RCF Lite

Your coding agent ships code fast, but what the product is supposed to do lives in prompt history, and prompt history is not a spec: weeks later nobody can say what is covered, what is tested, or what breaks when something changes. RCF Lite keeps those answers machine-checkable, as a live chain running from what you asked for, through requirements, user stories, acceptance criteria and tests, into the code itself.

rcf-lite is the tooling: one npm package, plain JSON files that live in your own repository, and an MCP server your agent drives. Your agent authors and maintains the chain. You wire it once per repository, then steer.

One prerequisite before anything else: Node.js 24 or newer. On an older Node the install appears to succeed but delivers a placeholder release with no CLI in it, so run node -v before you run npm install.

See it work in ten minutes Wire your agent Understand the method first

How you and your agent judge the tree together

The tool referees structure. rcf define validate proves every document matches its schema and every reference resolves; rcf audit coverage counts a test case only when its pointer resolves to a real test. Adequacy, whether the tree captures what you actually want built, is judged by you and your agent together, at two named moments:

• The docs-review gate, once, before the first build. Your agent has drawn a chain of requirements out of a conversation with you, and it stops to walk you through it. You open the tree in the viewer, read what the agent captured against what you meant, and push back on what is missing. Reviewing the tree covers how.
• The review stage of each build cycle, once per build item. Before an item is marked complete, your agent runs a fresh-context self-review that drives the running app against its acceptance criteria, a reviewer that uses the product rather than reads the code.

The judgement is two-party by design: the tool can prove the tree is well formed, and you can prove the tree is right. Both matter.