feat(validation): v2 LLM dashboard validation — plugin, routes, services

Core implementation of the v2 LLM dashboard validation pipeline:
- LLM plugin with Path A (screenshots) and Path B (logs-only) execution
- Validation task management (CRUD, schedule, run)
- WebSocket task progress with Python 3.13 asyncio fix
- Cross-task runs listing (GET /validation-tasks/runs/all)
- RecordResponse schema for validation records
- JSON prompt helper, per-dashboard status aggregation
- Prompt templates with docs/git-commit/validation presets
- Migration: v2 validation models + description column
- Tests: plugin persistence, prompt templates, batch, payload, url

backend/src/api/routes/assistant/_tool_llm_validation.py

@@ -1,9 +1,14 @@
-# #region AssistantToolLlmValidation [C:3] [TYPE Module] [SEMANTICS assistant, tool, llm, validation]
-# @BRIEF Handler for the "run_llm_validation" tool — run LLM dashboard validation.
+# #region AssistantToolLlmValidation [C:3] [TYPE Module] [SEMANTICS assistant, tool, llm, validation, task]
+# @BRIEF Handler for the "run_llm_validation" tool — create a validation task (policy) and trigger a run.
 # @LAYER API
 # @RELATION DEPENDS_ON -> [AssistantToolRegistry]
+# @RELATION DEPENDS_ON -> [ValidationTaskService]
 # @RELATION DEPENDS_ON -> [LLMProviderService]
 # @RELATION DEPENDS_ON -> [TaskManager]
+# @RATIONALE v2 replaces ad-hoc task dispatch with persistent validation tasks (policies) managed via
+#            ValidationTaskService. This makes validations repeatable, schedulable, and auditable.
+# @REJECTED Ad-hoc task dispatch rejected — single-use tasks cannot be re-triggered, scheduled, or
+#           tracked across runs. Policy+Run model provides run history and aggregate status.
 
 from __future__ import annotations
 
@@ -16,7 +21,9 @@ from src.core.config_manager import ConfigManager
 from src.core.logger import belief_scope, logger
 from src.core.task_manager import TaskManager
 from src.schemas.auth import User
+from src.schemas.validation import ValidationTaskCreate
 from src.services.llm_provider import LLMProviderService
+from src.services.validation_service import ValidationTaskService
 
 from ._resolvers import (
     _resolve_dashboard_id_entity,
@@ -27,11 +34,11 @@ from ._schemas import AssistantAction
 from ._tool_registry import _check_any_permission, assistant_tool
 
 
-# #region handle_run_llm_validation [C:3] [TYPE Function]
+# #region handle_run_llm_validation [C:4] [TYPE Function]
 @assistant_tool(
     operation="run_llm_validation",
     domain="llm",
-    description="Run LLM dashboard validation by dashboard id/slug/title",
+    description="Create and run an LLM dashboard validation task — creates a validation policy and triggers an immediate run",
     optional_entities=["dashboard_ref", "environment", "provider"],
     risk_level="guarded",
     requires_confirmation=False,
@@ -45,7 +52,7 @@ async def handle_run_llm_validation(
     config_manager: ConfigManager,
     db: Session,
 ) -> tuple[str, str | None, list[AssistantAction]]:
-    """Run LLM dashboard validation."""
+    """Create a validation task (policy) and trigger an immediate run."""
     _check_any_permission(current_user, [("plugin:llm_dashboard_validation", "EXECUTE")])
     entities = intent.get("entities", {})
     env_token = entities.get("environment")
@@ -66,26 +73,43 @@ async def handle_run_llm_validation(
             status_code=422,
             detail="Missing dashboard_id/environment/provider. Укажите ID/slug дашборда или окружение.",
         )
-    provider = LLMProviderService(db).get_provider(provider_id)
-    if not provider or not bool(provider.is_multimodal):
-        raise HTTPException(
-            status_code=422,
-            detail="Selected provider model is not multimodal for dashboard validation. Выберите мультимодальную модель (например, gpt-4o).",
-        )
-    task = await task_manager.create_task(
-        plugin_id="llm_dashboard_validation",
-        params={
-            "dashboard_id": str(dashboard_id),
-            "environment_id": env_id,
-            "provider_id": provider_id,
-        },
-        user_id=current_user.id,
+
+    # Create a validation task (policy) via ValidationTaskService
+    validation_svc = ValidationTaskService(
+        db=db, config_manager=config_manager, username=current_user.username
     )
+    task_payload = ValidationTaskCreate(
+        name=f"LLM Validation - Dashboard {dashboard_id}",
+        environment_id=env_id,
+        dashboard_ids=[str(dashboard_id)],
+        provider_id=provider_id,
+        screenshot_enabled=True,
+    )
+    try:
+        task = validation_svc.create_task(task_payload)
+    except ValueError as e:
+        raise HTTPException(status_code=422, detail=str(e))
+    logger.info(
+        "[AssistantToolLlmValidation] Created validation task '%s' (id=%s)",
+        task.name, task.id,
+    )
+
+    # Trigger an immediate run
+    try:
+        result = await validation_svc.trigger_run(task.id, task_manager)
+    except ValueError as e:
+        raise HTTPException(status_code=422, detail=str(e))
+    spawned_task_id = result["task_id"]
+    logger.info(
+        "[AssistantToolLlmValidation] Triggered run for task '%s' spawned=%s run=%s",
+        task.name, spawned_task_id, result.get("run_id"),
+    )
+
     return (
-        f"LLM-валидация запущена. task_id={task.id}",
-        task.id,
+        f"LLM-валидация запущена. task_id={spawned_task_id}",
+        spawned_task_id,
         [
-            AssistantAction(type="open_task", label="Open Task", target=task.id),
+            AssistantAction(type="open_task", label="Open Task", target=spawned_task_id),
             AssistantAction(
                 type="open_reports", label="Open Reports", target="/reports"
             ),