bugfix: model config robustness, default embedder, flexible json mode decision

python/configs/agents/research_agent.yaml

@@ -18,10 +18,10 @@ models:
   # Embedding model for knowledge base
   # Note: If not specified, will auto-select from available providers
   embedding:
-    # model_id: "openai/text-embedding-3-small"  # Optional, uses provider default
-    # provider: "openrouter"  # Optional, auto-detects if not specified
+    model_id: "Qwen/Qwen3-Embedding-4B"  # Optional, uses provider default
+    provider: "siliconflow"  # Optional, auto-detects if not specified
     parameters:
-      dimensions: ${EMBEDDER_DIMENSION:1536}
+      dimensions: ${EMBEDDER_DIMENSION:2560}
       # encoding_format: "float"
 
 # Environment Variable Overrides
@@ -35,5 +35,6 @@ env_overrides:
 
   # Embedding model overrides (backward compatible with existing env vars)
   EMBEDDER_MODEL_ID: "models.embedding.model_id"
+  EMBEDDER_MODEL_PROVIDER: "models.embedding.provider"
   EMBEDDER_DIMENSION: "models.embedding.parameters.dimensions"
 

python/configs/agents/super_agent.yaml

@@ -23,10 +23,6 @@ models:
     model_id: "anthropic/claude-haiku-4.5"
     provider: "openrouter"  # Can be: openrouter, siliconflow, or null (uses system primary_provider)
 
-
-
-
-
 # Environment Variable Overrides
 # Format: ENV_VAR_NAME -> config.path
 # These allow runtime configuration through environment variables
@@ -56,9 +52,8 @@ capabilities:
 advanced:
   # Debug mode (can be overridden by AGENT_DEBUG_MODE env var)
   debug_mode: false
-  
+
   # Output format
-  use_json_mode: true
   markdown: false
 
   # Context settings

python/valuecell/adapters/models/factory.py

@@ -443,6 +443,53 @@ def create_model_for_agent(
             **merged_params,
         )
 
+    def create_embedder_for_agent(
+        self, agent_name: str, use_fallback: bool = True, **kwargs
+    ):
+        """
+        Create an embedder for a specific agent using its configuration.
+
+        This method will use the agent's `embedding_model` configuration when
+        present. If no embedding config is provided for the agent it will
+        attempt to use the agent's primary model provider and select the
+        provider's default embedding model.
+        """
+        # Get agent configuration
+        agent_config = self.config_manager.get_agent_config(agent_name)
+
+        if not agent_config:
+            raise ValueError(f"Agent configuration not found: {agent_name}")
+
+        if not agent_config.enabled:
+            raise ValueError(f"Agent is disabled: {agent_name}")
+
+        # If agent specifies an embedding_model, use it
+        if agent_config.embedding_model:
+            emb = agent_config.embedding_model
+            merged_params = {**emb.parameters, **kwargs}
+            logger.info(
+                f"Creating embedder for agent '{agent_name}': model_id={emb.model_id}, provider={emb.provider}, params={merged_params}"
+            )
+            return self.create_embedder(
+                model_id=emb.model_id or None,
+                provider=emb.provider,
+                use_fallback=use_fallback,
+                **merged_params,
+            )
+
+        # Fallback: use primary model's provider and let factory pick provider default
+        primary = agent_config.primary_model
+        merged_params = {**primary.parameters, **kwargs}
+        logger.info(
+            f"Creating embedder for agent '{agent_name}' using primary provider: {primary.provider}"
+        )
+        return self.create_embedder(
+            model_id=None,
+            provider=primary.provider,
+            use_fallback=use_fallback,
+            **merged_params,
+        )
+
     def get_available_providers(self) -> list[str]:
         """
         Get list of available providers (with valid credentials)
@@ -752,3 +799,18 @@ def create_embedder(
     """
     factory = get_model_factory()
     return factory.create_embedder(model_id, provider, **kwargs)
+
+
+def create_embedder_for_agent(agent_name: str, **kwargs):
+    """
+    Convenience function to create an embedder configured for a specific agent.
+
+    Args:
+        agent_name: Agent name
+        **kwargs: Override parameters
+
+    Returns:
+        Embedder instance
+    """
+    factory = get_model_factory()
+    return factory.create_embedder_for_agent(agent_name, **kwargs)

python/valuecell/agents/research_agent/vdb.py

@@ -17,16 +17,16 @@
 from agno.vectordb.lancedb import LanceDb
 from agno.vectordb.search import SearchType
 
+import valuecell.utils.model as model_utils_mod
 from valuecell.utils.db import resolve_lancedb_uri
-from valuecell.utils.model import get_embedder
 
 # Create embedder using the configuration system
 # This will:
 # - Check EMBEDDER_MODEL_ID env var first
 # - Auto-select provider with embedding support (e.g., SiliconFlow if SILICONFLOW_API_KEY is set)
 # - Use provider's default embedding model if not specified
 # - Fall back to other providers if primary fails
-embedder = get_embedder("EMBEDDER_MODEL_ID")
+embedder = model_utils_mod.get_embedder_for_agent("research_agent")
 
 # Alternative usage examples:
 # embedder = get_embedder()  # Use default env key

python/valuecell/config/manager.py

@@ -271,10 +271,11 @@ def get_agent_config(self, agent_name: str) -> Optional[AgentConfig]:
                 model_id = provider_config.default_model
 
         # Get parameters
-        parameters = primary.get("parameters", {})
+        parameters = primary.get("parameters") or {}
 
         # Merge with global defaults
-        global_defaults = self._config.get("models", {}).get("defaults", {})
+        global_models = self._config.get("models") or {}
+        global_defaults = global_models.get("defaults") or {}
         merged_params = {**global_defaults, **parameters}
 
         primary_model = AgentModelConfig(

python/valuecell/core/coordinate/tests/test_e2e_persistence.py

@@ -20,7 +20,7 @@ async def test_orchestrator_buffer_store_e2e(tmp_path, monkeypatch):
         factory_mod, "create_embedder", lambda *args, **kwargs: "stub-embedder"
     )
     monkeypatch.setattr(
-        model_utils_mod, "create_model", lambda *args, **kwargs: "stub-model"
+        model_utils_mod, "get_model_for_agent", lambda *args, **kwargs: "stub-model"
     )
     monkeypatch.setattr(
         model_utils_mod, "create_embedder", lambda *args, **kwargs: "stub-embedder"

python/valuecell/core/plan/planner.py

@@ -19,12 +19,12 @@
 from agno.agent import Agent
 from agno.db.in_memory import InMemoryDb
 
+import valuecell.utils.model as model_utils_mod
 from valuecell.core.agent.connect import RemoteConnections
 from valuecell.core.task.models import Task, TaskStatus
 from valuecell.core.types import UserInput
 from valuecell.utils import generate_uuid
 from valuecell.utils.env import agent_debug_mode_enabled
-from valuecell.utils.model import get_model
 from valuecell.utils.uuid import generate_conversation_id
 
 from .models import ExecutionPlan, PlannerInput, PlannerResponse
@@ -88,8 +88,10 @@ def __init__(
         agent_connections: RemoteConnections,
     ):
         self.agent_connections = agent_connections
+        # Fetch model via utils module reference so tests can monkeypatch it reliably
+        model = model_utils_mod.get_model_for_agent("super_agent")
         self.planner_agent = Agent(
-            model=get_model("PLANNER_MODEL_ID"),
+            model=model,
             tools=[
                 # TODO: enable UserControlFlowTools when stable
                 # UserControlFlowTools(),
@@ -101,6 +103,7 @@ def __init__(
             markdown=False,
             output_schema=PlannerResponse,
             expected_output=PLANNER_EXPECTED_OUTPUT,
+            use_json_mode=model_utils_mod.model_should_use_json_mode(model),
             # context
             db=InMemoryDb(),
             add_datetime_to_context=True,

python/valuecell/core/plan/tests/test_planner.py

@@ -69,7 +69,7 @@ def continue_run(self, *args, **kwargs):
 
     monkeypatch.setattr(planner_mod, "Agent", FakeAgent)
     monkeypatch.setattr(
-        model_utils_mod, "create_model", lambda *args, **kwargs: "stub-model"
+        model_utils_mod, "get_model_for_agent", lambda *args, **kwargs: "stub-model"
     )
     monkeypatch.setattr(planner_mod, "agent_debug_mode_enabled", lambda: False)
 
@@ -120,7 +120,7 @@ def run(self, *args, **kwargs):
 
     monkeypatch.setattr(planner_mod, "Agent", FakeAgent)
     monkeypatch.setattr(
-        model_utils_mod, "create_model", lambda *args, **kwargs: "stub-model"
+        model_utils_mod, "get_model_for_agent", lambda *args, **kwargs: "stub-model"
     )
     monkeypatch.setattr(planner_mod, "agent_debug_mode_enabled", lambda: False)
 
@@ -142,7 +142,7 @@ async def callback(request):
 def test_tool_get_enabled_agents_formats_cards(monkeypatch: pytest.MonkeyPatch):
     # Mock create_model to avoid API key validation in CI
     monkeypatch.setattr(
-        model_utils_mod, "create_model", lambda *args, **kwargs: "stub-model"
+        model_utils_mod, "get_model_for_agent", lambda *args, **kwargs: "stub-model"
     )
     monkeypatch.setattr(planner_mod, "agent_debug_mode_enabled", lambda: False)
 

python/valuecell/core/super_agent/core.py

@@ -1,21 +1,28 @@
 import asyncio
-import logging
 from enum import Enum
 from typing import Optional
 
 from agno.agent import Agent
 from agno.db.in_memory import InMemoryDb
 from pydantic import BaseModel, Field
 
+import valuecell.utils.model as model_utils_mod
 from valuecell.core.super_agent.prompts import (
     SUPER_AGENT_EXPECTED_OUTPUT,
     SUPER_AGENT_INSTRUCTION,
 )
 from valuecell.core.types import UserInput
 from valuecell.utils.env import agent_debug_mode_enabled
-from valuecell.utils.model import get_model, get_model_for_agent
 
-logger = logging.getLogger(__name__)
+
+# Backward-compatible helper so tests can monkeypatch `get_model` on this module
+def get_model(agent_name: str):
+    """Return a model for the given agent via the centralized utils module.
+
+    Exposed at module-level to support test monkeypatching without touching
+    global provider configuration.
+    """
+    return model_utils_mod.get_model_for_agent(agent_name)
 
 
 class SuperAgentDecision(str, Enum):
@@ -48,15 +55,8 @@ class SuperAgent:
     def __init__(self) -> None:
         # Try to use super_agent specific configuration first,
         # fallback to PLANNER_MODEL_ID for backward compatibility
-        try:
-            model = get_model_for_agent("super_agent")
-        except Exception:
-            # Fallback to old behavior for backward compatibility
-            logger.warning(
-                "Failed to create model for super_agent, falling back to PLANNER_MODEL_ID"
-            )
-            model = get_model("PLANNER_MODEL_ID")
-
+        # Use module-level get_model indirection so tests can stub it easily
+        model = get_model("super_agent")
         self.agent = Agent(
             model=model,
             # TODO: enable tools when needed
@@ -67,6 +67,7 @@ def __init__(self) -> None:
             # output format
             expected_output=SUPER_AGENT_EXPECTED_OUTPUT,
             output_schema=SuperAgentOutcome,
+            use_json_mode=model_utils_mod.model_should_use_json_mode(model),
             # context
             db=InMemoryDb(),
             add_datetime_to_context=True,

python/valuecell/utils/model.py

@@ -13,15 +13,31 @@
 import os
 from typing import Optional
 
+from agno.models.base import Model as AgnoModel
+from agno.models.google import Gemini as AgnoGeminiModel
+
 from valuecell.adapters.models.factory import (
     create_embedder,
+    create_embedder_for_agent,
     create_model,
     create_model_for_agent,
 )
 
 logger = logging.getLogger(__name__)
 
 
+def model_should_use_json_mode(model: AgnoModel) -> bool:
+    try:
+        provider = getattr(model, "provider", None)
+        name = getattr(model, "name", None)
+        if provider == AgnoGeminiModel.provider and name == AgnoGeminiModel.name:
+            return True
+    except Exception:
+        # Any unexpected condition falls back to standard (non-JSON) mode
+        return False
+    return False
+
+
 def get_model(env_key: str, **kwargs):
     """
     Get model instance using configuration system with environment variable override.
@@ -230,6 +246,32 @@ def get_embedder(env_key: str = "EMBEDDER_MODEL_ID", **kwargs):
         raise
 
 
+def get_embedder_for_agent(agent_name: str, **kwargs):
+    """
+    Get an embedder instance configured specifically for an agent.
+
+    This mirrors `get_model_for_agent` but for embedders. It delegates to
+    the adapters/models factory which will resolve the agent's embedding
+    configuration and provider using the three-tier configuration system.
+
+    Args:
+        agent_name: Agent name matching the config file
+        **kwargs: Override parameters for this specific call
+
+    Returns:
+        Embedder instance configured for the agent
+
+    Raises:
+        ValueError: If agent configuration not found or embedder creation fails
+    """
+
+    try:
+        return create_embedder_for_agent(agent_name, **kwargs)
+    except Exception as e:
+        logger.error(f"Failed to create embedder for agent '{agent_name}': {e}")
+        raise
+
+
 def create_embedder_with_provider(
     provider: str, model_id: Optional[str] = None, **kwargs
 ):