Refactor pipelines

Refactor is Zingle AI's flagship engine for cleaning up an existing dbt project at scale. Instead of a single chat, a refactor is a pipeline - a set of tasks organized into phases and executed in dependency order, with a human reviewing and approving each step.

The pipeline model

Configuring a refactor

Before running, you configure the refactor per repository:

• Phases - which classes of transformation to include.
• Business domains - logical groupings that guide how models are organized.
• Locked models - models the pipeline must not touch.

Task types

A pipeline is built from task types, each handled by a specialized agent:

Task type	What it does
merge_models	Combine models that should be one.
dedup_models	Remove duplicate models.
divergent_models	Reconcile models that diverged from a common ancestor.
rename_models	Rename models to a consistent convention.
apply_layering	Enforce staging / intermediate / mart layering.
improve_staging_layer	Clean up the staging layer specifically.
discard_models	Remove models that are no longer needed.
build_semantic_models	Generate semantic-layer definitions.
organize_macros	Tidy and organize macros.
remove_nested_folders	Flatten unnecessary folder nesting.
clean_project_yml	Clean up dbt_project.yml.
correct_seed_structure	Fix seed file structure.
rebuild_configs	Rebuild folder-level configs.
rewrite_formatted_yml	Reformat YAML consistently.

Working through tasks

Each task moves through a lifecycle you control:

1. Plan - the pipeline proposes what the task will do; you can review and edit the plan.
2. Approve & start - kick off execution.
3. Execute - the agent performs the change; you can watch logs.
4. Impact - review what changed and its downstream impact.
5. Confirm / skip / retry - accept the result, skip the task, or retry it. You can also fast-forward through tasks.

Pipelines can be resumed or cancelled, and you can always see the current task and the full task tree.

Pull-request strategies

How changes land in GitHub is configurable:

Strategy	Behavior
single_pr	One PR for the whole pipeline.
batched_pr	Group related tasks into batched PRs.
per_subtask_pr	A PR per subtask for the finest-grained review.

When to use Refactor vs Optimize

• Use Refactor for structural change - reorganizing, renaming, merging, re-layering, and building the semantic layer.
• Use Optimize to scan for discrete issues (performance, materialization, DRY, tests) and apply targeted fixes.