feat(agent): add sandbox client for gVisor-isolated code execution

[EtnDiaz]

What

Adds flymyai.agent — a thin client over the agent-sandbox REST API for running code in isolated Kubernetes pods.

Usage

from flymyai.agent import execute_code

result = execute_code(
    sandbox_url="http://agent-sandbox.sandboxes.svc:8080",
    code="print(sum(range(100)))",
)
print(result.stdout)    # 4950
assert result.ok

Async:

from flymyai.agent import async_execute_code

result = await async_execute_code(sandbox_url="...", code="print(42)")

Context managers (SandboxClient, AsyncSandboxClient) are also available for multiple executions sharing one HTTP client.

Files

• flymyai/agent/sandbox.py — SandboxClient, AsyncSandboxClient, SandboxResult, execute_code(), async_execute_code()
• flymyai/agent/__init__.py — public exports
• README.md — Agent Sandbox section

Notes

• No new dependencies — uses httpx which is already in pyproject.toml
• The sandbox service (agent-sandbox) must be deployed separately: gitlab.flymy.ai/fma/core/agent-sandbox
• Supported languages: python, bash, javascript, sh

[EtnDiaz]

Closing — the client should call custom MCP tools through the backend API, not directly to agent-sandbox. Will reopen with the correct approach after the backend endpoint lands.