A developer built a GitHub Actions pipeline that detects post-deploy regressions, statistically triages causation, and opens fix PRs via an open-source coding agent automatically.
A developer published a technical breakdown of a self-healing deployment pipeline built for their GTM Agent on LangSmith Deployments. The system uses a GitHub Action triggered on every deploy to main, captures build and server logs, applies statistical regression detection (comparing 7-day baselines to 60-minute post-deploy windows with normalized error signatures), and routes confirmed regressions to Open SWE — an async open-source coding agent — to write a fix and open a PR with no human in the loop until review. The pipeline handles both build failures (narrow diff + error logs) and server-side regressions (probabilistic causation scoring) as separate paths.
The real signal here is the architecture: decouple regression detection from coding agent execution using a statistical triage layer so your agent only fires on causal regressions, not background noise. The 7-day baseline + 60-minute post-deploy window with normalized error signatures is a concrete, replicable pattern — not a research concept. Open SWE is open-source, meaning you can swap in your own coding agent or wire this to existing infra without waiting for a vendor.
Fork the Open SWE repo this week, wire a GitHub Action to your main branch that posts your last Docker build log to Open SWE on failure — you'll have an automated PR-opening agent on build errors in under a day before writing a single line of regression detection logic.
Go to your repo on GitHub and create .github/workflows/self-heal.yml
Tags
Signals by role
Also today
Tools mentioned