agentrun
Inspect, validate, and run AgentRun workflow files
TLDR
SYNOPSIS
agentrun command [args]
DESCRIPTION
agentrun is the command installed by the npm package @parcha/agentrun-dsl. It needs Node.js 22.19 or newer. npx agentrun runs the same binary without a global install.The tool reads AgentRun workflow documents: JSON descriptions of steps that can call tools, code, typed judgments, and agents. demo uses scripted judgments. inspect only reads structure. validate, dry-run, and run can execute JavaScript from the file, and they refuse to start until --trusted is passed.run calls the interpreter with an empty adapter set, so a workflow that needs an agent, a Jev decision, or a host effect cannot be completed from this command. Supply those functions through the package's JavaScript API. A finished run prints JSON. Status complete exits 0. Any other status exits 2. Usage errors and thrown failures exit 1.Slash commands such as /agentrun inside the Pi editor come from the separate @parcha/agentrun-pi extension. They are not this binary.
PARAMETERS
--trusted
Required by validate, dry-run, and run. Confirms that JavaScript stored in the workflow may run, including during validation.--json
On demo and inspect, print JSON instead of the text report.--scenario NAME
On demo, select billing, technical, or ambiguous.-h, --help
Print the command summary. agentrun help does the same.
SUBCOMMANDS
demo [--scenario name] [--json]
Run the bundled triage workflow with scripted judgments. Makes no network calls. The scenario is billing (default), technical, or ambiguous.example [file.json]
Write that triage workflow as JSON. The default path is workflow.json. An existing file is left unchanged and the command fails.inspect workflow.json [--json]
Print the workflow structure. Does not execute workflow code and does not require --trusted.validate workflow.json [input.json] --trusted
Check the workflow. When an input file is given, also check it against the input contract. Exits 1 when the result is not ok.dry-run workflow.json [input.json] --trusted
Execute with the given input object, or with input synthesized from the declared input schema. With no input schema, the input starts as {}. Exits 1 when the result is not ok.run workflow.json input.json --trusted
Run the workflow with no host adapters. Agent steps, Jev judgments, and tool effects need the JavaScript API. Exits 2 when the status is not complete.
CAVEATS
A different project, Serverless-Devs agentrun-cli, also installs a command named agentrun (alias ar) for Alibaba Cloud AgentRun. That binary reads ~/.agentrun/config.json and manages sandboxes and hosted agents. Whichever agentrun appears first on PATH is the one the shell runs. agentrun --help from @parcha/agentrun-dsl lists demo, example, inspect, validate, dry-run, and run. The cloud CLI lists groups such as sandbox and super-agent.--trusted records that workflow JavaScript may run in this process. It does not isolate that code. Run untrusted workflows only inside a sandbox you control.example creates the target with an exclusive write. If the file already exists, it is not replaced.The package is published as a beta, and the command set can change between releases.
HISTORY
AgentRun is published by Parcha Labs under the Apache-2.0 license. The agentrun binary ships with @parcha/agentrun-dsl.
