Seed Assistent personas as overlay folders and tighten Ollama/Assistent glue.

gpu-rent now writes personas/<id>/ on the VM (not legacy personas.json), adds seed-personas/doctor checks, and shortens mid/high keep-alive now that Assistent parks the LLM before Generate.

Co-authored-by: Cursor <cursoragent@cursor.com>

assistent-personas.example.yaml

@@ -1,6 +1,8 @@
 # Copy to assistent-personas.yaml (never commit the working copy).
-# Synced to the VM on each up → /mnt/swarm_data/Assistent/personas.yaml + personas.json
+# Synced to the VM on each up / seed-personas →
-# Assistent reads personas.json (overlay). Yaml is the laptop source for gpu-rent seed.
+#   /mnt/swarm_data/Assistent/personas/<id>/{persona.json,extra.md}
+#   /mnt/swarm_data/Assistent/_base/assistant.json (default_persona, num_ctx)
+# Yaml is the laptop source for gpu-rent seed; Assistent reads overlay folders.
 
 default: neutral
 

docs/autocomplete.md

@@ -18,6 +18,8 @@
 
 Файл маленький (danbooru.csv ≈ 3–4 MB). Killer из‑за него не откладываем, но на первом bootstrap качаем **до** старта UI, вместе с extensions.
 
+Swarm Assistent индексирует тот же csv в FTS-таблицу `tags` внутри `Assistent/memory/assistent.sqlite` (без эмбедов). Первый чат после смены sha занимает несколько секунд; дальше `lookup_tags` ищет канон/алиасы. В `memory_hits` csv не попадает.
+
 ## Откуда файл
 
 Дефолт — то, на что ссылается сама документация SwarmUI:

docs/cli.md

@@ -75,6 +75,7 @@ gpu-rent up --yes --ollama
 | `gpu-rent ssh` | Оболочка на VM |
 | `gpu-rent seed-models` | Докачать новые строки Civitai-манифеста на живой диск |
 | `gpu-rent seed-extensions` | Доклонировать/обновить git-репы; restart swarmui |
+| `gpu-rent seed-personas` | Пуш `assistent-personas.yaml` → overlay на VM (без полного up) |
 | `gpu-rent push` / `push-models` | Локальные деревья → VM |
 | `gpu-rent pull-output` | VM `Output/` → `./Output` |
 | `gpu-rent capture` / `capture all` | Инвентарь VM → merge **ссылок** в `models.yaml` + `extensions.yaml` (веса не качать) |

docs/decisions.md

@@ -10,7 +10,7 @@
 | Civitai хост | Дефолт API `civitai.red` (полный каталог). `.com` — SFW-витрина, NSFW с неё часто 404. Ссылки `.com`/`.red`/`.green` в манифесте принимаем. 404 → один retry на второй хост. Один токен на оба домена |
 | Пул GPU | Перед `up`/`flavors` сканируем `SCAN_POOLS` (дефолт `ru-6,ru-7`) по compute URL. Сегменты `ru-6a/b/c` берём из Nova `availability_zones`, не отдельными пулами. Пресет размера `FLAVOR_SIZE_PRESET` (дефолт `cheap` = 1×GPU, min vCPU/RAM). Диск уже есть — AZ диска, не дешёвый чужой сегмент. Первый запуск — дешёвый SKU в сегменте, где есть карта из `FLAVOR_PREFERENCE`. `.env` автоматом не пишем |
 | Манифест моделей | `<repo>/models.yaml`, типы: checkpoint / lora / vae / embedding / controlnet / upscaler. В git только `models.example.yaml` |
-| Расширения | `<repo>/extensions.yaml`: git-репы `swarmui` → `src/Extensions`, `comfy` → DLNodes. Поле `requires: none\|ollama\|any-llm` фильтрует по `LLM_RUNTIME`. Пример: swarm-assistent с `requires: ollama` |
+| Расширения | `<repo>/extensions.yaml`: git-репы `swarmui` → `src/Extensions`, `comfy` → DLNodes. Поле `requires: none\|ollama\|any-llm` фильтрует по `LLM_RUNTIME`. Пример: swarm-assistent с `requires: ollama`. Personas: `assistent-personas.yaml` → overlay `Assistent/personas/<id>/` на `up` / `seed-personas` |
 | Autocomplete | До первого старта: скачать word-list в `Data/Autocompletions`, прописать `DefaultUser.AutoComplete.Source`. На каждом `up` сверить GitHub blob sha и обновить файл, если изменился. Дефолт: `tags/danbooru.csv` из a1111-sd-webui-tagcomplete (как в доке SwarmUI) |
 | Доступ | Браузер на туннеле; MCP переключается на облако, пока оно живо; HTTP API SwarmUI через тот же туннель |
 | Preemptible | По умолчанию всегда. Обычный сервер — только `--no-spot` |

docs/extensions.md

@@ -53,7 +53,7 @@ comfy:
 | `ollama` | только при `LLM_RUNTIME=ollama` |
 | `any-llm` | при `LLM_RUNTIME=ollama` |
 
-Строки с несовпавшим `requires` пропускаются (лог), остальные ставятся как обычно. В `extensions.example.yaml` по умолчанию — **swarm-assistent** с `requires: ollama` (чат + доска Generate/Ref + persona/Cards, ветка `main`). На `up` сидится `assistent-personas.yaml` → `/mnt/swarm_data/Assistent/personas.yaml` **и** `personas.json` (Assistent читает только **json**).
+Строки с несовпавшим `requires` пропускаются (лог), остальные ставятся как обычно. В `extensions.example.yaml` по умолчанию — **swarm-assistent** с `requires: ollama` (чат + доска Generate/Ref + persona/Cards, ветка `main`). На `up` (и `gpu-rent seed-personas`) сидится `assistent-personas.yaml` → overlay `/mnt/swarm_data/Assistent/personas/<id>/persona.json` + `extra.md`, плюс `_base/assistant.json` (`default_persona`, опционально `num_ctx` с GPU tier). Legacy `personas.json` больше не пишется.
 
 В git репозитория gpu-rent не коммитить рабочий список с лишними приватными URL сверх примера.
 

docs/llm.md

@@ -81,18 +81,20 @@ $env:OLLAMA_HOST = "http://127.0.0.1:17811"
 
 После успешного тега — **warmup**: 1-token `POST /api/chat`, чтобы веса (~6 GB VL) легли в VRAM до первого сообщения в Assistent. То же при `gpu-rent tunnel`, если `/api/ps` пуст. Промах warmup — warning, GPU не гасим.
 
+Qwen3-VL (`huihui_ai/qwen3-vl-abliterated:…`) требует **Ollama ≥ 0.12.7**. Тег `:8b-instruct` — чат Instruct; `:latest` у этой библиотеки — Thinking, не ставить дефолтом.
+
 ### Пресеты (меню)
 
 | # | ключ | tag / смысл |
 | --- | --- | --- |
-| 1 | **recommended** | `huihui_ai/qwen2.5-vl-abliterated:7b` — RU + vision, ~6 GB |
+| 1 | **recommended** | `huihui_ai/qwen3-vl-abliterated:8b-instruct` — RU + vision, ~6.1 GB; запасной `huihui_ai/qwen2.5-vl-abliterated:7b` |
 | 2 | light | `…:3b` — то же, мало VRAM (~3 GB) |
 | 3 | text | `huihui_ai/qwen2.5-abliterate:7b` — RU, без vision (~5 GB) |
 | 4 | big | `huihui_ai/qwen2.5-vl-abliterated:32b` — RU + vision (~21 GB) |
 | 5 | empty | только runtime |
 | — | keep | не трогать yaml |
 
-Все пресеты (кроме `empty`) — **abliterated** Qwen2.5 с нормальным русским. Официальные censored-теги (`qwen2.5vl:…`) в меню нет.
+Рекомендуемый пресет — **abliterated Qwen3-VL Instruct**; light / text / big пока Qwen2.5. Официальные censored-теги (`qwen2.5vl:…`) в меню нет.
 
 `default: true` в yaml — preferred в логе; pull идёт по всему списку.
 
@@ -105,18 +107,22 @@ Unit `gpu-rent-ollama` читает `/mnt/swarm_data/.gpu-rent-gpu.json`:
 | Tier (VRAM) | Flash Attn | KEEP_ALIVE | KV cache | GPU_OVERHEAD | CONTEXT |
 | --- | --- | --- | --- | --- | --- |
 | low (<16 GiB) | off | 2m | q4_0 | 6 GiB | 8k |
-| mid (16–23) | on* | 15m | q8_0 | 10 GiB | 16k |
+| mid (16–23) | on* | 5m | q8_0 | 10 GiB | 16k |
-| high (24–47) | on* | 15m | q8_0 | 14 GiB | 16k |
+| high (24–47) | on* | 5m | q8_0 | 14 GiB | 16k |
 | ultra (≥48) | on* | 30m | q8_0 | 20 GiB | 32k |
 
-\*Flash на Ampere+ (compute ≥ 8.0). `NUM_PARALLEL=2`, `MAX_LOADED_MODELS=2` (chat VL + memory embed). Memory models use `use: memory` and a CPU Modelfile (`num_gpu 0`) so embed does not steal VRAM from the chat model. Ollama default `num_ctx` is 4096; we set `OLLAMA_CONTEXT_LENGTH` so Assistent + vision fits. Env: `/mnt/swarm_data/.gpu-rent-ollama.env`.
+\*Flash на Ampere+ (compute ≥ 8.0). `NUM_PARALLEL=2`, `MAX_LOADED_MODELS=2` (chat VL + memory embed). Memory models use `use: memory` and a CPU Modelfile (`num_gpu 0`) so embed does not steal VRAM from the chat model. Ollama default `num_ctx` is 4096; we set `OLLAMA_CONTEXT_LENGTH` so Assistent + vision fits. Mid/high keep-alive is **5m** because Assistent parks the chat model before Generate (`AssistentParkLlm`) and warms after. Env: `/mnt/swarm_data/.gpu-rent-ollama.env`.
+
+Personas: laptop `assistent-personas.yaml` → on `up` / `seed-personas` overlay folders under `/mnt/swarm_data/Assistent/personas/<id>/` (+ `_base/assistant.json` with `default_persona` / `num_ctx`). See [extensions.md](extensions.md).
 
 `ollama-models.yaml` entries:
 
 ```yaml
-- name: huihui_ai/qwen2.5-vl-abliterated:7b
+- name: huihui_ai/qwen3-vl-abliterated:8b-instruct
   use: chat
   default: true
+- name: huihui_ai/qwen2.5-vl-abliterated:7b
+  use: chat
 - name: nomic-embed-text
   use: memory
 ```
@@ -129,7 +135,8 @@ Roles are written to `/mnt/swarm_data/Assistent/ollama-roles.json` for the Assis
 
 Busy (не гасить GPU):
 
-- `ollama pull` (маркер младше ~45 мин; старше сбрасывается);
+- очередь / loading SwarmUI (Generate) — даже если Ollama `/api/ps` пуст после ParkLlm;
+- `ollama pull` (маркер младше ~3 ч; refresh во время pull);
 - загруженная модель в Ollama.
 
 Ошибка установки LLM на `up` — **fail** (не тихий лог).

ollama-models.example.yaml

@@ -2,13 +2,18 @@
 # name = exact tag for `ollama pull`.
 # use: chat — Assistent header select; use: memory — settings memory model.
 # Requirement: uncensored (abliterated) + Russian for chat. Empty models: [] → no pull.
+# Qwen3-VL tags need Ollama ≥ 0.12.7. Pin :8b-instruct — :latest is Thinking.
 
 models:
-  # Recommended (~6GB): vision + RU/EN, abliterated
+  # Recommended (~6.1GB): Qwen3-VL Instruct, vision + RU/EN, abliterated
-  - name: huihui_ai/qwen2.5-vl-abliterated:7b
+  - name: huihui_ai/qwen3-vl-abliterated:8b-instruct
     use: chat
     default: true
 
+  # Previous default — still in Assistent chat select
+  - name: huihui_ai/qwen2.5-vl-abliterated:7b
+    use: chat
+
   # Vector memory (CPU via num_gpu 0 Modelfile after pull) — not shown in chat select
   - name: nomic-embed-text
     use: memory

src/gpu_rent/access_card.py

@@ -40,6 +40,11 @@ def collect_access_links(cfg: Config, *, tunneled: bool) -> list[AccessLink]:
             links.extend(
                 [
                     AccessLink("SwarmUI UI", base, "браузер"),
+                    AccessLink(
+                        "Assistent",
+                        base,
+                        "вкладка Assistent в SwarmUI",
+                    ),
                     AccessLink("SwarmUI API", f"{base}/API/", "HTTP JSON"),
                     AccessLink("SwarmUI MCP", f"{base}/mcp", "Cursor mcp.json"),
                 ]

src/gpu_rent/cli.py

@@ -324,6 +324,16 @@ def status() -> None:
                 table.add_row("idle-killer", f"{killer_line} · arm ok (сессия)")
             else:
                 table.add_row("idle-killer", killer_line)
+            try:
+                from gpu_rent.provision import count_wanted_models_on_vm
+
+                wanted_n = count_wanted_models_on_vm(cfg, state.floating_ip)
+                table.add_row(
+                    "Assistent wanted",
+                    f"{wanted_n} в очереди" if wanted_n else "пусто",
+                )
+            except Exception:
+                table.add_row("Assistent wanted", "—")
         except GpuRentError as exc:
             table.add_row("диск used/free", f"SSH: {exc}")
             table.add_row("idle-killer", "нет SSH")
@@ -932,6 +942,18 @@ def seed_extensions_cmd() -> None:
         _die(exc)
 
 
+@app.command("seed-personas")
+def seed_personas_cmd() -> None:
+    """Push assistent-personas.yaml → VM Assistent overlay (без полного up)."""
+    try:
+        from gpu_rent.provision import seed_assistent_personas
+
+        cfg, host = _live()
+        seed_assistent_personas(cfg, host, log)
+    except GpuRentError as exc:
+        _die(exc)
+
+
 capture_app = typer.Typer(
     help=(
         "Снять с VM инвентарь → локальные манифесты (только ссылки, без весов). "

src/gpu_rent/doctor.py

@@ -15,7 +15,7 @@ from gpu_rent.inventory import (
     pick_volume_type,
     rank_flavors,
 )
-from gpu_rent.manifests import parse_extensions, parse_models
+from gpu_rent.manifests import parse_extensions, parse_models, repo_dirname, repo_matches_runtime
 from gpu_rent.os_client import (
     compute_quotas,
     connect,
@@ -243,10 +243,72 @@ def run_doctor() -> list[Check]:
     _civitai(cfg, checks)
     _huggingface(cfg, checks)
     _local_manifests(cfg, checks)
+    _llm_assistent(cfg, checks)
     _local_folders(cfg, checks)
     return checks
 
 
+def _llm_assistent(cfg: Config, checks: list[Check]) -> None:
+    """When Ollama is on, warn if Assistent stack pieces are missing locally."""
+    from gpu_rent.llm_runtime import normalize_runtime, parse_ollama_models
+    from gpu_rent.paths import ollama_models_manifest_path
+
+    runtime = normalize_runtime(getattr(cfg, "llm_runtime", "none"))
+    if runtime != "ollama":
+        checks.append(Check("LLM / Assistent", True, False, f"LLM_RUNTIME={runtime}"))
+        return
+
+    ollama_path = Path(
+        getattr(cfg, "ollama_models_manifest", None) or ollama_models_manifest_path()
+    )
+    if not ollama_path.is_file():
+        checks.append(
+            Check(
+                "ollama-models.yaml",
+                False,
+                True,
+                f"нет {ollama_path} — Assistent будет пустой. gpu-rent setup / скопируй example",
+            )
+        )
+    else:
+        try:
+            entries = parse_ollama_models(ollama_path)
+            chat = [e for e in entries if getattr(e, "use", "chat") != "memory"]
+            checks.append(
+                Check(
+                    "ollama-models.yaml",
+                    True,
+                    True,
+                    f"{len(entries)} тег(ов), chat={len(chat)}",
+                )
+            )
+        except Exception as exc:
+            checks.append(Check("ollama-models.yaml", False, True, str(exc)))
+
+    try:
+        repos = parse_extensions(cfg.extensions_manifest)
+    except ConfigError as exc:
+        checks.append(Check("Assistent ext", False, True, str(exc)))
+        return
+    has_assistent = any(
+        "assistent" in repo_dirname(r).lower()
+        or "assistent" in (r.url or "").lower()
+        for r in repos
+        if repo_matches_runtime(r, runtime)
+    )
+    if has_assistent:
+        checks.append(Check("Assistent ext", True, False, "swarm-assistent в extensions.yaml"))
+    else:
+        checks.append(
+            Check(
+                "Assistent ext",
+                True,
+                False,
+                "нет swarm-assistent (requires:ollama) в extensions.yaml — чат не поставится",
+            )
+        )
+
+
 def _civitai(cfg: Config, checks: list[Check]) -> None:
     if not cfg.civitai_api_token:
         checks.append(

src/gpu_rent/llm_runtime.py

@@ -21,10 +21,13 @@ VALID_RUNTIMES = frozenset({"none", "ollama"})
 MEMORY_EMBED_MODEL = "nomic-embed-text"
 
 OLLAMA_PRESETS: dict[str, list[str]] = {
-    # Requirement: uncensored (abliterated) + solid Russian. Qwen2.5 family.
+    # Requirement: uncensored (abliterated) + solid Russian. Vision for prompt help.
-    # Vision tags preferred for SwarmUI prompt help with images.
+    # Pin :8b-instruct — :latest on this library is Thinking, not chat.
-    # Memory embed (nomic) is appended separately with use: memory.
+    # Qwen3-VL needs Ollama ≥ 0.12.7. Memory embed (nomic) appended with use: memory.
-    "recommended": ["huihui_ai/qwen2.5-vl-abliterated:7b"],  # ~6 GB
+    "recommended": [
+        "huihui_ai/qwen3-vl-abliterated:8b-instruct",  # ~6.1 GB, default
+        "huihui_ai/qwen2.5-vl-abliterated:7b",  # ~6 GB, second chat tag
+    ],
     "light": ["huihui_ai/qwen2.5-vl-abliterated:3b"],  # ~3 GB
     "text": ["huihui_ai/qwen2.5-abliterate:7b"],  # ~5 GB, no vision
     "big": ["huihui_ai/qwen2.5-vl-abliterated:32b"],  # ~21 GB
@@ -32,7 +35,7 @@ OLLAMA_PRESETS: dict[str, list[str]] = {
 }
 
 OLLAMA_PRESET_LABELS: dict[str, str] = {
-    "recommended": "VL 7B abliterate — RU + картинки (~6GB)",
+    "recommended": "VL 8B Instruct abliterate — RU + картинки (~6GB)",
     "light": "VL 3B abliterate — RU + картинки, мало VRAM (~3GB)",
     "text": "7B abliterate text — RU, без vision (~5GB)",
     "big": "VL 32B abliterate — RU + картинки (~21GB)",
@@ -40,10 +43,8 @@ OLLAMA_PRESET_LABELS: dict[str, str] = {
     "keep": "не менять ollama-models.yaml",
 }
 
-# Deprecated text blob — prefer menu helpers below.
+# Deprecated: use ollama_preset_menu / OLLAMA_PRESET_LABELS
-PRESET_HELP = "\n".join(
+PRESET_HELP = ""  # kept empty; menus use OLLAMA_PRESET_LABELS
-    f"{k} — {v}" for k, v in OLLAMA_PRESET_LABELS.items() if k != "keep"
-)
 
 LLM_RUNTIME_LABELS: dict[str, str] = {
     "none": "только SwarmUI",

src/gpu_rent/perf_tiers.py

@@ -77,16 +77,16 @@ def ollama_tune_for(info: GpuInfo) -> OllamaTune:
         note = "ultra: flash+q8 KV, 20GiB reserved for Swarm, ctx 32k, keep 30m"
     elif info.tier == TIER_HIGH:
         overhead = 14 * 1024**3
-        keep = "15m"
+        keep = "5m"  # Assistent parks LLM before Generate; short keep is enough
         kv = "q8_0"
         ctx = 16384
-        note = "high: flash+q8 KV, 14GiB reserved for Swarm, ctx 16k, keep 15m"
+        note = "high: flash+q8 KV, 14GiB reserved for Swarm, ctx 16k, keep 5m"
     elif info.tier == TIER_MID:
         overhead = 10 * 1024**3
-        keep = "15m"
+        keep = "5m"
         kv = "q8_0"
         ctx = 16384
-        note = "mid: flash+q8 KV, 10GiB reserved for Swarm, ctx 16k, keep 15m"
+        note = "mid: flash+q8 KV, 10GiB reserved for Swarm, ctx 16k, keep 5m"
     else:
         overhead = 6 * 1024**3
         keep = "2m"

src/gpu_rent/provision.py

@@ -710,8 +710,55 @@ def seed_swarmui_api_keys(cfg: Config, host: str, log: Log) -> None:
         run_ssh(cfg, host, "rm -f /tmp/gpu-rent-swarm-api-keys.json", check=False)
 
 
+def _safe_persona_id(raw: object) -> str | None:
+    s = str(raw or "").strip().replace("\\", "/")
+    if not s or ".." in s or "/" in s:
+        return None
+    import re
+
+    if not re.fullmatch(r"[A-Za-z0-9][A-Za-z0-9_\-]{0,63}", s):
+        return None
+    return s
+
+
+def _overlay_num_ctx(cfg: Config, host: str) -> int | None:
+    """Prefer GPU-tier context from .gpu-rent-gpu.json when present."""
+    raw = run_ssh(
+        cfg,
+        host,
+        f"test -f {DATA}/.gpu-rent-gpu.json && cat {DATA}/.gpu-rent-gpu.json || true",
+        check=False,
+        timeout=15,
+    ).strip()
+    if not raw:
+        return None
+    try:
+        probe = json.loads(raw)
+        vram = int(probe.get("vram_mib") or probe.get("memory_total_mib") or 0)
+    except (json.JSONDecodeError, TypeError, ValueError):
+        return None
+    if vram <= 0:
+        return None
+    from gpu_rent.perf_tiers import GpuInfo, ollama_tune_for, tier_for_vram_mib
+
+    info = GpuInfo(
+        name=str(probe.get("name") or "gpu"),
+        vram_mib=vram,
+        compute_cap=str(probe.get("compute_cap") or "0.0"),
+        uuid=str(probe.get("uuid") or ""),
+        tier=tier_for_vram_mib(vram),
+    )
+    return ollama_tune_for(info).context_length
+
+
 def seed_assistent_personas(cfg: Config, host: str, log: Log) -> None:
-    """Push local assistent-personas.yaml → VM Assistent/personas.yaml + personas.json."""
+    """Push local assistent-personas.yaml → VM Assistent/personas/<id>/ overlay.
+
+    Laptop yaml stays the editor; on VM we write persona.json + extra.md and
+    overlay assistant.json (default_persona, optional num_ctx). Does not clobber
+    voice/likes/dislikes/rules already on disk. No longer writes legacy
+    personas.json / personas.yaml dumps.
+    """
     from gpu_rent.paths import assistent_personas_example_path, assistent_personas_manifest_path
 
     local = Path(
@@ -728,24 +775,96 @@ def seed_assistent_personas(cfg: Config, host: str, log: Log) -> None:
     if not text.strip():
         log("assistent-personas: пустой файл — skip")
         return
-    remote_dir = f"{DATA}/Assistent"
-    remote_yaml = f"{remote_dir}/personas.yaml"
-    remote_json = f"{remote_dir}/personas.json"
-    run_ssh(cfg, host, f"mkdir -p {shlex.quote(remote_dir)}", check=False)
-    put_text(cfg, host, remote_yaml, text if text.endswith("\n") else text + "\n")
     try:
         import yaml
 
         data = yaml.safe_load(text) or {}
-        put_text(
-            cfg,
-            host,
-            remote_json,
-            json.dumps(data, ensure_ascii=False, indent=2) + "\n",
-        )
     except Exception as exc:
-        log(f"assistent-personas: json convert — {exc}")
+        log(f"assistent-personas: yaml parse — {exc}")
-    log(f"assistent-personas → {remote_yaml}")
+        return
+    if not isinstance(data, dict):
+        log("assistent-personas: корень должен быть mapping — skip")
+        return
+
+    remote_dir = f"{DATA}/Assistent"
+    run_ssh(cfg, host, f"mkdir -p {shlex.quote(remote_dir)}/personas {shlex.quote(remote_dir)}/_base", check=False)
+
+    default_id = _safe_persona_id(data.get("default")) or "neutral"
+    personas = data.get("personas") or []
+    written = 0
+    if isinstance(personas, list):
+        for row in personas:
+            if not isinstance(row, dict):
+                continue
+            pid = _safe_persona_id(row.get("id"))
+            if not pid:
+                continue
+            title = str(row.get("title") or pid).strip() or pid
+            prompt = str(row.get("prompt") or "").strip()
+            pdir = f"{remote_dir}/personas/{pid}"
+            run_ssh(cfg, host, f"mkdir -p {shlex.quote(pdir)}", check=False)
+            meta = {"title": title, "tagline": title, "accent": "#8b949e"}
+            put_text(
+                cfg,
+                host,
+                f"{pdir}/persona.json",
+                json.dumps(meta, ensure_ascii=False, indent=2) + "\n",
+            )
+            if prompt:
+                put_text(
+                    cfg,
+                    host,
+                    f"{pdir}/extra.md",
+                    prompt if prompt.endswith("\n") else prompt + "\n",
+                )
+            written += 1
+
+    assistant_overlay: dict = {"default_persona": default_id}
+    num_ctx = _overlay_num_ctx(cfg, host)
+    if num_ctx:
+        assistant_overlay["num_ctx"] = num_ctx
+    put_text(
+        cfg,
+        host,
+        f"{remote_dir}/_base/assistant.json",
+        json.dumps(assistant_overlay, ensure_ascii=False, indent=2) + "\n",
+    )
+    # Drop legacy dumps so Assistent does not double-inject prompts.
+    run_ssh(
+        cfg,
+        host,
+        f"rm -f {shlex.quote(remote_dir + '/personas.json')} "
+        f"{shlex.quote(remote_dir + '/personas.yaml')}",
+        check=False,
+    )
+    ctx_note = f", num_ctx={num_ctx}" if num_ctx else ""
+    log(f"assistent-personas → overlay personas/{written} (default={default_id}{ctx_note})")
+
+
+def count_wanted_models_on_vm(cfg: Config, host: str) -> int:
+    """Count entries in Assistent wanted queue on the VM (best-effort)."""
+    remote = f"{DATA}/.gpu-rent-wanted-models.yaml"
+    raw = run_ssh(
+        cfg,
+        host,
+        f"test -f {shlex.quote(remote)} && cat {shlex.quote(remote)} || true",
+        check=False,
+        timeout=20,
+    ).strip()
+    if not raw:
+        return 0
+    try:
+        import yaml
+
+        data = yaml.safe_load(raw) or {}
+    except Exception:
+        return sum(1 for line in raw.splitlines() if line.strip().startswith("- url:"))
+    n = 0
+    if isinstance(data, dict):
+        for rows in data.values():
+            if isinstance(rows, list):
+                n += sum(1 for r in rows if isinstance(r, dict) and r.get("url"))
+    return n
 
 
 def merge_wanted_models_from_vm(cfg: Config, host: str, log: Log) -> int:
@@ -1072,6 +1191,8 @@ def provision_llm(cfg: Config, host: str, log: Log) -> None:
             )
 
     # Drop LLM units that should not hold VRAM for this runtime.
+    # Tombstone: still stop gpu-rent-llamacpp if an old disk left that unit behind
+    # (llamacpp runtime was removed; do not reinstall it).
     if runtime == "none":
         log("LLM: none — останавливаю gpu-rent-ollama / gpu-rent-llamacpp если были")
         _stop_units("gpu-rent-ollama", "gpu-rent-llamacpp")
@@ -1083,7 +1204,7 @@ def provision_llm(cfg: Config, host: str, log: Log) -> None:
         return
     still: list[str] = []
     if runtime == "ollama":
-        _stop_units("gpu-rent-llamacpp")
+        _stop_units("gpu-rent-llamacpp")  # tombstone: disable leftover llamacpp unit
         log("LLM: ставим/запускаем Ollama")
         run_script_sudo(
             cfg,

src/gpu_rent/remote/idle_killer.py

@@ -129,7 +129,11 @@ def swarm_busy(swarm_url: str, timeout: float = 8.0) -> tuple[bool, str]:
 
 
 def llm_busy(timeout: float = 3.0) -> tuple[bool, str]:
-    """Ollama pull / loaded models count as busy."""
+    """Ollama pull / loaded models count as busy.
+
+    Empty /api/ps after AssistentParkLlm is NOT busy by itself — Swarm queue
+    is checked separately in main() and keeps the GPU alive during Generate.
+    """
     pull_marker = DATA / ".gpu-rent-ollama-pulling"
     if pull_marker.is_file():
         try:

src/gpu_rent/remote/install_ollama.sh

@@ -47,9 +47,9 @@ except ValueError:
 if gib >= 48:
     tier, overhead, keep, kv, flash, ctx = "ultra", 20 * 1024**3, "30m", "q8_0", True, 32768
 elif gib >= 24:
-    tier, overhead, keep, kv, flash, ctx = "high", 14 * 1024**3, "15m", "q8_0", True, 16384
+    tier, overhead, keep, kv, flash, ctx = "high", 14 * 1024**3, "5m", "q8_0", True, 16384
 elif gib >= 16:
-    tier, overhead, keep, kv, flash, ctx = "mid", 10 * 1024**3, "15m", "q8_0", True, 16384
+    tier, overhead, keep, kv, flash, ctx = "mid", 10 * 1024**3, "5m", "q8_0", True, 16384
 else:
     tier, overhead, keep, kv, flash, ctx = "low", 6 * 1024**3, "2m", "q4_0", False, 8192
 flash = bool(flash and (ampere or "A100" in name.upper() or "H100" in name.upper() or gib >= 16))

tests/test_access_card.py

@@ -16,9 +16,10 @@ def test_collect_links_swarm_and_ollama(monkeypatch):
     links = collect_access_links(_Cfg(), tunneled=True)
     labels = [x.label for x in links]
     assert "SwarmUI UI" in labels
+    assert "Assistent" in labels
     assert "SwarmUI MCP" in labels
     assert "Ollama API" in labels
-    assert any("17811" in x.url for x in links)
+    assert any(x.label == "Assistent" and "вкладка" in x.note for x in links)
 
 
 def test_collect_links_no_tunnel():
@@ -57,7 +58,7 @@ def test_access_panel_hides_stale_llm_error_when_ollama_ok(monkeypatch):
                         {
                             "name": "ollama",
                             "ok": True,
-                            "detail": "1 models (huihui_ai/qwen2.5-vl-abliterated:7b)",
+                            "detail": "1 models (huihui_ai/qwen3-vl-abliterated:8b-instruct)",
                         }
                     ],
                 }

tests/test_assistent_personas_seed.py

@@ -0,0 +1,81 @@
+"""Unit tests for assistent personas overlay seed helpers."""
+
+from __future__ import annotations
+
+import json
+from pathlib import Path
+from unittest.mock import MagicMock
+
+import pytest
+
+from gpu_rent.provision import _safe_persona_id, count_wanted_models_on_vm, seed_assistent_personas
+
+
+def test_safe_persona_id():
+    assert _safe_persona_id("neutral") == "neutral"
+    assert _safe_persona_id("lewd_v2") == "lewd_v2"
+    assert _safe_persona_id("../x") is None
+    assert _safe_persona_id("a/b") is None
+    assert _safe_persona_id("") is None
+
+
+def test_seed_assistent_personas_writes_overlay(tmp_path: Path, monkeypatch):
+    yaml_path = tmp_path / "assistent-personas.yaml"
+    yaml_path.write_text(
+        "default: cinema\n"
+        "personas:\n"
+        "  - id: cinema\n"
+        "    title: Кино\n"
+        "    prompt: |\n"
+        "      You are a DP.\n"
+        "  - id: neutral\n"
+        "    title: Нейтральный\n"
+        "    prompt: Calm.\n",
+        encoding="utf-8",
+    )
+    puts: dict[str, str] = {}
+    ssh_cmds: list[str] = []
+
+    def fake_put(cfg, host, remote, text, mode=0o644):
+        puts[remote] = text
+
+    def fake_ssh(cfg, host, cmd, check=False, timeout=60):
+        ssh_cmds.append(cmd)
+        if ".gpu-rent-gpu.json" in cmd:
+            return json.dumps({"vram_mib": 24576, "name": "4090", "compute_cap": "8.9"})
+        return ""
+
+    monkeypatch.setattr("gpu_rent.provision.put_text", fake_put)
+    monkeypatch.setattr("gpu_rent.provision.run_ssh", fake_ssh)
+    monkeypatch.setattr(
+        "gpu_rent.paths.assistent_personas_manifest_path", lambda: yaml_path
+    )
+    monkeypatch.setattr(
+        "gpu_rent.paths.assistent_personas_example_path", lambda: tmp_path / "missing"
+    )
+
+    cfg = MagicMock()
+    cfg.assistent_personas_manifest = str(yaml_path)
+    logs: list[str] = []
+    seed_assistent_personas(cfg, "1.2.3.4", logs.append)
+
+    assert any("personas/cinema/persona.json" in p for p in puts)
+    assert any("personas/cinema/extra.md" in p for p in puts)
+    assert "/mnt/swarm_data/Assistent/_base/assistant.json" in puts
+    asst = json.loads(puts["/mnt/swarm_data/Assistent/_base/assistant.json"])
+    assert asst["default_persona"] == "cinema"
+    assert asst["num_ctx"] == 16384
+    assert any("rm -f" in c and "personas.json" in c for c in ssh_cmds)
+    assert not any(p.endswith("personas.yaml") for p in puts)
+    assert not any(p.endswith("personas.json") for p in puts)
+    assert any("overlay personas/2" in m for m in logs)
+
+
+def test_count_wanted_models(monkeypatch):
+    raw = "lora:\n  - url: https://civitai.red/x?modelVersionId=1\n    title: A\n"
+
+    monkeypatch.setattr(
+        "gpu_rent.provision.run_ssh",
+        lambda *a, **k: raw,
+    )
+    assert count_wanted_models_on_vm(MagicMock(), "h") == 1

tests/test_doctor_llm.py

@@ -0,0 +1,61 @@
+"""Doctor LLM / Assistent local checks."""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+from gpu_rent.doctor import Check, _llm_assistent
+
+
+class _Cfg:
+    llm_runtime = "ollama"
+    ollama_models_manifest = None
+    extensions_manifest = None
+
+
+def test_llm_assistent_warns_without_ollama_yaml(tmp_path: Path, monkeypatch):
+    cfg = _Cfg()
+    cfg.ollama_models_manifest = tmp_path / "missing-ollama.yaml"
+    cfg.extensions_manifest = tmp_path / "extensions.yaml"
+    cfg.extensions_manifest.write_text(
+        "swarmui:\n"
+        "  - url: https://gitea.example/swarm-assistent.git\n"
+        "    dir: swarm-assistent\n"
+        "    requires: ollama\n",
+        encoding="utf-8",
+    )
+    monkeypatch.setattr(
+        "gpu_rent.paths.ollama_models_manifest_path",
+        lambda: cfg.ollama_models_manifest,
+    )
+    checks: list[Check] = []
+    _llm_assistent(cfg, checks)
+    names = {c.name: c for c in checks}
+    assert names["ollama-models.yaml"].ok is False
+    assert names["ollama-models.yaml"].blocking is True
+    assert names["Assistent ext"].ok is True
+
+
+def test_llm_assistent_ok_with_manifests(tmp_path: Path, monkeypatch):
+    ollama = tmp_path / "ollama-models.yaml"
+    ollama.write_text(
+        "models:\n  - name: foo:7b\n    use: chat\n    default: true\n",
+        encoding="utf-8",
+    )
+    ext = tmp_path / "extensions.yaml"
+    ext.write_text(
+        "swarmui:\n"
+        "  - url: https://gitea.example/swarm-assistent.git\n"
+        "    dir: swarm-assistent\n"
+        "    requires: ollama\n",
+        encoding="utf-8",
+    )
+    cfg = _Cfg()
+    cfg.ollama_models_manifest = ollama
+    cfg.extensions_manifest = ext
+    cfg.llm_runtime = "ollama"
+    checks: list[Check] = []
+    _llm_assistent(cfg, checks)
+    by = {c.name: c for c in checks}
+    assert by["ollama-models.yaml"].ok
+    assert "swarm-assistent" in by["Assistent ext"].detail

tests/test_llm_runtime.py

@@ -56,7 +56,9 @@ def test_write_preset(tmp_path: Path):
     path = tmp_path / "out.yaml"
     write_ollama_models_preset(path, "recommended")
     entries = parse_ollama_models(path)
-    assert entries[0].name == "huihui_ai/qwen2.5-vl-abliterated:7b"
+    assert entries[0].name == "huihui_ai/qwen3-vl-abliterated:8b-instruct"
+    assert entries[0].default is True
+    assert entries[1].name == "huihui_ai/qwen2.5-vl-abliterated:7b"
     write_ollama_models_preset(path, "big")
     assert parse_ollama_models(path)[0].name == "huihui_ai/qwen2.5-vl-abliterated:32b"
     write_ollama_models_preset(path, "text")

tests/test_ollama_warmup.py

@@ -43,10 +43,10 @@ def test_warmup_skips_when_ps_has_model(monkeypatch):
     def fake_urlopen(req, timeout=None):
         url = getattr(req, "full_url", str(req))
         assert "/api/ps" in url
-        return FakeResp('{"models":[{"name":"huihui_ai/qwen2.5-vl-abliterated:7b"}]}')
+        return FakeResp('{"models":[{"name":"huihui_ai/qwen3-vl-abliterated:8b-instruct"}]}')
 
     monkeypatch.setattr(_mod.urllib.request, "urlopen", fake_urlopen)
-    msg = _mod.warmup("huihui_ai/qwen2.5-vl-abliterated:7b", keep_alive="15m", num_ctx=16384, timeout=5)
+    msg = _mod.warmup("huihui_ai/qwen3-vl-abliterated:8b-instruct", keep_alive="15m", num_ctx=16384, timeout=5)
     assert "skip" in msg
     assert "VRAM" in msg
 

tests/test_perf_tiers.py

@@ -82,6 +82,6 @@ def test_ollama_mid_4090_context_16k():
     )
     tune = ollama_tune_for(info)
     assert tune.context_length == 16384
-    assert tune.keep_alive == "15m"
+    assert tune.keep_alive == "5m"
     assert "OLLAMA_CONTEXT_LENGTH=16384" in "\n".join(ollama_env_lines(tune))
-    assert "OLLAMA_KEEP_ALIVE=15m" in "\n".join(ollama_env_lines(tune))
+    assert "OLLAMA_KEEP_ALIVE=5m" in "\n".join(ollama_env_lines(tune))

tests/test_warm_up_skips.py

@@ -86,7 +86,7 @@ def test_install_ollama_skips_restart_when_unit_unchanged():
     text = files("gpu_rent.remote").joinpath("install_ollama.sh").read_text(encoding="utf-8")
     assert "cmp -s" in text
     assert "skip restart" in text
-    assert '"mid", 10 * 1024**3, "15m"' in text
+    assert '"mid", 10 * 1024**3, "5m"' in text
 
 
 def test_provision_llm_skips_on_api_tags_not_cli_list():