Method format - Method

A .method file is YAML (JSON also works). New files use format: method/3.4. Files in method/3.1 to method/3.3 load and run unchanged. The specification and JSON Schema define the format. method schema prints a short field reference.

A complete Method

The quickstart example (download) runs one script and one check script:

format: method/3.4
name: Save and check a text file
goal: Preserve the exact supplied text in a saved file.
inputs:
  text:
    type: text
    description: Exact text to save, including whitespace and the final newline.
steps:
  save:
    name: Save the file
    purpose: Save the supplied text and verify its full contents.
    reading:
      check_name: Check the saved text
      check: Read the saved file and compare its contents with the supplied text.
    in:
      text: inputs.text
    do:
      kind: run
      runtime: node
      entrypoint: save-text.mjs
    limits:
      timeout_ms: 10000
    out:
      file:
        type: file
        description: The file containing the complete supplied text.
        format: text
    check:
      kind: run
      runtime: node
      entrypoint: check-text.mjs
result: file

format, name, goal, steps, and result are required. Everything else is optional.

Top-level fields

The top of the shipped support-triage example:

run_data: device
models:
  writer: anthropic/claude-haiku-4.5
secrets:
  HELPDESK_TOKEN: API token for the help desk. read-ticket.mjs and set-ticket.mjs send it as a Bearer token.

Field	What it holds
`id`	The Method’s ID (`wf_` and 32 hex digits). The CLI writes it as the second line at the first signed-in run. It is not part of the content digest. See Publish and versions.
`models`	Model names for steps. A value is a model ID (`openai/gpt-6-luna`), `{model, max_output_tokens?, reasoning_effort?}`, or `{agent: codex \| claude, model?, reasoning_effort?}` for a local agent.
`secrets`	Names and purposes of keys that scripts need. See Secrets.
`limits`	Run limits, for example `{timeout_ms: 600000, max_model_requests: 20, step: {timeout_ms: 120000}}`. `limits.step` sets the defaults for steps. Without `limits`, the runtime’s defaults apply.
`tools`	Tools that `agent` steps name in their `tools:` list. A script tool is part of the saved package.
`inputs`, `state`	Typed values that a run starts with. State is saved in one `state.json`.
`environment`	Connections, each with a `type`: `files`, `service`, `browser`, `desktop`, or `tool`. `role: observer` makes a connection visible only to effect observers.
`files`	Helper files to save with the version. Script entrypoints are included automatically.
`run_data`	`account` (default) or `device`. See Privacy.
`run_prompt`	Plain text for the agent that starts a saved Method.

Steps

Each step has exactly one of do or ask. The do kinds are:

run: a script, with runtime (node or python), entrypoint, and optional args.
call: one model request, with prompt and an optional model.
agent: a model with tools, with prompt, tools, and an optional model.
classify: a classifier answer. See Classify steps.

A step’s model is a name from models:, a model ID, or absent (the account’s default hosted model). Any other name is the error unknown_model. Other step fields: name, purpose, in (local names for references, such as text: inputs.text), out (typed outputs), check, limits, changes, effects, when (a boolean reference), after, each (once for each item of a list), concurrency (items at the same time, with each), repeat, and accept (accepted warnings, see The issue checker). Types are text, number, boolean, record (with fields), list (with items or fields), and file (a value with path and sha256). Prompts can insert declared inputs with {{name}} or {{customer.name}}.

Effects

An external change is confirmed by reading the changed system, not by trusting the script’s receipt.

A files connection in changes needs nothing: the runtime reads the folder before and after the step.
A step that changes another connection (a service, email, a browser that sends) needs effects, or a one-sentence no_effect_reason.
An effect has an intent, optional in bindings, and an observe reader. The built-in observers http, sqlite, and file need no script.

From the support-triage example:

    changes: [environment.helpdesk]
    effects:
      updated:
        intent: The ticket has the chosen priority.
        in: {ticket_id: inputs.ticket_id, priority: decision.priority}
        observe:
          kind: http
          path: /tickets/{inputs.ticket_id}
          expect: {fields: {priority: "{inputs.priority}"}}

A contradicted effect fails the run. When no reading confirms the change in time, the run is unconfirmed and the CLI exits 3. method observe makes later readings. Observed effects are a different thing: a record of what a script process did. See Observed effects.

​A complete Method

​Top-level fields

​Steps

​Effects

A complete Method

Top-level fields

Steps

Effects