Runtime image
The runtime image is the container the agent works in. Every turn's sandbox starts from it, under gVisor and without network access, so everything a repository's build and tests need must already be in the image or in the repository.
The default image
| Base | Debian bookworm with Python 3.12 |
| Node.js | 22, with npm and npx |
| Python tools | uv, and the packages Yi Code runs on, including pytest |
| Other tools | git, bash, tmux |
| Agent user | agent, uid 1000, not root |
| Platform | Linux amd64 |
It has no Go, Rust, Java or Gradle toolchain and no make. Repositories that need those get "Unverified" with "The execution environment lacks the required toolchain." See Verification.
Each sandbox gets 2 CPUs and 4 GiB of memory.
Choosing the image
The deployment administrator sets the image with YI_RUNTIME_IMAGE, pinned by its digest. Each turn records the image it started with.
- Cloud service: the image is ours. Organizations can't change it.
- Private deployment: the bundle sets
YI_RUNTIME_IMAGEto the bundled image when it installs. You can build your own, for example with your language toolchains and the dependencies your tests need preinstalled.
A custom image must be verified again
A custom image changes what runs inside the sandbox. Before any pool uses it, it must pass the same checks as ours on a real execution host:
- The native tool transport works under gVisor (
runsc), with terminal and file editing. - The agent runs as uid 1000 and can't write the protected Git baseline or the original tests.
- The sandbox can't reach the internet, the cloud metadata service or listening ports on the host.
- Verification and export work, and the sandbox is removed afterwards.
An execution pool is enabled only after its host passes these checks. A container build that passes in CI is not a substitute.