Add multi-turn Assistent session diagnostics to the Debug API.

POST /assistent/session + /chat with rich per-turn traces (patch, Exact merge, compact_context); chat-eval wraps one session turn. Docs playbook and unit/HTTP tests included.

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

docs/cli.md

@@ -133,31 +133,59 @@ Exit 0 → можно `up`. Exit 1 → причина в таблице / кра
 | Путь | Зачем |
 | --- | --- |
 | `/assistent` | Сводка: extension + overlay + roles + memory + live API + hints + playbook |
+| `/assistent/diagnose` | То же + journal по умолчанию + session store meta |
 | `/assistent/extension` | DLL, Sqlite deps, Tab/Assets, git HEAD |
 | `/assistent/overlay` | personas на VM vs локальный `assistent-personas/` |
 | `/assistent/roles` | `ollama-roles.json` ↔ `/api/tags` (`default_chat`) |
 | `/assistent/memory` | `assistent.sqlite` + counts |
 | `/assistent/api` | `AssistentListPersonas/Models/Memory/Chats` через туннель (или SSH) |
 | `/assistent/api?chat_smoke=1` | + 1-token Ollama `/api/chat` (кратко грузит модель) |
-| `/assistent/chat-eval` | **Opt-in** `GetNewSession` → `AssistentChat` (полный ответ ассистента) |
 | `/assistent/logs` | journalctl swarmui: build/load/Sqlite |
 | `/assistent?logs=1` | Сводка + journal |
+| `/assistent/session` | **POST** — multi-turn debug chat (in-memory, TTL ~45m) |
+| `/assistent/session/{id}/chat` | **POST** — ход + объект `trace` |
+| `/assistent/session/{id}` | **GET** / **DELETE** — состояние / сброс |
+| `/assistent/chat-eval` | One-shot: session + один chat + delete |
 
 Симптомы → куда смотреть: поле `playbook` в `/assistent`.
 
-**`/assistent/chat-eval`** — не входит в `/snapshot`. Может загрузить chat-модель в VRAM и записать чат в Sqlite (если DLL на месте). Предпочтительно `POST` с JSON; для агентов удобен и `GET` с query.
+**Session / chat-eval** не входят в `/snapshot`. Могут грузить VRAM, писать sqlite chat, тратить биллинг GPU. `compactContext` / Exact merge в Assistent HTTP API **нет** — sidecar реконструирует best-effort (`trace.gaps`).
+
+#### Agent playbook (curl)
 
 ```bash
-# POST (предпочтительно)
+BASE=http://127.0.0.1:17821
-curl -sS -X POST http://127.0.0.1:17821/assistent/chat-eval \
-  -H "Content-Type: application/json" \
-  -d "{\"message\":\"какой checkpoint и какие steps/cfg должны быть\",\"persona\":\"leonid\",\"timeout\":120}"
 
-# GET (агенты / быстрый smoke)
+# Static+live health (no chat)
-curl -sS "http://127.0.0.1:17821/assistent/chat-eval?message=ping&timeout=90"
+curl -sS "$BASE/assistent/diagnose" | jq '{ok,hints,playbook,session_store}'
+
+# Multi-turn conversational diagnostics
+SID=$(curl -sS -X POST "$BASE/assistent/session" \
+  -H 'Content-Type: application/json' \
+  -d '{"persona":"neutral","pack":"ordinary","context":{"krea_profile":"turbo","checkpoint":"krea-turbo"}}' \
+  | jq -r .debug_session_id)
+
+curl -sS -X POST "$BASE/assistent/session/$SID/chat" \
+  -H 'Content-Type: application/json' \
+  -d '{"message":"какой checkpoint и какие steps/cfg должны быть"}' \
+  | jq '{ok,reply,trace:{timings_ms,patch,exact_merge,compact_context,system_layers,gaps}}'
+
+curl -sS -X POST "$BASE/assistent/session/$SID/chat" \
+  -H 'Content-Type: application/json' \
+  -d '{"message":"сделай generate с этими params"}' \
+  | jq '{ok, trace:{patch,exact_merge}}'
+
+curl -sS "$BASE/assistent/session/$SID" | jq .session
+curl -sS -X DELETE "$BASE/assistent/session/$SID"
+
+# One-shot convenience (session + one chat + delete)
+curl -sS -X POST "$BASE/assistent/chat-eval" \
+  -H 'Content-Type: application/json' \
+  -d '{"message":"какие steps/cfg для turbo?","persona":"leonid","timeout":120}' \
+  | jq '{ok,reply,patch,trace}'
 ```
 
-Ответ: `ok`, `ms` / `chat_ms`, `model` / `preferred`, `reply` / `reply_prose`, `patch` (prompt/steps/cfg/aspect/actions…), `errors`, `hints`. Туннель `:swarmui_local_port` (обычно 17801), иначе SSH на VM `:7801`.
+Ответ chat: `ok`, `reply`, `trace` (`timings_ms`, `patch`, `exact_merge`, `compact_context`, `system_chars` / `system_layers`, `skills`, `hops`, `gaps`, `errors`/`warnings`/`hints`). Sqlite SaveChat fail — soft error, если текст ответа есть.
 
 После `up --no-tunnel` процесс завершается и sidecar гаснет — держи отдельно:
 
@@ -165,7 +193,7 @@ curl -sS "http://127.0.0.1:17821/assistent/chat-eval?message=ping&timeout=90"
 gpu-rent debug
 ```
 
-Мутаций конфига/GPU нет (restart/hold/stop — как раньше через CLI). `chat_smoke` только пингует Ollama; `chat-eval` — полноценный AssistentChat (opt-in).
+Мутаций конфига/GPU нет (restart/hold/stop — как раньше через CLI). `chat_smoke` только пингует Ollama; session/chat-eval — полноценный AssistentChat (opt-in).
 
 ---
 

src/gpu_rent/debug_api.py

@@ -1,8 +1,8 @@
 """Localhost debug HTTP sidecar for gpu-rent.
 
 Bind only 127.0.0.1. Started from `up` / `tunnel` / `debug`.
-Mostly read-only; `/assistent/chat-eval` is an opt-in AssistentChat probe
+Mostly read-only; Assistent session / chat-eval endpoints are opt-in
-(may load VRAM / persist chat) and is not part of `/snapshot`.
+(may load VRAM / persist chat) and are not part of `/snapshot`.
 """
 
 from __future__ import annotations
@@ -63,16 +63,25 @@ _CHAT_EVAL_PARAMS = [
     },
 ]
 
+_WARN_VRAM = (
+    "WARNING: may load chat model into VRAM, may write assistent.sqlite chat, "
+    "and incurs GPU billing if the VM is up. Not part of /snapshot. "
+    "compactContext / Exact merge are reconstructed client-side "
+    "(Assistent HTTP API does not return them)."
+)
+
 _OPENAPI: dict[str, Any] = {
     "openapi": "3.0.3",
     "info": {
         "title": "gpu-rent Debug API",
-        "version": "1.1.0",
+        "version": "1.2.0",
         "description": (
             "Localhost diagnostics for installer progress and "
-            "SwarmUI / Comfy / Ollama / Assistent. Mostly read-only; "
+            "SwarmUI / Comfy / Ollama / Assistent. Mostly read-only. "
-            "POST/GET /assistent/chat-eval is opt-in AssistentChat "
+            "Assistent multi-turn: POST /assistent/session + "
-            "(may load VRAM / write chat) and is skipped from /snapshot."
+            "/assistent/session/{id}/chat (trace object). "
+            "One-shot: POST/GET /assistent/chat-eval. "
+            + _WARN_VRAM
         ),
     },
     "servers": [{"url": "http://127.0.0.1:17821"}],
@@ -95,7 +104,9 @@ _OPENAPI: dict[str, Any] = {
             }
         },
         "/snapshot": {
-            "get": {"summary": "Cheap aggregate (no full diag, no chat-eval)"}
+            "get": {
+                "summary": "Cheap aggregate (no full diag, no Assistent chat/session)"
+            }
         },
         "/status": {"get": {"summary": "JSON mirror of gpu-rent status"}},
         "/state": {"get": {"summary": "state.json without secrets"}},
@@ -142,6 +153,27 @@ _OPENAPI: dict[str, Any] = {
                 ],
             }
         },
+        "/assistent/diagnose": {
+            "get": {
+                "summary": (
+                    "Full static+live Assistent health (logs on by default). "
+                    + _WARN_VRAM
+                ),
+                "parameters": [
+                    {
+                        "name": "chat_smoke",
+                        "in": "query",
+                        "schema": {"type": "boolean"},
+                    },
+                    {
+                        "name": "logs",
+                        "in": "query",
+                        "schema": {"type": "boolean", "default": True},
+                        "description": "Pass logs=0 to skip journal",
+                    },
+                ],
+            }
+        },
         "/assistent/extension": {
             "get": {"summary": "DLL/Sqlite deps/Tab assets under Extensions/swarm-assistent"}
         },
@@ -171,18 +203,92 @@ _OPENAPI: dict[str, Any] = {
         "/assistent/logs": {
             "get": {"summary": "journalctl swarmui filtered for Assistent/Sqlite/build"}
         },
+        "/assistent/session": {
+            "post": {
+                "summary": (
+                    "Start multi-turn Assistent debug session "
+                    "(GetNewSession + optional GetConfig). " + _WARN_VRAM
+                ),
+                "requestBody": {
+                    "required": False,
+                    "content": {
+                        "application/json": {
+                            "schema": {
+                                "type": "object",
+                                "properties": {
+                                    "persona": {"type": "string"},
+                                    "pack": {"type": "string"},
+                                    "model": {"type": "string"},
+                                    "skills": {
+                                        "type": "array",
+                                        "items": {"type": "string"},
+                                    },
+                                    "context": {
+                                        "type": "object",
+                                        "description": (
+                                            "Synthetic compactContext fields "
+                                            "(krea_profile, checkpoint, "
+                                            "recommended_params, …)"
+                                        ),
+                                    },
+                                    "probe_config": {
+                                        "type": "boolean",
+                                        "default": True,
+                                    },
+                                },
+                            }
+                        }
+                    },
+                },
+            }
+        },
+        "/assistent/session/{id}": {
+            "get": {"summary": "In-memory session summary + last trace"},
+            "delete": {"summary": "Drop in-memory debug session"},
+        },
+        "/assistent/session/{id}/chat": {
+            "post": {
+                "summary": (
+                    "Send user message; returns reply + trace "
+                    "(patch, Exact merge, compact_context sizes, system_layers). "
+                    + _WARN_VRAM
+                ),
+                "requestBody": {
+                    "required": True,
+                    "content": {
+                        "application/json": {
+                            "schema": {
+                                "type": "object",
+                                "required": ["message"],
+                                "properties": {
+                                    "message": {"type": "string"},
+                                    "timeout": {"type": "number"},
+                                    "persona": {"type": "string"},
+                                    "pack": {"type": "string"},
+                                    "model": {"type": "string"},
+                                    "skills": {
+                                        "type": "array",
+                                        "items": {"type": "string"},
+                                    },
+                                    "context": {"type": "object"},
+                                },
+                            }
+                        }
+                    },
+                },
+            }
+        },
         "/assistent/chat-eval": {
             "get": {
                 "summary": (
-                    "Opt-in AssistentChat eval (GetNewSession→AssistentChat); "
+                    "One-shot AssistentChat (session+chat+delete). " + _WARN_VRAM
-                    "may load VRAM / write chat — not in /snapshot"
                 ),
                 "parameters": _CHAT_EVAL_PARAMS,
             },
             "post": {
                 "summary": (
                     "Same as GET; prefer JSON body "
-                    "{message,persona,pack,model,timeout}"
+                    "{message,persona,pack,model,timeout,context}"
                 ),
                 "requestBody": {
                     "required": False,
@@ -196,6 +302,7 @@ _OPENAPI: dict[str, Any] = {
                                     "pack": {"type": "string"},
                                     "model": {"type": "string"},
                                     "timeout": {"type": "number"},
+                                    "context": {"type": "object"},
                                 },
                             }
                         }
@@ -207,6 +314,21 @@ _OPENAPI: dict[str, Any] = {
 }
 
 
+def _is_assistent_session_chat(path: str) -> bool:
+    parts = [p for p in path.split("/") if p]
+    return (
+        len(parts) == 4
+        and parts[0] == "assistent"
+        and parts[1] == "session"
+        and parts[3] == "chat"
+    )
+
+
+def _is_assistent_session_id(path: str) -> bool:
+    parts = [p for p in path.split("/") if p]
+    return len(parts) == 3 and parts[0] == "assistent" and parts[1] == "session"
+
+
 @dataclass
 class DebugHub:
     """Shared state for the sidecar process."""
@@ -397,7 +519,9 @@ def _make_handler(hub: DebugHub):
                 f"<p>base <code>{hub.base_url}</code> · mostly read-only · 127.0.0.1</p>"
                 f"<p>Agent: start at <a href='/openapi.json'>/openapi.json</a> "
                 f"then <a href='/snapshot'>/snapshot</a>. "
-                f"Opt-in chat: <code>/assistent/chat-eval</code> (not in snapshot).</p>"
+                f"Opt-in Assistent: <code>/assistent/session</code> (multi-turn) · "
+                f"<code>/assistent/chat-eval</code> (one-shot) — not in snapshot; "
+                f"may load VRAM.</p>"
                 f"<ul>{rows}</ul></body></html>"
             )
             self._send(200, html.encode("utf-8"), "text/html; charset=utf-8")
@@ -445,6 +569,19 @@ def _make_handler(hub: DebugHub):
                     },
                 )
 
+        def do_DELETE(self) -> None:  # noqa: N802
+            try:
+                self._dispatch("DELETE")
+            except Exception as exc:
+                self._json(
+                    500,
+                    {
+                        "ok": False,
+                        "error": str(exc)[:300],
+                        "trace": traceback.format_exc()[-800:],
+                    },
+                )
+
         def _dispatch(self, method: str = "GET") -> None:
             parsed = urlparse(self.path)
             path = parsed.path.rstrip("/") or "/"
@@ -464,12 +601,27 @@ def _make_handler(hub: DebugHub):
                 doc["servers"] = [{"url": hub.base_url}]
                 self._json(200, doc)
                 return
-            if method == "POST" and path != "/assistent/chat-eval":
+            if method == "POST" and path not in {
+                "/assistent/chat-eval",
+                "/assistent/session",
+            } and not _is_assistent_session_chat(path):
                 self._json(
                     405,
                     {
                         "ok": False,
-                        "error": "POST only allowed for /assistent/chat-eval",
+                        "error": (
+                            "POST only for /assistent/chat-eval, "
+                            "/assistent/session, /assistent/session/{id}/chat"
+                        ),
+                    },
+                )
+                return
+            if method == "DELETE" and not _is_assistent_session_id(path):
+                self._json(
+                    405,
+                    {
+                        "ok": False,
+                        "error": "DELETE only for /assistent/session/{id}",
                     },
                 )
                 return
@@ -592,16 +744,38 @@ def _make_handler(hub: DebugHub):
                 return
             if path == "/assistent" or path.startswith("/assistent/"):
                 from gpu_rent import debug_assistent
+                from gpu_rent import debug_assistent_session as das
 
-                def _flag(name: str) -> bool:
+                def _flag(name: str, *, default: bool = False) -> bool:
-                    raw = (qs.get(name) or ["0"])[0].strip().lower()
+                    vals = qs.get(name)
+                    if not vals:
+                        return default
+                    raw = vals[0].strip().lower()
                     return raw in {"1", "true", "yes", "on"}
 
                 chat_smoke = _flag("chat_smoke")
                 include_logs = _flag("logs")
                 sub = path[len("/assistent") :].lstrip("/") or ""
+                try_list = [
+                    "/assistent",
+                    "/assistent/diagnose",
+                    "/assistent/extension",
+                    "/assistent/overlay",
+                    "/assistent/roles",
+                    "/assistent/memory",
+                    "/assistent/api",
+                    "/assistent/wanted",
+                    "/assistent/logs",
+                    "/assistent/session",
+                    "/assistent/session/{id}",
+                    "/assistent/session/{id}/chat",
+                    "/assistent/chat-eval",
+                ]
 
                 if sub == "":
+                    if method != "GET":
+                        self._json(405, {"ok": False, "error": "GET only"})
+                        return
                     cache_key = f"assistent:{int(chat_smoke)}:{int(include_logs)}"
                     payload = hub.cached(
                         cache_key,
@@ -612,6 +786,29 @@ def _make_handler(hub: DebugHub):
                             include_logs=include_logs,
                         ),
                     )
+                elif sub == "diagnose":
+                    if method != "GET":
+                        self._json(405, {"ok": False, "error": "GET only"})
+                        return
+                    # logs default on for diagnose; logs=0 to skip
+                    diag_logs = _flag("logs", default=True)
+                    if qs.get("logs") and qs["logs"][0].strip().lower() in {
+                        "0",
+                        "false",
+                        "no",
+                        "off",
+                    }:
+                        diag_logs = False
+                    cache_key = f"assistent/diagnose:{int(chat_smoke)}:{int(diag_logs)}"
+                    payload = hub.cached(
+                        cache_key,
+                        CACHE_TTL_DEFAULT,
+                        lambda: das.collect_assistent_diagnose(
+                            hub.cfg,
+                            chat_smoke=chat_smoke,
+                            include_logs=diag_logs,
+                        ),
+                    )
                 elif sub == "extension":
                     payload = hub.cached(
                         "assistent/extension",
@@ -657,8 +854,78 @@ def _make_handler(hub: DebugHub):
                         CACHE_TTL_LOGS,
                         lambda: debug_assistent.collect_assistent_logs(hub.cfg),
                     )
+                elif sub == "session" and method == "POST":
+                    try:
+                        body = self._read_json_body()
+                    except ValueError as exc:
+                        self._json(400, {"ok": False, "error": str(exc)})
+                        return
+                    skills = body.get("skills")
+                    if skills is not None and not isinstance(skills, list):
+                        self._json(
+                            400, {"ok": False, "error": "skills must be an array"}
+                        )
+                        return
+                    ctx = body.get("context")
+                    if ctx is not None and not isinstance(ctx, dict):
+                        self._json(
+                            400, {"ok": False, "error": "context must be an object"}
+                        )
+                        return
+                    probe = body.get("probe_config", True)
+                    payload = das.create_debug_session(
+                        hub.cfg,
+                        persona=body.get("persona"),
+                        pack=body.get("pack"),
+                        model=body.get("model"),
+                        context=ctx,
+                        skills=skills,
+                        probe_config=bool(probe),
+                    )
+                elif _is_assistent_session_id(path):
+                    sid = sub.split("/", 1)[1]
+                    if method == "GET":
+                        payload = das.get_debug_session(sid)
+                    elif method == "DELETE":
+                        payload = das.delete_debug_session(sid)
+                    else:
+                        self._json(405, {"ok": False, "error": "GET or DELETE"})
+                        return
+                elif _is_assistent_session_chat(path):
+                    if method != "POST":
+                        self._json(405, {"ok": False, "error": "POST only"})
+                        return
+                    sid = sub.split("/")[1]
+                    try:
+                        body = self._read_json_body()
+                    except ValueError as exc:
+                        self._json(400, {"ok": False, "error": str(exc)})
+                        return
+                    skills = body.get("skills")
+                    if skills is not None and not isinstance(skills, list):
+                        self._json(
+                            400, {"ok": False, "error": "skills must be an array"}
+                        )
+                        return
+                    ctx = body.get("context")
+                    if ctx is not None and not isinstance(ctx, dict):
+                        self._json(
+                            400, {"ok": False, "error": "context must be an object"}
+                        )
+                        return
+                    payload = das.chat_debug_session(
+                        hub.cfg,
+                        sid,
+                        message=body.get("message"),
+                        timeout=body.get("timeout"),
+                        context=ctx,
+                        pack=body.get("pack"),
+                        model=body.get("model"),
+                        skills=skills,
+                        persona=body.get("persona"),
+                    )
                 elif sub == "chat-eval":
-                    body: dict[str, Any] = {}
+                    body = {}
                     if method == "POST":
                         try:
                             body = self._read_json_body()
@@ -666,12 +933,21 @@ def _make_handler(hub: DebugHub):
                             self._json(400, {"ok": False, "error": str(exc)})
                             return
 
-                    def _q(name: str) -> str | None:
+                    def _q(name: str) -> Any:
                         if name in body and body[name] is not None:
                             return body[name]
                         vals = qs.get(name)
                         return vals[0] if vals else None
 
+                    ctx = _q("context")
+                    if isinstance(ctx, str):
+                        try:
+                            ctx = json.loads(ctx)
+                        except json.JSONDecodeError:
+                            ctx = None
+                    if ctx is not None and not isinstance(ctx, dict):
+                        ctx = None
+
                     # Never cache — live AssistentChat / VRAM side effects.
                     payload = debug_assistent.run_assistent_chat_eval(
                         hub.cfg,
@@ -680,6 +956,7 @@ def _make_handler(hub: DebugHub):
                         pack=_q("pack"),
                         model=_q("model"),
                         timeout=_q("timeout"),
+                        context=ctx,
                     )
                 else:
                     self._json(
@@ -687,17 +964,7 @@ def _make_handler(hub: DebugHub):
                         {
                             "ok": False,
                             "error": f"unknown path {path}",
-                            "try": [
+                            "try": try_list,
-                                "/assistent",
-                                "/assistent/extension",
-                                "/assistent/overlay",
-                                "/assistent/roles",
-                                "/assistent/memory",
-                                "/assistent/api",
-                                "/assistent/wanted",
-                                "/assistent/logs",
-                                "/assistent/chat-eval",
-                            ],
                         },
                     )
                     return
@@ -759,7 +1026,9 @@ def start_debug_server(
             f"           {base}/openapi.json",
             f"           {base}/snapshot",
             f"           {base}/assistent",
-            f"           {base}/assistent/chat-eval  (opt-in; not in snapshot)",
+            f"           {base}/assistent/diagnose",
+            f"           {base}/assistent/session  (multi-turn; not in snapshot)",
+            f"           {base}/assistent/chat-eval  (one-shot; not in snapshot)",
         ]
         if log:
             for line in lines:
@@ -796,7 +1065,8 @@ def run_debug_blocking(
     try:
         if log:
             log("Debug API слушает (Ctrl+C — выход). "
-                "Мутаций конфига/GPU нет; /assistent/chat-eval — opt-in AssistentChat.")
+                "Мутаций конфига/GPU нет; "
+                "/assistent/session + /chat-eval — opt-in AssistentChat.")
         while True:
             time.sleep(3600)
     except KeyboardInterrupt:

src/gpu_rent/debug_assistent.py

@@ -2,7 +2,8 @@
 
 Probes extension compile/load markers, overlay personas, ollama-roles,
 sqlite, live SwarmUI Assistent* APIs, optional 1-token Ollama chat smoke,
-and opt-in AssistentChat evaluation (/assistent/chat-eval).
+opt-in AssistentChat evaluation (/assistent/chat-eval), and multi-turn
+debug sessions (/assistent/session*).
 """
 
 from __future__ import annotations
@@ -437,107 +438,23 @@ def run_assistent_chat_eval(
     pack: str | None = None,
     model: str | None = None,
     timeout: float | int | str | None = None,
+    context: dict | None = None,
 ) -> dict[str, Any]:
-    """Opt-in AssistentChat round-trip via local Swarm tunnel, else SSH :7801.
+    """Opt-in AssistentChat round-trip (session create → one chat → delete).
 
     May load the chat model into VRAM and write chat history if Sqlite works.
-    Not part of /snapshot.
+    Not part of /snapshot. Prefer POST /assistent/session for multi-turn.
     """
-    text = (message or DEFAULT_CHAT_EVAL_MESSAGE).strip() or DEFAULT_CHAT_EVAL_MESSAGE
+    from gpu_rent.debug_assistent_session import run_assistent_chat_eval_via_session
-    persona_id = (persona or "").strip() or "neutral"
-    pack_name = (pack or "").strip() or "ordinary"
-    t_chat = _clamp_chat_eval_timeout(timeout)
-    hints: list[str] = [
-        "opt-in AssistentChat eval — may load VRAM; may SaveChat if Sqlite works",
-        "not included in /snapshot",
-    ]
-    errors: list[str] = []
-    t0 = time.perf_counter()
 
-    base, via = _swarm_base(cfg)
+    return run_assistent_chat_eval_via_session(
-    if not base:
-        return _assistent_chat_eval_via_ssh(
         cfg,
-            message=text,
+        message=message,
-            persona=persona_id,
+        persona=persona,
-            pack=pack_name,
+        pack=pack,
-            model=(model or "").strip() or None,
+        model=model,
-            timeout=t_chat,
+        timeout=timeout,
-            hints=hints,
+        context=context,
-            t0=t0,
-        )
-
-    sid, err, ms_sess = _session_id(base)
-    if not sid:
-        return {
-            "ok": False,
-            "via": via,
-            "ms": round((time.perf_counter() - t0) * 1000, 1),
-            "session_ms": round(ms_sess, 1),
-            "error": f"GetNewSession failed: {err}",
-            "errors": [f"GetNewSession: {err}"],
-            "hints": hints,
-            "message": text,
-            "persona": persona_id,
-            "pack": pack_name,
-        }
-
-    chosen, preferred = _resolve_chat_model(cfg, base, sid, model=model)
-    if not chosen:
-        return {
-            "ok": False,
-            "via": via,
-            "ms": round((time.perf_counter() - t0) * 1000, 1),
-            "session_ms": round(ms_sess, 1),
-            "error": "no model (AssistentListModels.preferred / default_chat)",
-            "errors": ["model required"],
-            "hints": hints + ["GET /assistent/roles + /ollama — нет preferred chat model"],
-            "message": text,
-            "persona": persona_id,
-            "pack": pack_name,
-            "preferred": preferred,
-        }
-
-    payload = {
-        "session_id": sid,
-        "baseUrl": "http://127.0.0.1:11434",
-        "model": chosen,
-        "pack": pack_name,
-        "persona": persona_id,
-        "includeBase": True,
-        "messages": [{"role": "user", "content": text}],
-        "context_json": json.dumps(
-            {
-                "persona": persona_id,
-                "debug_eval": True,
-                "has_vision_image": False,
-                "images_in_request": False,
-            },
-            ensure_ascii=False,
-        ),
-        "skills": [],
-    }
-    ok, data, ms_call = _http_json(
-        f"{base}/API/AssistentChat",
-        method="POST",
-        body=payload,
-        timeout=t_chat,
-    )
-    return _finish_chat_eval(
-        ok=ok,
-        data=data,
-        ms_call=ms_call,
-        ms_total=(time.perf_counter() - t0) * 1000,
-        via=via,
-        session_ms=ms_sess,
-        message=text,
-        persona=persona_id,
-        pack=pack_name,
-        model=chosen,
-        preferred=preferred,
-        hints=hints,
-        errors=errors,
-        timeout=t_chat,
     )
 
 
@@ -623,157 +540,6 @@ def _finish_chat_eval(
     return out
 
 
-def _assistent_chat_eval_via_ssh(
-    cfg: Config,
-    *,
-    message: str,
-    persona: str,
-    pack: str,
-    model: str | None,
-    timeout: float,
-    hints: list[str],
-    t0: float,
-) -> dict[str, Any]:
-    state = load_state()
-    if not debug_checks.ssh_ready(cfg, state):
-        return {
-            **debug_checks._ssh_fail(state.phase),
-            "ok": False,
-            "ms": round((time.perf_counter() - t0) * 1000, 1),
-            "message": message,
-            "persona": persona,
-            "pack": pack,
-            "model": model,
-            "errors": [debug_checks.SSH_UNAVAILABLE],
-            "hints": hints + ["туннель SwarmUI закрыт и SSH нет — gpu-rent tunnel / up"],
-        }
-    host = debug_checks.ssh_host(state)
-    assert host is not None
-    # Resolve model on laptop if possible (roles via SSH fs), else let remote ListModels.
-    chosen = model
-    preferred = None
-    if not chosen:
-        roles = collect_assistent_roles(cfg)
-        preferred = (roles.get("roles") or {}).get("default_chat")
-        if isinstance(preferred, str) and preferred.strip():
-            chosen = preferred.strip()
-            preferred = chosen
-
-    payload_model = chosen or ""
-    # Remote script: GetNewSession → optional ListModels → AssistentChat
-    script = f'''
-import json, urllib.request, time
-MSG = {json.dumps(message, ensure_ascii=False)}
-PERSONA = {json.dumps(persona, ensure_ascii=False)}
-PACK = {json.dumps(pack, ensure_ascii=False)}
-MODEL = {json.dumps(payload_model, ensure_ascii=False)}
-TIMEOUT = {float(timeout)}
-
-def post(path, payload, timeout=TIMEOUT):
-    t0 = time.time()
-    req = urllib.request.Request(
-        "http://127.0.0.1:7801" + path,
-        data=json.dumps(payload).encode(),
-        headers={{"Content-Type": "application/json"}},
-        method="POST",
-    )
-    try:
-        with urllib.request.urlopen(req, timeout=timeout) as resp:
-            raw = resp.read().decode("utf-8", "replace")
-            ms = (time.time() - t0) * 1000
-            try:
-                data = json.loads(raw)
-            except Exception:
-                data = raw[:500]
-            return True, data, ms
-    except Exception as e:
-        return False, str(e), (time.time() - t0) * 1000
-
-ok, sess, ms = post("/API/GetNewSession", {{}}, timeout=20)
-if not ok or not isinstance(sess, dict) or not sess.get("session_id"):
-    print(json.dumps({{"ok": False, "error": sess, "session_ms": ms, "via": "ssh"}}))
-    raise SystemExit(0)
-sid = sess["session_id"]
-preferred = None
-model = MODEL.strip()
-if not model:
-    cok, mdata, _ = post("/API/AssistentListModels", {{
-        "session_id": sid, "baseUrl": "http://127.0.0.1:11434"
-    }}, timeout=25)
-    if cok and isinstance(mdata, dict):
-        preferred = mdata.get("preferred")
-        model = (preferred or "").strip()
-if not model:
-    print(json.dumps({{
-        "ok": False, "error": "no model", "session_ms": ms, "via": "ssh",
-        "preferred": preferred
-    }}))
-    raise SystemExit(0)
-payload = {{
-    "session_id": sid,
-    "baseUrl": "http://127.0.0.1:11434",
-    "model": model,
-    "pack": PACK,
-    "persona": PERSONA,
-    "includeBase": True,
-    "messages": [{{"role": "user", "content": MSG}}],
-    "context_json": json.dumps({{
-        "persona": PERSONA, "debug_eval": True,
-        "has_vision_image": False, "images_in_request": False
-    }}, ensure_ascii=False),
-    "skills": [],
-}}
-cok, data, cms = post("/API/AssistentChat", payload, timeout=TIMEOUT)
-print(json.dumps({{
-    "ok": cok, "data": data, "chat_ms": cms, "session_ms": ms,
-    "via": "ssh", "model": model, "preferred": preferred or model
-}}, ensure_ascii=False))
-'''
-    try:
-        from gpu_rent.ssh_ops import run_ssh
-
-        out = run_ssh(
-            cfg,
-            host,
-            "python3 - <<'PY'\n" + script + "\nPY",
-            check=False,
-            timeout=int(timeout) + 40,
-        ).strip()
-        remote = json.loads(out.splitlines()[-1])
-    except Exception as exc:
-        return {
-            "ok": False,
-            "via": "ssh",
-            "ms": round((time.perf_counter() - t0) * 1000, 1),
-            "message": message,
-            "persona": persona,
-            "pack": pack,
-            "model": model,
-            "error": str(exc)[:240],
-            "errors": [str(exc)[:240]],
-            "hints": hints,
-        }
-
-    data = remote.get("data")
-    ok = bool(remote.get("ok"))
-    return _finish_chat_eval(
-        ok=ok,
-        data=data if ok or isinstance(data, dict) else remote.get("error") or data,
-        ms_call=float(remote.get("chat_ms") or 0),
-        ms_total=(time.perf_counter() - t0) * 1000,
-        via="ssh",
-        session_ms=float(remote.get("session_ms") or 0),
-        message=message,
-        persona=persona,
-        pack=pack,
-        model=remote.get("model") or chosen,
-        preferred=remote.get("preferred") or preferred,
-        hints=hints,
-        errors=[],
-        timeout=timeout,
-    )
-
-
 _FS_CACHE: tuple[float, str, dict[str, Any]] | None = None
 _FS_TTL = 8.0
 
@@ -1366,6 +1132,15 @@ def collect_assistent_deep(
             "memory_broken": "GET /assistent/memory + /assistent/api (ListMemory)",
             "personas_missing": "GET /assistent/overlay + ListPersonas in /assistent/api",
             "chat_smoke": "GET /assistent/api?chat_smoke=1",
-            "chat_eval": "POST /assistent/chat-eval (opt-in AssistentChat; not in /snapshot)",
+            "chat_eval": "POST /assistent/chat-eval (one-shot; not in /snapshot)",
+            "session": (
+                "POST /assistent/session → POST .../chat (multi-turn + trace) "
+                "→ DELETE /assistent/session/{id}"
+            ),
+            "diagnose": "GET /assistent/diagnose or /assistent?logs=1",
+            "compact_context_gap": (
+                "AssistentChat does not return compactContext; "
+                "trace fills from GetConfig Exact + synthetic context_json (see trace.gaps)"
+            ),
         },
     }

tests/test_assistent_session.py

@@ -0,0 +1,419 @@
+"""Unit tests for Assistent patch extract + in-memory session store (no GPU)."""
+
+from __future__ import annotations
+
+import time
+
+from gpu_rent.debug_assistent import extract_assistent_patch
+from gpu_rent import debug_assistent_session as das
+
+
+def test_extract_assistent_patch_last_fence():
+    text = (
+        "Сначала prose.\n"
+        "```json\n"
+        '{"steps": 8, "cfg": 1}\n'
+        "```\n"
+        "ещё текст\n"
+        "```json\n"
+        '{"prompt": "a cat", "actions": ["generate"], "aspect": "16:9"}\n'
+        "```\n"
+    )
+    out = extract_assistent_patch(text)
+    assert out["patch"] is not None
+    assert out["patch"]["prompt"] == "a cat"
+    assert out["patch"]["generate"] is True
+    assert out["patch"]["aspect"] == "16:9"
+    assert "Сначала prose" in out["prose"]
+
+
+def test_extract_assistent_patch_ignores_non_patch_json():
+    text = 'hello\n```json\n{"foo": 1}\n```\n'
+    out = extract_assistent_patch(text)
+    assert out["patch"] is None
+
+
+def test_analyze_exact_merge_fills_omitted_params():
+    patch = {"actions": ["generate"], "prompt": "x"}
+    exact = {
+        "generation": {"profile": "turbo", "steps": 8, "cfg": 1, "sigma_shift": 1.15},
+        "profiles": {
+            "turbo": {"steps": 8, "cfg": 1, "sigma_shift": 1.15},
+            "raw": {"steps": 28, "cfg": 4.5, "sigma_shift": 1.15},
+        },
+    }
+    result = das.analyze_exact_merge(
+        patch,
+        krea_profile="turbo",
+        exact=exact,
+    )
+    assert result["wants_generate"] is True
+    assert result["would_fill"]["steps"] == 8
+    assert result["would_fill"]["cfg"] == 1
+    assert "steps omitted" in " ".join(result["hints"])
+
+
+def test_analyze_exact_merge_forces_foreign_leftovers():
+    patch = {"generate": True, "steps": 20, "cfg": 7}
+    exact = {
+        "profiles": {"turbo": {"steps": 8, "cfg": 1, "sigma_shift": 1.15}},
+        "generation": {"profile": "turbo"},
+    }
+    result = das.analyze_exact_merge(
+        patch, krea_profile="turbo", exact=exact, user_param_intent=False
+    )
+    assert result["would_force_to_exact"]["steps"]["to"] == 8
+    assert result["would_force_to_exact"]["cfg"]["to"] == 1
+
+
+def test_summarize_compact_context_sizes():
+    summary = das.summarize_compact_context(
+        {"persona": "neutral", "krea_profile": "turbo", "recommended_params": {"steps": 8}}
+    )
+    assert "persona" in summary["keys"]
+    assert summary["chars"] > 0
+    assert summary["token_ish"] >= 1
+    assert "does not return compactContext" in summary["note"]
+
+
+def test_session_store_create_get_delete_without_swarm(monkeypatch):
+    store = das.reset_session_store_for_tests()
+
+    class Cfg:
+        swarmui_local_port = 17801
+
+    # No local tunnel
+    monkeypatch.setattr(das, "_swarm_base", lambda cfg: (None, "none"))
+    monkeypatch.setattr(
+        das, "_resolve_chat_model", lambda cfg, base, sid, model=None: (None, None)
+    )
+
+    created = das.create_debug_session(
+        Cfg(),  # type: ignore[arg-type]
+        persona="neutral",
+        pack="ordinary",
+        probe_config=False,
+    )
+    assert created["ok"] is True
+    sid = created["debug_session_id"]
+    assert sid in store.list_ids()
+
+    got = das.get_debug_session(sid)
+    assert got["ok"] is True
+    assert got["session"]["persona"] == "neutral"
+    assert got["session"]["turn_count"] == 0
+
+    deleted = das.delete_debug_session(sid)
+    assert deleted["ok"] is True
+    assert das.get_debug_session(sid)["ok"] is False
+
+
+def test_session_store_ttl_eviction(monkeypatch):
+    store = das.reset_session_store_for_tests()
+    store.ttl_sec = 0.05
+
+    class Cfg:
+        swarmui_local_port = 17801
+
+    monkeypatch.setattr(das, "_swarm_base", lambda cfg: (None, "none"))
+    monkeypatch.setattr(
+        das, "_resolve_chat_model", lambda cfg, base, sid, model=None: ("m", "m")
+    )
+
+    created = das.create_debug_session(Cfg(), probe_config=False)  # type: ignore[arg-type]
+    sid = created["debug_session_id"]
+    time.sleep(0.08)
+    assert das.get_debug_session(sid)["ok"] is False
+
+
+def test_session_store_max_cap(monkeypatch):
+    store = das.reset_session_store_for_tests()
+    store.max_sessions = 2
+
+    class Cfg:
+        swarmui_local_port = 17801
+
+    monkeypatch.setattr(das, "_swarm_base", lambda cfg: (None, "none"))
+    monkeypatch.setattr(
+        das, "_resolve_chat_model", lambda cfg, base, sid, model=None: ("m", "m")
+    )
+
+    ids = []
+    for _ in range(3):
+        created = das.create_debug_session(Cfg(), probe_config=False)  # type: ignore[arg-type]
+        ids.append(created["debug_session_id"])
+        time.sleep(0.01)
+    assert len(store.list_ids()) == 2
+    assert ids[0] not in store.list_ids()
+
+
+def test_clamp_message():
+    short, trunc = das.clamp_message("hi", max_chars=10)
+    assert short == "hi" and trunc is False
+    long, trunc = das.clamp_message("x" * 20, max_chars=10)
+    assert len(long) == 10 and trunc is True
+
+
+def test_multi_turn_chat_keeps_history(monkeypatch):
+    store = das.reset_session_store_for_tests()
+
+    class Cfg:
+        swarmui_local_port = 17801
+
+    monkeypatch.setattr(
+        das, "_swarm_base", lambda cfg: ("http://127.0.0.1:17801", "local")
+    )
+    monkeypatch.setattr(das, "_session_id", lambda base: ("swarm-1", None, 3.0))
+    monkeypatch.setattr(
+        das,
+        "_resolve_chat_model",
+        lambda cfg, base, sid, model=None: ("qwen", "qwen"),
+    )
+
+    payloads: list[dict] = []
+
+    def fake_http(url, *, method="GET", body=None, timeout=12.0):
+        if url.endswith("/API/AssistentGetConfig"):
+            return True, {
+                "success": True,
+                "persona": "neutral",
+                "exact": {
+                    "profiles": {"turbo": {"steps": 8, "cfg": 1, "sigma_shift": 1.15}},
+                    "generation": {"profile": "turbo"},
+                },
+                "skills": [],
+                "enabled_skills": [],
+            }, 5.0
+        if url.endswith("/API/AssistentChat"):
+            payloads.append(body or {})
+            n = len(payloads)
+            reply = f'turn{n}\n```json\n{{"steps": {4 + n}, "actions": ["generate"]}}\n```'
+            return True, {"success": True, "reply": reply, "model": "qwen"}, 20.0
+        return False, "unexpected " + url, 1.0
+
+    monkeypatch.setattr(das, "_http_json", fake_http)
+
+    created = das.create_debug_session(Cfg(), probe_config=True)  # type: ignore[arg-type]
+    sid = created["debug_session_id"]
+    t1 = das.chat_debug_session(Cfg(), sid, message="first")  # type: ignore[arg-type]
+    t2 = das.chat_debug_session(Cfg(), sid, message="second")  # type: ignore[arg-type]
+    assert t1["ok"] and t2["ok"]
+    assert len(payloads) == 2
+    assert len(payloads[0]["messages"]) == 1
+    assert len(payloads[1]["messages"]) == 3  # user, assistant, user
+    assert payloads[1]["messages"][0]["content"] == "first"
+    assert payloads[1]["messages"][2]["content"] == "second"
+    assert t2["trace"]["patch"]["steps"] == 6
+    assert t2["trace"]["exact_merge"]["wants_generate"] is True
+    assert "gaps" in t2["trace"]
+    assert store.get(sid).turn_count == 2
+    das.delete_debug_session(sid)
+
+
+def test_build_chat_trace_soft_sqlite():
+    reply = 'ok\n```json\n{"steps": 8, "actions": ["generate"]}\n```'
+    trace = das.build_chat_trace(
+        ok=True,
+        timings_ms={"chat": 12.0, "total": 20.0},
+        model="qwen",
+        preferred="qwen",
+        persona="neutral",
+        pack="ordinary",
+        context={"krea_profile": "turbo", "recommended_params": {"steps": 8, "cfg": 1}},
+        skills=[],
+        config_probe={
+            "ok": True,
+            "_exact": {
+                "profiles": {"turbo": {"steps": 8, "cfg": 1, "sigma_shift": 1.15}}
+            },
+            "data": {"enabled_skills": []},
+        },
+        reply=reply,
+        response={
+            "reply": reply,
+            "system_chars": 1000,
+            "system_layers": {"core": 100, "pack": 50},
+            "error": "SaveChat sqlite fail",
+        },
+        errors=[],
+        warnings=[],
+        hints=[],
+    )
+    assert trace["ok"] is True
+    assert trace["patch"]["steps"] == 8
+    assert trace["exact_merge"]["wants_generate"] is True
+    assert any("Sqlite" in e or "sqlite" in e.lower() for e in trace["errors"])
+    assert "gaps" in trace and len(trace["gaps"]) >= 3
+
+
+def test_openapi_lists_session_paths(monkeypatch, tmp_path):
+    import socket
+    import urllib.request
+
+    from gpu_rent import debug_api
+    from gpu_rent.config import load_config
+
+    # Reuse same env helper pattern as test_debug_api
+    monkeypatch.setenv("HOME", str(tmp_path))
+    monkeypatch.setenv("USERPROFILE", str(tmp_path))
+    monkeypatch.chdir(tmp_path)
+    for key, val in {
+        "OS_AUTH_URL": "https://example/v3",
+        "OS_USER_DOMAIN_NAME": "default",
+        "OS_USERNAME": "u",
+        "OS_PASSWORD": "secret-password",
+        "OS_PROJECT_ID": "proj",
+        "OS_REGION_NAME": "ru-7",
+        "GPU_RENT_AZ": "ru-7a",
+        "CIVITAI_API_TOKEN": "civ",
+        "HF_TOKEN": "hf",
+        "GIT_TOKEN": "git",
+        "SELECTEL_API_TOKEN": "sel",
+    }.items():
+        monkeypatch.setenv(key, val)
+    cfg = load_config(require_auth=True)
+
+    sock = socket.socket()
+    sock.bind(("127.0.0.1", 0))
+    port = sock.getsockname()[1]
+    sock.close()
+    srv = debug_api.start_debug_server(cfg, port=port, print_urls=False)
+    assert srv is not None
+    try:
+        with urllib.request.urlopen(
+            f"http://127.0.0.1:{port}/openapi.json", timeout=5
+        ) as resp:
+            doc = __import__("json").loads(resp.read().decode())
+        assert "/assistent/session" in doc["paths"]
+        assert "/assistent/session/{id}/chat" in doc["paths"]
+        assert "/assistent/diagnose" in doc["paths"]
+        assert "VRAM" in doc["info"]["description"]
+    finally:
+        debug_api.stop_debug_server(srv)
+
+
+def test_http_session_chat_mocked(monkeypatch, tmp_path):
+    """POST session → chat → GET → DELETE via Debug API (mocked Swarm)."""
+    import json
+    import socket
+    import urllib.request
+
+    from gpu_rent import debug_api
+    from gpu_rent.config import load_config
+
+    monkeypatch.setenv("HOME", str(tmp_path))
+    monkeypatch.setenv("USERPROFILE", str(tmp_path))
+    monkeypatch.chdir(tmp_path)
+    for key, val in {
+        "OS_AUTH_URL": "https://example/v3",
+        "OS_USER_DOMAIN_NAME": "default",
+        "OS_USERNAME": "u",
+        "OS_PASSWORD": "secret",
+        "OS_PROJECT_ID": "proj",
+        "OS_REGION_NAME": "ru-7",
+        "GPU_RENT_AZ": "ru-7a",
+        "CIVITAI_API_TOKEN": "civ",
+        "HF_TOKEN": "hf",
+        "GIT_TOKEN": "git",
+        "SELECTEL_API_TOKEN": "sel",
+    }.items():
+        monkeypatch.setenv(key, val)
+    cfg = load_config(require_auth=True)
+    das.reset_session_store_for_tests()
+
+    monkeypatch.setattr(
+        das, "_swarm_base", lambda cfg: ("http://127.0.0.1:17801", "local")
+    )
+    monkeypatch.setattr(das, "_session_id", lambda base: ("swarm-sid", None, 3.0))
+    monkeypatch.setattr(
+        das,
+        "_resolve_chat_model",
+        lambda cfg, base, sid, model=None: ("qwen3-vl:8b", "qwen3-vl:8b"),
+    )
+
+    def fake_http(url, *, method="GET", body=None, timeout=12.0):
+        if url.endswith("/API/AssistentGetConfig"):
+            return (
+                True,
+                {
+                    "exact": {
+                        "profiles": {"turbo": {"steps": 8, "cfg": 1, "sigma_shift": 1.15}},
+                        "generation": {"profile": "turbo"},
+                    },
+                    "enabled_skills": [],
+                },
+                4.0,
+            )
+        if url.endswith("/API/AssistentChat"):
+            n = len((body or {}).get("messages") or [])
+            reply = (
+                f"turn-{n}\n"
+                '```json\n{"prompt":"cat","steps":8,"cfg":1,"actions":["generate"]}\n```'
+            )
+            return True, {"reply": reply, "system_chars": 1200, "model": "qwen3-vl:8b"}, 20.0
+        return False, "unexpected " + url, 1.0
+
+    monkeypatch.setattr(das, "_http_json", fake_http)
+
+    sock = socket.socket()
+    sock.bind(("127.0.0.1", 0))
+    port = sock.getsockname()[1]
+    sock.close()
+    srv = debug_api.start_debug_server(cfg, port=port, print_urls=False)
+    assert srv is not None
+    base = f"http://127.0.0.1:{port}"
+
+    def post(path: str, payload: dict) -> dict:
+        req = urllib.request.Request(
+            base + path,
+            data=json.dumps(payload).encode(),
+            headers={"Content-Type": "application/json"},
+            method="POST",
+        )
+        with urllib.request.urlopen(req, timeout=10) as resp:
+            return json.loads(resp.read().decode())
+
+    def get(path: str) -> dict:
+        with urllib.request.urlopen(base + path, timeout=10) as resp:
+            return json.loads(resp.read().decode())
+
+    def delete(path: str) -> dict:
+        req = urllib.request.Request(base + path, method="DELETE")
+        with urllib.request.urlopen(req, timeout=10) as resp:
+            return json.loads(resp.read().decode())
+
+    try:
+        created = post(
+            "/assistent/session",
+            {
+                "persona": "neutral",
+                "pack": "ordinary",
+                "context": {"krea_profile": "turbo"},
+            },
+        )
+        assert created["ok"] is True
+        sid = created["debug_session_id"]
+
+        turn1 = post(f"/assistent/session/{sid}/chat", {"message": "какие steps?"})
+        assert turn1["ok"] is True
+        assert turn1["trace"]["patch"]["steps"] == 8
+        assert turn1["trace"]["exact_merge"]["wants_generate"] is True
+        assert "compact_context" in turn1["trace"]
+        assert turn1["turn_count"] == 1
+
+        turn2 = post(f"/assistent/session/{sid}/chat", {"message": "ещё раз"})
+        assert turn2["ok"] is True
+        assert turn2["turn_count"] == 2
+        assert turn2["session"]["turn_count"] == 2
+
+        got = get(f"/assistent/session/{sid}")
+        assert got["ok"] is True
+        assert got["session"]["turn_count"] == 2
+        assert got["last_trace"]["ok"] is True
+
+        deleted = delete(f"/assistent/session/{sid}")
+        assert deleted["ok"] is True
+        gone = get(f"/assistent/session/{sid}")
+        assert gone["ok"] is False
+    finally:
+        debug_api.stop_debug_server(srv)

tests/test_debug_api.py

@@ -193,24 +193,41 @@ def test_extract_assistent_patch():
 
 def test_chat_eval_mocked_http(monkeypatch, tmp_path):
     from gpu_rent import debug_assistent
+    from gpu_rent import debug_assistent_session as das
 
     _auth_env(monkeypatch, tmp_path)
     cfg = load_config(require_auth=True)
+    das.reset_session_store_for_tests()
 
     monkeypatch.setattr(
-        debug_assistent,
+        das,
         "_swarm_base",
         lambda cfg: ("http://127.0.0.1:17801", "local"),
     )
     monkeypatch.setattr(
-        debug_assistent,
+        das,
         "_session_id",
         lambda base: ("sess-1", None, 5.0),
     )
+    monkeypatch.setattr(
+        das,
+        "_resolve_chat_model",
+        lambda cfg, base, sid, model=None: ("qwen3-vl:8b", "qwen3-vl:8b"),
+    )
 
     calls: list[tuple[str, dict]] = []
 
     def fake_http(url, *, method="GET", body=None, timeout=12.0):
+        if url.endswith("/API/AssistentGetConfig"):
+            return True, {
+                "success": True,
+                "persona": "leonid",
+                "exact": {
+                    "profiles": {"turbo": {"steps": 8, "cfg": 1, "sigma_shift": 1.15}}
+                },
+                "skills": [],
+                "enabled_skills": [],
+            }, 8.0
         if url.endswith("/API/AssistentListModels"):
             return True, {"preferred": "qwen3-vl:8b", "models": ["qwen3-vl:8b"]}, 10.0
         if url.endswith("/API/AssistentChat"):
@@ -223,7 +240,7 @@ def test_chat_eval_mocked_http(monkeypatch, tmp_path):
             return True, {"success": True, "reply": reply, "model": "qwen3-vl:8b"}, 42.0
         return False, "unexpected " + url, 1.0
 
-    monkeypatch.setattr(debug_assistent, "_http_json", fake_http)
+    monkeypatch.setattr(das, "_http_json", fake_http)
 
     out = debug_assistent.run_assistent_chat_eval(
         cfg,
@@ -240,6 +257,7 @@ def test_chat_eval_mocked_http(monkeypatch, tmp_path):
     assert out["patch"]["cfg"] == 1
     assert out["patch"]["aspect"] == "3:4"
     assert "turbo" in (out["reply_prose"] or "").lower() or "turbo" in (out["reply"] or "").lower()
+    assert out.get("trace") and out["trace"].get("exact_merge")
     assert calls and calls[0][1]["session_id"] == "sess-1"
     assert calls[0][1]["messages"][0]["content"] == "какой checkpoint?"
     assert calls[0][1]["includeBase"] is True