AI code bug fixing: from user report to merged fix

“AI wrote the code” changes what a bug costs. You can’t lean on the memory of having written the module; every investigation starts cold. The fix itself is cheap — an agent can write it in seconds — but only once someone has worked out what actually broke and where.

Mushi Mushi is built around that bottleneck. It’s the comprehension layer for AI-built apps: it turns raw user reports into diagnoses an agent (or you) can act on immediately.

The pipeline

Capture. Users report bugs in your app — widget tap on web, shake on mobile. The report keeps the screenshot, the page, and the user’s own words.

Diagnose. Mushi reads the report against your repo and writes a plain-English explanation: what the user hit, the likely cause, which files are involved, how severe it is. Classification lands in about ten seconds.

Deduplicate. Twenty reports about the same broken button collapse into one row, so your queue reflects distinct problems, not complaint volume.

Fix. Each diagnosis ships with a fix prompt scoped to the code involved. Paste it into Cursor or Claude Code, or let the editor pull it over MCP. Optionally, an agent opens a draft PR on your repo — you merge, edit, or ignore it.

Learn. Merged fixes become lessons (.mushi/lessons.json) that your editor sees on future PRs, so the same class of bug doesn’t come back next sprint.

Where this fits with your existing tools

Keep your error monitor for crashes and your test suite for regressions. Mushi covers the third category — the bugs only users find — and closes the loop back to your editor instead of a ticket queue.

Start with one command

npx mushi-mushi

The wizard installs the right SDK, writes env vars, and can file a test report so you see the loop before real users do. 50 diagnoses a month free; self-host if you prefer.

Next: Incident loop quickstart · How the pieces fit · Debug Cursor-built apps