Skip to main content

Monty sandbox provider

The repository's npm ci installs the pinned dependency as a plugin workspace. The bridge needs Node.js (20+ for the SDK, 22+ for this repository), including when the compositor is compiled with Bun. AGENT_NODE_BIN selects its executable. For an independently copied plugin, run npm ci --prefix plugins/monty --workspaces=false. The application loads the plugin without importing the native dependency; Monty is started lazily for each execution. Missing dependencies return an actionable tool error. Neither the core nor other plugins import Monty.

run_python supplies Python source and optional JSON inputs to service.sandbox.run. The final expression is the output; print is captured. The default backend uses the pinned pure-Monty worker from @pydantic/monty, with no directory mounts, OS callback, shell, network or environment access. This sandbox runs Python; it does not isolate arbitrary host shell commands. Trusted TypeScript plugin modules still run on the host.

Defaults: 32 MB interpreter memory, 10 seconds interpreter execution per feed, 100 recursion depth, 1000 suspensions, no sleep, 64 KB captured print output, and a 30-second wall deadline. Callers can set timeoutMs (100–300000) and memoryBytes (1–128 MB). Limits apply to interpreter allocation, not the complete host process RSS. RPC messages are limited to 1 MB and total worker output to 2 MB. Timeouts and cancellation terminate the bridge's entire process group, including the native worker on supported POSIX systems.

Trusted host callers may explicitly supply a functions record. Each callback receives (args: unknown[], signal: AbortSignal). Generated code can only call these named capabilities. Callbacks must observe cancellation and enforce their own permissions. They are process-local values, not a wire format.

The workflows plugin supplies task and log; task subagents use the normal permission chain and inherit cancellation. Python does not gain shell access from having a task callback. A child can request shell tools subject to the ordinary permission gate; those tools still run on the host.

To replace Monty, implement service.sandbox.run in another provider plugin and set serviceProviders in application configuration:

{"serviceProviders":{"service.sandbox.run":"other-sandbox"}}

The contract returns {value:{output,stdout}} or {deny:"reason"}. Use disabledPlugins: ["monty"] as well if the replacement supplies its own tools. No installed provider means execution fails closed.