[FEATURE] Add GuardrailConverseContent Support for Bedrock Guardrails to supportive selective Guarding

[kethamak]

Problem Statement

Feature: Bedrock Converse – support GuardrailConverseContent blocks + escape hatch

Background

Bedrock's Converse API lets callers wrap only the text that should be moderated in

{ "type": "guardrailConverseContent", "guardContent": { "text": "…" } }

This is critical for low-latency chat UIs because the guardrail ignores prior turns.
Docs: https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails-use-converse-api.html#evaluate-specific-content

Current behaviour -- All history is sent; no way to wrap only the latest message in GuardrailConverseContentBlock

from strands import Agent
from strands.models import BedrockModel

bedrock_model = BedrockModel(
    model_id="us.amazon.nova-pro-v1:0",
    guardrail_id="gr-abc123",
    guardrail_version="DRAFT",
)
agent = Agent(model=bedrock_model)
response = agent("User message")

1. Strands serialises every message as a plain text block.
2. Result: the guardrail evaluates the entire conversation, adding cost & latency.

Proposed Solution

Here's the Expected / requested behaviour

Add a guard_last_turn_only flag to BedrockModel that wraps only the latest user message in a guardrailConverseContent block, leaving conversation history unguarded.

bedrock_model = BedrockModel(
    model_id="us.amazon.nova-pro-v1:0",
    guardrail_id="gr-abc123",
    guardrail_version="DRAFT",
    guard_last_turn_only=True  # <-- NEW FLAG
)
agent = Agent(model=bedrock_model)
response = agent("User message")

Advanced control when needed

Add a raw_blocks kwarg that lets advanced users pass a list of fully-formed Bedrock content-block dicts; Strands will relay them unchanged, enabling any combination of ContentBlock types—including custom guardrail
wrapping—without waiting for wrapper updates.

bedrock_model = BedrockModel(
    model_id="us.amazon.nova-pro-v1:0",
    guardrail_id="gr-abc123",
    guardrail_version="DRAFT",
)
agent = Agent(model=bedrock_model)
response = agent.run(
    raw_blocks=[
        {"type":"text","text":"System prompt"},
        {"type":"text","text":"Un-guarded history"},
        {"type":"guardrailConverseContent",
         "guardContent":{"text":"Only this gets moderated"}}
    ]
)

Use Case

I'm building a low-latency, multi-turn chat product on Bedrock with Strands, and I can't use Bedrock's selective-guarding feature because the wrapper always sends the whole conversation; adding a simple
guard_last_turn_only flag (plus an escape-hatch raw_blocks param) would let me moderate just the latest user input or selectively guard inputs, cutting latency, cost, and false-positive risk.

Alternatives Solutions

No response

Additional Context

The same issue has already been fixed in LiteLLM and LangChain: BerriAI/litellm#12676, langchain-ai/langchain-aws#540

[mbuotidem]

Struggling with this as well, e.g, when using structured_output() to parse user questions, I need to guard only the user's input, not my system prompt:

user_input = "How do I reset my credentials?"

result = agent.structured_output(
    Schema,
    raw_blocks=[
        {"type": "text", "text": "Classify this question into categories."},
        {"type": "guardrailConverseContent", "guardContent": {"text": user_input}},
        {"type": "text", "text": """Categories: [HR, Security]"""}
    ]
)

[joeladdison]

This enhancement will help out a lot, letting us only apply guardrails to new content and skipping previously checked content. Looking forward to seeing this added.