Proposal: Introduce a compatible Graph Authoring API

[laipz8200]

Summary

Introduce a typed Graph Authoring API for building Graphon workflow definitions in Python while continuing to emit the existing runtime graph_config format.

Goals

• Provide a clearer, user-facing API for common graph construction flows.
• Preserve compatibility with the current runtime graph schema and execution behavior.
• Keep legacy compatibility code isolated so it can be removed after downstream migration.
• Avoid exposing internal/runtime-only nodes as first-class authoring concepts.

Non-goals

• This proposal does not change the runtime graph format.
• This proposal does not cover the LLM protocol rename or standard Slim LLM runtime work, which is being handled separately.
• This proposal does not attempt to expose every legacy node in the first stage.

Example

A basic three-node chat workflow should be expressible with straightforward Python authoring code:

graph = Graph()
start = graph.start("start", inputs={"query": "paragraph"})
reply = graph.llm(
    "reply",
    model=OpenAIChat("gpt-4.1-mini"),
    system="You are a helpful assistant.",
    user=Prompt("{{ query }}", query=start.output("query")),
    after=start,
)
graph.answer("answer", text=reply.output("text"), after=reply)

config = graph.to_config()

Rollout

The plan is to land the API incrementally with focused runtime validation tests and concise documentation. Compatibility aliases and migration-only code should be clearly marked and kept in dedicated compatibility boundaries.