runward runs from your agent.
Nobody wants to install runward by hand. An AI agent should discover it, set it up and operate it without you touching the keyboard. 0.19.0 makes that true: a baseline that privileges no tool, a command that recognises yours or asks which one it is, and output an agent reads without guessing.
One path, carried by the baseline
You run npx runward init, and that is all. From there, runward check works in any agent: the AGENTS.md baseline, always written, is enough, with nothing wired. And the default no longer privileges any tool. Init used to lean toward one agent; now it lays down the neutral baseline, and a channel is only added if you decide to.
runward wire: it recognises, or it asks
To have the check fire on its own, your agent runs runward wire. It recognises the tool running it, Claude Code and Cowork, Gemini, Cursor, by the signal it leaves; otherwise by a config file already present. If it recognises nothing, it does not guess: it asks you which one it is. And it never wires anything on its own. That is your gesture, or your agent's on your approval.
Driven on data, not on text
runward check --json returns a machine verdict, stable and deterministic: the current gate, what is missing, the conformance gaps. The agent branches on that instead of reading coloured text and guessing. It is the difference between a tool you drive and a tool you interpret.
None of this depends on perfect detection: the baseline runs everywhere, and wire proposes automation where it recognises the tool. When it does not, it asks you. That is honest, and it is sturdier than pretending to guess.
Install: npx runward init
Release v0.19.0 on GitHub · The documentation · CHANGELOG
← All news