> ## Documentation Index
> Fetch the complete documentation index at: https://docs.mandala.md/llms.txt
> Use this file to discover all available pages before exploring further.

# Sandbox

> Test agents against real workspace data without committing business side effects.

Sandbox runs the real agent, compiler, policies, connectors, and review flow against workspace data while keeping business side effects temporary or simulated.

It is a safety mode, not a separate toy agent and not a synonym for fixture data.

## What Sandbox preserves

A Sandbox run uses the same important inputs as a normal run:

* the selected workspace and acting user;
* the installed agent version and compiled capability bindings;
* real cataloged connector data when using the real-data path;
* the normal deterministic rules and evidence requirements;
* the review, approval, and audit model; and
* Context Engine retrieval when it is enabled and eligible.

This makes Sandbox useful for proving that mappings, freshness checks, duplicate detection, and evidence coverage work before activation.

## What Sandbox blocks

Sandbox prevents real business writes from escaping the safety boundary. The included state-changing executor produces mock outcomes only; it does not send a purchase order to a vendor or ERP.

Temporary session state is separated from durable business state, and every action still has to pass capability, policy, role, warning, and approval checks.

<Warning>
  Sandbox does not make stale or incomplete data safe. Freshness, source coverage, schema compatibility, and required-field failures still block the run.
</Warning>

## Open a real-data Sandbox session

In the interactive terminal, use:

```text theme={null}
/sandbox
```

The scripted equivalent is:

```bash theme={null}
mandala sandbox open --limit 25
```

This opens a temporary session over a bounded set of cataloged workspace data.

## Test a skill file

```bash theme={null}
mandala sandbox run \
  --skill skills/procurement-reorder/SKILL.md \
  --confirm-mappings
```

`--confirm-mappings` is required because Mandala creates declarative mappings between the skill's capability contract and the workspace data catalog. Review those mappings before accepting them.

Installed agents can also be tested from `/agents` with the `Test in Sandbox` action.

## Fixtures versus real data

Mandala supports both paths and labels them separately:

| Path              | Use                                                                                                           |
| ----------------- | ------------------------------------------------------------------------------------------------------------- |
| Synthetic fixture | Repeatable development and regression scenarios using Mandala Bean Co. data                                   |
| Real-data Sandbox | Read current workspace records through the real catalog and connector bindings while blocking external writes |

List and run synthetic scenarios with:

```text theme={null}
/fixtures
/run-fixture <row-or-id>
```

Scripted equivalents are `mandala workflow fixture list` and `mandala workflow fixture run <scenario-id>`.

## Workspace setting

Check the current setting with:

```bash theme={null}
mandala sandbox status
```

Owners and admins can change it with a reason and configuration version. Turning Sandbox off requires explicit confirmation because it weakens the workspace safety posture:

```bash theme={null}
mandala sandbox set off \
  --expected-version <n> \
  --reason "Approved production transition" \
  --confirm
```