DevMate AI

Inspiration

Developers often spend more time understanding an unfamiliar codebase than writing the actual change. A small request—such as adding authentication, improving tests, or updating documentation—can require tracing routes, models, dependencies, security concerns, and existing conventions across many files.

I wanted to build something that feels less like a generic AI chat window and more like a careful software-engineering teammate. That idea became DevMate AI: a workspace that helps developers understand a repository, investigate safely, plan changes before editing, and review the real Git diff afterward.

What it does

DevMate AI is an AI Software Engineering Copilot for local Git repositories.

A developer connects a repository, then can:

• Ask Codex to explain the architecture or a specific file.
• Review security and code-quality risks.
• Generate a repository-aware implementation plan.
• Review affected files, risks, testing, and documentation work before approving.
• Run Codex only after explicit approval.
• Inspect the actual Git status and diff produced by the approved change.

The core workflow is:

Understand → Investigate → Plan → Approve → Implement → Review diff

How we built it

DevMate is a lightweight local web application built with:

• HTML, CSS, and JavaScript for the engineering-workspace interface.
• Node.js for the local server, repository validation, source indexing, Codex gateway, and Git-diff retrieval.
• Git for transparent status and diff review.
• Codex with GPT-5.6 for repository investigation, planning, and approved code changes.

The interface includes a repository dashboard, architecture view, source explorer, security signals, plan approval screen, changes page, persistent assistant panel, command palette, file upload option, and dark plus white-and-lavender light themes.

For runtime functionality, DevMate connects to the developer’s locally signed-in Codex CLI. Read-only tasks use a read-only sandbox for architecture explanations, security reviews, and implementation planning. Workspace-write access is used only after the developer explicitly approves a plan.

How Codex and GPT-5.6 were used

Codex and GPT-5.6 were used both to build DevMate and inside DevMate’s engineering workflow.

During development, Codex accelerated implementation of the workspace UI, local repository workflow, Codex CLI bridge, planning and approval states, Git diff review, diagnostics, theme support, and testing refinements.

Inside the app, GPT-5.6 and Codex help DevMate:

• Inspect real repository files and explain architecture with file references.
• Create focused, multi-file implementation plans.
• Consider security, tests, documentation, and review steps before editing.
• Apply approved changes inside the connected Git workspace.
• Return activity signals and enable review of the real generated Git diff.

Challenges we faced

The most significant challenge was making the local Codex CLI integration reliable on Windows.

The project needed to work without asking users to paste an API key into the interface, while still keeping repository changes safe and transparent. Windows process launching, PowerShell execution policy, command forwarding, and the Codex CLI’s input behavior created several integration issues.

We solved this by creating a local PowerShell bridge for the signed-in Codex CLI, passing arguments safely, closing unused standard input so Codex would not wait for additional input, adding a built-in Test Codex connection check, and adding server diagnostics with run IDs. This made the workflow easier to verify before a user asks Codex to investigate or change a repository.

Another challenge was balancing a polished hackathon UI with a trustworthy engineering workflow. The result was an approval-first design: DevMate never presents an AI-generated change as automatic or hidden. The developer sees the plan before editing and the Git diff after editing.

What we learned

This project reinforced that agentic developer tools need more than an LLM response. They need clear boundaries, visible progress, reliable local integrations, and developer control.

We also learned that the most useful AI workflow is not “generate code immediately.” For real projects, the better sequence is to understand the repository first, plan the smallest complete change, obtain approval, and make every modification reviewable.

DevMate AI is our attempt to make that workflow visible, practical, and approachable.

Built With

• codex
• codexcli
• css3
• git
• gpt-5.6
• html5
• javascript
• node.js
• powershell