diff --git a/mcp-server/tools/media/analyze_image.js b/mcp-server/tools/media/analyze_image.js
index 9982e06..9386dcd 100644
--- a/mcp-server/tools/media/analyze_image.js
+++ b/mcp-server/tools/media/analyze_image.js
@@ -57,11 +57,16 @@ function resolveChatPreviewPath(imageUrl) {
     const fileParam = qs.get("file");
     if (!fileParam) return null;
 
-    // Sanitizar: evitar traversal — solo nombre base permitido
-    const safeName = path.basename(fileParam);
-    if (!safeName || safeName === "." || safeName === "..") return null;
+    // Sanitizar contra traversal PRESERVANDO el subdirectorio de usuario
+    // (el file= es "<username>/<archivo>"; basename lo perdía → not found).
+    if (fileParam.includes("..") || fileParam.startsWith("/") || fileParam.includes("\\")) return null;
 
-    return path.join(CHAT_UPLOADS_DIR, safeName);
+    const fullPath = path.join(CHAT_UPLOADS_DIR, fileParam);
+    // Asegurar que queda dentro de CHAT_UPLOADS_DIR.
+    const base = path.resolve(CHAT_UPLOADS_DIR);
+    if (!path.resolve(fullPath).startsWith(base + path.sep)) return null;
+
+    return fullPath;
 }
 
 /**
diff --git a/src/adapters/base.py b/src/adapters/base.py
index ac62f50..7062444 100644
--- a/src/adapters/base.py
+++ b/src/adapters/base.py
@@ -57,6 +57,10 @@ class ModelConfig:
     max_tokens: int = 4096
     temperature: float = 0.3
     stop_sequences: list[str] = field(default_factory=list)
+    # Nivel de razonamiento (minimal|low|medium|high). Vacío = sin razonamiento
+    # explícito. LiteLLM lo traduce por proveedor; modelos que no lo soportan lo
+    # ignoran (litellm.drop_params=True).
+    reasoning_effort: str = ""
     extra: dict[str, Any] = field(default_factory=dict)
 
 
diff --git a/src/adapters/litellm_adapter.py b/src/adapters/litellm_adapter.py
index 51eaa67..a985986 100644
--- a/src/adapters/litellm_adapter.py
+++ b/src/adapters/litellm_adapter.py
@@ -59,9 +59,25 @@ class LiteLLMAdapter(OpenAIAdapter):
 
     async def _acreate(self, kwargs: dict[str, Any]):
         kwargs = dict(kwargs)
-        kwargs["model"] = self._litellm_model
-        if self._api_key:
-            kwargs["api_key"] = self._api_key
-        if self._api_base:
-            kwargs["api_base"] = self._api_base
+        # Respetar el model_id por request (resuelto dinámicamente en
+        # send_message). Solo se honra si trae prefijo de proveedor
+        # ("deepseek/...", "openrouter/..."); cualquier otro valor (default
+        # no-litellm, vacío) cae al modelo por defecto del adapter — preserva el
+        # comportamiento previo para llamadas internas (planner, completions).
+        model = kwargs.get("model") or ""
+        if "/" not in model:
+            model = self._litellm_model
+        kwargs["model"] = model
+
+        if model.startswith("openrouter/"):
+            # OpenRouter: LiteLLM enruta con OPENROUTER_API_KEY del entorno y su
+            # base propia. NO forzar api_key/api_base del proxy DeepSeek — lo
+            # sobreescribirían y romperían el routing.
+            kwargs.pop("api_key", None)
+            kwargs.pop("api_base", None)
+        else:
+            if self._api_key:
+                kwargs["api_key"] = self._api_key
+            if self._api_base:
+                kwargs["api_base"] = self._api_base
         return await litellm.acompletion(**kwargs)
diff --git a/src/adapters/openai_adapter.py b/src/adapters/openai_adapter.py
index bac7896..24fd7cf 100644
--- a/src/adapters/openai_adapter.py
+++ b/src/adapters/openai_adapter.py
@@ -77,6 +77,8 @@ class OpenAIAdapter(ModelAdapter):
             "stream": True,
             "stream_options": {"include_usage": True},
         }
+        if getattr(config, "reasoning_effort", ""):
+            kwargs["reasoning_effort"] = config.reasoning_effort
         if tools:
             kwargs["tools"] = self._format_tools(tools)
 
@@ -266,6 +268,8 @@ class OpenAIAdapter(ModelAdapter):
             "temperature": config.temperature,
             "messages": self._to_openai_messages(messages),
         }
+        if getattr(config, "reasoning_effort", ""):
+            kwargs["reasoning_effort"] = config.reasoning_effort
         if tools:
             kwargs["tools"] = self._format_tools(tools)
         # Fuerza al modelo a usar un tool concreto para garantizar JSON por schema
@@ -428,8 +432,9 @@ class OpenAIAdapter(ModelAdapter):
                 if tool_calls:
                     m["tool_calls"] = tool_calls
                 out.append(m)
-            else:  # user (puede traer tool_result blocks)
+            else:  # user (puede traer tool_result blocks, texto e imágenes)
                 text_parts = []
+                image_blocks: list[dict[str, Any]] = []
                 for b in content:
                     if not isinstance(b, dict):
                         continue
@@ -442,7 +447,18 @@ class OpenAIAdapter(ModelAdapter):
                         })
                     elif t == "text":
                         text_parts.append(b.get("text", ""))
-                if text_parts:
+                    elif t == "image_url":
+                        # Visión nativa: preservar el bloque en formato multimodal OpenAI.
+                        image_blocks.append({"type": "image_url", "image_url": b.get("image_url") or {}})
+                if image_blocks:
+                    # Content como lista de bloques (texto + imágenes).
+                    parts: list[dict[str, Any]] = []
+                    joined = "\n".join(p for p in text_parts if p)
+                    if joined:
+                        parts.append({"type": "text", "text": joined})
+                    parts.extend(image_blocks)
+                    out.append({"role": "user", "content": parts})
+                elif text_parts:
                     out.append({"role": "user", "content": "\n".join(text_parts)})
         # Guard defensivo: el compactor ya garantiza el invariante tool_use ↔
         # tool_result (`_enforce_tool_pairing`), pero si algo se escapa el
diff --git a/src/api/routes.py b/src/api/routes.py
index 6e759b0..064777e 100644
--- a/src/api/routes.py
+++ b/src/api/routes.py
@@ -46,6 +46,10 @@ class SendMessageRequest(BaseModel):
     message: str
     stream: bool = False
     agent_id: str | None = None
+    # Imágenes para visión nativa: bloques listos para el modelo
+    # {"type":"image_url","image_url":{"url":"data:<mime>;base64,..."}}. Solo se
+    # envían cuando el modelo activo es multimodal (lo decide acai-app).
+    attachments: list[dict[str, Any]] | None = None
     # 'off' (default): la tool acai_plan no se expone al modelo, ejecuta directo.
     # 'force': system prompt obliga a llamar acai_plan antes de ejecutar.
     # 'auto' (legacy): se trata como 'off'. UI: toggle en ChatPanel.
@@ -335,6 +339,25 @@ async def send_message(
         if not agent_profile:
             agent_profile = agent_reg.get(agent_reg.default_agent_id)
 
+    # Resolución dinámica del modelo (Fase 2): override por-usuario (metadata de
+    # la sesión) → default global (Redis acai:config:ai:*). Si resuelve, se
+    # inyecta en una COPIA del profile para no mutar el del registry (singleton).
+    if agent_profile is not None:
+        from ..orchestrator.model_resolver import resolve_session_model
+        resolved = await resolve_session_model(session)
+        update = {}
+        if resolved.get("model_id"):
+            update["model_id"] = resolved["model_id"]
+        if resolved.get("reasoning_effort"):
+            update["reasoning_effort"] = resolved["reasoning_effort"]
+        if update:
+            agent_profile = agent_profile.model_copy(update=update)
+            logger.info(
+                "Session %s: modelo resuelto -> %s (reasoning=%s)",
+                session_id, update.get("model_id", "(default)"),
+                update.get("reasoning_effort", "off"),
+            )
+
     # Plan mode controlado por el usuario desde el toggle del ChatPanel.
     # 'auto' (default): heuristica del modelo trivial-vs-complex.
     # 'force': el agente DEBE llamar acai_plan como primera accion.
@@ -359,7 +382,7 @@ async def send_message(
 
     if body.stream:
         task = asyncio.create_task(
-            _execute_and_persist(orchestrator, storage, session, body.message)
+            _execute_and_persist(orchestrator, storage, session, body.message, body.attachments)
         )
         _running_executions[session_id] = task
         # Auto-limpieza del registro al terminar (solo si seguimos siendo la
@@ -377,11 +400,11 @@ async def send_message(
             "stream_url": f"/sessions/{session_id}/stream",
         }
 
-    result = await _execute_and_persist(orchestrator, storage, session, body.message)
+    result = await _execute_and_persist(orchestrator, storage, session, body.message, body.attachments)
     return result
 
 
-async def _execute_and_persist(orchestrator, storage, session, message) -> dict[str, Any]:
+async def _execute_and_persist(orchestrator, storage, session, message, attachments=None) -> dict[str, Any]:
     # Acquire exclusive lock — prevents concurrent execution on same session
     async with storage.session_lock(session.session_id) as acquired:
         if not acquired:
@@ -392,7 +415,7 @@ async def _execute_and_persist(orchestrator, storage, session, message) -> dict[
             }
 
         try:
-            result = await orchestrator.process_message(session, message)
+            result = await orchestrator.process_message(session, message, attachments)
             return result
         except asyncio.CancelledError:
             # Ejecución abortada por el usuario (stop) o preemptada por un
@@ -401,6 +424,24 @@ async def _execute_and_persist(orchestrator, storage, session, message) -> dict[
             # que el `await task` de la cancelación complete. El `finally`
             # persiste el estado y el `session_lock` se libera al salir.
             logger.info("Execution cancelled for session %s", session.session_id)
+            # Persistir el turno del usuario aunque se cancele: si no, un
+            # "vuelve a intentarlo" posterior se queda sin contexto de lo pedido.
+            # Guardamos su mensaje (+ imagen) y un marcador de interrupción para
+            # mantener la alternancia user/assistant.
+            try:
+                task = session.current_task
+                if task and (task.objective or "").strip():
+                    session.recent_messages = orchestrator._append_recent_messages(
+                        session.recent_messages,
+                        message=task.objective,
+                        conversation=[{
+                            "role": "assistant",
+                            "content": "[Respuesta interrumpida por el usuario antes de completarse]",
+                        }],
+                        image_attachments=task.image_attachments,
+                    )
+            except Exception:
+                logger.exception("No se pudo persistir el turno cancelado")
             session.status = SessionStatus.ACTIVE
             session.current_task = None
             raise
diff --git a/src/context/compactor.py b/src/context/compactor.py
index fc5e0ee..406ea76 100644
--- a/src/context/compactor.py
+++ b/src/context/compactor.py
@@ -890,6 +890,10 @@ class ContextCompactor:
                 elif btype == "tool_result":
                     tc = block.get("content", "")
                     tokens += estimate_tokens(tc if isinstance(tc, str) else str(tc))
+                elif btype == "image_url":
+                    # Una imagen ~1500 tokens. NO medir el base64 como texto, que
+                    # lo contaría como ~30k y reventaría presupuestos/trim.
+                    tokens += 1500
                 else:
                     tokens += estimate_tokens(str(block))
         else:
diff --git a/src/context/engine.py b/src/context/engine.py
index f3b9001..f8d23ca 100644
--- a/src/context/engine.py
+++ b/src/context/engine.py
@@ -924,8 +924,19 @@ class ContextEngine:
             messages.append({"role": "user", "content": "\n".join(history_lines)})
             messages.append({"role": "assistant", "content": "Entendido, tengo el contexto del historial. ¿En qué puedo ayudarte ahora?"})
 
-        # Current user message
-        messages.append({"role": "user", "content": user_content})
+        # Current user message — con imágenes adjuntas (visión nativa) si las hay.
+        # En ese caso el content pasa a ser lista de bloques [texto, image_url...].
+        image_attachments = []
+        if session.current_task and getattr(session.current_task, "image_attachments", None):
+            image_attachments = [
+                b for b in session.current_task.image_attachments if isinstance(b, dict)
+            ]
+        if image_attachments:
+            content_blocks = [{"type": "text", "text": user_content}]
+            content_blocks.extend(image_attachments)
+            messages.append({"role": "user", "content": content_blocks})
+        else:
+            messages.append({"role": "user", "content": user_content})
 
         # Append real conversation (assistant messages + tool results from current step)
         if conversation:
@@ -1037,6 +1048,10 @@ class ContextEngine:
                 elif btype == "tool_result":
                     tc = block.get("content", "")
                     total += estimate_tokens(tc if isinstance(tc, str) else str(tc))
+                elif btype == "image_url":
+                    # Heurística conservadora: una imagen ~1500 tokens (no se
+                    # cuenta el base64 como texto, que infla muchísimo).
+                    total += 1500
                 else:
                     total += estimate_tokens(str(block))
             return total
diff --git a/src/models/agent.py b/src/models/agent.py
index 1d14468..810e34b 100644
--- a/src/models/agent.py
+++ b/src/models/agent.py
@@ -20,6 +20,7 @@ class AgentProfile(BaseModel):
     allowed_tools: list[str] = Field(default_factory=list)
     model_id: str | None = None
     planner_model_id: str | None = None  # override del modelo solo para el sub-loop del planner
+    reasoning_effort: str | None = None  # nivel de razonamiento (minimal|low|medium|high) resuelto por sesión
     temperature: float | None = None
     max_tokens: int | None = None
     context_sections: list[str] = Field(
diff --git a/src/models/session.py b/src/models/session.py
index e91fc9c..61bf858 100644
--- a/src/models/session.py
+++ b/src/models/session.py
@@ -46,6 +46,9 @@ class TaskState(BaseModel):
 
     task_id: str = Field(default_factory=lambda: uuid.uuid4().hex[:12])
     objective: str
+    # Imágenes adjuntas a la petición actual (visión nativa). Cada item es un
+    # bloque listo para el modelo: {"type":"image_url","image_url":{"url":"data:..."}}.
+    image_attachments: list[dict[str, Any]] = Field(default_factory=list)
     status: TaskStatus = TaskStatus.PENDING
     plan: list[TaskStep] = Field(default_factory=list)
     current_step_index: int = 0
@@ -94,8 +97,8 @@ class SessionState(BaseModel):
     updated_at: datetime = Field(default_factory=lambda: datetime.now(timezone.utc))
     metadata: dict[str, Any] = Field(default_factory=dict)
 
-    def begin_task(self, objective: str) -> TaskState:
-        task = TaskState(objective=objective)
+    def begin_task(self, objective: str, image_attachments: list[dict[str, Any]] | None = None) -> TaskState:
+        task = TaskState(objective=objective, image_attachments=image_attachments or [])
         self.current_task = task
         self.status = SessionStatus.EXECUTING
         self.turn_count += 1
diff --git a/src/orchestrator/agents/base.py b/src/orchestrator/agents/base.py
index 533dd7e..2e557b2 100644
--- a/src/orchestrator/agents/base.py
+++ b/src/orchestrator/agents/base.py
@@ -93,6 +93,7 @@ class BaseAgent:
                 model_id=self.profile.model_id or "",
                 max_tokens=self.profile.max_tokens or 4096,
                 temperature=self.profile.temperature or 0.3,
+                reasoning_effort=self.profile.reasoning_effort or "",
             )
 
             # Snapshot del numero de tool_executions ya acumulados ANTES del
diff --git a/src/orchestrator/cost.py b/src/orchestrator/cost.py
new file mode 100644
index 0000000..196ca23
--- /dev/null
+++ b/src/orchestrator/cost.py
@@ -0,0 +1,121 @@
+"""Cálculo de coste por modelo (Fase 2).
+
+Prioridad de fuentes de precio (para que el coste registrado en
+`consumo_acaicode` coincida con lo que muestra el Forge Admin Panel):
+  1. Catálogo OpenRouter cacheado por el panel en Redis db 0
+     (`acai:config:ai:models_cache:openrouter` → price_in_1m / price_out_1m).
+  2. Price map de LiteLLM (conoce muchos modelos deepseek/, anthropic/, etc.).
+  3. Coste fijo de `settings` (comportamiento previo).
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+
+import redis.asyncio as redis
+
+from ..config import settings
+
+logger = logging.getLogger(__name__)
+
+# Caches de catálogo que publica el Forge Admin Panel en Redis db 0, por proveedor.
+# El id se guarda SIN el prefijo de proveedor de litellm (p.ej.
+# "moonshotai/kimi-k2.7-code", "deepseek-v4-pro").
+_CACHE_KEYS = {
+    "openrouter": "acai:config:ai:models_cache:openrouter",
+    "deepseek": "acai:config:ai:models_cache:deepseek",
+}
+_CONFIG_DB = 0
+_cfg_redis: "redis.Redis | None" = None
+
+
+def _get_cfg_redis() -> "redis.Redis":
+    global _cfg_redis
+    if _cfg_redis is None:
+        _cfg_redis = redis.Redis(
+            host=settings.redis_host,
+            port=settings.redis_port,
+            db=_CONFIG_DB,
+            password=settings.redis_password or None,
+            decode_responses=True,
+        )
+    return _cfg_redis
+
+
+async def _catalog_price_per_1m(model_id: str | None):
+    """(price_in_1m, price_out_1m) del catálogo del panel, o None.
+
+    model_id viene en formato litellm ("<provider>/<id>"). Separamos el prefijo
+    de proveedor para elegir el cache y buscar por el id catalogado.
+    """
+    if not model_id or "/" not in model_id:
+        return None
+    provider, _, raw_id = model_id.partition("/")
+    cache_key = _CACHE_KEYS.get(provider)
+    if not cache_key:
+        return None
+    try:
+        cached = await _get_cfg_redis().get(cache_key)
+        if not cached:
+            return None
+        models = json.loads(cached)
+    except Exception as e:  # pragma: no cover - defensivo
+        logger.warning("catálogo %s no disponible para coste: %s", provider, e)
+        return None
+    for m in models:
+        if m.get("id") == raw_id:
+            pin = m.get("price_in_1m")
+            pout = m.get("price_out_1m")
+            if pin is not None and pout is not None:
+                return (float(pin), float(pout))
+    return None
+
+
+async def compute_cost(model_id: str | None, input_tokens: int, output_tokens: int) -> dict:
+    """Coste de una ejecución para `model_id` y los tokens dados.
+
+    Devuelve {"cost_usd", "input_cost_1m", "output_cost_1m"} — el coste total y
+    las tarifas por 1M tokens REALMENTE aplicadas (se almacenan en
+    `consumo_acaicode.input_cost_1M` / `output_cost_1M`).
+    """
+    input_tokens = int(input_tokens or 0)
+    output_tokens = int(output_tokens or 0)
+
+    def _result(in_1m: float, out_1m: float) -> dict:
+        return {
+            "cost_usd": (input_tokens / 1_000_000) * in_1m + (output_tokens / 1_000_000) * out_1m,
+            "input_cost_1m": round(in_1m, 6),
+            "output_cost_1m": round(out_1m, 6),
+        }
+
+    # 1. Precio del catálogo OpenRouter (fuente que muestra el admin).
+    prices = await _catalog_price_per_1m(model_id)
+    if prices:
+        return _result(prices[0], prices[1])
+
+    # 2. Price map de LiteLLM (deepseek/, anthropic/, etc.).
+    if model_id and "/" in model_id:
+        try:
+            import litellm
+
+            prompt_cost, completion_cost = litellm.cost_per_token(
+                model=model_id,
+                prompt_tokens=input_tokens,
+                completion_tokens=output_tokens,
+            )
+            total = (prompt_cost or 0.0) + (completion_cost or 0.0)
+            if total > 0:
+                # Derivar tarifa por 1M a partir del coste por-token de litellm.
+                in_1m = (prompt_cost / input_tokens) * 1_000_000 if input_tokens else 0.0
+                out_1m = (completion_cost / output_tokens) * 1_000_000 if output_tokens else 0.0
+                return {
+                    "cost_usd": total,
+                    "input_cost_1m": round(in_1m, 6),
+                    "output_cost_1m": round(out_1m, 6),
+                }
+        except Exception as e:
+            logger.warning("cost_per_token(%s) falló, uso coste fijo: %s", model_id, e)
+
+    # 3. Coste fijo configurado.
+    return _result(settings.cost_per_1m_input, settings.cost_per_1m_output)
diff --git a/src/orchestrator/engine.py b/src/orchestrator/engine.py
index e073f25..0a52ff3 100644
--- a/src/orchestrator/engine.py
+++ b/src/orchestrator/engine.py
@@ -52,11 +52,16 @@ class OrchestratorEngine:
         self,
         session: SessionState,
         message: str,
+        image_attachments: list[dict[str, Any]] | None = None,
     ) -> dict[str, Any]:
-        """Process a user message. Single agent execution with timeout."""
+        """Process a user message. Single agent execution with timeout.
+
+        `image_attachments`: bloques image_url (visión nativa) para el turno del
+        usuario, cuando el modelo activo es multimodal.
+        """
         try:
             return await asyncio.wait_for(
-                self._run(session, message),
+                self._run(session, message, image_attachments),
                 timeout=settings.max_execution_timeout_seconds,
             )
         except asyncio.TimeoutError:
@@ -86,6 +91,7 @@ class OrchestratorEngine:
         self,
         session: SessionState,
         message: str,
+        image_attachments: list[dict[str, Any]] | None = None,
     ) -> dict[str, Any]:
         """Execute: message → agent → response."""
 
@@ -113,8 +119,8 @@ class OrchestratorEngine:
                 f"Peticion del usuario:\n{message}"
             )
 
-        # Create task
-        task = session.begin_task(objective=message)
+        # Create task (con imágenes adjuntas si las hay — visión nativa)
+        task = session.begin_task(objective=message, image_attachments=image_attachments)
         task.status = TaskStatus.EXECUTING
 
         # Reset del contador de invocaciones de `acai_plan` por turno (Fase 5).
@@ -154,6 +160,9 @@ class OrchestratorEngine:
             session.recent_messages,
             message=message,
             conversation=result.get("conversation", []),
+            image_attachments=(
+                session.current_task.image_attachments if session.current_task else None
+            ),
         )
 
         session.task_history.append(
@@ -182,13 +191,18 @@ class OrchestratorEngine:
         task.status = TaskStatus.COMPLETED
         session.complete_task()
 
-        # Calculate cost
+        # Calculate cost — por modelo realmente usado (Fase 2). El model_id
+        # efectivo vive en el agent_profile (resuelto en send_message).
         total_input = usage.get("input_tokens", 0)
         total_output = usage.get("output_tokens", 0)
-        cost_usd = (
-            (total_input / 1_000_000) * settings.cost_per_1m_input
-            + (total_output / 1_000_000) * settings.cost_per_1m_output
+        model_used = (
+            self.agent_profile.model_id
+            or settings.litellm_model
+            or settings.default_model_id
         )
+        from .cost import compute_cost
+        cost_info = await compute_cost(model_used, total_input, total_output)
+        cost_usd = cost_info["cost_usd"]
 
         await self.sse.emit(
             EventType.EXECUTION_COMPLETED,
@@ -201,6 +215,19 @@ class OrchestratorEngine:
                 "status": "completed",
                 "usage": usage,
                 "total_cost_usd": round(cost_usd, 6),
+                # Modelo + tarifas usadas → se propagan a consumo_acaicode via
+                # _report_usage (columnas input_cost_1M / output_cost_1M).
+                "model": model_used,
+                "modelUsage": {
+                    model_used: {
+                        "inputTokens": total_input,
+                        "outputTokens": total_output,
+                        "costUSD": round(cost_usd, 6),
+                        "inputCost1M": cost_info["input_cost_1m"],
+                        "outputCost1M": cost_info["output_cost_1m"],
+                        "reasoningEffort": self.agent_profile.reasoning_effort or "",
+                    }
+                },
             },
             session_id=session.session_id,
         )
@@ -246,13 +273,21 @@ class OrchestratorEngine:
         existing: list[dict[str, Any]],
         message: str,
         conversation: list[dict[str, Any]],
+        image_attachments: list[dict[str, Any]] | None = None,
     ) -> list[dict[str, Any]]:
         merged = [OrchestratorEngine._sanitize_recent_message(m) for m in existing]
         merged = [m for m in merged if m]
 
         current_turn: list[dict[str, Any]] = []
-        if message.strip():
-            current_turn.append({"role": "user", "content": message})
+        if message.strip() or image_attachments:
+            if image_attachments:
+                # Guardar el turno con la imagen como bloques para que PERSISTA
+                # en el contexto de turnos siguientes (visión nativa multimodal).
+                content_blocks = [{"type": "text", "text": message}]
+                content_blocks.extend(image_attachments)
+                current_turn.append({"role": "user", "content": content_blocks})
+            else:
+                current_turn.append({"role": "user", "content": message})
 
         for message_obj in conversation:
             sanitized = OrchestratorEngine._sanitize_recent_message(message_obj)
diff --git a/src/orchestrator/model_resolver.py b/src/orchestrator/model_resolver.py
new file mode 100644
index 0000000..bc37fd7
--- /dev/null
+++ b/src/orchestrator/model_resolver.py
@@ -0,0 +1,113 @@
+"""Resolución dinámica del modelo IA por sesión (Fase 2).
+
+Prioridad:
+  1. Override por-usuario: `session.metadata["ai_provider"|"ai_model"]`. Lo
+     inyecta acai-app via self-read del WS (`getAcaiCodeUserAiModel`) al crear
+     la sesión.
+  2. Default global: Redis `acai:config:ai:provider` / `acai:config:ai:model`,
+     que escribe el Forge Admin Panel. OJO: esas keys NO llevan el prefijo
+     `agentic` (son globales del stack Acai).
+  3. Sin configuración → None: no se toca el modelo y el adapter usa su default
+     (comportamiento previo).
+
+Solo aplica cuando el provider activo es `litellm` — los providers del catálogo
+(openrouter, deepseek) se enrutan por LiteLLM. Para claude/openai no se toca.
+"""
+
+from __future__ import annotations
+
+import logging
+
+import redis.asyncio as redis
+
+from ..config import settings
+
+logger = logging.getLogger(__name__)
+
+# Keys del Forge Admin Panel (globales, SIN prefijo agentic).
+_GLOBAL_PROVIDER_KEY = "acai:config:ai:provider"
+_GLOBAL_MODEL_KEY = "acai:config:ai:model"
+_GLOBAL_REASONING_KEY = "acai:config:ai:reasoning_effort"
+
+# Niveles de razonamiento válidos (lo demás se ignora → sin razonamiento).
+_VALID_EFFORTS = {"minimal", "low", "medium", "high"}
+
+# El Forge Admin Panel escribe la config global en Redis db 0 (REDIS_DB=0 del
+# admin). El agentic usa db 1 para sus sesiones, así que para leer la config
+# global necesitamos una conexión dedicada a db 0 (misma instancia Redis).
+_GLOBAL_CONFIG_DB = 0
+_global_redis: "redis.Redis | None" = None
+
+
+def _get_global_redis() -> "redis.Redis":
+    global _global_redis
+    if _global_redis is None:
+        _global_redis = redis.Redis(
+            host=settings.redis_host,
+            port=settings.redis_port,
+            db=_GLOBAL_CONFIG_DB,
+            password=settings.redis_password or None,
+            decode_responses=True,
+        )
+    return _global_redis
+
+
+def to_litellm_model(provider: str | None, model: str | None) -> str:
+    """Mapea {provider, model} del catálogo a un model string de LiteLLM."""
+    provider = (provider or "").strip().lower()
+    model = (model or "").strip()
+    if not model:
+        return ""
+    if provider == "openrouter":
+        # Los ids de OpenRouter ya vienen como "vendor/name" → prefijo openrouter/.
+        return model if model.startswith("openrouter/") else f"openrouter/{model}"
+    if provider == "deepseek":
+        return model if model.startswith("deepseek/") else f"deepseek/{model}"
+    # Provider desconocido: respetar el id tal cual (puede traer ya su prefijo).
+    return model
+
+
+def _norm_effort(value) -> str | None:
+    v = (value or "").strip().lower()
+    return v if v in _VALID_EFFORTS else None
+
+
+async def resolve_session_model(session) -> dict:
+    """Resuelve modelo + razonamiento efectivos para la sesión.
+
+    Devuelve {"model_id": str|None, "reasoning_effort": str|None}. El effort se
+    toma de la MISMA fuente que el modelo (override de usuario o default global),
+    para que sean coherentes. model_id None = sin override (adapter usa default).
+    """
+    none = {"model_id": None, "reasoning_effort": None}
+    if settings.default_model_provider != "litellm":
+        return none
+
+    # 1. Override por-usuario (metadata de la sesión).
+    meta = getattr(session, "metadata", None) or {}
+    provider = meta.get("ai_provider")
+    model = meta.get("ai_model")
+    if provider and model:
+        return {
+            "model_id": to_litellm_model(provider, model) or None,
+            "reasoning_effort": _norm_effort(meta.get("ai_reasoning_effort")),
+        }
+
+    # 2. Default global (Redis db 0, keys sin prefijo agentic).
+    try:
+        gr = _get_global_redis()
+        provider = await gr.get(_GLOBAL_PROVIDER_KEY)
+        model = await gr.get(_GLOBAL_MODEL_KEY)
+        effort = await gr.get(_GLOBAL_REASONING_KEY)
+    except Exception as e:  # pragma: no cover - defensivo
+        logger.warning("resolve_session_model: lectura Redis falló: %s", e)
+        return none
+
+    if provider and model:
+        return {
+            "model_id": to_litellm_model(provider, model) or None,
+            "reasoning_effort": _norm_effort(effort),
+        }
+
+    # 3. Sin configuración → sin override.
+    return none
diff --git a/src/orchestrator/planner.py b/src/orchestrator/planner.py
index c85b5aa..de49bf3 100644
--- a/src/orchestrator/planner.py
+++ b/src/orchestrator/planner.py
@@ -217,6 +217,8 @@ async def run_planner_subloop(
         max_tokens=settings.planner_max_tokens or 16000,
         # Temperatura mas baja que el agente principal — queremos JSON limpio.
         temperature=0.1,
+        # Mismo nivel de razonamiento resuelto por sesión que el agente principal.
+        reasoning_effort=agent_profile.reasoning_effort or "",
     )
 
     tool_defs = _build_planner_tools(mcp)
diff --git a/src/streaming/claude_format.py b/src/streaming/claude_format.py
index 467dc41..e9314e0 100644
--- a/src/streaming/claude_format.py
+++ b/src/streaming/claude_format.py
@@ -363,6 +363,8 @@ class ClaudeFormatEmitter:
                     "cache_creation_input_tokens": 0,
                 },
                 "total_cost_usd": data.get("total_cost_usd", 0),
+                # Modelo usado → acai-app lo registra en consumo_acaicode.
+                "modelUsage": data.get("modelUsage", {}),
             })
 
             # Done