Orchestrator

orchestrator is the required manifest field that configures the LLM driving a Quick App: the model, its system prompt, and how it handles attachments and many tools. This reference lists every field. It is for app developers who write Quick App manifests. It assumes you know orchestration.

{
  "orchestrator": {
    "deployment": {
      "deployment_id": "<MODEL_ID>",
      "parameters": { "temperature": 0.5, "seed": 820288 }
    },
    "system_prompt": { "type": "custom", "variables": { "company": "ACME" }, "content": "You are an assistant for {company}." },
    "max_iterations": 10,
    "attachment_strategy": { "type": "lazy_on_demand" },
    "tool_discovery": { "enabled": true }
  }
}

Fields

Field	Required	Default	Description
`deployment`	Yes	—	The orchestrator model. See Deployment. When the backend's `DEFAULT_ORCHESTRATOR_DEPLOYMENT_ID` is set, a manifest can leave it out.
`system_prompt`	No	An empty `custom` prompt	See System prompt
`max_iterations`	No	`15`	The most agent loop iterations in one turn. There is no unlimited setting; `0` or less fails the request before the first LLM call.
`propagate_stages`	No	`true`	Shows the model's own stages, such as reasoning steps, in the answer
`attachment_strategy`	No	—	How the model receives attachments. See Attachment strategy.
`tool_discovery`	No	—	Finds deferred tools on demand. See Tool discovery.

Deployment

Field	Required	Description
`deployment_id`	Yes	The DIAL model's deployment ID
`parameters`	No	Parameters sent to the model, listed below
`name`	—	Deprecated alias for `deployment_id`

Parameter	Description
`temperature`	Sampling temperature
`top_p`	Nucleus sampling
`seed`	A seed for more reproducible answers
`stop`	Up to 4 stop sequences
`n`	The number of choices to generate
`max_tokens`	The most tokens to generate
`presence_penalty`, `frequency_penalty`	Penalties between `-2.0` and `2.0`
`reasoning_effort`	`none`, `minimal`, `low`, `medium`, or `high`. Dropped, with a warning, when the model doesn't advertise `features.reasoningEfforts`.
`custom_fields`	`{ "configuration": {...} }`, the model's own configuration, such as Anthropic `betas`

{
  "deployment": {
    "deployment_id": "<CLAUDE_MODEL_ID>",
    "parameters": {
      "temperature": 1.0,
      "custom_fields": { "configuration": { "betas": ["token-efficient-tools-2025-02-19"] } }
    }
  }
}

System prompt

system_prompt takes one of three types.

`custom`

Field	Required	Description
`type`	—	`custom`
`content`	Yes	The prompt text. `{name}` placeholders are filled from `variables`.
`variables`	Yes	Values for the placeholders; `{}` when there are none

`dial`

Field	Required	Description
`type`	Yes	`dial`
`path`	Yes	The prompt's relative path in DIAL, bucket included
`variables`	No	Values for the prompt's placeholders

{ "system_prompt": { "type": "dial", "path": "<PROMPT_PATH>", "variables": { "language": "English" } } }

`predefined`

Field	Required	Description
`type`	Yes	`predefined`
`template`	Yes	The template name, such as `gpt_prompt`, `anthropic_prompt`, or `gemini_prompt` for ChatHub

A predefined prompt has no override. To change it, copy its text into a custom prompt.

Attachment strategy

Field	Required	Description
`type`	Yes	`lazy_on_demand`, the only strategy: the model loads attachments with `internal_attachments_get_content`
`accepted_types`	No	MIME types, such as `image/*`, that narrow what the model's deployment accepts. An empty list is rejected; leave the field out to accept everything the deployment does.

See Attachments for the orchestrator.

Tool discovery

Field	Default	Description
`enabled`	`false`	Leaves `deferred` Toolsets out of the first LLM request, for `internal_tool_search` to find
`service_model`	The orchestrator's model	The model for the search call
`min_tools_for_deferral`	`MIN_TOOLS_FOR_DEFERRAL`, `10`	Toolsets with fewer tools are always sent

See Dynamic tool discovery.

Fields​

Deployment​

System prompt​

custom​

dial​

predefined​

Attachment strategy​

Tool discovery​

Further reading​