ReviewOps AI

Inspiration

Customer feedback is scattered across forms, support messages, reviews, and team chats. Valuable reports often disappear because product teams still need to validate the issue, inspect the code, agree on a solution, test the change, and communicate the result.

We built ReviewOps AI to turn that fragmented process into a controlled workflow. The goal is not to let AI change a product by itself. It is to help a business convert customer evidence into a reviewable, tested improvement while keeping people responsible for the final decision.

What it does

The current MVP can:

• accept feedback from a product page, local API, JSON input, or team chat;
• validate submissions and detect duplicate deliveries;
• assign feedback to a configured organization and repository;
• use Codex to inspect a trusted codebase and prepare a code-grounded proposal;
• send the proposal to Discord for engineering discussion;
• let authorized reviewers request changes, approve a specific proposal version, or decline it;
• implement only the approved plan in an isolated Git checkout;
• run repository tests and generate a patch, test log, review summary, and fixed preview;
• preserve task history, approvals, failures, and delivery receipts in SQLite.

For the demo, a customer reports that exporting an empty contact list crashes instead of downloading a CSV. ReviewOps AI reproduces the problem, prepares a proposal, accepts an engineer's request to preserve the CSV headers, creates a revised plan, waits for explicit approval, applies the scoped fix, and verifies the result with tests.

Agent-manager architecture

A central workflow manager controls state, permissions, and handoffs across six specialist stages:

1. Intake and validation checks the payload, source ID, organization, and duplicate status.
2. Investigation gathers the captured error and relevant repository context.
3. Solution planning uses Codex to produce a versioned proposal with test criteria.
4. Human review lets engineers discuss the proposal and an authorized reviewer decide.
5. Implementation and verification applies only the approved plan in an isolated checkout and runs tests.
6. Delivery creates a review bundle and preview instead of silently deploying a change.

Several responsibilities are deterministic workflow services rather than independent language-model agents. This keeps the MVP reliable and auditable. A fuller multi-agent version is part of the roadmap.

How we built it

The application uses Node.js and TypeScript. SQLite stores tasks, plan versions, approvals, job state, audit events, and delivery receipts. The workflow engine enforces state transitions and rejects stale or unauthorized approvals.

Codex SDK supports repository investigation and approved implementation. Discord.js provides the engineering discussion and approval interface. The project also contains optional adapters for Slack, Inngest, OpenRouter, Brave Search, and a read-only Model Context Protocol service for task-scoped logs.

Implementation runs in a separate Git checkout. The original repository remains unchanged, and no change is deployed automatically. Secrets are excluded from the coding subprocess, and customer feedback is treated as untrusted evidence rather than executable instruction.

Challenges we ran into

The hardest part was not generating code; it was defining safe boundaries around generation. We had to make approval explicit, version proposals, reject stale approvals, recover interrupted jobs, prevent duplicate feedback tasks, restrict actions to a configured repository, and make failures visible instead of reporting false success.

Review authenticity is another difficult problem. The current MVP retains source identifiers, detects duplicate deliveries, and applies simple spam or grouping rules, but it does not claim to prove that a human review is true. Future versions will use provenance, cross-source corroboration, anomaly signals, and calibrated confidence labels.

Accomplishments that we're proud of

• A working feedback-to-tested-fix demonstration, not only a concept screen.
• Code-grounded proposals that stop for human review before implementation.
• Versioned approval and request-changes controls in Discord.
• Isolated Git execution with test output and review artifacts.
• Persistent audit history, duplicate-delivery handling, and restart recovery.
• Clear boundaries: no automatic merge, deployment, or unsupported claim of review authenticity.

What we learned

An AI product for business operations needs more than an accurate model response. It needs identity, authorization, durable state, evidence provenance, failure handling, and a human decision point. We also learned that a deterministic service is often better than another agent when a step requires predictable enforcement.

What's next

We plan to add support-platform and CRM connectors, evidence provenance and confidence scoring, specialist agents for clustering and prioritization, secure tenant isolation, restricted containers for untrusted tests, recoverable GitHub draft pull requests, and a customer-facing status view for acknowledged, planned, fixed, and released issues.

ReviewOps AI shortens the distance between customer evidence and product improvement without removing human accountability.

Built With

• codex
• discord.js
• github
• node.js
• sqlite
• typescript