ExecuteTool

The execute tool (design §4.33): runs a command in a SandboxBackend, and knows nothing else about where. Through the backend, every command runs under C4's Unverifiable guard: a completed command replays its stored result, one that may have run is refused rather than run again, and a Block cannot undo what a command did (rule 5).

• '''Arguments.''' command; workdir, a path such as /src (default /); timeout_seconds (default defaultTimeout, at most maxTimeout).
• '''Permissions.''' A call touches ToolResource(Operation.Execute, <command>), so a PermissionMiddleware rule matches on the command; the hints are destructive and openWorld.
• '''Order.''' A call holds backend.locks(workdir) (a workspace sandbox: its whole workspace), so a batch's commands and the edits of the files they may touch apply in call order.
• '''Bounds.''' Each stream is at most the backend's maxOutputBytes, the command at most SandboxBackend.MaxCommandBytes, the run at most the timeout; the tool itself is stopped maxTimeout plus a minute in, for a sandbox that never answers.

The result is the exit code, stdout and stderr, and a note when output was cut. A refusal (by the sandbox's policy, or of a replay) is a tool error; a failure whose outcome is unknown fails the task, so recover decides it from C4's records.