devagrawal09/stanley-code

Stanley is a command-line tool a developer or coding agent can hand judgment work to instead of doing it inline: ranking which files in a repo are relevant to a task, checking whether a diff actually matches its stated task and the project’s own rules, and triaging test failures or review comments. A request is described in plain language, and Jev routes it to one deterministic workflow, a built-in one, one trusted from another repository, or one Stanley wrote for itself, that gathers bounded evidence and asks Jev small fixed-choice questions about it. Code makes the final decision, not the model, and this path never starts a general agent.

When no workflow fits a request, Stanley falls back to the Pi coding agent if it’s installed, reports what Pi says it did without independently verifying that account, and separately asks an agent in the background to draft a new Stanley workflow for that kind of request. The candidate is validated and staged; once a person promotes it, the next such request runs without any agent. In one example in the README, Stanley catches a change meant to fix a crash that also silently skipped a test and removed an assertion, flagging both.

The project isn’t on npm yet; it’s built from source, needs Node.js 22.18 or newer, a Git repository, and a TypeSafe API key. CI tests on Linux; Windows is untested.

How it uses Jev

return client.systemOne(
  { state: request.state, questions: request.questions, model: request.model },
  { timeout: options.timeoutMs, ...(options.signal ? { signal: options.signal } : {}) },
);

View in src/adapters/jev.ts

Open on GitHub