feat: per-model inference parameters with extended thinking

backend/scripts/seed_bootstrap_data.py

@@ -175,6 +175,38 @@ def seed_default_quota_assignment(
     return result
 
 
+# Sensible inference-param defaults for general-purpose Claude chat models.
+# Temperature 0.7 mirrors the previous always-on default and gives admins a
+# starting point they can tighten per-model. Bounds match Anthropic's accepted
+# range. max_tokens is supported but left at no-default — the model's own cap
+# applies unless an admin explicitly sets one.
+CLAUDE_CHAT_SUPPORTED_PARAMS: dict[str, Any] = {
+    "params": {
+        "temperature": {
+            "supported": True,
+            "min": Decimal("0"),
+            "max": Decimal("1"),
+            "default": Decimal("0.7"),
+            "locked": False,
+        },
+        "top_p": {
+            "supported": True,
+            "min": Decimal("0"),
+            "max": Decimal("1"),
+            "default": None,
+            "locked": False,
+        },
+        "max_tokens": {
+            "supported": True,
+            "min": Decimal("1"),
+            "max": None,
+            "default": None,
+            "locked": False,
+        },
+    }
+}
+
+
 # Default Bedrock models to seed
 DEFAULT_MODELS: list[dict[str, Any]] = [
     {
@@ -190,9 +222,9 @@ def seed_default_quota_assignment(
         "outputPricePerMillionTokens": Decimal("5.00"),
         "cacheWritePricePerMillionTokens": Decimal("1.25"),
         "cacheReadPricePerMillionTokens": Decimal("0.10"),
-        "isReasoningModel": False,
         "supportsCaching": True,
         "isDefault": True,
+        "supportedParams": CLAUDE_CHAT_SUPPORTED_PARAMS,
     },
     {
         "modelId": "global.anthropic.claude-sonnet-4-6",
@@ -207,9 +239,9 @@ def seed_default_quota_assignment(
         "outputPricePerMillionTokens": Decimal("15.00"),
         "cacheWritePricePerMillionTokens": Decimal("3.75"),
         "cacheReadPricePerMillionTokens": Decimal("0.30"),
-        "isReasoningModel": False,
         "supportsCaching": True,
         "isDefault": False,
+        "supportedParams": CLAUDE_CHAT_SUPPORTED_PARAMS,
     },
     {
         "modelId": "amazon.nova-2-sonic-v1:0",
@@ -224,9 +256,11 @@ def seed_default_quota_assignment(
         "outputPricePerMillionTokens": Decimal("12.00"),
         "cacheWritePricePerMillionTokens": Decimal("0"),
         "cacheReadPricePerMillionTokens": Decimal("0"),
-        "isReasoningModel": False,
         "supportsCaching": False,
         "isDefault": False,
+        # Voice/bidi model: param shape differs from chat models. Leave
+        # supportedParams unset so the runtime passes through to whatever
+        # the BidiAgent path negotiates.
     },
 ]
 
@@ -289,13 +323,17 @@ def seed_default_models(
             "outputPricePerMillionTokens": model_def["outputPricePerMillionTokens"],
             "cacheWritePricePerMillionTokens": model_def["cacheWritePricePerMillionTokens"],
             "cacheReadPricePerMillionTokens": model_def["cacheReadPricePerMillionTokens"],
-            "isReasoningModel": model_def["isReasoningModel"],
             "supportsCaching": model_def["supportsCaching"],
             "isDefault": model_def["isDefault"],
             "createdAt": now,
             "updatedAt": now,
         }
 
+        # Optional: per-model inference parameter capabilities. Stored as a
+        # nested map; absence means "passthrough" at runtime.
+        if "supportedParams" in model_def and model_def["supportedParams"] is not None:
+            item["supportedParams"] = model_def["supportedParams"]
+
         try:
             table.put_item(Item=item)
             msg = f"Model '{model_def['modelName']}' ({model_id}) created"

backend/src/agents/main_agent/base_agent.py

@@ -57,6 +57,7 @@ def __init__(
         caching_enabled: Optional[bool] = None,
         provider: Optional[str] = None,
         max_tokens: Optional[int] = None,
+        inference_params: Optional[Dict[str, Any]] = None,
         skip_persistence: bool = False,
     ):
         """
@@ -68,11 +69,14 @@ def __init__(
             auth_token: Raw OIDC token for forwarding to external MCP tools (optional)
             enabled_tools: List of tool IDs to enable. If None, all tools are enabled.
             model_id: Model ID to use (format depends on provider)
-            temperature: Model temperature (0.0 - 1.0)
+            temperature: Legacy. Folded into ``inference_params['temperature']`` if set.
             system_prompt: System prompt text
             caching_enabled: Whether to enable prompt caching (Bedrock only)
             provider: LLM provider ("bedrock", "openai", or "gemini")
-            max_tokens: Maximum tokens to generate (optional)
+            max_tokens: Legacy. Folded into ``inference_params['max_tokens']`` if set.
+            inference_params: Canonical-name -> value map for inference params
+                (temperature, top_p, top_k, max_tokens, thinking, ...). Wins over
+                the legacy ``temperature``/``max_tokens`` kwargs when both are set.
             skip_persistence: If True, don't persist messages (for preview sessions)
         """
         # Basic state
@@ -82,9 +86,20 @@ def __init__(
         self.enabled_tools = enabled_tools
         self.agent = None
 
-        # Initialize model configuration
+        # Merge legacy temperature/max_tokens into the canonical dict. Explicit
+        # ``inference_params`` values win over the positional kwargs so callers
+        # migrating to the new shape get predictable precedence.
+        resolved_params: Dict[str, Any] = dict(inference_params or {})
+        if temperature is not None:
+            resolved_params.setdefault("temperature", temperature)
+        if max_tokens is not None:
+            resolved_params.setdefault("max_tokens", max_tokens)
+
         self.model_config = ModelConfig.from_params(
-            model_id=model_id, temperature=temperature, caching_enabled=caching_enabled, provider=provider, max_tokens=max_tokens
+            model_id=model_id,
+            caching_enabled=caching_enabled,
+            provider=provider,
+            inference_params=resolved_params,
         )
 
         # Frozen snapshot of agent-construction params, used when the turn
@@ -95,9 +110,8 @@ def __init__(
             "enabled_tools": enabled_tools,
             "model_id": model_id,
             "provider": provider,
-            "temperature": temperature,
             "caching_enabled": caching_enabled,
-            "max_tokens": max_tokens,
+            "inference_params": dict(resolved_params),
         }
 
         # Load retry configuration from environment variables

backend/src/agents/main_agent/config/constants.py

@@ -70,7 +70,6 @@ class Defaults:
 
     # --- Model ---
     MODEL_ID = "us.anthropic.claude-haiku-4-5-20251001-v1:0"
-    TEMPERATURE = 0.7
     CACHING_ENABLED = True
 
     # --- AWS ---