---
title: "The approver's guide"
description: "A guide for Superusers who approve requests in Alpacon. Learn to read a work session or token request, decide well, and follow it through the audit trail."
---

> Full documentation index: https://docs.alpacax.com/llms.txt

# The approver's guide

Most of Alpacon's docs speak to the person asking for access. This guide speaks to you—the reviewer. If you hold the **Superuser** role, sensitive requests from across your workspace land in your queue, and it's your call whether they proceed. Here's how a request moves through your day: what reaches you, how to read it, how to decide, and where to look afterward.

An approval reviewer decides whether a requested action may proceed. The separate [`reviewer` role](/use/audit/governance/periodic-reports/#reviewer-role) records decisions about findings in periodic reports; that role alone does not grant approval authority.

## Your role

When a workspace member needs to do something sensitive, the request routes to **Approvals** for a person to decide. What reaches you:

- **Work sessions** and **work session modifications**: scoped, time-bounded access to specific servers.
- **Session extensions**: requests to push out a running session's end time past what was originally approved.
- **Service tokens** and **token modifications**: issuing new credentials to an application, or changing an existing token's scopes or state.
- **Username** and **group name** changes: identity changes that require approval.

See [Request types](/reference/approvals/) for what you see and can change in each.

Reviewing the queue requires the **Superuser** role, and a few things follow from that:

- **More than one person can hold it.** By default any Superuser can review any pending request, so approvals aren't blocked on one person. An access rule can name other approvers instead.
- **Admins can't approve.** The **Admin** role alone doesn't grant approval. Members who aren't approvers can open their own requests from **My approval requests** on the dashboard, but can't review other people's requests.
- **Notices go to people who can decide.** Managing a target server alone does not make you an approver or a recipient. A person barred from deciding gets no notices or reminders, even if they are a Superuser.
- **You can't approve through the channel that made the request.** A request submitted from the CLI is decided in the **Approvals** queue of the web console (or via [Slack](/integrate/slack/)), never from the CLI. This keeps the person asking and the person deciding on separate sides.
- **Your own session is the exception.** When you request a session for yourself, it's approved the moment it's made—unless the requester is an AI agent (see [below](#when-the-requester-is-an-ai-agent)).

## Reading a request

Open a request from the [queue](/use/approvals/) to see its details: the requester's **justification**, the **requested resources** (or the before/after changes for a modification), a **risk level**, and—for work sessions—an **AI pre-review**. For a work session, that means the requested **duration**, **features**, **target servers**, and why they're asking.

### The AI pre-review

Every work session request carries an **AI pre-review** card with a recommended **risk level** (Low, Medium, High, or Critical risk), the **reasoning** behind it, and **recommended features** for the session. The recommendations are highlighted in the feature editor, but nothing is applied automatically—treat the pre-review as a second opinion, then decide. See [Reviewing requests](/use/approvals/reviewing/#ai-pre-review).

### The risk rating

Each request shows a **risk level** so you can gauge sensitivity at a glance:

- **Service tokens**: risk is computed from the requested scopes, with a per-scope breakdown and reason.
- **Work sessions**: the risk level comes from the AI pre-review.

Risk is color-coded, so a High or Critical request stands out.

### What you can adjust

Before approving a work session in the console, you can:

- **Servers**: remove requested servers.
- **Connect as**: narrow the run-as account when the request names one and the session is not assigned to a token.
- **Access**: remove requested features.
- **Session length** (or **Scheduled** for a scheduled session): shorten the duration or bring the end time forward.
- **Recommendations**: add recommendations for the assignee.

You cannot add servers or features, widen the account, or lengthen the window.

Approving sends your adjustments along with the decision. See [Request types](/reference/approvals/#work-session).

### Watch the clock

The queue shows a time-left meter to the decision deadline for every pending request, including immediate work sessions. The row turns to a warning when 30 minutes or less remain.

In the decision panel, notices appear when a scheduled start is within 60 minutes, when that start has passed, or when the request will expire within 30 minutes if nobody decides. All three notices use the same warning tone.

A work session whose requested end has passed cannot be approved (`work_session_expired`). An automatic cleanup cancels these pending sessions; a refused approval does not itself cancel them.

See [urgency warnings](/use/approvals/reviewing/#urgency-warnings).

## Deciding well

A pending request shows **Approve** and **Reject**. Either way, the requester is notified.

**Trim before you reject.** Because you can remove requested features and servers before granting a work session, an over-broad request often doesn't need to go back. Narrow it to what the work actually needs and approve the smaller scope. Reject requests that should not proceed. A request whose end time has passed cannot be approved; ask for a new request.

**Give sudo the most scrutiny.** Privilege elevation (sudo) is the heaviest scope: for the session window it grants elevated, sudo-level access across the session's other features. Including it in a session is only the first gate—each time the operator actually uses sudo, the action is checked again at that moment, may be held for a person to approve, and may prompt for MFA, with every sudo action recorded. Before you grant sudo, make sure the purpose genuinely calls for it. See [privilege elevation (sudo)](/reference/work-session-features/#privilege-elevation-sudo).

**Acting in bulk.** You can select multiple pending requests and approve or reject them in bulk. Bulk actions skip individual review and cannot be undone, so verify your selection first. Critical, blocked, or script requests stay pending for individual review.

## Cutting repeat work with approval policies

Reviewing every request by hand doesn't scale. Some requests are meant to clear without waiting for you:

- A Superuser's own session request is approved the moment it's made.
- A workspace can set [access rules](/use/approvals/#access-rules) that decide, based on a request's risk and who's asking, which actions are allowed and which wait for a person. Low-risk sudo actions, for instance, can be cleared automatically while riskier ones are held for approval.

The idea is to **govern by policy and review the exceptions**: let routine, low-risk requests flow through a defined lane, and spend your attention on the ones the policy holds back. One default stands out—sessions requested by an AI agent require a human by default (see [below](#when-the-requester-is-an-ai-agent)).

See [automatic approval](/use/approvals/#automatic-approval).

## While a session runs

An approval isn't the end of your visibility. **Live activity** (under **Execution**) shows the sessions running right now so you can watch and step in.

Active sessions appear as a table, one row each, carrying the title and purpose, the risk so far, the last activity, the requester, and the target servers and allowed features. Sessions that had a command blocked by policy, or whose risk reached High or Critical, sort to the top, so you do not have to read down the list to find the ones that need attention. Click a row for the live detail view, which streams:

- **Connection status**: Live, Reconnecting, Closed, or Connection error, with the latest snapshot time.
- **Status history**: the lifecycle from Requested through Approved and Active, on to Completed, Expired, or Revoked.
- **Activity timeline**: a real-time feed of commands, terminal opens and closes, sudo requests and grants, and file transfers, filterable by category.

When you need to intervene, two actions are your kill switch:

- **Complete session**: end a running session immediately.
- **Revoke session**: end an approved session before it becomes active.

See [Live activity](/use/live-activity/).

## After the session

Once a session ends, the record is where accountability lives.

**Session history.** **Audit** → **Session history** lists the sessions that have ended, with who ran each one, the servers in play, how it ended, and a **risk** level from the AI analysis. Open a row for the full timeline, the approval record, and the analysis. See [Session history](/use/audit/sessions/).

**Recordings and AI risk analysis.** A finished session's detail page replays its activity on the **History** tab (Websh terminal recordings are masked and can be played back), and the **Analysis** tab adds an AI review once the session ends: a **risk level**, a plain-language **summary**, key metrics, a **threat analysis** (including a **MITRE ATT&CK** technique list when anything is found), and a **response guide**. See [After a session](/use/audit/sessions/review/).

**Approval history.** Decisions are recorded. Superusers on paid plans can review closed requests in **Approval history** under **Audit > Governance**. It shows what was asked, who asked and who decided, the outcome, and the requester's reason—useful for questions like "why was this approved?" and "what's our approval-vs-rejection rate for sudo?" See [Approval history](/use/audit/governance/approvals/).

## When the requester is an AI agent

Alpacon lets AI agents request sessions from the CLI, but an agent's request lands on a human by default—you. By default, a session requested by an AI agent waits for a person even when a Superuser makes the request; the automatic approval of a Superuser's own session doesn't apply to it.

So an agent request gets the same care as any other, with a few things worth extra attention:

- **The agent's name is not verified.** The agent name on a request (for example `claude-code`) is what its client reported; the check on the icon covers only the fact that an agent sent it. Judge the request by its purpose and scopes, not by the name. See [Live activity](/use/live-activity/) for how it is shown.
- **No automatic approval applies by default**, so read the purpose and requested scopes yourself rather than assuming a policy vetted them.
- **Lean on the AI pre-review's** risk level and reasoning to spot a request that asks for more than the stated task needs.
- **Trim scope before granting**, especially sudo, so the agent gets exactly the features and servers the work requires and nothing more.

See [Request a session](/use/work-sessions/create/) for how agent requests are made.

## Related docs

- [Approvals overview](/use/approvals/)
- [Request types](/reference/approvals/)
- [Reviewing requests](/use/approvals/reviewing/)
- [Live activity](/use/live-activity/)
- [After a session](/use/audit/sessions/review/)
- [Session history](/use/audit/sessions/)
- [Approval history](/use/audit/governance/approvals/)

Source: https://docs.alpacax.com/learn/approver-guide/
