CLI Commands

Place scripts in a folder.

After writing, tell the user which command fits what they want to do:

• wmill script preview <script_path> — default when iterating on a local script. Runs the local file without deploying.
• wmill script run <path> — runs the script already deployed in the workspace. Use only when the user explicitly wants to test the deployed version, not local edits.
• wmill generate-metadata — regenerate the local .script.yaml (input schema) and .lock (resolved dependencies) for scripts you changed, and refresh their content hashes in wmill-lock.yaml. Local files only — not a deploy. See "Keep metadata in sync" below.
• Deploy local changes to the workspace — via git push or wmill sync push depending on how the repo is wired (see the Deploying section in AGENTS.wmill.md). Only suggest/run a deploy when the user explicitly asks to deploy/publish/push — not when they say "run", "try", or "test".

Preview vs run — choose by intent, not habit

If the user says "run the script", "try it", "test it", "does it work" while there are local edits to the script file, use script preview. Do NOT push the script to then script run it — pushing is a deploy, and deploying just to test overwrites the workspace version with untested changes.

Only use script run when:

• The user explicitly says "run the deployed version" / "run what's on the server".
• There is no local script being edited (you're just invoking an existing script).

Only use sync push when:

• The user explicitly asks to deploy, publish, push, or ship.
• The preview has already validated the change and the user wants it in the workspace.

Keep metadata in sync after editing

wmill-lock.yaml tracks a content hash for each item. Editing a script's content — most importantly adding or removing an import or changing main's arguments — invalidates that hash and leaves the .lock, the .script.yaml input schema, and the hash row out of date. Run wmill generate-metadata (scoped to what you touched) after such edits so the resolved lock, the auto-generated args UI (driven by .script.yaml), and wmill-lock.yaml all match the code. Leaving them stale produces spurious diffs in git-sync and CI.

This only writes local files (it is not a deploy), but it re-resolves dependencies, so it can bump unpinned versions (the same as deploying from the UI; expected, not a bug). So by default offer it and run it once the user agrees, rather than running it silently after every edit — unless the project's AGENTS.md opts into running metadata automatically (see the "Keeping metadata in sync" preference there). Either way YOU run the command, not the user. After running it, diff the regenerated .lock / .script.lock files and tell the user which dependency versions changed (e.g. requests 2.31.0 → 2.32.0), so they can catch an unwanted bump before deploying — even under Metadata: auto, since it's information, not a confirmation gate. Pin versions in code to keep them fixed.

With no path argument, generate-metadata regenerates only the items whose content hash drifted — not everything. Imports propagate: editing a script that others import marks every importer stale too, so a one-line change to a shared module can regenerate many locks (by design — their locks must reflect the imported code). If it touches more than you expect, run wmill generate-metadata --dry-run — it lists each stale item with a reason (content changed or depends on <path>) without changing anything — then narrow with a path argument (wmill generate-metadata f/foo) or --strict-folder-boundaries.

If the on-disk .lock and .script.yaml are already correct and only wmill-lock.yaml needs its hashes refreshed (hash drift, or bootstrapping missing entries), use wmill generate-metadata rehash — it re-records hashes from disk with no backend round-trip and no dependency changes.

After writing — offer to test, don't wait passively

If the user hasn't already told you to run/test/preview the script, offer it as a one-sentence next step (e.g. "Want me to run wmill script preview with sample args?"). Do not present a multi-option menu.

If the user already asked to test/run/try the script in their original request, skip the offer and just execute wmill script preview <path> -d '<args>' directly — pick plausible args from the script's declared parameters. The shape varies by language: main(...) for code languages, the SQL dialect's own placeholder syntax ($1 for PostgreSQL, ? for MySQL/Snowflake, @P1 for MSSQL, @name for BigQuery, etc.), positional $1, $2, … for Bash, param(...) for PowerShell.

wmill script preview does not deploy, but it still executes script code and may cause side effects; run it yourself when the user asked to test/preview (or after confirming that execution is intended). wmill generate-metadata does not deploy either — it only writes local files (locks, schemas, hashes) — but offer it before running (or run automatically if the project's AGENTS.md opts in), per "Keep metadata in sync" above. Deploying to the workspace (git push or wmill sync push depending on how the repo is wired — see the Deploying section) is the only step that mutates remote state — do it only when the user explicitly asks to deploy/publish/push.

For a visual open-the-script-in-the-dev-page preview (rather than script preview's run-and-print-result), use the preview skill.

Use wmill resource-type list --schema to discover available resource types.

Windmill Script Writing Guide

General Principles

• A script's inputs are its parameters. Credentials and configuration come in as resource-typed parameters, never hard-coded or read from the environment; the language section below shows how that language declares parameters
• When no resource type exists for a service, still take its connection as one resource-typed parameter, of a custom type holding the connection fields (base URL, account identifiers, credentials), its secret in a variable the resource references. A custom type is named c_<name>, which keeps it apart from a hub type synced later: the parameter takes RT.C<Name> in TypeScript, a c_<name> TypedDict in Python. A plain string parameter or a variable read inside the code only when the user asks for it: the resource keeps a connection's fields together, lets whoever runs the script pick an existing connection, and keeps the secret out of the arguments a run records. If the instance is reported as not having synced the Windmill Hub's resource types, mention that a superadmin can sync them, since the hub has types for most services. Otherwise, or when the user prefers it to syncing, propose the resource type and ask before creating it
• A Windmill Hub script takes a resource type the hub publishes, so on an instance that has not synced the hub's types that type is missing: say so, and that a superadmin can sync them, before asking for a resource of it. Hub scripts often declare the type inline (type Stripe = { token: string }), which is still the stripe resource type; in code you write, name it the way the language section shows (RT.Stripe in TypeScript) rather than copying the inline declaration
• Before relying on a resource type, a hub script's included, run wmill resource-type list: it shows whether the workspace has the type, and warns when the instance has not synced the hub's types
• Libraries are installed automatically - do not show installation instructions
• In a language with an entrypoint function (TypeScript, Python, Go, Rust, PHP, R, …), name it main (Main in C#) and do not call it; in TypeScript it must be async. SQL, GraphQL, Bash, PowerShell and Ansible scripts have no main: their language section shows how they take arguments
• Where the language has a Windmill client (wmill), use it to interact with the platform
• A script's input schema may carry a top-level prompt_for_ai string: its author's instructions to an AI choosing the inputs. Follow it when you pick arguments to run that script, and keep it when you rewrite the schema

Return Values

• A script can return any JSON-serializable value; a SQL script returns the rows its query produces
• Return values become available to subsequent flow steps via results.step_id

Preprocessor Scripts

Preprocessor scripts process raw trigger data from various sources (webhook, custom HTTP route, SQS, WebSocket, Kafka, NATS, MQTT, AMQP, Postgres, GCP Pub/Sub, Azure, or email) before passing it to the flow. This separates the trigger logic from the flow logic and keeps the auto-generated UI clean.

A preprocessor is written in TypeScript or Python: its function is named preprocessor instead of main, and it receives a single parameter called event (the language section gives its type).

The returned object determines the parameter values passed to the flow. e.g., { b: 1, a: 2 } calls the flow with a = 2 and b = 1, assuming the flow has two inputs called a and b.

Java

The script must contain a Main public class with a public static main() method:

public class Main {
    public static Object main(String name, int count) {
        java.util.Map<String, Object> result = new java.util.HashMap<>();
        result.put("name", name);
        result.put("count", count);
        return result;
    }
}

Important:

• Class must be named Main
• Method must be public static Object main(...)
• Return type is Object or void

Maven Dependencies

Add dependencies using comments at the top:

//requirements:
//com.google.code.gson:gson:2.10.1
//org.apache.httpcomponents:httpclient:4.5.14

import com.google.gson.Gson;

public class Main {
    public static Object main(String input) {
        Gson gson = new Gson();
        return gson.fromJson(input, Object.class);
    }
}