org.llm4s.agent.harness.SandboxBackend
See theSandboxBackend companion object
@Experimental
Where commands run. Its target is Unverifiable (C4): every execute runs through ReplayGuard.unverifiable under ctx.effectKey("sandbox.execute", command, workdir), never as idempotent, so a completed command replays its stored ExecResult and one whose run may have begun without recording its outcome is refused (GraphError.EffectReplayRefused, a tool error), never run again. Each stream is cut, on a character boundary, so that its JSON encoding - what the guard records - is at most maxOutputBytes, before the outcome is recorded; so a recorded ExecResult always fits C4's record, whatever the output's characters (rule 12). A command's effects stay when a Block retracts its turn (rule 5).
Attributes
- Companion
- object
- Graph
-
- Supertypes
-
class Objecttrait Matchableclass Any
Members list
In this article