The latest articles on DEV Community by Harsh Mangukia (@harshmangukia123). https://dev.to/harshmangukia123 https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fthepracticaldev.s3.amazonaws.com%2Fi%2F99mvlsfu5tfj9m7ku25d.png https://dev.to/harshmangukia123 en Harsh Mangukia Mon, 05 Oct 2026 00:36:09 +0000 https://dev.to/harshmangukia123/autonomous-repository-support-agent-2cd https://dev.to/harshmangukia123/autonomous-repository-support-agent-2cd <p><em>This is a submission for the <a href="https://dev.to/challenges/hacktoberfest-weekend-2026-10-01">Hacktoberfest Weekend Challenge: Build for a Friend</a></em></p> <h2> What I Built </h2> <p>I built an <strong>Autonomous GitHub Issue Debugger Agent</strong> for a close friend of mine who maintains an open-source library and constantly gets overwhelmed sorting through bug reports. </p> <p>Instead of manually digging through repositories or guessing error traces, my friend can trigger an autonomous agent via <strong>Temporal workflows</strong>. The agent uses <strong>Local RAG (MongoDB Atlas + Ollama)</strong> to scan the local codebase context, queries live data via <strong>SerpApi</strong> when external package updates break things, reasons through multi-step logic using <strong>Qwen via OpenRouter</strong>, and writes out a precise TypeScript code fix.</p> <h2> Code </h2> <div class="ltag-github-readme-tag"> <div class="readme-overview"> <h2> <img src="https://assets.dev.to/assets/github-logo-5a155e1f9a670af7944dd5e12375bc76ed542ea80224905ecaf878b9157cdefc.svg" alt="GitHub logo"> <a href="https://github.com/harsh-mangukia44" rel="noopener noreferrer"> harsh-mangukia44 </a> / <a href="https://github.com/harsh-mangukia44/Autonomous-Repository-Support-Agent" rel="noopener noreferrer"> Autonomous-Repository-Support-Agent </a> </h2> <h3> A TypeScript prototype that combines a Mastra AI agent, Temporal workflows, MongoDB Atlas Vector Search, Ollama embeddings, SerpApi web search, and Sentry monitoring. The agent is designed to investigate a technical issue using repository context and web search, then return a suggested fix. </h3> </div> <div class="ltag-github-body"> <div id="readme" class="md"><div class="markdown-heading"> <h1 class="heading-element">Autonomous Repository Support Agent</h1> </div> <p>A TypeScript prototype that combines a Mastra AI agent, Temporal workflows, MongoDB Atlas Vector Search, Ollama embeddings, SerpApi web search, and Sentry monitoring. The agent is designed to investigate a technical issue using repository context and web search, then return a suggested fix.</p> <div class="markdown-heading"> <h2 class="heading-element">How it works</h2> </div> <ol> <li> <code>ingest.ts</code> reads TypeScript files from <code>src/</code>, splits them into overlapping chunks, creates embeddings with the local Ollama <code>nomic-embed-text</code> model, and stores the chunks in MongoDB.</li> <li>The <code>techSupportAgent</code> in <code>src/mastra/index.ts</code> can use <ul> <li> <strong>Search Codebase</strong> to find relevant chunks in MongoDB Atlas Vector Search.</li> <li> <strong>Web Search</strong> to retrieve current technical references through SerpApi.</li> </ul> </li> <li> <code>src/workflows/workflow.ts</code> defines a Temporal workflow that calls the agent activity. Temporal retries activity failures up to five times.</li> <li> <code>src/worker.ts</code> starts the Temporal worker on the <code>agent-tasks</code> task queue and reports uncaught worker failures to Sentry.</li> </ol> <div class="markdown-heading"> <h2 class="heading-element">Requirements</h2> </div> <ul> <li>Node.js with npm</li> <li>A running Temporal server, reachable at <code>localhost:7233</code> unless configured…</li> </ul></div> </div> <div class="gh-btn-container"><a class="gh-btn" href="https://github.com/harsh-mangukia44/Autonomous-Repository-Support-Agent" rel="noopener noreferrer">View on GitHub</a></div> </div> <h2> How I Built It </h2> <ul> <li><p>Workflow Orchestration: Used Temporal to guarantee durable, reliable agent state transitions and task queues (worker.ts and trigger.ts).</p></li> <li><p>Agentic Framework: Built on Mastra to manage tools, agent instructions, and execution steps.</p></li> <li><p>Local RAG: Embedded repository files into MongoDB Atlas using RecursiveCharacterTextSplitter and local Ollama embeddings (nomic-embed-text).</p></li> <li><p>External Tooling: Integrated SerpApi for real-time web debugging when issues involve outdated third-party library specs.</p></li> <li><p>Observability: Tracked performance and execution spans using Sentry.</p></li> </ul> <h2> Why Does Open Innovation Matter? </h2> <p>Building this with open-source and open-weight AI was essential. Relying on closed, proprietary cloud APIs exposes developer repositories to strict key-validation blocks, rate limits, and opaque error structures.By keeping the embedding pipeline local and the agent orchestration open my friend maintains total privacy over their codebase context, can swap models instantly without rewriting code, and runs the entire diagnostic engine with zero licensing costs.</p> <h2> Prize Categories </h2> <p>Best Use of Gemma / Open-Weight Models<br> Best Use of Temporal<br> Best Use of MongoDB<br> Best Use of Sentry</p> devchallenge weekendchallenge hf26challenge github