Node types

A workflow is a graph of nodes in YAML (schema 2.0). Expressions use CEL; there is no JavaScript eval anywhere.

Node	Does	Uses a model?
`tool`	Calls a registered tool through the gateway (HTTP API, MCP server). Arguments can use `ref` and CEL; `project` keeps only the fields you list	No
`script`	Runs Python (via uv) or Bun on a worker, with input, `output_schema` and `limits` (time, memory on Linux)	No
`agent`	Runs a profile (instructions, model, limits) on an executor, with gateway-only tools, validated output and a budget	Yes
`retrieve`	Hybrid search over a knowledge dataset (`top_k`), returning excerpts with ids for citation	No (lexical-hash embedder in alpha)
`condition`	Branches on a CEL expression	No
`parallel`	Bounded fan-out of branches	No
`loop`	Repeats one `tool` or `script`, up to `max_iterations`, until a CEL `exit`	No
`subworkflow`	Runs a published workflow as a child run, with its own plan, ledger and run page	Depends on the child
`approval`	Waits for a person to approve or reject, optionally with structured decision data	No
`report`	Renders a Markdown template into a run artifact	No
`notify`	Sends a message (for example Slack), recorded in the ledger	No

Only the agent node spends tokens. Everything else is deterministic.

A taste of the syntax

nodes:
  - id: ci_runs
    type: tool
    tool: ci.list-runs@1
    arguments: {team: {ref: inputs.team}, since: {cel: "string(now - duration('336h'))"}}
    project: [id, status, branch, started_at, finished_at, failed_tests]
  - id: metrics
    type: script
    runtime: python
    entrypoint: scripts/metrics.py
    input: {map: "{'runs': nodes.ci_runs.output}"}
    output_schema: schemas/metrics.json
    limits: {time: 2m, memory_mb: 512}
  - id: analyse
    type: agent
    profile: quality-analyst@1
    datasets: [quality-guidelines@approved]
    output_schema: schemas/analysis.json
    budget: {max_output_tokens: 4000, max_tool_calls: 0}

(From examples/quality-report/workflow.yaml, abridged.)

Loops

- id: refine
  type: loop
  initial: {n: 0}
  max_iterations: 5            # 1 to 1000
  exit: "state.n >= 3"         # CEL, checked after every iteration
  on_max: fail                 # or continue; default fail
  node: {type: script, runtime: python, entrypoint: scripts/step.py, input: {map: "{'n': state.n}"}}

Each iteration is an ordinary attempt with its own ledger ordinal, so writes inside a loop are deduplicated per iteration. Agents as a loop body, and several steps in one body, are not supported; use a subworkflow.

Subworkflows

- id: review
  type: subworkflow
  workflow: doubler            # a workflow id, or id@version
  input: {x: 21}

The child is a run of its own, started as a Temporal child workflow. Cancelling the parent cancels the child. A workflow cannot call itself directly or through its chain, nesting stops at four levels, and the child’s output is treated as untrusted for the taint rule. Parallel fan-out of subworkflows and quorum joins are not built.