RFD 1037: Composite models as taskweft domains
Problem
Five catalog entries name one task and run several networks. seethrough_layer_decomposition runs nine. A Python script that calls them in order hides the order, the guards, and the failure points.
A script also cannot replan. When a stage fails, the caller repeats the whole job.
Decision
Model each composite as a taskweft domain and a paired problem. taskweft is an HTN planner at github.com/taskweft/taskweft, and it serves plan and validate over MCP at https://taskweft-mcp.fly.dev/mcp.
Write the domain in the Elixir DSL. The plan tool takes format of dsl by default, and JSON-LD is the fallback. Each file is real Elixir, thus Code.string_to_quoted/1 checks it with no planner.
The model image then calls the plan, and it runs one action per step. The order lives in the domain, and not in the Python.
See DETAILS.md for the domain shape and the type rules. It also covers why the solved plan is a checked-in file, why replanning is the payoff, and the table of all five composites.