AI assistant and automation

The AI assistant helps create Workspaces, install game services, inspect logs, and maintain configuration. Its actions follow the current account, Workspace permissions, conversation mode, and approval policy.

Start a task

1. Open a conversation in the target Workspace. Use a Panel conversation to create Workspaces or arrange work across several of them.
2. Choose an available model and mode: Plan to investigate first, or Agent to perform actions.
3. State the goal and limits, such as “Find out why the server will not start, and ask before editing files.”
4. Read the result and answer questions or approvals. After installation, check Console and try connecting to the game.

If no model is available, ask a super administrator to complete the setup below.

In the conversation's add menu, Upload images adds images to the current message. Upload files saves any file type to Asset library, where you can track and retry uploads. File cards in the current conversation show upload feedback; the files are not automatically sent as message attachments to the AI. Select assets attaches existing images from the library.

Give the assistant a library file's name to read UTF-8 text or view an image, or ask it to import the file into the current Workspace with approval. Tool results and imported files are visible to conversation participants.

Configure available models

A super administrator opens Agent Management → Models and providers:

1. Add a provider and enter its endpoint and credentials.
2. Discover models, check their capabilities, and enable the ones you need. A listed model is not necessarily enabled.
3. Assign conversation, approval, memory, and title models under Model roles.
4. Check task budgets, approvals, and memory under Runtime policy.

Set Language style to describe reply preferences, such as “Give the result first, then the steps.” Changes apply to new tasks and affect tone, wording, and formatting without changing permissions or the interface language.

More model settings

• Plans without presets, including OpenAI compatible and Ollama, require newly discovered models to be enabled manually.
• Provider request timeout defaults to 2 minutes and accepts 1 second to 2 hours. Streaming limits cover waiting for responses and data interruptions; the whole task has a separate runtime limit.
• For input/output budgets, step limits, and daily reasoning quotas by role, 0 means unlimited.

Choose a conversation scope

Conversation	Visibility	Suitable tasks
Panel	Creator only	Select nodes, create Workspaces, and manage Workspace metadata
Workspace	Current members with access	Install and maintain services, inspect logs, edit files, and use Shell and console

Panel conversations cannot directly operate game files, logs, or terminals. For installation, startup, or troubleshooting, the assistant hands a task summary to a conversation in the target Workspace.

One Panel task can handle several Workspaces and hand off each task separately. The Panel task continues after handoff. Open the conversation from its handoff result to check progress: successful dispatch does not confirm that installation finished or that the server is reachable.

Model, reasoning effort, and mode are saved per conversation and locked while a task runs. Reselect the model if it becomes unavailable.

• Plan mode: analyze requirements and prepare steps, with permitted read-only diagnostics in Workspaces.
• Agent mode: call authorized tools to make approved changes and control services.

Approvals and task controls

Approval mode	Behavior
Manual approval	You confirm operations requiring approval
Review for me	An approval model reviews operations within your grant
Full access	Operations within the delegation run automatically, retaining required human approvals

Available modes depend on platform settings and account permissions. Approvals cannot expand access. Assistant writes or deletions of Workspace Secrets always require human confirmation, with identity verification when required.

Review the target, file changes, and effects before confirming. Reopen collapsed questions or approvals from their pending notice; collapsing does not submit an answer. Changed conditions may require fresh confirmation.

Approving an action with Review for me uses one identity verification when required. Selecting Manual approval revokes your interactive grants for that conversation across login sessions.

While a task runs, add requirements or queue independent tasks. Queues pause after failure, interruption, or archiving. Resolve the issue, then select Resume queue in an unarchived conversation. Stopping requests cancellation of active operations; completed changes need separate reversal.

Completed tasks can be forked into a new conversation. Forks copy conversation context; Workspace files remain shared.

Skills and Workspace rules

After installing plugins, type @ in the composer to select a skill. Skills provide instructions without expanding permissions. See Use plugins.

Keep shared startup instructions and directory conventions in /workspace/WORKSPACE.md. The assistant reads it once per new task, so edits apply to the next task. Include only information suitable for Workspace members to share.

Manage memory

Open Memory in a Panel or Workspace conversation to manage the current account's personal memory. On a phone, open it from the conversation's More menu.

Only the owner can directly view and manage personal memory documents. The assistant uses them in Panel and Workspace tasks, so replies may expose their contents to Workspace members. Tasks from external plugin bindings also use the binding owner's memory. Use a dedicated account for bindings and keep its memory suitable for sharing.

Add, edit, merge, download, or export documents. Saved memory survives deletion of its source conversation. Clearing affects only the current account and requires any identity verification configured by the platform.

Super administrators control automatic generation, retention, and per-user quotas. External tool context is excluded from generation by default. Existing documents remain manageable when memory is disabled.

For database migration requirements, see Upgrade and backup.

Create an AI schedule

Create a one-time or cron task in a Workspace and specify its goal and execution policy. Each run creates a separate Workspace conversation using the owner's current permissions and model policy.

• By default, the previous run stops before a new run starts; you can allow it to continue.
• The latest 10 run conversations are retained by default, configurable from 1 to 100. Excess older conversations stop before deletion.
• Unattended questions use their default or recommended option. Requests requiring Secrets or human approval without a safe automatic resolution end with a visible policy error.
• Disabled owners or loss of write access prevent recovery under the previous authorization. Disabling a schedule prevents future runs; cancelling a run does not reverse completed actions.

Automatic approval and full access also follow unattended-execution policy. Check platform policy if a mode is unavailable. API keys cannot create or start AI schedules.

For model, queue, and approval issues, see Troubleshooting.