Project Steward

Overview

Project Steward is a local CLI for making AI-assisted development easier to review. Instead of treating chat output as project memory, it records what happened: what was claimed, what evidence supported it, what a reviewer said, what decision was made, and whether anything should become part of the project record.

I built it as a TypeScript and Node.js tool with local project records. The key boundary is that model output is advisory by default. It can help review, summarise, or produce a result, but project memory changes only through explicit records and human decisions.

What I Built

The core model treats development work as records rather than loose chat:

• turn records and project events
• claims and evidence
• reviewer notes and reviewer results
• synthesis records
• steward decisions
• provider execution records

That gives the project a way to remember what happened without depending on the conversation alone. It also keeps review work inspectable: a result can be traced back to the prompt, provider path, evidence, and decision that produced it.

Guarded Reviewer Path

The OpenAI reviewer path is deliberately gated. It requires explicit confirmation, checks provider settings, reads the API key only after the gate, sends a single reviewer request, and records the result as provider output.

That boundary is the important part. The reviewer path does not run the whole project pipeline, update durable memory, make final decisions, or verify claims on its own. It produces a review result that can be inspected and acted on later.

Validation

The current Project Steward dossier pass validated the project with local showcase checks, make doctor, make check, and finch-standards verify .. This case study does not claim public usage, production deployment, automatic security auditing, vulnerability discovery, formal security hardening, or compliance.

Why It Matters

AI-assisted development becomes risky when generated text blends into validated project records without a trail. Project Steward pushes the other way: generated output stays useful but advisory, and project records change only through evidence and explicit decisions.

That matters for developer tooling because the hard part is not only getting a model response. The hard part is making the response inspectable later: what prompt produced it, what evidence supported it, what decision followed, and what should remain unresolved.

Project Steward is a local-first active build. This case study focuses on the implemented governance model, guarded reviewer path, and evidence-record structure, not on public deployment or automatic security auditing.