Skip to main content
This page covers the settings that change how an expert does the work for you. They live server-side (one config row per user) and apply across sessions and across devices.
The tab labelled “Personalization” inside the app holds your memory list (About You), not the settings on this page. For memories, see Personal memory.

​
Tool run mode

Decides whether the expert asks before running tools (command execution, MCP calls, file writes, and so on). Settings → Tool Mode, three choices: With “Use Allowlist” you maintain a tool list on the same page, autocompleted from the tools currently available. When a confirmation is needed, the tool card offers:
  • Accept — approve this one call
  • Edit parameters — adjust and then run (a malformed edit returns “Invalid format, please try again”)
  • Cancel — reject; you can attach a reason, which the expert reads and adapts to
There is no auto-approval. A pending tool call waits indefinitely for you — it will not time out into running, nor into rejection. So clear pending confirmations before stepping away, or you will come back to a task frozen exactly where you left it.

​
Model preferences

Model choice is stored per modality, five of them, independently: Beyond the model itself, image and video also remember per-model parameter preferences (resolution and similar), reused on the next generation.
Model selection lives on the server, not in your browser. Switching devices or clearing your cache will not lose it. Conversely, changing the model on one device changes it on the next device you open.
The available model list is operational data (it moves with upstream providers), so the docs do not hardcode specific model names — doing so would create a list guaranteed to go stale. The model selector is the authoritative list.One rule is stable: model visibility is tiered by plan, ordered free(0) < pro(1) < max(2) < flagship(3) < team(4) < enterprise(5). A model requiring max is available to that tier and above. Upgrading unlocks more models; downgrading hides some.
When the platform changes its defaults it carries a version number, and a newer version overrides your row. In practice: modalities you never explicitly chose follow the platform default; ones you did choose keep your choice.

​
Expert team and multi-task

Both live under Settings → Expert Team, but they solve different problems.

​
Expert team (cross-expert delegation)

You can pick up to 10 experts as your team. Team members appear automatically in the delegatable roster of every conversation, and the lead expert can hand subtasks to them when it judges that useful. Delegations run in the background with a fixed ceiling of 900 seconds (15 minutes) each; a timeout produces an explicit error rather than a silent empty result.
@-mentions outrank the setting. Mentioning an expert with @ in a message adds it to this conversation’s roster temporarily — whether or not it is on your team. That is deliberate: an explicit mention is a stronger statement of intent than a settings toggle.When you mention two or more experts, the expert is required to delegate to each of them rather than picking one.
An empty team auto-injects nobody, but @-mentions still work.
Roster construction failures never interrupt the conversation — delegation is an enhancement, so if the roster cannot be built the run continues single-expert. Which means “my team member wasn’t consulted” can mean the roster failed to build, not that the expert decided against it.

​
Multi-task (parallel subtasks)

Controls whether an expert may split complex work into parallel subtasks. Off by default; you have to opt in. Once on, you can set a “default sub-task model”:
  • Left empty = follow the main conversation model
  • Set to a model = subtasks use it
If the model you picked has been retired, the system does not silently substitute the platform default — it falls back to the main conversation model. Substituting “something similar” is deliberately avoided: the sub-task model is an explicit choice you made for cost or capability reasons, and replacing it would quietly overturn that decision.
Subtasks cannot nest indefinitely: subagents never receive the task or delegate tools, and delegated specialists have subtasks disabled outright. Recursion is structurally impossible.

​
Knowledge connections

Connecting an external knowledge platform (Feishu / DingTalk / Yuque / IMA) lets experts with the knowledge plugin enabled search and cite your knowledge base during a conversation. Credentials are stored encrypted and used only in your own conversations. See Knowledge connectors.

​
The plugin panel is not a persistent setting

The plugin panel above the chat input (Sites, Movie, Canvas, Browser, and the other user-selectable plugins) is a per-session temporary choice, not an account setting:
  • Selections apply to the current session only
  • They are lost on page refresh — nothing is written to local storage or synced to the server
  • A new session starts from scratch
This is categorically different from everything else on this page. Tick the plugin you need in the current session, and do not expect it to be remembered.
There is a second layer: creator-side plugin disabling. An expert’s creator can turn off certain built-in plugins in the studio, and that layer is neither visible nor changeable from your plugin panel. What you can tick is whatever the creator has not disabled.

​
Failures and fixes

​
Scope of each setting

​
Keep reading

Personal memory

The memory list behind the “Personalization” tab

Multi-expert delegation

How delegation is decided and executed

Knowledge connectors

Connecting the four external knowledge platforms

Plugin catalog

Which plugins need to be ticked manually