Code and shell execution
SandboxPolicy
Section titled “SandboxPolicy”Limits on the processes that the Python and shell tools start, on Linux and macOS.
Pass one as CodeExecutionConfig(sandbox_policy=...) or ShellConfig(sandbox_policy=...). Without
one, those processes run with no sandbox. The sandbox controls:
- which files they can read and write
- which network they can reach
- which environment variables they see
- how much memory, CPU time, and how many processes they can use
macOS applies the file and network rules but not the resource limits. The sandbox reference explains how each rule is enforced.
The defaults match the CLI’s restricted profile.
SandboxPolicy.__init__
Section titled “SandboxPolicy.__init__”__init__( max_memory_mb: int = 2048, max_cpu_secs: int = 600, max_procs: int = 64, max_open_fds: int = 1024, max_file_sz_mb: int = 256, network: NetworkMode = NetworkMode.Loopback, extra_fs_read: list[str] = [], extra_fs_write: list[str] = [], extra_env: list[str] = [], strict: bool = False,) -> NoneParameters
| Name | Type | Default | Description |
|---|---|---|---|
max_memory_mb |
int |
2048 |
Memory (address space) limit, in MB. |
max_cpu_secs |
int |
600 |
CPU time limit, in seconds. On Linux, a limit shorter than the tool’s timeout is raised to match it. |
max_procs |
int |
64 |
Limit on processes and threads. |
max_open_fds |
int |
1024 |
Limit on open file descriptors. |
max_file_sz_mb |
int |
256 |
Largest file a process can write, in MB. |
network |
NetworkMode |
NetworkMode.Loopback |
Network access. See NetworkMode. |
extra_fs_read |
list[str] |
[] |
More paths the process can read. |
extra_fs_write |
list[str] |
[] |
More paths the process can read and write. |
extra_env |
list[str] |
[] |
More environment variable names to pass through. Their values come from the current environment. |
strict |
bool |
False |
Fail at startup if a requested file or network restriction isn’t available, instead of skipping it. |
CodeExecutionConfig
Section titled “CodeExecutionConfig”Turns on the built-in Python tool, execute_python.
Pass it as Runner(code_execution_config=...), then set
ChatCompletionRequest.enable_code_execution=True on each request that may run code.
CodeExecutionConfig.__init__
Section titled “CodeExecutionConfig.__init__”__init__( python_path: str | None = None, timeout_secs: int | None = None, working_directory: str | None = None, sandbox_policy: SandboxPolicy | None = None, permission: CodeExecutionPermission | None = None, approval_callback: Callable[[dict[str, object]], bool] | None = None,) -> NoneParameters
| Name | Type | Default | Description |
|---|---|---|---|
python_path |
str | None |
None |
Interpreter to run. Defaults to python on Windows and python3 elsewhere. |
timeout_secs |
int | None |
None |
Time limit per call, in seconds. Defaults to 60. |
working_directory |
str | None |
None |
One directory that every session runs in. Defaults to a new temporary directory per session. |
sandbox_policy |
SandboxPolicy | None |
None |
Sandbox for the interpreter. None runs it without one. |
permission |
CodeExecutionPermission | None |
None |
Older form of ChatCompletionRequest.agent_permission, which you should prefer. Defaults to Auto. |
approval_callback |
Callable[[dict[str, object]], bool] | None |
None |
Older, code-only approval callback. Prefer ChatCompletionRequest.agent_approval_callback. It receives a dict with approval_id, session_id, code, outputs, and working_directory, and returns True to run the code. An exception counts as a denial. |
ShellConfig
Section titled “ShellConfig”Turns on the built-in shell tool.
Pass it as Runner(shell_config=...), then set ChatCompletionRequest.enable_shell=True or pass
shell_skills on each request that may run commands.
ShellConfig.__init__
Section titled “ShellConfig.__init__”__init__( shell_path: str | None = None, timeout_secs: int | None = None, working_directory: str | None = None, sandbox_policy: SandboxPolicy | None = None, permission: AgentPermission | None = None,) -> NoneParameters
| Name | Type | Default | Description |
|---|---|---|---|
shell_path |
str | None |
None |
Shell to run. Defaults to cmd on Windows and /bin/sh elsewhere. |
timeout_secs |
int | None |
None |
Time limit per call, in seconds. Defaults to 600. |
working_directory |
str | None |
None |
Parent directory for sessions. Each session runs in its own <working_directory>/<session_id>. Defaults to a new temporary directory per session. |
sandbox_policy |
SandboxPolicy | None |
None |
Sandbox for the shell. None runs it without one. |
permission |
AgentPermission | None |
None |
Default permission for shell commands. Defaults to Auto. For per-request control, use ChatCompletionRequest.agent_permission. |
ShellSkillMount
Section titled “ShellSkillMount”A local skill directory to mount for the shell tool, in the OpenAI Skill layout (a folder with a SKILL.md).
Pass instances in ChatCompletionRequest.shell_skills. With the HTTP server, upload skills through
/v1/skills instead. Each skill appears at skills/<name> in the shell’s working directory.
ShellSkillMount.__init__
Section titled “ShellSkillMount.__init__”__init__(name: str, description: str, source_path: str) -> NoneParameters
| Name | Type | Default | Description |
|---|---|---|---|
name |
str |
required | Skill name, shown to the model. |
description |
str |
required | What the skill does, shown to the model. |
source_path |
str |
required | Local directory that holds the skill. |
Generated from mistralrs-pyo3/mistralrs.pyi.