Hey Product Hunt 👋

I build most of my software with coding agents now. I give them the task and the architecture and let them run. A few times, when I pushed an agent to improve a feature or squeeze out performance, it rewrote core code I never wanted touched, and the tests still passed. The diff showed what changed, but not why.

I wanted the agent’s decisions and assumptions put in front of me, so I could judge the work without reading hours of session. So I built Sonah.

What it does: it reads the agent’s recorded session along with the code, and gives you one page ordered by what needs your attention:

Did it do what you asked? Each ask is marked done, partly done, not done or not checked.
What did it decide on its own? Decisions and assumptions, for you to confirm or overrule.
Is “done” true? Claims the agent’s own results contradict, like “all tests pass” after a failing test.
What’s wrong in the code? A read-only code review agent checks the change against your asks and the agent’s claims.

Every quote is checked word for word against the session or the code. There’s no score and no verdict; you make the call.

How it runs: it’s a CLI that runs locally through your own Claude Code login, so it uses the plan you already have. No account, no telemetry. It writes an HTML page for you and a Markdown report you can drop into a PR.

npm install -g getsonah
sonah doctor
cd your-project && sonah review --format html --output ~/review.html

I’d love to hear what you think, especially about the decisions it flags, and what it should catch that it doesn’t. I’m here all day to answer questions.