Repository navigation
[FEATURE] Amazon Nova tool caching with cachePoint in messages #449
Description
Problem Statement
I would like strands to cache tool definitions when using Amazon Nova models.
Today Amazon Nova models do not accept prompt cache checkpoints in tools field. However Amazon Nova will cache the tool configuration if there is a cache checkpoint set in the messages payload. Strands should set that cache checkpoint for me when I specify cache_tools so that the tool configuration is cached.
Proposed Solution
When setting cache_tools with BedrockModel and model_id is an Amazon Nova model, i.e.
bedrock_model = BedrockModel(
model_id="us.amazon.nova-premier-v1:0",
cache_tools="default"
)then strands should add a messages { cachePoint: { type: ‘default’ } } before adding the user messages to cache the tool configuration.
Make the following Bedrock API call when cache_tools is set and the model is an Amazon Nova model:
tool_config = { ...}
messages = [
{
"role": "user",
"content": [
{"text": "You are a helpful AI assistant. Use the tools to solve the users questions."},
{ "cachePoint": { "type": 'default' } },
.... other messages
],
},
]
response = client.converse(
modelId=[...],
messages=messages,
toolConfig=tool_config
)Use Case
This is useful to cache long tool definitions.
Alternatives Solutions
No response
Additional Context
No response
Activity
Unshure commentedon Jul 16, 2025
Hi @malte-aws, thanks for reaching out!
I'm hesitant to add the feature to automatically add a
cachePointmessage to the messages array based on the user setting thecache_toolsince the behavior of acachePointmessage caching a tool_specs seems to be an undocumented behavior (I couldnt find it in the docs you shared), and adding acachePointmessage can lead to unexpected behaviors for customers.Similar to the sudo-code you shared above, you can pass in a
cachePointmessage yourself as part of the messages array passed to the agent: https://strandsagents.com/latest/documentation/docs/user-guide/concepts/model-providers/amazon-bedrock/#messages-cachingLet me know if this helps!
fix(tests): pin MCP SDK to 1.25.2 (strands-agents#449)