deepagents-contrib-aws 0.1.2

AWS contributions for Deep Agents

deepagents-contrib-aws

AWS backend implementations for the deepagents framework.

Backends

S3Backend

An S3-backed implementation of the BackendProtocol that persists agent workspace files in Amazon S3. Supports all eight protocol methods: ls, read, write, edit, grep, glob, upload_files, download_files.

Installation

pip install deepagents-contrib-aws

Quick Start

From constructor

Minimal — S3 client is created implicitly; region and credentials are picked up from environment variables or ~/.aws/config:

from deepagents_contrib_aws import S3Backend

backend = S3Backend(bucket="my-bucket", prefix="agent/workspace/")

Explicit — pass a pre-configured boto3 client (useful for custom endpoints like LocalStack, S3-compatible storage, or reusing an existing session):

import boto3
from deepagents_contrib_aws import S3Backend

client = boto3.client("s3", region_name="us-west-2")
backend = S3Backend(
    bucket="my-bucket",
    prefix="agent/workspace/",
    client=client,
)

From environment variables

export S3_BACKEND_BUCKET=my-bucket
export S3_BACKEND_PREFIX=agent/workspace/
export AWS_REGION=us-west-2  # or AWS_DEFAULT_REGION

from deepagents_contrib_aws import S3Backend

backend = S3Backend.from_env()

With deepagents

from deepagents import create_deep_agent
from deepagents_contrib_aws import S3Backend

backend = S3Backend.from_env()
agent = create_deep_agent(backend=backend)

Basic operations

backend = S3Backend(bucket="my-bucket", prefix="demo/")

# Write a file (errors if file already exists)
result = backend.write("/hello.py", "print('hello')")

# Read it back (with line-based pagination)
result = backend.read("/hello.py")

# Edit it (exact string replacement)
result = backend.edit("/hello.py", "hello", "world")

# List directory
result = backend.ls("/")

# Search file contents (literal text, not regex)
result = backend.grep("world", path="/")

# Find files by pattern
result = backend.glob("*.py")

# Bulk upload
result = backend.upload_files([("/a.txt", b"content a"), ("/b.txt", b"content b")])

# Bulk download
result = backend.download_files(["/a.txt", "/b.txt"])

Configuration

Constructor parameters

Parameter	Type	Default	Description
bucket	str	required	S3 bucket name
prefix	str	""	Key prefix for all objects (e.g. "agent/workspace/")
client	boto3 S3 client	None	Optional pre-configured boto3 client
region_name	str	None	AWS region for the default client

Environment variables

Variable	Required	Default	Description
S3_BACKEND_BUCKET	Yes (for from_env())	--	S3 bucket name
S3_BACKEND_PREFIX	No	""	Key prefix for all objects
AWS_REGION	No	--	AWS region (checked first; auto-set by AWS Lambda)
AWS_DEFAULT_REGION	No	--	AWS region (standard boto3 fallback if AWS_REGION is not set)

AWS credentials are resolved via the standard boto3 credential chain: environment variables (AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY, AWS_SESSION_TOKEN), shared credentials file (~/.aws/credentials), AWS SSO, or IAM role.

Development

Setup

# Install uv
curl -LsSf https://astral.sh/uv/install.sh | sh

# Create venv and install all dependencies
uv sync

# Lint
uv run ruff check src/ tests/

# Build
uv build

Testing

The project has two test suites: unit tests (mocked S3 via moto) and integration tests (real S3 bucket).

Unit tests

Unit tests use moto to mock all S3 API calls. No AWS credentials or network access required.

uv run pytest

This runs 58 tests covering all 8 protocol methods, path helpers, error mapping, from_env() factory, and edge cases (binary files, empty files, pagination, partial batch failures).

Integration tests

Integration tests run against a real S3 bucket to validate actual AWS behavior (pagination, error codes, eventual consistency). They are skipped by default.

Prerequisites:

1. An existing S3 bucket with read/write access
2. AWS credentials configured via any standard method:
   • Environment variables (AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY)
   • Shared credentials file (~/.aws/credentials)
   • AWS SSO (aws sso login)
   • IAM role (EC2, ECS, Lambda)
3. IAM permissions on the test bucket:
   • s3:GetObject
   • s3:PutObject
   • s3:DeleteObject (for cleanup after tests)
   • s3:ListBucket

Running:

S3_TEST_BUCKET=your-bucket-name uv run pytest -m integration

Optionally set the region if the bucket is not in us-east-1:

S3_TEST_BUCKET=your-bucket-name AWS_DEFAULT_REGION=us-west-2 uv run pytest -m integration

Run both unit and integration tests together:

S3_TEST_BUCKET=your-bucket-name uv run pytest -m ""

Test isolation: Integration tests create objects under a unique prefix (integration-test-<uuid>/) and clean up after themselves. They will not leave artifacts in your bucket.

License

MIT