autofix-watch
Log watcher that drafts its own fixes
Problem
Client apps break at night and the first sign is often a message from the client. Fully automatic AI fixes are worse: they patch config errors that aren't code bugs, or ship to prod with nobody looking.
Approach
- Polls dev and prod logs over SSH every 2 minutes, reading only new lines (file offset, journald cursor or docker --since), plus health checks with a restart button.
- Fingerprints errors (normalised message + first app stack frame → SHA-1) so the same bug is one alert with a count, and a fixed bug that returns raises a “came back” alert.
- Telegram bot with inline buttons: let Claude fix, mute, ignore, retry with a hint, deploy to prod, roll back. Secrets, emails, JWTs and card numbers are redacted before anything leaves the server.
- Guard rails around Claude Code: a throwaway git worktree, tests before dev deploy, one run at a time, 4 runs an hour, 15-minute timeout, and no push, SSH or network tools for the AI.
Result
Errors reach my phone within two minutes, grouped and redacted. Claude drafts and tests a fix on a branch, dev gets it automatically, and prod changes only after one tap from me.
Stack
The preview above replays the hardest moment of this project's Experience, drawn from the same simulation code. Click it to try the whole thing yourself.
Under the hood
A simplified excerpt of the core logic and how a request flows through the system.
More projects
Healthcare Platform
Telemedicine & Doctor Booking
HRMS System
Attendance, payroll & payslips
Jira-style Tool
Sprint & task management
Ticketing Platform
Event seat selection & ticketing
Hotel PMS
Cloud Hotel Property Management System
DIT
Document Injection Tool · AI document assistant
Job Consultancy Portal
Job listings, applications & admin