diff --git a/agents/acai/system.md b/agents/acai/system.md
index 0365538..570491c 100644
--- a/agents/acai/system.md
+++ b/agents/acai/system.md
@@ -128,6 +128,17 @@ Cuando creas una tabla con `enlace` (noticias, vacantes, blog), añade por defec
 
 NO añadas un campo "estado" calculado si ya tienes `visible` + fechas.
 
+## Schema (.ini.php) — NUNCA editar a mano
+
+Los `cms/data/schema/*.ini.php` se modifican **exclusivamente** con las tools de schema: `create_table`, `update_table_metadata`, `delete_table`, `reorder_tables`, `create_field`, `update_field`, `delete_field`, `reorder_fields`. NO uses `acai-write` ni `acai-line-replace` sobre estos archivos:
+
+- Saltarías validaciones (regex, tipos, etc.)
+- No invalidas la cache de schemas — el frontend ve schema viejo
+- No sincronizas con MySQL (no crea/borra columnas reales)
+- Puedes romper el formato INI con un escape mal puesto
+
+Para subcampos de un `multitext`, llama a `update_field` con `props.descriptionjson` como **string JSON** del array `[{id_campo, nombre_campo, tipo}, ...]`. La tool docu lo explica.
+
 ## Formularios embebidos en detalles
 
 Si un detalle necesita un formulario (postular, pedir info), embebe el módulo del formulario **dentro** de la sección general pasándole el `num`:
diff --git a/mcp-server/tools/records/getModuleConfigVars.js b/mcp-server/tools/records/getModuleConfigVars.js
index ed36e8e..85b986b 100644
--- a/mcp-server/tools/records/getModuleConfigVars.js
+++ b/mcp-server/tools/records/getModuleConfigVars.js
@@ -1,17 +1,32 @@
 import { z } from "zod";
-import { withAuth, getSessionCredentials } from "../../auth/index.js";
-import { handleApiResponse, handleToolError, validateRequired } from "../helpers/errorHandler.js";
-import { AcaiHttpClient } from "../helpers/acaiHttpClient.js";
+import { withAuth } from "../../auth/index.js";
 import { withAuthParams } from "../helpers/authSchema.js";
+import { handleToolError, validateRequired } from "../helpers/errorHandler.js";
+import { pythonGet } from "../helpers/pythonServerClient.js";
+import { getCurrentProjectInfo } from "../files/helpers.js";
+
+// get_module_config_vars
+//
+// Pasa por el server Python (/api/creator/get-module-vars) en lugar de ir
+// directamente al PHP. Asi obtenemos `varsMeta`: para cada variable del
+// modulo, su ubicacion fisica en `builder_custom` (recordNum + columna real).
+// Sin esto el agente solo conoce el nombre humano de la variable (ej.
+// `titulo`) y al hacer create_or_update_record adivina el campo y suele
+// fallar silenciosamente porque la columna real es algo como `title2`.
 
 export function registerGetModuleConfigVarsTool(server) {
     server.tool(
         "get_module_config_vars",
-        `Get the current configuration variable values for a module instance on a page record. Returns resolved values (text, HTML, etc.) for simple vars and arrays of objects for multi/repeater vars.
+        `Get the current configuration variable values for a module instance on a page record. Returns:
+
+- vars: resolved values (text, HTML, etc.) for simple vars and arrays for multi/repeater vars
+- varsMeta: per-var physical location { tableName: 'builder_custom', recordNum, fieldName, type }. USE THIS to know exactly which row + column to update with create_or_update_record. The variable's display name (e.g. 'titulo') is NOT the same as the physical column name (e.g. 'title2').
+- uploadFields: per-var upload location for upload_record_image / replace_record_image
+- moduleId, sectionId
 
 Required params:
-- tableName (string) without 'cms_' prefix
-- recordNum (number) record primary key ('num' field, never 'id')
+- tableName (string) without 'cms_' prefix (parent page table, e.g. 'apartados')
+- recordNum (number) parent record primary key ('num', never 'id')
 - sectionId (string) section ID of the module instance`,
         withAuthParams({
             tableName: z.string().describe("Parent table name (e.g. 'apartados')"),
@@ -19,44 +34,62 @@ Required params:
             sectionId: z.string().describe("Section ID of the module instance"),
         }),
         { readOnlyHint: true, destructiveHint: false },
-        withAuth(async ({ tableName, recordNum, sectionId }, extra) => {
+        withAuth(async ({ tableName, recordNum, sectionId }, _extra) => {
             try {
-                const validationError = validateRequired({ tableName, recordNum, sectionId }, ['tableName', 'recordNum', 'sectionId'], 'get_module_config_vars');
+                const validationError = validateRequired(
+                    { tableName, recordNum, sectionId },
+                    ["tableName", "recordNum", "sectionId"],
+                    "get_module_config_vars"
+                );
                 if (validationError) return validationError;
 
-                const sessionId = extra.sessionId;
-                const credentials = await getSessionCredentials(sessionId);
-                const payload = {
-                    tableName,
-                    recordNum,
-                    sectionId
-                };
+                const { projectSlug } = getCurrentProjectInfo();
+                const result = await pythonGet("/api/creator/get-module-vars", {
+                    project: projectSlug,
+                    table: tableName,
+                    num: recordNum,
+                    sectionId,
+                });
 
-                const response = await AcaiHttpClient.getModuleConfigVars(
-                    credentials,
-                    credentials.token,
-                    credentials.tokenHash,
-                    payload
-                );
-
-                const apiError = handleApiResponse(response.data, 'get_module_config_vars');
-                if (apiError) return apiError;
+                // El endpoint Python devuelve la respuesta del PHP envuelta:
+                // { success, data: { moduleId, vars, configVars, uploadFields, varsMeta, ... } }
+                // o directamente las claves en raiz. Normalizamos.
+                const inner = result?.data && typeof result.data === "object" ? result.data : result;
+                if (!inner || inner.success === false) {
+                    return {
+                        content: [{
+                            type: "text",
+                            text: JSON.stringify({
+                                success: false,
+                                error: inner?.error || "Could not read module config vars",
+                            }),
+                        }],
+                        isError: true,
+                    };
+                }
 
                 return {
                     content: [{
                         type: "text",
                         text: JSON.stringify({
                             success: true,
-                            action: 'get_module_config_vars',
+                            action: "get_module_config_vars",
                             tableName,
                             recordNum,
                             sectionId,
-                            data: response.data?.data ?? response.data
-                        }, null, 2)
-                    }]
+                            data: {
+                                moduleId: inner.moduleId,
+                                sectionId: inner.sectionId,
+                                vars: inner.vars || {},
+                                varsMeta: inner.varsMeta || {},
+                                uploadFields: inner.uploadFields || {},
+                                configVars: inner.configVars || {},
+                            },
+                        }, null, 2),
+                    }],
                 };
             } catch (error) {
-                return handleToolError(error, 'get_module_config_vars', { tableName, recordNum, sectionId });
+                return handleToolError(error, "get_module_config_vars", { tableName, recordNum, sectionId });
             }
         })
     );
diff --git a/mcp-server/tools/tables/createField.js b/mcp-server/tools/tables/createField.js
index 03374b5..eeb9df0 100644
--- a/mcp-server/tools/tables/createField.js
+++ b/mcp-server/tools/tables/createField.js
@@ -31,7 +31,48 @@ Field types:
  - separator: visual separator in the form (no data column)
 
 'initialProps' is optional; use it to override defaults (e.g. {isRequired:1, maxLength:100}).
-Table names WITHOUT 'cms_' prefix. Primary key is always 'num'.`,
+Table names WITHOUT 'cms_' prefix. Primary key is always 'num'.
+
+MULTITEXT FIELDS (type='multitext') — initialProps shape:
+A multitext field is a repeater of N sub-fields per row. Each row in the
+admin form gives the user a set of sub-fields to fill (e.g. a list of FAQ
+items where each one has 'question' + 'answer' + 'category').
+ - tablaAuxiliar (string, opcional): nombre de tabla auxiliar para el repeater.
+ - descriptionjson (REQUIRED): JSON STRING (not object) with the array of
+   sub-fields. Each sub-field is an object with:
+   - id_campo (string): clave tecnica, slug. Estable: NO se cambia despues.
+   - nombre_campo (string): etiqueta visible en la UI.
+   - tipo (string '0'..'5'): '0'=texto, '1'=tabla CMS, '3'=fecha, '4'=color, '5'=icono.
+   - When tipo='1': also pass tabla (target table without 'cms_'),
+     campo_valor (default 'num'), campo_muestra (label field).
+   Example (correct):
+     "descriptionjson": "[{\\"id_campo\\":\\"pregunta\\",\\"nombre_campo\\":\\"Pregunta\\",\\"tipo\\":\\"0\\"},{\\"id_campo\\":\\"respuesta\\",\\"nombre_campo\\":\\"Respuesta\\",\\"tipo\\":\\"0\\"}]"
+   ⚠ It MUST be a JSON-encoded string. If you pass an object the backend
+     rejects it. Strings, single-quoted JSON, or other formats break the editor.
+
+LIST FIELDS (type='list') — initialProps shape:
+ - listType (REQUIRED): one of 'pulldown' | 'radios' | 'pulldownMulti' | 'checkboxes'.
+   NOT 'select' nor 'dropdown' — use 'pulldown'.
+ - optionsType (REQUIRED): one of 'text' | 'table' | 'query'.
+
+ - When optionsType='text': pass 'optionsText' as a SINGLE string with one
+   option per line, separated by REAL NEWLINE CHARACTERS ('\\n' in JSON).
+   Each line is either 'value|Label' (preferred) or just 'label' (value=label).
+   ⚠ Do NOT separate options with commas. Commas inside an option are valid
+   data — Acai uses '\\n' as the option delimiter, period.
+   Example (correct):
+     "optionsText": "indefinido|Indefinido\\ntemporal|Temporal\\nfreelance|Freelance"
+   Example (WRONG, will store all 4 as a single option):
+     "optionsText": "Indefinido,Temporal,Prácticas,Freelance"
+
+ - When optionsType='table': pass 'optionsTablename' (target table without cms_),
+   'optionsValueField' (default 'num'), 'optionsLabelField'. Optional 'filterField'
+   is a SQL WHERE clause without the WHERE keyword (e.g. "visible=1").
+
+ - When optionsType='query': pass 'optionsQuery' as raw SQL. Acai uses POSITIONAL
+   columns: column 0 is the value, column 1 is the label. So write
+   "SELECT num, titulo FROM cms_xxx WHERE active=1" — the 'AS value/label'
+   aliases have NO effect.`,
         withAuthParams({
             tableName: z.string().describe("Table name without 'cms_' prefix"),
             fieldName: z.string().describe("New field name (SQL-safe identifier)"),
diff --git a/mcp-server/tools/tables/updateField.js b/mcp-server/tools/tables/updateField.js
index eded563..2dcde2c 100644
--- a/mcp-server/tools/tables/updateField.js
+++ b/mcp-server/tools/tables/updateField.js
@@ -29,7 +29,27 @@ Destructive cases:
    drops HTML). The backend returns 'warnings' in the response — surface them
    to the user.
 
-Table names WITHOUT 'cms_' prefix.`,
+Table names WITHOUT 'cms_' prefix.
+
+MULTITEXT FIELDS — when 'props' touches multitext config:
+ - descriptionjson: JSON STRING (not object) with the array of sub-fields.
+   Each: {id_campo, nombre_campo, tipo} where tipo is '0'(texto)|'1'(tabla)|
+   '3'(fecha)|'4'(color)|'5'(icono). For tipo='1' also include tabla,
+   campo_valor, campo_muestra.
+   ⚠ MUST be JSON-encoded string. Backend rejects objects directly.
+   Example: "descriptionjson":"[{\\"id_campo\\":\\"pregunta\\",\\"nombre_campo\\":\\"Pregunta\\",\\"tipo\\":\\"0\\"}]"
+
+LIST FIELDS — when 'props' touches list config:
+ - listType: 'pulldown' | 'radios' | 'pulldownMulti' | 'checkboxes' (NOT 'select').
+ - optionsType: 'text' | 'table' | 'query'.
+ - optionsText (for optionsType='text'): one option per LINE, separated by
+   '\\n' (real newline). Each line is 'value|Label' or just 'label'.
+   ⚠ Do NOT use commas as the option separator — commas are valid inside
+   a label. Example (correct): "indef|Indefinido\\ntemp|Temporal".
+ - optionsTablename / optionsValueField / optionsLabelField / filterField
+   for optionsType='table'.
+ - optionsQuery for optionsType='query' — column 0 is the value, column 1
+   the label (positional, 'AS value/label' aliases are ignored).`,
         withAuthParams({
             tableName: z.string().describe("Table name without 'cms_' prefix"),
             fieldName: z.string().describe("Current field name"),
diff --git a/src/adapters/claude_adapter.py b/src/adapters/claude_adapter.py
index c8842e6..7890186 100644
--- a/src/adapters/claude_adapter.py
+++ b/src/adapters/claude_adapter.py
@@ -5,6 +5,8 @@ from __future__ import annotations
 import asyncio
 import json
 import logging
+import re
+import uuid
 from typing import Any, AsyncIterator
 
 import anthropic
@@ -15,6 +17,75 @@ from .base import ModelAdapter, ModelConfig, ModelResponse, StreamChunk
 logger = logging.getLogger(__name__)
 
 
+# Algunos fine-tunes (sobre todo MiniMax) ocasionalmente emiten las tool calls
+# como texto literal en lugar de usar los `tool_use` blocks nativos:
+#   <minimax:tool_call>
+#     <invoke name="acai_code__acai_view">
+#       <parameter name="file_path">...</parameter>
+#     </invoke>
+#   </minimax:tool_call>
+#
+# Cuando eso pasa el orquestador ve "texto" y la tool nunca se ejecuta — el
+# usuario ve el XML crudo en el chat. Detectamos y convertimos a tool_use
+# sintetico mientras streameamos. Es un parche defensivo: el caso normal
+# (tool_use blocks) sigue por el camino estandar.
+_TOOL_CALL_OPEN_RE = re.compile(r"<(?:minimax:tool_call|invoke\s+name)", re.IGNORECASE)
+_INVOKE_RE = re.compile(
+    r"<invoke\s+name=\"([^\"]+)\"\s*>(.*?)</invoke>",
+    re.IGNORECASE | re.DOTALL,
+)
+_PARAM_RE = re.compile(
+    r"<parameter\s+name=\"([^\"]+)\"\s*>(.*?)</parameter>",
+    re.IGNORECASE | re.DOTALL,
+)
+
+
+def _safe_emit_split(buf: str) -> str:
+    """Devuelve el prefijo del buffer que es seguro emitir como texto sin
+    perder un posible inicio de tag XML que esta llegando fragmentado.
+
+    Mantenemos en hold los ultimos 30 chars si terminan con `<` o con un
+    prefijo parcial de `<minimax:tool_call` / `<invoke`. Si el buffer es
+    largo y no termina con `<`, todo es seguro.
+    """
+    if not buf:
+        return ""
+    # Buscar el ultimo `<` y comprobar si lo que sigue puede ser apertura.
+    idx = buf.rfind("<")
+    if idx == -1:
+        return buf
+    tail = buf[idx:]
+    # Si el tail ya tiene `>` cerrado, es un tag normal — emitir todo.
+    if ">" in tail:
+        return buf
+    # Si el tail puede ser inicio de tool_call/invoke, retenerlo.
+    candidates = ("<minimax:tool_call", "<invoke")
+    for cand in candidates:
+        if cand.startswith(tail.lower()) or tail.lower().startswith(cand[:len(tail)].lower()):
+            return buf[:idx]
+    # No coincide con ninguna apertura sospechosa — emitir todo.
+    return buf
+
+
+def _parse_xml_tool_calls(text: str) -> list[dict[str, Any]]:
+    """Extrae tool calls del texto. Devuelve lista de {id, name, arguments}.
+    Si no encuentra patrones validos devuelve []."""
+    calls = []
+    for m in _INVOKE_RE.finditer(text):
+        name = m.group(1).strip()
+        body = m.group(2)
+        args = {}
+        for p in _PARAM_RE.finditer(body):
+            args[p.group(1).strip()] = p.group(2).strip()
+        if name:
+            calls.append({
+                "id": "xml_{}".format(uuid.uuid4().hex[:12]),
+                "name": name,
+                "arguments": args,
+            })
+    return calls
+
+
 # Errores transitorios del proxy del modelo (MiniMax/Anthropic). Reintentamos
 # con backoff exponencial: 1s, 3s, 9s. 529 es overloaded_error de Anthropic;
 # 429 rate-limit; 503 service unavailable.
@@ -98,6 +169,14 @@ class ClaudeAdapter(ModelAdapter):
                     current_tool_name = ""
                     accumulated_args = ""
                     input_tokens = 0
+                    # Buffer + flag para detectar XML tool calls inline (MiniMax).
+                    # En modo "text", emitimos delta directamente. Si vemos `<invoke`
+                    # o `<minimax:tool_call`, pasamos a modo "buffer" y dejamos de
+                    # emitir hasta cerrar el bloque o terminar el mensaje. Al final
+                    # parseamos y emitimos un tool_use sintetico.
+                    text_buffer = ""
+                    in_xml_capture = False
+                    xml_buffer = ""
 
                     async for event in stream:
                         yielded_any = True
@@ -121,7 +200,29 @@ class ClaudeAdapter(ModelAdapter):
                         if event.type == "content_block_delta":
                             delta = event.delta
                             if delta.type == "text_delta":
-                                yield StreamChunk(delta=delta.text)
+                                text_buffer += delta.text
+                                if in_xml_capture:
+                                    xml_buffer += delta.text
+                                else:
+                                    # Detectar inicio del bloque XML antes de emitir.
+                                    m = _TOOL_CALL_OPEN_RE.search(text_buffer)
+                                    if m:
+                                        # Emitir el texto previo al match (texto
+                                        # legitimo que el modelo escribio antes del XML).
+                                        prev = text_buffer[:m.start()]
+                                        if prev:
+                                            yield StreamChunk(delta=prev)
+                                        in_xml_capture = True
+                                        xml_buffer = text_buffer[m.start():]
+                                        text_buffer = ""
+                                    else:
+                                        # Holdback: si el final del buffer parece
+                                        # arrancar una apertura ('<', '<i', '<inv'...)
+                                        # esperamos al siguiente delta antes de emitir.
+                                        safe = _safe_emit_split(text_buffer)
+                                        if safe:
+                                            yield StreamChunk(delta=safe)
+                                            text_buffer = text_buffer[len(safe):]
                             elif delta.type == "input_json_delta":
                                 accumulated_args += delta.partial_json
                                 yield StreamChunk(
@@ -145,9 +246,40 @@ class ClaudeAdapter(ModelAdapter):
                             continue
 
                         if event.type == "message_delta":
+                            # Antes de cerrar, vaciar buffers.
+                            if in_xml_capture and xml_buffer:
+                                # Parsear el XML capturado y emitir tool_use sinteticos.
+                                calls = _parse_xml_tool_calls(xml_buffer)
+                                if calls:
+                                    logger.info(
+                                        "Detected %d inline XML tool call(s) — converting to tool_use",
+                                        len(calls),
+                                    )
+                                    for c in calls:
+                                        yield StreamChunk(
+                                            tool_call_id=c["id"],
+                                            tool_name=c["name"],
+                                        )
+                                        yield StreamChunk(
+                                            tool_call_id=c["id"],
+                                            tool_name=c["name"],
+                                            tool_arguments=json.dumps(c["arguments"]),
+                                            finish_reason="tool_use",
+                                        )
+                                else:
+                                    # No se pudo parsear — devolver al usuario el
+                                    # texto crudo para no perderlo silenciosamente.
+                                    yield StreamChunk(delta=xml_buffer)
+                                xml_buffer = ""
+                                in_xml_capture = False
+                            elif text_buffer:
+                                yield StreamChunk(delta=text_buffer)
+                                text_buffer = ""
                             output_tokens = getattr(event.usage, "output_tokens", 0) if event.usage else 0
+                            # Si convertimos XML a tool_use, override el stop_reason.
+                            stop_reason = event.delta.stop_reason or ""
                             yield StreamChunk(
-                                finish_reason=event.delta.stop_reason or "",
+                                finish_reason=stop_reason,
                                 usage={
                                     "input_tokens": input_tokens,
                                     "output_tokens": output_tokens,
diff --git a/src/orchestrator/agents/base.py b/src/orchestrator/agents/base.py
index c8f051d..1888897 100644
--- a/src/orchestrator/agents/base.py
+++ b/src/orchestrator/agents/base.py
@@ -64,8 +64,6 @@ class BaseAgent:
         total_output_tokens = 0
         # Real conversation history: assistant messages + tool results
         conversation: list[dict[str, Any]] = []
-        tool_fingerprints: dict[str, ToolExecution] = {}
-        all_duplicates_streak = 0  # consecutive steps where ALL calls are duplicates
 
         for step in range(max_steps):
             # Build context with real conversation
@@ -205,14 +203,19 @@ class BaseAgent:
             ]
             conversation.append(assistant_msg)
 
-            # Execute tool calls and add COMPLETE results to conversation
-            duplicates_this_step = 0
+            # Execute tool calls and add COMPLETE results to conversation.
+            # Antes habia dos capas anti-duplicado: (a) cachear resultado y
+            # devolver "[DUPLICADO]" en lugar de re-ejecutar y (b) cortar el
+            # step si TODAS las llamadas del paso eran duplicadas. Las quitamos
+            # porque en conversaciones largas el agente puede LEGITIMAMENTE
+            # repetir una llamada (p.ej. re-leer un fichero tras editarlo) y
+            # las heuristicas bloqueaban acciones validas. El usuario prefiere
+            # libertad — runaway loops se mitigan con limit de steps externo.
             for tc in tool_calls:
                 # Si los args no se pudieron parsear (p.ej. truncados por max_tokens),
                 # NO ejecutamos la tool. En su lugar devolvemos un mensaje al modelo
                 # explicando el problema para que pueda ajustar el siguiente intento
-                # (dividir el contenido, acortar, etc.). Fingerprint incluye el hash
-                # del raw para distinguir fallos distintos.
+                # (dividir el contenido, acortar, etc.).
                 if tc.get("parse_error"):
                     pe = tc["parse_error"]
                     conversation.append({
@@ -229,24 +232,6 @@ class BaseAgent:
                         ),
                     })
                     continue
-                fp_raw = f"{tc['name']}:{json.dumps(tc.get('parsed_arguments', {}), sort_keys=True)}"
-                fp = hashlib.md5(fp_raw.encode()).hexdigest()
-
-                if fp in tool_fingerprints:
-                    prev_exec = tool_fingerprints[fp]
-                    tool_executions.append(prev_exec)
-                    duplicates_this_step += 1
-                    # Return cached result as tool message
-                    conversation.append({
-                        "role": "tool",
-                        "tool_call_id": tc["id"],
-                        "content": (
-                            "[DUPLICADO] Ya ejecutada con mismos argumentos. Resultado: "
-                            f"{prev_exec.raw_output[:settings.tool_raw_output_max_chars]}"
-                        ),
-                    })
-                    logger.warning("Duplicate tool call skipped: %s (fingerprint: %s)", tc["name"], fp[:8])
-                    continue
 
                 tool_exec = await self._execute_tool(
                     session=session,
@@ -255,7 +240,6 @@ class BaseAgent:
                     artifacts=artifacts,
                     tool_call_id=tc["id"],
                 )
-                tool_fingerprints[fp] = tool_exec
                 tool_executions.append(tool_exec)
 
                 # COMPLETE result in conversation (truncated to safe limit)
@@ -269,32 +253,6 @@ class BaseAgent:
                     ),
                 })
 
-            # Loop detection: if ALL tool calls in this step were duplicates
-            if duplicates_this_step == len(tool_calls):
-                all_duplicates_streak += 1
-                if all_duplicates_streak >= 2:
-                    logger.warning("Loop detected: %d consecutive steps with all duplicate calls. Breaking.", all_duplicates_streak)
-                    conversation.append({
-                        "role": "user",
-                        "content": "[SISTEMA] Se detectaron llamadas repetidas. Ya tienes toda la información necesaria. Genera tu respuesta final ahora.",
-                    })
-                    # One more chance to generate a final response
-                    ctx = await self.context.build_context(
-                        session=session, agent=self.profile,
-                        artifacts=artifacts, conversation=conversation,
-                    )
-                    async for chunk in self.model.stream(
-                        messages=ctx.to_messages(),
-                        config=config,
-                    ):
-                        if chunk.delta:
-                            accumulated_content += chunk.delta
-                        if chunk.finish_reason:
-                            break
-                    break
-            else:
-                all_duplicates_streak = 0
-
         return {
             "content": accumulated_content,
             "artifacts": artifacts,