Graph Document Format

Subfork examples use the portable subfork.graph/1 JSON envelope. An agent can author or adapt an example without access to the Subfork source tree, but it must consult the production public node catalog for node-specific ports, parameters, versions, and worker requirements. A local development catalog is used only while testing unreleased changes.

Minimal Shape

{
  "format": "subfork.graph/1",
  "tags": ["example", "text"],
  "definition": {
    "version": "0.1.0",
    "name": "Example name",
    "description": "What the graph does and its important limits.",
    "nodes": [
      {
        "node_instance_id": "message",
        "node_id": "n_text_value",
        "node_version": "1.0.0",
        "title": "Message",
        "params": {
          "text": "Hello"
        }
      }
    ],
    "edges": [],
    "graph_outputs": {
      "text": {
        "node_instance_id": "message",
        "output_name": "text"
      }
    }
  }
}

Keep the envelope and exported field names exactly as shown. Start from the closest existing example or a fresh Subfork export when possible. Do not invent node IDs, parameter names, ports, or versions from memory.

tags is envelope metadata carried by export and import. Use at most 12 unique, lowercase slugs of 1–32 letters, numbers, hyphens, or underscores. Tags describe the example’s durable category and capabilities; imports apply them when creating a graph, while merge and replace preserve the destination graph’s existing tags.

Definition Fields

• version, name, and description identify the graph document. Give the graph a concise name and describe network access, side effects, or important limits.
• nodes is a nonempty array of node instances. Every node_instance_id must be unique within the graph. node_id selects a node type from the target catalog, and node_version records the version used by the export.
• params contains node-specific configuration. Use only fields and values supported by the node manifest.
• edges connect an exact source_output to an exact target_input. The source and target IDs are node instance IDs, not node type IDs.
• graph_outputs gives stable names to selected node outputs. Each binding names an existing node instance and one of its outputs.
• metadata carries authoring and presentation hints. Existing examples use metadata.template for category, requirements, and a composite interface, and metadata.canvas for positions and recommended panels. Copy these shapes from a nearby example and omit optional hints that do not apply.

Some exported nodes include custom_inputs, exposed_inputs, exposed_outputs, and matching visibility flags. Preserve these fields when adapting an export. Add dynamic inputs only when the node manifest and a working example show that the node supports them.

Node Contracts

Fetch the production catalog before choosing or changing a node:

curl --fail --silent --show-error --max-time 15 --max-filesize 1048576 \
  https://subfork.com/api/v1/nodes/n_text_value

The single-node endpoint returns the matching manifest directly. Use the full /api/v1/nodes collection when discovering available node IDs.

The manifest is authoritative for:

• node_id and current version;
• input and output port names and types;
• parameters, defaults, required values, and enumerated options;
• execution-profile schema and required worker capabilities;
• allowed placement.

The catalog at subfork.com is authoritative for released examples. A local development instance may temporarily expose unreleased nodes during testing. The examples repository does not vendor a catalog snapshot because it would become stale and could incorrectly claim compatibility with production.

Variables And Secrets

Configuration that differs by graph or execution belongs in graph-scoped execution profiles. Examples use template references such as:

Use these references only in parameters that support template expansion. List every required variable and secret in the companion notes and in metadata.template.requires when that metadata is present. Never put a secret value, private URL, account identifier, cookie, or local path in graph JSON.

Credential-bearing HTTP requests should keep the provider origin fixed. Vary bounded resource identifiers through variables instead of accepting an arbitrary URL that could receive the credential.

Agent Authoring Workflow

1. Read the root and category-specific AGENTS.md files.
2. Choose the closest existing graph and read both its JSON and companion notes.
3. Fetch every proposed node manifest from subfork.com, or from local development when testing an unreleased node.
4. Build nodes with unique instance IDs and exact manifest parameters.
5. Connect exact output and input port names, checking their types.
6. Expose only useful graph outputs and add presentation metadata when it helps.
7. Add companion notes covering setup, network calls, bounds, costs, side effects, outputs, limitations, and test status.

8. Run offline validation and target-catalog validation:

   python3 scripts/validate_examples.py
   python3 scripts/validate_examples.py \
     --node-catalog-dir ../subfork-new/nodes
   python3 scripts/validate_examples.py \
     --node-catalog-url https://subfork.com/api/v1/nodes
   

9. Import the result into a fresh graph in local development, then inspect it and run the documented happy path before describing it as tested.

Repository validation checks the envelope, references, and common credential patterns. The local catalog option also checks versions, parameters, exposed and custom ports, edge types, literal collisions, and graph outputs. The production catalog option checks released node-type availability. Provider behavior, costs, and side effects still require a fresh import and an appropriate manual run.