Steps

Soma uses a deliberately small step-list format — not an intermediate representation yet. A step list is a list of maps, and the executor is strictly sequential. The runtime must not depend on where steps came from: every future planner (the LFE DSL, an LLM planner) compiles down to exactly this format.

The step-list format

A step is a map with id, tool, args, and timeout_ms. The args may reference the output of a prior step with a simple from_step reference.

[
  #{id => read,  tool => file_read, args => #{path => <<"in.txt">>}, timeout_ms => 1000},
  #{id => up,    tool => echo,      args => #{from_step => read},     timeout_ms => 1000},
  #{id => write, tool => file_write, args => #{path => <<"out.txt">>, from_step => up}, timeout_ms => 1000}
]

The demo run that proves the v0.1 contract is exactly this shape: file_read → echo → file_write.

from_step wiring

The only data flow between steps is the from_step reference. When a step’s args carry from_step => SomePriorId, the runtime substitutes the recorded output of that prior step. This is the whole variable model — there are no named variables, no expressions.

What is deliberately out of scope

The executor is sequential: validate the step, start the tool call, wait for the result, record the event, advance to the next step. There is no branching, no loops, no DAG, no variables beyond from_step. Parallelism and richer planning are roadmap items that still compile down to this same list.