The issue checker

method validate FILE checks the Method and this computer’s setup without running the work. It lists issues. Each issue has a level, a code, where it is (at), a one-sentence message, and a one-sentence fix.

Level	What it means	Blocks a run or a publish
error	The run would be wrong or unsafe.	Yes. `validate` exits 1 and `publish` publishes nothing.
warning	A likely problem. Fix it, or accept it with a reason.	Never.
note	Advice. Shown only with `--notes`.	Never.

validate lists warnings only for steps whose text changed since the last validate, at most 5. --all lists every warning.

Codes

Code	Level	When
format codes (for example `unknown_model`)	error	A field, reference, type, or dependency is wrong. `unknown_model`: a step’s `model` is not in `models:`, not a model ID, and not `default`.
`legacy_runtime_json`, `legacy_sidecar`	error	A `runtime.json` or `FILE.method.json` is next to the Method. See Migrate from 0.13.
`secret_value`	error	The Method or one of its files holds a value of this computer’s secrets. The issue never shows the value.
`missing_secret`	warning in validate, error at run	A declared secret has no value on this computer. See Secrets.
`model_not_private`	warning	A step names a hosted model that no private provider serves. Hosted runs refuse it. Choose one from `method models`.
`untrusted_content_can_act`	warning	An agent with a browser can also change things.
`unused_output`	warning	No step, check, effect, result, or run label uses an output.
`script_calls_model`	warning	A script appears to call a model. Give each prompt its own `call`, `agent`, or `classify` step.
`accept_unused`	note	A step accepts a code that no longer fires on it.

Model checks

Some checks ask a model a yes/no question about a step, through your Method account. They run for changed steps, and their answers are cached. validate waits at most 3 seconds and reports the rest as pending. A new model check starts in shadow mode: it is computed and recorded on this computer, but not shown. It is shown only after its precision on a labeled test set reaches its level (0.90 for a warning, 0.75 for a note). Today only script_calls_model is shown. prompt_multiple_tasks, check_enforces_wording, classify_needs_math, classify_no_unclear, and prompt_no_missing_input run in shadow mode.

Accept a warning

When a warning is not a problem for this task, accept it on the step with the user’s reason. The shipped support-triage example accepts one:

  move:
    # ...
    accept:
      unused_output: The receipt stays in the run record. The result cannot name it, because the step is skipped when the team is clear.

The issue stays in the list with accepted set to the reason. Errors cannot be accepted. Do not add a check only to remove a warning.

Start

Method lifecycle

Worked examples

The issue checker

Codes

Model checks

Accept a warning

​Codes

​Model checks

​Accept a warning

Codes

Model checks

Accept a warning