Connectors › AI steps

AI compose

A model writes the step's declared fields from a template over the record's fields, in batches, with fetched fields fenced as data.

Connector	`ai/compose`
Role	compose
Records	any type
Tier	built in
Credential	none Optional: `ANTHROPIC_API_KEY`, `ANTHROPIC_WORKSPACE_ID`
Cost	unset in the manifest; `gtme plan` prints what the step declares at run time
Source	internal/adapters

Use it

It ships inside the binary. In a pipeline file:

steps:
  - id: compose
    use: ai/compose
    with:
      template: TEMPLATE

Replace the placeholder:

TEMPLATE: the operator's text (ADR-057): a template rendered once per step over config.* — the shared block the records follow (SPEC §10.3). The runner loads {file:} and renders before OPEN; the adapter receives text.

gtme plan pipeline.yaml checks the credential is set and that every field this step needs is provided by a step before it, without calling anyone.

Needs

Whatever fields the step's uses: line names. The runner checks them at plan time.

Provides

first_line, ps_line

Config

Key	Type	What it does
`template` required	string	The operator's text (ADR-057): a template rendered once per step over config.* — the shared block the records follow (SPEC §10.3). The runner loads {file:} and renders before OPEN; the adapter receives text.
`batch_size`	integer
`model`	string
`max_tokens`	integer
`fields`	array of string	Restrict what the model sees to these fields (default: everything known)
`fence`	boolean	ADR-035: wrap externally fetched fields in a delimiter labelled as subject-supplied data (default true); false opts out
`deferred`	boolean	ADR-038: submit the batch to the engine's batch surface (the Message Batches API, half price) and end the run in flight; the next `gtme run` of the pipeline collects. Valid only on the pipeline's last step.

On this page, connector means what gtme calls an adapter. This one is a process adapter compiled into the binary. Two adapter tiers explains the difference.