exec_core_cmd is an explicit exception to the V4/V5 rule that shell execution normally belongs to an Endpoint Provider.
exec_core_cmdruns only on the Core Service host throughcore.local.- Its working directory is fixed to the Core process
Path.cwd(). - It does not add arbitrary Core file access, Workspace I/O, or local MCP lifecycle.
exec_sys_cmdkeeps its existing Desktop/Endpoint meaning and continues to require Endpoint shell capability routing.
Core command execution is controlled by:
core_shell_exec_enabled, defaulttrue.core_cmd_policy_path, defaultuser/core_cmd_policy.json.core_command_timeout_seconds, default120.core_command_output_max_chars, default20000.
Environment overrides:
MEETYOU_CORE_SHELL_EXEC_ENABLEDMEETYOU_CORE_CMD_POLICY_PATHMEETYOU_CORE_COMMAND_TIMEOUT_SECONDSMEETYOU_CORE_COMMAND_OUTPUT_MAX_CHARS
If the Core policy file is missing or invalid, Core uses the built-in whitelist. It must not fall back to mode:none.
The built-in policy is whitelist-first. It allows low-risk host inspection commands such as dir, echo, hostname, whoami, date, time, ver, where, version probes, and read-only Git inspection (git status, git log, git diff, git branch).
Network read helpers are allowed only as stdout reads:
curl/curl.exewget -qO-/wget.exe -qO-
The shared hard guard always runs before whitelist matching. It rejects shell control and redirection operators, encoded PowerShell, destructive delete/shutdown/registry/permission/firewall patterns, destructive Git and container cleanup commands, and curl/wget options that write files, upload data, read local files, or switch to non-HTTP local/file protocols.
exec_core_cmdis registered as a Core-owned tool oncore.local.- It is marked
destructive, not parallel-safe, and order-required. - It is a default/basic assistant tool, so it is visible across assistant modes and endpoint basic bundles when enabled.
- It is hidden when
core_shell_exec_enabled=false; direct calls are rejected by policy as well.