Autodev/Autoresearch: Containerized execution for autonomous goals

[Delqhi]

Problem

Autonomous goals run shell commands in the host workspace. A buggy or malicious goal can corrupt the repo, delete files, or exfiltrate data. SOTA autonomous agents (Devin, OpenHands) run in containers.

What to do

1. internal/autonomy/container.go:
   • Detect Docker/Podman/OrbStack
   • Build lightweight sin-code-runner image
   • Run verify command inside container with workspace bind-mount
2. sin-code daemon --container flag
3. Config: autonomy.container_engine (docker|podman|orbstack)
4. Container lifecycle: create per goal, reuse for continuations, destroy after verified
5. Isolated network by default; allow egress via --network=host opt-in
6. Support for E2B-style cloud sandbox as future extension

Acceptance criteria

• Detect available container engine
• Run goal in container
• Verify command executes inside container
• Workspace bind-mount works
• Container cleaned up after goal