Amazon Bedrock Connectivity

Relevant source files

The following files were used as context for generating this wiki page:

• CLAUDE.md
• src/agent/setup.ts
• deploy/cloudformation.yml
• src/main.ts
• package.json
• src/agent/prompt.ts

Amazon Bedrock Connectivity serves as the core Large Language Model (LLM) integration layer for the pi-digby Slack bot. It utilizes the @mariozechner/pi-ai provider to interface with Bedrock models, specifically Claude Sonnet 4.6, providing the agent with its reasoning and tool-use capabilities.

The system is designed to run within AWS infrastructure, leveraging IAM roles for secure, credential-less authentication when deployed on ECS Fargate. It supports streaming responses, tool execution loops, and multimodal input (image processing) through the Bedrock API.

Architecture and Integration

The connectivity layer acts as a bridge between the local agent logic and the AWS Bedrock service. It is initialized during the agent setup phase and is consumed by the Agent class to process prompts.

Component Relationship

The following diagram illustrates how Amazon Bedrock integrates with the internal agent components:

The integration is primarily managed in src/agent/setup.ts, where the model is selected and the authentication strategy is defined.
Sources: src/agent/setup.ts:40-45, CLAUDE.md:1-15

Core Components Table

Component	Responsibility	Source Reference
Model Provider	amazon-bedrock via @mariozechner/pi-ai	src/agent/setup.ts:42
Model ID	us.anthropic.claude-sonnet-4-6	src/agent/setup.ts:42
Auth Strategy	ECS Task Role / IAM Credentials	src/agent/setup.ts:47-61
Streaming	Enabled by default for real-time Slack updates	CLAUDE.md:58

Authentication and Security

Connectivity to Bedrock is secured through AWS Identity and Access Management (IAM). The application supports multiple credential resolution paths, prioritizing secure environment-based roles over static keys.

Credential Resolution

The getBedrockApiKey function checks for credentials in the following order:

1. AWS_PROFILE environment variable.
2. AWS_ACCESS_KEY_ID and AWS_SECRET_ACCESS_KEY pair.
3. AWS_CONTAINER_CREDENTIALS_RELATIVE_URI (automatically provided by ECS Fargate Task Roles).

Sources: src/agent/setup.ts:47-61, CLAUDE.md:50-52

IAM Policy Requirements

For the agent to function, the ECS Task Role (pi-digby-task) must have explicit permissions to invoke Bedrock models.

# Simplified policy from CloudFormation
Policies:
  - PolicyName: bedrock-invoke
    PolicyDocument:
      Statement:
        - Effect: Allow
          Action:
            - bedrock:InvokeModel
            - bedrock:InvokeModelWithResponseStream
          Resource: "*"

Sources: deploy/cloudformation.yml:195-204

Data Flow and Message Processing

When a user sends a message in Slack, the system prepares a payload for Bedrock that includes system instructions, conversation history, and any relevant media.

Request Orchestration

The following sequence diagram shows the flow of a single request to Amazon Bedrock:

Image Handling

Bedrock connectivity supports multimodal inputs. Images uploaded to Slack are:

1. Downloaded to the local /data/attachments directory.
2. Resized via utils/image-resize.js.
3. Converted to Base64 and wrapped in ImageContent objects before being sent to Bedrock.

Sources: src/agent/setup.ts:223-260, src/main.ts:105-125

Infrastructure Configuration

The connectivity is supported by specific AWS infrastructure settings defined in the deployment templates.

Parameter	Value / Description	Source
Region	us-east-1 (Default)	deploy/cloudformation.yml:345
Compute	ECS Fargate (vCPUs: 1024, RAM: 2048)	deploy/cloudformation.yml:333-334
Environment	AWS_DEFAULT_REGION passed to container	deploy/cloudformation.yml:344-345

Error Handling and Timeouts

The connectivity layer implements robust error handling to manage the ephemeral nature of API calls.

• Retries: While the Bedrock provider handles low-level retries, the AgentSession manages the tool loop stability.
• Timeouts: A run timeout is enforced (configured via getRunTimeout()) to prevent hanging connections. If a timeout occurs, session.abort() is called to terminate the Bedrock stream.
• Exception Mapping: API errors (like max_tokens or timeout) are caught in the run method and translated into user-friendly Slack notifications (e.g., "Ran out of space mid-response").

Sources: src/agent/setup.ts:285-315, src/main.ts:153-158

Conclusion

Amazon Bedrock Connectivity is the foundational intelligence engine of pi-digby. By integrating Claude 3.5 Sonnet through the AWS SDK (via pi-ai), the system achieves a balance of high-reasoning capability and secure, scalable infrastructure execution. The use of streaming and multimodal support allows for a responsive and feature-rich Slack assistant experience.