Workflow DSL

The current workflow DSL is a deliberately limited Rust builder chain. It favors static extraction over arbitrary runtime behavior.

Supported Calls

• workflow(...)
• .input_schema(...)
• .output_schema(...)
• .required_model(...)
• .api_step(...)
• .tool_step(...)
• .thread_step(...)
• .openai_chat_input(...)

Arguments must be string literals or META.id.

API Step

.api_step("draft", "node.llm_prompt", "openai.chat")

This maps to:

/ctx/home/<uid>/api/openai.chat/inbox/draft.req.json
/ctx/home/<uid>/api/openai.chat/outbox/draft.resp.json

Tool Step

.tool_step("read_context", "node.read_context", "filesystem.read")

This maps to CortexFS tool invocation paths.

Thread Step

.thread_step("append_turn", "node.thread_message", "design-review")

This maps to a user-visible CortexFS thread inbox.

Request Template

.openai_chat_input("llm.planner", "prompt")

If the caller does not provide an explicit submit body, LightFlow renders an OpenAI chat request from inputs.prompt.

Current Limits

The DSL does not yet express:

• step output references
• DAG dependencies
• branches
• retries and timeouts
• streaming outputs
• artifact naming
• complete schema validation
• complete route context mapping for model aliases

These limits keep preview deterministic while the core backend contract stabilizes.