Linkmortem
An open source agent that finds the dead links and dead sections in a website or in the docs of a GitHub repository, explains why each one died, recovers what the page said from the Internet Archive, finds where the content lives now and writes a patch with only the replacements it can prove. Code decides everything it can verify, and Gemma 4, an open weights model, reads only the pages code can't judge. It runs at linkmortem.tech, from the command line and as a GitHub Action that opens a pull request and keeps an issue with the report, without ever merging on its own. Submitted to the Hacktoberfest Open Source AI Challenge on DEV, where the full story of the project is told.
Metrics
- Model
- Gemma 4
- Links settled without the model
- 90%
- Costly errors of the production model
- 0
- Open models compared
- 9
- Automated tests
- 243
- Challenge
- Hacktoberfest DEV 2026
The problem
The web keeps losing its memory. 38 percent of the pages that existed in 2013 were gone ten years later, and awesome lists, READMEs and documentation slowly fill up with links that lead nowhere. Common link checkers only look at the status code, so they call a parked domain or a renamed section alive, and they never say what the lost page said or where it lives now. Fixing them by hand means checking every link, finding an archived copy and confirming the replacement is the same content.
The solution
Linkmortem performs an autopsy on every link in six stages. It finds up to 150 links, diagnoses each one down to the section it points to, recovers the last healthy copy from the Wayback Machine, looks for the content's new home, writes the patch and preserves living links that were never archived. Code settles status codes, parked domains, bot checks, Markdown anchors built the way GitHub builds them and moves announced by the archived page itself. Gemma 4 only decides whether a page that answers is alive and whether a candidate is the same page, and without evidence a link is never declared dead. MongoDB Atlas stores shared autopsies and a graveyard of resolved links searched with Vector Search, SerpApi searches live within a daily budget and Sentry traces every scan as an agent.
Technical stack
- Frontend
- React · TypeScript · Vite · Tailwind CSS · Three.js
- Backend
- Python · FastAPI · Pydantic · Gemma 4 · Ollama · Backboard · MongoDB Atlas · SerpApi · Sentry · Docker · Render
Challenges
Discovering that two models scoring 0 percent in the first evaluation had never answered, because the provider returned an error inside a successful response, and asking for JSON in the prompt to take Llama from 0 to 93 percent
Detecting dead sections, the most common break in awesome lists, by building each Markdown file's anchors the way GitHub does, with 427 links to sections across seven lists and no false ghosts
Noticing that Enzyme's archived copy announced its move while the search failed, and reading those announcements before searching to find the new home with no model and no search spent
Designing a bot that knows when to stay quiet, with one pull request that updates, whole link replacements that never touch code and no action on branches or pull requests a person touched
Closing security gaps such as addresses with credentials, invalid ports, a rate limit that a forged header could dodge and requests with no size limit
Learnings
Letting the model propose while code decides made the agent trustworthy, and almost every bug ended up moving a decision from the model into code
A number without evidence is worth nothing, and the evaluation only became useful once it recorded the reason behind every invalid answer
Measuring open models on your own cases lets you choose by costly errors and not only by accuracy
Honesty is a product feature, and saying how much was checked and what couldn't be confirmed matters as much as finding ghosts
An agent's traces tell what tests don't show, which is where the false 0 percent and a span with no input came from