PrismetSage Deutschle

PrismCode

Mission control for Claude Code, Codex, and DeepSeek

Building

Scroll to explore · select an image to enlarge

1 / 2

Compare two proposed changesScripted demo
Inspect the diffScripted demo

Inside the work

Make agent work comparable.

Scripted demo

My contribution

Design and engineering of a desktop workbench for Claude Code, Codex and DeepSeek, with shared session and review surfaces.

Engineering decisions

Normalize at the adapter boundary
Agent-specific adapters emit one event contract for streamed text, tool activity, file changes and permission requests, keeping provider details out of the shared interface.
Put review before adoption
The comparison workflow presents proposed changes side by side with explicit choices to apply one result or keep both. A mock-agent loop supports interface development.

What you can inspect

The scripted demonstrations show the comparison and diff views. They illustrate the review workflow; they do not establish real agent performance or production usage.

Explore the project images ↑

A desktop IDE where three coding agents work side by side over one workspace, each normalized into a single event stream. PRISM A/B races Claude against Codex on the same prompt in separate git worktrees, then diffs the results so you keep the winner. Sessions resume, accounts hot-swap, and a usage tracker shows spend per account.

Highlights

One adapter per agent, one AgentEvent contract: adding an agent needed no UI work
Git worktree racing with compare, apply-winner, and keep-both
Zero-token dev loop with a mock agent

How the agents share one workspace

PRISM A/B: two git worktrees, one prompt Compare, apply the winner, or keep both

The PRISM A/B compare view for one prompt: Claude changed 2 files (+47) and Codex 1 file (+16), their diffs side by side, with Apply Claude, Apply Codex, Keep both and Discard buttons.