Rules

Rules P001 to P005: when each fires, its severity, its exact message, and how suggestions are chosen.

Each rule also has a page with an example and how to fix or allow it: P001, P002, P003, P004 and P005.

The rules

ID	Name	Severity	Trigger
P001	inline-tool	error, or tolerated via baseline, or allowed	A `run:` step calls a listed tool directly.
P002	unknown-task	error	An invoker names a task the runner does not define.
P003	internal-task	error	An invoker names an internal task: `internal: true` (Taskfile), private (just) or hidden (mise).
P004	regression	error	A workflow’s P001 count exceeds its baseline, or a locked workflow has any P001. One finding per workflow, located at line 1 of the workflow.
P005	unsupported	warning	Unsupported shell, shell parse error, dynamic command name, allow comment without reason.

Message wording (normative, because users grep for it). With another runner (runners), its command takes the place of task in task <name>: make lint, just lint, mise run lint.

P001: step "<step>" in job "<job>" runs '<command>' directly followed, when a suggestion exists, by ; already available as 'task <name>'. For composite actions use in action "<path>" instead of in job.
P002: 'task <name>' is not defined in <taskfile> followed, when a suggestion exists, by (closest: <name>).
P003: '<name>' is marked internal: true and should not be invoked from CI. With just it reads is private, and with mise is hidden, in place of is marked internal: true.
P004: <workflow> has <current> inline-tool findings, baseline allows <baseline> or <workflow> is locked at 0 and has <current> inline-tool findings.
P005: step uses shell '<shell>'; not analyzed, composite step has no shell; not analyzed, cannot parse step script: <error>, dynamic command name '<word>' cannot be analyzed, allow comment has no reason.

A step without a name: is referred to as step <index> (1-based) in place of step "<step>", unquoted: step 4 in job "web" runs 'pnpm test' directly.

Suggestions

P001: normalize each Taskfile command string (trim, collapse whitespace). If it equals the inline command’s normalized form, suggest task <name>, never suggesting internal tasks and choosing the shortest name. Ties go to the shorter name, then the alphabetically first. Only the canonical task name is suggested, not an alias. Treat pnpm run X as equal to pnpm X and yarn run X as equal to yarn X.

P001, from v0.3: both commands are also unwrapped once before comparing. A command that starts with a package runner below drops the runner words and the options after them; options are the arguments starting with - before the first other one, -- ends them, --flag=value carries its value, and a listed flag also drops the argument after it. Nothing left means no unwrapping. An inline command and a task command match when they are equal as written or unwrapped, on either side. Only suggestions change.

Runner words	Options that take a separate value	Not unwrapped when
`npx`, `pnpx`, `pnx`, `bunx`	`-p --package`	`-c` or `--call` is given
`pnpm exec`, `pnpm dlx`	none
`uv run`, `uvx`	`--with --with-editable --with-requirements --python -p --project --directory --extra --group --package --env-file --from --index --default-index`
`dotnet tool run`	none

P001: the task must also run in the step’s working directory (analysis), and calling the runner from that directory must read the same root file, so the suggested name exists there. Where each runner’s tasks run is in runners. When either is unknown, nothing is suggested, so fix leaves the step alone.
P002: a task name that has the unknown name as a prefix wins, shortest first. Otherwise suggest the name with the lowest Levenshtein distance, when that distance is at most 3 and less than the length of the unknown name. Internal tasks are never suggested.

Rules

The rules

Suggestions

See also