Inspiration

I was inspired by coding agents and harnesses such as Claude Code, Hermes Agent, PewDiePie’s Odysius, and OpenCode. They showed me how much a model’s capabilities can grow when it has the right tools and a thoughtful harness around it.

I was also interested in discussions of DeepSeek’s harness and KV-cache optimizations. That pushed me to build a strong, flexible harness for custom models—and to explore how an agent could make code changes while providing evidence that they work.

What it does

AngChat’s PatchProof workflow turns a coding task into an inspectable patch. The agent uses NVIDIA Nemotron through Nebius Token Factory to inspect a selected repository, read and edit files, and work through a coding task.

It can run the repository’s configured checks locally and save a JSON verification receipt with the task, model, scripts, results, timings, and output. Developers can review that evidence alongside the Git diff.

How we built it

We built the app with Node.js and Express, SQLite for sessions and settings, and a browser-based interface. A streaming agent loop connects the model to workspace tools for finding, reading, and editing files, running commands, and inspecting Git status and diffs.

The verifier runs only configured npm scripts from an allowlist: test, lint, typecheck, and build. It records the actual results in local receipts. The app also supports multiple model providers and optional local model inference.

Challenges we ran into

A useful harness has to do more than send a prompt to a model. We had to connect model responses to real tools, keep the agent’s file operations within the selected project, and make sure verification results reflect commands that actually ran.

We also worked through differences between model providers, streaming responses, tool calls, and the context needed to understand a codebase. Showing verification results clearly—and distinguishing a failed check from a successful tool call—was another important design challenge.

Accomplishments that we're proud of

We built a working repository-aware coding agent that can inspect and edit a real project, run its checks locally, and save a reviewable receipt. The receipt and Git diff give developers concrete material to inspect instead of asking them to rely only on the agent’s summary.

We’re also proud that the harness can work with hosted and local models, and that the demo repositories give us small, repeatable scenarios for showing the complete workflow.

What we learned

The model is only one part of an AI coding system. The harness matters: clear tool boundaries, focused instructions, useful feedback from checks, and a way to inspect the result all shape how helpful the agent is.

We also learned that verification should be visible and specific. A check’s name, output, timing, and result make it easier for a developer to decide what to trust and what still needs review.

What's next for AngChat

Next, we want to broaden verification beyond npm-based projects, make receipts easier to compare across runs, and continue improving the agent’s ability to recover from relevant check failures.

We also want to make AngChat easier to try through a dependable hosted demo, while keeping the workflow clear about which model handled the code and which checks ran locally.