Transforms¶

Transforms are in-memory document mutations that run before rendering or, in graph mode, act as edges between formats.

Built-in document transforms¶

Transform	What it does	Source
`EmojiTransform`	Replaces emoji with `[emoji]` for non-HTML targets	`src/transforms/emoji.rs`
`VariableSubstitutionTransform`	Replaces `{{key}}` placeholders from `variables`	`src/transforms/variable.rs`
`SyntaxHighlightTransform`	Normalizes fenced-code language tags to lowercase	`src/transforms/syntax_highlight.rs`

Order of execution¶

The standard registry always registers built-ins in this order:

emoji
variables
syntax-highlight normalization

That order is defined in register_transforms.

Command transforms¶

YAML-defined command transforms are loaded from transforms: files and can run either:

stdin/stdout mode when no {input} / {output} placeholders exist
file mode when placeholders are present

Example:

transforms:
  - name: md-to-html
    program: pandoc
    args: ["{input}", "-o", "{output}"]
    from: markdown
    to: html
    cost: 0.5
    quality: 0.95

AI transforms¶

AI transforms add fields such as ai, model, prompt, endpoint, api_key_env, cache_path, and prompt_version.

transforms:
  - name: summarize
    ai: ollama
    model: mistral
    prompt: "Summarize this document:

{input}"
    from: markdown
    to: markdown
    cost: 2.0
    quality: 0.8

Aggregation transforms¶

Setting input_kind: collection marks a transform as an aggregation edge. Those transforms combine multiple inputs into one output and are registered in an AggregationRegistry.

Failure modes¶

TransformRegistry supports:

FailFast — stop on the first error
ContinueOnError — log the failure and pass the unmodified input to the next transform

Warning

A transform YAML entry must supply exactly one of program, plugin, or ai from a schema perspective. In practice, graph execution currently materializes command and AI transforms directly; plugin-backed graph edges are not yet wired by the graph builder.