Add opt-in /assistent/chat-eval for AssistentChat via Debug API.

Lets agents POST/GET a real Assistent turn (tunnel or SSH) without folding it into cheap /snapshot; documents VRAM/Sqlite side effects.

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

tests/test_debug_api.py

@@ -144,6 +144,13 @@ def test_debug_server_routes(monkeypatch, tmp_path):
         code, data = get("/openapi.json")
         assert "/assistent/extension" in data["paths"]
         assert "/assistent/api" in data["paths"]
+        assert "/assistent/chat-eval" in data["paths"]
+        assert "post" in data["paths"]["/assistent/chat-eval"]
+
+        code, data = get("/assistent/chat-eval")
+        assert "ok" in data
+        assert data.get("error") == debug_checks.SSH_UNAVAILABLE or not data.get("ok")
+        assert "hints" in data
 
         with pytest.raises(Exception):
             urllib.request.urlopen(f"http://127.0.0.1:{port}/nope", timeout=2)
@@ -151,6 +158,105 @@ def test_debug_server_routes(monkeypatch, tmp_path):
         debug_api.stop_debug_server(srv)
 
 
+def test_extract_assistent_patch():
+    from gpu_rent.debug_assistent import extract_assistent_patch
+
+    bare = extract_assistent_patch("просто текст без json")
+    assert bare["patch"] is None
+    assert "просто текст" in bare["prose"]
+
+    text = (
+        "Ок, вот параметры:\n"
+        "```json\n"
+        '{"prompt":"a cat","steps":4,"cfg":1.0,"aspect":"1:1","actions":["generate"]}\n'
+        "```\n"
+    )
+    got = extract_assistent_patch(text)
+    assert got["patch"] is not None
+    assert got["patch"]["steps"] == 4
+    assert got["patch"]["cfg"] == 1.0
+    assert got["patch"]["aspect"] == "1:1"
+    assert got["patch"]["generate"] is True
+    assert "Ок" in got["prose"]
+    assert "```" not in got["prose"]
+
+    # last fence wins
+    multi = (
+        "```json\n{\"steps\":1}\n```\n"
+        "mid\n"
+        "```json\n{\"steps\":8,\"cfg\":2}\n```"
+    )
+    got2 = extract_assistent_patch(multi)
+    assert got2["patch"]["steps"] == 8
+    assert got2["patch"]["cfg"] == 2
+
+
+def test_chat_eval_mocked_http(monkeypatch, tmp_path):
+    from gpu_rent import debug_assistent
+
+    _auth_env(monkeypatch, tmp_path)
+    cfg = load_config(require_auth=True)
+
+    monkeypatch.setattr(
+        debug_assistent,
+        "_swarm_base",
+        lambda cfg: ("http://127.0.0.1:17801", "local"),
+    )
+    monkeypatch.setattr(
+        debug_assistent,
+        "_session_id",
+        lambda base: ("sess-1", None, 5.0),
+    )
+
+    calls: list[tuple[str, dict]] = []
+
+    def fake_http(url, *, method="GET", body=None, timeout=12.0):
+        if url.endswith("/API/AssistentListModels"):
+            return True, {"preferred": "qwen3-vl:8b", "models": ["qwen3-vl:8b"]}, 10.0
+        if url.endswith("/API/AssistentChat"):
+            calls.append((url, body or {}))
+            reply = (
+                "Используй turbo.\n"
+                '```json\n{"prompt":"portrait","steps":4,"cfg":1,"aspect":"3:4",'
+                '"actions":["generate"]}\n```'
+            )
+            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)
+
+    out = debug_assistent.run_assistent_chat_eval(
+        cfg,
+        message="какой checkpoint?",
+        persona="leonid",
+        timeout=60,
+    )
+    assert out["ok"] is True
+    assert out["via"] == "local"
+    assert out["model"] == "qwen3-vl:8b"
+    assert out["preferred"] == "qwen3-vl:8b"
+    assert out["persona"] == "leonid"
+    assert out["patch"]["steps"] == 4
+    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 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
+
+
+def test_chat_eval_timeout_clamp():
+    from gpu_rent.debug_assistent import (
+        MAX_CHAT_EVAL_TIMEOUT,
+        MIN_CHAT_EVAL_TIMEOUT,
+        _clamp_chat_eval_timeout,
+    )
+
+    assert _clamp_chat_eval_timeout(5) == MIN_CHAT_EVAL_TIMEOUT
+    assert _clamp_chat_eval_timeout(9999) == MAX_CHAT_EVAL_TIMEOUT
+    assert _clamp_chat_eval_timeout("90") == 90.0
+
+
 def test_assistent_compact_and_roles_local(monkeypatch, tmp_path):
     from gpu_rent import debug_assistent