Voice Dev Loop

Inspiration

Software requirements often move through meetings, notes, Jira, and GitHub by manual copying. That process loses context and slows delivery. Voice Dev Loop tests a smaller idea: speak one feature request, review what the system understood, and let a controlled workflow prepare a tested draft pull request.

What it does

A user records a short English requirement through a push-to-talk interface. The app uploads the recording only after it stops, transcribes it, and lets the user correct the transcript. It then converts the transcript into a structured requirement with a title, description, acceptance criteria, constraints, and open questions.

Nothing changes externally until the user reviews and confirms the requirement. After approval, Voice Dev Loop creates a Jira story and dispatches a GitHub Actions workflow. GitHub Copilot CLI prepares the code change inside a restricted path. A separate validation job checks the changed paths and runs lint, TypeScript, tests, and a production build without network access or CI secrets. The result is a draft pull request for human review. It never merges automatically.

How we built it

The controller is a minimal Next.js and TypeScript application. Groq Whisper handles recorded English speech. GitHub Models structures each requirement. Jira Cloud stores the approved story. GitHub repository dispatch starts the coding workflow in a separate playground repository.

The GitHub workflow gives the coding agent access only to the demo source and test folders. It validates the generated diff before running checks in an isolated container. Jira, the Actions run, and the draft pull request remain visible so a reviewer can follow the request from speech to code.

How we used Codex and GPT-5.6

We used Codex and GPT-5.6 to design and implement the controller, workflow boundaries, input validation, restricted edit policy, and test sandbox. Codex helped keep the project small while still covering the complete path from spoken request to reviewed code. The GitHub workflow uses Copilot CLI to generate the controlled repository change.

Challenges

The main challenge was connecting several systems without giving the coding agent broad access. We kept human approval before external writes, restricted editable paths, separated code generation from verification, removed network access from the test container, and prevented automatic merge or deployment.

A second challenge was keeping the demo reliable. Provider quotas and long workflow waits can distract from the core idea, so the interface exposes clear status links and keeps the workflow narrow.

Accomplishments

The working demo records speech, produces an editable transcript, structures the requirement, creates a real Jira story, starts a real GitHub workflow, validates and tests the generated change, and opens a draft pull request.

What we learned

A useful agent workflow needs clear boundaries more than extra features. Small approval gates, explicit allowed paths, and independent tests make automation easier to trust and explain.

What's next

The next step is a hosted deployment with durable workflow state, trace identifiers, signed Jira events, stronger duplicate protection, and a GitHub App with short-lived credentials. The current version stays intentionally small and ends at a tested draft pull request.

Built With

• next.js