Workflows and the run command

xtalforge_agent::workflow provides Workflow/Step (model version 1), Workflow::from_json/from_yaml/from_text/validate, schema_json/schema_text, expand_inputs, run_workflow with RunOptions/WorkflowReport, and the WorkflowCheckpoint resume format. The same model backs the CLI xtalforge run <workflow> --inputs ... --jobs ..., the agent verbs workflow_schema, workflow_validate and workflow_run, and the MCP tools with the same names.

The workflow steps are the agent verbs, so a workflow and an agent session share one vocabulary. The JSON Schema is returned by workflow schema and is also available at docs/public/workflow.schema.json.

YAML

YAML is a second syntax for the same model; JSON stays canonical. The parser is a small pure-Rust subset: block mappings and sequences, compact - name: value entries, flow lists/maps of simple values, quoted and plain scalars, comments, null/booleans/numbers and an optional leading ---. Anchors, aliases, tags, merge keys, block scalars, multiline plain scalars and multi-document files are rejected with a line and column. .yaml/.yml paths select YAML; other paths try JSON first and fall back to YAML. The JSON Schema marks YAML as accepted.

In-memory execution

run_workflow_memory(&workflow, &files) takes a BTreeMap<String, String> of input paths to text and returns MemoryWorkflowReport { report, outputs } without touching a filesystem. The WASM export run_workflow_memory wraps it, so the browser runs the same steps and returns written documents as a map. Checkpoints are not produced because there is no disk to resume from.

Parity key: workflow.