+            "44": {"class_type": "UNETLoader",      "inputs": {"unet_name": unet, "weight_dtype": "default"}},
+            "45": {"class_type": "CLIPLoader",       "inputs": {"clip_name": clip, "type": "stable_diffusion", "device": "default"}},
+            "15": {"class_type": "VAELoader",        "inputs": {"vae_name": vae}},
+            "28": {"class_type": "EmptyLatentImage", "inputs": {"width": 1024, "height": 720, "batch_size": 1}},
             "11": {"class_type": "CLIPTextEncode",   "inputs": {"text": positive,  "clip": ["45", 0]}},
             "12": {"class_type": "CLIPTextEncode",   "inputs": {"text": negative,  "clip": ["45", 0]}},
             "19": {
@@ -68,9 +260,24 @@ def _build_workflow(positive: str, negative: str) -> dict:
             "8":  {"class_type": "VAEDecode", "inputs": {"samples": ["19", 0], "vae": ["15", 0]}},
             "9":  {"class_type": "SaveImage", "inputs": {"filename_prefix": "chatbot", "images": ["8", 0]}},
         }
-    # Standard checkpoint workflow (Pony / SDXL)
+        if SD_STYLE_LORA:
+            workflow["46"] = {
+                "class_type": "LoraLoader",
+                "inputs": {
+                    "lora_name": SD_STYLE_LORA,
+                    "model": ["44", 0],
+                    "clip": ["45", 0],
+                    "strength_model": SD_STYLE_LORA_WEIGHT,
+                    "strength_clip": SD_STYLE_LORA_WEIGHT,
+                },
+            }
+            workflow["19"]["inputs"]["model"] = ["46", 0]
+            workflow["11"]["inputs"]["clip"] = ["46", 1]
+            workflow["12"]["inputs"]["clip"] = ["46", 1]
+        return workflow
+    ckpt = o.get("checkpoint") or SD_CHECKPOINT
     return {
-        "4":  {"class_type": "CheckpointLoaderSimple", "inputs": {"ckpt_name": SD_CHECKPOINT}},
+        "4":  {"class_type": "CheckpointLoaderSimple", "inputs": {"ckpt_name": ckpt}},
         "5":  {"class_type": "EmptyLatentImage",       "inputs": {"width": 832, "height": 1216, "batch_size": 1}},
         "6":  {"class_type": "CLIPTextEncode",         "inputs": {"text": positive, "clip": ["4", 1]}},
         "7":  {"class_type": "CLIPTextEncode",         "inputs": {"text": negative, "clip": ["4", 1]}},
@@ -89,24 +296,78 @@ def _build_workflow(positive: str, negative: str) -> dict:
     }
 
 
+async def comfy_api_request(
+    method: str,
+    path: str,
+    *,
+    params: dict | None = None,
+    json_body: dict | None = None,
+    timeout: float = 60,
+) -> tuple[int, dict | str, dict]:
+    """
+    Raw Comfy API call for debug. Returns (status_code, parsed_json_or_text, response_headers_subset).
+    """
+    async with _make_comfy_client(timeout=timeout) as client:
+        await _prime_comfy_gateway(client)
+        token = SD_QUERY_PARAMS.get("token")
+        use_vast = _is_vast_gateway() and (bool(token) or SD_BASIC_AUTH is not None)
+        req_params = _merge_params(params) if use_vast else (params or None)
+        req_kwargs: dict = {}
+        if json_body is not None and method.upper() not in ("GET", "HEAD"):
+            req_kwargs["json"] = json_body
+        resp = await _comfy_request(
+            client,
+            method.upper(),
+            path,
+            params=req_params,
+            **req_kwargs,
+        )
+        headers = {
+            k: resp.headers.get(k)
+            for k in ("content-type", "location", "www-authenticate")
+            if resp.headers.get(k)
+        }
+        try:
+            body = resp.json()
+        except Exception:
+            body = resp.text[:8000]
+        return resp.status_code, body, headers
+
+
+async def fetch_object_info() -> dict:
+    status, body, _ = await comfy_api_request("GET", "/object_info", timeout=120)
+    if status != 200 or not isinstance(body, dict):
+        raise RuntimeError(f"object_info failed: HTTP {status} {body!s:.300}")
+    return body
+
+
 async def check_sd() -> bool:
     try:
-        async with httpx.AsyncClient(timeout=5) as client:
-            r = await client.get(f"{SD_BASE_URL}/system_stats")
+        async with _make_comfy_client(timeout=15) as client:
+            await _prime_comfy_gateway(client)
+            r = await _comfy_request(client, "GET", "/system_stats")
             return r.status_code == 200
     except Exception:
         return False
 
 
-async def txt2img(prompt: str, negative_prompt: str | None = None) -> tuple[bytes, str]:
+async def txt2img(
+    prompt: str,
+    negative_prompt: str | None = None,
+    *,
+    overrides: dict | None = None,
+) -> tuple[bytes, str]:
     neg = negative_prompt or SD_DEFAULT_NEGATIVE
-    workflow = _build_workflow(prompt, neg)
+    workflow = _build_workflow(prompt, neg, overrides)
     client_id = uuid.uuid4().hex
 
-    logger.info("ComfyUI request → %s  prompt: %.120s", SD_BASE_URL, prompt)
-    async with httpx.AsyncClient(timeout=300) as client:
-        resp = await client.post(
-            f"{SD_BASE_URL}/prompt",
+    logger.info("ComfyUI request → %s  prompt: %.120s", _log_comfy_target(), prompt)
+    async with _make_comfy_client() as client:
+        await _prime_comfy_gateway(client)
+        resp = await _comfy_request(
+            client,
+            "POST",
+            "/prompt",
             json={"prompt": workflow, "client_id": client_id},
         )
         resp.raise_for_status()
@@ -115,7 +376,7 @@ async def txt2img(prompt: str, negative_prompt: str | None = None) -> tuple[byte
 
         for _ in range(300):
             await asyncio.sleep(1)
-            hist = await client.get(f"{SD_BASE_URL}/history/{prompt_id}")
+            hist = await _comfy_request(client, "GET", f"/history/{prompt_id}")
             data = hist.json()
             if prompt_id in data:
                 entry = data[prompt_id]
@@ -127,9 +388,15 @@ async def txt2img(prompt: str, negative_prompt: str | None = None) -> tuple[byte
                 for node_output in outputs.values():
                     if "images" in node_output:
                         img_info = node_output["images"][0]
-                        img_resp = await client.get(
-                            f"{SD_BASE_URL}/view",
-                            params={"filename": img_info["filename"], "subfolder": img_info.get("subfolder", ""), "type": img_info.get("type", "output")},
+                        img_resp = await _comfy_request(
+                            client,
+                            "GET",
+                            "/view",
+                            params={
+                                "filename": img_info["filename"],
+                                "subfolder": img_info.get("subfolder", ""),
+                                "type": img_info.get("type", "output"),
+                            },
                         )
                         img_resp.raise_for_status()
                         image_bytes = img_resp.content
@@ -145,11 +412,43 @@ async def txt2img(prompt: str, negative_prompt: str | None = None) -> tuple[byte
     raise RuntimeError("ComfyUI generation timed out or produced no output")
 
 
-async def generate_from_full_prompt(full_prompt: str) -> tuple[str | None, str | None]:
+async def generate_from_full_prompt(
+    full_prompt: str,
+    *,
+    overrides: dict | None = None,
+) -> tuple[str | None, str | None]:
     positive, negative = split_prompt_and_negative(full_prompt)
     try:
-        _, rel_path = await txt2img(positive, negative)
+        _, rel_path = await txt2img(positive, negative, overrides=overrides)
         return rel_path, None
+    except httpx.HTTPStatusError as e:
+        code = e.response.status_code
+        if code == 401:
+            logger.error(
+                "ComfyUI 401: Vast Caddy needs SD_COMFY_TOKEN (or ?token= in SD_BASE_URL) "
+                "and/or SD_COMFY_HTTP_BASIC=user:pass from the instance page. "
+                "Test: curl -u user:pass http://IP:PORT/system_stats "
+                "or open /?token=… in browser then curl with cookies. "
+                "Alternative: trycloudflare URL for localhost:8188 in Port Mapping."
+            )
+        elif code in (301, 302, 303, 307, 308):
+            logger.error(
+                "ComfyUI %s: wrong URL — use trycloudflare tunnel for 8188, not web UI link. "
+                "SD_BASE_URL=https://reviewer-relief-edmonton-specializing.trycloudflare.com "
+                "(no ?token=). Location: %s",
+                code,
+                e.response.headers.get("location"),
+            )
+        else:
+            logger.error("ComfyUI HTTP %s: %s", code, e)
+        return None, str(e)
+    except httpx.ConnectError as e:
+        logger.error(
+            "ComfyUI connect failed (%s): IP:8188 is often not exposed on Vast. "
+            "Use trycloudflare URL from Port Mapping for localhost:8188.",
+            e,
+        )
+        return None, str(e)
     except Exception as e:
         logger.error("ComfyUI error: %s", e)
         return None, str(e)
diff --git a/services/session_identity.py b/services/session_identity.py
new file mode 100644
index 0000000..671ae6f
--- /dev/null
+++ b/services/session_identity.py
@@ -0,0 +1,31 @@
+import logging
+
+from services.memory import get_session
+
+logger = logging.getLogger(__name__)
+
+
+async def resolve_session_persona(
+    session_id: str,
+    requested: str | None = None,
+    *,
+    create_persona: str | None = None,
+) -> str:
+    """
+    Session.persona_id is the source of truth.
+    requested is ignored when it disagrees (logged). create_persona used only if session missing.
+    """
+    session = await get_session(session_id)
+    if not session:
+        return (create_persona or requested or "default").strip() or "default"
+
+    bound = (session.get("persona_id") or "default").strip() or "default"
+    req = (requested or "").strip()
+    if req and req != bound:
+        logger.warning(
+            "persona_id mismatch session=%s bound=%s requested=%s (using bound)",
+            session_id,
+            bound,
+            req,
+        )
+    return bound
diff --git a/services/system_message_migration.py b/services/system_message_migration.py
new file mode 100644
index 0000000..4eef6fd
--- /dev/null
+++ b/services/system_message_migration.py
@@ -0,0 +1,21 @@
+import logging
+
+from services.chat_prompt import get_system_prompt
+from services.memory import get_all_sessions, get_history, upsert_static_system_message
+
+logger = logging.getLogger(__name__)
+
+
+async def migrate_static_system_messages() -> int:
+    """Rebuild stored system rows from sessions.persona_id (strip legacy RPG text)."""
+    updated = 0
+    for session in await get_all_sessions():
+        sid = session["session_id"]
+        persona_id = session.get("persona_id") or "default"
+        history = await get_history(sid)
+        static = await get_system_prompt(persona_id, history, "")
+        if await upsert_static_system_message(sid, static, history):
+            updated += 1
+    if updated:
+        logger.info("Migrated %s session system message(s) to static persona prompt", updated)
+    return updated
diff --git a/static/avatars/card_carrie_0578318a_dbee17bc.png b/static/avatars/card_carrie_0578318a_dbee17bc.png
new file mode 100644
index 0000000..8ef6eb0
Binary files /dev/null and b/static/avatars/card_carrie_0578318a_dbee17bc.png differ
diff --git a/static/avatars/card_carrie_926629e6_09f02497.png b/static/avatars/card_carrie_926629e6_09f02497.png
new file mode 100644
index 0000000..8ef6eb0
Binary files /dev/null and b/static/avatars/card_carrie_926629e6_09f02497.png differ
diff --git a/static/avatars/card_carrie_926629e6_8b9b8891.png b/static/avatars/card_carrie_926629e6_8b9b8891.png
new file mode 100644
index 0000000..8ef6eb0
Binary files /dev/null and b/static/avatars/card_carrie_926629e6_8b9b8891.png differ
diff --git a/static/avatars/card_carrie_926629e6_e01546a5.png b/static/avatars/card_carrie_926629e6_e01546a5.png
new file mode 100644
index 0000000..8ef6eb0
Binary files /dev/null and b/static/avatars/card_carrie_926629e6_e01546a5.png differ
diff --git a/static/css/app.css b/static/css/app.css
index 6930b96..03af736 100644
--- a/static/css/app.css
+++ b/static/css/app.css
@@ -156,6 +156,88 @@ header h1 { font-size: 1.1rem; color: #e94560; }
 #systemBlobRefresh.spinning { animation: spin 0.6s linear infinite; }
 @keyframes spin { to { transform: rotate(360deg); } }
 .blob-changed { background: rgba(255, 200, 50, 0.15); border-radius: 3px; transition: background 2s ease; }
+.blob-context-warn { color: #e9a045; }
+
+.rp-dialogue { color: #e94560; }
+.rp-action { font-style: italic; color: #a8a8b8; }
+
+.rp-choice {
+    display: inline-block;
+    padding: 4px 10px;
+    border-radius: 6px;
+    background: rgba(155, 127, 212, 0.15);
+    border: 1px solid rgba(155, 127, 212, 0.35);
+}
+.rp-choice-tag {
+    font-size: 0.72rem;
+    text-transform: uppercase;
+    letter-spacing: 0.04em;
+    color: #9b7fd4;
+    font-weight: 600;
+}
+.rp-choice-label { color: #d4c8f0; }
+
+.dice-user-override {
+    margin-top: 8px;
+    padding-top: 8px;
+    border-top: 1px dashed rgba(255, 255, 255, 0.12);
+}
+.dice-user-badge {
+    font-size: 0.7rem;
+    text-transform: uppercase;
+    letter-spacing: 0.05em;
+    margin-bottom: 6px;
+    opacity: 0.9;
+}
+.dice-user-row {
+    display: flex;
+    flex-wrap: wrap;
+    align-items: baseline;
+    gap: 6px;
+    line-height: 1.45;
+}
+.dice-intent-struck {
+    text-decoration: line-through;
+    opacity: 0.55;
+    color: #888;
+}
+.dice-intent-arrow { color: #9b7fd4; font-weight: bold; }
+.dice-intent-resolved { color: #c8e6c9; font-style: italic; }
+.message.user.has-dice-override .bubble { border-color: rgba(155, 127, 212, 0.4); }
+
+.quest-panel-actions {
+    display: flex;
+    gap: 6px;
+    padding: 8px 10px;
+    border-top: 1px solid #333;
+}
+.quest-panel-actions button {
+    flex: 1;
+    font-size: 0.75rem;
+    padding: 6px 8px;
+    border-radius: 4px;
+    border: 1px solid #444;
+    background: #2a2a3e;
+    color: #ccc;
+    cursor: pointer;
+}
+.quest-panel-actions button:hover:not(:disabled) { background: #3a3a52; }
+.quest-panel-actions button:disabled { opacity: 0.4; cursor: default; }
+.quest-panel-actions .quest-btn-done:hover:not(:disabled) { border-color: #2ecc71; color: #2ecc71; }
+.quest-panel-actions .quest-btn-fail:hover:not(:disabled) { border-color: #e74c3c; color: #e74c3c; }
+
+.format-btn {
+    margin-top: 4px;
+    padding: 2px 8px;
+    font-size: 0.75rem;
+    border: 1px solid #444;
+    border-radius: 4px;
+    background: transparent;
+    color: #aaa;
+    cursor: pointer;
+    align-self: flex-start;
+}
+.format-btn:hover { background: #4a90d9; color: white; border-color: #4a90d9; }
 .system-blob-content {
     white-space: pre-wrap;
     word-break: break-word;
@@ -283,7 +365,16 @@ header h1 { font-size: 1.1rem; color: #e94560; }
 .translate-btn:hover { background: #4a90d9; color: white; }
 .translate-btn:disabled { opacity: 0.5; cursor: default; }
 
-.chat-image { margin-top: 8px; max-width: 100%; border-radius: 8px; border: 1px solid #0f3460; }
+.chat-image-wrap { margin-top: 8px; }
+.chat-image-label {
+    font-size: 0.75rem;
+    color: #888;
+    margin-bottom: 4px;
+    text-transform: uppercase;
+    letter-spacing: 0.04em;
+}
+.chat-image { max-width: 100%; border-radius: 8px; border: 1px solid #0f3460; display: block; }
+.image-prompt-blocks .image-prompt-block + .image-prompt-block { margin-top: 8px; }
 
 .image-generating {
     display: flex;
@@ -312,10 +403,41 @@ header h1 { font-size: 1.1rem; color: #e94560; }
 .image-error { margin-top: 6px; font-size: 0.75rem; color: #888; }
 
 .choice-row {
+    display: flex;
+    flex-direction: column;
+    gap: 10px;
+    margin-top: 8px;
+    position: relative;
+    z-index: 1;
+}
+.choice-section-btns {
     display: flex;
     flex-wrap: wrap;
     gap: 6px;
-    margin-top: 8px;
+}
+.choice-section-label {
+    font-size: 0.72rem;
+    text-transform: uppercase;
+    letter-spacing: 0.04em;
+    color: #c9a227;
+    margin-bottom: 4px;
+}
+.choice-section-label-generic {
+    color: #888;
+}
+.choice-section-plot {
+    padding: 8px 10px;
+    border-radius: 10px;
+    border: 1px solid rgba(201, 162, 39, 0.45);
+    background: rgba(201, 162, 39, 0.08);
+}
+.choice-beat-teaser {
+    font-size: 0.75rem;
+    color: #b8a888;
+    font-style: italic;
+    line-height: 1.35;
+    margin-bottom: 6px;
+    white-space: pre-wrap;
 }
 .choice-btn {
     background: #16213e;
@@ -326,10 +448,20 @@ header h1 { font-size: 1.1rem; color: #e94560; }
     padding: 6px 10px;
     cursor: pointer;
 }
+.choice-btn-plot {
+    border-color: rgba(201, 162, 39, 0.55);
+    background: rgba(201, 162, 39, 0.12);
+    color: #e8d5a3;
+}
 .choice-btn:hover {
     border-color: #e94560;
     color: #e94560;
 }
+.choice-btn-plot:hover {
+    border-color: #c9a227;
+    color: #f0e0b0;
+    background: rgba(201, 162, 39, 0.2);
+}
 
 
 .typing {
@@ -499,6 +631,14 @@ textarea:focus { border-color: #e94560; }
 .outcome-crit-success .dice-outcome { color: #f1c40f; }
 .narrator-text { white-space: pre-wrap; line-height: 1.5; }
 
+.narrator-activity-hint {
+    margin-top: 6px;
+    font-size: 0.72rem;
+    color: #9b7fd4;
+    opacity: 0.9;
+    line-height: 1.35;
+}
+
 /* Affinity display in header */
 .affinity-display {
     font-size: 0.8rem; padding: 4px 10px;
@@ -508,21 +648,137 @@ textarea:focus { border-color: #e94560; }
 .affinity-display.affinity-high { border-color: #e94560; color: #e94560; }
 .affinity-display.affinity-low { border-color: #555; color: #666; }
 
+.stats-display {
+    font-size: 0.72rem;
+    padding: 4px 8px;
+    border: 1px solid #0f3460;
+    border-radius: 10px;
+    color: #9b9bb8;
+    white-space: nowrap;
+}
+.stats-display.stats-warn { border-color: #e9a045; color: #e9a045; }
+.stats-display.stats-critical { border-color: #e74c3c; color: #e74c3c; }
+
+.rpg-debug-panel {
+    margin-top: 12px;
+    padding: 10px;
+    border: 1px dashed #3a3a52;
+    border-radius: 8px;
+    background: rgba(0, 0, 0, 0.15);
+}
+.rpg-debug-panel summary {
+    cursor: pointer;
+    font-size: 0.85rem;
+    color: #9b7fd4;
+}
+.rpg-debug-grid {
+    display: grid;
+    grid-template-columns: 1fr 1fr;
+    gap: 8px;
+    margin: 10px 0;
+}
+.rpg-debug-grid label {
+    font-size: 0.78rem;
+    color: #aaa;
+    display: flex;
+    flex-direction: column;
+    gap: 4px;
+}
+.rpg-debug-grid input {
+    background: #1a1a2e;
+    border: 1px solid #0f3460;
+    border-radius: 6px;
+    color: #e0e0e0;
+    padding: 6px 8px;
+}
+.rpg-debug-apply {
+    background: #3a3a52;
+    border: 1px solid #9b7fd4;
+    border-radius: 8px;
+    color: #e0e0e0;
+    padding: 6px 14px;
+    cursor: pointer;
+    font-size: 0.8rem;
+}
+.rpg-debug-apply:hover { background: #4a4a62; }
+.rpg-debug-status {
+    margin-left: 8px;
+    font-size: 0.75rem;
+    color: #2ecc71;
+}
+
+.context-editor-modal { max-width: 640px; }
+.context-editor-body label {
+    display: block;
+    font-size: 0.78rem;
+    color: #aaa;
+    margin-bottom: 10px;
+}
+.context-editor-body textarea,
+.context-editor-body input[type="number"] {
+    width: 100%;
+    margin-top: 4px;
+    background: #1a1a2e;
+    border: 1px solid #0f3460;
+    border-radius: 8px;
+    color: #e0e0e0;
+    padding: 8px 10px;
+    font-family: Consolas, 'Segoe UI', monospace;
+    font-size: 0.78rem;
+    resize: vertical;
+}
+.context-editor-stats {
+    display: grid;
+    grid-template-columns: repeat(4, 1fr);
+    gap: 8px;
+}
+#contextEditorOpen {
+    background: transparent;
+    border: 1px solid #0f3460;
+    border-radius: 8px;
+    color: #888;
+    padding: 2px 8px;
+    cursor: pointer;
+    font-size: 0.9rem;
+}
+#contextEditorOpen:hover { border-color: #9b7fd4; color: #9b7fd4; }
+
 /* Quest panel in sidebar */
 .quest-panel {
     border-top: 1px solid #0f3460;
     padding: 10px 14px;
     flex-shrink: 0;
+    display: flex;
+    flex-direction: column;
+    max-height: min(42vh, 280px);
+    min-height: 0;
+}
+#questList {
+    overflow-y: auto;
+    flex: 1;
+    min-height: 0;
 }
 .quest-panel-header {
     font-size: 0.75rem; color: #888;
     text-transform: uppercase; letter-spacing: 0.05em;
-    margin-bottom: 6px;
+    margin-bottom: 4px;
+}
+.quest-panel-hint {
+    font-size: 0.68rem;
+    color: #555;
+    margin: 0 0 6px;
+    line-height: 1.3;
 }
 .quest-item {
-    font-size: 0.8rem; padding: 4px 0;
+    font-size: 0.8rem; padding: 4px 6px;
     color: #bbb; line-height: 1.4;
     border-bottom: 1px solid #0f3460;
+    cursor: pointer;
+    user-select: none;
+}
+.quest-item.quest-selected {
+    background: rgba(233, 69, 96, 0.12);
+    border-radius: 4px;
 }
 .quest-item:last-child { border-bottom: none; }
 .quest-done { color: #555; text-decoration: line-through; }
@@ -587,6 +843,11 @@ textarea:focus { border-color: #e94560; }
     flex-direction: row !important;
     padding: 8px 0;
 }
+.hint-text {
+    font-size: 0.8rem;
+    color: #888;
+    margin: 0 0 8px;
+}
 .chat-settings-meta {
     margin-top: 12px; padding: 10px;
     background: #1a1a2e; border-radius: 8px;
@@ -610,7 +871,9 @@ textarea:focus { border-color: #e94560; }
     outline: none; border-bottom: 1px dashed #e94560;
 }
 .message-actions {
-    display: flex; gap: 4px; margin-top: 4px; flex-wrap: wrap;
+    display: flex; gap: 4px; margin-top: 6px; flex-wrap: wrap;
+    position: relative;
+    z-index: 2;
 }
 .message-actions button {
     background: #0f3460; border: none; border-radius: 6px;
diff --git a/static/css/debug.css b/static/css/debug.css
new file mode 100644
index 0000000..e0bcf19
--- /dev/null
+++ b/static/css/debug.css
@@ -0,0 +1,209 @@
+/* app.css sets body { overflow: hidden; height: 100vh } for chat layout */
+html:has(body.debug-page),
+body.debug-page {
+    height: auto;
+    min-height: 100vh;
+    overflow-x: hidden;
+    overflow-y: auto;
+}
+
+.debug-page {
+    background: #0f0f1a;
+    color: #ddd;
+    min-height: 100vh;
+    padding-bottom: 48px;
+}
+
+.debug-header {
+    display: flex;
+    align-items: center;
+    gap: 16px;
+    padding: 12px 20px;
+    border-bottom: 1px solid #1a2744;
+    background: #16213e;
+}
+
+.debug-header a {
+    color: #9b7fd4;
+    text-decoration: none;
+}
+
+.debug-header h1 {
+    flex: 1;
+    margin: 0;
+    font-size: 1.1rem;
+}
+
+.debug-tabs {
+    display: flex;
+    gap: 4px;
+    padding: 8px 16px;
+    background: #12121f;
+    border-bottom: 1px solid #1a2744;
+    flex-wrap: wrap;
+}
+
+.debug-tabs button {
+    background: transparent;
+    border: 1px solid #2a3a5c;
+    color: #aaa;
+    padding: 8px 14px;
+    border-radius: 8px;
+    cursor: pointer;
+}
+
+.debug-tabs button.active {
+    background: #1a2744;
+    color: #e94560;
+    border-color: #e94560;
+}
+
+.debug-main {
+    padding: 16px 20px 40px;
+    max-width: 1200px;
+    margin: 0 auto;
+    overflow: visible;
+}
+
+.debug-panel {
+    display: none;
+}
+
+.debug-panel.active {
+    display: block;
+}
+
+.debug-grid {
+    display: grid;
+    grid-template-columns: repeat(auto-fill, minmax(200px, 1fr));
+    gap: 12px;
+    margin-bottom: 12px;
+}
+
+.debug-grid label,
+.debug-main > label {
+    display: flex;
+    flex-direction: column;
+    gap: 4px;
+    font-size: 0.85rem;
+    color: #aaa;
+    margin-bottom: 10px;
+}
+
+.debug-grid input,
+.debug-grid select,
+.debug-main textarea,
+.debug-main input,
+.debug-main select {
+    background: #1a1a2e;
+    border: 1px solid #0f3460;
+    color: #eee;
+    border-radius: 6px;
+    padding: 8px;
+    font-family: inherit;
+}
+
+.debug-main textarea {
+    width: 100%;
+    box-sizing: border-box;
+    font-family: ui-monospace, monospace;
+    font-size: 0.85rem;
+}
+
+.debug-btn {
+    background: #1a2744;
+    border: 1px solid #3a5080;
+    color: #ccc;
+    padding: 8px 16px;
+    border-radius: 8px;
+    cursor: pointer;
+    margin-bottom: 12px;
+}
+
+.debug-btn.primary {
+    background: #e94560;
+    border-color: #e94560;
+    color: #fff;
+}
+
+.debug-btn:hover {
+    filter: brightness(1.1);
+}
+
+.debug-row {
+    display: flex;
+    flex-wrap: wrap;
+    gap: 8px;
+    margin-bottom: 8px;
+}
+
+.debug-out {
+    background: #0a0a14;
+    border: 1px solid #1a2744;
+    border-radius: 8px;
+    padding: 12px;
+    overflow: auto;
+    max-height: 420px;
+    font-size: 0.8rem;
+    white-space: pre-wrap;
+    word-break: break-word;
+}
+
+.debug-out.compact {
+    max-height: 160px;
+}
+
+.debug-out.small {
+    max-height: 240px;
+}
+
+.debug-split {
+    display: grid;
+    grid-template-columns: 1fr 1fr;
+    gap: 12px;
+}
+
+@media (max-width: 800px) {
+    .debug-split {
+        grid-template-columns: 1fr;
+    }
+}
+
+.debug-split h3,
+.debug-main h3 {
+    font-size: 0.9rem;
+    color: #9b7fd4;
+    margin: 16px 0 8px;
+}
+
+.debug-img-wrap {
+    margin: 12px 0;
+}
+
+.debug-img-wrap img {
+    max-width: 100%;
+    max-height: 512px;
+    border-radius: 8px;
+    border: 1px solid #333;
+}
+
+.debug-img-wrap.hidden {
+    display: none;
+}
+
+.model-list-block {
+    margin-bottom: 8px;
+}
+
+.model-list-block summary {
+    cursor: pointer;
+    color: #9b7fd4;
+}
+
+.model-list-block ul {
+    margin: 4px 0 0;
+    padding-left: 1.2rem;
+    font-size: 0.8rem;
+    max-height: 120px;
+    overflow: auto;
+}
diff --git a/static/debug.html b/static/debug.html
new file mode 100644
index 0000000..6fff947
--- /dev/null
+++ b/static/debug.html
@@ -0,0 +1,134 @@
+<!DOCTYPE html>
+<html lang="ru">
+<head>
+    <meta charset="UTF-8">
+    <meta name="viewport" content="width=device-width, initial-scale=1.0">
+    <title>Debug — AI ChatBot</title>
+    <link rel="stylesheet" href="/static/css/app.css">
+    <link rel="stylesheet" href="/static/css/debug.css">
+</head>
+<body class="debug-page">
+<header class="debug-header">
+    <a href="/">← Чат</a>
+    <h1>Debug</h1>
+    <button type="button" id="btnReloadConfig">↻ Config</button>
+</header>
+
+<nav class="debug-tabs" id="debugTabs">
+    <button type="button" class="active" data-tab="config">Config</button>
+    <button type="button" data-tab="sdprompt">SD Prompt</button>
+    <button type="button" data-tab="llm">LLM</button>
+    <button type="button" data-tab="comfy">ComfyUI</button>
+</nav>
+
+<main class="debug-main">
+    <section class="debug-panel active" id="panel-config">
+        <pre id="configOut" class="debug-out">Загрузка…</pre>
+    </section>
+
+    <section class="debug-panel" id="panel-sdprompt">
+        <div class="debug-grid">
+            <label>Персонаж
+                <select id="sdPersona"></select>
+            </label>
+            <label>Outfit JSON
+                <input type="text" id="sdOutfit" value="[]" placeholder='["dress", "barefoot"]'>
+            </label>
+            <label>Appearance override (опц.)
+                <input type="text" id="sdAppearance" placeholder="оставьте пустым — из карточки">
+            </label>
+            <label>
+                <input type="checkbox" id="sdUseProse"> Использовать prose для Anima
+            </label>
+        </div>
+        <label>Чат (user: / assistant:)
+            <textarea id="sdChat" rows="8" placeholder="user: привет&#10;assistant: *улыбается*"></textarea>
+        </label>
+        <button type="button" class="debug-btn primary" id="btnSdPrompt">Собрать промпт (SD_PROMPT_MODEL)</button>
+        <div class="debug-split">
+            <div>
+                <h3>Scene JSON</h3>
+                <pre id="sdScene" class="debug-out">—</pre>
+            </div>
+            <div>
+                <h3>Теги / гибрид</h3>
+                <pre id="sdPrompts" class="debug-out">—</pre>
+            </div>
+        </div>
+        <details>
+            <summary>LLM raw + builder</summary>
+            <pre id="sdLlmRaw" class="debug-out small">—</pre>
+        </details>
+    </section>
+
+    <section class="debug-panel" id="panel-llm">
+        <div class="debug-grid">
+            <label>Model
+                <input type="text" id="llmModel" placeholder="пусто = SD_PROMPT_MODEL / SYSTEM">
+            </label>
+        </div>
+        <label>System
+            <textarea id="llmSystem" rows="4"></textarea>
+        </label>
+        <label>User
+            <textarea id="llmUser" rows="6"></textarea>
+        </label>
+        <button type="button" class="debug-btn primary" id="btnLlm">Отправить</button>
+        <pre id="llmOut" class="debug-out">—</pre>
+    </section>
+
+    <section class="debug-panel" id="panel-comfy">
+        <div class="debug-row">
+            <button type="button" class="debug-btn" id="btnComfyPing">Ping /system_stats</button>
+            <button type="button" class="debug-btn" id="btnComfyModels">Загрузить модели (/object_info)</button>
+        </div>
+        <pre id="comfyPingOut" class="debug-out compact">—</pre>
+
+        <h3>Модели в Comfy</h3>
+        <div class="debug-grid" id="comfyModelLists">—</div>
+
+        <h3>Генерация</h3>
+        <div class="debug-grid">
+            <label>UNET <select id="genUnet"><option value="">— env —</option></select></label>
+            <label>CLIP <select id="genClip"><option value="">— env —</option></select></label>
+            <label>VAE <select id="genVae"><option value="">— env —</option></select></label>
+            <label>Checkpoint <select id="genCkpt"><option value="">— env / Anima —</option></select></label>
+        </div>
+        <label>Positive
+            <textarea id="genPositive" rows="4"></textarea>
+        </label>
+        <label>Negative
+            <textarea id="genNegative" rows="2"></textarea>
+        </label>
+        <button type="button" class="debug-btn primary" id="btnComfyGen">Сгенерировать</button>
+        <div id="comfyImgWrap" class="debug-img-wrap hidden">
+            <img id="comfyImg" alt="result">
+        </div>
+        <pre id="comfyGenOut" class="debug-out compact">—</pre>
+
+        <h3>Raw API</h3>
+        <div class="debug-grid">
+            <label>Method
+                <select id="rawMethod">
+                    <option>GET</option>
+                    <option>POST</option>
+                </select>
+            </label>
+            <label>Path
+                <input type="text" id="rawPath" value="/system_stats">
+            </label>
+        </div>
+        <label>Query JSON
+            <textarea id="rawParams" rows="2">{}</textarea>
+        </label>
+        <label>Body JSON (POST)
+            <textarea id="rawBody" rows="4" placeholder="{}"></textarea>
+        </label>
+        <button type="button" class="debug-btn" id="btnComfyRaw">Выполнить</button>
+        <pre id="comfyRawOut" class="debug-out">—</pre>
+    </section>
+</main>
+
+<script type="module" src="/static/js/debug.js"></script>
+</body>
+</html>
diff --git a/static/index.html b/static/index.html
index 98b9646..d84b585 100644
--- a/static/index.html
+++ b/static/index.html
@@ -13,8 +13,10 @@
     <h1>🤖 AI Chat</h1>
     <span class="header-title" id="headerTitle">Новый чат</span>
     <span id="rpgBadge" class="rpg-badge hidden" title="RPG режим">RPG</span>
+    <a href="/debug" class="header-icon-btn" title="Debug" style="text-decoration:none">🛠</a>
     <button id="chatSettingsBtn" type="button" class="header-icon-btn" title="Настройки чата">⚙️</button>
-    <span id="affinityDisplay" class="affinity-display hidden"></span>
+    <span id="affinityDisplay" class="affinity-display hidden" title="Симпатия к игроку"></span>
+    <span id="statsDisplay" class="stats-display hidden" title="Шкалы lust / stamina / tension"></span>
 </header>
 
 <div class="app-body">
@@ -26,7 +28,12 @@
         <div class="session-list" id="sessionList"></div>
         <div class="quest-panel hidden" id="questPanel">
             <div class="quest-panel-header">Квесты</div>
+            <p class="quest-panel-hint" id="questPanelHint">Клик по 🔸 — выбор, затем кнопка ниже</p>
             <div id="questList"></div>
+            <div class="quest-panel-actions" id="questPanelActions">
+                <button type="button" id="questBtnDone" class="quest-btn-done" disabled>✓ Готово</button>
+                <button type="button" id="questBtnFail" class="quest-btn-fail" disabled>✗ Провал</button>
+            </div>
         </div>
     </aside>
 
@@ -36,6 +43,7 @@
             <div class="system-blob-header">
                 <span>System</span>
                 <button type="button" id="systemBlobRefresh" title="Обновить">↻</button>
+                <button type="button" id="contextEditorOpen" title="Редактировать контекст">✎</button>
                 <button type="button" id="systemBlobToggle">Скрыть</button>
             </div>
             <pre class="system-blob-content" id="systemBlobContent">—</pre>
@@ -289,6 +297,7 @@
                         <label><input type="checkbox" id="ncSettingNarrator" checked> 📖 Нарратор</label>
                         <label><input type="checkbox" id="ncSettingQuests" checked> 📜 Квесты</label>
                         <label><input type="checkbox" id="ncSettingAffinity" checked> 💖 Симпатия</label>
+                        <label><input type="checkbox" id="ncSettingStats"> 📊 Шкалы (lust/stamina/tension)</label>
                         <label><input type="checkbox" id="ncSettingChoices" checked> 🔘 Кнопки выбора</label>
                     </div>
                 </div>
@@ -314,6 +323,9 @@
             <label>Название чата
                 <input type="text" id="chatSettingsTitle">
             </label>
+            <p class="wizard-page-title">Персонаж чата</p>
+            <p class="hint-text">Смена персонажа перепривязывает этот чат. Историю можно сохранить или очистить.</p>
+            <div class="persona-pick-grid" id="chatSettingsPersonaGrid"></div>
             <label class="rpg-mode-option">
                 <input type="checkbox" id="chatSettingsRpg"> RPG режим
             </label>
@@ -334,9 +346,30 @@
                     <label><input type="checkbox" id="csSettingNarrator"> 📖 Нарратор</label>
                     <label><input type="checkbox" id="csSettingQuests"> 📜 Квесты</label>
                     <label><input type="checkbox" id="csSettingAffinity"> 💖 Симпатия</label>
+                    <label><input type="checkbox" id="csSettingStats"> 📊 Шкалы (lust/stamina/tension)</label>
                     <label><input type="checkbox" id="csSettingChoices"> 🔘 Кнопки выбора</label>
                 </div>
                 <div class="chat-settings-meta" id="chatSettingsMeta"></div>
+                <details class="rpg-debug-panel" id="chatSettingsRpgDebug">
+                    <summary>🧪 Отладка: симпатия и шкалы</summary>
+                    <p class="hint-text">Жёстко задаёт состояние для следующих реплик. Цель — текущий игрок (позже: пары персонаж×игрок).</p>
+                    <div class="rpg-debug-grid">
+                        <label>💖 Симпатия (−30…30)
+                            <input type="number" id="debugAffinity" min="-30" max="30" step="1">
+                        </label>
+                        <label>🔥 Lust (0–10)
+                            <input type="number" id="debugLust" min="0" max="10" step="1">
+                        </label>
+                        <label>⚡ Stamina (0–10)
+                            <input type="number" id="debugStamina" min="0" max="10" step="1">
+                        </label>
+                        <label>😰 Tension (0–10)
+                            <input type="number" id="debugTension" min="0" max="10" step="1">
+                        </label>
+                    </div>
+                    <button type="button" id="debugRpgStateApply" class="rpg-debug-apply">Применить к сессии</button>
+                    <span class="rpg-debug-status" id="debugRpgStateStatus"></span>
+                </details>
             </div>
         </div>
         <div class="modal-buttons modal-wizard-footer">
@@ -346,6 +379,54 @@
     </div>
 </div>
 
-<script type="module" src="/static/js/app.js?v=4"></script>
+<div class="modal-overlay" id="contextEditorModal">
+    <div class="modal modal-wizard context-editor-modal">
+        <div class="modal-wizard-header">
+            <h2>✎ Контекст сессии</h2>
+        </div>
+        <div class="modal-wizard-body context-editor-body">
+            <p class="hint-text">Правки применяются сразу к следующим сообщениям и генерации картинок. JSON-поля — массив или объект.</p>
+            <label>Status quo
+                <textarea id="ctxStatusQuo" rows="3" spellcheck="false"></textarea>
+            </label>
+            <label>Global plot
+                <textarea id="ctxGlobalPlot" rows="2" spellcheck="false"></textarea>
+            </label>
+            <label>Outfit (JSON array, danbooru-теги с цветом)
+                <textarea id="ctxOutfit" rows="4" spellcheck="false" placeholder='["black_sports_shorts", "white_torn_tank_top"]'></textarea>
+            </label>
+            <label>Scene (JSON)
+                <textarea id="ctxScene" rows="4" spellcheck="false"></textarea>
+            </label>
+            <label>Facts (JSON array)
+                <textarea id="ctxFacts" rows="3" spellcheck="false" placeholder='[{"text":"...", "rp_day":"день 1"}]'></textarea>
+            </label>
+            <label>Plot arc (JSON)
+                <textarea id="ctxPlotArc" rows="6" spellcheck="false"></textarea>
+            </label>
+            <div class="context-editor-stats">
+                <label>💖 Affinity
+                    <input type="number" id="ctxAffinity" min="-30" max="30" step="1">
+                </label>
+                <label>🔥 Lust
+                    <input type="number" id="ctxLust" min="0" max="10" step="1">
+                </label>
+                <label>⚡ Stamina
+                    <input type="number" id="ctxStamina" min="0" max="10" step="1">
+                </label>
+                <label>😰 Tension
+                    <input type="number" id="ctxTension" min="0" max="10" step="1">
+                </label>
+            </div>
+            <span class="rpg-debug-status" id="contextEditorStatus"></span>
+        </div>
+        <div class="modal-buttons modal-wizard-footer">
+            <button id="contextEditorCancel" type="button">Отмена</button>
+            <button id="contextEditorSave" type="button" style="background:#9b7fd4;color:white">Сохранить</button>
+        </div>
+    </div>
+</div>
+
+<script type="module" src="/static/js/app.js?v=17"></script>
 </body>
 </html>
diff --git a/static/js/app.js b/static/js/app.js
index 4945c8f..6faaa1c 100644
--- a/static/js/app.js
+++ b/static/js/app.js
@@ -2,8 +2,9 @@ import { toggleSidebar, dom } from './state.js';
 import { initSessions } from './sessions.js';
 import { openNewChatWizard, initNewChatWizard } from './newChatWizard.js';
 import { openChatSettings, initChatSettings } from './chatSettings.js';
+import { initContextEditor } from './contextEditor.js';
 import { loadPersonas, initPersonaModals } from './personas.js';
-import { sendMessage, clearHistory } from './chat.js';
+import { sendMessage, clearHistory, initQuestPanel } from './chat.js';
 
 document.getElementById('sidebarToggle').addEventListener('click', () => {
     const open = toggleSidebar();
@@ -36,5 +37,7 @@ dom.systemBlobToggle?.addEventListener('click', () => {
 initPersonaModals();
 initNewChatWizard();
 initChatSettings();
+initContextEditor();
+initQuestPanel();
 await initSessions();
 loadPersonas();
diff --git a/static/js/chat.js b/static/js/chat.js
index 02266d8..cdb28ad 100644
--- a/static/js/chat.js
+++ b/static/js/chat.js
@@ -1,9 +1,19 @@
 import { sessionId, currentPersona, dom } from './state.js';
-import { parseImagePromptFromContent, copyToClipboard } from './utils.js';
+import { parseImagePromptFromContent, copyToClipboard, splitSdPromptForCopy } from './utils.js';
+import {
+    attachFormatToggle,
+    initBubbleContent,
+    applyDiceOverrideToBubble,
+    isPlayerChoiceContent,
+} from './rpFormat.js';
+
+let _pendingUserBubble = null;
+let _selectedQuestId = null;
+let _questsCache = [];
 
 export async function initChat(options = {}) {
-    if (!sessionId || !currentPersona) return;
-    const payload = { message: '', session_id: sessionId, persona_id: currentPersona };
+    if (!sessionId) return;
+    const payload = { message: '', session_id: sessionId };
     if (options.first_mes_override?.trim()) payload.first_mes_override = options.first_mes_override.trim();
     const res = await fetch('/chat/init', { method: 'POST', headers: { 'Content-Type': 'application/json' }, body: JSON.stringify(payload) });
     if (!res.ok) return;
@@ -16,19 +26,22 @@ export function updateEmptyState() {
     dom.emptyState?.classList.toggle('hidden', !!hasMessages);
 }
 
-export function createImagePromptBlock(promptText) {
+function createImagePromptBlockSingle(label, promptText) {
     const block = document.createElement('div');
     block.className = 'image-prompt-block';
     const header = document.createElement('div');
     header.className = 'image-prompt-header';
-    header.innerHTML = '<span>🎨 SD prompt</span>';
+    header.innerHTML = `<span>🎨 ${label}</span>`;
 
     const copyBtn = document.createElement('button');
     copyBtn.type = 'button';
     copyBtn.className = 'copy-prompt-btn';
     copyBtn.textContent = 'Копировать';
-    copyBtn.addEventListener('click', async () => {
-        const ok = await copyToClipboard(promptText);
+    copyBtn.addEventListener('click', async (e) => {
+        e.preventDefault();
+        e.stopPropagation();
+        const full = textEl.textContent?.trim() || promptText || '';
+        const ok = await copyToClipboard(splitSdPromptForCopy(full));
         copyBtn.textContent = ok ? 'Скопировано' : 'Ошибка';
         setTimeout(() => { copyBtn.textContent = 'Копировать'; }, 1500);
     });
@@ -39,11 +52,10 @@ export function createImagePromptBlock(promptText) {
     regenBtn.className = 'copy-prompt-btn';
     regenBtn.textContent = '🖼 Перегенерировать';
     regenBtn.addEventListener('click', async () => {
-        const wrapper = block.parentElement;
+        const wrapper = block.closest('.message');
         regenBtn.disabled = true;
         regenBtn.textContent = '⏳…';
-        wrapper?.querySelector('.chat-image')?.remove();
-        wrapper?.querySelector('.image-error')?.remove();
+        wrapper?.querySelectorAll('.chat-image-wrap, .chat-image, .image-error').forEach(el => el.remove());
         showImageGenerating(wrapper);
         try {
             const res = await fetch('/images/generate', {
@@ -76,6 +88,26 @@ export function createImagePromptBlock(promptText) {
     return block;
 }
 
+export function createImagePromptBlock(promptText, promptAlt = null) {
+    const wrap = document.createElement('div');
+    wrap.className = 'image-prompt-blocks';
+    wrap.appendChild(createImagePromptBlockSingle('SD prompt', promptText));
+    const alt = (promptAlt || '').trim();
+    const main = (promptText || '').trim();
+    if (alt && alt !== main) {
+        wrap.appendChild(createImagePromptBlockSingle('SD prompt (только теги)', promptAlt));
+    }
+    return wrap;
+}
+
+/** Replace or create tag + optional hybrid prompt blocks under a message. */
+export function ensureImagePromptBlocks(wrapper, tagPrompt, altPrompt = null) {
+    if (!wrapper || !tagPrompt) return;
+    wrapper.querySelector('.image-prompt-blocks')?.remove();
+    wrapper.querySelectorAll('.image-prompt-block').forEach(el => el.remove());
+    wrapper.appendChild(createImagePromptBlock(tagPrompt, altPrompt || null));
+}
+
 const OUTCOME_CLASS = {
     'critical failure': 'outcome-crit-fail',
     'failure': 'outcome-fail',
@@ -113,26 +145,132 @@ function renderNarratorMessage(narrator) {
     return el;
 }
 
-function renderChoices(wrapper, choices) {
-    if (!choices?.length) return;
+export function removeChoiceRows(wrapper) {
+    wrapper?.querySelectorAll('.choice-row').forEach(el => el.remove());
+}
+
+export function ensureMessageActionsLast(wrapper) {
+    const actions = wrapper?.querySelector('.message-actions');
+    if (actions) wrapper.appendChild(actions);
+}
+
+export function renderChoices(wrapper, choices) {
+    if (!choices?.length || !wrapper) return;
+    removeChoiceRows(wrapper);
+
+    const plotChoices = choices.filter(c => c?.source === 'plot_beat');
+    const otherChoices = choices.filter(c => c?.source !== 'plot_beat');
+
     const row = document.createElement('div');
     row.className = 'choice-row';
-    for (const c of choices) {
+
+    const appendBtn = (container, c) => {
         const btn = document.createElement('button');
         btn.type = 'button';
-        btn.className = 'choice-btn';
-        btn.textContent = c.label;
-        btn.addEventListener('click', () => sendMessage(c.label, true));
-        row.appendChild(btn);
-    }
+        btn.className = c.source === 'plot_beat' ? 'choice-btn choice-btn-plot' : 'choice-btn';
+        const label = c.label || '';
+        btn.textContent = label;
+        if (c.beat_title) btn.title = c.beat_title;
+        btn.addEventListener('click', () => {
+            removeChoiceRows(wrapper);
+            sendMessage(label, true);
+        });
+        container.appendChild(btn);
+    };
+
+    const appendSection = (items, { plot } = {}) => {
+        if (!items.length) return;
+        const section = document.createElement('div');
+        section.className = plot ? 'choice-section choice-section-plot' : 'choice-section';
+        if (plot) {
+            const hdr = document.createElement('div');
+            hdr.className = 'choice-section-label';
+            const title = items[0].beat_title || 'Сюжетный beat';
+            hdr.textContent = `📜 ${title}`;
+            section.appendChild(hdr);
+            const inj = (items[0].beat_injection || '').trim();
+            if (inj) {
+                const teaser = document.createElement('div');
+                teaser.className = 'choice-beat-teaser';
+                teaser.textContent = inj.length > 220 ? `${inj.slice(0, 217)}…` : inj;
+                section.appendChild(teaser);
+            }
+        } else if (plotChoices.length) {
+            const hdr = document.createElement('div');
+            hdr.className = 'choice-section-label choice-section-label-generic';
+            hdr.textContent = '🔘 Дальше';
+            section.appendChild(hdr);
+        }
+        const btns = document.createElement('div');
+        btns.className = 'choice-section-btns';
+        for (const c of items) appendBtn(btns, c);
+        section.appendChild(btns);
+        row.appendChild(section);
+    };
+
+    appendSection(plotChoices, { plot: true });
+    appendSection(otherChoices);
+
     wrapper.appendChild(row);
+    ensureMessageActionsLast(wrapper);
+}
+
+function restoreChoicesFromHistory(messages) {
+    const visible = messages.filter(m => m.role !== 'system');
+    if (!visible.length) return;
+    const last = visible[visible.length - 1];
+    if (last.role !== 'assistant' || !last.choices_json) return;
+    let choices = [];
+    try {
+        choices = JSON.parse(last.choices_json);
+    } catch { return; }
+    if (!choices?.length) return;
+    const wrapper = dom.messagesEl.querySelector(
+        `.message.assistant[data-message-id="${last.id}"]`,
+    );
+    if (wrapper) renderChoices(wrapper, choices);
+}
+
+function finalizeAssistantMessage(wrapper, messageId, choices) {
+    if (!wrapper || !messageId) return;
+    wrapper.dataset.messageId = String(messageId);
+    attachMessageActions(wrapper, Number(messageId), 'assistant');
+    if (choices?.length) renderChoices(wrapper, choices);
+    else ensureMessageActionsLast(wrapper);
+}
+
+function showNarratorActivityHint(wrapper, meta) {
+    if (!wrapper || !meta) return;
+    wrapper.querySelector('.narrator-activity-hint')?.remove();
+    const parts = [];
+    if (meta.post_ok === false && meta.pre_ok === false) {
+        parts.push('⚠️ Narrator LLM не ответил — проверьте ROUTER_KEY / RPG_NARRATOR_MODEL');
+    } else {
+        if (meta.dice) parts.push('🎲 бросок');
+        if (meta.directives_count > 0) parts.push(`📋 ${meta.directives_count} указаний`);
+        if (meta.choices_count > 0) parts.push(`🔘 ${meta.choices_count} выборов`);
+        if (meta.status_quo) parts.push('🌍 status_quo');
+        if (meta.beats_replenished) parts.push(`📜 +${meta.beats_replenished} beats`);
+        if (meta.beat_mode === 'after_dice') parts.push('📜 beat (d20)');
+        if (meta.beat_mode === 'llm') parts.push('📜 beat (AI)');
+        if (meta.beat_mode === 'stuck_recovery') parts.push('📜 beat (recovery)');
+        if (meta.beat_mode === 'trigger') parts.push('📜 beat (keywords)');
+        if (meta.arc_pruned) parts.push(`🧹 −${meta.arc_pruned} beat`);
+        if (meta.facts_added) parts.push(`📌 +${meta.facts_added} фактов`);
+    }
+    if (!parts.length) return;
+    const el = document.createElement('div');
+    el.className = 'narrator-activity-hint';
+    el.textContent = `📖 Narrator: ${parts.join(' · ')}`;
+    wrapper.appendChild(el);
+    ensureMessageActionsLast(wrapper);
 }
 
 function renderDebugBlocks(wrapper, blocks) {
     if (!blocks?.length) return;
     for (const b of blocks) {
         if (!b?.text) continue;
-        if (b.type === 'narrator_injection') {
+        if (b.type === 'narrator_injection' || b.type === 'status_quo') {
             const w = document.createElement('div');
             w.className = 'message narrator';
             const lbl = document.createElement('div');
@@ -148,16 +286,106 @@ function renderDebugBlocks(wrapper, blocks) {
     }
 }
 
+function questIdEq(a, b) {
+    return Number(a) === Number(b);
+}
+
+function syncQuestActionButtons() {
+    const doneBtn = document.getElementById('questBtnDone');
+    const failBtn = document.getElementById('questBtnFail');
+    const active = _selectedQuestId != null
+        && _questsCache.some(q => questIdEq(q.id, _selectedQuestId) && q.status === 'active');
+    if (doneBtn) doneBtn.disabled = !active;
+    if (failBtn) failBtn.disabled = !active;
+}
+
 export function updateQuestPanel(quests) {
     const list = document.getElementById('questList');
+    const actions = document.getElementById('questPanelActions');
     if (!list) return;
+    _questsCache = quests || [];
     list.innerHTML = '';
-    for (const q of quests) {
+
+    if (!_questsCache.length) {
+        _selectedQuestId = null;
+        syncQuestActionButtons();
+        return;
+    }
+
+    if (_selectedQuestId != null && !_questsCache.some(
+        q => questIdEq(q.id, _selectedQuestId) && q.status === 'active',
+    )) {
+        _selectedQuestId = null;
+    }
+
+    for (const q of _questsCache) {
         const el = document.createElement('div');
         el.className = `quest-item quest-${q.status}`;
+        if (questIdEq(q.id, _selectedQuestId)) el.classList.add('quest-selected');
+        if (q.status === 'done') el.classList.add('quest-done');
+        if (q.status === 'failed') el.classList.add('quest-failed');
+        el.dataset.questId = String(q.id);
         el.textContent = (q.status === 'done' ? '✅ ' : q.status === 'failed' ? '❌ ' : '🔸 ') + q.title;
+        if (q.status === 'active') {
+            el.addEventListener('click', () => {
+                const qid = Number(q.id);
+                _selectedQuestId = questIdEq(_selectedQuestId, qid) ? null : qid;
+                updateQuestPanel(_questsCache);
+            });
+        }
         list.appendChild(el);
     }
+    syncQuestActionButtons();
+}
+
+async function patchSelectedQuest(status) {
+    if (!sessionId || _selectedQuestId == null) {
+        alert('Сначала выберите активный квест (🔸) в списке.');
+        return;
+    }
+    const q = _questsCache.find(x => questIdEq(x.id, _selectedQuestId));
+    if (!q || q.status !== 'active') return;
+
+    const verb = status === 'done' ? 'завершить' : 'отметить провалом';
+    if (!confirm(`${verb.charAt(0).toUpperCase() + verb.slice(1)} квест?\n\n«${q.title}»`)) return;
+
+    const doneBtn = document.getElementById('questBtnDone');
+    const failBtn = document.getElementById('questBtnFail');
+    doneBtn && (doneBtn.disabled = true);
+    failBtn && (failBtn.disabled = true);
+
+    try {
+        const res = await fetch(`/sessions/${sessionId}/quests/${Number(_selectedQuestId)}`, {
+            method: 'PATCH',
+            headers: { 'Content-Type': 'application/json' },
+            body: JSON.stringify({ status }),
+        });
+        if (!res.ok) {
+            const err = await res.json().catch(() => ({}));
+            alert(err.detail || res.statusText);
+            return;
+        }
+        _selectedQuestId = null;
+        const listRes = await fetch(`/sessions/${sessionId}/quests`);
+        if (listRes.ok) updateQuestPanel(await listRes.json());
+    } finally {
+        syncQuestActionButtons();
+    }
+}
+
+export function initQuestPanel() {
+    const panel = document.getElementById('questPanel');
+    if (!panel || panel.dataset.questBound === '1') return;
+    panel.dataset.questBound = '1';
+    panel.addEventListener('click', (e) => {
+        if (e.target.id === 'questBtnDone') {
+            e.preventDefault();
+            patchSelectedQuest('done');
+        } else if (e.target.id === 'questBtnFail') {
+            e.preventDefault();
+            patchSelectedQuest('failed');
+        }
+    });
 }
 
 export function updateAffinityDisplay(affinity) {
@@ -169,12 +397,38 @@ export function updateAffinityDisplay(affinity) {
     el.className = `affinity-display ${affinity > 5 ? 'affinity-high' : affinity < -3 ? 'affinity-low' : ''}`;
 }
 
-export function appendChatImage(wrapper, imagePath) {
+export function updateStatsDisplay(stats) {
+    const el = dom.statsDisplay;
+    if (!el || !stats) return;
+    const lust = Number(stats.lust ?? 0);
+    const stamina = Number(stats.stamina ?? 10);
+    const tension = Number(stats.tension ?? 0);
+    el.textContent = `🔥${lust} ⚡${stamina} 😰${tension}`;
+    el.className = 'stats-display';
+    if (stamina <= 2 || tension >= 8) el.classList.add('stats-critical');
+    else if (stamina <= 4 || tension >= 6) el.classList.add('stats-warn');
+    el.classList.remove('hidden');
+}
+
+export function hideStatsDisplay() {
+    dom.statsDisplay?.classList.add('hidden');
+}
+
+export function appendChatImage(wrapper, imagePath, label = '') {
     if (!imagePath) return;
+    const figure = document.createElement('figure');
+    figure.className = 'chat-image-wrap';
+    if (label) {
+        const cap = document.createElement('figcaption');
+        cap.className = 'chat-image-label';
+        cap.textContent = label;
+        figure.appendChild(cap);
+    }
     const img = document.createElement('img');
     img.className = 'chat-image';
     img.src = imagePath;
-    wrapper.appendChild(img);
+    figure.appendChild(img);
+    wrapper.appendChild(figure);
 }
 
 export function showImageGenerating(wrapper) {
@@ -194,6 +448,7 @@ export function removeImageGenerating(wrapper) {
 function attachMessageActions(wrapper, messageId, role) {
     if (!messageId) return;
     wrapper.dataset.messageId = String(messageId);
+    wrapper.querySelector('.message-actions')?.remove();
     const actions = document.createElement('div');
     actions.className = 'message-actions';
     const editBtn = document.createElement('button');
@@ -216,7 +471,28 @@ function attachMessageActions(wrapper, messageId, role) {
     branchBtn.title = 'Ветка отсюда';
     branchBtn.addEventListener('click', () => forkFromMessage(messageId));
     actions.appendChild(branchBtn);
+    const delBtn = document.createElement('button');
+    delBtn.type = 'button';
+    delBtn.textContent = '🗑';
+    delBtn.title = 'Удалить сообщение и всё после него';
+    delBtn.addEventListener('click', () => deleteMessageAndAfter(messageId, wrapper));
+    actions.appendChild(delBtn);
     wrapper.appendChild(actions);
+    ensureMessageActionsLast(wrapper);
+}
+
+async function deleteMessageAndAfter(messageId, wrapper) {
+    if (!sessionId || !messageId) return;
+    if (!confirm('Удалить это сообщение и всю переписку после него?')) return;
+    const res = await fetch(`/chat/messages/${messageId}`, { method: 'DELETE' });
+    if (!res.ok) {
+        const err = await res.json().catch(() => ({}));
+        alert(err.detail || res.statusText);
+        return;
+    }
+    await reloadChatFromServer(sessionId);
+    const listRes = await fetch(`/sessions/${sessionId}/quests`);
+    if (listRes.ok) updateQuestPanel(await listRes.json());
 }
 
 async function startEditMessage(wrapper, messageId) {
@@ -262,7 +538,7 @@ async function regenerateMessage(messageId, wrapper) {
         const res = await fetch('/chat/regenerate', {
             method: 'POST',
             headers: { 'Content-Type': 'application/json' },
-            body: JSON.stringify({ session_id: sessionId, persona_id: currentPersona, message_id: messageId }),
+            body: JSON.stringify({ session_id: sessionId, message_id: messageId }),
         });
         if (!res.ok) throw new Error('Ошибка: ' + res.status);
         removeTyping();
@@ -297,14 +573,25 @@ export async function reloadChatFromServer(id) {
     const messages = await histRes.json();
     clearMessages();
     messages.filter(m => m.role !== 'system').forEach(m => {
+        if (m.role === 'narrator') {
+            try {
+                const data = typeof m.content === 'string' ? JSON.parse(m.content) : m.content;
+                if (data?.text) renderNarratorMessage(data);
+            } catch { /* ignore bad narrator payload */ }
+            return;
+        }
         addMessage(
             m.role === 'user' ? 'user' : 'assistant',
             m.content,
             m.image_prompt,
             m.image_path ? `/static/${m.image_path}` : null,
             m.id,
+            m.image_prompt_alt,
+            m.image_path_alt ? `/static/${m.image_path_alt}` : null,
+            m.role === 'user' ? m.action_resolution : null,
         );
     });
+    restoreChoicesFromHistory(messages);
 }
 
 const IMAGE_PROMPT_RE = /\[IMAGE_PROMPT:.*?\]/gs;
@@ -330,6 +617,22 @@ async function consumeStream(res) {
             // Narrator arrives BEFORE chunks — render immediately
             if (data.narrator) {
                 renderNarratorMessage(data.narrator);
+                if (
+                    _pendingUserBubble
+                    && data.narrator.roll != null
+                    && data.narrator.original_intent
+                    && !isPlayerChoiceContent(_pendingUserBubble.dataset.raw || '')
+                ) {
+                    const resolution = {
+                        intent_text: data.narrator.original_intent,
+                        resolution_text: data.narrator.text,
+                        roll: data.narrator.roll,
+                        outcome: data.narrator.outcome,
+                    };
+                    _pendingUserBubble._diceResolution = resolution;
+                    applyDiceOverrideToBubble(_pendingUserBubble, resolution);
+                    _pendingUserBubble.closest('.message')?.classList.add('has-dice-override');
+                }
             }
 
             if (data.chunk !== undefined) {
@@ -344,8 +647,12 @@ async function consumeStream(res) {
             if (data.image_generating && bubble) {
                 bubble.classList.remove('typing-active');
                 const wrapper = bubble.parentElement;
-                if (data.image_prompt && !wrapper.querySelector('.image-prompt-block')) {
-                    wrapper.appendChild(createImagePromptBlock(data.image_prompt));
+                if (data.image_prompt) {
+                    ensureImagePromptBlocks(
+                        wrapper,
+                        data.image_prompt,
+                        data.image_prompt_alt || null,
+                    );
                 }
                 showImageGenerating(wrapper);
                 dom.messagesEl.scrollTop = dom.messagesEl.scrollHeight;
@@ -356,19 +663,24 @@ async function consumeStream(res) {
                 removeImageGenerating(wrapper);
                 bubble?.classList.remove('typing-active');
 
-                // Strip IMAGE_PROMPT tag from final text
+                // Strip IMAGE_PROMPT tag; apply server-side OOC strip if provided
                 if (bubble) {
-                    bubble.textContent = bubble.textContent.replace(IMAGE_PROMPT_RE, '').trim();
+                    const fromServer = data.assistant_content;
+                    const cleaned = (fromServer ?? bubble.textContent)
+                        .replace(IMAGE_PROMPT_RE, '')
+                        .trim();
+                    initBubbleContent(bubble, cleaned, { formatted: true });
                 }
 
-                if (data.image_prompt && wrapper && !wrapper.querySelector('.image-prompt-block')) {
-                    wrapper.appendChild(createImagePromptBlock(data.image_prompt));
+                if (data.image_prompt && wrapper) {
+                    ensureImagePromptBlocks(
+                        wrapper,
+                        data.image_prompt,
+                        data.image_prompt_alt || null,
+                    );
                 }
                 if (data.image_path && wrapper) {
-                    console.log('[image] appending', data.image_path, 'to', wrapper);
-                    appendChatImage(wrapper, data.image_path);
-                } else {
-                    console.log('[image] skip: image_path=', data.image_path, 'wrapper=', wrapper);
+                    appendChatImage(wrapper, data.image_path, '');
                 }
                 if (data.image_error && wrapper) {
                     const err = document.createElement('div');
@@ -376,11 +688,25 @@ async function consumeStream(res) {
                     err.textContent = '🖼 ' + data.image_error;
                     wrapper.appendChild(err);
                 }
-                if (data.choices?.length && bubble) renderChoices(bubble.parentElement, data.choices);
+                if (bubble?.parentElement && data.assistant_message_id) {
+                    finalizeAssistantMessage(
+                        bubble.parentElement,
+                        data.assistant_message_id,
+                        data.choices,
+                    );
+                } else if (data.choices?.length && bubble) {
+                    renderChoices(bubble.parentElement, data.choices);
+                }
                 if (data.debug) renderDebugBlocks(bubble?.parentElement || dom.messagesEl, data.debug);
+                if (data.narrator_meta && bubble?.parentElement) {
+                    showNarratorActivityHint(bubble.parentElement, data.narrator_meta);
+                }
                 if (data.affinity !== undefined) updateAffinityDisplay(data.affinity);
+                if (data.narrative_stats) updateStatsDisplay(data.narrative_stats);
                 if (data.quests?.length) updateQuestPanel(data.quests);
 
+                _pendingUserBubble = null;
+
                 const { loadSessions } = await import('./sessions.js');
                 loadSessions();
             }
@@ -388,7 +714,16 @@ async function consumeStream(res) {
     }
 }
 
-export function addMessage(role, content = '', imagePrompt = null, imagePath = null, messageId = null) {
+export function addMessage(
+    role,
+    content = '',
+    imagePrompt = null,
+    imagePath = null,
+    messageId = null,
+    imagePromptAlt = null,
+    imagePathAlt = null,
+    actionResolution = null,
+) {
     updateEmptyState();
     const wrapper = document.createElement('div');
     wrapper.className = `message ${role}`;
@@ -407,8 +742,18 @@ export function addMessage(role, content = '', imagePrompt = null, imagePath = n
 
     const bubble = document.createElement('div');
     bubble.className = 'bubble';
-    bubble.textContent = displayContent;
+    initBubbleContent(bubble, displayContent, {
+        formatted: !!displayContent,
+        actionResolution: role === 'user' ? actionResolution : null,
+    });
     wrapper.appendChild(bubble);
+    if (role === 'user' && actionResolution?.resolution_text) {
+        wrapper.classList.add('has-dice-override');
+    }
+
+    if (displayContent) {
+        attachFormatToggle(wrapper, bubble);
+    }
 
     if (role === 'assistant') {
         const translateBtn = document.createElement('button');
@@ -418,12 +763,12 @@ export function addMessage(role, content = '', imagePrompt = null, imagePath = n
         let originalText = null;
         translateBtn.addEventListener('click', async () => {
             if (originalText !== null) {
-                bubble.textContent = originalText;
+                initBubbleContent(bubble, originalText, { formatted: true });
                 originalText = null;
                 translateBtn.textContent = '🌐 RU';
                 return;
             }
-            originalText = bubble.textContent;
+            originalText = bubble.dataset.raw ?? bubble.textContent;
             translateBtn.disabled = true;
             translateBtn.textContent = '…';
             try {
@@ -434,7 +779,8 @@ export function addMessage(role, content = '', imagePrompt = null, imagePath = n
                 });
                 if (!res.ok) throw new Error(res.statusText);
                 const data = await res.json();
-                bubble.textContent = data.translated;
+                initBubbleContent(bubble, data.translated, { formatted: true });
+                bubble.dataset.raw = data.translated;
                 translateBtn.textContent = '↩ Оригинал';
             } catch {
                 originalText = null;
@@ -446,8 +792,9 @@ export function addMessage(role, content = '', imagePrompt = null, imagePath = n
         wrapper.appendChild(translateBtn);
     }
 
-    if (prompt) wrapper.appendChild(createImagePromptBlock(prompt));
-    if (imagePath) appendChatImage(wrapper, imagePath);
+    if (prompt) wrapper.appendChild(createImagePromptBlock(prompt, imagePromptAlt));
+    if (imagePath) appendChatImage(wrapper, imagePath, imagePathAlt ? 'Теги' : '');
+    if (imagePathAlt) appendChatImage(wrapper, imagePathAlt, 'Гибрид');
     attachMessageActions(wrapper, messageId, role);
     dom.messagesEl.appendChild(wrapper);
     dom.messagesEl.scrollTop = dom.messagesEl.scrollHeight;
@@ -481,13 +828,15 @@ export async function sendMessage(text, isNarratorChoice = false) {
     dom.inputEl.value = '';
     dom.inputEl.style.height = 'auto';
     dom.sendBtn.disabled = true;
-    addMessage('user', isNarratorChoice ? `[${text}]` : text);
+    const userContent = isNarratorChoice ? `[Player chose: ${text}]` : text;
+    dom.messagesEl.querySelectorAll('.message.assistant').forEach(w => removeChoiceRows(w));
+    _pendingUserBubble = addMessage('user', userContent);
     showTyping();
     try {
         const res = await fetch('/chat/stream', {
             method: 'POST',
             headers: { 'Content-Type': 'application/json' },
-            body: JSON.stringify({ message: text, session_id: sessionId, persona_id: currentPersona, is_narrator_choice: isNarratorChoice }),
+            body: JSON.stringify({ message: text, session_id: sessionId, is_narrator_choice: isNarratorChoice }),
         });
         if (!res.ok) throw new Error('Ошибка сервера: ' + res.status);
         removeTyping();
diff --git a/static/js/chatSettings.js b/static/js/chatSettings.js
index f907422..2c20878 100644
--- a/static/js/chatSettings.js
+++ b/static/js/chatSettings.js
@@ -1,7 +1,10 @@
-import { sessionId, currentPersona, dom } from './state.js';
+import { sessionId, currentPersona, setCurrentPersona, dom } from './state.js';
 import { GENRE_LABELS, bindGenreGrid, resetGenreGrid } from './utils.js';
+import { personaIndex } from './personas.js';
 
 const chatSettingsGenres = new Set();
+let chatSettingsPersonaId = 'default';
+let chatSettingsInitialPersonaId = 'default';
 
 function updateChatSettingsGenresLabel() {
     const el = document.getElementById('chatSettingsGenresLabel');
@@ -15,11 +18,32 @@ function updateChatSettingsGenresLabel() {
     }
 }
 
+function fillChatSettingsPersonaGrid() {
+    const grid = document.getElementById('chatSettingsPersonaGrid');
+    if (!grid) return;
+    grid.innerHTML = '';
+    for (const p of personaIndex.values()) {
+        const card = document.createElement('button');
+        card.type = 'button';
+        card.className = 'persona-pick-card' + (p.persona_id === chatSettingsPersonaId ? ' selected' : '');
+        card.dataset.id = p.persona_id;
+        card.innerHTML = `<span class="emoji">${p.emoji || '🤖'}</span>${p.name}`;
+        card.addEventListener('click', () => {
+            chatSettingsPersonaId = p.persona_id;
+            grid.querySelectorAll('.persona-pick-card').forEach(c => {
+                c.classList.toggle('selected', c.dataset.id === chatSettingsPersonaId);
+            });
+        });
+        grid.appendChild(card);
+    }
+}
+
 function loadRpgSettingsToDom(prefix, settings) {
     document.getElementById(`${prefix}SettingDice`).checked = settings.dice !== false;
     document.getElementById(`${prefix}SettingNarrator`).checked = settings.narrator !== false;
     document.getElementById(`${prefix}SettingQuests`).checked = settings.quests !== false;
     document.getElementById(`${prefix}SettingAffinity`).checked = settings.affinity !== false;
+    document.getElementById(`${prefix}SettingStats`).checked = settings.stats === true;
     document.getElementById(`${prefix}SettingChoices`).checked = settings.choices !== false;
 }
 
@@ -29,6 +53,7 @@ function readRpgSettingsFromDom(prefix) {
         narrator: document.getElementById(`${prefix}SettingNarrator`)?.checked ?? true,
         quests: document.getElementById(`${prefix}SettingQuests`)?.checked ?? true,
         affinity: document.getElementById(`${prefix}SettingAffinity`)?.checked ?? true,
+        stats: document.getElementById(`${prefix}SettingStats`)?.checked ?? false,
         choices: document.getElementById(`${prefix}SettingChoices`)?.checked ?? true,
     };
 }
@@ -51,6 +76,10 @@ async function bootstrapRpg(sid, personaId, genreValue, settings) {
     });
     if (res.ok) {
         const data = await res.json();
+        if (data.affinity !== undefined) {
+            const { updateAffinityDisplay } = await import('./chat.js');
+            updateAffinityDisplay(data.affinity);
+        }
         if (data.quests) updateQuestPanel(data.quests);
         if (data.plot_arc) {
             const title = data.plot_arc.title || '';
@@ -67,6 +96,10 @@ export async function openChatSettings() {
     const s = await res.json();
 
     document.getElementById('chatSettingsTitle').value = s.title || '';
+    chatSettingsPersonaId = s.persona_id || 'default';
+    chatSettingsInitialPersonaId = chatSettingsPersonaId;
+    fillChatSettingsPersonaGrid();
+
     const rpgOn = !!s.rpg_enabled;
     document.getElementById('chatSettingsRpg').checked = rpgOn;
     document.getElementById('chatSettingsRpgBlock').classList.toggle('hidden', !rpgOn);
@@ -91,12 +124,33 @@ export async function openChatSettings() {
         const arc = JSON.parse(s.plot_arc_json || '{}');
         phase = arc.phase || '';
     } catch { /* ignore */ }
+    let stats = { lust: 0, stamina: 10, tension: 0 };
+    try {
+        stats = { ...stats, ...JSON.parse(s.narrative_stats_json || '{}') };
+    } catch { /* ignore */ }
+
     document.getElementById('chatSettingsMeta').innerHTML = [
         `Симпатия: ${s.affinity ?? 0}`,
+        settings.stats
+            ? `Шкалы: lust ${stats.lust ?? 0}, stamina ${stats.stamina ?? 10}, tension ${stats.tension ?? 0}`
+            : '',
         s.genre ? `Жанр: ${(s.genre || '').split(',').map(g => GENRE_LABELS[g.trim()] || g).join(' + ')}` : '',
         phase ? `Фаза арки: ${phase}` : '',
     ].filter(Boolean).join('<br>');
 
+    const dbg = document.getElementById('chatSettingsRpgDebug');
+    if (dbg) dbg.open = rpgOn;
+    const affIn = document.getElementById('debugAffinity');
+    const lustIn = document.getElementById('debugLust');
+    const stamIn = document.getElementById('debugStamina');
+    const tensIn = document.getElementById('debugTension');
+    if (affIn) affIn.value = String(s.affinity ?? 0);
+    if (lustIn) lustIn.value = String(stats.lust ?? 0);
+    if (stamIn) stamIn.value = String(stats.stamina ?? 10);
+    if (tensIn) tensIn.value = String(stats.tension ?? 0);
+    const st = document.getElementById('debugRpgStateStatus');
+    if (st) st.textContent = '';
+
     document.getElementById('chatSettingsModal').classList.add('open');
 }
 
@@ -115,15 +169,94 @@ export function initChatSettings() {
         document.getElementById('chatSettingsModal').classList.remove('open');
     });
 
+    document.getElementById('debugRpgStateApply')?.addEventListener('click', async () => {
+        if (!sessionId) return;
+        const statusEl = document.getElementById('debugRpgStateStatus');
+        const body = {};
+        const aff = document.getElementById('debugAffinity')?.value;
+        const lust = document.getElementById('debugLust')?.value;
+        const stam = document.getElementById('debugStamina')?.value;
+        const tens = document.getElementById('debugTension')?.value;
+        if (aff !== '' && aff != null) body.affinity = Number(aff);
+        if (lust !== '' && lust != null) body.lust = Number(lust);
+        if (stam !== '' && stam != null) body.stamina = Number(stam);
+        if (tens !== '' && tens != null) body.tension = Number(tens);
+        try {
+            const res = await fetch(`/sessions/${sessionId}/rpg-state`, {
+                method: 'PATCH',
+                headers: { 'Content-Type': 'application/json' },
+                body: JSON.stringify(body),
+            });
+            if (!res.ok) {
+                const err = await res.json().catch(() => ({}));
+                if (statusEl) {
+                    statusEl.textContent = err.detail || 'Ошибка';
+                    statusEl.style.color = '#e74c3c';
+                }
+                return;
+            }
+            const data = await res.json();
+            const { updateAffinityDisplay, updateStatsDisplay } = await import('./chat.js');
+            if (data.affinity !== undefined) updateAffinityDisplay(data.affinity);
+            if (data.narrative_stats) updateStatsDisplay(data.narrative_stats);
+            if (statusEl) {
+                statusEl.textContent = 'Сохранено';
+                statusEl.style.color = '#2ecc71';
+            }
+            const blobRes = await fetch(`/chat/system/${sessionId}`);
+            if (blobRes.ok) {
+                const { renderSystemBlob } = await import('./sessions.js');
+                renderSystemBlob(await blobRes.json());
+            }
+        } catch {
+            if (statusEl) {
+                statusEl.textContent = 'Сеть';
+                statusEl.style.color = '#e74c3c';
+            }
+        }
+    });
+
     document.getElementById('chatSettingsSave')?.addEventListener('click', async () => {
         if (!sessionId) return;
-        const { loadSessions, applySessionUi } = await import('./sessions.js');
+        const { loadSessions, applySessionUi, renderSystemBlob } = await import('./sessions.js');
+        const { reloadChatFromServer } = await import('./chat.js');
+        const { highlightPersonaBar } = await import('./personas.js');
 
         const title = document.getElementById('chatSettingsTitle').value.trim();
         const rpgOn = document.getElementById('chatSettingsRpg').checked;
         const genreValue = [...chatSettingsGenres].join(',') || 'adventure';
         const settings = readRpgSettingsFromDom('cs');
 
+        if (chatSettingsPersonaId !== chatSettingsInitialPersonaId) {
+            const pName = personaIndex.get(chatSettingsPersonaId)?.name || chatSettingsPersonaId;
+            const keepHistory = confirm(
+                `Перепривязать чат к «${pName}»?\n\n`
+                + 'OK — сохранить историю сообщений (персонаж в старых репликах может не совпадать).\n'
+                + 'Отмена — очистить историю и начать с приветствия нового персонажа.',
+            );
+            const clearHistory = !keepHistory;
+
+            const rebindRes = await fetch(`/sessions/${sessionId}/rebind-persona`, {
+                method: 'POST',
+                headers: { 'Content-Type': 'application/json' },
+                body: JSON.stringify({
+                    persona_id: chatSettingsPersonaId,
+                    clear_history: clearHistory,
+                }),
+            });
+            if (!rebindRes.ok) {
+                const err = await rebindRes.json().catch(() => ({}));
+                alert(err.detail || 'Не удалось сменить персонажа');
+                return;
+            }
+            setCurrentPersona(chatSettingsPersonaId);
+            chatSettingsInitialPersonaId = chatSettingsPersonaId;
+            highlightPersonaBar(chatSettingsPersonaId);
+            await reloadChatFromServer(sessionId);
+            const blobRes = await fetch(`/chat/system/${sessionId}`);
+            if (blobRes.ok) renderSystemBlob(await blobRes.json());
+        }
+
         await fetch(`/sessions/${sessionId}`, {
             method: 'PATCH',
             headers: { 'Content-Type': 'application/json' },
@@ -141,7 +274,7 @@ export function initChatSettings() {
             let arc = {};
             try { arc = JSON.parse(s.plot_arc_json || '{}'); } catch { /* ignore */ }
             if (!arc || !Object.keys(arc).length) {
-                await bootstrapRpg(sessionId, currentPersona, genreValue, settings);
+                await bootstrapRpg(sessionId, chatSettingsPersonaId, genreValue, settings);
             }
         }
 
diff --git a/static/js/contextEditor.js b/static/js/contextEditor.js
new file mode 100644
index 0000000..1f2b1d5
--- /dev/null
+++ b/static/js/contextEditor.js
@@ -0,0 +1,105 @@
+import { sessionId } from './state.js';
+
+function fmtJson(raw, fallback) {
+    try {
+        return JSON.stringify(JSON.parse(raw || fallback), null, 2);
+    } catch {
+        return raw || fallback;
+    }
+}
+
+function parseStats(raw) {
+    try {
+        return { lust: 0, stamina: 10, tension: 0, ...JSON.parse(raw || '{}') };
+    } catch {
+        return { lust: 0, stamina: 10, tension: 0 };
+    }
+}
+
+export async function openContextEditor() {
+    if (!sessionId) return;
+    const res = await fetch(`/chat/system/${sessionId}`);
+    if (!res.ok) return;
+    const blob = await res.json();
+
+    document.getElementById('ctxStatusQuo').value = blob.status_quo || '';
+    document.getElementById('ctxGlobalPlot').value = blob.global_plot || '';
+    document.getElementById('ctxOutfit').value = fmtJson(blob.outfit_json, '[]');
+    document.getElementById('ctxScene').value = fmtJson(blob.scene_json, '{}');
+    document.getElementById('ctxFacts').value = fmtJson(blob.facts_json, '[]');
+    document.getElementById('ctxPlotArc').value = fmtJson(blob.plot_arc_json, '{}');
+    document.getElementById('ctxAffinity').value = String(blob.affinity ?? 0);
+
+    const stats = parseStats(blob.narrative_stats_json);
+    document.getElementById('ctxLust').value = String(stats.lust ?? 0);
+    document.getElementById('ctxStamina').value = String(stats.stamina ?? 10);
+    document.getElementById('ctxTension').value = String(stats.tension ?? 0);
+
+    const st = document.getElementById('contextEditorStatus');
+    if (st) {
+        st.textContent = '';
+        st.style.color = '';
+    }
+    document.getElementById('contextEditorModal')?.classList.add('open');
+}
+
+export function initContextEditor() {
+    document.getElementById('contextEditorOpen')?.addEventListener('click', () => {
+        openContextEditor();
+    });
+    document.getElementById('contextEditorCancel')?.addEventListener('click', () => {
+        document.getElementById('contextEditorModal')?.classList.remove('open');
+    });
+    document.getElementById('contextEditorSave')?.addEventListener('click', async () => {
+        if (!sessionId) return;
+        const statusEl = document.getElementById('contextEditorStatus');
+        const body = {
+            status_quo: document.getElementById('ctxStatusQuo')?.value ?? '',
+            global_plot: document.getElementById('ctxGlobalPlot')?.value ?? '',
+            outfit_json: document.getElementById('ctxOutfit')?.value ?? '[]',
+            scene_json: document.getElementById('ctxScene')?.value ?? '{}',
+            facts_json: document.getElementById('ctxFacts')?.value ?? '[]',
+            plot_arc_json: document.getElementById('ctxPlotArc')?.value ?? '{}',
+            affinity: Number(document.getElementById('ctxAffinity')?.value ?? 0),
+            lust: Number(document.getElementById('ctxLust')?.value ?? 0),
+            stamina: Number(document.getElementById('ctxStamina')?.value ?? 10),
+            tension: Number(document.getElementById('ctxTension')?.value ?? 0),
+        };
+        try {
+            const res = await fetch(`/sessions/${sessionId}/context`, {
+                method: 'PATCH',
+                headers: { 'Content-Type': 'application/json' },
+                body: JSON.stringify(body),
+            });
+            if (!res.ok) {
+                const err = await res.json().catch(() => ({}));
+                if (statusEl) {
+                    statusEl.textContent = err.detail || res.statusText;
+                    statusEl.style.color = '#e74c3c';
+                }
+                return;
+            }
+            const data = await res.json();
+            if (data.outfit_json) {
+                document.getElementById('ctxOutfit').value = fmtJson(data.outfit_json, '[]');
+            }
+            const { renderSystemBlob, applySessionUi } = await import('./sessions.js');
+            const { updateAffinityDisplay, updateStatsDisplay } = await import('./chat.js');
+            const blobRes = await fetch(`/chat/system/${sessionId}`);
+            if (blobRes.ok) renderSystemBlob(await blobRes.json());
+            const sessRes = await fetch(`/sessions/${sessionId}`);
+            if (sessRes.ok) applySessionUi(await sessRes.json());
+            if (data.affinity !== undefined) updateAffinityDisplay(data.affinity);
+            if (data.narrative_stats) updateStatsDisplay(data.narrative_stats);
+            if (statusEl) {
+                statusEl.textContent = 'Сохранено (outfit: цвета добавлены автоматически, если не указаны)';
+                statusEl.style.color = '#2ecc71';
+            }
+        } catch {
+            if (statusEl) {
+                statusEl.textContent = 'Ошибка сети';
+                statusEl.style.color = '#e74c3c';
+            }
+        }
+    });
+}
diff --git a/static/js/debug.js b/static/js/debug.js
new file mode 100644
index 0000000..4815a01
--- /dev/null
+++ b/static/js/debug.js
@@ -0,0 +1,217 @@
+const $ = (id) => document.getElementById(id);
+
+function fmt(obj) {
+    return typeof obj === 'string' ? obj : JSON.stringify(obj, null, 2);
+}
+
+async function api(path, opts = {}) {
+    const res = await fetch(path, {
+        headers: { 'Content-Type': 'application/json', ...(opts.headers || {}) },
+        ...opts,
+    });
+    const text = await res.text();
+    let data;
+    try {
+        data = JSON.parse(text);
+    } catch {
+        data = text;
+    }
+    if (!res.ok) {
+        const detail = data?.detail || text || res.statusText;
+        throw new Error(`${res.status}: ${detail}`);
+    }
+    return data;
+}
+
+function initTabs() {
+    const tabs = document.querySelectorAll('#debugTabs button');
+    tabs.forEach((btn) => {
+        btn.addEventListener('click', () => {
+            tabs.forEach((t) => t.classList.remove('active'));
+            btn.classList.add('active');
+            document.querySelectorAll('.debug-panel').forEach((p) => p.classList.remove('active'));
+            $(`panel-${btn.dataset.tab}`).classList.add('active');
+        });
+    });
+}
+
+async function loadConfig() {
+    const c = await api('/debug/config');
+    $('configOut').textContent = fmt(c);
+    $('llmModel').placeholder = c.sd_prompt_model || c.system_model;
+    return c;
+}
+
+async function loadPersonas() {
+    const list = await api('/debug/personas');
+    const sel = $('sdPersona');
+    sel.innerHTML = '';
+    for (const p of list) {
+        const opt = document.createElement('option');
+        opt.value = p.persona_id;
+        opt.textContent = `${p.name} (${p.persona_id})`;
+        sel.appendChild(opt);
+    }
+}
+
+async function runSdPrompt() {
+    $('sdScene').textContent = '…';
+    $('sdPrompts').textContent = '…';
+    const body = {
+        persona_id: $('sdPersona').value,
+        chat_excerpt: $('sdChat').value,
+        outfit_json: $('sdOutfit').value || '[]',
+        use_prose: $('sdUseProse') ? $('sdUseProse').checked : false,
+    };
+    const app = $('sdAppearance').value.trim();
+    if (app) body.appearance_override = app;
+
+    const data = await api('/debug/sd-prompt', { method: 'POST', body: JSON.stringify(body) });
+    $('sdScene').textContent = data.scene ? fmt(data.scene) : (data.error || '—');
+    const prompts = [];
+    if (data.tags_only_full) prompts.push('=== TAGS + POV (no prose) ===\n' + data.tags_only_full);
+    if (data.hybrid_full) prompts.push('\n=== HYBRID (Comfy) ===\n' + data.hybrid_full);
+    if (!data.tags_only_full && data.tag_full) prompts.push('=== PROMPT ===\n' + data.tag_full);
+    $('sdPrompts').textContent = prompts.join('\n') || data.error || '—';
+    $('sdLlmRaw').textContent = [
+        `model: ${data.sd_prompt_model}`,
+        `dual: ${data.anima_dual}`,
+        '',
+        '--- system ---',
+        data.builder_system || '',
+        '',
+        '--- user ---',
+        data.builder_user || '',
+        '',
+        '--- raw ---',
+        data.llm_raw || data.error || '',
+    ].join('\n');
+    if (data.tag_full || data.hybrid_full) {
+        const src = data.hybrid_full || data.tag_full;
+        const parts = src.includes('__NEGATIVE_PROMPT__') 
+            ? src.split('\n\n__NEGATIVE_PROMPT__\n\n') 
+            : src.includes('\n\nNegative prompt:') 
+                ? src.split('\n\nNegative prompt:') 
+                : [src, ''];
+        $('genPositive').value = parts[0] || '';
+        $('genNegative').value = parts[1] || '';
+    }
+}
+
+async function runLlm() {
+    $('llmOut').textContent = '…';
+    const data = await api('/debug/llm', {
+        method: 'POST',
+        body: JSON.stringify({
+            model: $('llmModel').value.trim(),
+            system: $('llmSystem').value,
+            user: $('llmUser').value,
+        }),
+    });
+    $('llmOut').textContent = `model: ${data.model}\n\n${data.response}`;
+}
+
+function fillModelSelect(sel, options, configured) {
+    const current = sel.querySelector('option')?.value ?? '';
+    sel.innerHTML = `<option value="">— env: ${configured || '—'} —</option>`;
+    for (const name of options || []) {
+        const opt = document.createElement('option');
+        opt.value = name;
+        opt.textContent = name;
+        if (name === configured) opt.selected = true;
+        sel.appendChild(opt);
+    }
+}
+
+async function loadComfyModels() {
+    $('comfyModelLists').textContent = 'Загрузка object_info…';
+    const data = await api('/debug/comfy/models');
+    const { models, configured } = data;
+    fillModelSelect($('genUnet'), models.unets, configured.unet);
+    fillModelSelect($('genClip'), models.clips, configured.clip);
+    fillModelSelect($('genVae'), models.vaes, configured.vae);
+    fillModelSelect($('genCkpt'), models.checkpoints, configured.checkpoint);
+
+    const wrap = $('comfyModelLists');
+    wrap.innerHTML = '';
+    for (const [key, list] of Object.entries(models)) {
+        const block = document.createElement('details');
+        block.className = 'model-list-block';
+        block.open = key === 'unets' || key === 'checkpoints';
+        block.innerHTML = `<summary>${key} (${list.length})</summary>`;
+        const ul = document.createElement('ul');
+        for (const item of list) {
+            const li = document.createElement('li');
+            li.textContent = item;
+            ul.appendChild(li);
+        }
+        block.appendChild(ul);
+        wrap.appendChild(block);
+    }
+}
+
+async function comfyPing() {
+    $('comfyPingOut').textContent = '…';
+    const data = await api('/debug/comfy/ping');
+    $('comfyPingOut').textContent = fmt(data);
+}
+
+async function comfyGenerate() {
+    $('comfyGenOut').textContent = 'Генерация…';
+    $('comfyImgWrap').classList.add('hidden');
+    const body = {
+        positive: $('genPositive').value,
+        negative: $('genNegative').value,
+    };
+    const u = $('genUnet').value;
+    const c = $('genClip').value;
+    const v = $('genVae').value;
+    const ck = $('genCkpt').value;
+    if (u) body.unet = u;
+    if (c) body.clip = c;
+    if (v) body.vae = v;
+    if (ck) body.checkpoint = ck;
+
+    const data = await api('/debug/comfy/generate', {
+        method: 'POST',
+        body: JSON.stringify(body),
+    });
+    $('comfyGenOut').textContent = fmt(data);
+    if (data.image_path) {
+        $('comfyImg').src = data.image_path + '?t=' + Date.now();
+        $('comfyImgWrap').classList.remove('hidden');
+    }
+}
+
+async function comfyRaw() {
+    $('comfyRawOut').textContent = '…';
+    const data = await api('/debug/comfy/raw', {
+        method: 'POST',
+        body: JSON.stringify({
+            method: $('rawMethod').value,
+            path: $('rawPath').value,
+            params_json: $('rawParams').value || '{}',
+            body_json: $('rawBody').value || '',
+        }),
+    });
+    $('comfyRawOut').textContent = fmt(data);
+}
+
+function bind() {
+    initTabs();
+    $('btnReloadConfig').addEventListener('click', loadConfig);
+    $('btnSdPrompt').addEventListener('click', () => runSdPrompt().catch(showErr));
+    $('btnLlm').addEventListener('click', () => runLlm().catch(showErr));
+    $('btnComfyPing').addEventListener('click', () => comfyPing().catch(showErr));
+    $('btnComfyModels').addEventListener('click', () => loadComfyModels().catch(showErr));
+    $('btnComfyGen').addEventListener('click', () => comfyGenerate().catch(showErr));
+    $('btnComfyRaw').addEventListener('click', () => comfyRaw().catch(showErr));
+}
+
+function showErr(e) {
+    alert(e.message || String(e));
+}
+
+bind();
+loadConfig().catch(showErr);
+loadPersonas().catch(showErr);
diff --git a/static/js/newChatWizard.js b/static/js/newChatWizard.js
index 37035f6..778bac3 100644
--- a/static/js/newChatWizard.js
+++ b/static/js/newChatWizard.js
@@ -1,4 +1,9 @@
-import { setSessionId, setCurrentPersona, currentPersona, dom } from './state.js';
+import {
+    setSessionId,
+    setCurrentPersona,
+    getNewChatDefaultPersona,
+    dom,
+} from './state.js';
 import {
     initWizard,
     GENRE_LABELS,
@@ -7,9 +12,9 @@ import {
     fillGreetingSelect,
     getSelectedGreeting,
 } from './utils.js';
-import { personaIndex, highlightPersona } from './personas.js';
+import { personaIndex } from './personas.js';
 
-let newChatPersonaId = currentPersona;
+let newChatPersonaId = getNewChatDefaultPersona();
 let newChatGreetingCtx = null;
 const newChatGenres = new Set();
 const newChatModalEl = document.getElementById('newChatModal');
@@ -84,7 +89,7 @@ function fillNewChatPersonaGrid() {
     const grid = document.getElementById('newChatPersonaGrid');
     if (!grid) return;
     grid.innerHTML = '';
-    newChatPersonaId = currentPersona;
+    newChatPersonaId = getNewChatDefaultPersona();
     for (const p of personaIndex.values()) {
         const card = document.createElement('button');
         card.type = 'button';
@@ -121,34 +126,8 @@ function updateNewChatGenresLabel() {
     }
 }
 
-async function bootstrapRpg(sid, personaId, genreValue, settings) {
-    const { updateQuestPanel, addMessage } = await import('./chat.js');
-    await fetch(`/sessions/${sid}`, {
-        method: 'PATCH',
-        headers: { 'Content-Type': 'application/json' },
-        body: JSON.stringify({
-            rpg_enabled: true,
-            genre: genreValue,
-            rpg_settings_json: JSON.stringify(settings),
-        }),
-    });
-    const res = await fetch('/chat/rpg/bootstrap', {
-        method: 'POST',
-        headers: { 'Content-Type': 'application/json' },
-        body: JSON.stringify({ session_id: sid, persona_id: personaId, genre: genreValue }),
-    });
-    if (res.ok) {
-        const data = await res.json();
-        if (data.quests) updateQuestPanel(data.quests);
-        if (data.plot_arc) {
-            const title = data.plot_arc.title || '';
-            const hint = data.plot_arc.next_beat_hint || '';
-            if (title || hint) addMessage('assistant', `📖 ${title}${hint ? '\n' + hint : ''}`);
-        }
-    }
-}
-
 export function openNewChatWizard() {
+    import('./personas.js').then(({ refreshPersonaBarHighlight }) => refreshPersonaBarHighlight());
     fillNewChatPersonaGrid();
     resetGenreGrid(document.getElementById('newChatGenreGrid'), newChatGenres);
     updateNewChatGenresLabel();
@@ -161,8 +140,17 @@ export function openNewChatWizard() {
 }
 
 export async function createNewChatFromWizard() {
-    const { clearMessages, initChat, reloadChatFromServer } = await import('./chat.js');
-    const { loadSessions, applySessionUi } = await import('./sessions.js');
+    const {
+        clearMessages,
+        initChat,
+        reloadChatFromServer,
+        showImageGenerating,
+        removeImageGenerating,
+        updateQuestPanel,
+        updateAffinityDisplay,
+        renderChoices,
+    } = await import('./chat.js');
+    const { loadSessions, applySessionUi, renderSystemBlob } = await import('./sessions.js');
 
     const sid = 'sess_' + Math.random().toString(36).slice(2, 10);
     setSessionId(sid);
@@ -176,10 +164,23 @@ export async function createNewChatFromWizard() {
     newChatWizard?.reset();
 
     try {
+        const sessionPatch = { persona_id: newChatPersonaId, rpg_enabled: rpg };
+        if (rpg) {
+            sessionPatch.genre = [...newChatGenres].join(',') || 'adventure';
+            sessionPatch.rpg_settings_json = JSON.stringify({
+                dice: document.getElementById('ncSettingDice')?.checked ?? true,
+                narrator: document.getElementById('ncSettingNarrator')?.checked ?? true,
+                quests: document.getElementById('ncSettingQuests')?.checked ?? true,
+                affinity: document.getElementById('ncSettingAffinity')?.checked ?? true,
+                stats: document.getElementById('ncSettingStats')?.checked ?? false,
+                choices: document.getElementById('ncSettingChoices')?.checked ?? true,
+            });
+        }
+
         await fetch(`/sessions/${sid}`, {
             method: 'PATCH',
             headers: { 'Content-Type': 'application/json' },
-            body: JSON.stringify({ persona_id: newChatPersonaId, rpg_enabled: rpg }),
+            body: JSON.stringify(sessionPatch),
         });
 
         if (customTitle) {
@@ -194,25 +195,57 @@ export async function createNewChatFromWizard() {
             dom.headerTitle.textContent = rpg ? `${pName} — RPG` : `${pName} — новый чат`;
         }
 
-        highlightPersona(newChatPersonaId);
+        const { highlightPersonaBar } = await import('./personas.js');
+        highlightPersonaBar(newChatPersonaId);
         const greetingOverride = getNewChatFirstMesOverride();
         await initChat(greetingOverride ? { first_mes_override: greetingOverride } : {});
 
-        if (rpg) {
-            const genreValue = [...newChatGenres].join(',') || 'adventure';
-            const settings = {
-                dice: document.getElementById('ncSettingDice')?.checked ?? true,
-                narrator: document.getElementById('ncSettingNarrator')?.checked ?? true,
-                quests: document.getElementById('ncSettingQuests')?.checked ?? true,
-                affinity: document.getElementById('ncSettingAffinity')?.checked ?? true,
-                choices: document.getElementById('ncSettingChoices')?.checked ?? true,
-            };
-            await bootstrapRpg(sid, newChatPersonaId, genreValue, settings);
+        const assistantWrapper = dom.messagesEl.querySelector('.message.assistant');
+        showImageGenerating(assistantWrapper);
+
+        let openingData = null;
+        try {
+            const openingRes = await fetch('/chat/opening/process', {
+                method: 'POST',
+                headers: { 'Content-Type': 'application/json' },
+                body: JSON.stringify({
+                    session_id: sid,
+                    persona_id: newChatPersonaId,
+                    rpg,
+                }),
+            });
+            openingData = await openingRes.json();
+            if (!openingRes.ok) {
+                console.error('opening/process failed:', openingData.detail || openingRes.statusText);
+            }
+        } finally {
+            removeImageGenerating(assistantWrapper);
         }
 
         await reloadChatFromServer(sid);
+
+        if (openingData?.quests?.length) {
+            updateQuestPanel(openingData.quests);
+        }
+        if (openingData?.affinity !== undefined) {
+            updateAffinityDisplay(openingData.affinity);
+        }
+        if (openingData?.image_error) {
+            const wrapper = dom.messagesEl.querySelector('.message.assistant');
+            if (wrapper) {
+                const err = document.createElement('div');
+                err.className = 'image-error';
+                err.textContent = '🖼 ' + openingData.image_error;
+                wrapper.appendChild(err);
+            }
+        }
+
         const sessionRes = await fetch(`/sessions/${sid}`);
         if (sessionRes.ok) applySessionUi(await sessionRes.json());
+
+        const blobRes = await fetch(`/chat/system/${sid}`);
+        if (blobRes.ok) renderSystemBlob(await blobRes.json());
+
         await loadSessions();
     } catch (e) {
         console.error('createNewChat error:', e);
diff --git a/static/js/personas.js b/static/js/personas.js
index 3b5628b..8546132 100644
--- a/static/js/personas.js
+++ b/static/js/personas.js
@@ -1,5 +1,9 @@
-import { currentPersona, setCurrentPersona, sessionId } from './state.js';
-import { initChat } from './chat.js';
+import {
+    currentPersona,
+    sessionId,
+    getNewChatDefaultPersona,
+    setNewChatDefaultPersona,
+} from './state.js';
 import { initWizard, fillGreetingSelect, getSelectedGreeting } from './utils.js';
 
 export let personaIndex = new Map();
@@ -21,12 +25,18 @@ let cardImportWizard;
 let cardPreview = null;
 let cardImportFile = null;
 
-export function highlightPersona(personaId) {
+export function highlightPersonaBar(personaId) {
     document.querySelectorAll('.persona-card').forEach(c => {
         c.classList.toggle('active', c.dataset.id === personaId);
     });
 }
 
+/** Active session → session persona; otherwise new-chat preset. */
+export function refreshPersonaBarHighlight() {
+    const id = sessionId ? currentPersona : getNewChatDefaultPersona();
+    highlightPersonaBar(id);
+}
+
 export async function loadPersonas() {
     const res = await fetch('/personas/');
     const personas = await res.json();
@@ -37,9 +47,11 @@ export async function loadPersonas() {
     const bar = document.getElementById('personaBar');
     bar.innerHTML = '';
 
+    const barActiveId = sessionId ? currentPersona : getNewChatDefaultPersona();
+
     personas.forEach(p => {
         const card = document.createElement('div');
-        card.className = 'persona-card' + (p.persona_id === currentPersona ? ' active' : '');
+        card.className = 'persona-card' + (p.persona_id === barActiveId ? ' active' : '');
         card.dataset.id = p.persona_id;
         const isCard = p.persona_id.startsWith('card_');
         const isCustomPersona = p.custom && !isCard;
@@ -131,16 +143,8 @@ export async function loadPersonas() {
 }
 
 export async function selectPersona(personaId) {
-    setCurrentPersona(personaId);
-    highlightPersona(personaId);
-    if (sessionId) {
-        await fetch(`/sessions/${sessionId}`, {
-            method: 'PATCH',
-            headers: { 'Content-Type': 'application/json' },
-            body: JSON.stringify({ persona_id: personaId }),
-        });
-        await initChat();
-    }
+    setNewChatDefaultPersona(personaId);
+    highlightPersonaBar(personaId);
 }
 
 function fillImpCardForm(preview) {
diff --git a/static/js/rpFormat.js b/static/js/rpFormat.js
new file mode 100644
index 0000000..4ab63b4
--- /dev/null
+++ b/static/js/rpFormat.js
@@ -0,0 +1,209 @@
+/** Escape for safe text nodes (not HTML injection). */
+function appendText(parent, text) {
+    if (!text) return;
+    parent.appendChild(document.createTextNode(text));
+}
+
+const PLAYER_CHOSE_RE = /^\[Player chose:\s*(.+)\]$/s;
+const RP_TOKEN_RE = /("(?:[^"\\]|\\.)*"|«[^»]*»|'(?:[^'\\]|\\.)*'|\*\*[^*]+\*\*|\*[^*]+\*)/g;
+
+const OUTCOME_CLASS = {
+    'critical failure': 'outcome-crit-fail',
+    failure: 'outcome-fail',
+    success: 'outcome-success',
+    'critical success': 'outcome-crit-success',
+};
+
+export function isPlayerChoiceContent(text) {
+    return PLAYER_CHOSE_RE.test((text || '').trim());
+}
+
+export function parsePlayerChoice(text) {
+    const m = (text || '').trim().match(PLAYER_CHOSE_RE);
+    return m ? m[1].trim() : null;
+}
+
+function appendPlayerChoice(frag, label) {
+    const wrap = document.createElement('span');
+    wrap.className = 'rp-choice';
+    const tag = document.createElement('span');
+    tag.className = 'rp-choice-tag';
+    tag.textContent = '🔘 Выбор';
+    wrap.appendChild(tag);
+    appendText(wrap, ' ');
+    const lab = document.createElement('span');
+    lab.className = 'rp-choice-label';
+    lab.textContent = label;
+    wrap.appendChild(lab);
+    frag.appendChild(wrap);
+}
+
+/**
+ * Build DOM fragment for RP formatting (dialogue, action, player choice).
+ */
+export function buildRpFormatFragment(text) {
+    const frag = document.createDocumentFragment();
+    if (!text) return frag;
+
+    const trimmed = text.trim();
+    const choiceLabel = parsePlayerChoice(trimmed);
+    if (choiceLabel !== null) {
+        appendPlayerChoice(frag, choiceLabel);
+        return frag;
+    }
+
+    let last = 0;
+    for (const m of trimmed.matchAll(RP_TOKEN_RE)) {
+        const idx = m.index ?? 0;
+        if (idx > last) appendText(frag, trimmed.slice(last, idx));
+
+        const token = m[0];
+        if (
+            (token.startsWith('"') && token.endsWith('"'))
+            || (token.startsWith('«') && token.endsWith('»'))
+            || (token.startsWith("'") && token.endsWith("'"))
+        ) {
+            const span = document.createElement('span');
+            span.className = 'rp-dialogue';
+            span.textContent = token;
+            frag.appendChild(span);
+        } else if (token.startsWith('**') && token.endsWith('**')) {
+            const em = document.createElement('em');
+            em.className = 'rp-action';
+            em.textContent = token.slice(2, -2);
+            frag.appendChild(em);
+        } else if (token.startsWith('*') && token.endsWith('*')) {
+            const em = document.createElement('em');
+            em.className = 'rp-action';
+            em.textContent = token.slice(1, -1);
+            frag.appendChild(em);
+        } else {
+            appendText(frag, token);
+        }
+        last = idx + token.length;
+    }
+    if (last < trimmed.length) appendText(frag, trimmed.slice(last));
+    return frag;
+}
+
+/**
+ * Dice override block on user bubble (intent struck through → resolution).
+ */
+export function buildDiceOverrideFragment(resolution) {
+    const { intent_text, resolution_text, roll, outcome } = resolution;
+    const wrap = document.createElement('div');
+    wrap.className = 'dice-user-override';
+
+    const badge = document.createElement('div');
+    badge.className = `dice-user-badge ${OUTCOME_CLASS[outcome] || ''}`;
+    badge.textContent = `🎲 ${roll} · ${outcome || 'roll'}`;
+    wrap.appendChild(badge);
+
+    const row = document.createElement('div');
+    row.className = 'dice-user-row';
+
+    const struck = document.createElement('span');
+    struck.className = 'dice-intent-struck';
+    struck.textContent = intent_text || '';
+    row.appendChild(struck);
+
+    const arrow = document.createElement('span');
+    arrow.className = 'dice-intent-arrow';
+    arrow.textContent = '→';
+    row.appendChild(arrow);
+
+    const resolved = document.createElement('span');
+    resolved.className = 'dice-intent-resolved';
+    resolved.textContent = resolution_text || '';
+    row.appendChild(resolved);
+
+    wrap.appendChild(row);
+    return wrap;
+}
+
+export function applyDiceOverrideToBubble(bubble, resolution) {
+    if (!resolution?.resolution_text) return;
+    bubble.dataset.diceOverride = '1';
+    const raw = bubble.dataset.raw ?? '';
+    const frag = document.createDocumentFragment();
+    if (raw && !isPlayerChoiceContent(raw)) {
+        frag.appendChild(buildRpFormatFragment(raw));
+        frag.appendChild(document.createElement('br'));
+    } else if (raw) {
+        frag.appendChild(buildRpFormatFragment(raw));
+        frag.appendChild(document.createElement('br'));
+    }
+    frag.appendChild(buildDiceOverrideFragment(resolution));
+    bubble.innerHTML = '';
+    bubble.appendChild(frag);
+    bubble.classList.add('rp-formatted', 'has-dice-override');
+}
+
+export function applyRpFormatToBubble(bubble) {
+    if (bubble.dataset.diceOverride === '1' && bubble._diceResolution) {
+        applyDiceOverrideToBubble(bubble, bubble._diceResolution);
+        return;
+    }
+    const raw = bubble.dataset.raw ?? bubble.textContent ?? '';
+    bubble.dataset.raw = raw;
+    bubble.innerHTML = '';
+    bubble.appendChild(buildRpFormatFragment(raw));
+    bubble.classList.add('rp-formatted');
+}
+
+export function applyRpPlainToBubble(bubble) {
+    const raw = bubble.dataset.raw ?? bubble.textContent ?? '';
+    bubble.dataset.raw = raw;
+    bubble.textContent = raw;
+    bubble.classList.remove('rp-formatted');
+}
+
+export function attachFormatToggle(wrapper, bubble) {
+    const btn = document.createElement('button');
+    btn.type = 'button';
+    btn.className = 'format-btn';
+    btn.textContent = '↩ Текст';
+    let showingPlain = false;
+
+    const syncBtn = () => {
+        btn.textContent = showingPlain ? '✨' : '↩ Текст';
+        btn.title = showingPlain ? 'Форматировать' : 'Показать оригинал';
+    };
+
+    btn.addEventListener('click', (e) => {
+        e.preventDefault();
+        e.stopPropagation();
+        if (showingPlain) {
+            if (bubble._diceResolution) {
+                applyDiceOverrideToBubble(bubble, bubble._diceResolution);
+            } else {
+                applyRpFormatToBubble(bubble);
+            }
+            showingPlain = false;
+        } else {
+            applyRpPlainToBubble(bubble);
+            showingPlain = true;
+        }
+        syncBtn();
+    });
+
+    syncBtn();
+    wrapper.appendChild(btn);
+    return btn;
+}
+
+export function initBubbleContent(bubble, rawText, { formatted = true, actionResolution = null } = {}) {
+    bubble.dataset.raw = rawText ?? '';
+    bubble._diceResolution = actionResolution || null;
+    if (actionResolution?.resolution_text && formatted) {
+        bubble.dataset.diceOverride = '1';
+        applyDiceOverrideToBubble(bubble, actionResolution);
+        return;
+    }
+    if (formatted && rawText) {
+        applyRpFormatToBubble(bubble);
+    } else {
+        bubble.textContent = rawText ?? '';
+        bubble.classList.remove('rp-formatted', 'has-dice-override');
+    }
+}
diff --git a/static/js/sessions.js b/static/js/sessions.js
index 1cf88af..ff200e9 100644
--- a/static/js/sessions.js
+++ b/static/js/sessions.js
@@ -1,8 +1,10 @@
 import {
     sessionId, setSessionId, setCurrentPersona, currentPersona, dom, setRpgEnabled,
 } from './state.js';
-import { updateQuestPanel, updateAffinityDisplay } from './chat.js';
-import { highlightPersona, personaIndex } from './personas.js';
+import {
+    updateQuestPanel, updateAffinityDisplay, updateStatsDisplay, hideStatsDisplay,
+} from './chat.js';
+import { highlightPersonaBar, personaIndex } from './personas.js';
 import { formatSessionDate } from './utils.js';
 import { openNewChatWizard } from './newChatWizard.js';
 
@@ -35,6 +37,16 @@ export function applySessionUi(session) {
         dom.affinityDisplay?.classList.add('hidden');
     }
 
+    if (rpgOn && settings.stats) {
+        try {
+            updateStatsDisplay(JSON.parse(session.narrative_stats_json || '{}'));
+        } catch {
+            hideStatsDisplay();
+        }
+    } else {
+        hideStatsDisplay();
+    }
+
     if (rpgOn && settings.quests) {
         fetch(`/sessions/${session.session_id}/quests`)
             .then(r => r.ok ? r.json() : [])
@@ -114,7 +126,7 @@ export async function loadChatHistory(id) {
         const s = await sessionRes.json();
         if (s.persona_id) {
             setCurrentPersona(s.persona_id);
-            highlightPersona(s.persona_id);
+            highlightPersonaBar(s.persona_id);
         }
         applySessionUi(s);
     }
@@ -155,7 +167,7 @@ export async function initSessions() {
 
 let _prevBlobSections = {};
 
-function renderSystemBlob(blob) {
+export function renderSystemBlob(blob) {
     const tryFmt = (str, fallback = '') => {
         try { return JSON.stringify(JSON.parse(str), null, 2); } catch { return str || fallback; }
     };
@@ -165,13 +177,26 @@ function renderSystemBlob(blob) {
         return `  ${icon} [${q.status}] ${q.title}`;
     }).join('\n');
 
+    const personaLine = blob.persona_id
+        ? `[persona] ${blob.persona_name || blob.persona_id} (${blob.persona_id})`
+        : '';
+
+    const ctx = blob.context_usage;
+    const ctxLine = ctx
+        ? `[context] ~${ctx.tokens_est} / ${ctx.max_tokens_est} tokens (${ctx.percent}%) · ${ctx.chars} chars`
+        : '';
+
     const sections = {
+        context:         ctxLine,
+        persona:         personaLine,
         system_prompt:   blob.system_prompt ? `[system_prompt]\n${blob.system_prompt}` : '',
         status_quo:      blob.status_quo    ? `[status_quo]\n${blob.status_quo}` : '',
         affinity:        blob.affinity != null ? `[affinity] ${blob.affinity}` : '',
+        scene:           blob.scene_json && blob.scene_json !== '{}' ? `[scene]\n${tryFmt(blob.scene_json)}` : '',
+        stats:           blob.narrative_stats_json && blob.narrative_stats_json !== '{}' ? `[stats]\n${tryFmt(blob.narrative_stats_json)}` : '',
         genre:           blob.genre         ? `[genre] ${blob.genre}` : '',
         rpg_settings:    blob.rpg_settings_json && blob.rpg_settings_json !== '{}' ? `[rpg_settings]\n${tryFmt(blob.rpg_settings_json)}` : '',
-        outfit:          blob.outfit_json && blob.outfit_json !== '[]' ? `[outfit]\n${tryFmt(blob.outfit_json)}` : '',
+        outfit:          `[outfit]\n${tryFmt(blob.outfit_json ?? '[]')}`,
         facts:           blob.facts_json && blob.facts_json !== '[]' ? `[facts]\n${tryFmt(blob.facts_json)}` : '',
         plot_arc:        blob.plot_arc_json && blob.plot_arc_json !== '{}' ? `[plot_arc]\n${tryFmt(blob.plot_arc_json)}` : '',
         quests:          questLines ? `[quests]\n${questLines}` : '',
@@ -184,7 +209,9 @@ function renderSystemBlob(blob) {
         if (!text) continue;
         const span = document.createElement('span');
         span.textContent = text;
-        if (_prevBlobSections[key] && _prevBlobSections[key] !== text) {
+        if (key === 'context' && ctx && ctx.percent > 80) {
+            span.className = 'blob-context-warn';
+        } else if (_prevBlobSections[key] && _prevBlobSections[key] !== text) {
             span.className = 'blob-changed';
             setTimeout(() => span.classList.remove('blob-changed'), 3000);
         }
diff --git a/static/js/state.js b/static/js/state.js
index 6818f22..e94ad55 100644
--- a/static/js/state.js
+++ b/static/js/state.js
@@ -1,17 +1,31 @@
 export let sessionId = localStorage.getItem('chat_session_id') || null;
-export let currentPersona = localStorage.getItem('persona_id') || 'default';
+/** Persona bound to the active session (from server, not global preset). */
+export let currentPersona = 'default';
 export let sidebarOpen = true;
 export let rpgEnabled = false;
+
+const NEW_CHAT_PERSONA_KEY = 'new_chat_persona_id';
+
 export function toggleSidebar() { sidebarOpen = !sidebarOpen; return sidebarOpen; }
 
+export function getNewChatDefaultPersona() {
+    return localStorage.getItem(NEW_CHAT_PERSONA_KEY)
+        || localStorage.getItem('persona_id')
+        || 'default';
+}
+
+export function setNewChatDefaultPersona(id) {
+    const pid = id || 'default';
+    localStorage.setItem(NEW_CHAT_PERSONA_KEY, pid);
+}
+
 export function setSessionId(id) {
     sessionId = id;
     if (id) localStorage.setItem('chat_session_id', id);
 }
 
 export function setCurrentPersona(id) {
-    currentPersona = id;
-    localStorage.setItem('persona_id', id);
+    currentPersona = id || 'default';
 }
 
 export function setRpgEnabled(v) { rpgEnabled = !!v; }
@@ -25,6 +39,7 @@ export const dom = {
     headerTitle: document.getElementById('headerTitle'),
     emptyState: document.getElementById('emptyState'),
     affinityDisplay: document.getElementById('affinityDisplay'),
+    statsDisplay: document.getElementById('statsDisplay'),
     rpgBadge: document.getElementById('rpgBadge'),
     systemBlob: document.getElementById('systemBlob'),
     systemBlobContent: document.getElementById('systemBlobContent'),
diff --git a/static/js/utils.js b/static/js/utils.js
index c169608..c4fcd79 100644
--- a/static/js/utils.js
+++ b/static/js/utils.js
@@ -6,12 +6,32 @@ export function parseImagePromptFromContent(content) {
     return { text, prompt };
 }
 
+export function splitSdPromptForCopy(fullPrompt) {
+    if (!fullPrompt) return '';
+    const marker = '\n\nNegative prompt:';
+    const i = fullPrompt.indexOf(marker);
+    return (i >= 0 ? fullPrompt.slice(0, i) : fullPrompt).trim();
+}
+
 export async function copyToClipboard(text) {
+    if (!text) return false;
     try {
         await navigator.clipboard.writeText(text);
         return true;
     } catch {
-        return false;
+        try {
+            const ta = document.createElement('textarea');
+            ta.value = text;
+            ta.setAttribute('readonly', '');
+            ta.style.cssText = 'position:fixed;left:-9999px;top:0';
+            document.body.appendChild(ta);
+            ta.select();
+            const ok = document.execCommand('copy');
+            document.body.removeChild(ta);
+            return ok;
+        } catch {
+            return false;
+        }
     }
 }
 
diff --git a/tests/test_arc_stuck_recovery.py b/tests/test_arc_stuck_recovery.py
new file mode 100644
index 0000000..ed94eca
--- /dev/null
+++ b/tests/test_arc_stuck_recovery.py
@@ -0,0 +1,87 @@
+import asyncio
+
+from services.rpg_plot import (
+    prune_beats_for_done_quests,
+    process_arc_beats,
+    should_advance_arc_keywords,
+    pop_matching_beats,
+    dice_outcome_to_beat_trigger,
+)
+
+
+def test_prune_removes_beat_when_quest_done():
+    arc = {
+        "beats": [
+            {"id": "b2", "title": "Highway Howl", "trigger": "event_driven:travel", "choices": []},
+        ]
+    }
+    quests = [{"title": "Highway Howl", "status": "done"}]
+    arc, removed = prune_beats_for_done_quests(arc, quests)
+    assert removed[0]["title"] == "Highway Howl"
+    assert arc["beats"] == []
+
+
+def test_stuck_recovery_fires_when_no_active_quests():
+    arc = {
+        "beats": [
+            {"id": "b3", "title": "New Beat", "trigger": "event_driven:rest", "choices": [{"id": "a", "label": "A"}]},
+        ]
+    }
+    quests = [{"title": "Old", "status": "done"}]
+
+    async def run():
+        return await process_arc_beats(arc, quests, "hello")
+
+    arc2, fired, pruned, mode = asyncio.run(run())
+    assert mode == "stuck_recovery"
+    assert fired[0]["title"] == "New Beat"
+    assert arc2["beats"] == []
+
+
+def test_dice_outcome_maps_to_after_fail():
+    assert dice_outcome_to_beat_trigger("failure") == "event_driven:after_fail"
+    assert dice_outcome_to_beat_trigger("critical failure") == "event_driven:after_fail"
+    assert dice_outcome_to_beat_trigger("success") == "event_driven:after_success"
+
+
+def test_after_fail_beat_fires_on_dice_failure():
+    arc = {
+        "beats": [
+            {
+                "id": "b_fail",
+                "title": "Dust Yourself Off",
+                "trigger": "event_driven:after_fail",
+                "injection": "The stumble leaves you both shaken.",
+                "choices": [{"id": "a", "label": "Try again"}],
+            },
+            {
+                "id": "b_ok",
+                "title": "Victory Lap",
+                "trigger": "event_driven:after_success",
+                "choices": [],
+            },
+        ]
+    }
+
+    async def run():
+        return await process_arc_beats(
+            arc, [], "продолжаем разговор", last_dice_outcome="failure"
+        )
+
+    arc2, fired, _, mode = asyncio.run(run())
+    assert mode == "after_dice"
+    assert fired[0]["id"] == "b_fail"
+    assert len(arc2["beats"]) == 1
+    assert arc2["beats"][0]["id"] == "b_ok"
+
+
+def test_keyword_fallback_travel():
+    arc = {
+        "beats": [
+            {"id": "b2", "title": "Highway Howl", "trigger": "event_driven:travel", "choices": []},
+        ]
+    }
+    trig = should_advance_arc_keywords("едем на стадион")
+    assert trig == "event_driven:travel"
+    arc, fired = pop_matching_beats(arc, trig, max_beats=1)
+    assert fired[0]["title"] == "Highway Howl"
diff --git a/tests/test_dice_llm_payload.py b/tests/test_dice_llm_payload.py
new file mode 100644
index 0000000..6060352
--- /dev/null
+++ b/tests/test_dice_llm_payload.py
@@ -0,0 +1,54 @@
+from services.memory import narrator_message_content, parse_narrator_message
+from routers.chat import messages_for_llm
+from services.rpg_state import format_narrator_outcome_for_llm
+
+
+def test_messages_for_llm_includes_narrator_ruling_not_original_only():
+    resolution = (
+        "Luna's attempt backfires; people roll their eyes and an elderly man pats her head."
+    )
+    narrator_json = narrator_message_content(
+        {
+            "roll": 5,
+            "outcome": "failure",
+            "text": resolution,
+            "original_intent": "bulldoze through the line",
+        }
+    )
+    history = [
+        {"role": "system", "content": "static"},
+        {
+            "role": "user",
+            "content": "Луна влезла без очереди, распугивая всех",
+            "action_resolution": {
+                "intent_text": "Луна влезла без очереди",
+                "roll": 5,
+                "outcome": "failure",
+                "resolution_text": resolution,
+            },
+        },
+        {"role": "narrator", "content": narrator_json},
+    ]
+    llm = messages_for_llm(history, "system+runtime")
+    user_msgs = [m for m in llm if m["role"] == "user"]
+    assert len(user_msgs) == 2
+    assert "canonical outcome" in user_msgs[0]["content"].lower()
+    assert "Луна влезла" in user_msgs[0]["content"]
+    assert "MANDATORY" in user_msgs[1]["content"]
+    assert "backfires" in user_msgs[1]["content"]
+    assert "Do NOT write a success version" in user_msgs[1]["content"]
+
+
+def test_format_narrator_failure_wording():
+    text = format_narrator_outcome_for_llm(
+        {"roll": 5, "outcome": "failure", "text": "It failed."}
+    )
+    assert "FAILED" in text
+    assert "Do NOT write a success version" in text
+
+
+def test_parse_narrator_message_roundtrip():
+    raw = narrator_message_content({"roll": 12, "outcome": "success", "text": "OK"})
+    data = parse_narrator_message(raw)
+    assert data["roll"] == 12
+    assert data["text"] == "OK"
diff --git a/tests/test_outfit_tags.py b/tests/test_outfit_tags.py
new file mode 100644
index 0000000..450da17
--- /dev/null
+++ b/tests/test_outfit_tags.py
@@ -0,0 +1,18 @@
+from services.outfit_tags import enrich_outfit_tag, normalize_outfit_list, parse_and_normalize_outfit_json
+
+
+def test_enrich_shorts_and_tank():
+    out = normalize_outfit_list(["sports_shorts", "torn_tank_top", "championship_belt_collar"])
+    assert out == ["black_sports_shorts", "white_torn_tank_top", "gold_championship_belt_collar"]
+
+
+def test_keeps_existing_color():
+    assert enrich_outfit_tag("red_dress") == "red_dress"
+    assert "red" in normalize_outfit_list(["red_dress"])[0]
+
+
+def test_parse_json_string():
+    raw = '["sports_shorts", "torn_tank_top"]'
+    parsed = parse_and_normalize_outfit_json(raw)
+    assert "black_sports_shorts" in parsed
+    assert "white_torn_tank_top" in parsed
diff --git a/tests/test_plot_choices.py b/tests/test_plot_choices.py
new file mode 100644
index 0000000..045150c
--- /dev/null
+++ b/tests/test_plot_choices.py
@@ -0,0 +1,29 @@
+from services.rpg_plot import choices_from_beat, choices_from_narrator, normalize_choice
+
+
+def test_choices_from_beat_tags_source():
+    beat = {
+        "id": "b_new_1",
+        "title": "Rest Stop Confession",
+        "injection": "GPS alerts you to a rest stop.",
+        "choices": [
+            {"id": "a", "label": "Tease her about snores"},
+            {"id": "b", "label": "Ask about snacks"},
+        ],
+    }
+    out = choices_from_beat(beat)
+    assert len(out) == 2
+    assert out[0]["source"] == "plot_beat"
+    assert out[0]["beat_title"] == "Rest Stop Confession"
+    assert out[0]["beat_id"] == "b_new_1"
+    assert "beat_injection" in out[0]
+
+
+def test_choices_from_narrator_tags_source():
+    out = choices_from_narrator([{"id": "a", "label": "Look around"}])
+    assert out[0]["source"] == "narrator"
+    assert "beat_title" not in out[0]
+
+
+def test_normalize_choice_skips_empty_label():
+    assert normalize_choice({"id": "a", "label": "  "}, source="narrator") is None
diff --git a/tests/test_rp_sanitize.py b/tests/test_rp_sanitize.py
new file mode 100644
index 0000000..6e13a1f
--- /dev/null
+++ b/tests/test_rp_sanitize.py
@@ -0,0 +1,18 @@
+from services.rp_sanitize import strip_ooc_from_reply
+
+
+def test_strip_ps_block():
+    text = (
+        "Луна зевает и прижимается к тебе.\n\n"
+        "Статус кво? Она никогда не признает слабость.\n\n"
+        "P.S. Когда вы выйдете из сауны, она будет бурчать."
+    )
+    out = strip_ooc_from_reply(text)
+    assert "P.S." not in out
+    assert "Статус кво" not in out
+    assert "Луна зевает" in out
+
+
+def test_strip_keeps_in_character_body():
+    text = "— «Щенок…» — она бурчит, не открывая глаз."
+    assert strip_ooc_from_reply(text) == text
diff --git a/tests/test_rpg_facts.py b/tests/test_rpg_facts.py
new file mode 100644
index 0000000..7ba5e4e
--- /dev/null
+++ b/tests/test_rpg_facts.py
@@ -0,0 +1,42 @@
+from services.rpg_facts import (
+    merge_facts,
+    parse_facts_list,
+    facts_to_prompt,
+    facts_list_to_json,
+    dedupe_facts_fuzzy,
+    facts_are_similar,
+)
+
+
+def test_legacy_string_facts():
+    raw = '["Old fact", "Another"]'
+    facts = parse_facts_list(raw)
+    assert len(facts) == 2
+    assert facts[0]["text"] == "Old fact"
+
+
+def test_fuzzy_similar_near_duplicate():
+    assert facts_are_similar(
+        "Rin and Grigo found a magical glade with glowing flowers",
+        "Rin and Grigo found magical glade glowing flowers",
+    )
+
+
+def test_dedupe_collapses_duplicates():
+    raw = facts_list_to_json(
+        [
+            {"text": "Rin found a magical glade with glowing flowers", "rp_day": "day 1"},
+            {"text": "Rin found magical glade glowing flowers", "rp_day": "day 1"},
+            {"text": "Player name is Grigo", "rp_day": "day 1"},
+        ]
+    )
+    out = dedupe_facts_fuzzy(parse_facts_list(raw))
+    assert len(out) == 2
+
+
+def test_merge_with_rp_day():
+    existing = facts_list_to_json([{"text": "A", "rp_day": "день 1"}])
+    merged = parse_facts_list(
+        merge_facts(existing, [{"text": "B", "rp_day": ""}], rp_day_default="день 2")
+    )
+    assert merged[1]["rp_day"] == "день 2"
diff --git a/tests/test_rpg_state_prompts.py b/tests/test_rpg_state_prompts.py
new file mode 100644
index 0000000..72118cc
--- /dev/null
+++ b/tests/test_rpg_state_prompts.py
@@ -0,0 +1,17 @@
+from services.rpg_state import affinity_prompt_block, stats_prompt_block
+
+
+def test_affinity_mandatory_wording():
+    block = affinity_prompt_block(5)
+    assert "MANDATORY" in block
+    assert "current player" in block
+
+
+def test_stamina_2_exhausted_instruction():
+    block = stats_prompt_block({"lust": 0, "stamina": 2, "tension": 0})
+    assert "barely moves" in block.lower() or "exhausted" in block.lower()
+
+
+def test_stamina_1_collapse_instruction():
+    block = stats_prompt_block({"lust": 0, "stamina": 1, "tension": 0})
+    assert "collapse" in block.lower() or "pass out" in block.lower()
diff --git a/tests/test_sd_prompt.py b/tests/test_sd_prompt.py
new file mode 100644
index 0000000..6a28f20
--- /dev/null
+++ b/tests/test_sd_prompt.py
@@ -0,0 +1,243 @@
+"""Unit tests for layered Anima prompt assembly (no LLM)."""
+
+import os
+import sys
+from unittest.mock import patch
+
+import pytest
+
+# Ensure project root on path
+sys.path.insert(0, os.path.abspath(os.path.join(os.path.dirname(__file__), "..")))
+
+from services import sd_prompt as sp
+
+
+@pytest.fixture
+def anima():
+    with patch.object(sp, "_is_anima", return_value=True), patch.object(sp, "_is_pony", return_value=False):
+        yield
+
+
+PERSONA_WOLF = {
+    "appearance_tags": "wolfgirl, white_hair, golden_eyes, wolf_ears, tail, big_breast",
+    "appearance_prose": "",
+    "lora_name": "",
+}
+
+PERSONA_CARRIE = {
+    "appearance_tags": "short_hair, brown_hair, blue_eyes, skinny",
+    "appearance_prose": "",
+    "lora_name": "",
+}
+
+
+def test_walking_scene_includes_action_tags_and_contextual_pov(anima):
+    scene = sp._sanitize_scene_fields({
+        "shot_type": "first_person_pov",
+        "pov_cue": "walking_together",
+        "viewer_body_visible": False,
+        "action_tags": "holding_hands, walking, smiling, looking_at_each_other",
+        "environment_tags": "outdoors, sunlight, golden_hour",
+        "scene_description": "She walks beside you, laughter in the warm afternoon light.",
+    })
+    hybrid = sp.build_positive_prompt_hybrid(scene, PERSONA_WOLF, "")
+    assert "walking" in hybrid
+    assert "smiling" in hybrid
+    assert "holding_hands" not in hybrid
+    assert "looking_at_each_other" not in hybrid
+    assert "outdoors" in hybrid
+    assert "threshold" not in hybrid.lower()
+    assert "POV: walking beside you" in hybrid
+    assert "someone" not in hybrid.lower()
+    assert "both " not in hybrid.lower()
+
+
+def test_hybrid_differs_from_tags_only_when_prose_present(anima):
+    scene = {
+        "shot_type": "first_person_pov",
+        "pov_cue": "walking_together",
+        "viewer_body_visible": False,
+        "action_tags": "holding_hands, walking",
+        "environment_tags": "outdoors, sunlight",
+        "scene_description": "Shared laughter drifts through the golden afternoon.",
+    }
+    tags_only = sp.build_positive_prompt_tags_only(scene, PERSONA_WOLF, "")
+    hybrid = sp.build_positive_prompt_hybrid(scene, PERSONA_WOLF, "")
+    assert tags_only != hybrid
+    assert "Shared laughter" in hybrid
+    assert "Shared laughter" not in tags_only
+
+
+def test_carrie_doorway_scene(anima):
+    scene = {
+        "shot_type": "first_person_pov",
+        "pov_cue": "doorway_invite",
+        "viewer_body_visible": False,
+        "action_tags": "arms_out, inviting_hug, smirk, looking_at_viewer",
+        "environment_tags": "doorway, apartment, night, indoors",
+        "scene_description": "She waits in the doorway with playful hunger in half-lidded eyes.",
+    }
+    outfit = "crop_top, ripped_jeans, black_jeans"
+    hybrid = sp.build_positive_prompt_hybrid(scene, PERSONA_CARRIE, outfit)
+    assert "arms_out" in hybrid
+    assert "doorway" in hybrid
+    assert "crop_top" in hybrid
+    assert "threshold" not in hybrid.lower()
+    assert "POV: she blocks the doorway" in hybrid
+
+
+def test_pov_inferred_from_action_when_cue_missing(anima):
+    scene = {
+        "shot_type": "first_person_pov",
+        "action_tags": "holding_hands, walking, smiling",
+        "environment_tags": "outdoors, park",
+        "scene_description": "",
+    }
+    tags = sp.build_positive_prompt_tags_only(scene, PERSONA_WOLF, "")
+    assert "POV: walking beside you" in tags
+
+
+def test_negative_includes_interaction_block_for_pov_contact(anima):
+    scene = {
+        "shot_type": "first_person_pov",
+        "viewer_body_visible": False,
+        "action_tags": "arms_out, hug, inviting_hug",
+        "environment_tags": "doorway",
+    }
+    neg = sp._negative_for_scene(scene)
+    assert "duplicate" in neg
+    assert "extra_person" in neg
+    assert "third person" in neg
+
+
+def test_scene_should_generate_false():
+    assert sp._scene_should_generate({"should_generate": False}) is False
+    assert sp._scene_should_generate({"should_generate": True}) is True
+    assert sp._scene_should_generate({}) is True
+
+
+def test_format_builder_user_block_illustrate_vs_context(anima):
+    messages = [
+        {"role": "assistant", "content": "Long old first_mes " + ("x" * 900)},
+        {"role": "user", "content": "Hi"},
+        {"role": "assistant", "content": "*walks holding your hand*"},
+    ]
+    block = sp._format_builder_user_block(PERSONA_WOLF, messages, "[]")
+    assert "=== ILLUSTRATE" in block
+    assert "=== Context" in block
+    assert "*walks holding your hand*" in block
+    assert "Long old first_mes" in block
+    assert len(block.split("Long old first_mes")[1].split("assistant:")[0]) < 900
+
+
+def test_bundle_from_scene_anima_uses_hybrid_as_tag_full(anima):
+    scene = {
+        "should_generate": True,
+        "shot_type": "first_person_pov",
+        "pov_cue": "face_to_face",
+        "action_tags": "smiling",
+        "environment_tags": "indoors",
+        "scene_description": "A warm smile greets you.",
+    }
+    with patch.object(sp, "anima_dual_enabled", return_value=False):
+        bundle = sp._bundle_from_scene(scene, PERSONA_WOLF, "")
+    assert "A warm smile" in bundle.tag_full
+    assert bundle.desc_full is None
+
+
+def test_user_example_walking_llm_output_cleaned(anima):
+    """Regression: LLM prose/sentence leakage and second-person refs."""
+    scene = sp._sanitize_scene_fields({
+        "shot_type": "first_person_pov",
+        "pov_cue": "walking_together",
+        "action_tags": (
+            "holding_hands, walking, smiling, looking_at_each_other, "
+            "A wolfgirl walks hand in hand with someone, both smiling and chatting"
+        ),
+        "environment_tags": "outdoor, daylight, path",
+        "scene_description": (
+            "A wolfgirl walks hand in hand with someone, both smiling and chatting under the daylight."
+        ),
+    })
+    persona = {**PERSONA_WOLF, "appearance_tags": PERSONA_WOLF["appearance_tags"] + ", pumped_up"}
+    tags_only = sp.build_positive_prompt_tags_only(scene, persona, "")
+    hybrid = sp.build_positive_prompt_hybrid(scene, persona, "")
+    assert "pumped_up" not in tags_only
+    assert "someone" not in hybrid.lower()
+    assert "both " not in hybrid.lower()
+    assert ". A wolfgirl walks" not in tags_only
+    assert tags_only != hybrid or not scene.get("scene_description")
+
+
+def test_user_example_carrie_env_reconciled(anima):
+    scene = sp._sanitize_scene_fields({
+        "shot_type": "first_person_pov",
+        "pov_cue": "doorway_invite",
+        "action_tags": "arms_out, inviting_hug, smirk, half-lidded_eyes",
+        "environment_tags": "doorway, nighttime, outdoor",
+        "scene_description": (
+            "Carrie stands in her doorway at night, arms outstretched toward you with a mischievous smirk."
+        ),
+    })
+    hybrid = sp.build_positive_prompt_hybrid(
+        scene, PERSONA_CARRIE, "crop_top, ripped_jeans, black_jeans, jeans"
+    )
+    assert "outdoor" not in hybrid.lower() or "doorway" in hybrid
+    assert ", jeans," not in f", {hybrid},"
+    assert "someone" not in hybrid.lower()
+
+
+def test_long_first_mes_uses_final_beat(anima):
+    carrie_tail = (
+        "About an hour later...\n\n"
+        "Carrie stood at her front door, arms out, smirking. "
+        '"Come on, hug me. Now." It\'s getting cold out.'
+    )
+    long = ("She shops for clothes.\n\n" * 5) + carrie_tail
+    excerpt = sp._extract_illustrate_content(long)
+    assert "front door" in excerpt or "hug me" in excerpt
+    assert "shops for clothes" not in excerpt
+
+
+def test_hybrid_gets_fallback_when_no_scene_description(anima):
+    scene = sp._sanitize_scene_fields({
+        "shot_type": "first_person_pov",
+        "pov_cue": "walking_together",
+        "action_tags": "walking, smiling",
+        "environment_tags": "outdoor, daylight",
+        "scene_description": "",
+    })
+    tags_only = sp.build_positive_prompt_tags_only(scene, PERSONA_WOLF, "")
+    hybrid = sp.build_positive_prompt_hybrid(scene, PERSONA_WOLF, "")
+    assert hybrid != tags_only
+    assert "afternoon" in hybrid.lower() or "laughter" in hybrid.lower()
+
+
+def test_yuki_pov_drops_lifting_and_nose_rub(anima):
+    scene = sp._sanitize_scene_fields({
+        "shot_type": "first_person_pov",
+        "pov_cue": "face_to_face",
+        "action_tags": "arms_out, lifting, nose_rub, smiling",
+        "environment_tags": "indoors, warm_lighting",
+        "scene_description": "Her golden eyes soften with warmth toward the camera.",
+    })
+    hybrid = sp.build_positive_prompt_hybrid(scene, {**PERSONA_WOLF, "appearance_tags": "fox_girl, golden_eyes"}, "pink_sweater")
+    assert "lifting" not in hybrid
+    assert "nose_rub" not in hybrid
+    assert "golden" in hybrid.lower()
+
+
+def test_bundle_tags_only_alt_when_dual_compare(anima):
+    scene = {
+        "shot_type": "first_person_pov",
+        "pov_cue": "dialogue_close",
+        "action_tags": "smiling",
+        "environment_tags": "indoors",
+        "scene_description": "Soft light on her face.",
+    }
+    with patch.object(sp, "anima_dual_enabled", return_value=True):
+        bundle = sp._bundle_from_scene(scene, PERSONA_WOLF, "")
+    assert bundle.desc_full is not None
+    assert bundle.desc_full != bundle.tag_full
+    assert "Soft light" in bundle.tag_full
+    assert "Soft light" not in bundle.desc_full.split(sp.NEGATIVE_PROMPT_SEPARATOR)[0]