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

# Commands audit

> Review Tier 3 free-form runInSandbox invocations by risk class and disposition, and find high-frequency, low-risk commands to promote into the catalog.

**Commands Audit** (`/platform/commands/audit`) shows the Tier 3 free-form `runInSandbox` invocations that Trust Fabric captured, and the patterns among them that are candidates for promotion into the typed command catalog. It is the **Audit** tab of the **Commands** sidebar entry, after **Catalog**. Open it to see what agents ran outside the catalog and which of those commands deserve a catalog entry.

![Commands audit (light)](https://mintcdn.com/gloo-b243725a/xyq5perq6SLByPBn/forge/images/console/platform/commands-audit-light.png?fit=max\&auto=format\&n=xyq5perq6SLByPBn\&q=85\&s=7fea81257fd3814319e50d808bb27426)

Commands audit — Forge Console at `/platform/commands/audit`

## What you can do

* Pick a window with the **1h**, **24h**, **7d**, or **30d** chips. **7d** is the default; both sections follow the window.
* In **Recent Tier 3 invocations**, follow a run link to open that run.
* In **Promotion candidates**, **Open issue** opens a pre-filled issue in a new tab carrying the pattern's hash, risk class, count, and sample command, to track promoting it to a Tier 1 catalog command.

## Recent Tier 3 invocations

| Column | Shows |
| - | - |
| **Time** | When the invocation ran |
| **Agent** | The agent that issued it |
| **Run / Step** | The run (linked) and the step |
| **Command (hash)** | The hash of the command string |
| **Risk class** | `read-only`, `mutates-fs`, `network-egress`, `destructive`, or `privileged` |
| **Confidence** | Confidence of the risk-class inference |
| **Disposition** | `pass`, `allow`, `block`, or `held` |

`held` covers both `hold` and `hold-for-approval` dispositions. The table shows up to 200 invocations.

## Promotion candidates

A candidate is a command pattern seen at least five times in the window with a risk class of `read-only` or `mutates-fs`. Each card shows the hash prefix, risk class, the sample command (or "Sample command unavailable" when the execution log row cannot be joined), **Count**, and **Last seen**.

## Who can act

The page requires `audit:read`.

## Key state to watch

* `destructive` or `privileged` risk class in the invocations table: an agent ran a high-risk command outside the catalog. Open the run and check the disposition.
* `block` dispositions: Trust Fabric stopped the command; confirm the agent recovered or the run failed cleanly.
* `held` dispositions: the command was held rather than run; check the approvals queue for the matching run.
* A candidate with a high **Count** and recent **Last seen**: agents keep improvising the same command; promote it with **Open issue**.
* "No Tier 3 invocations in the selected window": widen the window before concluding nothing ran.

## Related surfaces

* [Commands](https://docs.gloo.com/forge/console/platform/commands)
* [Runs](https://docs.gloo.com/forge/console/production/runs.md)
* [Approvals](https://docs.gloo.com/forge/console/production/approvals.md)
* [Decisions](https://docs.gloo.com/forge/console/trust-fabric/decisions.md)


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.