Stitch

Inspiration

Broken CI at 2am usually means: someone notices the red badge, opens a chat, pastes logs, guesses at a fix, opens a PR, and hopes it passes. That loop is slow, manual, and easy to skip when nobody's watching. Stitch automates the loop. A webhook fires on failure — no human opens a chat session to start it.

What it does

Stitch is a webhook-triggered agent for broken CI. When GitHub Actions fails, Stitch pulls the job logs, runs a diagnosis pass (OpenAI, Claude, Gemini, or Copilot — configurable per workspace), then a separate fix-generation pass that writes a real unified diff against your repository. It validates the patch, opens a pull request (or comments on an existing PR on feature branches), notifies Slack/email, and records everything in a Fix Log, Issue Record, and Audit Trail.

The same failure on main vs release/* vs feature/* gets different behavior — Autopilot, Fix & propose, or Diagnose & suggest — because production and experiments shouldn't be treated the same.

We built it as a real SaaS app: React dashboard, PostgreSQL multi-tenant backend, role-based permissions, and a seeded demo plus a live test repo with intentional CI failures for judges to trigger end-to-end.

Codex built the pipeline, GitHub plugin, branch router, dashboard, and multi-model AI layer. OpenAI (and other providers) power diagnosis and fix generation at runtime as two separate steps.