[FEATURE] Support granular caching (cache points) in system prompts

[AlexanderRichey]

Problem Statement

Strands currently supports "default" caching of system prompts, but it does not allow users to control which parts of the system prompt should and should not be cached. When using dynamic system prompts, where some of the prompt is static, but a portion is dynamic, the ability to control which portion of the prompt gets cached is essential to be able to leverage prompt caching.

Proposed Solution

Bedrock supports granular caching today by accepting an array of content blocks in its system prompt parameter. For example:

[
  {"text": "static portion of my prompt"},
  {"cachePoint": {"type": "default"}},
  {"text": "dynamic portion of my prompt"},
]

Strands should consider accepting such an array for its own system prompt param. Here's how this could look:

a = Agent(
  model=BedrockModel(
    model_id="anthropic.claude-sonnet-4-20250514-v1:0",
  ),
  system_prompt=[
    {"text": "static portion of my prompt"},
    {"cachePoint": {"type": "default"}},
    {"text": f"dynamic portion of my prompt: {something_random()}"},
  ],
)

Use Case

See above.

Alternatives Solutions

No response

Additional Context

No response

[flowconic]

+1 for this feature

@AlexanderRichey , did you ever find a work-around meanwhile?

[flowconic]

Wondering if this would be as "simple" as updating the file strands/models/bedrock.py and method format_request
FROM

def format_request(
        self,
        messages: Messages,
        tool_specs: Optional[list[ToolSpec]] = None,
        system_prompt: Optional[str] = None,
    ) -> dict[str, Any]:
        """Format a Bedrock converse stream request.

        Args:
            messages: List of message objects to be processed by the model.
            tool_specs: List of tool specifications to make available to the model.
            system_prompt: System prompt to provide context to the model.

        Returns:
            A Bedrock converse stream request.
        """
        return {
            "modelId": self.config["model_id"],
            "messages": self._format_bedrock_messages(messages),
            "system": [
                *([{"text": system_prompt}] if system_prompt else []),
                *([{"cachePoint": {"type": self.config["cache_prompt"]}}] if self.config.get("cache_prompt") else []),
            ],
            **(
                {
                    "toolConfig": {
...

TO something like

def format_request(
        self,
        messages: Messages,
        tool_specs: Optional[list[ToolSpec]] = None,
        system_prompt: Optional[str] = None,
    ) -> dict[str, Any]:

        return {
            "modelId": self.config["model_id"],
            "messages": self._format_bedrock_messages(messages),

            # THIS part is proposed update
            "system": [
                *(
                    [{"text": system_prompt}] if isinstance(system_prompt, str) and system_prompt
                    else system_prompt if isinstance(system_prompt, list)
                    else []
                ),
                *([{"cachePoint": {"type": self.config["cache_prompt"]}}] if self.config.get("cache_prompt") else []),
            ],
            # Proposed update end

            **(
                {
                    "toolConfig": {
....

[haandol]

+1 for the PFR

[dbschmigelski]

Hey, @AlexanderRichey, this should be resolved, for Bedrock and LiteLLM models at least, after we released #1112. We are tracking applying this to other models in #1140