OpenShell/architecture/sandbox.md

Sandbox

A sandbox is the runtime boundary where agent code executes. It is created by a compute runtime and managed inside the workload by openshell-sandbox, the sandbox supervisor.

Runtime Model

Each sandbox workload has two trust levels:

Process	Role
Supervisor	Starts as root inside the workload, prepares isolation, runs the proxy, fetches config, injects credentials, serves the relay socket, and launches child processes.
Agent child	Runs as an unprivileged user with filesystem, process, and network restrictions applied.

The supervisor keeps enough privilege to manage the sandbox, but the agent child loses that privilege before user code runs.

Startup Flow

1. The compute runtime starts the workload with sandbox identity, callback endpoint, TLS or secret material, image metadata, and initial command.
2. The supervisor loads policy and runtime settings from local files or the gateway, depending on mode.
3. It prepares filesystem access, process restrictions, network namespace routing, trust stores, provider credential resolution, and inference routes.
4. It starts the policy proxy and local SSH server.
5. It opens a supervisor session back to the gateway for connect, exec, file sync, config polling, and log push.
6. It launches the agent command as the restricted sandbox user.

Isolation Layers

OpenShell uses overlapping controls rather than a single sandbox primitive:

Layer	Purpose
Filesystem policy	Landlock restricts the paths the agent can read or write.
Process policy	The child process runs as a non-root user with reduced privileges.
Seccomp	Blocks dangerous syscalls, including raw socket paths that bypass the proxy.
Network namespace	Forces ordinary agent egress through the local CONNECT proxy.
Policy proxy	Evaluates destination, binary identity, TLS/L7 rules, SSRF checks, and inference interception.

The supervisor may enrich baseline filesystem allowances for runtime-required paths, such as proxy support files or GPU device paths when a GPU is present.

Network and Inference

All ordinary agent egress is routed through the sandbox proxy. The proxy identifies the calling binary, checks trust-on-first-use binary identity, rejects unsafe internal destinations, and evaluates the active policy.

https://inference.local is special. It bypasses OPA network policy and is handled by the inference interception path:

1. The proxy terminates the local TLS connection with the sandbox CA.
2. It detects known OpenAI, Anthropic, and compatible inference request shapes.
3. It strips caller-supplied credentials and disallowed headers.
4. It forwards through openshell-router using the route bundle fetched from the gateway.

External inference endpoints that do not use inference.local are treated like ordinary network traffic and must be allowed by policy.

Credentials

Provider credentials are stored at the gateway and fetched by the supervisor at runtime. The supervisor injects resolved environment variables into the initial agent process and SSH child processes. Driver-controlled environment variables override template values so sandbox images cannot spoof identity, callback, or relay settings.

Credential placeholders in proxied HTTP requests can be resolved by the proxy when policy allows the target endpoint. Secrets must not be logged in OCSF or plain tracing output.

Connect and Logs

The supervisor runs an SSH server on a Unix socket inside the sandbox. The gateway reaches it through the outbound supervisor relay, not by dialing the sandbox workload directly. The relay supports:

• Interactive shell sessions.
• Command execution.
• Tar-based file sync.
• Port forwarding where supported by the CLI/TUI surface.

Sandbox logs are emitted locally and can also be pushed back to the gateway. Security-relevant sandbox behavior uses OCSF structured events; internal diagnostics use ordinary tracing.

Failure Behavior

• If gateway config polling fails, the sandbox keeps its last-known-good policy.
• If a live policy update is invalid, the supervisor rejects it and keeps the current policy.
• Existing raw byte streams are connection scoped. Dynamic policy changes apply to new connections or the next parsed HTTP request where the proxy can safely re-evaluate.
• If the supervisor relay drops, the sandbox can keep running, but connect and exec operations fail until the supervisor registers again.