Preference Reference

Preference Reference#

Warning

Project Neo is currently in Technical Preview. Features described in this section may change before general availability.

Behavior is shaped through a flat set of settings addressed by dotted paths, grouped into four namespaces: inference, capability, source, and trait. Four layers contribute values on these paths. The project and each agent contribute through prefers, and they also restrict what the layers above them may contribute through constrains and pins. A user and an individual request contribute preferences only. This page is the complete catalog of preference paths, with the type, allowed values, and default for each.

With the one exception noted below, every path listed here is a preference: a value that a user or a request can contribute, subject to the constraints and pins set by the project and the selected agent. Behavior that a request must never influence is policy, not a preference, and lives on a resource body instead. For those fields, such as the project context budget, see the Resource Reference page. For how a value is resolved across the request, user, agent, and project layers, see the Overview page. For the authoring primitives and the decision rule between prefers, constrains, and pins, see the Authoring Agents page.

Tri-State Toggles#

The capability and source switches accept three values rather than a plain boolean:

true

The capability is on. The value is taken as written and never re-derived, so turning on a capability backed by a resource kind you have not published gives the agent the capability with nothing to reach.
false

The capability is off. Again the value is taken as written, so a capability set to false stays off even when the resources that would back it are present. This is the difference from auto: only auto is derived.
auto

The default. The platform derives the effective value. For a switch backed by a resource kind, such as capability.mcp.enabled or source.knowledge_graph.enabled, auto turns the capability on when your configuration publishes at least one resource of that kind, and off otherwise. For a switch that is not backed by a resource, auto turns the capability on.

Inference Settings#

Path	Type	Default	Description
`inference.model`	Model id	none	The model used to generate responses. The value is the `metadata.id` of a `Model` resource. It has no built-in default, so every project must set it in `prefers` or `pins`. A project that does not fails both `genai lint` and the deploy.
`inference.reasoning_effort`	`disabled` / `low` / `medium` / `high`	unset	How much reasoning depth to trade for latency. When unset, the provider default applies. `disabled` suppresses reasoning on a reasoning-capable model, and `low`, `medium`, and `high` step the effort up. Effort applies only to a model whose catalog entry sets `enable_reasoning: true`.

Capability Settings#

Path	Type	Default	Description
`capability.code_interpreter.enabled`	`true` / `false` / `auto`	`auto`	Allow the agent to run code in a sandbox. Backed by the `CodeInterpreter` resource.
`capability.mcp.enabled`	`true` / `false` / `auto`	`auto`	Allow the agent to use tools from external MCP servers. Backed by the `MCP` resource.
`capability.subagents.enabled`	`true` / `false` / `auto`	`auto`	Allow the agent to delegate to other agents.
`capability.filesystem.enabled`	`true` / `false` / `auto`	`auto`	Allow the agent to read from the sandbox filesystem.
`capability.tasks.enabled`	`true` / `false` / `auto`	`true`	Allow the agent to create and manage scheduled tasks on behalf of the user. The platform controls this path, so writing it in `prefers` or `pins` fails the deploy. See the note below.
`capability.studio_mcp.enabled`	`true` / `false` / `auto`	`auto`	Allow the agent to use tools exposed by Squirro Studio plugins. Unlike the MCP switch, this reaches a cluster-internal channel rather than a resource you publish, which is why `auto` is on. Setting `false` withdraws every Studio plugin tool at once, so it is an operator-level switch rather than a way to pick among servers.

Note

capability.tasks.enabled is the one path in this table you cannot author. It is on for every project, and it is listed here because it governs whether users can create scheduled tasks by asking an agent. Every other path in the table is set through prefers, constrains, or pins in the usual way.

Source Settings#

Path	Type	Default	Description
`source.documents.enabled`	`true` / `false` / `auto`	`auto`	Use document retrieval for the conversation.
`source.documents.limit`	integer, 1 to 200	`20`	Maximum number of documents to retrieve.
`source.documents.filter`	string	`""`	Search filter expression applied to retrieval.
`source.knowledge_graph.enabled`	`true` / `false` / `auto`	`auto`	Use knowledge graph retrieval for the conversation. Backed by the `KnowledgeGraph` resource.
`source.knowledge_graph.default`	KnowledgeGraph id	unset	The knowledge graph the agent queries when a call does not name one. When unset, the agent picks a graph explicitly on each call.

Trait Settings#

The trait namespace is parametric: it has one path per TraitDimension your configuration declares.

Path	Type	Default	Description
`trait.<dimension>`	trait id	unset	The active trait for the dimension named `<dimension>`, whose value is the id of one trait within that dimension. When unset, the dimension contributes no prompt fragment.

For example, a TraitDimension with metadata.id: style is addressed as trait.style, and setting trait.style: concise activates the concise trait. For the authoring workflow, see the Authoring Agents page.