Direct code deploy: uv fails with invalid value 'aarch64-manylinux2014' for '--python-platform' when using duckdb >= 1.3.2 from macOS

[alexander-grun]

Summary

When using Amazon Bedrock AgentCore Runtime direct code deployment from macOS with duckdb >= 1.3.2 in requirements.txt, the dependency installation step with uv fails during cross-compilation for Linux ARM64.

The AgentCore CLI appears to invoke uv with:

--python-platform aarch64-manylinux2014

but the installed uv version does not recognize that value and exits with:

error: invalid value 'aarch64-manylinux2014' for '--python-platform <PYTHON_PLATFORM>'

tip: a similar value exists: 'aarch64-manylinux_2_40'

As a result, agentcore deploy / agentcore launch fails before the code is even uploaded/runs in the AgentCore runtime.

This makes it difficult to use native dependencies like duckdb >= 1.3.2 with direct code deploy from macOS.

---

Environment

Dev machine:

• macOS (Apple Silicon)
• Python 3.12
• uv installed via the official installer:
  curl -LsSf https://astral.sh/uv/install.sh | sh
• bedrock-agentcore / AgentCore CLI installed via pip

AgentCore configuration (excerpt from .bedrock_agentcore.yaml):

default_agent: cashflow_agent
agents:
  cashflow_agent:
    name: cashflow_agent
    entrypoint: cashflow_agent.py
    deployment_type: direct_code_deploy
    runtime_type: PYTHON_3_12
    platform: linux/arm64
    container_runtime: null
    source_path: .
    aws:
      execution_role: <redacted-iam-role-arn>
      execution_role_auto_create: true
      account: '<redacted-account-id>'
      region: <redacted-region>
      ecr_repository: null
      ecr_auto_create: false
      s3_path: <redacted-s3-path>
      s3_auto_create: false
      network_configuration:
        network_mode: PUBLIC
        network_mode_config: null
      protocol_configuration:
        server_protocol: HTTP
      observability:
        enabled: true
      lifecycle_configuration:
        idle_runtime_session_timeout: null
        max_lifetime: null
    bedrock_agentcore:
      agent_id: cashflow_agent-<redacted>
      agent_arn: arn:aws:bedrock-agentcore:<redacted-region>:<redacted-account-id>:runtime/cashflow_agent-<redacted>
      agent_session_id: null
    codebuild:
      project_name: null
      execution_role: null
      source_bucket: null
    memory:
      mode: NO_MEMORY
      memory_id: null
      memory_arn: null
      memory_name: null
      event_expiry_days: 30
      first_invoke_memory_check_done: false
      was_created_by_toolkit: false
    identity:
      credential_providers: []
      workload: null
    authorizer_configuration: null
    request_header_configuration: null
    oauth_configuration: null

(If you prefer, you can simplify this further to just show the relevant bits: deployment_type, runtime_type, platform.)

Dependencies (relevant part of requirements.txt):

duckdb>=1.3.2
bedrock-agentcore
# other deps...

---

Steps to Reproduce

1. On macOS (Apple Silicon), create an AgentCore project configured for direct code deployment with:

   • deployment_type: direct_code_deploy
   • runtime_type: PYTHON_3_12
   • platform: linux/arm64

2. Add duckdb>=1.3.2 to requirements.txt (along with bedrock-agentcore and other deps).

3. Run:

   agentcore deploy
   # or
   agentcore launch

4. Observe the output during the dependency installation step:

   Installing dependencies with uv (cross-compiling for Linux ARM64)...
   ❌ Failed to install dependencies with uv: error: invalid value 'aarch64-manylinux2014' for '--python-platform <PYTHON_PLATFORM>'
   
     tip: a similar value exists: 'aarch64-manylinux_2_40'
   
   For more information, try '--help'.
   

---

Actual Behavior

• The AgentCore CLI attempts to install dependencies using uv with a cross-platform target for Linux ARM64.

• It passes --python-platform aarch64-manylinux2014 to uv.

• The uv version installed on macOS rejects this value as invalid and terminates with:

  error: invalid value 'aarch64-manylinux2014' for '--python-platform <PYTHON_PLATFORM>'
  
  tip: a similar value exists: 'aarch64-manylinux_2_40'
  

• The deployment fails at the packaging / dependency-resolution stage, so:

  • No code is deployed to AgentCore Runtime.
  • The error happens before any code runs in the Amazon Linux ARM64 environment.

This makes it impossible (with the current tooling) to use duckdb>=1.3.2 with direct code deploy from macOS without manual workarounds.

---

Expected Behavior

From a user perspective, for direct code deployment with platform: linux/arm64 on macOS:

• agentcore deploy / agentcore launch should:

  • Successfully build a Linux ARM64-compatible deployment package (using uv or another supported tool), or
  • At least provide a clear, supported path / command to do so.

Concretely:

• Either:

  • AgentCore pins or bundles a uv version that accepts aarch64-manylinux2014, or
  • AgentCore uses a --python-platform value that is valid for current uv releases and matches the Amazon Linux 2023 ARM64 runtime’s glibc / manylinux baseline.

• And the documentation should clearly describe the recommended flow for:

  “I’m on macOS, using direct code deploy with platform: linux/arm64, and I depend on a package with native Linux ARM64 wheels (like duckdb>=1.3.2).”

---

Why This Matters

• Direct code deploy is positioned as the quickest way to iterate with AgentCore, including from developer laptops (macOS).
• As soon as a native dependency like duckdb>=1.3.2 is added, the current cross-compilation logic with uv fails on macOS.
• The error is not about my code or DuckDB per se, but about the hard-coded --python-platform value passed by the AgentCore tooling conflicting with current uv behavior.

This forces users to:

• Either downgrade DuckDB or avoid it,
• Or manually build dependencies in an Amazon Linux ARM64 environment (EC2 Graviton, Docker), which removes much of the convenience of direct code deploy.

---

Questions / Requests

1. What is the officially supported --python-platform / manylinux target for the AgentCore Runtime direct code environment (Amazon Linux 2023 ARM64)?

2. Can the AgentCore CLI be updated to:

   • Use a --python-platform value that is valid for current uv versions and matches the runtime, or
   • Pin / vendor a uv version that supports aarch64-manylinux2014, or
   • Provide a dedicated agentcore build --platform linux/arm64 command that uses Docker with the same base image as the runtime to build dependencies?

3. In the meantime, is there a recommended workaround for macOS users who:

   • Want to use direct code deploy,
   • Need duckdb >= 1.3.2 or similar native libraries,
   • And don’t want to switch to full container-based deployment?

[alexander-grun]

tl/dr: Direct code deploy: CLI uses deprecated aarch64-manylinux2014 uv platform tag (rejected by current uv)