Glasshouse-lite

Inspiration

AI coding agents can edit files, run commands, and modify entire repositories within minutes.

After they finish, developers usually only see the final Git diff. We lose the entire journey that led to those changes.

Questions like these are difficult to answer:

Which file was edited first?
Which command caused a failure?
When did the bug appear?
Has the session log been modified?

We wanted to make AI coding sessions as transparent as flight recorders are for airplanes.

What it does

Glasshouse-lite records observable actions performed by AI coding agents.

It captures events such as:

File reads
File writes
Terminal commands
Test execution
Command output
Session metadata

Every event is linked using a SHA-256 hash chain, making the session tamper-evident.

Developers can replay an entire coding session through an interactive timeline, inspect code diffs, verify cryptographic integrity, and understand exactly how the AI reached its final solution.

Importantly, Glasshouse never records hidden model reasoning—only observable operational events.

How we built it

The project consists of two major components:

A local recorder that wraps an AI coding agent and captures observable events.
A standalone replay viewer that visualizes sessions and verifies their integrity entirely in the browser.

Sessions are stored as portable JSON files or exported as self-contained HTML replay files that work offline with no server required.

Challenges we ran into

One major challenge was designing a universal session format that could support different AI coding agents.

Different agents expose different interfaces and event streams, so we created an adapter-based architecture that translates each agent into a common event model.

Another challenge was making replay files completely self-contained.

To ensure they work even when opened directly from disk (file://), we implemented SHA-256 verification entirely in JavaScript instead of relying on browser Web Crypto APIs that require secure contexts.