{
  "id": 20877113,
  "title": "I Let AI Agents Rebrand My Entire OSS Ecosystem in One Night. Two Pipelines Were Silently Dead.",
  "url": "https://urgent.news/2026/11/20/i-let-ai-agents-rebrand-my-entire-oss-ecosystem-in-one-night-two",
  "topic": "ai",
  "section": "AI",
  "published": "2026-11-20T09:07:59.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/ofri-peretz/i-let-ai-agents-rebrand-my-entire-oss-ecosystem-in-one-night-two-pipelines-were-silently-dead-49i6"
  },
  "original_language": "en",
  "account": "My open-source software ecosystem suffered from two dead pipelines. One site served an outdated two-week-old build while every deployment ran successfully. Another repository had merged four pull requests with zero CI checks since May. The issue came to light due to a rebranding effort. Within a single night, AI agents and I merged 13 pull requests across the five repositories and repainted the ecosystem from Tailwind violet to burnt orange. A dead pipeline is revealed when a shipped item is finally seen. Dead was the ESLint documentation site's deploy workflow, which had failed on every run for two weeks. The rebranding merged the changes, with 1,580 tests passing, and production continued serving the old violet build from mid-July. The root cause was using an unpinned npx vercel, which drifted to a new major version (58.4.4) that emitted functions referencing files outside the build output directory. The deploy shipped an archive of only that directory, causing the platform to fail with ENOENT on a file never packed. The fix was to pin every invocation to the CLI of the last green deploy, vercel@54.20.1, published on 6 July. The other pipeline had its PR workflows manually disabled during a May billing crunch. Disabled workflows do not fail but produce nothing, and the four merged PRs over two months showed an inviting check mark as if a single one had passed. Running a live check that verifies the live output against production is essential to catch dead pipelines. One failure mode it catches is built ≠ delivered, where the deploy dies after the build passes. Another is disabled ≠ passing, where no checks render as good checks. The third is downstream ≠ upstream, where the fix lands where the next sync overwrites it. The agents wrote the code for the changes, but they could not notice the issues. The division of labor between the machine and human checks is crucial. Trust in the AI-generated code is based on the harness that queries production, not a feeling about the model.",
  "summary": "Two of my five repositories had delivery pipelines that were silently dead. Not failing loudly — dead . One site had served a two-week-old build while every deploy \"ran\". One repo had merged four PRs with zero CI checks since May. I found out because of a rebrand. In one night, AI agents and I merged 13 pull requests across those five repos and repainted it from Tailwind violet to burnt orange.…",
  "key_points": [
    "AI agents merged 13 pull requests across five repositories in one night",
    "Dead ESLint documentation site pipeline failed for two weeks",
    "Pinning npx vercel@54.20.1 fixed the broken deploy pipeline"
  ],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}