[API Updates] Model / shield / memory-bank routing + agent persistence + support for private headers (#92)

This is yet another of those large PRs (hopefully we will have less and less of them as things mature fast). This one introduces substantial improvements and some simplifications to the stack.

Most important bits:

* Agents reference implementation now has support for session / turn persistence. The default implementation uses sqlite but there's also support for using Redis.

* We have re-architected the structure of the Stack APIs to allow for more flexible routing. The motivating use cases are:
  - routing model A to ollama and model B to a remote provider like Together
  - routing shield A to local impl while shield B to a remote provider like Bedrock
  - routing a vector memory bank to Weaviate while routing a keyvalue memory bank to Redis

* Support for provider specific parameters to be passed from the clients. A client can pass data using `x_llamastack_provider_data` parameter which can be type-checked and provided to the Adapter implementations.

llama_stack/providers/impls/meta_reference/agents/safety.py

@@ -4,51 +4,48 @@
 # This source code is licensed under the terms described in the LICENSE file in
 # the root directory of this source tree.
 
+import asyncio
+
 from typing import List
 
-from llama_models.llama3.api.datatypes import Message, Role, UserMessage
+from llama_models.llama3.api.datatypes import Message
 from termcolor import cprint
 
-from llama_stack.apis.safety import (
-    OnViolationAction,
-    Safety,
-    ShieldDefinition,
-    ShieldResponse,
-)
+from llama_stack.apis.safety import *  # noqa: F403
 
 
 class SafetyException(Exception):  # noqa: N818
-    def __init__(self, response: ShieldResponse):
-        self.response = response
-        super().__init__(response.violation_return_message)
+    def __init__(self, violation: SafetyViolation):
+        self.violation = violation
+        super().__init__(violation.user_message)
 
 
 class ShieldRunnerMixin:
     def __init__(
         self,
         safety_api: Safety,
-        input_shields: List[ShieldDefinition] = None,
-        output_shields: List[ShieldDefinition] = None,
+        input_shields: List[str] = None,
+        output_shields: List[str] = None,
     ):
         self.safety_api = safety_api
         self.input_shields = input_shields
         self.output_shields = output_shields
 
-    async def run_shields(
-        self, messages: List[Message], shields: List[ShieldDefinition]
-    ) -> List[ShieldResponse]:
-        messages = messages.copy()
-        # some shields like llama-guard require the first message to be a user message
-        # since this might be a tool call, first role might not be user
-        if len(messages) > 0 and messages[0].role != Role.user.value:
-            messages[0] = UserMessage(content=messages[0].content)
-
-        results = await self.safety_api.run_shields(
-            messages=messages,
-            shields=shields,
+    async def run_multiple_shields(
+        self, messages: List[Message], shields: List[str]
+    ) -> None:
+        responses = await asyncio.gather(
+            *[
+                self.safety_api.run_shield(
+                    shield_type=shield_type,
+                    messages=messages,
+                )
+                for shield_type in shields
+            ]
         )
-        for shield, r in zip(shields, results):
-            if r.is_violation:
+
+        for shield, r in zip(shields, responses):
+            if r.violation:
                 if shield.on_violation_action == OnViolationAction.RAISE:
                     raise SafetyException(r)
                 elif shield.on_violation_action == OnViolationAction.WARN:
@@ -56,5 +53,3 @@ class ShieldRunnerMixin:
                         f"[Warn]{shield.__class__.__name__} raised a warning",
                         color="red",
                     )
-
-        return results