Pipeline Engine

Copy Markdown View Source

Public API

The primary entry points are:

run/3 executes a pipeline from its start node. resume/3 restores execution from a checkpoint map, struct, or checkpoint.json path.

DOT Model

The parser converts DOT into these runtime structs:

This model is intentionally smaller than full Graphviz DOT. It prioritizes the subset needed to execute Attractor-style workflows reliably.

Handler Mapping

Node execution is determined by either an explicit type attribute or the node shape:

ShapeDefault handler type
Mdiamondstart
Msquareexit
diamondconditional
componentparallel
tripleoctagonparallel.fan_in
hexagonwait.human
parallelogramtool
housestack.manager_loop
boxcodergen

You can override this with type="...", and you can register new handler modules through AttractorEx.HandlerRegistry.

Routing Rules

The engine selects the next edge in this order:

  1. Matching condition=...
  2. Matching status=...
  3. Matching the handler outcome's preferred_label
  4. Matching the handler outcome's suggested_next_ids
  5. Unconditional edge by weight
  6. Lexical fallback

This is implemented in AttractorEx.Engine and matches the spec-driven behavior documented in the local compliance notes.

Retries And Goal Gates

AttractorEx.Engine supports:

  • Node retry policies with exponential backoff.
  • Graph-level and node-level retry targets.
  • Goal-gate nodes that must end in success or partial success.
  • Loop restart semantics via edge metadata.

Outcomes are represented as AttractorEx.Outcome values and serialized through AttractorEx.StatusContract.

Styling And Transforms

Before validation and execution, the engine applies:

  1. Model stylesheet resolution through AttractorEx.ModelStylesheet
  2. Variable expansion through AttractorEx.Transforms.VariableExpansion
  3. Any custom graph transforms passed in opts

At runtime, codergen stages may also receive synthesized context carryover text based on fidelity settings.

For the precise implementation details: