diff --git a/.env.example b/.env.example
index bb2a5da..9f71f53 100644
--- a/.env.example
+++ b/.env.example
@@ -53,6 +53,14 @@ ANTHROPIC_API_KEY=change_me
 AUTO_BI_ANTHROPIC_MODEL=claude-sonnet-5
 AUTO_BI_ANTHROPIC_MAX_TOKENS=16000
 
+# Direct Mistral API (opt-in alternative: AUTO_BI_LLM_PROVIDER=mistral)
+MISTRAL_API_KEY=change_me
+# Optional AUTO_BI-prefixed alias:
+# AUTO_BI_MISTRAL_API_KEY=change_me
+AUTO_BI_MISTRAL_MODEL=mistral-large-latest
+AUTO_BI_MISTRAL_URL=https://api.mistral.ai
+AUTO_BI_MISTRAL_MAX_TOKENS=16000
+
 # GraceKelly LLM service (local, opt-in alternative: AUTO_BI_LLM_PROVIDER=gracekelly)
 AUTO_BI_GRACEKELLY_URL=http://127.0.0.1:8011
 AUTO_BI_GRACEKELLY_MODEL=claude-sonnet-5
@@ -120,7 +128,7 @@ AUTO_BI_LLM_BUDGET_DAY_MAX_SECONDS=0
 AUTO_BI_LLM_BUDGET_DAY_MAX_COST_USD=0
 # cost price table, USD per 1000 tokens, "model:in/out,..." (example rates — set yours;
 # used only when a *_MAX_COST_USD limit above is set)
-AUTO_BI_LLM_BUDGET_PRICES=claude-opus-4-8:0.005/0.025,claude-sonnet-5:0.003/0.015,claude-sonnet-4-6:0.003/0.015,claude-haiku-4-5:0.001/0.005
+AUTO_BI_LLM_BUDGET_PRICES=claude-opus-4-8:0.005/0.025,claude-sonnet-5:0.003/0.015,claude-sonnet-4-6:0.003/0.015,claude-haiku-4-5:0.001/0.005,mistral-large-latest:0.0005/0.0015
 # Fail-closed remote bind (P0-3): required to serve on non-loopback with auth off
 # and without DEMO_AUTO_ONLY. Prefer AUTH_ENABLED=true on the public internet.
 AUTO_BI_ALLOW_INSECURE_REMOTE=false
diff --git a/.github/workflows/eval-live-sentinel.yml b/.github/workflows/eval-live-sentinel.yml
index 4ab76b4..7dbcc74 100644
--- a/.github/workflows/eval-live-sentinel.yml
+++ b/.github/workflows/eval-live-sentinel.yml
@@ -59,13 +59,15 @@ jobs:
         id: gate
         env:
           ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
+          MISTRAL_API_KEY: ${{ secrets.MISTRAL_API_KEY }}
+          AUTO_BI_MISTRAL_API_KEY: ${{ secrets.AUTO_BI_MISTRAL_API_KEY }}
           AUTO_BI_GRACEKELLY_URL: ${{ secrets.AUTO_BI_GRACEKELLY_URL }}
         run: |
-          if [ -n "${ANTHROPIC_API_KEY}" ] || [ -n "${AUTO_BI_GRACEKELLY_URL}" ]; then
+          if [ -n "${MISTRAL_API_KEY}" ] || [ -n "${AUTO_BI_MISTRAL_API_KEY}" ] || [ -n "${ANTHROPIC_API_KEY}" ] || [ -n "${AUTO_BI_GRACEKELLY_URL}" ]; then
             echo "run=true" >> "$GITHUB_OUTPUT"
           else
             echo "run=false" >> "$GITHUB_OUTPUT"
-            echo "No ANTHROPIC_API_KEY / AUTO_BI_GRACEKELLY_URL secret — sentinel skipped (offline replay still gates quality)."
+            echo "No Mistral / Anthropic / GraceKelly credential secret — sentinel skipped (offline replay still gates quality)."
           fi
 
       - name: Install uv
@@ -84,11 +86,22 @@ jobs:
         if: steps.gate.outputs.run == 'true'
         env:
           ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
+          MISTRAL_API_KEY: ${{ secrets.MISTRAL_API_KEY }}
+          AUTO_BI_MISTRAL_API_KEY: ${{ secrets.AUTO_BI_MISTRAL_API_KEY }}
           AUTO_BI_LLM_PROVIDER: ${{ secrets.AUTO_BI_LLM_PROVIDER }}
           AUTO_BI_GRACEKELLY_URL: ${{ secrets.AUTO_BI_GRACEKELLY_URL }}
           AUTO_BI_GRACEKELLY_MODEL: ${{ secrets.AUTO_BI_GRACEKELLY_MODEL }}
+          AUTO_BI_MISTRAL_MODEL: ${{ secrets.AUTO_BI_MISTRAL_MODEL }}
         run: |
-          export AUTO_BI_LLM_PROVIDER="${AUTO_BI_LLM_PROVIDER:-anthropic}"
+          if [ -z "${AUTO_BI_LLM_PROVIDER}" ]; then
+            if [ -n "${MISTRAL_API_KEY}" ] || [ -n "${AUTO_BI_MISTRAL_API_KEY}" ]; then
+              export AUTO_BI_LLM_PROVIDER="mistral"
+            elif [ -n "${ANTHROPIC_API_KEY}" ]; then
+              export AUTO_BI_LLM_PROVIDER="anthropic"
+            else
+              export AUTO_BI_LLM_PROVIDER="gracekelly"
+            fi
+          fi
           uv run auto_bi eval \
             --suite golden \
             --llm-mode live \
diff --git a/CHANGELOG.md b/CHANGELOG.md
index deb83cd..dd8178c 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -4,6 +4,18 @@
 
 ## [Unreleased]
 
+### Added
+
+- **Direct Mistral provider** — `AUTO_BI_LLM_PROVIDER=mistral` routes through
+  Mistral Chat Completions with standard `MISTRAL_API_KEY` (or the
+  `AUTO_BI_MISTRAL_API_KEY` alias), the shared structured-repair and budget
+  hooks, token-usage logging, safe HTTP errors, and live-sentinel secret
+  routing. Default model: `mistral-large-latest`.
+
+- **Local BYOK runbook** — `docs/LOCAL_BYOK.md`: первый локальный запуск после
+  clone со своим `ANTHROPIC_API_KEY` (ClickHouse + Superset через Compose,
+  `uv run auto_bi serve`, health/ready, остановка и troubleshooting).
+
 ## [0.5.0] - 2026-07-29
 
 ### Fixed
diff --git a/README.md b/README.md
index 638d537..e3fd523 100644
--- a/README.md
+++ b/README.md
@@ -4,12 +4,15 @@
 
 Агент «запрос → дашборд» поверх DM-слоя DWH. Принимает запрос **текстом, drag&drop-раскладкой полей витрин или авто-обзором витрины** (детерминированный курируемый дашборд без LLM), уточняет детали только при реальных расхождениях с данными, честно предупреждает о не предусмотренных витриной паттернах (engine-aware **Feasibility Advisor** — вплоть до «это запрос на новую витрину»), строит дашборд в выбранной BI и возвращает ссылку.
 
-**Скоуп v1 (RU-рынок, release-gated в CI):** ClickHouse (DM) + Apache Superset (BI). **v2 experimental:** Greengage/Greenplum (offline advisor/golden в CI; live DWH — operator stand) + Yandex DataLens (unit compile offline; live contract Mac-only, не default release gate). Универсальность — в швах (IR, адаптеры), не в имплементации.
-**LLM:** прямой Anthropic Messages API (по умолчанию — нужен только `ANTHROPIC_API_KEY`); локальный сервис GraceKelly — документированная опция (`AUTO_BI_LLM_PROVIDER=gracekelly`, см. [USER_GUIDE §6](docs/USER_GUIDE.md#6-конфигурация-переменные-окружения)).
+**Скоуп v1 (RU-рынок, release-gated в CI):** ClickHouse (DM) + Apache Superset (BI). **v2 experimental:** Greengage/Greenplum (offline advisor/golden в CI; live DWH — operator stand) + Yandex DataLens (offline compile contracts и повторный live contract **15/15** на Mac; не default release gate). Универсальность — в швах (IR, адаптеры), не в имплементации.
+**LLM:** прямой Anthropic Messages API по умолчанию; прямой Mistral Chat Completions (`AUTO_BI_LLM_PROVIDER=mistral`, `MISTRAL_API_KEY`) и локальный сервис GraceKelly — документированные опции (см. [USER_GUIDE §6](docs/USER_GUIDE.md#6-конфигурация-переменные-окружения)).
 
 ## Демо
 
-**Живое демо: <https://juliome20-auto-bi-demo.hf.space>** — публичная песочница (Hugging Face Space, один контейнер ClickHouse + Superset + Auto_BI): выберите витрину, соберите авто-обзор и откройте готовый дашборд в Superset без логина. Работает детерминированный путь без LLM; полный текстовый цикл — на видео ниже. Данные синтетические, всё пересоздаётся при рестарте (холодный старт ~3 мин).
+**Поддерживаемый путь без стенда:** офлайн golden path
+(`uv run python scripts/demo_golden_path.py`) — детерминированный IR/SQL/advisor
+без DWH, BI и LLM. Полный локальный запуск со своим API key —
+[docs/LOCAL_BYOK.md](docs/LOCAL_BYOK.md).
 
 ![Auto_BI — полный цикл: текст → уточнение → спецификация + advisor → сборка → дашборд Superset](docs/screenshots/demo.gif)
 
@@ -23,9 +26,14 @@
 
 ## Статус
 
-**Phase 0–4 + бэклог адекватности дашбордов (B1–B4) закрыты.** Работает end-to-end: текст/поля → spec → валидация → сборка дашборда. v1-стек (ClickHouse + Superset) и v2-стек (Greenplum/Greengage интроспекция + advisor; self-hosted DataLens-адаптер) live-проверены; web UI с двумя режимами ввода, итерациями, Feasibility Advisor, заявками владельцу DM и панелью наблюдаемости.
+**Phase 0–4 + бэклог адекватности дашбордов (B1–B4) закрыты.** Работает end-to-end: текст/поля → spec → валидация → сборка дашборда. v1-стек (ClickHouse + Superset) live-проверен на v0.5.0; v2 (Greenplum/Greengage advisor/golden; DataLens) — offline evidence/contracts, а DataLens Mac-only live contract повторно прошёл **15/15** на текущем self-hosted stand 2026-07-29 (experimental / non-default / non-closure; фактический seeded workbook задавался через `AUTO_BI_DATALENS_WORKBOOK_ID`); web UI с двумя режимами ввода, итерациями, Feasibility Advisor, заявками владельцу DM и панелью наблюдаемости.
 
-**Актуальное состояние и residual roadmap** — [docs/CURRENT_STATE.md](docs/CURRENT_STATE.md). История фаз — [docs/PLAN.md](docs/PLAN.md). Полный env inventory (generated) — [docs/ENV_REFERENCE.md](docs/ENV_REFERENCE.md).
+Все пять вынесенных external live validations завершены exact evidence:
+DataLens 15/15, direct Mistral sentinel 3/3, protected-tag rejection,
+intentional Trivy failure before promotion и process-restart reconciliation.
+Активной audit work не осталось.
+
+**Актуальное состояние и closure evidence** — [docs/CURRENT_STATE.md](docs/CURRENT_STATE.md). История фаз — [docs/PLAN.md](docs/PLAN.md). Полный env inventory (generated) — [docs/ENV_REFERENCE.md](docs/ENV_REFERENCE.md).
 
 ## Чем отличается
 
@@ -54,7 +62,7 @@ flowchart LR
 Установка, команды CLI, web UI, конфигурация — [docs/USER_GUIDE.md](docs/USER_GUIDE.md).
 Подключение новой витрины DWH за ≤ 1 ч — [docs/ONBOARDING_DWH.md](docs/ONBOARDING_DWH.md).
 
-Local-first — три ступени:
+Local-first — два поддерживаемых пути:
 
 1. **Офлайн golden path** — без DWH, BI, LLM и API-ключа:
 
@@ -62,9 +70,7 @@ Local-first — три ступени:
 uv run python scripts/demo_golden_path.py
 ```
 
-2. **HF Space** — детерминированный auto-only, пользовательский ключ не нужен; текстовый режим там намеренно недоступен (см. «Демо» выше).
-
-3. **Полный локальный путь.** Скопируйте `.env.example` в `.env` (`cp .env.example .env`; PowerShell: `Copy-Item .env.example .env`). Задайте свой `ANTHROPIC_API_KEY` **или** `AUTO_BI_LLM_PROVIDER=gracekelly` и `AUTO_BI_GRACEKELLY_URL`. Для DWH/BI — `AUTO_BI_CH_HOST`, `AUTO_BI_CH_PASSWORD`, `AUTO_BI_SUPERSET_URL`, `AUTO_BI_SUPERSET_PASSWORD` (полный inventory — [docs/ENV_REFERENCE.md](docs/ENV_REFERENCE.md)). `docker compose up -d` поднимает **только ClickHouse и Superset**, не Auto_BI; агент локально: `auto_bi serve` → http://127.0.0.1:8200.
+2. **Полный локальный путь.** Пошаговый Anthropic-пример — [docs/LOCAL_BYOK.md](docs/LOCAL_BYOK.md). Скопируйте `.env.example` в `.env` (`cp .env.example .env`; PowerShell: `Copy-Item .env.example .env`). Задайте свой `ANTHROPIC_API_KEY`; либо `AUTO_BI_LLM_PROVIDER=mistral` + `MISTRAL_API_KEY`; либо `AUTO_BI_LLM_PROVIDER=gracekelly` + `AUTO_BI_GRACEKELLY_URL`. Для DWH/BI — `AUTO_BI_CH_HOST`, `AUTO_BI_CH_PASSWORD`, `AUTO_BI_SUPERSET_URL`, `AUTO_BI_SUPERSET_PASSWORD` (полный inventory — [docs/ENV_REFERENCE.md](docs/ENV_REFERENCE.md)). `docker compose up -d` поднимает **только ClickHouse и Superset**, не Auto_BI; агент локально: `auto_bi serve` → http://127.0.0.1:8200.
 
 ```bash
 pip install autobi-agent                          # или pip install -e . из корня репозитория
@@ -90,6 +96,7 @@ uv run python scripts/demo_golden_path.py
 | Файл | Что внутри |
 |---|---|
 | [docs/USER_GUIDE.md](docs/USER_GUIDE.md) | Руководство пользователя: установка, команды CLI, web UI, два режима ввода, advisor, наблюдаемость, конфигурация |
+| [docs/LOCAL_BYOK.md](docs/LOCAL_BYOK.md) | Первый локальный запуск со своим Anthropic API key: clone, `.env`, Compose (CH+Superset), `uv run auto_bi serve`, health/ready |
 | [docs/ONBOARDING_DWH.md](docs/ONBOARDING_DWH.md) | Подключение нового DWH за ≤ 1 ч: доступы, `.env`, интроспекция, обогащение, проверка (ClickHouse + Greenplum) |
 | [docs/ARCHITECTURE.md](docs/ARCHITECTURE.md) | Архитектура: скоуп, IR-first, семантическая модель с физическим слоем, агент, Feasibility Advisor, адаптеры, LLM-слой, решения D1–D10, риски |
 | [docs/DEPLOYMENT.md](docs/DEPLOYMENT.md) | Деплой в проде: workers=1, reverse-proxy/TLS, готовность, docker-compose, бэкап SQLite, ротация логов, чеклист секретов |
diff --git a/auto_bi/cli.py b/auto_bi/cli.py
index 504c7f2..0dbfac2 100644
--- a/auto_bi/cli.py
+++ b/auto_bi/cli.py
@@ -726,11 +726,13 @@ def bi_healthcheck() -> AdapterHealth:
         return probe_health(adapter_for, TargetBI.SUPERSET)
 
     def llm_healthcheck() -> AdapterHealth:
-        if settings.llm_provider.strip().lower() != "gracekelly":
-            # Anthropic is a hosted API with no separate process to be "up/down" locally,
-            # and an actual completion call would cost tokens on every readiness probe —
-            # report configured-and-constructible (already proven by make_llm below).
-            return AdapterHealth(ok=True, message="anthropic: no live check (avoids token cost)")
+        provider = settings.llm_provider.strip().lower()
+        if provider != "gracekelly":
+            # Hosted providers (anthropic, mistral, …) have no separate process to be
+            # "up/down" locally, and an actual completion call would cost tokens on every
+            # readiness probe — report configured-and-constructible (already proven by
+            # make_llm below).
+            return AdapterHealth(ok=True, message=f"{provider}: no live check (avoids token cost)")
         import httpx
 
         try:
@@ -1010,9 +1012,11 @@ def _render(title: str, report: EvalReport) -> None:
             store = Store(settings.store_path)
             live_llm = make_llm(settings, store=store)
             provider = settings.llm_provider.strip().lower()
-            model_id = (
-                settings.gracekelly_model if provider == "gracekelly" else settings.anthropic_model
-            )
+            model_id = {
+                "gracekelly": settings.gracekelly_model,
+                "anthropic": settings.anthropic_model,
+                "mistral": settings.mistral_model,
+            }.get(provider, provider)
             provider_detail = (
                 f"{settings.gracekelly_url}, {model_id}" if provider == "gracekelly" else model_id
             )
diff --git a/auto_bi/config.py b/auto_bi/config.py
index f90e062..b1d8a19 100644
--- a/auto_bi/config.py
+++ b/auto_bi/config.py
@@ -5,7 +5,7 @@
 from functools import lru_cache
 from logging import Logger
 
-from pydantic import Field
+from pydantic import AliasChoices, Field
 from pydantic_settings import BaseSettings, SettingsConfigDict
 
 
@@ -154,14 +154,13 @@ class Settings(BaseSettings):
         ),
     )
 
-    # LLM provider seam (llm/factory.py): "anthropic" (default — direct Anthropic Messages
-    # API, works out of the box with just an API key) or "gracekelly" (local orchestration
-    # service, documented opt-in — ARCHITECTURE §3.6).
+    # LLM provider seam (llm/factory.py): direct Anthropic (default), direct Mistral,
+    # or GraceKelly (local orchestration service).
     llm_provider: str = Field(
         default="anthropic",
         description=(
-            'LLM backend selector: "anthropic" for direct API or "gracekelly" for the local'
-            " orchestration service."
+            'LLM backend selector: "anthropic" or "mistral" for direct API access, or'
+            ' "gracekelly" for the local orchestration service.'
         ),
     )
 
@@ -202,6 +201,33 @@ class Settings(BaseSettings):
         ),
     )
 
+    # Direct Mistral chat-completions API. The client also accepts the standard
+    # MISTRAL_API_KEY process env when this AUTO_BI-prefixed field is empty.
+    mistral_api_key: str = Field(
+        default="",
+        validation_alias=AliasChoices(
+            "mistral_api_key",
+            "AUTO_BI_MISTRAL_API_KEY",
+            "MISTRAL_API_KEY",
+        ),
+        description=(
+            "Mistral API key for direct calls; empty lets the client fall back to"
+            " MISTRAL_API_KEY."
+        ),
+    )
+    mistral_model: str = Field(
+        default="mistral-large-latest",
+        description="Mistral model id used when llm_provider is mistral.",
+    )
+    mistral_url: str = Field(
+        default="https://api.mistral.ai",
+        description="Mistral API base URL; the client appends /v1/chat/completions.",
+    )
+    mistral_max_tokens: int = Field(
+        default=16000,
+        description="Maximum output tokens requested from direct Mistral chat completions.",
+    )
+
     # plan_sol step 2 / audit P0-2: DWH values (top-N) leave the process only on
     # explicit opt-in. Default false — clean install never sends samples to an
     # external LLM. Set true only for public/internal classes after classification.
@@ -380,7 +406,7 @@ class Settings(BaseSettings):
         ),
     )
     # cost price table (USD per 1000 tokens), "model:in/out,...". List prices as of
-    # 2026-07-18; override for your provider contract. Used only when a *_max_cost_usd
+    # 2026-07-29; override for your provider contract. Used only when a *_max_cost_usd
     # limit is set — an unlisted model prices at 0, so add yours before relying on a cap.
     # Sonnet 5 carries a lower introductory rate through 2026-08-31; the table keeps the
     # standard rate so the guard errs toward over-estimating spend, not under.
@@ -389,7 +415,8 @@ class Settings(BaseSettings):
             "claude-opus-4-8:0.005/0.025,"
             "claude-sonnet-5:0.003/0.015,"
             "claude-sonnet-4-6:0.003/0.015,"
-            "claude-haiku-4-5:0.001/0.005"
+            "claude-haiku-4-5:0.001/0.005,"
+            "mistral-large-latest:0.0005/0.0015"
         ),
         description=(
             "USD-per-1k-token price table as model:in/out pairs; unlisted models price at 0 until"
@@ -533,8 +560,8 @@ def unknown_env_settings(environ: Mapping[str, str] | None = None) -> list[str]:
     `extra="ignore"` silently drops typos — `AUTO_BI_AUTH_ENABLE=true` leaves auth OFF
     with no trace. `serve` reports every returned name as a warning so a misspelled
     security flag is visible in the log instead of silently inert. Compares against
-    `Settings.model_fields` plus any explicit string validation_alias (none today;
-    AliasChoices would need unpacking if ever introduced).
+    `Settings.model_fields` plus any explicit string validation_alias. Field-derived
+    `AUTO_BI_*` names remain recognised when a field also accepts non-prefixed aliases.
     """
     env = os.environ if environ is None else environ
     prefix = str(Settings.model_config.get("env_prefix", "")).upper()
diff --git a/auto_bi/llm/_structured.py b/auto_bi/llm/_structured.py
index 3822569..f51a9bf 100644
--- a/auto_bi/llm/_structured.py
+++ b/auto_bi/llm/_structured.py
@@ -1,6 +1,6 @@
 """Shared structured-output machinery for LLM clients (transport-agnostic).
 
-Both GraceKellyClient and AnthropicClient turn text-in/text-out completions into
+GraceKellyClient, AnthropicClient and MistralClient turn text-in/text-out completions into
 schema-validated objects via the SAME JSON-extraction + repair loop (invariant 1:
 the LLM emits only DashboardSpec/etc. as JSON; we parse and validate it here, never
 trusting native formats). Keeping this here means the two clients differ only in
@@ -134,8 +134,9 @@ def append_llm_log(
     """Append one LLM-call record to the jsonl log and (if present) the durable Store.
 
     The prompt itself is NEVER logged — only its sha256 prefix and length (security §4).
-    `input_tokens`/`output_tokens` are real usage from providers that report it (Anthropic);
-    None where the provider returns no usage (GraceKelly) or the call failed before a response.
+    `input_tokens`/`output_tokens` are real usage from providers that report it
+    (Anthropic/Mistral); None where the provider returns no usage (GraceKelly) or
+    the call failed before a response.
     Logging is best-effort: a failure here must never kill the pipeline.
     """
     prompt_sha256 = hashlib.sha256(prompt.encode()).hexdigest()[:16]
diff --git a/auto_bi/llm/budget.py b/auto_bi/llm/budget.py
index 0624c07..0496b6b 100644
--- a/auto_bi/llm/budget.py
+++ b/auto_bi/llm/budget.py
@@ -17,7 +17,8 @@
 Usage is read back from the existing `llm_calls` ledger (Store), which already records
 every attempt with tokens/latency (`Store.log_llm_call`), so budgets survive across
 requests and restarts without a parallel table. Tokens are real where the provider
-reports them (Anthropic) and char-estimated (chars / 4) where it does not (GraceKelly).
+reports them (Anthropic/Mistral) and char-estimated (chars / 4) where it does not
+(GraceKelly).
 
 Fail closed: `check` raises `BudgetExceeded` BEFORE issuing the call that would cross a
 limit, naming the exceeded dimension. Opt-in, off by default (AUTO_BI_LLM_BUDGET_ENABLED),
diff --git a/auto_bi/llm/factory.py b/auto_bi/llm/factory.py
index c4f787f..a8d498b 100644
--- a/auto_bi/llm/factory.py
+++ b/auto_bi/llm/factory.py
@@ -29,8 +29,13 @@ def make_llm(settings: Settings, store: Store | None = None) -> LLMClient:
         from auto_bi.llm.anthropic import AnthropicClient
 
         return AnthropicClient(settings, store=store, budget=budget)
+    if provider == "mistral":
+        from auto_bi.llm.mistral import MistralClient
+
+        return MistralClient(settings, store=store, budget=budget)
     raise ValueError(
-        f"unknown AUTO_BI_LLM_PROVIDER {settings.llm_provider!r} (use 'gracekelly' or 'anthropic')"
+        f"unknown AUTO_BI_LLM_PROVIDER {settings.llm_provider!r} "
+        "(use 'gracekelly', 'anthropic', or 'mistral')"
     )
 
 
diff --git a/auto_bi/llm/fixture.py b/auto_bi/llm/fixture.py
index a3b8a3a..fd4aa28 100644
--- a/auto_bi/llm/fixture.py
+++ b/auto_bi/llm/fixture.py
@@ -27,7 +27,7 @@
       "format_version": 2,
       "template_version": "<16 hex of prompt templates>",
       "schema_version": "<16 hex of IR JSON schemas>",
-      "provider": "anthropic|gracekelly|fixture-refresh|...",
+      "provider": "anthropic|mistral|gracekelly|fixture-refresh|...",
       "model_id": "...",
       "calls": [
         {"step": "...", "schema": "...", "prompt_sha256": "...", "response": {...}}
diff --git a/auto_bi/llm/mistral.py b/auto_bi/llm/mistral.py
new file mode 100644
index 0000000..5abff07
--- /dev/null
+++ b/auto_bi/llm/mistral.py
@@ -0,0 +1,211 @@
+"""Direct Mistral chat-completions client behind the shared LLMClient seam."""
+
+from __future__ import annotations
+
+import os
+import time
+from collections.abc import Callable
+from pathlib import Path
+from typing import TYPE_CHECKING, Any, TypeVar, cast
+
+import httpx
+from pydantic import BaseModel
+
+from auto_bi.config import Settings
+from auto_bi.llm._structured import append_llm_log, complete_with_repair
+from auto_bi.llm.base import LLMError
+
+if TYPE_CHECKING:
+    from auto_bi.llm.budget import LLMBudget
+    from auto_bi.store import Store
+
+T = TypeVar("T", bound=BaseModel)
+
+_MAX_RATE_LIMIT_RETRIES = 4
+_BACKOFF_BASE_SECONDS = 2.0
+_BACKOFF_CAP_SECONDS = 30.0
+
+
+def _extract_usage(data: dict[str, Any]) -> tuple[int | None, int | None]:
+    usage = data.get("usage")
+    if not isinstance(usage, dict):
+        return None, None
+    prompt = usage.get("prompt_tokens")
+    completion = usage.get("completion_tokens")
+    return (
+        prompt if isinstance(prompt, int) else None,
+        completion if isinstance(completion, int) else None,
+    )
+
+
+def _extract_choice(data: dict[str, Any]) -> tuple[str, str]:
+    """Return (text, finish_reason) for string or current chunked Mistral content."""
+    choices = data.get("choices")
+    if not isinstance(choices, list) or not choices or not isinstance(choices[0], dict):
+        return "", "unknown"
+    first = choices[0]
+    finish_reason = first.get("finish_reason")
+    status = finish_reason if isinstance(finish_reason, str) and finish_reason else "unknown"
+    message = first.get("message")
+    if not isinstance(message, dict):
+        return "", status
+    content = message.get("content")
+    if isinstance(content, str):
+        return content, status
+    if not isinstance(content, list):
+        return "", status
+    parts: list[str] = []
+    for chunk in content:
+        if isinstance(chunk, dict) and isinstance(chunk.get("text"), str):
+            parts.append(chunk["text"])
+    return "".join(parts), status
+
+
+def _retry_after_seconds(response: httpx.Response, attempt: int) -> float:
+    header = response.headers.get("Retry-After")
+    if header:
+        try:
+            return min(_BACKOFF_CAP_SECONDS, max(0.0, float(header)))
+        except ValueError:
+            pass
+    return min(_BACKOFF_CAP_SECONDS, _BACKOFF_BASE_SECONDS * (2.0**attempt))
+
+
+class MistralClient:
+    """Sync text-in/JSON-out client for Mistral's `/v1/chat/completions` API."""
+
+    def __init__(
+        self,
+        settings: Settings,
+        http: httpx.Client | None = None,
+        log_path: str | Path = "logs/llm_calls.jsonl",
+        store: Store | None = None,
+        budget: LLMBudget | None = None,
+        sleep: Callable[[float], None] = time.sleep,
+    ) -> None:
+        self._settings = settings
+        api_key = settings.mistral_api_key or os.environ.get("MISTRAL_API_KEY", "")
+        if not api_key:
+            raise LLMError(
+                "Mistral API key is not configured; set MISTRAL_API_KEY or "
+                "AUTO_BI_MISTRAL_API_KEY"
+            )
+        self._http = http or httpx.Client(
+            base_url=settings.mistral_url,
+            timeout=httpx.Timeout(300.0, connect=10.0),
+            transport=httpx.HTTPTransport(retries=2),
+            headers={"Authorization": f"Bearer {api_key}"},
+        )
+        self._log_path = Path(log_path)
+        self._store = store
+        self._budget = budget
+        self._sleep = sleep
+
+    def close(self) -> None:
+        """Release the owned or injected HTTP pool."""
+        self._http.close()
+
+    def complete(
+        self,
+        prompt: str,
+        schema: type[T],
+        *,
+        reasoning: bool = False,
+        session_id: str | None = None,
+        step: str = "",
+    ) -> T:
+        return cast(
+            T,
+            complete_with_repair(
+                lambda value: self._call(
+                    value,
+                    reasoning=reasoning,
+                    session_id=session_id,
+                    step=step,
+                ),
+                prompt,
+                schema,
+                on_attempt=self._budget_hook(session_id),
+            ),
+        )
+
+    def _budget_hook(self, session_id: str | None) -> Callable[[], None] | None:
+        if self._budget is None:
+            return None
+        budget = self._budget
+        model = self._settings.mistral_model
+        return lambda: budget.check(session_id=session_id, model=model)
+
+    def _post(self, payload: dict[str, Any]) -> dict[str, Any]:
+        for attempt in range(1 + _MAX_RATE_LIMIT_RETRIES):
+            response = self._http.post("/v1/chat/completions", json=payload)
+            if response.status_code == 429:
+                if attempt < _MAX_RATE_LIMIT_RETRIES:
+                    self._sleep(_retry_after_seconds(response, attempt))
+                    continue
+                raise LLMError(
+                    "Mistral API still rate-limited after " f"{_MAX_RATE_LIMIT_RETRIES} retries"
+                )
+            if response.status_code >= 400:
+                # Provider bodies can echo request/auth diagnostics; never expose them.
+                raise LLMError(f"Mistral API HTTP {response.status_code}")
+            try:
+                data = response.json()
+            except ValueError as exc:
+                raise LLMError("Mistral API returned invalid JSON") from exc
+            if not isinstance(data, dict):
+                raise LLMError("Mistral API returned a non-object response")
+            return data
+        raise LLMError("Mistral API retry loop exited unexpectedly")
+
+    def _call(
+        self,
+        prompt: str,
+        *,
+        reasoning: bool,
+        session_id: str | None,
+        step: str,
+    ) -> str:
+        payload = {
+            "model": self._settings.mistral_model,
+            "messages": [{"role": "user", "content": prompt}],
+            "max_tokens": self._settings.mistral_max_tokens,
+            "temperature": 0,
+        }
+        started = time.monotonic()
+        status = "transport_error"
+        completion_chars = 0
+        input_tokens: int | None = None
+        output_tokens: int | None = None
+        try:
+            data = self._post(payload)
+            input_tokens, output_tokens = _extract_usage(data)
+            text, finish_reason = _extract_choice(data)
+            status = (
+                "completed"
+                if finish_reason in {"stop", "length", "model_length"}
+                else finish_reason
+            )
+            completion_chars = len(text)
+            if not text:
+                raise LLMError(f"Mistral returned no text (finish_reason={finish_reason})")
+            return text
+        except LLMError:
+            raise
+        except httpx.HTTPError as exc:
+            raise LLMError(f"Mistral transport error: {exc}") from exc
+        finally:
+            append_llm_log(
+                self._log_path,
+                self._store,
+                model=self._settings.mistral_model,
+                prompt=prompt,
+                reasoning=reasoning,
+                status=status,
+                latency_ms=round((time.monotonic() - started) * 1000),
+                session_id=session_id,
+                step=step,
+                completion_chars=completion_chars,
+                input_tokens=input_tokens,
+                output_tokens=output_tokens,
+            )
diff --git a/auto_bi/store/db.py b/auto_bi/store/db.py
index 162c1fe..b5957ab 100644
--- a/auto_bi/store/db.py
+++ b/auto_bi/store/db.py
@@ -11,10 +11,10 @@
 of agent steps (grounding/propose/advisor/approve) and build phases.
 
 Schema v5 (token accounting, E2): `llm_calls` gained nullable `input_tokens` /
-`output_tokens`. The Anthropic Messages API returns `usage.input_tokens/output_tokens`,
-so calls on that provider carry real tokens; GraceKelly reports no usage and a transport
-error has no response, so those rows stay NULL (NULL = "no usage reported", distinct from
-a real zero — `completion_chars` remains the universal size proxy for every call).
+`output_tokens`. Anthropic and Mistral return input/output token usage, so calls on those
+providers carry real tokens; GraceKelly reports no usage and a transport error has no
+response, so those rows stay NULL (NULL = "no usage reported", distinct from a real zero
+— `completion_chars` remains the universal size proxy for every call).
 
 Schema v6 (B-4 hardening): `auth_tokens.token` now stores sha256(raw token) hex, not the
 raw bearer token — a stolen SQLite file no longer yields live sessions directly. The
@@ -183,7 +183,7 @@ def _encode_spec_snapshot(spec_json: dict[str, Any]) -> tuple[str, str]:
     step          TEXT NOT NULL DEFAULT '',
     completion_chars INTEGER NOT NULL DEFAULT 0,
     input_tokens  INTEGER,  -- NULL = provider reported no usage (GraceKelly / transport error)
-    output_tokens INTEGER   -- real tokens only where the provider returns usage (Anthropic)
+    output_tokens INTEGER   -- real tokens where the provider returns usage (Anthropic/Mistral)
 );
 -- llm/budget.py reads this ledger per session and per rolling window on every provider
 -- round-trip; index the two scope keys (created via always-run CREATE IF NOT EXISTS, no
@@ -309,7 +309,7 @@ def _migrate(self) -> None:
             # v4: dm_change_requests carries the advisor's concrete fix artifact (DDL)
             self._add_column("dm_change_requests", "remediation", "TEXT NOT NULL DEFAULT ''")
         if version < 5:
-            # v5: real token usage on providers that report it (Anthropic); nullable so
+            # v5: real token usage on providers that report it (Anthropic/Mistral); nullable so
             # legacy/GraceKelly rows stay NULL rather than a misleading zero
             self._add_column("llm_calls", "input_tokens", "INTEGER")
             self._add_column("llm_calls", "output_tokens", "INTEGER")
@@ -1019,10 +1019,10 @@ def llm_usage_summary(
     ) -> dict[str, Any]:
         """Aggregates for the LLM-usage dashboard. Char volumes are a universal size
         proxy (every call has them). Real `input_tokens`/`output_tokens` are summed
-        NULL-ignoring — they are populated only on providers that report usage (Anthropic);
-        GraceKelly reports none, so its rows stay NULL. `token_calls` counts the rows that
-        carry real tokens, so callers can show token figures only when they exist rather
-        than presenting a NULL-driven 0 as if it were measured.
+        NULL-ignoring — they are populated only on providers that report usage
+        (Anthropic/Mistral); GraceKelly reports none, so its rows stay NULL. `token_calls`
+        counts the rows that carry real tokens, so callers can show token figures only when
+        they exist rather than presenting a NULL-driven 0 as if it were measured.
 
         `owner` (P1-4): when set, only calls whose session is owned by that username —
         used for non-admin observability so a user never sees foreign spend.
@@ -1082,9 +1082,10 @@ def _llm_usage(self, where: str, params: tuple[Any, ...]) -> dict[str, Any]:
 
         Returns `calls`, `latency_ms`, total estimated `tokens`, and a per-`model`
         breakdown so the enforcer can price cost. Tokens are the provider's real usage
-        where reported (Anthropic), else char-estimated (chars / 4) so a token budget
-        still bites on GraceKelly, which reports none. Every attempt is counted (a repair
-        is a distinct row), independent of status — a budget must see all round-trips.
+        where reported (Anthropic/Mistral), else char-estimated (chars / 4) so a token
+        budget still bites on GraceKelly, which reports none. Every attempt is counted
+        (a repair is a distinct row), independent of status — a budget must see all
+        round-trips.
         """
         rows = self._rows(
             "SELECT model,"
diff --git a/docs/ARCHITECTURE.md b/docs/ARCHITECTURE.md
index e275497..0d186c9 100644
--- a/docs/ARCHITECTURE.md
+++ b/docs/ARCHITECTURE.md
@@ -208,7 +208,7 @@ Factory — единственное место выбора конкретно
 contracts закреплены unit и live contract tests. Crash-recovery contract:
 [ADR 0002](adr/0002-durable-build-attempt-reconciliation.md).
 
-### 3.6 LLM Layer — Anthropic (default) + GraceKelly (opt-in)
+### 3.6 LLM Layer — Anthropic (default) + Mistral / GraceKelly (opt-in)
 
 Agent core зависит только от:
 
@@ -216,7 +216,8 @@ Agent core зависит только от:
 LLMClient.complete(prompt, schema) -> ValidatedModel
 ```
 
-Factory выбирает direct Anthropic Messages API или локальный GraceKelly.
+Factory выбирает direct Anthropic Messages API, direct Mistral Chat Completions
+API или локальный GraceKelly.
 Structured repair loop общий: JSON extraction, Pydantic validation, bounded
 feedback retries и durable call logging.
 
@@ -442,7 +443,7 @@ Superset guard проверяет фактически исполняемый so
 |---|---|---|
 | D1 | IR-first | Мульти-BI остаётся тестируемым и управляемым |
 | D2 | Superset primary, DataLens secondary | Два реальных адаптера на общей границе |
-| D3 | `LLMClient` с direct Anthropic default | Внешняя установка не зависит от локального orchestration service |
+| D3 | `LLMClient` с direct Anthropic default и direct Mistral option | Внешняя установка не зависит от локального orchestration service |
 | D4 | Semantic model в versioned YAML | Review, diff и ручное владение семантикой |
 | D5 | LLM думает, код исполняет | Валидация и native payloads остаются детерминированными |
 | D6 | Python, FastAPI, Pydantic, sqlglot, httpx | Простая state machine не требует тяжёлого agent framework |
diff --git a/docs/CURRENT_STATE.md b/docs/CURRENT_STATE.md
index c0766f9..db724ca 100644
--- a/docs/CURRENT_STATE.md
+++ b/docs/CURRENT_STATE.md
@@ -12,11 +12,13 @@
 **Версия пакета:** см. `auto_bi.__version__` / `pyproject.toml` (ratchet в
 `tests/test_docs_defaults.py`).
 
-**Режим:** closure candidate (не fully closed). Software release **v0.5.0**
-complete на exact `main`/tag SHA
-`e78076d2d00ddc1748bf6e22f13cf7cb93fc6515`. Закрываемый scope и residual —
-[PROJECT_CLOSURE.md](PROJECT_CLOSURE.md). Полное закрытие ждёт mandatory
-external gate: public HF demo sync to closing SHA or decommission.
+**Режим:** software release **v0.5.0** complete на exact `main`/tag SHA
+`e78076d2d00ddc1748bf6e22f13cf7cb93fc6515`. Локальный/software closure scope
+закрыт; пять вынесенных external validations получили exact live evidence
+2026-07-29. Активная audit work = **0**. HF полностью исключён из current
+project scope и backlog; historical token `owner-de-scoped` сохраняется только
+для reconciliation исходных 24 строк. Scope и disposition —
+[PROJECT_CLOSURE.md](PROJECT_CLOSURE.md).
 
 ## Продукт
 
@@ -27,8 +29,7 @@ external gate: public HF demo sync to closing SHA or decommission.
 | Greenplum advisor + golden | **offline contract in CI** | advisor GP + golden GP replay |
 | Greenplum live DWH | experimental | operator stand |
 | DataLens compile path | offline contracts **passed** | contract suite |
-| DataLens live stand | **experimental** / unavailable (Mac-only stand absent) | not default release gate |
-| Public HF demo | **out of sync** | health reports `0.4.0`, `demo_auto_only=false`, no capabilities; live assertion fails; v0.5.0 publish dry-run OK; sync blocked (no local `HF_TOKEN`, no repo HF secret/workflow). Gate: sync to closing SHA or decommission |
+| DataLens live stand | **`closed`** 2026-07-29: Mac-only self-hosted contract **15/15 passed**; текущий image seed использует workbook `z4wtz6tg5194o`, переданный через supported env override | experimental / non-default / **non-closure** |
 
 ## Безопасность и runtime (plan_sol 1–4)
 
@@ -41,6 +42,18 @@ external gate: public HF demo sync to closing SHA or decommission.
 
 - GitHub `main` + `v*` tag rulesets **active without bypass**; Dependabot open
   queue empty after sequential disposition.
+- Protected-tag mutation rejection **live-проверен 2026-07-29**: canary
+  [`v-retag-smoke-20260729`](https://github.com/brownjuly2003-code/Auto_BI/tree/v-retag-smoke-20260729)
+  создан на remote `main` `13fc855`; попытка force-update на другой remote
+  object отклонена GitHub с HTTP 422 (`Cannot update this protected ref` /
+  `Cannot force-push to this tag`), ref остался неизменным; release workflow
+  для canary не запускался.
+- Intentional Trivy-fail-before-promotion **live-проверен** run
+  [30512999822](https://github.com/brownjuly2003-code/Auto_BI/actions/runs/30512999822):
+  Trivy отклонил **23** исправимых HIGH/CRITICAL findings в probe-image,
+  workflow завершился ожидаемым failure, а GHCR `:latest` сохранил digest
+  `sha256:bff75bcef9d894e86c2be63a584284c02c2b2546425ac11a15e83ad83e4e84f1`;
+  временная remote branch удалена.
 - **v0.5.0 external evidence (complete):**
   - post-merge CI, CodeQL, Gitleaks, Demo image — passed on closing SHA
     `e78076d2d00ddc1748bf6e22f13cf7cb93fc6515`;
@@ -98,7 +111,7 @@ resume / **fields DnD seed**). Post-merge CI on v0.5.0 closing SHA passed.
 |---|---|
 | Online SQLite backup + integrity | `Store.backup_to` / `integrity_check`; `scripts/store_backup.py` |
 | Restore drill | script + `tests/test_store_backup.py` |
-| Pre-return BI crash recovery | RR-4 durable attempts + exact Superset/DataLens cleanup ([ADR 0002](adr/0002-durable-build-attempt-reconciliation.md)) |
+| Pre-return BI crash recovery | RR-4 durable attempts + exact Superset/DataLens cleanup ([ADR 0002](adr/0002-durable-build-attempt-reconciliation.md)); live DataLens process-death smoke: child exit 97, remote delivery confirmed, discovered/deleted 3/3, session failed |
 | Offline perf baseline | `tests/test_perf_baseline.py` (soft abs + relative ratios) |
 | Property / metamorphic | `tests/test_property_quality.py` (guard, validate, normalize, **RBAC**, Superset native-filter scope) |
 | Cumulative bounded mutation gate | CI targets `auto_bi/agent/sql_guard.py`, `auto_bi/ir/validate.py`, `auto_bi/agent/dataset_plan.py`, `auto_bi/agent/cleanup.py`; weak outcomes rejected; detailed snapshot evidence in [operations/SLO.md](operations/SLO.md) |
@@ -115,20 +128,63 @@ container cold start **6341 ms**, PID1 RSS **89.594 MiB**, cgroup memory
 
 Evidence: [operations/REAUDIT_plan_sol_23_07_26.md](operations/REAUDIT_plan_sol_23_07_26.md).
 Offline score **8.8/10** (was 8.2). Finding R1 fixed (SafeError test fake vs `build(spec, ctx)`).
-GHA / protected `v0.5.0` tag / package-image publish evidence complete. Remaining
-external residual: public HF demo sync to closing SHA or decommission (not
-decommissioned; sync blocked by missing HF token/secret/workflow).
+GHA / protected `v0.5.0` tag / package-image publish evidence complete. Scope,
+live evidence и final accounting закрыты; excluded external demo paths не
+являются residual или current product path.
 
 ## Closure disposition
 
-Прежний open-ended residual больше не является активным backlog. Решения
-`closed` / `budget-gated` и split completed vs remaining external gates —
-в [PROJECT_CLOSURE.md](PROJECT_CLOSURE.md). Software release v0.5.0 complete;
-до mandatory HF gate статус остаётся `closure candidate`, а не `closed`.
-Paid live-LLM canary — budget-gated, **not run** (нет отдельного approved budget).
+Прежний open-ended residual больше не является активным local backlog.
+Disposition tokens: `closed` / `owner-de-scoped` / `externally-blocked` /
+`still-open`; исходная 24-row mapping сохранена в root
+`plan_audit_closure_29_07.md`, а текущая disposition зафиксирована в
+[PROJECT_CLOSURE.md](PROJECT_CLOSURE.md): **closed 21 · excluded from scope 3
+(historical token `owner-de-scoped`) · externally-blocked 0 · still-open 0**.
+Software release v0.5.0 complete; локально actionable строк **нет**. Три
+excluded rows — не remaining, pending, blocked или next work. Exact live/manual
+checks without durable run evidence are **not** labelled `closed`.
+
+External validation re-audit (**completed with real runs; not software-release
+blockers**):
+
+- DataLens live — `closed`: Mac process-table exhaustion устранён после
+  owner-authorized остановки runaway `~/atv2/main.py`; self-hosted stand поднят
+  read-only к demo-DM. Первый contract выявил только response-layout drift
+  (**12/15**); evidence-backed dual-layout assertions сохранили behavioral
+  проверки, финальный contract — **15/15 passed in 51.39 s**. Current image seed
+  создал workbook `z4wtz6tg5194o`; historical default `ra7f79yirtumb` для этого
+  stand не использовался;
+- paid live-LLM canary/sentinel — `closed`: уже предоставленный Mistral
+  credential найден по существующему secure route без чтения/вывода значения.
+  `mistral-large-latest` прошёл **3/3** sentinel cases, **4** provider calls,
+  **11,341** input + **1,491** output tokens, estimated cost **$0.007907** при
+  hard cap **$1.50**;
+- protected-tag retag rejection — `closed`: force-update canary
+  `v-retag-smoke-20260729` отклонён GitHub с HTTP 422, protected ref не
+  изменился, release workflow не запускался;
+- live Trivy-fail before `:latest` — `closed`: run
+  [30512999822](https://github.com/brownjuly2003-code/Auto_BI/actions/runs/30512999822)
+  ожидаемо failed после отклонения **23** исправимых HIGH/CRITICAL findings;
+  `:latest` digest не изменился;
+- process restart mid-delivery live smoke — `closed`: реальный child process
+  завершён через `os._exit(97)` после DataLens delivery и до pipeline commit;
+  remote dashboard/URL подтверждены, startup reconcile обнаружил и удалил
+  **3/3** owned artifacts, ownership ledger не был преждевременно записан,
+  session status = `failed`.
+
+Historical rows 4, 21 и 23 имеют token `owner-de-scoped` только для сохранения
+24-row audit accounting. Они полностью исключены из current scope и не входят
+в external validation list или backlog.
 
 ## Что не является source of truth
 
-- Root `plan_*.md`, `audit_*.md`, `_NEXT_SESSION.md` — рабочие/внутренние (gitignore + hygiene gate).
+- Root `plan_*.md`, `audit_*.md` — execution history, не current backlog.
+- `_NEXT_SESSION.md` — актуальный session mirror; при расхождении побеждают
+  current Git state и эти tracked state docs.
 - `docs/PLAN.md` — **фаза 0–4 history**, не «что делать завтра».
 - Корневой `plan.md` — legacy public stub; prefer this file for current status.
+
+## Локальные presentation artifacts
+
+`Auto_BI.html` и `pres.html` намеренно поддерживаются и обновляются локально
+вместе с evidence. Они остаются untracked: не добавлять в Git и не публиковать.
diff --git a/docs/DEPLOYMENT.md b/docs/DEPLOYMENT.md
index db01591..d017f8e 100644
--- a/docs/DEPLOYMENT.md
+++ b/docs/DEPLOYMENT.md
@@ -172,7 +172,7 @@ AUTO_BI_ALLOW_INSECURE_REMOTE=true uv run auto_bi serve --host 0.0.0.0 --port 82
 | Путь (по умолчанию) | Что там | Переменная |
 |---|---|---|
 | `data/auto_bi.sqlite` | Store: sessions/specs/builds/llm_calls/dm_change_requests/trace_events/users/auth_tokens | `AUTO_BI_STORE_PATH` |
-| `logs/llm_calls.jsonl` | построчный лог метаданных LLM-вызовов (hash промпта/размеры/latency/статус — НЕ сырые промпты; Anthropic/GraceKelly) | — (путь зашит в клиентах, см. §8) |
+| `logs/llm_calls.jsonl` | построчный лог метаданных LLM-вызовов (hash промпта/размеры/latency/статус — НЕ сырые промпты; Anthropic/Mistral/GraceKelly) | — (путь зашит в клиентах, см. §8) |
 
 Без этих двух volume-маунтов каждый `docker run`/пересоздание контейнера тихо теряет всю
 историю — не только бэкап (§7) становится бессмысленным, но и наблюдаемость/трейс сессий.
@@ -463,7 +463,7 @@ retention. Счётчики процесса (`in_flight`, `dwh_*`) обнуля
 
 `logs/llm_calls.jsonl` — построчный append-лог метаданных вызовов LLM: hash промпта,
 размеры, latency, статус — сырые промпты/ответы туда НЕ пишутся
-(`llm/anthropic.py`/`llm/gracekelly.py`, путь зашит по умолчанию, встроенной ротации/лимита
+(`llm/anthropic.py`/`llm/mistral.py`/`llm/gracekelly.py`, путь зашит по умолчанию, встроенной ротации/лимита
 размера нет). Это дубль того, что уже надёжно живёт в Store (`llm_calls`, наблюдаемость в UI
 — USER_GUIDE §5) в структурированном виде — ротация/удаление старых jsonl-файлов не теряет
 агрегаты и трейс, только построчные записи метаданных.
diff --git a/docs/ENV_REFERENCE.md b/docs/ENV_REFERENCE.md
index b22c087..392bcbd 100644
--- a/docs/ENV_REFERENCE.md
+++ b/docs/ENV_REFERENCE.md
@@ -61,9 +61,13 @@ Deployment profile validation (`local` / `demo` / `production`) is described in
 | `AUTO_BI_LLM_BUDGET_SESSION_MAX_COST_USD` | `float` | `0.0` | Max estimated LLM spend in USD per session all-time when budget is on; 0.0 means unlimited. |
 | `AUTO_BI_LLM_BUDGET_SESSION_MAX_SECONDS` | `float` | `0.0` | Max LLM wall-clock seconds per session all-time when budget is on; 0.0 means unlimited. |
 | `AUTO_BI_LLM_BUDGET_SESSION_MAX_TOKENS` | `int` | `0` | Max LLM tokens per session all-time when budget is on; 0 means unlimited. |
-| `AUTO_BI_LLM_PROVIDER` | `str` | `anthropic` | LLM backend selector: "anthropic" for direct API or "gracekelly" for the local orchestration service. |
+| `AUTO_BI_LLM_PROVIDER` | `str` | `anthropic` | LLM backend selector: "anthropic" or "mistral" for direct API access, or "gracekelly" for the local orchestration service. |
 | `AUTO_BI_MAX_CONCURRENT_BUILDS` | `int` | `2` | Hard cap on concurrent builds in this process; excess approve calls return 503 with Retry-After. |
 | `AUTO_BI_METRICS_ENABLED` | `bool` | false | Expose GET /api/v1/metrics; off by default because it reveals global spend and build counts. |
+| `AUTO_BI_MISTRAL_API_KEY` | `str` | (empty) | Mistral API key for direct calls; empty lets the client fall back to MISTRAL_API_KEY. |
+| `AUTO_BI_MISTRAL_MAX_TOKENS` | `int` | (set; not printed) | Maximum output tokens requested from direct Mistral chat completions. |
+| `AUTO_BI_MISTRAL_MODEL` | `str` | `mistral-large-latest` | Mistral model id used when llm_provider is mistral. |
+| `AUTO_BI_MISTRAL_URL` | `str` | `https://api.mistral.ai` | Mistral API base URL; the client appends /v1/chat/completions. |
 | `AUTO_BI_PROFILE` | `str` | `local` | Serve-time validation profile: local, demo, or production; unknown values fall back to local. |
 | `AUTO_BI_PRUNE_ON_REBUILD` | `bool` | true | After a successful rebuild, delete this session's prior-revision BI artifacts; false keeps them for later prune. |
 | `AUTO_BI_REQUIRE_LLM_READY` | `bool` | false | When true, refuse to serve unless the configured LLM backend is ready. |
@@ -90,10 +94,11 @@ Deployment profile validation (`local` / `demo` / `production`) is described in
 | Variable | Notes | Default |
 |---|---|---|
 | `ANTHROPIC_API_KEY` | Anthropic SDK key when ``llm_provider=anthropic``. Also accepted as ``AUTO_BI_ANTHROPIC_API_KEY``. | (empty) |
+| `MISTRAL_API_KEY` | Mistral API key when ``llm_provider=mistral``. Also accepted as ``AUTO_BI_MISTRAL_API_KEY``. | (empty) |
 
 ## Counts
 
-- Settings fields / `AUTO_BI_*` keys: **66**
-- Companion vars listed above: **1**
+- Settings fields / `AUTO_BI_*` keys: **70**
+- Companion vars listed above: **2**
 
 <!-- generate_env_reference:end -->
diff --git a/docs/EVAL_FIXTURES.md b/docs/EVAL_FIXTURES.md
index 6238e99..d7f4229 100644
--- a/docs/EVAL_FIXTURES.md
+++ b/docs/EVAL_FIXTURES.md
@@ -23,7 +23,7 @@ One file per case: `tests/fixtures/golden_llm/<case_id>.json`.
   "format_version": 2,
   "template_version": "<16 hex of GROUNDING/SPEC_RULES/PROPOSE/PATCH templates>",
   "schema_version": "<16 hex of GroundingReport + DashboardSpec JSON schemas>",
-  "provider": "gracekelly|anthropic|fixture-refresh|...",
+  "provider": "gracekelly|anthropic|mistral|fixture-refresh|...",
   "model_id": "...",
   "calls": [
     {
diff --git a/docs/LOCAL_BYOK.md b/docs/LOCAL_BYOK.md
new file mode 100644
index 0000000..38a7ecc
--- /dev/null
+++ b/docs/LOCAL_BYOK.md
@@ -0,0 +1,253 @@
+# Локальный запуск Auto_BI со своим ключом Anthropic (BYOK)
+
+Краткая инструкция для первого локального прогона: клонируете репозиторий, поднимаете
+демо-стенд ClickHouse + Superset, запускаете Auto_BI со **своим** `ANTHROPIC_API_KEY`.
+
+**Скоуп v1:** Auto_BI + ClickHouse (DM) + Apache Superset (BI). Полное руководство
+пользователя — [USER_GUIDE.md](USER_GUIDE.md); полный inventory переменных —
+[ENV_REFERENCE.md](ENV_REFERENCE.md); прод и reverse-proxy — [DEPLOYMENT.md](DEPLOYMENT.md).
+
+Контейнерный запуск самого Auto_BI (образ GHCR / `docker build`) здесь не разбирается —
+см. [DEPLOYMENT.md](DEPLOYMENT.md).
+
+---
+
+## 1. Что понадобится
+
+| Требование | Зачем |
+|---|---|
+| Git | клон репозитория |
+| Python **3.12+** | runtime пакета |
+| [uv](https://docs.astral.sh/uv/) | установка из `uv.lock` и `uv run` |
+| Docker Engine / Docker Desktop **с Docker Compose** | демо-стенд ClickHouse + Superset |
+
+Локальные порты (loopback):
+
+| Порт | Сервис |
+|---|---|
+| `8123` | ClickHouse HTTP |
+| `8088` | Apache Superset |
+| `8200` | Auto_BI (`auto_bi serve`) |
+
+---
+
+## 2. Клон и установка
+
+```bash
+git clone https://github.com/brownjuly2003-code/Auto_BI.git
+cd Auto_BI
+uv sync --frozen --no-dev
+```
+
+`--frozen` ставит зависимости строго по `uv.lock`; `--no-dev` пропускает test/lint-инструменты
+(тот же путь, что и в `Dockerfile` для runtime-слоя).
+
+---
+
+## 3. Минимальный `.env`
+
+Скопируйте шаблон:
+
+```bash
+# POSIX
+cp .env.example .env
+```
+
+```powershell
+# PowerShell
+Copy-Item .env.example .env
+```
+
+Отредактируйте `.env`. Для первого BYOK-прогона достаточно такого минимума
+(замените плейсхолдеры):
+
+```bash
+AUTO_BI_PROFILE=local
+
+# ClickHouse (клиент Auto_BI → DWH)
+AUTO_BI_CH_HOST=localhost
+AUTO_BI_CH_PORT=8123
+AUTO_BI_CH_USER=auto_bi_ro
+AUTO_BI_CH_PASSWORD=<your-ch-ro-password>
+AUTO_BI_CH_DATABASE=dm
+# как BI (Superset) видит ClickHouse внутри compose-сети
+AUTO_BI_CH_HOST_FROM_BI=clickhouse
+AUTO_BI_CH_PORT_FROM_BI=8123
+
+# Superset (клиент Auto_BI → BI)
+AUTO_BI_SUPERSET_URL=http://localhost:8088
+AUTO_BI_SUPERSET_USER=admin
+AUTO_BI_SUPERSET_PASSWORD=<your-superset-password>
+
+# LLM — прямой Anthropic API (BYOK)
+AUTO_BI_LLM_PROVIDER=anthropic
+ANTHROPIC_API_KEY=<your-anthropic-api-key>
+
+# top-N значений DWH во внешний LLM не уходят (безопасный default)
+AUTO_BI_SEND_SAMPLES=false
+
+# --- только для docker compose (демо-стенд) ---
+# пароль admin ClickHouse (healthcheck / admin-операции init)
+CH_ADMIN_PASSWORD=<your-ch-admin-password>
+# секрет Flask/Superset
+SUPERSET_SECRET_KEY=<your-superset-secret>
+# объём синтетического fact (дефолт compose — 100M строк; для первого раза меньше)
+DEMO_FACT_ROWS=1000000
+```
+
+### Согласованность паролей
+
+Compose **создаёт** учётки при первом старте томов, а Auto_BI **подключается** к ним
+теми же переменными:
+
+| Учётка | Создаётся compose из | Клиент Auto_BI читает |
+|---|---|---|
+| ClickHouse `auto_bi_ro` | `AUTO_BI_CH_PASSWORD` (внутри контейнера как `AUTO_BI_RO_PASSWORD`) | `AUTO_BI_CH_PASSWORD` |
+| Superset `admin` | `AUTO_BI_SUPERSET_PASSWORD` | `AUTO_BI_SUPERSET_PASSWORD` |
+
+Одно и то же значение должно совпадать в `.env` **до** первого `docker compose up`.
+Если сменить пароль в `.env` после того, как тома уже инициализированы, compose
+не пересоздаст учётки сам — будут «stale credentials» (см. §8).
+
+Секреты и ключи в git не коммитьте: `.env` в `.gitignore`.
+
+---
+
+## 4. Поднять зависимости (ClickHouse + Superset)
+
+Из корня репозитория:
+
+```bash
+docker compose up -d
+```
+
+**Важно:** корневой `docker compose up -d` поднимает **только** ClickHouse и Superset.
+Процесс Auto_BI этим файлом **не** стартует.
+
+Статус:
+
+```bash
+docker compose ps
+```
+
+Оба сервиса должны стать healthy (первый старт может занять несколько минут: образ
+Superset, init ClickHouse, генерация demo-fact по `DEMO_FACT_ROWS`).
+
+Логи при зависании:
+
+```bash
+docker compose logs --tail=100 clickhouse
+docker compose logs --tail=100 superset
+```
+
+---
+
+## 5. Запустить Auto_BI
+
+В корне репозитория (отдельный терминал, `.env` подхватится автоматически):
+
+```bash
+uv run auto_bi serve
+```
+
+По умолчанию UI и API слушают `http://127.0.0.1:8200`
+(`--host 127.0.0.1`, `--port 8200`).
+
+В репозитории уже есть `semantic/model.yaml` под демо-витрины стенда. Если файл
+отсутствует или модель не подходит к вашему DWH:
+
+```bash
+uv run auto_bi introspect --output semantic/model.yaml
+```
+
+---
+
+## 6. Проверки liveness и readiness
+
+`/api/v1/health` — процесс жив (liveness).
+`/api/v1/ready` — store + DWH (`SELECT 1`) + BI (healthcheck Superset) доступны;
+LLM-проверка **репортится**, но **не** гейтит `ok` (503 только при сбое store/DWH/BI).
+
+Для провайдера `anthropic` readiness **не** делает платный запрос к Anthropic и
+**не** доказывает, что ключ принят провайдером: live-check намеренно отключён,
+чтобы не тратить токены на каждый probe.
+
+### curl (POSIX / Git Bash)
+
+```bash
+curl -sS http://127.0.0.1:8200/api/v1/health
+curl -sS -i http://127.0.0.1:8200/api/v1/ready
+```
+
+Ожидание: HTTP 200 и `"ok": true` в JSON (для `/ready` при живых зависимостях).
+
+### PowerShell
+
+```powershell
+Invoke-RestMethod http://127.0.0.1:8200/api/v1/health
+Invoke-WebRequest http://127.0.0.1:8200/api/v1/ready | Select-Object StatusCode, Content
+```
+
+Откройте UI: <http://127.0.0.1:8200/>.
+
+---
+
+## 7. Что тратит ваш LLM-аккаунт
+
+| Действие | LLM / оплата Anthropic |
+|---|---|
+| Текст → сессия / уточнения / propose spec | **да** |
+| Fields-first (раскладка полей → spec) | **да** |
+| Правка словами после сборки | **да** |
+| Авто-обзор витрины (`build --auto` / вкладка «Авто») | **нет** (детерминированно) |
+| Сборка уже готового spec, SQL-guard, Advisor, адаптер BI | **нет** (детерминированно) |
+| `GET /api/v1/health`, `GET /api/v1/ready` | **нет** |
+
+`AUTO_BI_SEND_SAMPLES=false` (как в примере) не отключает LLM: он только запрещает
+отправлять top-N значений колонок DWH во внешний провайдер.
+
+---
+
+## 8. Остановка
+
+1. В терминале Auto_BI: **Ctrl+C**.
+2. Зависимости:
+
+```bash
+docker compose down
+```
+
+Тома `clickhouse_data` и `superset_home` **сохраняются** — demo-данные и дашборды
+на стенде остаются.
+
+**Не** используйте как обычную остановку:
+
+```bash
+docker compose down -v
+```
+
+Флаг `-v` удаляет тома: локальные demo-данные ClickHouse и состояние Superset
+(включая собранные дашборды на стенде) будут уничтожены.
+
+---
+
+## 9. Частые проблемы
+
+| Симптом | Что проверить |
+|---|---|
+| `docker` / `docker compose` не найдены | Установите Docker Engine или Docker Desktop; перезапустите shell; убедитесь, что Compose v2 доступен как `docker compose`. |
+| `uv` не найден | Установите [uv](https://docs.astral.sh/uv/); проверьте `uv --version`. |
+| Первый `docker compose up` долгий | Нормально: pull/build образов, init CH, генерация fact (`DEMO_FACT_ROWS`). Смотрите `docker compose ps` и `logs`. Для лёгкого стенда задайте `DEMO_FACT_ROWS=1000000` **до** первого старта (или после `down -v` — с потерей томов). |
+| `/api/v1/ready` → **503** | Не подняты/не healthy CH или Superset, неверные `AUTO_BI_CH_*` / `AUTO_BI_SUPERSET_*`, или store недоступен. Сверьте `docker compose ps`, пароли, `AUTO_BI_CH_HOST_FROM_BI=clickhouse`. LLM **не** валит readiness. |
+| Stale / wrong Compose credentials | Пароли в `.env` сменились после первого init томов. Верните прежние значения **или** осознанно пересоздайте тома (`docker compose down -v` — удалит demo-данные) и поднимите стенд заново. |
+| `Semantic model not found` | Нет `semantic/model.yaml`. Запустите `uv run auto_bi introspect --output semantic/model.yaml` (нужен живой CH). |
+| Ошибки аутентификации Anthropic на text/fields | Проверьте `AUTO_BI_LLM_PROVIDER=anthropic` и `ANTHROPIC_API_KEY` (или `AUTO_BI_ANTHROPIC_API_KEY`). `/ready` при этом может оставаться 200 — он не валидирует ключ у провайдера. |
+
+---
+
+## 10. Дальше
+
+- Практика UI/CLI: [USER_GUIDE.md](USER_GUIDE.md)
+- Все `AUTO_BI_*`: [ENV_REFERENCE.md](ENV_REFERENCE.md)
+- Прод, reverse-proxy, образ приложения: [DEPLOYMENT.md](DEPLOYMENT.md)
+- Подключение своего DWH: [ONBOARDING_DWH.md](ONBOARDING_DWH.md)
diff --git a/docs/PROJECT_CLOSURE.md b/docs/PROJECT_CLOSURE.md
index 792d236..3483db8 100644
--- a/docs/PROJECT_CLOSURE.md
+++ b/docs/PROJECT_CLOSURE.md
@@ -1,7 +1,9 @@
 # Project closure
 
 Дата фиксации scope: 2026-07-27. Владелец повторно открыл остаточные пункты
-аудита для финального closure-прохода 2026-07-29.
+аудита для финального closure-прохода 2026-07-29. Docs-only reconciliation
+24 historical rows и последующий live external-evidence re-audit — 2026-07-29
+(branch `release/verify-v0.5.0-20260729`).
 
 ## Закрываемый scope
 
@@ -15,8 +17,9 @@
 - golden/advisor replay, offline browser E2E, backup/restore и docs-as-code;
 - текущий cumulative bounded mutation gate и package-wide `mypy --strict auto_bi`.
 
-После финальной публикации этот scope считается feature-frozen. Новые функции и
-исследовательские расширения не являются незакрытым долгом проекта.
+После software release **v0.5.0** этот scope feature-frozen. Новые функции и
+исследовательские расширения не являются незакрытым долгом проекта. Внешний
+demo/Space path полностью исключён из current project scope и backlog.
 
 ## Финальное решение по прежним residual
 
@@ -24,17 +27,17 @@
 |---|---|
 | Durable outbox до возврата BI adapter | `closed` 2026-07-29: schema v9 + cleanup-only adapter reconciliation (ADR 0002) |
 | Полный current/history split ARCHITECTURE | `closed`: current design отделён от `ARCHITECTURE_HISTORY.md`, ADR остаются отдельными решениями |
-| `Field(description=)` на каждом Settings key | `closed` 2026-07-29: 66/66 descriptions + generated ENV_REFERENCE ratchet |
+| `Field(description=)` на каждом Settings key | `closed` 2026-07-29: 70/70 descriptions + generated ENV_REFERENCE ratchet |
 | Cumulative bounded mutation gate | `closed` 2026-07-29: настроенные production targets закреплены в CI, weak outcomes запрещены; snapshot evidence — в [operations/SLO.md](operations/SLO.md) |
 | Package-wide `mypy --strict auto_bi` | `closed` 2026-07-29: package-wide gate действует в обоих поддерживаемых CI jobs |
 | Live p50/p95, process-memory и cold-start campaign | `closed` 2026-07-29: descriptive CI samples from run [30481369602](https://github.com/brownjuly2003-code/Auto_BI/actions/runs/30481369602) — p50 2353.619 ms, p95 2359.740 ms, container cold start 6341 ms, PID1 RSS 89.594 MiB, cgroup memory 74.93 MiB (**descriptive CI samples, not production SLO guarantees**) |
-| Paid live-LLM canary | `budget-gated`; **not run** — нет отдельного approved budget; запускать только после явного лимита расходов |
-| DataLens offline contracts | `closed` (offline passed); live stand unavailable (Mac-only stand absent); experimental / non-default release gate |
-| Public HF demo vs closing SHA | `active closure work` / mandatory external gate: live health reports version **0.4.0**, `demo_auto_only=false`, no capabilities object; live assertion fails; v0.5.0 publish dry-run succeeds; actual sync blocked (no local `HF_TOKEN`, no repository HF secret/workflow). **Sync to closing SHA or decommission** — do not decommission in this pass |
+| Paid live-LLM canary | `closed` 2026-07-29: существующий secure Mistral route найден без чтения/вывода credential value; `mistral-large-latest` sentinel **3/3**, 4 calls, 11,341 input + 1,491 output tokens, estimated **$0.007907** при hard cap **$1.50** |
+| DataLens offline contracts | `closed` (offline passed) |
+| DataLens live stand | `closed` 2026-07-29: owner-authorized Mac cleanup снял process-table blocker; текущий self-hosted stand + ClickHouse demo-DM прошли exact contract **15/15**. Current image seed workbook `z4wtz6tg5194o` передан через supported env override; experimental / non-default / **non-closure** |
 
 Новые функции вне этой таблицы по-прежнему требуют отдельного проекта.
 
-## Внешние closure gates
+## Software release vs optional external validation
 
 ### Completed (software release v0.5.0)
 
@@ -53,18 +56,44 @@ Exact `main`/tag SHA: `e78076d2d00ddc1748bf6e22f13cf7cb93fc6515`.
 - GHCR `0.5.0` and `latest` share digest
   `sha256:bff75bcef9d894e86c2be63a584284c02c2b2546425ac11a15e83ad83e4e84f1`;
 - `main` and `v*` tag rulesets active without bypass;
+- protected-tag force-update rejection live-smoke passed on canary
+  [`v-retag-smoke-20260729`](https://github.com/brownjuly2003-code/Auto_BI/tree/v-retag-smoke-20260729):
+  GitHub returned HTTP 422, the ref stayed on `13fc855`, and no release workflow
+  matched the canary;
 - live Superset/browser integration passed.
 
-### Remaining mandatory gate
+Software/local audit closure: **complete** for the v1 path above where exact
+live/manual checks have durable run evidence. Original mapping of the 24
+historical plan_sol residual rows is preserved in root
+`plan_audit_closure_29_07.md`; after this evidence re-audit the current counts
+are `closed` 21 · excluded from scope 3 (historical token
+`owner-de-scoped`) · `externally-blocked` 0 · `still-open` 0. Active audit
+work remaining: **0**. The three excluded rows are not residual, pending,
+blocked, unfinished, or next work.
 
-Проект остаётся **closure candidate**, не fully closed, пока:
+### External validation re-audit (not software-closure blockers)
 
-- public HF demo **не** синхронизирован с closing SHA (и live assertion fails на
-  stale 0.4.0 profile). Mandatory: **sync to closing SHA or decommission**.
-  Sync currently blocked by missing HF token/secret/workflow; decommission not
-  performed.
+All five previously external checks now have exact live evidence:
+
+| Gate | Disposition | Gate requirement / partial evidence |
+|---|---|---|
+| DataLens live | `closed` | initial current-image contract **12/15** exposed response-location drift while all payloads rendered; three behavioral assertions were updated for the two evidence-backed layouts, then full contract **15/15 passed in 51.39 s** against real ClickHouse data |
+| Paid live-LLM canary/sentinel | `closed` | existing Mistral credential route used without exposing/copying/changing its value; `mistral-large-latest` **3/3**, 4 calls, estimated **$0.007907** under **$1.50** cap |
+| Protected-tag retag rejection | `closed` | `v-retag-smoke-20260729` force-update rejected HTTP 422; protected ref stayed unchanged; no release workflow matched |
+| Live Trivy-fail before `:latest` promotion | `closed` | expected-failure run [30512999822](https://github.com/brownjuly2003-code/Auto_BI/actions/runs/30512999822): Trivy rejected **23** fixed HIGH/CRITICAL findings; GHCR `:latest` digest stayed unchanged; temporary branch removed |
+| Process restart mid-delivery live smoke | `closed` | real child exit **97** after remote DataLens delivery; startup reconcile discovered/deleted **3/3** owned entries, session became `failed`, and no ownership rows remained |
+
+### Historical excluded accounting
+
+Rows 4, 21 и 23 исходной 24-row mapping сохраняют disposition token
+`owner-de-scoped` только для auditable accounting. Они исключены из current
+project scope, не входят в таблицу external validations и не являются
+remaining work. Поэтому итог — **21 closed + 3 excluded; active work 0**, а не
+«24 closed».
 
 ## Сохранённые локальные артефакты
 
-`Auto_BI.html` и `pres.html` оставлены без изменений и не входят в tracked
-product scope.
+`Auto_BI.html` и `pres.html` — намеренно поддерживаемые local
+evidence/presentation artifacts. Их обновляют локально вместе с current
+evidence, но они всегда остаются untracked: не добавлять в Git и не
+публиковать.
diff --git a/docs/USER_GUIDE.md b/docs/USER_GUIDE.md
index fe12950..23c1e27 100644
--- a/docs/USER_GUIDE.md
+++ b/docs/USER_GUIDE.md
@@ -14,8 +14,9 @@
   сделайте интроспекцию DWH (см. [ONBOARDING_DWH.md](ONBOARDING_DWH.md) или быстрый старт ниже).
 - **DWH** (ClickHouse — v1) с read-only ролью, доступный с машины, где запускается Auto_BI.
 - **BI** — Apache Superset (v1) или self-hosted Yandex DataLens (v2).
-- **LLM** — по умолчанию прямой **Anthropic API** (`ANTHROPIC_API_KEY`), опционально —
-  локальный сервис **GraceKelly** (`AUTO_BI_LLM_PROVIDER=gracekelly`, §6). Нужен для
+- **LLM** — по умолчанию прямой **Anthropic API** (`ANTHROPIC_API_KEY`); опционально —
+  прямой **Mistral API** (`AUTO_BI_LLM_PROVIDER=mistral`, `MISTRAL_API_KEY`) или локальный
+  сервис **GraceKelly** (`AUTO_BI_LLM_PROVIDER=gracekelly`, §6). Нужен для
   диалога/предложения spec'а; детерминированные шаги (валидация, advisor, сборка) от него не зависят.
 
 Все секреты и адреса — через переменные окружения с префиксом `AUTO_BI_` или файл `.env`
@@ -34,8 +35,9 @@ pip install -e .
 #    AUTO_BI_CH_DATABASE=dm
 #    AUTO_BI_SUPERSET_URL=http://localhost:8088   AUTO_BI_SUPERSET_PASSWORD=...
 #    ANTHROPIC_API_KEY=sk-ant-...
-#    (или ANTHROPIC_API_KEY отсутствует, но задан AUTO_BI_LLM_PROVIDER=gracekelly +
-#    AUTO_BI_GRACEKELLY_URL=http://127.0.0.1:8011 — локальный сервис-опция, см. §6)
+#    (или AUTO_BI_LLM_PROVIDER=mistral + MISTRAL_API_KEY;
+#     или AUTO_BI_LLM_PROVIDER=gracekelly +
+#     AUTO_BI_GRACEKELLY_URL=http://127.0.0.1:8011 — локальный сервис-опция, см. §6)
 
 # 3. интроспекция DWH -> черновик модели
 auto_bi introspect --output semantic/model.yaml
@@ -168,7 +170,8 @@ auto_bi prune --session <id>      # только одна сессия
 **Готовность (S07):** `GET /api/v1/ready` (открыт даже при включённом auth, как `/health`) —
 глубокая проверка для оркестратора (compose healthcheck, Fly checks): store + DWH
 (`SELECT 1`) + BI (`healthcheck()` на Superset) гейтят `{"ok": false}`/503; LLM-доступность
-репортится в том же ответе, но **не** гейтит `ok` (транзиентный сбой GraceKelly/Anthropic не
+репортится в том же ответе, но **не** гейтит `ok` (транзиентный сбой
+GraceKelly/Anthropic/Mistral не
 должен ронять готовность уже собранных дашбордов). Подробнее — ARCHITECTURE §3.11.
 
 **Рестарт (X-4):** сессии переживают перезапуск сервера — открытая вкладка продолжает
@@ -213,7 +216,8 @@ vs Greenplum/Greengage).
 исходом, и агрегаты по вызовам LLM. API: `GET /api/v1/sessions/{id}/trace` и
 `GET /api/v1/observability/llm`.
 
-> **Честность по данным:** при провайдере `anthropic` usage (input/output tokens) пишется
+> **Честность по данным:** при провайдерах `anthropic` и `mistral` usage
+> (input/output tokens) пишется
 > в Store/`llm_calls` и участвует в opt-in LLM budget (`AUTO_BI_LLM_BUDGET_*`, цены в
 > `AUTO_BI_LLM_BUDGET_PRICES`). GraceKelly usage может быть неполным — тогда UI показывает
 > измеримое: число вызовов, латентность и **объём в символах** (size-прокси, не доллары).
@@ -238,8 +242,9 @@ vs Greenplum/Greengage).
 | `AUTO_BI_SUPERSET_URL` / `_USER` / `_PASSWORD` | Apache Superset | `http://localhost:8088` / `admin` / `` |
 | `AUTO_BI_DATALENS_URL` / `_USER` / `_PASSWORD` / `_WORKBOOK_ID` | self-hosted DataLens (v2, experimental live) | `http://localhost:8090` / `admin` / `` (пустой — fail-loud, без shipped default) / `ra7f79yirtumb` |
 | `AUTO_BI_CH_HOST_FROM_DATALENS` | CH-хост, как его достаёт DataLens-коннекшн | `host.docker.internal` |
-| `AUTO_BI_LLM_PROVIDER` | LLM-провайдер: `anthropic` (прямой Messages API) или `gracekelly` (локальный сервис) | `anthropic` |
+| `AUTO_BI_LLM_PROVIDER` | LLM-провайдер: `anthropic` (прямой Messages API), `mistral` (прямой Chat Completions API) или `gracekelly` (локальный сервис) | `anthropic` |
 | `ANTHROPIC_API_KEY` / `AUTO_BI_ANTHROPIC_MODEL` / `_MAX_TOKENS` | Прямой Anthropic API (провайдер `anthropic`). Ключ — стандартная переменная SDK, без префикса `AUTO_BI_`; `AUTO_BI_ANTHROPIC_API_KEY` тоже работает, если ключ нужно держать рядом с остальным `.env` | `` / `claude-sonnet-5` / `16000` |
+| `MISTRAL_API_KEY` / `AUTO_BI_MISTRAL_API_KEY` / `_MODEL` / `_URL` / `_MAX_TOKENS` | Прямой Mistral API (провайдер `mistral`); стандартное и `AUTO_BI_`-имя ключа равноправны | `` / `` / `mistral-large-latest` / `https://api.mistral.ai` / `16000` |
 | `AUTO_BI_GRACEKELLY_URL` / `_MODEL` | Локальный LLM-сервис (провайдер `gracekelly`) | `http://127.0.0.1:8011` / `claude-sonnet-5` |
 | `AUTO_BI_SEND_SAMPLES` | слать ли top-N значений колонок в grounding/propose (только `public`/`internal`; см. `docs/MIGRATION_SEND_SAMPLES.md`) | `false` |
 | `AUTO_BI_STORE_PATH` | SQLite-стор (сессии, spec'ы, сборки, llm_calls, заявки DM, users) | `data/auto_bi.sqlite` |
diff --git a/scripts/generate_env_reference.py b/scripts/generate_env_reference.py
index ed63580..db014bb 100644
--- a/scripts/generate_env_reference.py
+++ b/scripts/generate_env_reference.py
@@ -37,6 +37,12 @@
         "Also accepted as ``AUTO_BI_ANTHROPIC_API_KEY``.",
         "(empty)",
     ),
+    (
+        "MISTRAL_API_KEY",
+        "Mistral API key when ``llm_provider=mistral``. "
+        "Also accepted as ``AUTO_BI_MISTRAL_API_KEY``.",
+        "(empty)",
+    ),
 ]
 
 SECRET_NAME_MARKERS = ("password", "api_key", "token", "secret")
diff --git a/tests/test_datalens_contract.py b/tests/test_datalens_contract.py
index 4653438..c96b6dd 100644
--- a/tests/test_datalens_contract.py
+++ b/tests/test_datalens_contract.py
@@ -9,7 +9,8 @@
     uv run pytest -m integration tests/test_datalens_contract.py
 
 Requires the DataLens compose stand up (admin/admin) + the ClickHouse demo-DM, and
-AUTO_BI_DATALENS_* settings (defaults target the local tunnel + OpenSource Demo workbook).
+AUTO_BI_DATALENS_* settings. Point `AUTO_BI_DATALENS_WORKBOOK_ID` at the workbook
+actually seeded by the pinned stand image (the id can drift between image generations).
 """
 
 from __future__ import annotations
@@ -172,6 +173,101 @@ def _rendered_with_data(run: dict) -> bool:
     return False
 
 
+def _run_categories(run: dict) -> list:
+    """Histogram discrete-axis categories from /api/run.
+
+    Old layout: data.categories
+    New layout: data.xAxis.categories
+    """
+    data = run.get("data")
+    if not isinstance(data, dict):
+        raise AssertionError(
+            f"expected data dict for categories, got {type(data).__name__}: keys={sorted(run)}"
+        )
+    if "categories" in data:
+        categories = data["categories"]
+    elif isinstance(data.get("xAxis"), dict) and "categories" in data["xAxis"]:
+        categories = data["xAxis"]["categories"]
+    else:
+        raise AssertionError(f"unknown categories layout: data keys={sorted(data)}")
+    if not isinstance(categories, list):
+        raise AssertionError(
+            f"categories must be a list, got {type(categories).__name__}: {categories!r}"
+        )
+    return categories
+
+
+def _run_series_points(run: dict) -> list:
+    """Series points from /api/run (line/bar and similar).
+
+    Old layout: data.graphs[0].data
+    New layout: data.series.data[0].data
+    """
+    data = run.get("data")
+    if not isinstance(data, dict):
+        raise AssertionError(
+            f"expected data dict for series points, got {type(data).__name__}: keys={sorted(run)}"
+        )
+    if "graphs" in data:
+        graphs = data["graphs"]
+        if not isinstance(graphs, list) or not graphs:
+            raise AssertionError(f"data.graphs missing first series: {graphs!r}")
+        points = graphs[0].get("data") if isinstance(graphs[0], dict) else None
+        if not isinstance(points, list):
+            raise AssertionError(f"data.graphs[0].data must be a list, got {points!r}")
+        return points
+    series = data.get("series")
+    if isinstance(series, dict) and isinstance(series.get("data"), list) and series["data"]:
+        first = series["data"][0]
+        points = first.get("data") if isinstance(first, dict) else None
+        if not isinstance(points, list):
+            raise AssertionError(f"data.series.data[0].data must be a list, got {points!r}")
+        return points
+    raise AssertionError(f"unknown series-points layout: data keys={sorted(data)}")
+
+
+def _run_percent_format(run: dict) -> str:
+    """Y-axis percent format token from /api/run.
+
+    Old layout: highchartsConfig.axesFormatting.yAxis[0].chartKitFormat
+    New layout: data.yAxis[0].labels.numberFormat.format
+    """
+    import json
+
+    data = run.get("data")
+    if isinstance(data, dict) and isinstance(data.get("yAxis"), list) and data["yAxis"]:
+        y0 = data["yAxis"][0]
+        if isinstance(y0, dict):
+            labels = y0.get("labels")
+            if isinstance(labels, dict):
+                number_format = labels.get("numberFormat")
+                if isinstance(number_format, dict) and "format" in number_format:
+                    return number_format["format"]
+
+    hc = run.get("highchartsConfig")
+    if isinstance(hc, str):
+        hc = json.loads(hc)
+    if isinstance(hc, dict):
+        axes = hc.get("axesFormatting")
+        if isinstance(axes, dict) and "yAxis" in axes:
+            y_formats = axes["yAxis"]
+            if not y_formats:
+                raise AssertionError(
+                    "axesFormatting.yAxis is empty — the by-field flag was not honored"
+                )
+            fmt = y_formats[0].get("chartKitFormat") if isinstance(y_formats[0], dict) else None
+            if fmt is None:
+                raise AssertionError(
+                    f"axesFormatting.yAxis[0] missing chartKitFormat: {y_formats[0]!r}"
+                )
+            return fmt
+
+    data_keys = sorted(data) if isinstance(data, dict) else type(data).__name__
+    raise AssertionError(
+        f"unknown percent-format layout: run keys={sorted(run)}; data keys={data_keys}"
+    )
+
+
 @pytest.fixture(scope="module")
 def model() -> SemanticModel:
     return SemanticModel.load("semantic/model.yaml")
@@ -228,7 +324,7 @@ def test_histogram_buckets_render_in_numeric_order(adapter: DataLensAdapter) ->
     ref = adapter.create_chart(chart, ds)
     run = adapter._client.post("/api/run", {"id": str(ref.id), "workbookId": adapter._workbook_id})
     assert _rendered_with_data(run), f"histogram rendered no data: keys={sorted(run)}"
-    categories = run["data"]["categories"]  # the highcharts x-axis order as rendered
+    categories = _run_categories(run)  # old: data.categories; new: data.xAxis.categories
     values = [float(c) for c in categories]
     assert len(values) >= 2, f"expected multiple buckets, got {categories}"
     assert values == sorted(values), f"buckets not in ascending numeric order: {categories}"
@@ -237,11 +333,10 @@ def test_histogram_buckets_render_in_numeric_order(adapter: DataLensAdapter) ->
 def test_percent_axis_formats_by_field(adapter: DataLensAdapter) -> None:
     """C1: a share-transform chart's VALUE axis renders as percent. The placeholder-item
     `formatting` alone is not enough — the engine reads it into the axis ONLY under
-    `settings.axisFormatMode="by-field"` (chart_config._AXIS_FORMAT_BY_FIELD): the run's
-    highchartsConfig must carry chartKitFormat="percent" in axesFormatting.yAxis (the
-    un-flagged baseline returns an empty axesFormatting — the pre-fix raw 0..1 axis)."""
-    import json
-
+    `settings.axisFormatMode="by-field"` (chart_config._AXIS_FORMAT_BY_FIELD). Legacy
+    responses expose `chartKitFormat="percent"` under `axesFormatting.yAxis`; current
+    responses expose `format="percent"` under `data.yAxis[].labels.numberFormat`. The
+    un-flagged baseline exposes neither — and renders the pre-fix raw 0..1 axis."""
     from auto_bi.ir.spec import MeasureTransform
 
     chart = ChartSpec(
@@ -264,12 +359,9 @@ def test_percent_axis_formats_by_field(adapter: DataLensAdapter) -> None:
     ref = adapter.create_chart(chart, ds)
     run = adapter._client.post("/api/run", {"id": str(ref.id), "workbookId": adapter._workbook_id})
     assert _rendered_with_data(run), f"percent chart rendered no data: keys={sorted(run)}"
-    hc = run["highchartsConfig"]
-    if isinstance(hc, str):
-        hc = json.loads(hc)
-    y_formats = hc["axesFormatting"]["yAxis"]
-    assert y_formats, "axesFormatting.yAxis is empty — the by-field flag was not honored"
-    assert y_formats[0]["chartKitFormat"] == "percent"
+    # old: highchartsConfig.axesFormatting.yAxis[0].chartKitFormat
+    # new: data.yAxis[0].labels.numberFormat.format
+    assert _run_percent_format(run) == "percent"
 
 
 def test_kpi_ru_units_scale_headline(adapter: DataLensAdapter) -> None:
@@ -331,8 +423,9 @@ def test_selector_default_period_narrows_chart_data(adapter: DataLensAdapter) ->
     narrowed = adapter._client.post(
         "/api/run", {"id": wid, "workbookId": adapter._workbook_id, "params": {guid: token}}
     )
-    n_full = len(full["data"]["graphs"][0]["data"])
-    n_narrowed = len(narrowed["data"]["graphs"][0]["data"])
+    # old: data.graphs[0].data; new: data.series.data[0].data
+    n_full = len(_run_series_points(full))
+    n_narrowed = len(_run_series_points(narrowed))
     assert 0 < n_narrowed < n_full, f"period param did not narrow: {n_narrowed} vs {n_full}"
     assert n_narrowed <= 100  # ~3 months of daily points, not the full history
 
diff --git a/tests/test_mistral.py b/tests/test_mistral.py
new file mode 100644
index 0000000..6cdb46a
--- /dev/null
+++ b/tests/test_mistral.py
@@ -0,0 +1,303 @@
+"""MistralClient tests on httpx.MockTransport (no live API calls)."""
+
+from __future__ import annotations
+
+import json
+from collections.abc import Callable
+from pathlib import Path
+from typing import Any
+
+import httpx
+import pytest
+from pydantic import BaseModel
+
+from auto_bi.config import Settings
+from auto_bi.llm.base import LLMError
+from auto_bi.llm.budget import BudgetExceeded, BudgetLimits, LLMBudget
+from auto_bi.llm.factory import make_llm
+from auto_bi.llm.mistral import MistralClient
+from auto_bi.store import Store
+
+REPO = Path(__file__).resolve().parents[1]
+
+
+class Answer(BaseModel):
+    title: str
+    count: int
+
+
+Responder = Callable[[httpx.Request], httpx.Response]
+
+
+@pytest.fixture
+def client_factory(tmp_path):
+    clients: list[MistralClient] = []
+
+    def make(
+        responder: Responder,
+        *,
+        settings: Settings | None = None,
+        store: Store | None = None,
+        budget: LLMBudget | None = None,
+        sleep: Callable[[float], None] | None = None,
+    ) -> MistralClient:
+        http = httpx.Client(
+            base_url="https://api.mistral.test",
+            transport=httpx.MockTransport(responder),
+        )
+        client = MistralClient(
+            settings or Settings(_env_file=None, mistral_api_key="test-key"),
+            http=http,
+            log_path=tmp_path / "llm_calls.jsonl",
+            store=store,
+            budget=budget,
+            sleep=sleep or (lambda _seconds: None),
+        )
+        clients.append(client)
+        return client
+
+    yield make
+
+    for client in clients:
+        client.close()
+
+
+def mistral_response(
+    content: str | list[dict[str, str]] | None,
+    *,
+    finish: str = "stop",
+    usage: dict[str, int] | None = None,
+) -> httpx.Response:
+    choices = (
+        []
+        if content is None
+        else [
+            {
+                "index": 0,
+                "message": {"role": "assistant", "content": content},
+                "finish_reason": finish,
+            }
+        ]
+    )
+    body: dict[str, Any] = {"choices": choices}
+    if usage is not None:
+        body["usage"] = usage
+    return httpx.Response(200, json=body)
+
+
+def test_complete_uses_official_chat_completions_shape(client_factory) -> None:
+    def responder(request: httpx.Request) -> httpx.Response:
+        body = json.loads(request.content)
+        assert request.url.path == "/v1/chat/completions"
+        assert body["model"] == "mistral-large-latest"
+        assert body["messages"] == [{"role": "user", "content": "сделай"}]
+        assert body["temperature"] == 0
+        assert body["max_tokens"] == 16000
+        return mistral_response('```json\n{"title": "ok", "count": 5}\n```')
+
+    result = client_factory(responder).complete("сделай", Answer)
+    assert result == Answer(title="ok", count=5)
+
+
+def test_list_content_chunks_are_joined(client_factory) -> None:
+    response = [
+        {"type": "text", "text": '{"title": "ok", '},
+        {"type": "text", "text": '"count": 6}'},
+    ]
+    result = client_factory(lambda _request: mistral_response(response)).complete("сделай", Answer)
+    assert result == Answer(title="ok", count=6)
+
+
+def test_complete_uses_shared_repair_loop(client_factory) -> None:
+    prompts: list[str] = []
+
+    def responder(request: httpx.Request) -> httpx.Response:
+        prompts.append(json.loads(request.content)["messages"][0]["content"])
+        if len(prompts) == 1:
+            return mistral_response('{"title": "ok", "count": "не число"}')
+        return mistral_response('{"title": "ok", "count": 7}')
+
+    result = client_factory(responder).complete("сделай", Answer)
+    assert result.count == 7
+    assert len(prompts) == 2
+    assert "не прошёл валидацию" in prompts[1]
+
+
+def test_standard_environment_key_is_used_without_logging_value(
+    monkeypatch: pytest.MonkeyPatch, tmp_path
+) -> None:
+    monkeypatch.setenv("MISTRAL_API_KEY", "test-standard-key")
+    client = MistralClient(
+        Settings(_env_file=None),
+        log_path=tmp_path / "llm_calls.jsonl",
+    )
+    try:
+        assert client._http.headers["Authorization"] == "Bearer test-standard-key"
+        assert not (tmp_path / "llm_calls.jsonl").exists()
+    finally:
+        client.close()
+
+
+def test_missing_key_fails_before_transport(monkeypatch: pytest.MonkeyPatch, tmp_path) -> None:
+    monkeypatch.delenv("MISTRAL_API_KEY", raising=False)
+    with pytest.raises(LLMError, match="API key is not configured"):
+        MistralClient(Settings(_env_file=None), log_path=tmp_path / "llm_calls.jsonl")
+
+
+def test_standard_key_name_is_loaded_from_dotenv(tmp_path) -> None:
+    env_file = tmp_path / ".env"
+    env_file.write_text("MISTRAL_API_KEY=test-dotenv-key\n", encoding="utf-8")
+    settings = Settings(_env_file=env_file)
+    assert settings.mistral_api_key == "test-dotenv-key"
+
+
+def test_http_error_omits_provider_response_body(client_factory) -> None:
+    marker = "MISTRAL_SECRET_RESPONSE_MARKER"
+
+    def responder(_request: httpx.Request) -> httpx.Response:
+        return httpx.Response(401, json={"api_key": marker})
+
+    with pytest.raises(LLMError) as exc_info:
+        client_factory(responder).complete("сделай", Answer)
+    message = str(exc_info.value)
+    assert "HTTP 401" in message
+    assert marker not in message
+
+
+def test_retries_429_with_retry_after_then_succeeds(client_factory) -> None:
+    calls: list[int] = []
+    sleeps: list[float] = []
+
+    def responder(_request: httpx.Request) -> httpx.Response:
+        calls.append(1)
+        if len(calls) <= 2:
+            return httpx.Response(429, headers={"Retry-After": "0.25"})
+        return mistral_response('{"title": "ok", "count": 3}')
+
+    result = client_factory(responder, sleep=sleeps.append).complete("сделай", Answer)
+    assert result.count == 3
+    assert len(calls) == 3
+    assert sleeps == [0.25, 0.25]
+
+
+def test_persistent_429_stops_after_bounded_retries(client_factory) -> None:
+    calls: list[int] = []
+
+    def responder(_request: httpx.Request) -> httpx.Response:
+        calls.append(1)
+        return httpx.Response(429, json={"message": "rate limited"})
+
+    with pytest.raises(LLMError, match="rate-limited after 4 retries"):
+        client_factory(responder).complete("сделай", Answer)
+    assert len(calls) == 5
+
+
+def test_usage_and_completed_status_are_written_to_store(client_factory, tmp_path) -> None:
+    store = Store(tmp_path / "usage.sqlite")
+    session_id = store.create_session("r")
+    usage = {"prompt_tokens": 120, "completion_tokens": 45}
+    client = client_factory(
+        lambda _request: mistral_response(
+            '{"title": "ok", "count": 5}',
+            usage=usage,
+        ),
+        store=store,
+    )
+    client.complete("сделай", Answer, session_id=session_id, step="propose_spec")
+    (call,) = store.llm_calls(session_id)
+    assert call["step"] == "propose_spec"
+    assert call["input_tokens"] == 120
+    assert call["output_tokens"] == 45
+    assert call["model"] == "mistral-large-latest"
+    assert call["status"] == "completed"
+    store.close()
+
+
+def test_budget_blocks_repair_before_third_provider_call(client_factory, tmp_path) -> None:
+    store = Store(tmp_path / "budget.sqlite")
+    session_id = store.create_session("r")
+    budget = LLMBudget(
+        store,
+        session_limits=BudgetLimits(max_calls=2),
+        day_limits=BudgetLimits(),
+    )
+    calls: list[int] = []
+
+    def responder(_request: httpx.Request) -> httpx.Response:
+        calls.append(1)
+        return mistral_response(f'{{"title": "ok", "count": "bad-{len(calls)}"}}')
+
+    with pytest.raises(BudgetExceeded, match="calls"):
+        client_factory(responder, store=store, budget=budget).complete(
+            "сделай",
+            Answer,
+            session_id=session_id,
+        )
+    assert len(calls) == 2
+    store.close()
+
+
+def test_factory_routes_mistral_and_wires_budget(tmp_path) -> None:
+    store = Store(tmp_path / "factory.sqlite")
+    settings = Settings(
+        _env_file=None,
+        llm_provider="mistral",
+        mistral_api_key="test-key",
+        llm_budget_enabled=True,
+        llm_budget_session_max_calls=5,
+    )
+    client = make_llm(settings, store=store)
+    try:
+        assert isinstance(client, MistralClient)
+        assert client._budget is not None
+    finally:
+        client.close()
+        store.close()
+
+
+def test_close_releases_injected_http_pool(client_factory) -> None:
+    client = client_factory(lambda _request: mistral_response('{"title": "ok", "count": 1}'))
+    http = client._http
+    client.close()
+    assert http.is_closed
+
+
+def test_live_sentinel_routes_mistral_secret_and_selects_provider() -> None:
+    workflow = (REPO / ".github" / "workflows" / "eval-live-sentinel.yml").read_text(
+        encoding="utf-8"
+    )
+    assert "MISTRAL_API_KEY: ${{ secrets.MISTRAL_API_KEY }}" in workflow
+    assert "AUTO_BI_MISTRAL_API_KEY: ${{ secrets.AUTO_BI_MISTRAL_API_KEY }}" in workflow
+    assert "AUTO_BI_MISTRAL_MODEL: ${{ secrets.AUTO_BI_MISTRAL_MODEL }}" in workflow
+    assert 'AUTO_BI_LLM_PROVIDER="mistral"' in workflow
+    assert (
+        'if [ -n "${MISTRAL_API_KEY}" ] || ' '[ -n "${AUTO_BI_MISTRAL_API_KEY}" ]; then'
+    ) in workflow
+
+
+def test_example_and_budget_table_cover_mistral() -> None:
+    env_example = (REPO / ".env.example").read_text(encoding="utf-8")
+    assert "MISTRAL_API_KEY=change_me" in env_example
+    assert "AUTO_BI_MISTRAL_MODEL=mistral-large-latest" in env_example
+    prices = Settings(_env_file=None).llm_budget_prices
+    assert "mistral-large-latest:0.0005/0.0015" in prices
+
+
+def test_cli_hosted_llm_readiness_message_uses_configured_provider() -> None:
+    """llm_healthcheck must label the actual hosted provider, not hardcode anthropic.
+
+    Nested closure in serve wiring — source-ratchet (same style as workflow checks above).
+    """
+    source = (REPO / "auto_bi" / "cli.py").read_text(encoding="utf-8")
+    assert 'message="anthropic: no live check (avoids token cost)"' not in source
+    assert 'message=f"{provider}: no live check (avoids token cost)"' in source
+    assert "provider = settings.llm_provider.strip().lower()" in source
+
+
+def test_auto_bi_prefixed_mistral_key_alias_populates_settings(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    monkeypatch.delenv("MISTRAL_API_KEY", raising=False)
+    monkeypatch.setenv("AUTO_BI_MISTRAL_API_KEY", "test-auto-bi-mistral-key")
+    settings = Settings(_env_file=None)
+    assert settings.mistral_api_key == "test-auto-bi-mistral-key"
diff --git a/tests/test_readme_local_first_ladder.py b/tests/test_readme_local_first_ladder.py
index 6e6cfa7..7da46d1 100644
--- a/tests/test_readme_local_first_ladder.py
+++ b/tests/test_readme_local_first_ladder.py
@@ -1,4 +1,7 @@
-"""Docs ratchet: README local-first ladder (offline → HF Space → full local).
+"""Docs ratchet: README exposes only the two current local-first paths.
+
+Supported paths: offline golden path + full LOCAL_BYOK. Excluded external demo
+paths must not reappear as onboarding or current product status.
 
 Pathlib-only: no app imports, no network, no Settings.
 """
@@ -21,14 +24,20 @@ def _howto_section() -> str:
 def test_readme_local_first_ladder() -> None:
     section = _howto_section()
 
+    # Offline golden path (supported).
     assert "uv run python scripts/demo_golden_path.py" in section
     assert "без DWH, BI, LLM и API-ключа" in section
 
-    assert "HF Space" in section
-    assert "auto-only" in section
-    assert "пользовательский ключ не нужен" in section
-    assert "текстовый режим там намеренно недоступен" in section
+    # Excluded external demo paths are not onboarding steps or current status.
+    assert "HF Space" not in section
+    assert "Hugging Face" not in section
+    assert "hf.space" not in section
+
+    # Negative: superseded claims from the removed external demo route.
+    assert "пользовательский ключ не нужен" not in section
+    assert "текстовый режим там намеренно недоступен" not in section
 
+    # Full LOCAL_BYOK path (supported).
     assert ".env.example" in section
     assert "cp .env.example .env" in section
     assert "Copy-Item .env.example .env" in section